2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/hashtable.h>
46 #include "workqueue_sched.h"
52 * A bound gcwq is either associated or disassociated with its CPU.
53 * While associated (!DISASSOCIATED), all workers are bound to the
54 * CPU and none has %WORKER_UNBOUND set and concurrency management
57 * While DISASSOCIATED, the cpu may be offline and all workers have
58 * %WORKER_UNBOUND set and concurrency management disabled, and may
59 * be executing on any CPU. The gcwq behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex of all pools on the gcwq to avoid changing binding
63 * state while create_worker() is in progress.
65 GCWQ_DISASSOCIATED = 1 << 0, /* cpu can't serve workers */
66 GCWQ_FREEZING = 1 << 1, /* freeze in progress */
69 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
70 POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */
73 WORKER_STARTED = 1 << 0, /* started */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
80 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND |
83 NR_WORKER_POOLS = 2, /* # worker pools per gcwq */
85 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
87 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
88 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
90 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
91 /* call for help after 10ms
93 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
94 CREATE_COOLDOWN = HZ, /* time to breath after fail */
97 * Rescue workers are used only on emergencies and shared by
100 RESCUER_NICE_LEVEL = -20,
101 HIGHPRI_NICE_LEVEL = -20,
105 * Structure fields follow one of the following exclusion rules.
107 * I: Modifiable by initialization/destruction paths and read-only for
110 * P: Preemption protected. Disabling preemption is enough and should
111 * only be modified and accessed from the local cpu.
113 * L: gcwq->lock protected. Access with gcwq->lock held.
115 * X: During normal operation, modification requires gcwq->lock and
116 * should be done only from local cpu. Either disabling preemption
117 * on local cpu or grabbing gcwq->lock is enough for read access.
118 * If GCWQ_DISASSOCIATED is set, it's identical to L.
120 * F: wq->flush_mutex protected.
122 * W: workqueue_lock protected.
129 * The poor guys doing the actual heavy lifting. All on-duty workers
130 * are either serving the manager role, on idle list or on busy hash.
133 /* on idle list while idle, on busy hash table while busy */
135 struct list_head entry; /* L: while idle */
136 struct hlist_node hentry; /* L: while busy */
139 struct work_struct *current_work; /* L: work being processed */
140 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
141 struct list_head scheduled; /* L: scheduled works */
142 struct task_struct *task; /* I: worker task */
143 struct worker_pool *pool; /* I: the associated pool */
144 /* 64 bytes boundary on 64bit, 32 on 32bit */
145 unsigned long last_active; /* L: last active timestamp */
146 unsigned int flags; /* X: flags */
147 int id; /* I: worker id */
149 /* for rebinding worker to CPU */
150 struct work_struct rebind_work; /* L: for busy worker */
154 struct global_cwq *gcwq; /* I: the owning gcwq */
155 unsigned int flags; /* X: flags */
157 struct list_head worklist; /* L: list of pending works */
158 int nr_workers; /* L: total number of workers */
160 /* nr_idle includes the ones off idle_list for rebinding */
161 int nr_idle; /* L: currently idle ones */
163 struct list_head idle_list; /* X: list of idle workers */
164 struct timer_list idle_timer; /* L: worker idle timeout */
165 struct timer_list mayday_timer; /* L: SOS timer for workers */
167 struct mutex assoc_mutex; /* protect GCWQ_DISASSOCIATED */
168 struct ida worker_ida; /* L: for worker IDs */
172 * Global per-cpu workqueue. There's one and only one for each cpu
173 * and all works are queued and processed here regardless of their
177 spinlock_t lock; /* the gcwq lock */
178 unsigned int cpu; /* I: the associated cpu */
179 unsigned int flags; /* L: GCWQ_* flags */
181 /* workers are chained either in busy_hash or pool idle_list */
182 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
183 /* L: hash of busy workers */
185 struct worker_pool pools[NR_WORKER_POOLS];
186 /* normal and highpri pools */
187 } ____cacheline_aligned_in_smp;
190 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
191 * work_struct->data are used for flags and thus cwqs need to be
192 * aligned at two's power of the number of flag bits.
194 struct cpu_workqueue_struct {
195 struct worker_pool *pool; /* I: the associated pool */
196 struct workqueue_struct *wq; /* I: the owning workqueue */
197 int work_color; /* L: current color */
198 int flush_color; /* L: flushing color */
199 int nr_in_flight[WORK_NR_COLORS];
200 /* L: nr of in_flight works */
201 int nr_active; /* L: nr of active works */
202 int max_active; /* L: max active works */
203 struct list_head delayed_works; /* L: delayed works */
207 * Structure used to wait for workqueue flush.
210 struct list_head list; /* F: list of flushers */
211 int flush_color; /* F: flush color waiting for */
212 struct completion done; /* flush completion */
216 * All cpumasks are assumed to be always set on UP and thus can't be
217 * used to determine whether there's something to be done.
220 typedef cpumask_var_t mayday_mask_t;
221 #define mayday_test_and_set_cpu(cpu, mask) \
222 cpumask_test_and_set_cpu((cpu), (mask))
223 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
224 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
225 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
226 #define free_mayday_mask(mask) free_cpumask_var((mask))
228 typedef unsigned long mayday_mask_t;
229 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
230 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
231 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
232 #define alloc_mayday_mask(maskp, gfp) true
233 #define free_mayday_mask(mask) do { } while (0)
237 * The externally visible workqueue abstraction is an array of
238 * per-CPU workqueues:
240 struct workqueue_struct {
241 unsigned int flags; /* W: WQ_* flags */
243 struct cpu_workqueue_struct __percpu *pcpu;
244 struct cpu_workqueue_struct *single;
246 } cpu_wq; /* I: cwq's */
247 struct list_head list; /* W: list of all workqueues */
249 struct mutex flush_mutex; /* protects wq flushing */
250 int work_color; /* F: current work color */
251 int flush_color; /* F: current flush color */
252 atomic_t nr_cwqs_to_flush; /* flush in progress */
253 struct wq_flusher *first_flusher; /* F: first flusher */
254 struct list_head flusher_queue; /* F: flush waiters */
255 struct list_head flusher_overflow; /* F: flush overflow list */
257 mayday_mask_t mayday_mask; /* cpus requesting rescue */
258 struct worker *rescuer; /* I: rescue worker */
260 int nr_drainers; /* W: drain in progress */
261 int saved_max_active; /* W: saved cwq max_active */
262 #ifdef CONFIG_LOCKDEP
263 struct lockdep_map lockdep_map;
265 char name[]; /* I: workqueue name */
268 struct workqueue_struct *system_wq __read_mostly;
269 EXPORT_SYMBOL_GPL(system_wq);
270 struct workqueue_struct *system_highpri_wq __read_mostly;
271 EXPORT_SYMBOL_GPL(system_highpri_wq);
272 struct workqueue_struct *system_long_wq __read_mostly;
273 EXPORT_SYMBOL_GPL(system_long_wq);
274 struct workqueue_struct *system_unbound_wq __read_mostly;
275 EXPORT_SYMBOL_GPL(system_unbound_wq);
276 struct workqueue_struct *system_freezable_wq __read_mostly;
277 EXPORT_SYMBOL_GPL(system_freezable_wq);
279 #define CREATE_TRACE_POINTS
280 #include <trace/events/workqueue.h>
282 #define for_each_worker_pool(pool, gcwq) \
283 for ((pool) = &(gcwq)->pools[0]; \
284 (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
286 #define for_each_busy_worker(worker, i, pos, gcwq) \
287 hash_for_each(gcwq->busy_hash, i, pos, worker, hentry)
289 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
292 if (cpu < nr_cpu_ids) {
294 cpu = cpumask_next(cpu, mask);
295 if (cpu < nr_cpu_ids)
299 return WORK_CPU_UNBOUND;
301 return WORK_CPU_NONE;
304 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
305 struct workqueue_struct *wq)
307 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
313 * An extra gcwq is defined for an invalid cpu number
314 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
315 * specific CPU. The following iterators are similar to
316 * for_each_*_cpu() iterators but also considers the unbound gcwq.
318 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
319 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
320 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
321 * WORK_CPU_UNBOUND for unbound workqueues
323 #define for_each_gcwq_cpu(cpu) \
324 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
325 (cpu) < WORK_CPU_NONE; \
326 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
328 #define for_each_online_gcwq_cpu(cpu) \
329 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
330 (cpu) < WORK_CPU_NONE; \
331 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
333 #define for_each_cwq_cpu(cpu, wq) \
334 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
335 (cpu) < WORK_CPU_NONE; \
336 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
338 #ifdef CONFIG_DEBUG_OBJECTS_WORK
340 static struct debug_obj_descr work_debug_descr;
342 static void *work_debug_hint(void *addr)
344 return ((struct work_struct *) addr)->func;
348 * fixup_init is called when:
349 * - an active object is initialized
351 static int work_fixup_init(void *addr, enum debug_obj_state state)
353 struct work_struct *work = addr;
356 case ODEBUG_STATE_ACTIVE:
357 cancel_work_sync(work);
358 debug_object_init(work, &work_debug_descr);
366 * fixup_activate is called when:
367 * - an active object is activated
368 * - an unknown object is activated (might be a statically initialized object)
370 static int work_fixup_activate(void *addr, enum debug_obj_state state)
372 struct work_struct *work = addr;
376 case ODEBUG_STATE_NOTAVAILABLE:
378 * This is not really a fixup. The work struct was
379 * statically initialized. We just make sure that it
380 * is tracked in the object tracker.
382 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
383 debug_object_init(work, &work_debug_descr);
384 debug_object_activate(work, &work_debug_descr);
390 case ODEBUG_STATE_ACTIVE:
399 * fixup_free is called when:
400 * - an active object is freed
402 static int work_fixup_free(void *addr, enum debug_obj_state state)
404 struct work_struct *work = addr;
407 case ODEBUG_STATE_ACTIVE:
408 cancel_work_sync(work);
409 debug_object_free(work, &work_debug_descr);
416 static struct debug_obj_descr work_debug_descr = {
417 .name = "work_struct",
418 .debug_hint = work_debug_hint,
419 .fixup_init = work_fixup_init,
420 .fixup_activate = work_fixup_activate,
421 .fixup_free = work_fixup_free,
424 static inline void debug_work_activate(struct work_struct *work)
426 debug_object_activate(work, &work_debug_descr);
429 static inline void debug_work_deactivate(struct work_struct *work)
431 debug_object_deactivate(work, &work_debug_descr);
434 void __init_work(struct work_struct *work, int onstack)
437 debug_object_init_on_stack(work, &work_debug_descr);
439 debug_object_init(work, &work_debug_descr);
441 EXPORT_SYMBOL_GPL(__init_work);
443 void destroy_work_on_stack(struct work_struct *work)
445 debug_object_free(work, &work_debug_descr);
447 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
450 static inline void debug_work_activate(struct work_struct *work) { }
451 static inline void debug_work_deactivate(struct work_struct *work) { }
454 /* Serializes the accesses to the list of workqueues. */
455 static DEFINE_SPINLOCK(workqueue_lock);
456 static LIST_HEAD(workqueues);
457 static bool workqueue_freezing; /* W: have wqs started freezing? */
460 * The almighty global cpu workqueues. nr_running is the only field
461 * which is expected to be used frequently by other cpus via
462 * try_to_wake_up(). Put it in a separate cacheline.
464 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
465 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]);
468 * Global cpu workqueue and nr_running counter for unbound gcwq. The
469 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
470 * workers have WORKER_UNBOUND set.
472 static struct global_cwq unbound_global_cwq;
473 static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = {
474 [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
477 static int worker_thread(void *__worker);
479 static int worker_pool_pri(struct worker_pool *pool)
481 return pool - pool->gcwq->pools;
484 static struct global_cwq *get_gcwq(unsigned int cpu)
486 if (cpu != WORK_CPU_UNBOUND)
487 return &per_cpu(global_cwq, cpu);
489 return &unbound_global_cwq;
492 static atomic_t *get_pool_nr_running(struct worker_pool *pool)
494 int cpu = pool->gcwq->cpu;
495 int idx = worker_pool_pri(pool);
497 if (cpu != WORK_CPU_UNBOUND)
498 return &per_cpu(pool_nr_running, cpu)[idx];
500 return &unbound_pool_nr_running[idx];
503 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
504 struct workqueue_struct *wq)
506 if (!(wq->flags & WQ_UNBOUND)) {
507 if (likely(cpu < nr_cpu_ids))
508 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
509 } else if (likely(cpu == WORK_CPU_UNBOUND))
510 return wq->cpu_wq.single;
514 static unsigned int work_color_to_flags(int color)
516 return color << WORK_STRUCT_COLOR_SHIFT;
519 static int get_work_color(struct work_struct *work)
521 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
522 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
525 static int work_next_color(int color)
527 return (color + 1) % WORK_NR_COLORS;
531 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
532 * contain the pointer to the queued cwq. Once execution starts, the flag
533 * is cleared and the high bits contain OFFQ flags and CPU number.
535 * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
536 * and clear_work_data() can be used to set the cwq, cpu or clear
537 * work->data. These functions should only be called while the work is
538 * owned - ie. while the PENDING bit is set.
540 * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
541 * a work. gcwq is available once the work has been queued anywhere after
542 * initialization until it is sync canceled. cwq is available only while
543 * the work item is queued.
545 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
546 * canceled. While being canceled, a work item may have its PENDING set
547 * but stay off timer and worklist for arbitrarily long and nobody should
548 * try to steal the PENDING bit.
550 static inline void set_work_data(struct work_struct *work, unsigned long data,
553 BUG_ON(!work_pending(work));
554 atomic_long_set(&work->data, data | flags | work_static(work));
557 static void set_work_cwq(struct work_struct *work,
558 struct cpu_workqueue_struct *cwq,
559 unsigned long extra_flags)
561 set_work_data(work, (unsigned long)cwq,
562 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
565 static void set_work_cpu_and_clear_pending(struct work_struct *work,
569 * The following wmb is paired with the implied mb in
570 * test_and_set_bit(PENDING) and ensures all updates to @work made
571 * here are visible to and precede any updates by the next PENDING
575 set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0);
578 static void clear_work_data(struct work_struct *work)
580 smp_wmb(); /* see set_work_cpu_and_clear_pending() */
581 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
584 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
586 unsigned long data = atomic_long_read(&work->data);
588 if (data & WORK_STRUCT_CWQ)
589 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
594 static struct global_cwq *get_work_gcwq(struct work_struct *work)
596 unsigned long data = atomic_long_read(&work->data);
599 if (data & WORK_STRUCT_CWQ)
600 return ((struct cpu_workqueue_struct *)
601 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq;
603 cpu = data >> WORK_OFFQ_CPU_SHIFT;
604 if (cpu == WORK_CPU_NONE)
607 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
608 return get_gcwq(cpu);
611 static void mark_work_canceling(struct work_struct *work)
613 struct global_cwq *gcwq = get_work_gcwq(work);
614 unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE;
616 set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) | WORK_OFFQ_CANCELING,
617 WORK_STRUCT_PENDING);
620 static bool work_is_canceling(struct work_struct *work)
622 unsigned long data = atomic_long_read(&work->data);
624 return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
628 * Policy functions. These define the policies on how the global worker
629 * pools are managed. Unless noted otherwise, these functions assume that
630 * they're being called with gcwq->lock held.
633 static bool __need_more_worker(struct worker_pool *pool)
635 return !atomic_read(get_pool_nr_running(pool));
639 * Need to wake up a worker? Called from anything but currently
642 * Note that, because unbound workers never contribute to nr_running, this
643 * function will always return %true for unbound gcwq as long as the
644 * worklist isn't empty.
646 static bool need_more_worker(struct worker_pool *pool)
648 return !list_empty(&pool->worklist) && __need_more_worker(pool);
651 /* Can I start working? Called from busy but !running workers. */
652 static bool may_start_working(struct worker_pool *pool)
654 return pool->nr_idle;
657 /* Do I need to keep working? Called from currently running workers. */
658 static bool keep_working(struct worker_pool *pool)
660 atomic_t *nr_running = get_pool_nr_running(pool);
662 return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
665 /* Do we need a new worker? Called from manager. */
666 static bool need_to_create_worker(struct worker_pool *pool)
668 return need_more_worker(pool) && !may_start_working(pool);
671 /* Do I need to be the manager? */
672 static bool need_to_manage_workers(struct worker_pool *pool)
674 return need_to_create_worker(pool) ||
675 (pool->flags & POOL_MANAGE_WORKERS);
678 /* Do we have too many workers and should some go away? */
679 static bool too_many_workers(struct worker_pool *pool)
681 bool managing = pool->flags & POOL_MANAGING_WORKERS;
682 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
683 int nr_busy = pool->nr_workers - nr_idle;
686 * nr_idle and idle_list may disagree if idle rebinding is in
687 * progress. Never return %true if idle_list is empty.
689 if (list_empty(&pool->idle_list))
692 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
699 /* Return the first worker. Safe with preemption disabled */
700 static struct worker *first_worker(struct worker_pool *pool)
702 if (unlikely(list_empty(&pool->idle_list)))
705 return list_first_entry(&pool->idle_list, struct worker, entry);
709 * wake_up_worker - wake up an idle worker
710 * @pool: worker pool to wake worker from
712 * Wake up the first idle worker of @pool.
715 * spin_lock_irq(gcwq->lock).
717 static void wake_up_worker(struct worker_pool *pool)
719 struct worker *worker = first_worker(pool);
722 wake_up_process(worker->task);
726 * wq_worker_waking_up - a worker is waking up
727 * @task: task waking up
728 * @cpu: CPU @task is waking up to
730 * This function is called during try_to_wake_up() when a worker is
734 * spin_lock_irq(rq->lock)
736 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
738 struct worker *worker = kthread_data(task);
740 if (!(worker->flags & WORKER_NOT_RUNNING)) {
741 WARN_ON_ONCE(worker->pool->gcwq->cpu != cpu);
742 atomic_inc(get_pool_nr_running(worker->pool));
747 * wq_worker_sleeping - a worker is going to sleep
748 * @task: task going to sleep
749 * @cpu: CPU in question, must be the current CPU number
751 * This function is called during schedule() when a busy worker is
752 * going to sleep. Worker on the same cpu can be woken up by
753 * returning pointer to its task.
756 * spin_lock_irq(rq->lock)
759 * Worker task on @cpu to wake up, %NULL if none.
761 struct task_struct *wq_worker_sleeping(struct task_struct *task,
764 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
765 struct worker_pool *pool = worker->pool;
766 atomic_t *nr_running = get_pool_nr_running(pool);
768 if (worker->flags & WORKER_NOT_RUNNING)
771 /* this can only happen on the local cpu */
772 BUG_ON(cpu != raw_smp_processor_id());
775 * The counterpart of the following dec_and_test, implied mb,
776 * worklist not empty test sequence is in insert_work().
777 * Please read comment there.
779 * NOT_RUNNING is clear. This means that we're bound to and
780 * running on the local cpu w/ rq lock held and preemption
781 * disabled, which in turn means that none else could be
782 * manipulating idle_list, so dereferencing idle_list without gcwq
785 if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
786 to_wakeup = first_worker(pool);
787 return to_wakeup ? to_wakeup->task : NULL;
791 * worker_set_flags - set worker flags and adjust nr_running accordingly
793 * @flags: flags to set
794 * @wakeup: wakeup an idle worker if necessary
796 * Set @flags in @worker->flags and adjust nr_running accordingly. If
797 * nr_running becomes zero and @wakeup is %true, an idle worker is
801 * spin_lock_irq(gcwq->lock)
803 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
806 struct worker_pool *pool = worker->pool;
808 WARN_ON_ONCE(worker->task != current);
811 * If transitioning into NOT_RUNNING, adjust nr_running and
812 * wake up an idle worker as necessary if requested by
815 if ((flags & WORKER_NOT_RUNNING) &&
816 !(worker->flags & WORKER_NOT_RUNNING)) {
817 atomic_t *nr_running = get_pool_nr_running(pool);
820 if (atomic_dec_and_test(nr_running) &&
821 !list_empty(&pool->worklist))
822 wake_up_worker(pool);
824 atomic_dec(nr_running);
827 worker->flags |= flags;
831 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
833 * @flags: flags to clear
835 * Clear @flags in @worker->flags and adjust nr_running accordingly.
838 * spin_lock_irq(gcwq->lock)
840 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
842 struct worker_pool *pool = worker->pool;
843 unsigned int oflags = worker->flags;
845 WARN_ON_ONCE(worker->task != current);
847 worker->flags &= ~flags;
850 * If transitioning out of NOT_RUNNING, increment nr_running. Note
851 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
852 * of multiple flags, not a single flag.
854 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
855 if (!(worker->flags & WORKER_NOT_RUNNING))
856 atomic_inc(get_pool_nr_running(pool));
860 * find_worker_executing_work - find worker which is executing a work
861 * @gcwq: gcwq of interest
862 * @work: work to find worker for
864 * Find a worker which is executing @work on @gcwq. This function is
865 * identical to __find_worker_executing_work() except that this
866 * function calculates @bwh itself.
869 * spin_lock_irq(gcwq->lock).
872 * Pointer to worker which is executing @work if found, NULL
875 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
876 struct work_struct *work)
878 struct worker *worker;
879 struct hlist_node *tmp;
881 hash_for_each_possible(gcwq->busy_hash, worker, tmp, hentry, (unsigned long)work)
882 if (worker->current_work == work)
889 * move_linked_works - move linked works to a list
890 * @work: start of series of works to be scheduled
891 * @head: target list to append @work to
892 * @nextp: out paramter for nested worklist walking
894 * Schedule linked works starting from @work to @head. Work series to
895 * be scheduled starts at @work and includes any consecutive work with
896 * WORK_STRUCT_LINKED set in its predecessor.
898 * If @nextp is not NULL, it's updated to point to the next work of
899 * the last scheduled work. This allows move_linked_works() to be
900 * nested inside outer list_for_each_entry_safe().
903 * spin_lock_irq(gcwq->lock).
905 static void move_linked_works(struct work_struct *work, struct list_head *head,
906 struct work_struct **nextp)
908 struct work_struct *n;
911 * Linked worklist will always end before the end of the list,
912 * use NULL for list head.
914 list_for_each_entry_safe_from(work, n, NULL, entry) {
915 list_move_tail(&work->entry, head);
916 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
921 * If we're already inside safe list traversal and have moved
922 * multiple works to the scheduled queue, the next position
923 * needs to be updated.
929 static void cwq_activate_delayed_work(struct work_struct *work)
931 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
933 trace_workqueue_activate_work(work);
934 move_linked_works(work, &cwq->pool->worklist, NULL);
935 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
939 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
941 struct work_struct *work = list_first_entry(&cwq->delayed_works,
942 struct work_struct, entry);
944 cwq_activate_delayed_work(work);
948 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
949 * @cwq: cwq of interest
950 * @color: color of work which left the queue
952 * A work either has completed or is removed from pending queue,
953 * decrement nr_in_flight of its cwq and handle workqueue flushing.
956 * spin_lock_irq(gcwq->lock).
958 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
960 /* ignore uncolored works */
961 if (color == WORK_NO_COLOR)
964 cwq->nr_in_flight[color]--;
967 if (!list_empty(&cwq->delayed_works)) {
968 /* one down, submit a delayed one */
969 if (cwq->nr_active < cwq->max_active)
970 cwq_activate_first_delayed(cwq);
973 /* is flush in progress and are we at the flushing tip? */
974 if (likely(cwq->flush_color != color))
977 /* are there still in-flight works? */
978 if (cwq->nr_in_flight[color])
981 /* this cwq is done, clear flush_color */
982 cwq->flush_color = -1;
985 * If this was the last cwq, wake up the first flusher. It
986 * will handle the rest.
988 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
989 complete(&cwq->wq->first_flusher->done);
993 * try_to_grab_pending - steal work item from worklist and disable irq
994 * @work: work item to steal
995 * @is_dwork: @work is a delayed_work
996 * @flags: place to store irq state
998 * Try to grab PENDING bit of @work. This function can handle @work in any
999 * stable state - idle, on timer or on worklist. Return values are
1001 * 1 if @work was pending and we successfully stole PENDING
1002 * 0 if @work was idle and we claimed PENDING
1003 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1004 * -ENOENT if someone else is canceling @work, this state may persist
1005 * for arbitrarily long
1007 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1008 * interrupted while holding PENDING and @work off queue, irq must be
1009 * disabled on entry. This, combined with delayed_work->timer being
1010 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1012 * On successful return, >= 0, irq is disabled and the caller is
1013 * responsible for releasing it using local_irq_restore(*@flags).
1015 * This function is safe to call from any context including IRQ handler.
1017 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1018 unsigned long *flags)
1020 struct global_cwq *gcwq;
1022 local_irq_save(*flags);
1024 /* try to steal the timer if it exists */
1026 struct delayed_work *dwork = to_delayed_work(work);
1029 * dwork->timer is irqsafe. If del_timer() fails, it's
1030 * guaranteed that the timer is not queued anywhere and not
1031 * running on the local CPU.
1033 if (likely(del_timer(&dwork->timer)))
1037 /* try to claim PENDING the normal way */
1038 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1042 * The queueing is in progress, or it is already queued. Try to
1043 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1045 gcwq = get_work_gcwq(work);
1049 spin_lock(&gcwq->lock);
1050 if (!list_empty(&work->entry)) {
1052 * This work is queued, but perhaps we locked the wrong gcwq.
1053 * In that case we must see the new value after rmb(), see
1054 * insert_work()->wmb().
1057 if (gcwq == get_work_gcwq(work)) {
1058 debug_work_deactivate(work);
1061 * A delayed work item cannot be grabbed directly
1062 * because it might have linked NO_COLOR work items
1063 * which, if left on the delayed_list, will confuse
1064 * cwq->nr_active management later on and cause
1065 * stall. Make sure the work item is activated
1068 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1069 cwq_activate_delayed_work(work);
1071 list_del_init(&work->entry);
1072 cwq_dec_nr_in_flight(get_work_cwq(work),
1073 get_work_color(work));
1075 spin_unlock(&gcwq->lock);
1079 spin_unlock(&gcwq->lock);
1081 local_irq_restore(*flags);
1082 if (work_is_canceling(work))
1089 * insert_work - insert a work into gcwq
1090 * @cwq: cwq @work belongs to
1091 * @work: work to insert
1092 * @head: insertion point
1093 * @extra_flags: extra WORK_STRUCT_* flags to set
1095 * Insert @work which belongs to @cwq into @gcwq after @head.
1096 * @extra_flags is or'd to work_struct flags.
1099 * spin_lock_irq(gcwq->lock).
1101 static void insert_work(struct cpu_workqueue_struct *cwq,
1102 struct work_struct *work, struct list_head *head,
1103 unsigned int extra_flags)
1105 struct worker_pool *pool = cwq->pool;
1107 /* we own @work, set data and link */
1108 set_work_cwq(work, cwq, extra_flags);
1111 * Ensure that we get the right work->data if we see the
1112 * result of list_add() below, see try_to_grab_pending().
1116 list_add_tail(&work->entry, head);
1119 * Ensure either worker_sched_deactivated() sees the above
1120 * list_add_tail() or we see zero nr_running to avoid workers
1121 * lying around lazily while there are works to be processed.
1125 if (__need_more_worker(pool))
1126 wake_up_worker(pool);
1130 * Test whether @work is being queued from another work executing on the
1131 * same workqueue. This is rather expensive and should only be used from
1134 static bool is_chained_work(struct workqueue_struct *wq)
1136 unsigned long flags;
1139 for_each_gcwq_cpu(cpu) {
1140 struct global_cwq *gcwq = get_gcwq(cpu);
1141 struct worker *worker;
1142 struct hlist_node *pos;
1145 spin_lock_irqsave(&gcwq->lock, flags);
1146 for_each_busy_worker(worker, i, pos, gcwq) {
1147 if (worker->task != current)
1149 spin_unlock_irqrestore(&gcwq->lock, flags);
1151 * I'm @worker, no locking necessary. See if @work
1152 * is headed to the same workqueue.
1154 return worker->current_cwq->wq == wq;
1156 spin_unlock_irqrestore(&gcwq->lock, flags);
1161 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1162 struct work_struct *work)
1164 struct global_cwq *gcwq;
1165 struct cpu_workqueue_struct *cwq;
1166 struct list_head *worklist;
1167 unsigned int work_flags;
1168 unsigned int req_cpu = cpu;
1171 * While a work item is PENDING && off queue, a task trying to
1172 * steal the PENDING will busy-loop waiting for it to either get
1173 * queued or lose PENDING. Grabbing PENDING and queueing should
1174 * happen with IRQ disabled.
1176 WARN_ON_ONCE(!irqs_disabled());
1178 debug_work_activate(work);
1180 /* if dying, only works from the same workqueue are allowed */
1181 if (unlikely(wq->flags & WQ_DRAINING) &&
1182 WARN_ON_ONCE(!is_chained_work(wq)))
1185 /* determine gcwq to use */
1186 if (!(wq->flags & WQ_UNBOUND)) {
1187 struct global_cwq *last_gcwq;
1189 if (cpu == WORK_CPU_UNBOUND)
1190 cpu = raw_smp_processor_id();
1193 * It's multi cpu. If @work was previously on a different
1194 * cpu, it might still be running there, in which case the
1195 * work needs to be queued on that cpu to guarantee
1198 gcwq = get_gcwq(cpu);
1199 last_gcwq = get_work_gcwq(work);
1201 if (last_gcwq && last_gcwq != gcwq) {
1202 struct worker *worker;
1204 spin_lock(&last_gcwq->lock);
1206 worker = find_worker_executing_work(last_gcwq, work);
1208 if (worker && worker->current_cwq->wq == wq)
1211 /* meh... not running there, queue here */
1212 spin_unlock(&last_gcwq->lock);
1213 spin_lock(&gcwq->lock);
1216 spin_lock(&gcwq->lock);
1219 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1220 spin_lock(&gcwq->lock);
1223 /* gcwq determined, get cwq and queue */
1224 cwq = get_cwq(gcwq->cpu, wq);
1225 trace_workqueue_queue_work(req_cpu, cwq, work);
1227 if (WARN_ON(!list_empty(&work->entry))) {
1228 spin_unlock(&gcwq->lock);
1232 cwq->nr_in_flight[cwq->work_color]++;
1233 work_flags = work_color_to_flags(cwq->work_color);
1235 if (likely(cwq->nr_active < cwq->max_active)) {
1236 trace_workqueue_activate_work(work);
1238 worklist = &cwq->pool->worklist;
1240 work_flags |= WORK_STRUCT_DELAYED;
1241 worklist = &cwq->delayed_works;
1244 insert_work(cwq, work, worklist, work_flags);
1246 spin_unlock(&gcwq->lock);
1250 * queue_work_on - queue work on specific cpu
1251 * @cpu: CPU number to execute work on
1252 * @wq: workqueue to use
1253 * @work: work to queue
1255 * Returns %false if @work was already on a queue, %true otherwise.
1257 * We queue the work to a specific CPU, the caller must ensure it
1260 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1261 struct work_struct *work)
1264 unsigned long flags;
1266 local_irq_save(flags);
1268 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1269 __queue_work(cpu, wq, work);
1273 local_irq_restore(flags);
1276 EXPORT_SYMBOL_GPL(queue_work_on);
1279 * queue_work - queue work on a workqueue
1280 * @wq: workqueue to use
1281 * @work: work to queue
1283 * Returns %false if @work was already on a queue, %true otherwise.
1285 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1286 * it can be processed by another CPU.
1288 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1290 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1292 EXPORT_SYMBOL_GPL(queue_work);
1294 void delayed_work_timer_fn(unsigned long __data)
1296 struct delayed_work *dwork = (struct delayed_work *)__data;
1297 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1299 /* should have been called from irqsafe timer with irq already off */
1300 __queue_work(dwork->cpu, cwq->wq, &dwork->work);
1302 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1304 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1305 struct delayed_work *dwork, unsigned long delay)
1307 struct timer_list *timer = &dwork->timer;
1308 struct work_struct *work = &dwork->work;
1311 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1312 timer->data != (unsigned long)dwork);
1313 WARN_ON_ONCE(timer_pending(timer));
1314 WARN_ON_ONCE(!list_empty(&work->entry));
1317 * If @delay is 0, queue @dwork->work immediately. This is for
1318 * both optimization and correctness. The earliest @timer can
1319 * expire is on the closest next tick and delayed_work users depend
1320 * on that there's no such delay when @delay is 0.
1323 __queue_work(cpu, wq, &dwork->work);
1327 timer_stats_timer_set_start_info(&dwork->timer);
1330 * This stores cwq for the moment, for the timer_fn. Note that the
1331 * work's gcwq is preserved to allow reentrance detection for
1334 if (!(wq->flags & WQ_UNBOUND)) {
1335 struct global_cwq *gcwq = get_work_gcwq(work);
1338 * If we cannot get the last gcwq from @work directly,
1339 * select the last CPU such that it avoids unnecessarily
1340 * triggering non-reentrancy check in __queue_work().
1345 if (lcpu == WORK_CPU_UNBOUND)
1346 lcpu = raw_smp_processor_id();
1348 lcpu = WORK_CPU_UNBOUND;
1351 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1354 timer->expires = jiffies + delay;
1356 if (unlikely(cpu != WORK_CPU_UNBOUND))
1357 add_timer_on(timer, cpu);
1363 * queue_delayed_work_on - queue work on specific CPU after delay
1364 * @cpu: CPU number to execute work on
1365 * @wq: workqueue to use
1366 * @dwork: work to queue
1367 * @delay: number of jiffies to wait before queueing
1369 * Returns %false if @work was already on a queue, %true otherwise. If
1370 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1373 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1374 struct delayed_work *dwork, unsigned long delay)
1376 struct work_struct *work = &dwork->work;
1378 unsigned long flags;
1380 /* read the comment in __queue_work() */
1381 local_irq_save(flags);
1383 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1384 __queue_delayed_work(cpu, wq, dwork, delay);
1388 local_irq_restore(flags);
1391 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1394 * queue_delayed_work - queue work on a workqueue after delay
1395 * @wq: workqueue to use
1396 * @dwork: delayable work to queue
1397 * @delay: number of jiffies to wait before queueing
1399 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1401 bool queue_delayed_work(struct workqueue_struct *wq,
1402 struct delayed_work *dwork, unsigned long delay)
1404 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1406 EXPORT_SYMBOL_GPL(queue_delayed_work);
1409 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1410 * @cpu: CPU number to execute work on
1411 * @wq: workqueue to use
1412 * @dwork: work to queue
1413 * @delay: number of jiffies to wait before queueing
1415 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1416 * modify @dwork's timer so that it expires after @delay. If @delay is
1417 * zero, @work is guaranteed to be scheduled immediately regardless of its
1420 * Returns %false if @dwork was idle and queued, %true if @dwork was
1421 * pending and its timer was modified.
1423 * This function is safe to call from any context including IRQ handler.
1424 * See try_to_grab_pending() for details.
1426 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1427 struct delayed_work *dwork, unsigned long delay)
1429 unsigned long flags;
1433 ret = try_to_grab_pending(&dwork->work, true, &flags);
1434 } while (unlikely(ret == -EAGAIN));
1436 if (likely(ret >= 0)) {
1437 __queue_delayed_work(cpu, wq, dwork, delay);
1438 local_irq_restore(flags);
1441 /* -ENOENT from try_to_grab_pending() becomes %true */
1444 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1447 * mod_delayed_work - modify delay of or queue a delayed work
1448 * @wq: workqueue to use
1449 * @dwork: work to queue
1450 * @delay: number of jiffies to wait before queueing
1452 * mod_delayed_work_on() on local CPU.
1454 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1455 unsigned long delay)
1457 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1459 EXPORT_SYMBOL_GPL(mod_delayed_work);
1462 * worker_enter_idle - enter idle state
1463 * @worker: worker which is entering idle state
1465 * @worker is entering idle state. Update stats and idle timer if
1469 * spin_lock_irq(gcwq->lock).
1471 static void worker_enter_idle(struct worker *worker)
1473 struct worker_pool *pool = worker->pool;
1474 struct global_cwq *gcwq = pool->gcwq;
1476 BUG_ON(worker->flags & WORKER_IDLE);
1477 BUG_ON(!list_empty(&worker->entry) &&
1478 (worker->hentry.next || worker->hentry.pprev));
1480 /* can't use worker_set_flags(), also called from start_worker() */
1481 worker->flags |= WORKER_IDLE;
1483 worker->last_active = jiffies;
1485 /* idle_list is LIFO */
1486 list_add(&worker->entry, &pool->idle_list);
1488 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1489 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1492 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1493 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1494 * nr_running, the warning may trigger spuriously. Check iff
1495 * unbind is not in progress.
1497 WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
1498 pool->nr_workers == pool->nr_idle &&
1499 atomic_read(get_pool_nr_running(pool)));
1503 * worker_leave_idle - leave idle state
1504 * @worker: worker which is leaving idle state
1506 * @worker is leaving idle state. Update stats.
1509 * spin_lock_irq(gcwq->lock).
1511 static void worker_leave_idle(struct worker *worker)
1513 struct worker_pool *pool = worker->pool;
1515 BUG_ON(!(worker->flags & WORKER_IDLE));
1516 worker_clr_flags(worker, WORKER_IDLE);
1518 list_del_init(&worker->entry);
1522 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1525 * Works which are scheduled while the cpu is online must at least be
1526 * scheduled to a worker which is bound to the cpu so that if they are
1527 * flushed from cpu callbacks while cpu is going down, they are
1528 * guaranteed to execute on the cpu.
1530 * This function is to be used by rogue workers and rescuers to bind
1531 * themselves to the target cpu and may race with cpu going down or
1532 * coming online. kthread_bind() can't be used because it may put the
1533 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1534 * verbatim as it's best effort and blocking and gcwq may be
1535 * [dis]associated in the meantime.
1537 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1538 * binding against %GCWQ_DISASSOCIATED which is set during
1539 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1540 * enters idle state or fetches works without dropping lock, it can
1541 * guarantee the scheduling requirement described in the first paragraph.
1544 * Might sleep. Called without any lock but returns with gcwq->lock
1548 * %true if the associated gcwq is online (@worker is successfully
1549 * bound), %false if offline.
1551 static bool worker_maybe_bind_and_lock(struct worker *worker)
1552 __acquires(&gcwq->lock)
1554 struct global_cwq *gcwq = worker->pool->gcwq;
1555 struct task_struct *task = worker->task;
1559 * The following call may fail, succeed or succeed
1560 * without actually migrating the task to the cpu if
1561 * it races with cpu hotunplug operation. Verify
1562 * against GCWQ_DISASSOCIATED.
1564 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1565 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1567 spin_lock_irq(&gcwq->lock);
1568 if (gcwq->flags & GCWQ_DISASSOCIATED)
1570 if (task_cpu(task) == gcwq->cpu &&
1571 cpumask_equal(¤t->cpus_allowed,
1572 get_cpu_mask(gcwq->cpu)))
1574 spin_unlock_irq(&gcwq->lock);
1577 * We've raced with CPU hot[un]plug. Give it a breather
1578 * and retry migration. cond_resched() is required here;
1579 * otherwise, we might deadlock against cpu_stop trying to
1580 * bring down the CPU on non-preemptive kernel.
1588 * Rebind an idle @worker to its CPU. worker_thread() will test
1589 * list_empty(@worker->entry) before leaving idle and call this function.
1591 static void idle_worker_rebind(struct worker *worker)
1593 struct global_cwq *gcwq = worker->pool->gcwq;
1595 /* CPU may go down again inbetween, clear UNBOUND only on success */
1596 if (worker_maybe_bind_and_lock(worker))
1597 worker_clr_flags(worker, WORKER_UNBOUND);
1599 /* rebind complete, become available again */
1600 list_add(&worker->entry, &worker->pool->idle_list);
1601 spin_unlock_irq(&gcwq->lock);
1605 * Function for @worker->rebind.work used to rebind unbound busy workers to
1606 * the associated cpu which is coming back online. This is scheduled by
1607 * cpu up but can race with other cpu hotplug operations and may be
1608 * executed twice without intervening cpu down.
1610 static void busy_worker_rebind_fn(struct work_struct *work)
1612 struct worker *worker = container_of(work, struct worker, rebind_work);
1613 struct global_cwq *gcwq = worker->pool->gcwq;
1615 if (worker_maybe_bind_and_lock(worker))
1616 worker_clr_flags(worker, WORKER_UNBOUND);
1618 spin_unlock_irq(&gcwq->lock);
1622 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1623 * @gcwq: gcwq of interest
1625 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1626 * is different for idle and busy ones.
1628 * Idle ones will be removed from the idle_list and woken up. They will
1629 * add themselves back after completing rebind. This ensures that the
1630 * idle_list doesn't contain any unbound workers when re-bound busy workers
1631 * try to perform local wake-ups for concurrency management.
1633 * Busy workers can rebind after they finish their current work items.
1634 * Queueing the rebind work item at the head of the scheduled list is
1635 * enough. Note that nr_running will be properly bumped as busy workers
1638 * On return, all non-manager workers are scheduled for rebind - see
1639 * manage_workers() for the manager special case. Any idle worker
1640 * including the manager will not appear on @idle_list until rebind is
1641 * complete, making local wake-ups safe.
1643 static void rebind_workers(struct global_cwq *gcwq)
1645 struct worker_pool *pool;
1646 struct worker *worker, *n;
1647 struct hlist_node *pos;
1650 lockdep_assert_held(&gcwq->lock);
1652 for_each_worker_pool(pool, gcwq)
1653 lockdep_assert_held(&pool->assoc_mutex);
1655 /* dequeue and kick idle ones */
1656 for_each_worker_pool(pool, gcwq) {
1657 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1659 * idle workers should be off @pool->idle_list
1660 * until rebind is complete to avoid receiving
1661 * premature local wake-ups.
1663 list_del_init(&worker->entry);
1666 * worker_thread() will see the above dequeuing
1667 * and call idle_worker_rebind().
1669 wake_up_process(worker->task);
1673 /* rebind busy workers */
1674 for_each_busy_worker(worker, i, pos, gcwq) {
1675 struct work_struct *rebind_work = &worker->rebind_work;
1676 struct workqueue_struct *wq;
1678 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1679 work_data_bits(rebind_work)))
1682 debug_work_activate(rebind_work);
1685 * wq doesn't really matter but let's keep @worker->pool
1686 * and @cwq->pool consistent for sanity.
1688 if (worker_pool_pri(worker->pool))
1689 wq = system_highpri_wq;
1693 insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
1694 worker->scheduled.next,
1695 work_color_to_flags(WORK_NO_COLOR));
1699 static struct worker *alloc_worker(void)
1701 struct worker *worker;
1703 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1705 INIT_LIST_HEAD(&worker->entry);
1706 INIT_LIST_HEAD(&worker->scheduled);
1707 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1708 /* on creation a worker is in !idle && prep state */
1709 worker->flags = WORKER_PREP;
1715 * create_worker - create a new workqueue worker
1716 * @pool: pool the new worker will belong to
1718 * Create a new worker which is bound to @pool. The returned worker
1719 * can be started by calling start_worker() or destroyed using
1723 * Might sleep. Does GFP_KERNEL allocations.
1726 * Pointer to the newly created worker.
1728 static struct worker *create_worker(struct worker_pool *pool)
1730 struct global_cwq *gcwq = pool->gcwq;
1731 const char *pri = worker_pool_pri(pool) ? "H" : "";
1732 struct worker *worker = NULL;
1735 spin_lock_irq(&gcwq->lock);
1736 while (ida_get_new(&pool->worker_ida, &id)) {
1737 spin_unlock_irq(&gcwq->lock);
1738 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1740 spin_lock_irq(&gcwq->lock);
1742 spin_unlock_irq(&gcwq->lock);
1744 worker = alloc_worker();
1748 worker->pool = pool;
1751 if (gcwq->cpu != WORK_CPU_UNBOUND)
1752 worker->task = kthread_create_on_node(worker_thread,
1753 worker, cpu_to_node(gcwq->cpu),
1754 "kworker/%u:%d%s", gcwq->cpu, id, pri);
1756 worker->task = kthread_create(worker_thread, worker,
1757 "kworker/u:%d%s", id, pri);
1758 if (IS_ERR(worker->task))
1761 if (worker_pool_pri(pool))
1762 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1765 * Determine CPU binding of the new worker depending on
1766 * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1767 * flag remains stable across this function. See the comments
1768 * above the flag definition for details.
1770 * As an unbound worker may later become a regular one if CPU comes
1771 * online, make sure every worker has %PF_THREAD_BOUND set.
1773 if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
1774 kthread_bind(worker->task, gcwq->cpu);
1776 worker->task->flags |= PF_THREAD_BOUND;
1777 worker->flags |= WORKER_UNBOUND;
1783 spin_lock_irq(&gcwq->lock);
1784 ida_remove(&pool->worker_ida, id);
1785 spin_unlock_irq(&gcwq->lock);
1792 * start_worker - start a newly created worker
1793 * @worker: worker to start
1795 * Make the gcwq aware of @worker and start it.
1798 * spin_lock_irq(gcwq->lock).
1800 static void start_worker(struct worker *worker)
1802 worker->flags |= WORKER_STARTED;
1803 worker->pool->nr_workers++;
1804 worker_enter_idle(worker);
1805 wake_up_process(worker->task);
1809 * destroy_worker - destroy a workqueue worker
1810 * @worker: worker to be destroyed
1812 * Destroy @worker and adjust @gcwq stats accordingly.
1815 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1817 static void destroy_worker(struct worker *worker)
1819 struct worker_pool *pool = worker->pool;
1820 struct global_cwq *gcwq = pool->gcwq;
1821 int id = worker->id;
1823 /* sanity check frenzy */
1824 BUG_ON(worker->current_work);
1825 BUG_ON(!list_empty(&worker->scheduled));
1827 if (worker->flags & WORKER_STARTED)
1829 if (worker->flags & WORKER_IDLE)
1832 list_del_init(&worker->entry);
1833 worker->flags |= WORKER_DIE;
1835 spin_unlock_irq(&gcwq->lock);
1837 kthread_stop(worker->task);
1840 spin_lock_irq(&gcwq->lock);
1841 ida_remove(&pool->worker_ida, id);
1844 static void idle_worker_timeout(unsigned long __pool)
1846 struct worker_pool *pool = (void *)__pool;
1847 struct global_cwq *gcwq = pool->gcwq;
1849 spin_lock_irq(&gcwq->lock);
1851 if (too_many_workers(pool)) {
1852 struct worker *worker;
1853 unsigned long expires;
1855 /* idle_list is kept in LIFO order, check the last one */
1856 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1857 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1859 if (time_before(jiffies, expires))
1860 mod_timer(&pool->idle_timer, expires);
1862 /* it's been idle for too long, wake up manager */
1863 pool->flags |= POOL_MANAGE_WORKERS;
1864 wake_up_worker(pool);
1868 spin_unlock_irq(&gcwq->lock);
1871 static bool send_mayday(struct work_struct *work)
1873 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1874 struct workqueue_struct *wq = cwq->wq;
1877 if (!(wq->flags & WQ_RESCUER))
1880 /* mayday mayday mayday */
1881 cpu = cwq->pool->gcwq->cpu;
1882 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1883 if (cpu == WORK_CPU_UNBOUND)
1885 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1886 wake_up_process(wq->rescuer->task);
1890 static void gcwq_mayday_timeout(unsigned long __pool)
1892 struct worker_pool *pool = (void *)__pool;
1893 struct global_cwq *gcwq = pool->gcwq;
1894 struct work_struct *work;
1896 spin_lock_irq(&gcwq->lock);
1898 if (need_to_create_worker(pool)) {
1900 * We've been trying to create a new worker but
1901 * haven't been successful. We might be hitting an
1902 * allocation deadlock. Send distress signals to
1905 list_for_each_entry(work, &pool->worklist, entry)
1909 spin_unlock_irq(&gcwq->lock);
1911 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1915 * maybe_create_worker - create a new worker if necessary
1916 * @pool: pool to create a new worker for
1918 * Create a new worker for @pool if necessary. @pool is guaranteed to
1919 * have at least one idle worker on return from this function. If
1920 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1921 * sent to all rescuers with works scheduled on @pool to resolve
1922 * possible allocation deadlock.
1924 * On return, need_to_create_worker() is guaranteed to be false and
1925 * may_start_working() true.
1928 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1929 * multiple times. Does GFP_KERNEL allocations. Called only from
1933 * false if no action was taken and gcwq->lock stayed locked, true
1936 static bool maybe_create_worker(struct worker_pool *pool)
1937 __releases(&gcwq->lock)
1938 __acquires(&gcwq->lock)
1940 struct global_cwq *gcwq = pool->gcwq;
1942 if (!need_to_create_worker(pool))
1945 spin_unlock_irq(&gcwq->lock);
1947 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1948 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1951 struct worker *worker;
1953 worker = create_worker(pool);
1955 del_timer_sync(&pool->mayday_timer);
1956 spin_lock_irq(&gcwq->lock);
1957 start_worker(worker);
1958 BUG_ON(need_to_create_worker(pool));
1962 if (!need_to_create_worker(pool))
1965 __set_current_state(TASK_INTERRUPTIBLE);
1966 schedule_timeout(CREATE_COOLDOWN);
1968 if (!need_to_create_worker(pool))
1972 del_timer_sync(&pool->mayday_timer);
1973 spin_lock_irq(&gcwq->lock);
1974 if (need_to_create_worker(pool))
1980 * maybe_destroy_worker - destroy workers which have been idle for a while
1981 * @pool: pool to destroy workers for
1983 * Destroy @pool workers which have been idle for longer than
1984 * IDLE_WORKER_TIMEOUT.
1987 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1988 * multiple times. Called only from manager.
1991 * false if no action was taken and gcwq->lock stayed locked, true
1994 static bool maybe_destroy_workers(struct worker_pool *pool)
1998 while (too_many_workers(pool)) {
1999 struct worker *worker;
2000 unsigned long expires;
2002 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2003 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2005 if (time_before(jiffies, expires)) {
2006 mod_timer(&pool->idle_timer, expires);
2010 destroy_worker(worker);
2018 * manage_workers - manage worker pool
2021 * Assume the manager role and manage gcwq worker pool @worker belongs
2022 * to. At any given time, there can be only zero or one manager per
2023 * gcwq. The exclusion is handled automatically by this function.
2025 * The caller can safely start processing works on false return. On
2026 * true return, it's guaranteed that need_to_create_worker() is false
2027 * and may_start_working() is true.
2030 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2031 * multiple times. Does GFP_KERNEL allocations.
2034 * false if no action was taken and gcwq->lock stayed locked, true if
2035 * some action was taken.
2037 static bool manage_workers(struct worker *worker)
2039 struct worker_pool *pool = worker->pool;
2042 if (pool->flags & POOL_MANAGING_WORKERS)
2045 pool->flags |= POOL_MANAGING_WORKERS;
2048 * To simplify both worker management and CPU hotplug, hold off
2049 * management while hotplug is in progress. CPU hotplug path can't
2050 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2051 * lead to idle worker depletion (all become busy thinking someone
2052 * else is managing) which in turn can result in deadlock under
2053 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2054 * manager against CPU hotplug.
2056 * assoc_mutex would always be free unless CPU hotplug is in
2057 * progress. trylock first without dropping @gcwq->lock.
2059 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2060 spin_unlock_irq(&pool->gcwq->lock);
2061 mutex_lock(&pool->assoc_mutex);
2063 * CPU hotplug could have happened while we were waiting
2064 * for assoc_mutex. Hotplug itself can't handle us
2065 * because manager isn't either on idle or busy list, and
2066 * @gcwq's state and ours could have deviated.
2068 * As hotplug is now excluded via assoc_mutex, we can
2069 * simply try to bind. It will succeed or fail depending
2070 * on @gcwq's current state. Try it and adjust
2071 * %WORKER_UNBOUND accordingly.
2073 if (worker_maybe_bind_and_lock(worker))
2074 worker->flags &= ~WORKER_UNBOUND;
2076 worker->flags |= WORKER_UNBOUND;
2081 pool->flags &= ~POOL_MANAGE_WORKERS;
2084 * Destroy and then create so that may_start_working() is true
2087 ret |= maybe_destroy_workers(pool);
2088 ret |= maybe_create_worker(pool);
2090 pool->flags &= ~POOL_MANAGING_WORKERS;
2091 mutex_unlock(&pool->assoc_mutex);
2096 * process_one_work - process single work
2098 * @work: work to process
2100 * Process @work. This function contains all the logics necessary to
2101 * process a single work including synchronization against and
2102 * interaction with other workers on the same cpu, queueing and
2103 * flushing. As long as context requirement is met, any worker can
2104 * call this function to process a work.
2107 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2109 static void process_one_work(struct worker *worker, struct work_struct *work)
2110 __releases(&gcwq->lock)
2111 __acquires(&gcwq->lock)
2113 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2114 struct worker_pool *pool = worker->pool;
2115 struct global_cwq *gcwq = pool->gcwq;
2116 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2117 work_func_t f = work->func;
2119 struct worker *collision;
2120 #ifdef CONFIG_LOCKDEP
2122 * It is permissible to free the struct work_struct from
2123 * inside the function that is called from it, this we need to
2124 * take into account for lockdep too. To avoid bogus "held
2125 * lock freed" warnings as well as problems when looking into
2126 * work->lockdep_map, make a copy and use that here.
2128 struct lockdep_map lockdep_map;
2130 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2133 * Ensure we're on the correct CPU. DISASSOCIATED test is
2134 * necessary to avoid spurious warnings from rescuers servicing the
2135 * unbound or a disassociated gcwq.
2137 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2138 !(gcwq->flags & GCWQ_DISASSOCIATED) &&
2139 raw_smp_processor_id() != gcwq->cpu);
2142 * A single work shouldn't be executed concurrently by
2143 * multiple workers on a single cpu. Check whether anyone is
2144 * already processing the work. If so, defer the work to the
2145 * currently executing one.
2147 collision = find_worker_executing_work(gcwq, work);
2148 if (unlikely(collision)) {
2149 move_linked_works(work, &collision->scheduled, NULL);
2153 /* claim and dequeue */
2154 debug_work_deactivate(work);
2155 hash_add(gcwq->busy_hash, &worker->hentry, (unsigned long)worker);
2156 worker->current_work = work;
2157 worker->current_cwq = cwq;
2158 work_color = get_work_color(work);
2160 list_del_init(&work->entry);
2163 * CPU intensive works don't participate in concurrency
2164 * management. They're the scheduler's responsibility.
2166 if (unlikely(cpu_intensive))
2167 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2170 * Unbound gcwq isn't concurrency managed and work items should be
2171 * executed ASAP. Wake up another worker if necessary.
2173 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2174 wake_up_worker(pool);
2177 * Record the last CPU and clear PENDING which should be the last
2178 * update to @work. Also, do this inside @gcwq->lock so that
2179 * PENDING and queued state changes happen together while IRQ is
2182 set_work_cpu_and_clear_pending(work, gcwq->cpu);
2184 spin_unlock_irq(&gcwq->lock);
2186 lock_map_acquire_read(&cwq->wq->lockdep_map);
2187 lock_map_acquire(&lockdep_map);
2188 trace_workqueue_execute_start(work);
2191 * While we must be careful to not use "work" after this, the trace
2192 * point will only record its address.
2194 trace_workqueue_execute_end(work);
2195 lock_map_release(&lockdep_map);
2196 lock_map_release(&cwq->wq->lockdep_map);
2198 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2199 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2200 " last function: %pf\n",
2201 current->comm, preempt_count(), task_pid_nr(current), f);
2202 debug_show_held_locks(current);
2206 spin_lock_irq(&gcwq->lock);
2208 /* clear cpu intensive status */
2209 if (unlikely(cpu_intensive))
2210 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2212 /* we're done with it, release */
2213 hash_del(&worker->hentry);
2214 worker->current_work = NULL;
2215 worker->current_cwq = NULL;
2216 cwq_dec_nr_in_flight(cwq, work_color);
2220 * process_scheduled_works - process scheduled works
2223 * Process all scheduled works. Please note that the scheduled list
2224 * may change while processing a work, so this function repeatedly
2225 * fetches a work from the top and executes it.
2228 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2231 static void process_scheduled_works(struct worker *worker)
2233 while (!list_empty(&worker->scheduled)) {
2234 struct work_struct *work = list_first_entry(&worker->scheduled,
2235 struct work_struct, entry);
2236 process_one_work(worker, work);
2241 * worker_thread - the worker thread function
2244 * The gcwq worker thread function. There's a single dynamic pool of
2245 * these per each cpu. These workers process all works regardless of
2246 * their specific target workqueue. The only exception is works which
2247 * belong to workqueues with a rescuer which will be explained in
2250 static int worker_thread(void *__worker)
2252 struct worker *worker = __worker;
2253 struct worker_pool *pool = worker->pool;
2254 struct global_cwq *gcwq = pool->gcwq;
2256 /* tell the scheduler that this is a workqueue worker */
2257 worker->task->flags |= PF_WQ_WORKER;
2259 spin_lock_irq(&gcwq->lock);
2261 /* we are off idle list if destruction or rebind is requested */
2262 if (unlikely(list_empty(&worker->entry))) {
2263 spin_unlock_irq(&gcwq->lock);
2265 /* if DIE is set, destruction is requested */
2266 if (worker->flags & WORKER_DIE) {
2267 worker->task->flags &= ~PF_WQ_WORKER;
2271 /* otherwise, rebind */
2272 idle_worker_rebind(worker);
2276 worker_leave_idle(worker);
2278 /* no more worker necessary? */
2279 if (!need_more_worker(pool))
2282 /* do we need to manage? */
2283 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2287 * ->scheduled list can only be filled while a worker is
2288 * preparing to process a work or actually processing it.
2289 * Make sure nobody diddled with it while I was sleeping.
2291 BUG_ON(!list_empty(&worker->scheduled));
2294 * When control reaches this point, we're guaranteed to have
2295 * at least one idle worker or that someone else has already
2296 * assumed the manager role.
2298 worker_clr_flags(worker, WORKER_PREP);
2301 struct work_struct *work =
2302 list_first_entry(&pool->worklist,
2303 struct work_struct, entry);
2305 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2306 /* optimization path, not strictly necessary */
2307 process_one_work(worker, work);
2308 if (unlikely(!list_empty(&worker->scheduled)))
2309 process_scheduled_works(worker);
2311 move_linked_works(work, &worker->scheduled, NULL);
2312 process_scheduled_works(worker);
2314 } while (keep_working(pool));
2316 worker_set_flags(worker, WORKER_PREP, false);
2318 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2322 * gcwq->lock is held and there's no work to process and no
2323 * need to manage, sleep. Workers are woken up only while
2324 * holding gcwq->lock or from local cpu, so setting the
2325 * current state before releasing gcwq->lock is enough to
2326 * prevent losing any event.
2328 worker_enter_idle(worker);
2329 __set_current_state(TASK_INTERRUPTIBLE);
2330 spin_unlock_irq(&gcwq->lock);
2336 * rescuer_thread - the rescuer thread function
2337 * @__wq: the associated workqueue
2339 * Workqueue rescuer thread function. There's one rescuer for each
2340 * workqueue which has WQ_RESCUER set.
2342 * Regular work processing on a gcwq may block trying to create a new
2343 * worker which uses GFP_KERNEL allocation which has slight chance of
2344 * developing into deadlock if some works currently on the same queue
2345 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2346 * the problem rescuer solves.
2348 * When such condition is possible, the gcwq summons rescuers of all
2349 * workqueues which have works queued on the gcwq and let them process
2350 * those works so that forward progress can be guaranteed.
2352 * This should happen rarely.
2354 static int rescuer_thread(void *__wq)
2356 struct workqueue_struct *wq = __wq;
2357 struct worker *rescuer = wq->rescuer;
2358 struct list_head *scheduled = &rescuer->scheduled;
2359 bool is_unbound = wq->flags & WQ_UNBOUND;
2362 set_user_nice(current, RESCUER_NICE_LEVEL);
2364 set_current_state(TASK_INTERRUPTIBLE);
2366 if (kthread_should_stop()) {
2367 __set_current_state(TASK_RUNNING);
2372 * See whether any cpu is asking for help. Unbounded
2373 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2375 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2376 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2377 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2378 struct worker_pool *pool = cwq->pool;
2379 struct global_cwq *gcwq = pool->gcwq;
2380 struct work_struct *work, *n;
2382 __set_current_state(TASK_RUNNING);
2383 mayday_clear_cpu(cpu, wq->mayday_mask);
2385 /* migrate to the target cpu if possible */
2386 rescuer->pool = pool;
2387 worker_maybe_bind_and_lock(rescuer);
2390 * Slurp in all works issued via this workqueue and
2393 BUG_ON(!list_empty(&rescuer->scheduled));
2394 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2395 if (get_work_cwq(work) == cwq)
2396 move_linked_works(work, scheduled, &n);
2398 process_scheduled_works(rescuer);
2401 * Leave this gcwq. If keep_working() is %true, notify a
2402 * regular worker; otherwise, we end up with 0 concurrency
2403 * and stalling the execution.
2405 if (keep_working(pool))
2406 wake_up_worker(pool);
2408 spin_unlock_irq(&gcwq->lock);
2416 struct work_struct work;
2417 struct completion done;
2420 static void wq_barrier_func(struct work_struct *work)
2422 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2423 complete(&barr->done);
2427 * insert_wq_barrier - insert a barrier work
2428 * @cwq: cwq to insert barrier into
2429 * @barr: wq_barrier to insert
2430 * @target: target work to attach @barr to
2431 * @worker: worker currently executing @target, NULL if @target is not executing
2433 * @barr is linked to @target such that @barr is completed only after
2434 * @target finishes execution. Please note that the ordering
2435 * guarantee is observed only with respect to @target and on the local
2438 * Currently, a queued barrier can't be canceled. This is because
2439 * try_to_grab_pending() can't determine whether the work to be
2440 * grabbed is at the head of the queue and thus can't clear LINKED
2441 * flag of the previous work while there must be a valid next work
2442 * after a work with LINKED flag set.
2444 * Note that when @worker is non-NULL, @target may be modified
2445 * underneath us, so we can't reliably determine cwq from @target.
2448 * spin_lock_irq(gcwq->lock).
2450 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2451 struct wq_barrier *barr,
2452 struct work_struct *target, struct worker *worker)
2454 struct list_head *head;
2455 unsigned int linked = 0;
2458 * debugobject calls are safe here even with gcwq->lock locked
2459 * as we know for sure that this will not trigger any of the
2460 * checks and call back into the fixup functions where we
2463 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2464 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2465 init_completion(&barr->done);
2468 * If @target is currently being executed, schedule the
2469 * barrier to the worker; otherwise, put it after @target.
2472 head = worker->scheduled.next;
2474 unsigned long *bits = work_data_bits(target);
2476 head = target->entry.next;
2477 /* there can already be other linked works, inherit and set */
2478 linked = *bits & WORK_STRUCT_LINKED;
2479 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2482 debug_work_activate(&barr->work);
2483 insert_work(cwq, &barr->work, head,
2484 work_color_to_flags(WORK_NO_COLOR) | linked);
2488 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2489 * @wq: workqueue being flushed
2490 * @flush_color: new flush color, < 0 for no-op
2491 * @work_color: new work color, < 0 for no-op
2493 * Prepare cwqs for workqueue flushing.
2495 * If @flush_color is non-negative, flush_color on all cwqs should be
2496 * -1. If no cwq has in-flight commands at the specified color, all
2497 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2498 * has in flight commands, its cwq->flush_color is set to
2499 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2500 * wakeup logic is armed and %true is returned.
2502 * The caller should have initialized @wq->first_flusher prior to
2503 * calling this function with non-negative @flush_color. If
2504 * @flush_color is negative, no flush color update is done and %false
2507 * If @work_color is non-negative, all cwqs should have the same
2508 * work_color which is previous to @work_color and all will be
2509 * advanced to @work_color.
2512 * mutex_lock(wq->flush_mutex).
2515 * %true if @flush_color >= 0 and there's something to flush. %false
2518 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2519 int flush_color, int work_color)
2524 if (flush_color >= 0) {
2525 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2526 atomic_set(&wq->nr_cwqs_to_flush, 1);
2529 for_each_cwq_cpu(cpu, wq) {
2530 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2531 struct global_cwq *gcwq = cwq->pool->gcwq;
2533 spin_lock_irq(&gcwq->lock);
2535 if (flush_color >= 0) {
2536 BUG_ON(cwq->flush_color != -1);
2538 if (cwq->nr_in_flight[flush_color]) {
2539 cwq->flush_color = flush_color;
2540 atomic_inc(&wq->nr_cwqs_to_flush);
2545 if (work_color >= 0) {
2546 BUG_ON(work_color != work_next_color(cwq->work_color));
2547 cwq->work_color = work_color;
2550 spin_unlock_irq(&gcwq->lock);
2553 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2554 complete(&wq->first_flusher->done);
2560 * flush_workqueue - ensure that any scheduled work has run to completion.
2561 * @wq: workqueue to flush
2563 * Forces execution of the workqueue and blocks until its completion.
2564 * This is typically used in driver shutdown handlers.
2566 * We sleep until all works which were queued on entry have been handled,
2567 * but we are not livelocked by new incoming ones.
2569 void flush_workqueue(struct workqueue_struct *wq)
2571 struct wq_flusher this_flusher = {
2572 .list = LIST_HEAD_INIT(this_flusher.list),
2574 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2578 lock_map_acquire(&wq->lockdep_map);
2579 lock_map_release(&wq->lockdep_map);
2581 mutex_lock(&wq->flush_mutex);
2584 * Start-to-wait phase
2586 next_color = work_next_color(wq->work_color);
2588 if (next_color != wq->flush_color) {
2590 * Color space is not full. The current work_color
2591 * becomes our flush_color and work_color is advanced
2594 BUG_ON(!list_empty(&wq->flusher_overflow));
2595 this_flusher.flush_color = wq->work_color;
2596 wq->work_color = next_color;
2598 if (!wq->first_flusher) {
2599 /* no flush in progress, become the first flusher */
2600 BUG_ON(wq->flush_color != this_flusher.flush_color);
2602 wq->first_flusher = &this_flusher;
2604 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2606 /* nothing to flush, done */
2607 wq->flush_color = next_color;
2608 wq->first_flusher = NULL;
2613 BUG_ON(wq->flush_color == this_flusher.flush_color);
2614 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2615 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2619 * Oops, color space is full, wait on overflow queue.
2620 * The next flush completion will assign us
2621 * flush_color and transfer to flusher_queue.
2623 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2626 mutex_unlock(&wq->flush_mutex);
2628 wait_for_completion(&this_flusher.done);
2631 * Wake-up-and-cascade phase
2633 * First flushers are responsible for cascading flushes and
2634 * handling overflow. Non-first flushers can simply return.
2636 if (wq->first_flusher != &this_flusher)
2639 mutex_lock(&wq->flush_mutex);
2641 /* we might have raced, check again with mutex held */
2642 if (wq->first_flusher != &this_flusher)
2645 wq->first_flusher = NULL;
2647 BUG_ON(!list_empty(&this_flusher.list));
2648 BUG_ON(wq->flush_color != this_flusher.flush_color);
2651 struct wq_flusher *next, *tmp;
2653 /* complete all the flushers sharing the current flush color */
2654 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2655 if (next->flush_color != wq->flush_color)
2657 list_del_init(&next->list);
2658 complete(&next->done);
2661 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2662 wq->flush_color != work_next_color(wq->work_color));
2664 /* this flush_color is finished, advance by one */
2665 wq->flush_color = work_next_color(wq->flush_color);
2667 /* one color has been freed, handle overflow queue */
2668 if (!list_empty(&wq->flusher_overflow)) {
2670 * Assign the same color to all overflowed
2671 * flushers, advance work_color and append to
2672 * flusher_queue. This is the start-to-wait
2673 * phase for these overflowed flushers.
2675 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2676 tmp->flush_color = wq->work_color;
2678 wq->work_color = work_next_color(wq->work_color);
2680 list_splice_tail_init(&wq->flusher_overflow,
2681 &wq->flusher_queue);
2682 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2685 if (list_empty(&wq->flusher_queue)) {
2686 BUG_ON(wq->flush_color != wq->work_color);
2691 * Need to flush more colors. Make the next flusher
2692 * the new first flusher and arm cwqs.
2694 BUG_ON(wq->flush_color == wq->work_color);
2695 BUG_ON(wq->flush_color != next->flush_color);
2697 list_del_init(&next->list);
2698 wq->first_flusher = next;
2700 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2704 * Meh... this color is already done, clear first
2705 * flusher and repeat cascading.
2707 wq->first_flusher = NULL;
2711 mutex_unlock(&wq->flush_mutex);
2713 EXPORT_SYMBOL_GPL(flush_workqueue);
2716 * drain_workqueue - drain a workqueue
2717 * @wq: workqueue to drain
2719 * Wait until the workqueue becomes empty. While draining is in progress,
2720 * only chain queueing is allowed. IOW, only currently pending or running
2721 * work items on @wq can queue further work items on it. @wq is flushed
2722 * repeatedly until it becomes empty. The number of flushing is detemined
2723 * by the depth of chaining and should be relatively short. Whine if it
2726 void drain_workqueue(struct workqueue_struct *wq)
2728 unsigned int flush_cnt = 0;
2732 * __queue_work() needs to test whether there are drainers, is much
2733 * hotter than drain_workqueue() and already looks at @wq->flags.
2734 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2736 spin_lock(&workqueue_lock);
2737 if (!wq->nr_drainers++)
2738 wq->flags |= WQ_DRAINING;
2739 spin_unlock(&workqueue_lock);
2741 flush_workqueue(wq);
2743 for_each_cwq_cpu(cpu, wq) {
2744 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2747 spin_lock_irq(&cwq->pool->gcwq->lock);
2748 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2749 spin_unlock_irq(&cwq->pool->gcwq->lock);
2754 if (++flush_cnt == 10 ||
2755 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2756 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2757 wq->name, flush_cnt);
2761 spin_lock(&workqueue_lock);
2762 if (!--wq->nr_drainers)
2763 wq->flags &= ~WQ_DRAINING;
2764 spin_unlock(&workqueue_lock);
2766 EXPORT_SYMBOL_GPL(drain_workqueue);
2768 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2770 struct worker *worker = NULL;
2771 struct global_cwq *gcwq;
2772 struct cpu_workqueue_struct *cwq;
2775 gcwq = get_work_gcwq(work);
2779 spin_lock_irq(&gcwq->lock);
2780 if (!list_empty(&work->entry)) {
2782 * See the comment near try_to_grab_pending()->smp_rmb().
2783 * If it was re-queued to a different gcwq under us, we
2784 * are not going to wait.
2787 cwq = get_work_cwq(work);
2788 if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
2791 worker = find_worker_executing_work(gcwq, work);
2794 cwq = worker->current_cwq;
2797 insert_wq_barrier(cwq, barr, work, worker);
2798 spin_unlock_irq(&gcwq->lock);
2801 * If @max_active is 1 or rescuer is in use, flushing another work
2802 * item on the same workqueue may lead to deadlock. Make sure the
2803 * flusher is not running on the same workqueue by verifying write
2806 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2807 lock_map_acquire(&cwq->wq->lockdep_map);
2809 lock_map_acquire_read(&cwq->wq->lockdep_map);
2810 lock_map_release(&cwq->wq->lockdep_map);
2814 spin_unlock_irq(&gcwq->lock);
2819 * flush_work - wait for a work to finish executing the last queueing instance
2820 * @work: the work to flush
2822 * Wait until @work has finished execution. @work is guaranteed to be idle
2823 * on return if it hasn't been requeued since flush started.
2826 * %true if flush_work() waited for the work to finish execution,
2827 * %false if it was already idle.
2829 bool flush_work(struct work_struct *work)
2831 struct wq_barrier barr;
2833 lock_map_acquire(&work->lockdep_map);
2834 lock_map_release(&work->lockdep_map);
2836 if (start_flush_work(work, &barr)) {
2837 wait_for_completion(&barr.done);
2838 destroy_work_on_stack(&barr.work);
2844 EXPORT_SYMBOL_GPL(flush_work);
2846 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2848 unsigned long flags;
2852 ret = try_to_grab_pending(work, is_dwork, &flags);
2854 * If someone else is canceling, wait for the same event it
2855 * would be waiting for before retrying.
2857 if (unlikely(ret == -ENOENT))
2859 } while (unlikely(ret < 0));
2861 /* tell other tasks trying to grab @work to back off */
2862 mark_work_canceling(work);
2863 local_irq_restore(flags);
2866 clear_work_data(work);
2871 * cancel_work_sync - cancel a work and wait for it to finish
2872 * @work: the work to cancel
2874 * Cancel @work and wait for its execution to finish. This function
2875 * can be used even if the work re-queues itself or migrates to
2876 * another workqueue. On return from this function, @work is
2877 * guaranteed to be not pending or executing on any CPU.
2879 * cancel_work_sync(&delayed_work->work) must not be used for
2880 * delayed_work's. Use cancel_delayed_work_sync() instead.
2882 * The caller must ensure that the workqueue on which @work was last
2883 * queued can't be destroyed before this function returns.
2886 * %true if @work was pending, %false otherwise.
2888 bool cancel_work_sync(struct work_struct *work)
2890 return __cancel_work_timer(work, false);
2892 EXPORT_SYMBOL_GPL(cancel_work_sync);
2895 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2896 * @dwork: the delayed work to flush
2898 * Delayed timer is cancelled and the pending work is queued for
2899 * immediate execution. Like flush_work(), this function only
2900 * considers the last queueing instance of @dwork.
2903 * %true if flush_work() waited for the work to finish execution,
2904 * %false if it was already idle.
2906 bool flush_delayed_work(struct delayed_work *dwork)
2908 local_irq_disable();
2909 if (del_timer_sync(&dwork->timer))
2910 __queue_work(dwork->cpu,
2911 get_work_cwq(&dwork->work)->wq, &dwork->work);
2913 return flush_work(&dwork->work);
2915 EXPORT_SYMBOL(flush_delayed_work);
2918 * cancel_delayed_work - cancel a delayed work
2919 * @dwork: delayed_work to cancel
2921 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2922 * and canceled; %false if wasn't pending. Note that the work callback
2923 * function may still be running on return, unless it returns %true and the
2924 * work doesn't re-arm itself. Explicitly flush or use
2925 * cancel_delayed_work_sync() to wait on it.
2927 * This function is safe to call from any context including IRQ handler.
2929 bool cancel_delayed_work(struct delayed_work *dwork)
2931 unsigned long flags;
2935 ret = try_to_grab_pending(&dwork->work, true, &flags);
2936 } while (unlikely(ret == -EAGAIN));
2938 if (unlikely(ret < 0))
2941 set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work));
2942 local_irq_restore(flags);
2945 EXPORT_SYMBOL(cancel_delayed_work);
2948 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2949 * @dwork: the delayed work cancel
2951 * This is cancel_work_sync() for delayed works.
2954 * %true if @dwork was pending, %false otherwise.
2956 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2958 return __cancel_work_timer(&dwork->work, true);
2960 EXPORT_SYMBOL(cancel_delayed_work_sync);
2963 * schedule_work_on - put work task on a specific cpu
2964 * @cpu: cpu to put the work task on
2965 * @work: job to be done
2967 * This puts a job on a specific cpu
2969 bool schedule_work_on(int cpu, struct work_struct *work)
2971 return queue_work_on(cpu, system_wq, work);
2973 EXPORT_SYMBOL(schedule_work_on);
2976 * schedule_work - put work task in global workqueue
2977 * @work: job to be done
2979 * Returns %false if @work was already on the kernel-global workqueue and
2982 * This puts a job in the kernel-global workqueue if it was not already
2983 * queued and leaves it in the same position on the kernel-global
2984 * workqueue otherwise.
2986 bool schedule_work(struct work_struct *work)
2988 return queue_work(system_wq, work);
2990 EXPORT_SYMBOL(schedule_work);
2993 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2995 * @dwork: job to be done
2996 * @delay: number of jiffies to wait
2998 * After waiting for a given time this puts a job in the kernel-global
2999 * workqueue on the specified CPU.
3001 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3002 unsigned long delay)
3004 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3006 EXPORT_SYMBOL(schedule_delayed_work_on);
3009 * schedule_delayed_work - put work task in global workqueue after delay
3010 * @dwork: job to be done
3011 * @delay: number of jiffies to wait or 0 for immediate execution
3013 * After waiting for a given time this puts a job in the kernel-global
3016 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3018 return queue_delayed_work(system_wq, dwork, delay);
3020 EXPORT_SYMBOL(schedule_delayed_work);
3023 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3024 * @func: the function to call
3026 * schedule_on_each_cpu() executes @func on each online CPU using the
3027 * system workqueue and blocks until all CPUs have completed.
3028 * schedule_on_each_cpu() is very slow.
3031 * 0 on success, -errno on failure.
3033 int schedule_on_each_cpu(work_func_t func)
3036 struct work_struct __percpu *works;
3038 works = alloc_percpu(struct work_struct);
3044 for_each_online_cpu(cpu) {
3045 struct work_struct *work = per_cpu_ptr(works, cpu);
3047 INIT_WORK(work, func);
3048 schedule_work_on(cpu, work);
3051 for_each_online_cpu(cpu)
3052 flush_work(per_cpu_ptr(works, cpu));
3060 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3062 * Forces execution of the kernel-global workqueue and blocks until its
3065 * Think twice before calling this function! It's very easy to get into
3066 * trouble if you don't take great care. Either of the following situations
3067 * will lead to deadlock:
3069 * One of the work items currently on the workqueue needs to acquire
3070 * a lock held by your code or its caller.
3072 * Your code is running in the context of a work routine.
3074 * They will be detected by lockdep when they occur, but the first might not
3075 * occur very often. It depends on what work items are on the workqueue and
3076 * what locks they need, which you have no control over.
3078 * In most situations flushing the entire workqueue is overkill; you merely
3079 * need to know that a particular work item isn't queued and isn't running.
3080 * In such cases you should use cancel_delayed_work_sync() or
3081 * cancel_work_sync() instead.
3083 void flush_scheduled_work(void)
3085 flush_workqueue(system_wq);
3087 EXPORT_SYMBOL(flush_scheduled_work);
3090 * execute_in_process_context - reliably execute the routine with user context
3091 * @fn: the function to execute
3092 * @ew: guaranteed storage for the execute work structure (must
3093 * be available when the work executes)
3095 * Executes the function immediately if process context is available,
3096 * otherwise schedules the function for delayed execution.
3098 * Returns: 0 - function was executed
3099 * 1 - function was scheduled for execution
3101 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3103 if (!in_interrupt()) {
3108 INIT_WORK(&ew->work, fn);
3109 schedule_work(&ew->work);
3113 EXPORT_SYMBOL_GPL(execute_in_process_context);
3115 int keventd_up(void)
3117 return system_wq != NULL;
3120 static int alloc_cwqs(struct workqueue_struct *wq)
3123 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3124 * Make sure that the alignment isn't lower than that of
3125 * unsigned long long.
3127 const size_t size = sizeof(struct cpu_workqueue_struct);
3128 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3129 __alignof__(unsigned long long));
3131 if (!(wq->flags & WQ_UNBOUND))
3132 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3137 * Allocate enough room to align cwq and put an extra
3138 * pointer at the end pointing back to the originally
3139 * allocated pointer which will be used for free.
3141 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3143 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3144 *(void **)(wq->cpu_wq.single + 1) = ptr;
3148 /* just in case, make sure it's actually aligned */
3149 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3150 return wq->cpu_wq.v ? 0 : -ENOMEM;
3153 static void free_cwqs(struct workqueue_struct *wq)
3155 if (!(wq->flags & WQ_UNBOUND))
3156 free_percpu(wq->cpu_wq.pcpu);
3157 else if (wq->cpu_wq.single) {
3158 /* the pointer to free is stored right after the cwq */
3159 kfree(*(void **)(wq->cpu_wq.single + 1));
3163 static int wq_clamp_max_active(int max_active, unsigned int flags,
3166 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3168 if (max_active < 1 || max_active > lim)
3169 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3170 max_active, name, 1, lim);
3172 return clamp_val(max_active, 1, lim);
3175 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3178 struct lock_class_key *key,
3179 const char *lock_name, ...)
3181 va_list args, args1;
3182 struct workqueue_struct *wq;
3186 /* determine namelen, allocate wq and format name */
3187 va_start(args, lock_name);
3188 va_copy(args1, args);
3189 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3191 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3195 vsnprintf(wq->name, namelen, fmt, args1);
3200 * Workqueues which may be used during memory reclaim should
3201 * have a rescuer to guarantee forward progress.
3203 if (flags & WQ_MEM_RECLAIM)
3204 flags |= WQ_RESCUER;
3206 max_active = max_active ?: WQ_DFL_ACTIVE;
3207 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3211 wq->saved_max_active = max_active;
3212 mutex_init(&wq->flush_mutex);
3213 atomic_set(&wq->nr_cwqs_to_flush, 0);
3214 INIT_LIST_HEAD(&wq->flusher_queue);
3215 INIT_LIST_HEAD(&wq->flusher_overflow);
3217 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3218 INIT_LIST_HEAD(&wq->list);
3220 if (alloc_cwqs(wq) < 0)
3223 for_each_cwq_cpu(cpu, wq) {
3224 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3225 struct global_cwq *gcwq = get_gcwq(cpu);
3226 int pool_idx = (bool)(flags & WQ_HIGHPRI);
3228 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3229 cwq->pool = &gcwq->pools[pool_idx];
3231 cwq->flush_color = -1;
3232 cwq->max_active = max_active;
3233 INIT_LIST_HEAD(&cwq->delayed_works);
3236 if (flags & WQ_RESCUER) {
3237 struct worker *rescuer;
3239 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3242 wq->rescuer = rescuer = alloc_worker();
3246 rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3248 if (IS_ERR(rescuer->task))
3251 rescuer->task->flags |= PF_THREAD_BOUND;
3252 wake_up_process(rescuer->task);
3256 * workqueue_lock protects global freeze state and workqueues
3257 * list. Grab it, set max_active accordingly and add the new
3258 * workqueue to workqueues list.
3260 spin_lock(&workqueue_lock);
3262 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3263 for_each_cwq_cpu(cpu, wq)
3264 get_cwq(cpu, wq)->max_active = 0;
3266 list_add(&wq->list, &workqueues);
3268 spin_unlock(&workqueue_lock);
3274 free_mayday_mask(wq->mayday_mask);
3280 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3283 * destroy_workqueue - safely terminate a workqueue
3284 * @wq: target workqueue
3286 * Safely destroy a workqueue. All work currently pending will be done first.
3288 void destroy_workqueue(struct workqueue_struct *wq)
3292 /* drain it before proceeding with destruction */
3293 drain_workqueue(wq);
3296 * wq list is used to freeze wq, remove from list after
3297 * flushing is complete in case freeze races us.
3299 spin_lock(&workqueue_lock);
3300 list_del(&wq->list);
3301 spin_unlock(&workqueue_lock);
3304 for_each_cwq_cpu(cpu, wq) {
3305 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3308 for (i = 0; i < WORK_NR_COLORS; i++)
3309 BUG_ON(cwq->nr_in_flight[i]);
3310 BUG_ON(cwq->nr_active);
3311 BUG_ON(!list_empty(&cwq->delayed_works));
3314 if (wq->flags & WQ_RESCUER) {
3315 kthread_stop(wq->rescuer->task);
3316 free_mayday_mask(wq->mayday_mask);
3323 EXPORT_SYMBOL_GPL(destroy_workqueue);
3326 * cwq_set_max_active - adjust max_active of a cwq
3327 * @cwq: target cpu_workqueue_struct
3328 * @max_active: new max_active value.
3330 * Set @cwq->max_active to @max_active and activate delayed works if
3334 * spin_lock_irq(gcwq->lock).
3336 static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
3338 cwq->max_active = max_active;
3340 while (!list_empty(&cwq->delayed_works) &&
3341 cwq->nr_active < cwq->max_active)
3342 cwq_activate_first_delayed(cwq);
3346 * workqueue_set_max_active - adjust max_active of a workqueue
3347 * @wq: target workqueue
3348 * @max_active: new max_active value.
3350 * Set max_active of @wq to @max_active.
3353 * Don't call from IRQ context.
3355 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3359 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3361 spin_lock(&workqueue_lock);
3363 wq->saved_max_active = max_active;
3365 for_each_cwq_cpu(cpu, wq) {
3366 struct global_cwq *gcwq = get_gcwq(cpu);
3368 spin_lock_irq(&gcwq->lock);
3370 if (!(wq->flags & WQ_FREEZABLE) ||
3371 !(gcwq->flags & GCWQ_FREEZING))
3372 cwq_set_max_active(get_cwq(gcwq->cpu, wq), max_active);
3374 spin_unlock_irq(&gcwq->lock);
3377 spin_unlock(&workqueue_lock);
3379 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3382 * workqueue_congested - test whether a workqueue is congested
3383 * @cpu: CPU in question
3384 * @wq: target workqueue
3386 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3387 * no synchronization around this function and the test result is
3388 * unreliable and only useful as advisory hints or for debugging.
3391 * %true if congested, %false otherwise.
3393 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3395 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3397 return !list_empty(&cwq->delayed_works);
3399 EXPORT_SYMBOL_GPL(workqueue_congested);
3402 * work_cpu - return the last known associated cpu for @work
3403 * @work: the work of interest
3406 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3408 unsigned int work_cpu(struct work_struct *work)
3410 struct global_cwq *gcwq = get_work_gcwq(work);
3412 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3414 EXPORT_SYMBOL_GPL(work_cpu);
3417 * work_busy - test whether a work is currently pending or running
3418 * @work: the work to be tested
3420 * Test whether @work is currently pending or running. There is no
3421 * synchronization around this function and the test result is
3422 * unreliable and only useful as advisory hints or for debugging.
3423 * Especially for reentrant wqs, the pending state might hide the
3427 * OR'd bitmask of WORK_BUSY_* bits.
3429 unsigned int work_busy(struct work_struct *work)
3431 struct global_cwq *gcwq = get_work_gcwq(work);
3432 unsigned long flags;
3433 unsigned int ret = 0;
3438 spin_lock_irqsave(&gcwq->lock, flags);
3440 if (work_pending(work))
3441 ret |= WORK_BUSY_PENDING;
3442 if (find_worker_executing_work(gcwq, work))
3443 ret |= WORK_BUSY_RUNNING;
3445 spin_unlock_irqrestore(&gcwq->lock, flags);
3449 EXPORT_SYMBOL_GPL(work_busy);
3454 * There are two challenges in supporting CPU hotplug. Firstly, there
3455 * are a lot of assumptions on strong associations among work, cwq and
3456 * gcwq which make migrating pending and scheduled works very
3457 * difficult to implement without impacting hot paths. Secondly,
3458 * gcwqs serve mix of short, long and very long running works making
3459 * blocked draining impractical.
3461 * This is solved by allowing a gcwq to be disassociated from the CPU
3462 * running as an unbound one and allowing it to be reattached later if the
3463 * cpu comes back online.
3466 /* claim manager positions of all pools */
3467 static void gcwq_claim_assoc_and_lock(struct global_cwq *gcwq)
3469 struct worker_pool *pool;
3471 for_each_worker_pool(pool, gcwq)
3472 mutex_lock_nested(&pool->assoc_mutex, pool - gcwq->pools);
3473 spin_lock_irq(&gcwq->lock);
3476 /* release manager positions */
3477 static void gcwq_release_assoc_and_unlock(struct global_cwq *gcwq)
3479 struct worker_pool *pool;
3481 spin_unlock_irq(&gcwq->lock);
3482 for_each_worker_pool(pool, gcwq)
3483 mutex_unlock(&pool->assoc_mutex);
3486 static void gcwq_unbind_fn(struct work_struct *work)
3488 struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3489 struct worker_pool *pool;
3490 struct worker *worker;
3491 struct hlist_node *pos;
3494 BUG_ON(gcwq->cpu != smp_processor_id());
3496 gcwq_claim_assoc_and_lock(gcwq);
3499 * We've claimed all manager positions. Make all workers unbound
3500 * and set DISASSOCIATED. Before this, all workers except for the
3501 * ones which are still executing works from before the last CPU
3502 * down must be on the cpu. After this, they may become diasporas.
3504 for_each_worker_pool(pool, gcwq)
3505 list_for_each_entry(worker, &pool->idle_list, entry)
3506 worker->flags |= WORKER_UNBOUND;
3508 for_each_busy_worker(worker, i, pos, gcwq)
3509 worker->flags |= WORKER_UNBOUND;
3511 gcwq->flags |= GCWQ_DISASSOCIATED;
3513 gcwq_release_assoc_and_unlock(gcwq);
3516 * Call schedule() so that we cross rq->lock and thus can guarantee
3517 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3518 * as scheduler callbacks may be invoked from other cpus.
3523 * Sched callbacks are disabled now. Zap nr_running. After this,
3524 * nr_running stays zero and need_more_worker() and keep_working()
3525 * are always true as long as the worklist is not empty. @gcwq now
3526 * behaves as unbound (in terms of concurrency management) gcwq
3527 * which is served by workers tied to the CPU.
3529 * On return from this function, the current worker would trigger
3530 * unbound chain execution of pending work items if other workers
3533 for_each_worker_pool(pool, gcwq)
3534 atomic_set(get_pool_nr_running(pool), 0);
3538 * Workqueues should be brought up before normal priority CPU notifiers.
3539 * This will be registered high priority CPU notifier.
3541 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3542 unsigned long action,
3545 unsigned int cpu = (unsigned long)hcpu;
3546 struct global_cwq *gcwq = get_gcwq(cpu);
3547 struct worker_pool *pool;
3549 switch (action & ~CPU_TASKS_FROZEN) {
3550 case CPU_UP_PREPARE:
3551 for_each_worker_pool(pool, gcwq) {
3552 struct worker *worker;
3554 if (pool->nr_workers)
3557 worker = create_worker(pool);
3561 spin_lock_irq(&gcwq->lock);
3562 start_worker(worker);
3563 spin_unlock_irq(&gcwq->lock);
3567 case CPU_DOWN_FAILED:
3569 gcwq_claim_assoc_and_lock(gcwq);
3570 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3571 rebind_workers(gcwq);
3572 gcwq_release_assoc_and_unlock(gcwq);
3579 * Workqueues should be brought down after normal priority CPU notifiers.
3580 * This will be registered as low priority CPU notifier.
3582 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3583 unsigned long action,
3586 unsigned int cpu = (unsigned long)hcpu;
3587 struct work_struct unbind_work;
3589 switch (action & ~CPU_TASKS_FROZEN) {
3590 case CPU_DOWN_PREPARE:
3591 /* unbinding should happen on the local CPU */
3592 INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3593 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3594 flush_work(&unbind_work);
3602 struct work_for_cpu {
3603 struct work_struct work;
3609 static void work_for_cpu_fn(struct work_struct *work)
3611 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3613 wfc->ret = wfc->fn(wfc->arg);
3617 * work_on_cpu - run a function in user context on a particular cpu
3618 * @cpu: the cpu to run on
3619 * @fn: the function to run
3620 * @arg: the function arg
3622 * This will return the value @fn returns.
3623 * It is up to the caller to ensure that the cpu doesn't go offline.
3624 * The caller must not hold any locks which would prevent @fn from completing.
3626 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3628 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3630 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3631 schedule_work_on(cpu, &wfc.work);
3632 flush_work(&wfc.work);
3635 EXPORT_SYMBOL_GPL(work_on_cpu);
3636 #endif /* CONFIG_SMP */
3638 #ifdef CONFIG_FREEZER
3641 * freeze_workqueues_begin - begin freezing workqueues
3643 * Start freezing workqueues. After this function returns, all freezable
3644 * workqueues will queue new works to their frozen_works list instead of
3648 * Grabs and releases workqueue_lock and gcwq->lock's.
3650 void freeze_workqueues_begin(void)
3654 spin_lock(&workqueue_lock);
3656 BUG_ON(workqueue_freezing);
3657 workqueue_freezing = true;
3659 for_each_gcwq_cpu(cpu) {
3660 struct global_cwq *gcwq = get_gcwq(cpu);
3661 struct workqueue_struct *wq;
3663 spin_lock_irq(&gcwq->lock);
3665 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3666 gcwq->flags |= GCWQ_FREEZING;
3668 list_for_each_entry(wq, &workqueues, list) {
3669 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3671 if (cwq && wq->flags & WQ_FREEZABLE)
3672 cwq->max_active = 0;
3675 spin_unlock_irq(&gcwq->lock);
3678 spin_unlock(&workqueue_lock);
3682 * freeze_workqueues_busy - are freezable workqueues still busy?
3684 * Check whether freezing is complete. This function must be called
3685 * between freeze_workqueues_begin() and thaw_workqueues().
3688 * Grabs and releases workqueue_lock.
3691 * %true if some freezable workqueues are still busy. %false if freezing
3694 bool freeze_workqueues_busy(void)
3699 spin_lock(&workqueue_lock);
3701 BUG_ON(!workqueue_freezing);
3703 for_each_gcwq_cpu(cpu) {
3704 struct workqueue_struct *wq;
3706 * nr_active is monotonically decreasing. It's safe
3707 * to peek without lock.
3709 list_for_each_entry(wq, &workqueues, list) {
3710 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3712 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3715 BUG_ON(cwq->nr_active < 0);
3716 if (cwq->nr_active) {
3723 spin_unlock(&workqueue_lock);
3728 * thaw_workqueues - thaw workqueues
3730 * Thaw workqueues. Normal queueing is restored and all collected
3731 * frozen works are transferred to their respective gcwq worklists.
3734 * Grabs and releases workqueue_lock and gcwq->lock's.
3736 void thaw_workqueues(void)
3740 spin_lock(&workqueue_lock);
3742 if (!workqueue_freezing)
3745 for_each_gcwq_cpu(cpu) {
3746 struct global_cwq *gcwq = get_gcwq(cpu);
3747 struct worker_pool *pool;
3748 struct workqueue_struct *wq;
3750 spin_lock_irq(&gcwq->lock);
3752 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3753 gcwq->flags &= ~GCWQ_FREEZING;
3755 list_for_each_entry(wq, &workqueues, list) {
3756 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3758 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3761 /* restore max_active and repopulate worklist */
3762 cwq_set_max_active(cwq, wq->saved_max_active);
3765 for_each_worker_pool(pool, gcwq)
3766 wake_up_worker(pool);
3768 spin_unlock_irq(&gcwq->lock);
3771 workqueue_freezing = false;
3773 spin_unlock(&workqueue_lock);
3775 #endif /* CONFIG_FREEZER */
3777 static int __init init_workqueues(void)
3781 /* make sure we have enough bits for OFFQ CPU number */
3782 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
3785 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3786 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3788 /* initialize gcwqs */
3789 for_each_gcwq_cpu(cpu) {
3790 struct global_cwq *gcwq = get_gcwq(cpu);
3791 struct worker_pool *pool;
3793 spin_lock_init(&gcwq->lock);
3795 gcwq->flags |= GCWQ_DISASSOCIATED;
3797 hash_init(gcwq->busy_hash);
3799 for_each_worker_pool(pool, gcwq) {
3801 INIT_LIST_HEAD(&pool->worklist);
3802 INIT_LIST_HEAD(&pool->idle_list);
3804 init_timer_deferrable(&pool->idle_timer);
3805 pool->idle_timer.function = idle_worker_timeout;
3806 pool->idle_timer.data = (unsigned long)pool;
3808 setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3809 (unsigned long)pool);
3811 mutex_init(&pool->assoc_mutex);
3812 ida_init(&pool->worker_ida);
3816 /* create the initial worker */
3817 for_each_online_gcwq_cpu(cpu) {
3818 struct global_cwq *gcwq = get_gcwq(cpu);
3819 struct worker_pool *pool;
3821 if (cpu != WORK_CPU_UNBOUND)
3822 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3824 for_each_worker_pool(pool, gcwq) {
3825 struct worker *worker;
3827 worker = create_worker(pool);
3829 spin_lock_irq(&gcwq->lock);
3830 start_worker(worker);
3831 spin_unlock_irq(&gcwq->lock);
3835 system_wq = alloc_workqueue("events", 0, 0);
3836 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3837 system_long_wq = alloc_workqueue("events_long", 0, 0);
3838 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3839 WQ_UNBOUND_MAX_ACTIVE);
3840 system_freezable_wq = alloc_workqueue("events_freezable",
3842 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3843 !system_unbound_wq || !system_freezable_wq);
3846 early_initcall(init_workqueues);