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>
45 #include "workqueue_sched.h"
51 * A bound gcwq is either associated or disassociated with its CPU.
52 * While associated (!DISASSOCIATED), all workers are bound to the
53 * CPU and none has %WORKER_UNBOUND set and concurrency management
56 * While DISASSOCIATED, the cpu may be offline and all workers have
57 * %WORKER_UNBOUND set and concurrency management disabled, and may
58 * be executing on any CPU. The gcwq behaves as an unbound one.
60 * Note that DISASSOCIATED can be flipped only while holding
61 * managership of all pools on the gcwq to avoid changing binding
62 * state while create_worker() is in progress.
64 GCWQ_DISASSOCIATED = 1 << 0, /* cpu can't serve workers */
65 GCWQ_FREEZING = 1 << 1, /* freeze in progress */
68 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
71 WORKER_STARTED = 1 << 0, /* started */
72 WORKER_DIE = 1 << 1, /* die die die */
73 WORKER_IDLE = 1 << 2, /* is idle */
74 WORKER_PREP = 1 << 3, /* preparing to run works */
75 WORKER_REBIND = 1 << 5, /* mom is home, come back */
76 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
77 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_REBIND | WORKER_UNBOUND |
82 NR_WORKER_POOLS = 2, /* # worker pools per gcwq */
84 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
85 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
86 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
88 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
89 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
92 /* call for help after 10ms
94 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
95 CREATE_COOLDOWN = HZ, /* time to breath after fail */
98 * Rescue workers are used only on emergencies and shared by
101 RESCUER_NICE_LEVEL = -20,
102 HIGHPRI_NICE_LEVEL = -20,
106 * Structure fields follow one of the following exclusion rules.
108 * I: Modifiable by initialization/destruction paths and read-only for
111 * P: Preemption protected. Disabling preemption is enough and should
112 * only be modified and accessed from the local cpu.
114 * L: gcwq->lock protected. Access with gcwq->lock held.
116 * X: During normal operation, modification requires gcwq->lock and
117 * should be done only from local cpu. Either disabling preemption
118 * on local cpu or grabbing gcwq->lock is enough for read access.
119 * If GCWQ_DISASSOCIATED is set, it's identical to L.
121 * F: wq->flush_mutex protected.
123 * W: workqueue_lock protected.
131 * The poor guys doing the actual heavy lifting. All on-duty workers
132 * are either serving the manager role, on idle list or on busy hash.
135 /* on idle list while idle, on busy hash table while busy */
137 struct list_head entry; /* L: while idle */
138 struct hlist_node hentry; /* L: while busy */
141 struct work_struct *current_work; /* L: work being processed */
142 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
143 struct list_head scheduled; /* L: scheduled works */
144 struct task_struct *task; /* I: worker task */
145 struct worker_pool *pool; /* I: the associated pool */
146 /* 64 bytes boundary on 64bit, 32 on 32bit */
147 unsigned long last_active; /* L: last active timestamp */
148 unsigned int flags; /* X: flags */
149 int id; /* I: worker id */
151 /* for rebinding worker to CPU */
152 struct idle_rebind *idle_rebind; /* L: for idle worker */
153 struct work_struct rebind_work; /* L: for busy worker */
157 struct global_cwq *gcwq; /* I: the owning gcwq */
158 unsigned int flags; /* X: flags */
160 struct list_head worklist; /* L: list of pending works */
161 int nr_workers; /* L: total number of workers */
162 int nr_idle; /* L: currently idle ones */
164 struct list_head idle_list; /* X: list of idle workers */
165 struct timer_list idle_timer; /* L: worker idle timeout */
166 struct timer_list mayday_timer; /* L: SOS timer for workers */
168 struct mutex manager_mutex; /* mutex manager should hold */
169 struct ida worker_ida; /* L: for worker IDs */
173 * Global per-cpu workqueue. There's one and only one for each cpu
174 * and all works are queued and processed here regardless of their
178 spinlock_t lock; /* the gcwq lock */
179 unsigned int cpu; /* I: the associated cpu */
180 unsigned int flags; /* L: GCWQ_* flags */
182 /* workers are chained either in busy_hash or pool idle_list */
183 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
184 /* L: hash of busy workers */
186 struct worker_pool pools[NR_WORKER_POOLS];
187 /* normal and highpri pools */
189 wait_queue_head_t rebind_hold; /* rebind hold wait */
190 } ____cacheline_aligned_in_smp;
193 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
194 * work_struct->data are used for flags and thus cwqs need to be
195 * aligned at two's power of the number of flag bits.
197 struct cpu_workqueue_struct {
198 struct worker_pool *pool; /* I: the associated pool */
199 struct workqueue_struct *wq; /* I: the owning workqueue */
200 int work_color; /* L: current color */
201 int flush_color; /* L: flushing color */
202 int nr_in_flight[WORK_NR_COLORS];
203 /* L: nr of in_flight works */
204 int nr_active; /* L: nr of active works */
205 int max_active; /* L: max active works */
206 struct list_head delayed_works; /* L: delayed works */
210 * Structure used to wait for workqueue flush.
213 struct list_head list; /* F: list of flushers */
214 int flush_color; /* F: flush color waiting for */
215 struct completion done; /* flush completion */
219 * All cpumasks are assumed to be always set on UP and thus can't be
220 * used to determine whether there's something to be done.
223 typedef cpumask_var_t mayday_mask_t;
224 #define mayday_test_and_set_cpu(cpu, mask) \
225 cpumask_test_and_set_cpu((cpu), (mask))
226 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
227 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
228 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
229 #define free_mayday_mask(mask) free_cpumask_var((mask))
231 typedef unsigned long mayday_mask_t;
232 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
233 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
234 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
235 #define alloc_mayday_mask(maskp, gfp) true
236 #define free_mayday_mask(mask) do { } while (0)
240 * The externally visible workqueue abstraction is an array of
241 * per-CPU workqueues:
243 struct workqueue_struct {
244 unsigned int flags; /* W: WQ_* flags */
246 struct cpu_workqueue_struct __percpu *pcpu;
247 struct cpu_workqueue_struct *single;
249 } cpu_wq; /* I: cwq's */
250 struct list_head list; /* W: list of all workqueues */
252 struct mutex flush_mutex; /* protects wq flushing */
253 int work_color; /* F: current work color */
254 int flush_color; /* F: current flush color */
255 atomic_t nr_cwqs_to_flush; /* flush in progress */
256 struct wq_flusher *first_flusher; /* F: first flusher */
257 struct list_head flusher_queue; /* F: flush waiters */
258 struct list_head flusher_overflow; /* F: flush overflow list */
260 mayday_mask_t mayday_mask; /* cpus requesting rescue */
261 struct worker *rescuer; /* I: rescue worker */
263 int nr_drainers; /* W: drain in progress */
264 int saved_max_active; /* W: saved cwq max_active */
265 #ifdef CONFIG_LOCKDEP
266 struct lockdep_map lockdep_map;
268 char name[]; /* I: workqueue name */
271 struct workqueue_struct *system_wq __read_mostly;
272 EXPORT_SYMBOL_GPL(system_wq);
273 struct workqueue_struct *system_highpri_wq __read_mostly;
274 EXPORT_SYMBOL_GPL(system_highpri_wq);
275 struct workqueue_struct *system_long_wq __read_mostly;
276 EXPORT_SYMBOL_GPL(system_long_wq);
277 struct workqueue_struct *system_nrt_wq __read_mostly;
278 EXPORT_SYMBOL_GPL(system_nrt_wq);
279 struct workqueue_struct *system_unbound_wq __read_mostly;
280 EXPORT_SYMBOL_GPL(system_unbound_wq);
281 struct workqueue_struct *system_freezable_wq __read_mostly;
282 EXPORT_SYMBOL_GPL(system_freezable_wq);
283 struct workqueue_struct *system_nrt_freezable_wq __read_mostly;
284 EXPORT_SYMBOL_GPL(system_nrt_freezable_wq);
286 #define CREATE_TRACE_POINTS
287 #include <trace/events/workqueue.h>
289 #define for_each_worker_pool(pool, gcwq) \
290 for ((pool) = &(gcwq)->pools[0]; \
291 (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
293 #define for_each_busy_worker(worker, i, pos, gcwq) \
294 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
295 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
297 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
300 if (cpu < nr_cpu_ids) {
302 cpu = cpumask_next(cpu, mask);
303 if (cpu < nr_cpu_ids)
307 return WORK_CPU_UNBOUND;
309 return WORK_CPU_NONE;
312 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
313 struct workqueue_struct *wq)
315 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
321 * An extra gcwq is defined for an invalid cpu number
322 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
323 * specific CPU. The following iterators are similar to
324 * for_each_*_cpu() iterators but also considers the unbound gcwq.
326 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
327 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
328 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
329 * WORK_CPU_UNBOUND for unbound workqueues
331 #define for_each_gcwq_cpu(cpu) \
332 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
333 (cpu) < WORK_CPU_NONE; \
334 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
336 #define for_each_online_gcwq_cpu(cpu) \
337 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
338 (cpu) < WORK_CPU_NONE; \
339 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
341 #define for_each_cwq_cpu(cpu, wq) \
342 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
343 (cpu) < WORK_CPU_NONE; \
344 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
346 #ifdef CONFIG_DEBUG_OBJECTS_WORK
348 static struct debug_obj_descr work_debug_descr;
350 static void *work_debug_hint(void *addr)
352 return ((struct work_struct *) addr)->func;
356 * fixup_init is called when:
357 * - an active object is initialized
359 static int work_fixup_init(void *addr, enum debug_obj_state state)
361 struct work_struct *work = addr;
364 case ODEBUG_STATE_ACTIVE:
365 cancel_work_sync(work);
366 debug_object_init(work, &work_debug_descr);
374 * fixup_activate is called when:
375 * - an active object is activated
376 * - an unknown object is activated (might be a statically initialized object)
378 static int work_fixup_activate(void *addr, enum debug_obj_state state)
380 struct work_struct *work = addr;
384 case ODEBUG_STATE_NOTAVAILABLE:
386 * This is not really a fixup. The work struct was
387 * statically initialized. We just make sure that it
388 * is tracked in the object tracker.
390 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
391 debug_object_init(work, &work_debug_descr);
392 debug_object_activate(work, &work_debug_descr);
398 case ODEBUG_STATE_ACTIVE:
407 * fixup_free is called when:
408 * - an active object is freed
410 static int work_fixup_free(void *addr, enum debug_obj_state state)
412 struct work_struct *work = addr;
415 case ODEBUG_STATE_ACTIVE:
416 cancel_work_sync(work);
417 debug_object_free(work, &work_debug_descr);
424 static struct debug_obj_descr work_debug_descr = {
425 .name = "work_struct",
426 .debug_hint = work_debug_hint,
427 .fixup_init = work_fixup_init,
428 .fixup_activate = work_fixup_activate,
429 .fixup_free = work_fixup_free,
432 static inline void debug_work_activate(struct work_struct *work)
434 debug_object_activate(work, &work_debug_descr);
437 static inline void debug_work_deactivate(struct work_struct *work)
439 debug_object_deactivate(work, &work_debug_descr);
442 void __init_work(struct work_struct *work, int onstack)
445 debug_object_init_on_stack(work, &work_debug_descr);
447 debug_object_init(work, &work_debug_descr);
449 EXPORT_SYMBOL_GPL(__init_work);
451 void destroy_work_on_stack(struct work_struct *work)
453 debug_object_free(work, &work_debug_descr);
455 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
458 static inline void debug_work_activate(struct work_struct *work) { }
459 static inline void debug_work_deactivate(struct work_struct *work) { }
462 /* Serializes the accesses to the list of workqueues. */
463 static DEFINE_SPINLOCK(workqueue_lock);
464 static LIST_HEAD(workqueues);
465 static bool workqueue_freezing; /* W: have wqs started freezing? */
468 * The almighty global cpu workqueues. nr_running is the only field
469 * which is expected to be used frequently by other cpus via
470 * try_to_wake_up(). Put it in a separate cacheline.
472 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
473 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]);
476 * Global cpu workqueue and nr_running counter for unbound gcwq. The
477 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
478 * workers have WORKER_UNBOUND set.
480 static struct global_cwq unbound_global_cwq;
481 static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = {
482 [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
485 static int worker_thread(void *__worker);
487 static int worker_pool_pri(struct worker_pool *pool)
489 return pool - pool->gcwq->pools;
492 static struct global_cwq *get_gcwq(unsigned int cpu)
494 if (cpu != WORK_CPU_UNBOUND)
495 return &per_cpu(global_cwq, cpu);
497 return &unbound_global_cwq;
500 static atomic_t *get_pool_nr_running(struct worker_pool *pool)
502 int cpu = pool->gcwq->cpu;
503 int idx = worker_pool_pri(pool);
505 if (cpu != WORK_CPU_UNBOUND)
506 return &per_cpu(pool_nr_running, cpu)[idx];
508 return &unbound_pool_nr_running[idx];
511 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
512 struct workqueue_struct *wq)
514 if (!(wq->flags & WQ_UNBOUND)) {
515 if (likely(cpu < nr_cpu_ids))
516 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
517 } else if (likely(cpu == WORK_CPU_UNBOUND))
518 return wq->cpu_wq.single;
522 static unsigned int work_color_to_flags(int color)
524 return color << WORK_STRUCT_COLOR_SHIFT;
527 static int get_work_color(struct work_struct *work)
529 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
530 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
533 static int work_next_color(int color)
535 return (color + 1) % WORK_NR_COLORS;
539 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
540 * contain the pointer to the queued cwq. Once execution starts, the flag
541 * is cleared and the high bits contain OFFQ flags and CPU number.
543 * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
544 * and clear_work_data() can be used to set the cwq, cpu or clear
545 * work->data. These functions should only be called while the work is
546 * owned - ie. while the PENDING bit is set.
548 * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
549 * a work. gcwq is available once the work has been queued anywhere after
550 * initialization until it is sync canceled. cwq is available only while
551 * the work item is queued.
553 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
554 * canceled. While being canceled, a work item may have its PENDING set
555 * but stay off timer and worklist for arbitrarily long and nobody should
556 * try to steal the PENDING bit.
558 static inline void set_work_data(struct work_struct *work, unsigned long data,
561 BUG_ON(!work_pending(work));
562 atomic_long_set(&work->data, data | flags | work_static(work));
565 static void set_work_cwq(struct work_struct *work,
566 struct cpu_workqueue_struct *cwq,
567 unsigned long extra_flags)
569 set_work_data(work, (unsigned long)cwq,
570 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
573 static void set_work_cpu_and_clear_pending(struct work_struct *work,
577 * The following wmb is paired with the implied mb in
578 * test_and_set_bit(PENDING) and ensures all updates to @work made
579 * here are visible to and precede any updates by the next PENDING
583 set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0);
586 static void clear_work_data(struct work_struct *work)
588 smp_wmb(); /* see set_work_cpu_and_clear_pending() */
589 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
592 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
594 unsigned long data = atomic_long_read(&work->data);
596 if (data & WORK_STRUCT_CWQ)
597 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
602 static struct global_cwq *get_work_gcwq(struct work_struct *work)
604 unsigned long data = atomic_long_read(&work->data);
607 if (data & WORK_STRUCT_CWQ)
608 return ((struct cpu_workqueue_struct *)
609 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq;
611 cpu = data >> WORK_OFFQ_CPU_SHIFT;
612 if (cpu == WORK_CPU_NONE)
615 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
616 return get_gcwq(cpu);
619 static void mark_work_canceling(struct work_struct *work)
621 struct global_cwq *gcwq = get_work_gcwq(work);
622 unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE;
624 set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) | WORK_OFFQ_CANCELING,
625 WORK_STRUCT_PENDING);
628 static bool work_is_canceling(struct work_struct *work)
630 unsigned long data = atomic_long_read(&work->data);
632 return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
636 * Policy functions. These define the policies on how the global worker
637 * pools are managed. Unless noted otherwise, these functions assume that
638 * they're being called with gcwq->lock held.
641 static bool __need_more_worker(struct worker_pool *pool)
643 return !atomic_read(get_pool_nr_running(pool));
647 * Need to wake up a worker? Called from anything but currently
650 * Note that, because unbound workers never contribute to nr_running, this
651 * function will always return %true for unbound gcwq as long as the
652 * worklist isn't empty.
654 static bool need_more_worker(struct worker_pool *pool)
656 return !list_empty(&pool->worklist) && __need_more_worker(pool);
659 /* Can I start working? Called from busy but !running workers. */
660 static bool may_start_working(struct worker_pool *pool)
662 return pool->nr_idle;
665 /* Do I need to keep working? Called from currently running workers. */
666 static bool keep_working(struct worker_pool *pool)
668 atomic_t *nr_running = get_pool_nr_running(pool);
670 return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
673 /* Do we need a new worker? Called from manager. */
674 static bool need_to_create_worker(struct worker_pool *pool)
676 return need_more_worker(pool) && !may_start_working(pool);
679 /* Do I need to be the manager? */
680 static bool need_to_manage_workers(struct worker_pool *pool)
682 return need_to_create_worker(pool) ||
683 (pool->flags & POOL_MANAGE_WORKERS);
686 /* Do we have too many workers and should some go away? */
687 static bool too_many_workers(struct worker_pool *pool)
689 bool managing = mutex_is_locked(&pool->manager_mutex);
690 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
691 int nr_busy = pool->nr_workers - nr_idle;
693 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
700 /* Return the first worker. Safe with preemption disabled */
701 static struct worker *first_worker(struct worker_pool *pool)
703 if (unlikely(list_empty(&pool->idle_list)))
706 return list_first_entry(&pool->idle_list, struct worker, entry);
710 * wake_up_worker - wake up an idle worker
711 * @pool: worker pool to wake worker from
713 * Wake up the first idle worker of @pool.
716 * spin_lock_irq(gcwq->lock).
718 static void wake_up_worker(struct worker_pool *pool)
720 struct worker *worker = first_worker(pool);
723 wake_up_process(worker->task);
727 * wq_worker_waking_up - a worker is waking up
728 * @task: task waking up
729 * @cpu: CPU @task is waking up to
731 * This function is called during try_to_wake_up() when a worker is
735 * spin_lock_irq(rq->lock)
737 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
739 struct worker *worker = kthread_data(task);
741 if (!(worker->flags & WORKER_NOT_RUNNING))
742 atomic_inc(get_pool_nr_running(worker->pool));
746 * wq_worker_sleeping - a worker is going to sleep
747 * @task: task going to sleep
748 * @cpu: CPU in question, must be the current CPU number
750 * This function is called during schedule() when a busy worker is
751 * going to sleep. Worker on the same cpu can be woken up by
752 * returning pointer to its task.
755 * spin_lock_irq(rq->lock)
758 * Worker task on @cpu to wake up, %NULL if none.
760 struct task_struct *wq_worker_sleeping(struct task_struct *task,
763 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
764 struct worker_pool *pool = worker->pool;
765 atomic_t *nr_running = get_pool_nr_running(pool);
767 if (worker->flags & WORKER_NOT_RUNNING)
770 /* this can only happen on the local cpu */
771 BUG_ON(cpu != raw_smp_processor_id());
774 * The counterpart of the following dec_and_test, implied mb,
775 * worklist not empty test sequence is in insert_work().
776 * Please read comment there.
778 * NOT_RUNNING is clear. This means that we're bound to and
779 * running on the local cpu w/ rq lock held and preemption
780 * disabled, which in turn means that none else could be
781 * manipulating idle_list, so dereferencing idle_list without gcwq
784 if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
785 to_wakeup = first_worker(pool);
786 return to_wakeup ? to_wakeup->task : NULL;
790 * worker_set_flags - set worker flags and adjust nr_running accordingly
792 * @flags: flags to set
793 * @wakeup: wakeup an idle worker if necessary
795 * Set @flags in @worker->flags and adjust nr_running accordingly. If
796 * nr_running becomes zero and @wakeup is %true, an idle worker is
800 * spin_lock_irq(gcwq->lock)
802 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
805 struct worker_pool *pool = worker->pool;
807 WARN_ON_ONCE(worker->task != current);
810 * If transitioning into NOT_RUNNING, adjust nr_running and
811 * wake up an idle worker as necessary if requested by
814 if ((flags & WORKER_NOT_RUNNING) &&
815 !(worker->flags & WORKER_NOT_RUNNING)) {
816 atomic_t *nr_running = get_pool_nr_running(pool);
819 if (atomic_dec_and_test(nr_running) &&
820 !list_empty(&pool->worklist))
821 wake_up_worker(pool);
823 atomic_dec(nr_running);
826 worker->flags |= flags;
830 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
832 * @flags: flags to clear
834 * Clear @flags in @worker->flags and adjust nr_running accordingly.
837 * spin_lock_irq(gcwq->lock)
839 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
841 struct worker_pool *pool = worker->pool;
842 unsigned int oflags = worker->flags;
844 WARN_ON_ONCE(worker->task != current);
846 worker->flags &= ~flags;
849 * If transitioning out of NOT_RUNNING, increment nr_running. Note
850 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
851 * of multiple flags, not a single flag.
853 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
854 if (!(worker->flags & WORKER_NOT_RUNNING))
855 atomic_inc(get_pool_nr_running(pool));
859 * busy_worker_head - return the busy hash head for a work
860 * @gcwq: gcwq of interest
861 * @work: work to be hashed
863 * Return hash head of @gcwq for @work.
866 * spin_lock_irq(gcwq->lock).
869 * Pointer to the hash head.
871 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
872 struct work_struct *work)
874 const int base_shift = ilog2(sizeof(struct work_struct));
875 unsigned long v = (unsigned long)work;
877 /* simple shift and fold hash, do we need something better? */
879 v += v >> BUSY_WORKER_HASH_ORDER;
880 v &= BUSY_WORKER_HASH_MASK;
882 return &gcwq->busy_hash[v];
886 * __find_worker_executing_work - find worker which is executing a work
887 * @gcwq: gcwq of interest
888 * @bwh: hash head as returned by busy_worker_head()
889 * @work: work to find worker for
891 * Find a worker which is executing @work on @gcwq. @bwh should be
892 * the hash head obtained by calling busy_worker_head() with the same
896 * spin_lock_irq(gcwq->lock).
899 * Pointer to worker which is executing @work if found, NULL
902 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
903 struct hlist_head *bwh,
904 struct work_struct *work)
906 struct worker *worker;
907 struct hlist_node *tmp;
909 hlist_for_each_entry(worker, tmp, bwh, hentry)
910 if (worker->current_work == work)
916 * find_worker_executing_work - find worker which is executing a work
917 * @gcwq: gcwq of interest
918 * @work: work to find worker for
920 * Find a worker which is executing @work on @gcwq. This function is
921 * identical to __find_worker_executing_work() except that this
922 * function calculates @bwh itself.
925 * spin_lock_irq(gcwq->lock).
928 * Pointer to worker which is executing @work if found, NULL
931 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
932 struct work_struct *work)
934 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
939 * move_linked_works - move linked works to a list
940 * @work: start of series of works to be scheduled
941 * @head: target list to append @work to
942 * @nextp: out paramter for nested worklist walking
944 * Schedule linked works starting from @work to @head. Work series to
945 * be scheduled starts at @work and includes any consecutive work with
946 * WORK_STRUCT_LINKED set in its predecessor.
948 * If @nextp is not NULL, it's updated to point to the next work of
949 * the last scheduled work. This allows move_linked_works() to be
950 * nested inside outer list_for_each_entry_safe().
953 * spin_lock_irq(gcwq->lock).
955 static void move_linked_works(struct work_struct *work, struct list_head *head,
956 struct work_struct **nextp)
958 struct work_struct *n;
961 * Linked worklist will always end before the end of the list,
962 * use NULL for list head.
964 list_for_each_entry_safe_from(work, n, NULL, entry) {
965 list_move_tail(&work->entry, head);
966 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
971 * If we're already inside safe list traversal and have moved
972 * multiple works to the scheduled queue, the next position
973 * needs to be updated.
979 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
981 struct work_struct *work = list_first_entry(&cwq->delayed_works,
982 struct work_struct, entry);
984 trace_workqueue_activate_work(work);
985 move_linked_works(work, &cwq->pool->worklist, NULL);
986 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
991 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
992 * @cwq: cwq of interest
993 * @color: color of work which left the queue
994 * @delayed: for a delayed work
996 * A work either has completed or is removed from pending queue,
997 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1000 * spin_lock_irq(gcwq->lock).
1002 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1005 /* ignore uncolored works */
1006 if (color == WORK_NO_COLOR)
1009 cwq->nr_in_flight[color]--;
1013 if (!list_empty(&cwq->delayed_works)) {
1014 /* one down, submit a delayed one */
1015 if (cwq->nr_active < cwq->max_active)
1016 cwq_activate_first_delayed(cwq);
1020 /* is flush in progress and are we at the flushing tip? */
1021 if (likely(cwq->flush_color != color))
1024 /* are there still in-flight works? */
1025 if (cwq->nr_in_flight[color])
1028 /* this cwq is done, clear flush_color */
1029 cwq->flush_color = -1;
1032 * If this was the last cwq, wake up the first flusher. It
1033 * will handle the rest.
1035 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1036 complete(&cwq->wq->first_flusher->done);
1040 * try_to_grab_pending - steal work item from worklist and disable irq
1041 * @work: work item to steal
1042 * @is_dwork: @work is a delayed_work
1043 * @flags: place to store irq state
1045 * Try to grab PENDING bit of @work. This function can handle @work in any
1046 * stable state - idle, on timer or on worklist. Return values are
1048 * 1 if @work was pending and we successfully stole PENDING
1049 * 0 if @work was idle and we claimed PENDING
1050 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1051 * -ENOENT if someone else is canceling @work, this state may persist
1052 * for arbitrarily long
1054 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1055 * preempted while holding PENDING and @work off queue, preemption must be
1056 * disabled on entry. This ensures that we don't return -EAGAIN while
1057 * another task is preempted in this function.
1059 * On successful return, >= 0, irq is disabled and the caller is
1060 * responsible for releasing it using local_irq_restore(*@flags).
1062 * This function is safe to call from any context other than IRQ handler.
1063 * An IRQ handler may run on top of delayed_work_timer_fn() which can make
1064 * this function return -EAGAIN perpetually.
1066 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1067 unsigned long *flags)
1069 struct global_cwq *gcwq;
1071 WARN_ON_ONCE(in_irq());
1073 local_irq_save(*flags);
1075 /* try to steal the timer if it exists */
1077 struct delayed_work *dwork = to_delayed_work(work);
1079 if (likely(del_timer(&dwork->timer)))
1083 /* try to claim PENDING the normal way */
1084 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1088 * The queueing is in progress, or it is already queued. Try to
1089 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1091 gcwq = get_work_gcwq(work);
1095 spin_lock(&gcwq->lock);
1096 if (!list_empty(&work->entry)) {
1098 * This work is queued, but perhaps we locked the wrong gcwq.
1099 * In that case we must see the new value after rmb(), see
1100 * insert_work()->wmb().
1103 if (gcwq == get_work_gcwq(work)) {
1104 debug_work_deactivate(work);
1105 list_del_init(&work->entry);
1106 cwq_dec_nr_in_flight(get_work_cwq(work),
1107 get_work_color(work),
1108 *work_data_bits(work) & WORK_STRUCT_DELAYED);
1110 spin_unlock(&gcwq->lock);
1114 spin_unlock(&gcwq->lock);
1116 local_irq_restore(*flags);
1117 if (work_is_canceling(work))
1124 * insert_work - insert a work into gcwq
1125 * @cwq: cwq @work belongs to
1126 * @work: work to insert
1127 * @head: insertion point
1128 * @extra_flags: extra WORK_STRUCT_* flags to set
1130 * Insert @work which belongs to @cwq into @gcwq after @head.
1131 * @extra_flags is or'd to work_struct flags.
1134 * spin_lock_irq(gcwq->lock).
1136 static void insert_work(struct cpu_workqueue_struct *cwq,
1137 struct work_struct *work, struct list_head *head,
1138 unsigned int extra_flags)
1140 struct worker_pool *pool = cwq->pool;
1142 /* we own @work, set data and link */
1143 set_work_cwq(work, cwq, extra_flags);
1146 * Ensure that we get the right work->data if we see the
1147 * result of list_add() below, see try_to_grab_pending().
1151 list_add_tail(&work->entry, head);
1154 * Ensure either worker_sched_deactivated() sees the above
1155 * list_add_tail() or we see zero nr_running to avoid workers
1156 * lying around lazily while there are works to be processed.
1160 if (__need_more_worker(pool))
1161 wake_up_worker(pool);
1165 * Test whether @work is being queued from another work executing on the
1166 * same workqueue. This is rather expensive and should only be used from
1169 static bool is_chained_work(struct workqueue_struct *wq)
1171 unsigned long flags;
1174 for_each_gcwq_cpu(cpu) {
1175 struct global_cwq *gcwq = get_gcwq(cpu);
1176 struct worker *worker;
1177 struct hlist_node *pos;
1180 spin_lock_irqsave(&gcwq->lock, flags);
1181 for_each_busy_worker(worker, i, pos, gcwq) {
1182 if (worker->task != current)
1184 spin_unlock_irqrestore(&gcwq->lock, flags);
1186 * I'm @worker, no locking necessary. See if @work
1187 * is headed to the same workqueue.
1189 return worker->current_cwq->wq == wq;
1191 spin_unlock_irqrestore(&gcwq->lock, flags);
1196 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1197 struct work_struct *work)
1199 struct global_cwq *gcwq;
1200 struct cpu_workqueue_struct *cwq;
1201 struct list_head *worklist;
1202 unsigned int work_flags;
1203 unsigned int req_cpu = cpu;
1206 * While a work item is PENDING && off queue, a task trying to
1207 * steal the PENDING will busy-loop waiting for it to either get
1208 * queued or lose PENDING. Grabbing PENDING and queueing should
1209 * happen with IRQ disabled.
1211 WARN_ON_ONCE(!irqs_disabled());
1213 debug_work_activate(work);
1215 /* if dying, only works from the same workqueue are allowed */
1216 if (unlikely(wq->flags & WQ_DRAINING) &&
1217 WARN_ON_ONCE(!is_chained_work(wq)))
1220 /* determine gcwq to use */
1221 if (!(wq->flags & WQ_UNBOUND)) {
1222 struct global_cwq *last_gcwq;
1224 if (cpu == WORK_CPU_UNBOUND)
1225 cpu = raw_smp_processor_id();
1228 * It's multi cpu. If @wq is non-reentrant and @work
1229 * was previously on a different cpu, it might still
1230 * be running there, in which case the work needs to
1231 * be queued on that cpu to guarantee non-reentrance.
1233 gcwq = get_gcwq(cpu);
1234 if (wq->flags & WQ_NON_REENTRANT &&
1235 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1236 struct worker *worker;
1238 spin_lock(&last_gcwq->lock);
1240 worker = find_worker_executing_work(last_gcwq, work);
1242 if (worker && worker->current_cwq->wq == wq)
1245 /* meh... not running there, queue here */
1246 spin_unlock(&last_gcwq->lock);
1247 spin_lock(&gcwq->lock);
1250 spin_lock(&gcwq->lock);
1253 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1254 spin_lock(&gcwq->lock);
1257 /* gcwq determined, get cwq and queue */
1258 cwq = get_cwq(gcwq->cpu, wq);
1259 trace_workqueue_queue_work(req_cpu, cwq, work);
1261 if (WARN_ON(!list_empty(&work->entry))) {
1262 spin_unlock(&gcwq->lock);
1266 cwq->nr_in_flight[cwq->work_color]++;
1267 work_flags = work_color_to_flags(cwq->work_color);
1269 if (likely(cwq->nr_active < cwq->max_active)) {
1270 trace_workqueue_activate_work(work);
1272 worklist = &cwq->pool->worklist;
1274 work_flags |= WORK_STRUCT_DELAYED;
1275 worklist = &cwq->delayed_works;
1278 insert_work(cwq, work, worklist, work_flags);
1280 spin_unlock(&gcwq->lock);
1284 * queue_work_on - queue work on specific cpu
1285 * @cpu: CPU number to execute work on
1286 * @wq: workqueue to use
1287 * @work: work to queue
1289 * Returns %false if @work was already on a queue, %true otherwise.
1291 * We queue the work to a specific CPU, the caller must ensure it
1294 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1295 struct work_struct *work)
1298 unsigned long flags;
1300 local_irq_save(flags);
1302 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1303 __queue_work(cpu, wq, work);
1307 local_irq_restore(flags);
1310 EXPORT_SYMBOL_GPL(queue_work_on);
1313 * queue_work - queue work on a workqueue
1314 * @wq: workqueue to use
1315 * @work: work to queue
1317 * Returns %false if @work was already on a queue, %true otherwise.
1319 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1320 * it can be processed by another CPU.
1322 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1324 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1326 EXPORT_SYMBOL_GPL(queue_work);
1328 void delayed_work_timer_fn(unsigned long __data)
1330 struct delayed_work *dwork = (struct delayed_work *)__data;
1331 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1333 local_irq_disable();
1334 __queue_work(dwork->cpu, cwq->wq, &dwork->work);
1337 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1339 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1340 struct delayed_work *dwork, unsigned long delay)
1342 struct timer_list *timer = &dwork->timer;
1343 struct work_struct *work = &dwork->work;
1346 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1347 timer->data != (unsigned long)dwork);
1348 BUG_ON(timer_pending(timer));
1349 BUG_ON(!list_empty(&work->entry));
1351 timer_stats_timer_set_start_info(&dwork->timer);
1354 * This stores cwq for the moment, for the timer_fn. Note that the
1355 * work's gcwq is preserved to allow reentrance detection for
1358 if (!(wq->flags & WQ_UNBOUND)) {
1359 struct global_cwq *gcwq = get_work_gcwq(work);
1362 * If we cannot get the last gcwq from @work directly,
1363 * select the last CPU such that it avoids unnecessarily
1364 * triggering non-reentrancy check in __queue_work().
1369 if (lcpu == WORK_CPU_UNBOUND)
1370 lcpu = raw_smp_processor_id();
1372 lcpu = WORK_CPU_UNBOUND;
1375 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1378 timer->expires = jiffies + delay;
1380 if (unlikely(cpu != WORK_CPU_UNBOUND))
1381 add_timer_on(timer, cpu);
1387 * queue_delayed_work_on - queue work on specific CPU after delay
1388 * @cpu: CPU number to execute work on
1389 * @wq: workqueue to use
1390 * @dwork: work to queue
1391 * @delay: number of jiffies to wait before queueing
1393 * Returns %false if @work was already on a queue, %true otherwise. If
1394 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1397 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1398 struct delayed_work *dwork, unsigned long delay)
1400 struct work_struct *work = &dwork->work;
1402 unsigned long flags;
1405 return queue_work_on(cpu, wq, &dwork->work);
1407 /* read the comment in __queue_work() */
1408 local_irq_save(flags);
1410 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1411 __queue_delayed_work(cpu, wq, dwork, delay);
1415 local_irq_restore(flags);
1418 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1421 * queue_delayed_work - queue work on a workqueue after delay
1422 * @wq: workqueue to use
1423 * @dwork: delayable work to queue
1424 * @delay: number of jiffies to wait before queueing
1426 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1428 bool queue_delayed_work(struct workqueue_struct *wq,
1429 struct delayed_work *dwork, unsigned long delay)
1431 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1433 EXPORT_SYMBOL_GPL(queue_delayed_work);
1436 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1437 * @cpu: CPU number to execute work on
1438 * @wq: workqueue to use
1439 * @dwork: work to queue
1440 * @delay: number of jiffies to wait before queueing
1442 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1443 * modify @dwork's timer so that it expires after @delay. If @delay is
1444 * zero, @work is guaranteed to be scheduled immediately regardless of its
1447 * Returns %false if @dwork was idle and queued, %true if @dwork was
1448 * pending and its timer was modified.
1450 * This function is safe to call from any context other than IRQ handler.
1451 * See try_to_grab_pending() for details.
1453 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1454 struct delayed_work *dwork, unsigned long delay)
1456 unsigned long flags;
1460 ret = try_to_grab_pending(&dwork->work, true, &flags);
1461 } while (unlikely(ret == -EAGAIN));
1463 if (likely(ret >= 0)) {
1464 __queue_delayed_work(cpu, wq, dwork, delay);
1465 local_irq_restore(flags);
1468 /* -ENOENT from try_to_grab_pending() becomes %true */
1471 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1474 * mod_delayed_work - modify delay of or queue a delayed work
1475 * @wq: workqueue to use
1476 * @dwork: work to queue
1477 * @delay: number of jiffies to wait before queueing
1479 * mod_delayed_work_on() on local CPU.
1481 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1482 unsigned long delay)
1484 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1486 EXPORT_SYMBOL_GPL(mod_delayed_work);
1489 * worker_enter_idle - enter idle state
1490 * @worker: worker which is entering idle state
1492 * @worker is entering idle state. Update stats and idle timer if
1496 * spin_lock_irq(gcwq->lock).
1498 static void worker_enter_idle(struct worker *worker)
1500 struct worker_pool *pool = worker->pool;
1501 struct global_cwq *gcwq = pool->gcwq;
1503 BUG_ON(worker->flags & WORKER_IDLE);
1504 BUG_ON(!list_empty(&worker->entry) &&
1505 (worker->hentry.next || worker->hentry.pprev));
1507 /* can't use worker_set_flags(), also called from start_worker() */
1508 worker->flags |= WORKER_IDLE;
1510 worker->last_active = jiffies;
1512 /* idle_list is LIFO */
1513 list_add(&worker->entry, &pool->idle_list);
1515 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1516 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1519 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1520 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1521 * nr_running, the warning may trigger spuriously. Check iff
1522 * unbind is not in progress.
1524 WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
1525 pool->nr_workers == pool->nr_idle &&
1526 atomic_read(get_pool_nr_running(pool)));
1530 * worker_leave_idle - leave idle state
1531 * @worker: worker which is leaving idle state
1533 * @worker is leaving idle state. Update stats.
1536 * spin_lock_irq(gcwq->lock).
1538 static void worker_leave_idle(struct worker *worker)
1540 struct worker_pool *pool = worker->pool;
1542 BUG_ON(!(worker->flags & WORKER_IDLE));
1543 worker_clr_flags(worker, WORKER_IDLE);
1545 list_del_init(&worker->entry);
1549 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1552 * Works which are scheduled while the cpu is online must at least be
1553 * scheduled to a worker which is bound to the cpu so that if they are
1554 * flushed from cpu callbacks while cpu is going down, they are
1555 * guaranteed to execute on the cpu.
1557 * This function is to be used by rogue workers and rescuers to bind
1558 * themselves to the target cpu and may race with cpu going down or
1559 * coming online. kthread_bind() can't be used because it may put the
1560 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1561 * verbatim as it's best effort and blocking and gcwq may be
1562 * [dis]associated in the meantime.
1564 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1565 * binding against %GCWQ_DISASSOCIATED which is set during
1566 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1567 * enters idle state or fetches works without dropping lock, it can
1568 * guarantee the scheduling requirement described in the first paragraph.
1571 * Might sleep. Called without any lock but returns with gcwq->lock
1575 * %true if the associated gcwq is online (@worker is successfully
1576 * bound), %false if offline.
1578 static bool worker_maybe_bind_and_lock(struct worker *worker)
1579 __acquires(&gcwq->lock)
1581 struct global_cwq *gcwq = worker->pool->gcwq;
1582 struct task_struct *task = worker->task;
1586 * The following call may fail, succeed or succeed
1587 * without actually migrating the task to the cpu if
1588 * it races with cpu hotunplug operation. Verify
1589 * against GCWQ_DISASSOCIATED.
1591 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1592 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1594 spin_lock_irq(&gcwq->lock);
1595 if (gcwq->flags & GCWQ_DISASSOCIATED)
1597 if (task_cpu(task) == gcwq->cpu &&
1598 cpumask_equal(¤t->cpus_allowed,
1599 get_cpu_mask(gcwq->cpu)))
1601 spin_unlock_irq(&gcwq->lock);
1604 * We've raced with CPU hot[un]plug. Give it a breather
1605 * and retry migration. cond_resched() is required here;
1606 * otherwise, we might deadlock against cpu_stop trying to
1607 * bring down the CPU on non-preemptive kernel.
1614 struct idle_rebind {
1615 int cnt; /* # workers to be rebound */
1616 struct completion done; /* all workers rebound */
1620 * Rebind an idle @worker to its CPU. During CPU onlining, this has to
1621 * happen synchronously for idle workers. worker_thread() will test
1622 * %WORKER_REBIND before leaving idle and call this function.
1624 static void idle_worker_rebind(struct worker *worker)
1626 struct global_cwq *gcwq = worker->pool->gcwq;
1628 /* CPU must be online at this point */
1629 WARN_ON(!worker_maybe_bind_and_lock(worker));
1630 if (!--worker->idle_rebind->cnt)
1631 complete(&worker->idle_rebind->done);
1632 spin_unlock_irq(&worker->pool->gcwq->lock);
1634 /* we did our part, wait for rebind_workers() to finish up */
1635 wait_event(gcwq->rebind_hold, !(worker->flags & WORKER_REBIND));
1639 * Function for @worker->rebind.work used to rebind unbound busy workers to
1640 * the associated cpu which is coming back online. This is scheduled by
1641 * cpu up but can race with other cpu hotplug operations and may be
1642 * executed twice without intervening cpu down.
1644 static void busy_worker_rebind_fn(struct work_struct *work)
1646 struct worker *worker = container_of(work, struct worker, rebind_work);
1647 struct global_cwq *gcwq = worker->pool->gcwq;
1649 if (worker_maybe_bind_and_lock(worker))
1650 worker_clr_flags(worker, WORKER_REBIND);
1652 spin_unlock_irq(&gcwq->lock);
1656 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1657 * @gcwq: gcwq of interest
1659 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1660 * is different for idle and busy ones.
1662 * The idle ones should be rebound synchronously and idle rebinding should
1663 * be complete before any worker starts executing work items with
1664 * concurrency management enabled; otherwise, scheduler may oops trying to
1665 * wake up non-local idle worker from wq_worker_sleeping().
1667 * This is achieved by repeatedly requesting rebinding until all idle
1668 * workers are known to have been rebound under @gcwq->lock and holding all
1669 * idle workers from becoming busy until idle rebinding is complete.
1671 * Once idle workers are rebound, busy workers can be rebound as they
1672 * finish executing their current work items. Queueing the rebind work at
1673 * the head of their scheduled lists is enough. Note that nr_running will
1674 * be properbly bumped as busy workers rebind.
1676 * On return, all workers are guaranteed to either be bound or have rebind
1677 * work item scheduled.
1679 static void rebind_workers(struct global_cwq *gcwq)
1680 __releases(&gcwq->lock) __acquires(&gcwq->lock)
1682 struct idle_rebind idle_rebind;
1683 struct worker_pool *pool;
1684 struct worker *worker;
1685 struct hlist_node *pos;
1688 lockdep_assert_held(&gcwq->lock);
1690 for_each_worker_pool(pool, gcwq)
1691 lockdep_assert_held(&pool->manager_mutex);
1694 * Rebind idle workers. Interlocked both ways. We wait for
1695 * workers to rebind via @idle_rebind.done. Workers will wait for
1696 * us to finish up by watching %WORKER_REBIND.
1698 init_completion(&idle_rebind.done);
1700 idle_rebind.cnt = 1;
1701 INIT_COMPLETION(idle_rebind.done);
1703 /* set REBIND and kick idle ones, we'll wait for these later */
1704 for_each_worker_pool(pool, gcwq) {
1705 list_for_each_entry(worker, &pool->idle_list, entry) {
1706 if (worker->flags & WORKER_REBIND)
1709 /* morph UNBOUND to REBIND */
1710 worker->flags &= ~WORKER_UNBOUND;
1711 worker->flags |= WORKER_REBIND;
1714 worker->idle_rebind = &idle_rebind;
1716 /* worker_thread() will call idle_worker_rebind() */
1717 wake_up_process(worker->task);
1721 if (--idle_rebind.cnt) {
1722 spin_unlock_irq(&gcwq->lock);
1723 wait_for_completion(&idle_rebind.done);
1724 spin_lock_irq(&gcwq->lock);
1725 /* busy ones might have become idle while waiting, retry */
1730 * All idle workers are rebound and waiting for %WORKER_REBIND to
1731 * be cleared inside idle_worker_rebind(). Clear and release.
1732 * Clearing %WORKER_REBIND from this foreign context is safe
1733 * because these workers are still guaranteed to be idle.
1735 for_each_worker_pool(pool, gcwq)
1736 list_for_each_entry(worker, &pool->idle_list, entry)
1737 worker->flags &= ~WORKER_REBIND;
1739 wake_up_all(&gcwq->rebind_hold);
1741 /* rebind busy workers */
1742 for_each_busy_worker(worker, i, pos, gcwq) {
1743 struct work_struct *rebind_work = &worker->rebind_work;
1744 struct workqueue_struct *wq;
1746 /* morph UNBOUND to REBIND */
1747 worker->flags &= ~WORKER_UNBOUND;
1748 worker->flags |= WORKER_REBIND;
1750 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1751 work_data_bits(rebind_work)))
1754 debug_work_activate(rebind_work);
1757 * wq doesn't really matter but let's keep @worker->pool
1758 * and @cwq->pool consistent for sanity.
1760 if (worker_pool_pri(worker->pool))
1761 wq = system_highpri_wq;
1765 insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
1766 worker->scheduled.next,
1767 work_color_to_flags(WORK_NO_COLOR));
1771 static struct worker *alloc_worker(void)
1773 struct worker *worker;
1775 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1777 INIT_LIST_HEAD(&worker->entry);
1778 INIT_LIST_HEAD(&worker->scheduled);
1779 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1780 /* on creation a worker is in !idle && prep state */
1781 worker->flags = WORKER_PREP;
1787 * create_worker - create a new workqueue worker
1788 * @pool: pool the new worker will belong to
1790 * Create a new worker which is bound to @pool. The returned worker
1791 * can be started by calling start_worker() or destroyed using
1795 * Might sleep. Does GFP_KERNEL allocations.
1798 * Pointer to the newly created worker.
1800 static struct worker *create_worker(struct worker_pool *pool)
1802 struct global_cwq *gcwq = pool->gcwq;
1803 const char *pri = worker_pool_pri(pool) ? "H" : "";
1804 struct worker *worker = NULL;
1807 spin_lock_irq(&gcwq->lock);
1808 while (ida_get_new(&pool->worker_ida, &id)) {
1809 spin_unlock_irq(&gcwq->lock);
1810 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1812 spin_lock_irq(&gcwq->lock);
1814 spin_unlock_irq(&gcwq->lock);
1816 worker = alloc_worker();
1820 worker->pool = pool;
1823 if (gcwq->cpu != WORK_CPU_UNBOUND)
1824 worker->task = kthread_create_on_node(worker_thread,
1825 worker, cpu_to_node(gcwq->cpu),
1826 "kworker/%u:%d%s", gcwq->cpu, id, pri);
1828 worker->task = kthread_create(worker_thread, worker,
1829 "kworker/u:%d%s", id, pri);
1830 if (IS_ERR(worker->task))
1833 if (worker_pool_pri(pool))
1834 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1837 * Determine CPU binding of the new worker depending on
1838 * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1839 * flag remains stable across this function. See the comments
1840 * above the flag definition for details.
1842 * As an unbound worker may later become a regular one if CPU comes
1843 * online, make sure every worker has %PF_THREAD_BOUND set.
1845 if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
1846 kthread_bind(worker->task, gcwq->cpu);
1848 worker->task->flags |= PF_THREAD_BOUND;
1849 worker->flags |= WORKER_UNBOUND;
1855 spin_lock_irq(&gcwq->lock);
1856 ida_remove(&pool->worker_ida, id);
1857 spin_unlock_irq(&gcwq->lock);
1864 * start_worker - start a newly created worker
1865 * @worker: worker to start
1867 * Make the gcwq aware of @worker and start it.
1870 * spin_lock_irq(gcwq->lock).
1872 static void start_worker(struct worker *worker)
1874 worker->flags |= WORKER_STARTED;
1875 worker->pool->nr_workers++;
1876 worker_enter_idle(worker);
1877 wake_up_process(worker->task);
1881 * destroy_worker - destroy a workqueue worker
1882 * @worker: worker to be destroyed
1884 * Destroy @worker and adjust @gcwq stats accordingly.
1887 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1889 static void destroy_worker(struct worker *worker)
1891 struct worker_pool *pool = worker->pool;
1892 struct global_cwq *gcwq = pool->gcwq;
1893 int id = worker->id;
1895 /* sanity check frenzy */
1896 BUG_ON(worker->current_work);
1897 BUG_ON(!list_empty(&worker->scheduled));
1899 if (worker->flags & WORKER_STARTED)
1901 if (worker->flags & WORKER_IDLE)
1904 list_del_init(&worker->entry);
1905 worker->flags |= WORKER_DIE;
1907 spin_unlock_irq(&gcwq->lock);
1909 kthread_stop(worker->task);
1912 spin_lock_irq(&gcwq->lock);
1913 ida_remove(&pool->worker_ida, id);
1916 static void idle_worker_timeout(unsigned long __pool)
1918 struct worker_pool *pool = (void *)__pool;
1919 struct global_cwq *gcwq = pool->gcwq;
1921 spin_lock_irq(&gcwq->lock);
1923 if (too_many_workers(pool)) {
1924 struct worker *worker;
1925 unsigned long expires;
1927 /* idle_list is kept in LIFO order, check the last one */
1928 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1929 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1931 if (time_before(jiffies, expires))
1932 mod_timer(&pool->idle_timer, expires);
1934 /* it's been idle for too long, wake up manager */
1935 pool->flags |= POOL_MANAGE_WORKERS;
1936 wake_up_worker(pool);
1940 spin_unlock_irq(&gcwq->lock);
1943 static bool send_mayday(struct work_struct *work)
1945 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1946 struct workqueue_struct *wq = cwq->wq;
1949 if (!(wq->flags & WQ_RESCUER))
1952 /* mayday mayday mayday */
1953 cpu = cwq->pool->gcwq->cpu;
1954 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1955 if (cpu == WORK_CPU_UNBOUND)
1957 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1958 wake_up_process(wq->rescuer->task);
1962 static void gcwq_mayday_timeout(unsigned long __pool)
1964 struct worker_pool *pool = (void *)__pool;
1965 struct global_cwq *gcwq = pool->gcwq;
1966 struct work_struct *work;
1968 spin_lock_irq(&gcwq->lock);
1970 if (need_to_create_worker(pool)) {
1972 * We've been trying to create a new worker but
1973 * haven't been successful. We might be hitting an
1974 * allocation deadlock. Send distress signals to
1977 list_for_each_entry(work, &pool->worklist, entry)
1981 spin_unlock_irq(&gcwq->lock);
1983 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1987 * maybe_create_worker - create a new worker if necessary
1988 * @pool: pool to create a new worker for
1990 * Create a new worker for @pool if necessary. @pool is guaranteed to
1991 * have at least one idle worker on return from this function. If
1992 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1993 * sent to all rescuers with works scheduled on @pool to resolve
1994 * possible allocation deadlock.
1996 * On return, need_to_create_worker() is guaranteed to be false and
1997 * may_start_working() true.
2000 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2001 * multiple times. Does GFP_KERNEL allocations. Called only from
2005 * false if no action was taken and gcwq->lock stayed locked, true
2008 static bool maybe_create_worker(struct worker_pool *pool)
2009 __releases(&gcwq->lock)
2010 __acquires(&gcwq->lock)
2012 struct global_cwq *gcwq = pool->gcwq;
2014 if (!need_to_create_worker(pool))
2017 spin_unlock_irq(&gcwq->lock);
2019 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2020 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2023 struct worker *worker;
2025 worker = create_worker(pool);
2027 del_timer_sync(&pool->mayday_timer);
2028 spin_lock_irq(&gcwq->lock);
2029 start_worker(worker);
2030 BUG_ON(need_to_create_worker(pool));
2034 if (!need_to_create_worker(pool))
2037 __set_current_state(TASK_INTERRUPTIBLE);
2038 schedule_timeout(CREATE_COOLDOWN);
2040 if (!need_to_create_worker(pool))
2044 del_timer_sync(&pool->mayday_timer);
2045 spin_lock_irq(&gcwq->lock);
2046 if (need_to_create_worker(pool))
2052 * maybe_destroy_worker - destroy workers which have been idle for a while
2053 * @pool: pool to destroy workers for
2055 * Destroy @pool workers which have been idle for longer than
2056 * IDLE_WORKER_TIMEOUT.
2059 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2060 * multiple times. Called only from manager.
2063 * false if no action was taken and gcwq->lock stayed locked, true
2066 static bool maybe_destroy_workers(struct worker_pool *pool)
2070 while (too_many_workers(pool)) {
2071 struct worker *worker;
2072 unsigned long expires;
2074 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2075 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2077 if (time_before(jiffies, expires)) {
2078 mod_timer(&pool->idle_timer, expires);
2082 destroy_worker(worker);
2090 * manage_workers - manage worker pool
2093 * Assume the manager role and manage gcwq worker pool @worker belongs
2094 * to. At any given time, there can be only zero or one manager per
2095 * gcwq. The exclusion is handled automatically by this function.
2097 * The caller can safely start processing works on false return. On
2098 * true return, it's guaranteed that need_to_create_worker() is false
2099 * and may_start_working() is true.
2102 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2103 * multiple times. Does GFP_KERNEL allocations.
2106 * false if no action was taken and gcwq->lock stayed locked, true if
2107 * some action was taken.
2109 static bool manage_workers(struct worker *worker)
2111 struct worker_pool *pool = worker->pool;
2114 if (!mutex_trylock(&pool->manager_mutex))
2117 pool->flags &= ~POOL_MANAGE_WORKERS;
2120 * Destroy and then create so that may_start_working() is true
2123 ret |= maybe_destroy_workers(pool);
2124 ret |= maybe_create_worker(pool);
2126 mutex_unlock(&pool->manager_mutex);
2131 * process_one_work - process single work
2133 * @work: work to process
2135 * Process @work. This function contains all the logics necessary to
2136 * process a single work including synchronization against and
2137 * interaction with other workers on the same cpu, queueing and
2138 * flushing. As long as context requirement is met, any worker can
2139 * call this function to process a work.
2142 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2144 static void process_one_work(struct worker *worker, struct work_struct *work)
2145 __releases(&gcwq->lock)
2146 __acquires(&gcwq->lock)
2148 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2149 struct worker_pool *pool = worker->pool;
2150 struct global_cwq *gcwq = pool->gcwq;
2151 struct hlist_head *bwh = busy_worker_head(gcwq, work);
2152 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2153 work_func_t f = work->func;
2155 struct worker *collision;
2156 #ifdef CONFIG_LOCKDEP
2158 * It is permissible to free the struct work_struct from
2159 * inside the function that is called from it, this we need to
2160 * take into account for lockdep too. To avoid bogus "held
2161 * lock freed" warnings as well as problems when looking into
2162 * work->lockdep_map, make a copy and use that here.
2164 struct lockdep_map lockdep_map;
2166 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2169 * Ensure we're on the correct CPU. DISASSOCIATED test is
2170 * necessary to avoid spurious warnings from rescuers servicing the
2171 * unbound or a disassociated gcwq.
2173 WARN_ON_ONCE(!(worker->flags & (WORKER_UNBOUND | WORKER_REBIND)) &&
2174 !(gcwq->flags & GCWQ_DISASSOCIATED) &&
2175 raw_smp_processor_id() != gcwq->cpu);
2178 * A single work shouldn't be executed concurrently by
2179 * multiple workers on a single cpu. Check whether anyone is
2180 * already processing the work. If so, defer the work to the
2181 * currently executing one.
2183 collision = __find_worker_executing_work(gcwq, bwh, work);
2184 if (unlikely(collision)) {
2185 move_linked_works(work, &collision->scheduled, NULL);
2189 /* claim and dequeue */
2190 debug_work_deactivate(work);
2191 hlist_add_head(&worker->hentry, bwh);
2192 worker->current_work = work;
2193 worker->current_cwq = cwq;
2194 work_color = get_work_color(work);
2196 list_del_init(&work->entry);
2199 * CPU intensive works don't participate in concurrency
2200 * management. They're the scheduler's responsibility.
2202 if (unlikely(cpu_intensive))
2203 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2206 * Unbound gcwq isn't concurrency managed and work items should be
2207 * executed ASAP. Wake up another worker if necessary.
2209 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2210 wake_up_worker(pool);
2213 * Record the last CPU and clear PENDING which should be the last
2214 * update to @work. Also, do this inside @gcwq->lock so that
2215 * PENDING and queued state changes happen together while IRQ is
2218 set_work_cpu_and_clear_pending(work, gcwq->cpu);
2220 spin_unlock_irq(&gcwq->lock);
2222 lock_map_acquire_read(&cwq->wq->lockdep_map);
2223 lock_map_acquire(&lockdep_map);
2224 trace_workqueue_execute_start(work);
2227 * While we must be careful to not use "work" after this, the trace
2228 * point will only record its address.
2230 trace_workqueue_execute_end(work);
2231 lock_map_release(&lockdep_map);
2232 lock_map_release(&cwq->wq->lockdep_map);
2234 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2235 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2236 " last function: %pf\n",
2237 current->comm, preempt_count(), task_pid_nr(current), f);
2238 debug_show_held_locks(current);
2242 spin_lock_irq(&gcwq->lock);
2244 /* clear cpu intensive status */
2245 if (unlikely(cpu_intensive))
2246 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2248 /* we're done with it, release */
2249 hlist_del_init(&worker->hentry);
2250 worker->current_work = NULL;
2251 worker->current_cwq = NULL;
2252 cwq_dec_nr_in_flight(cwq, work_color, false);
2256 * process_scheduled_works - process scheduled works
2259 * Process all scheduled works. Please note that the scheduled list
2260 * may change while processing a work, so this function repeatedly
2261 * fetches a work from the top and executes it.
2264 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2267 static void process_scheduled_works(struct worker *worker)
2269 while (!list_empty(&worker->scheduled)) {
2270 struct work_struct *work = list_first_entry(&worker->scheduled,
2271 struct work_struct, entry);
2272 process_one_work(worker, work);
2277 * worker_thread - the worker thread function
2280 * The gcwq worker thread function. There's a single dynamic pool of
2281 * these per each cpu. These workers process all works regardless of
2282 * their specific target workqueue. The only exception is works which
2283 * belong to workqueues with a rescuer which will be explained in
2286 static int worker_thread(void *__worker)
2288 struct worker *worker = __worker;
2289 struct worker_pool *pool = worker->pool;
2290 struct global_cwq *gcwq = pool->gcwq;
2292 /* tell the scheduler that this is a workqueue worker */
2293 worker->task->flags |= PF_WQ_WORKER;
2295 spin_lock_irq(&gcwq->lock);
2298 * DIE can be set only while idle and REBIND set while busy has
2299 * @worker->rebind_work scheduled. Checking here is enough.
2301 if (unlikely(worker->flags & (WORKER_REBIND | WORKER_DIE))) {
2302 spin_unlock_irq(&gcwq->lock);
2304 if (worker->flags & WORKER_DIE) {
2305 worker->task->flags &= ~PF_WQ_WORKER;
2309 idle_worker_rebind(worker);
2313 worker_leave_idle(worker);
2315 /* no more worker necessary? */
2316 if (!need_more_worker(pool))
2319 /* do we need to manage? */
2320 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2324 * ->scheduled list can only be filled while a worker is
2325 * preparing to process a work or actually processing it.
2326 * Make sure nobody diddled with it while I was sleeping.
2328 BUG_ON(!list_empty(&worker->scheduled));
2331 * When control reaches this point, we're guaranteed to have
2332 * at least one idle worker or that someone else has already
2333 * assumed the manager role.
2335 worker_clr_flags(worker, WORKER_PREP);
2338 struct work_struct *work =
2339 list_first_entry(&pool->worklist,
2340 struct work_struct, entry);
2342 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2343 /* optimization path, not strictly necessary */
2344 process_one_work(worker, work);
2345 if (unlikely(!list_empty(&worker->scheduled)))
2346 process_scheduled_works(worker);
2348 move_linked_works(work, &worker->scheduled, NULL);
2349 process_scheduled_works(worker);
2351 } while (keep_working(pool));
2353 worker_set_flags(worker, WORKER_PREP, false);
2355 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2359 * gcwq->lock is held and there's no work to process and no
2360 * need to manage, sleep. Workers are woken up only while
2361 * holding gcwq->lock or from local cpu, so setting the
2362 * current state before releasing gcwq->lock is enough to
2363 * prevent losing any event.
2365 worker_enter_idle(worker);
2366 __set_current_state(TASK_INTERRUPTIBLE);
2367 spin_unlock_irq(&gcwq->lock);
2373 * rescuer_thread - the rescuer thread function
2374 * @__wq: the associated workqueue
2376 * Workqueue rescuer thread function. There's one rescuer for each
2377 * workqueue which has WQ_RESCUER set.
2379 * Regular work processing on a gcwq may block trying to create a new
2380 * worker which uses GFP_KERNEL allocation which has slight chance of
2381 * developing into deadlock if some works currently on the same queue
2382 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2383 * the problem rescuer solves.
2385 * When such condition is possible, the gcwq summons rescuers of all
2386 * workqueues which have works queued on the gcwq and let them process
2387 * those works so that forward progress can be guaranteed.
2389 * This should happen rarely.
2391 static int rescuer_thread(void *__wq)
2393 struct workqueue_struct *wq = __wq;
2394 struct worker *rescuer = wq->rescuer;
2395 struct list_head *scheduled = &rescuer->scheduled;
2396 bool is_unbound = wq->flags & WQ_UNBOUND;
2399 set_user_nice(current, RESCUER_NICE_LEVEL);
2401 set_current_state(TASK_INTERRUPTIBLE);
2403 if (kthread_should_stop())
2407 * See whether any cpu is asking for help. Unbounded
2408 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2410 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2411 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2412 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2413 struct worker_pool *pool = cwq->pool;
2414 struct global_cwq *gcwq = pool->gcwq;
2415 struct work_struct *work, *n;
2417 __set_current_state(TASK_RUNNING);
2418 mayday_clear_cpu(cpu, wq->mayday_mask);
2420 /* migrate to the target cpu if possible */
2421 rescuer->pool = pool;
2422 worker_maybe_bind_and_lock(rescuer);
2425 * Slurp in all works issued via this workqueue and
2428 BUG_ON(!list_empty(&rescuer->scheduled));
2429 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2430 if (get_work_cwq(work) == cwq)
2431 move_linked_works(work, scheduled, &n);
2433 process_scheduled_works(rescuer);
2436 * Leave this gcwq. If keep_working() is %true, notify a
2437 * regular worker; otherwise, we end up with 0 concurrency
2438 * and stalling the execution.
2440 if (keep_working(pool))
2441 wake_up_worker(pool);
2443 spin_unlock_irq(&gcwq->lock);
2451 struct work_struct work;
2452 struct completion done;
2455 static void wq_barrier_func(struct work_struct *work)
2457 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2458 complete(&barr->done);
2462 * insert_wq_barrier - insert a barrier work
2463 * @cwq: cwq to insert barrier into
2464 * @barr: wq_barrier to insert
2465 * @target: target work to attach @barr to
2466 * @worker: worker currently executing @target, NULL if @target is not executing
2468 * @barr is linked to @target such that @barr is completed only after
2469 * @target finishes execution. Please note that the ordering
2470 * guarantee is observed only with respect to @target and on the local
2473 * Currently, a queued barrier can't be canceled. This is because
2474 * try_to_grab_pending() can't determine whether the work to be
2475 * grabbed is at the head of the queue and thus can't clear LINKED
2476 * flag of the previous work while there must be a valid next work
2477 * after a work with LINKED flag set.
2479 * Note that when @worker is non-NULL, @target may be modified
2480 * underneath us, so we can't reliably determine cwq from @target.
2483 * spin_lock_irq(gcwq->lock).
2485 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2486 struct wq_barrier *barr,
2487 struct work_struct *target, struct worker *worker)
2489 struct list_head *head;
2490 unsigned int linked = 0;
2493 * debugobject calls are safe here even with gcwq->lock locked
2494 * as we know for sure that this will not trigger any of the
2495 * checks and call back into the fixup functions where we
2498 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2499 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2500 init_completion(&barr->done);
2503 * If @target is currently being executed, schedule the
2504 * barrier to the worker; otherwise, put it after @target.
2507 head = worker->scheduled.next;
2509 unsigned long *bits = work_data_bits(target);
2511 head = target->entry.next;
2512 /* there can already be other linked works, inherit and set */
2513 linked = *bits & WORK_STRUCT_LINKED;
2514 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2517 debug_work_activate(&barr->work);
2518 insert_work(cwq, &barr->work, head,
2519 work_color_to_flags(WORK_NO_COLOR) | linked);
2523 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2524 * @wq: workqueue being flushed
2525 * @flush_color: new flush color, < 0 for no-op
2526 * @work_color: new work color, < 0 for no-op
2528 * Prepare cwqs for workqueue flushing.
2530 * If @flush_color is non-negative, flush_color on all cwqs should be
2531 * -1. If no cwq has in-flight commands at the specified color, all
2532 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2533 * has in flight commands, its cwq->flush_color is set to
2534 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2535 * wakeup logic is armed and %true is returned.
2537 * The caller should have initialized @wq->first_flusher prior to
2538 * calling this function with non-negative @flush_color. If
2539 * @flush_color is negative, no flush color update is done and %false
2542 * If @work_color is non-negative, all cwqs should have the same
2543 * work_color which is previous to @work_color and all will be
2544 * advanced to @work_color.
2547 * mutex_lock(wq->flush_mutex).
2550 * %true if @flush_color >= 0 and there's something to flush. %false
2553 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2554 int flush_color, int work_color)
2559 if (flush_color >= 0) {
2560 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2561 atomic_set(&wq->nr_cwqs_to_flush, 1);
2564 for_each_cwq_cpu(cpu, wq) {
2565 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2566 struct global_cwq *gcwq = cwq->pool->gcwq;
2568 spin_lock_irq(&gcwq->lock);
2570 if (flush_color >= 0) {
2571 BUG_ON(cwq->flush_color != -1);
2573 if (cwq->nr_in_flight[flush_color]) {
2574 cwq->flush_color = flush_color;
2575 atomic_inc(&wq->nr_cwqs_to_flush);
2580 if (work_color >= 0) {
2581 BUG_ON(work_color != work_next_color(cwq->work_color));
2582 cwq->work_color = work_color;
2585 spin_unlock_irq(&gcwq->lock);
2588 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2589 complete(&wq->first_flusher->done);
2595 * flush_workqueue - ensure that any scheduled work has run to completion.
2596 * @wq: workqueue to flush
2598 * Forces execution of the workqueue and blocks until its completion.
2599 * This is typically used in driver shutdown handlers.
2601 * We sleep until all works which were queued on entry have been handled,
2602 * but we are not livelocked by new incoming ones.
2604 void flush_workqueue(struct workqueue_struct *wq)
2606 struct wq_flusher this_flusher = {
2607 .list = LIST_HEAD_INIT(this_flusher.list),
2609 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2613 lock_map_acquire(&wq->lockdep_map);
2614 lock_map_release(&wq->lockdep_map);
2616 mutex_lock(&wq->flush_mutex);
2619 * Start-to-wait phase
2621 next_color = work_next_color(wq->work_color);
2623 if (next_color != wq->flush_color) {
2625 * Color space is not full. The current work_color
2626 * becomes our flush_color and work_color is advanced
2629 BUG_ON(!list_empty(&wq->flusher_overflow));
2630 this_flusher.flush_color = wq->work_color;
2631 wq->work_color = next_color;
2633 if (!wq->first_flusher) {
2634 /* no flush in progress, become the first flusher */
2635 BUG_ON(wq->flush_color != this_flusher.flush_color);
2637 wq->first_flusher = &this_flusher;
2639 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2641 /* nothing to flush, done */
2642 wq->flush_color = next_color;
2643 wq->first_flusher = NULL;
2648 BUG_ON(wq->flush_color == this_flusher.flush_color);
2649 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2650 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2654 * Oops, color space is full, wait on overflow queue.
2655 * The next flush completion will assign us
2656 * flush_color and transfer to flusher_queue.
2658 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2661 mutex_unlock(&wq->flush_mutex);
2663 wait_for_completion(&this_flusher.done);
2666 * Wake-up-and-cascade phase
2668 * First flushers are responsible for cascading flushes and
2669 * handling overflow. Non-first flushers can simply return.
2671 if (wq->first_flusher != &this_flusher)
2674 mutex_lock(&wq->flush_mutex);
2676 /* we might have raced, check again with mutex held */
2677 if (wq->first_flusher != &this_flusher)
2680 wq->first_flusher = NULL;
2682 BUG_ON(!list_empty(&this_flusher.list));
2683 BUG_ON(wq->flush_color != this_flusher.flush_color);
2686 struct wq_flusher *next, *tmp;
2688 /* complete all the flushers sharing the current flush color */
2689 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2690 if (next->flush_color != wq->flush_color)
2692 list_del_init(&next->list);
2693 complete(&next->done);
2696 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2697 wq->flush_color != work_next_color(wq->work_color));
2699 /* this flush_color is finished, advance by one */
2700 wq->flush_color = work_next_color(wq->flush_color);
2702 /* one color has been freed, handle overflow queue */
2703 if (!list_empty(&wq->flusher_overflow)) {
2705 * Assign the same color to all overflowed
2706 * flushers, advance work_color and append to
2707 * flusher_queue. This is the start-to-wait
2708 * phase for these overflowed flushers.
2710 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2711 tmp->flush_color = wq->work_color;
2713 wq->work_color = work_next_color(wq->work_color);
2715 list_splice_tail_init(&wq->flusher_overflow,
2716 &wq->flusher_queue);
2717 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2720 if (list_empty(&wq->flusher_queue)) {
2721 BUG_ON(wq->flush_color != wq->work_color);
2726 * Need to flush more colors. Make the next flusher
2727 * the new first flusher and arm cwqs.
2729 BUG_ON(wq->flush_color == wq->work_color);
2730 BUG_ON(wq->flush_color != next->flush_color);
2732 list_del_init(&next->list);
2733 wq->first_flusher = next;
2735 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2739 * Meh... this color is already done, clear first
2740 * flusher and repeat cascading.
2742 wq->first_flusher = NULL;
2746 mutex_unlock(&wq->flush_mutex);
2748 EXPORT_SYMBOL_GPL(flush_workqueue);
2751 * drain_workqueue - drain a workqueue
2752 * @wq: workqueue to drain
2754 * Wait until the workqueue becomes empty. While draining is in progress,
2755 * only chain queueing is allowed. IOW, only currently pending or running
2756 * work items on @wq can queue further work items on it. @wq is flushed
2757 * repeatedly until it becomes empty. The number of flushing is detemined
2758 * by the depth of chaining and should be relatively short. Whine if it
2761 void drain_workqueue(struct workqueue_struct *wq)
2763 unsigned int flush_cnt = 0;
2767 * __queue_work() needs to test whether there are drainers, is much
2768 * hotter than drain_workqueue() and already looks at @wq->flags.
2769 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2771 spin_lock(&workqueue_lock);
2772 if (!wq->nr_drainers++)
2773 wq->flags |= WQ_DRAINING;
2774 spin_unlock(&workqueue_lock);
2776 flush_workqueue(wq);
2778 for_each_cwq_cpu(cpu, wq) {
2779 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2782 spin_lock_irq(&cwq->pool->gcwq->lock);
2783 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2784 spin_unlock_irq(&cwq->pool->gcwq->lock);
2789 if (++flush_cnt == 10 ||
2790 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2791 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2792 wq->name, flush_cnt);
2796 spin_lock(&workqueue_lock);
2797 if (!--wq->nr_drainers)
2798 wq->flags &= ~WQ_DRAINING;
2799 spin_unlock(&workqueue_lock);
2801 EXPORT_SYMBOL_GPL(drain_workqueue);
2803 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2804 bool wait_executing)
2806 struct worker *worker = NULL;
2807 struct global_cwq *gcwq;
2808 struct cpu_workqueue_struct *cwq;
2811 gcwq = get_work_gcwq(work);
2815 spin_lock_irq(&gcwq->lock);
2816 if (!list_empty(&work->entry)) {
2818 * See the comment near try_to_grab_pending()->smp_rmb().
2819 * If it was re-queued to a different gcwq under us, we
2820 * are not going to wait.
2823 cwq = get_work_cwq(work);
2824 if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
2826 } else if (wait_executing) {
2827 worker = find_worker_executing_work(gcwq, work);
2830 cwq = worker->current_cwq;
2834 insert_wq_barrier(cwq, barr, work, worker);
2835 spin_unlock_irq(&gcwq->lock);
2838 * If @max_active is 1 or rescuer is in use, flushing another work
2839 * item on the same workqueue may lead to deadlock. Make sure the
2840 * flusher is not running on the same workqueue by verifying write
2843 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2844 lock_map_acquire(&cwq->wq->lockdep_map);
2846 lock_map_acquire_read(&cwq->wq->lockdep_map);
2847 lock_map_release(&cwq->wq->lockdep_map);
2851 spin_unlock_irq(&gcwq->lock);
2856 * flush_work - wait for a work to finish executing the last queueing instance
2857 * @work: the work to flush
2859 * Wait until @work has finished execution. This function considers
2860 * only the last queueing instance of @work. If @work has been
2861 * enqueued across different CPUs on a non-reentrant workqueue or on
2862 * multiple workqueues, @work might still be executing on return on
2863 * some of the CPUs from earlier queueing.
2865 * If @work was queued only on a non-reentrant, ordered or unbound
2866 * workqueue, @work is guaranteed to be idle on return if it hasn't
2867 * been requeued since flush started.
2870 * %true if flush_work() waited for the work to finish execution,
2871 * %false if it was already idle.
2873 bool flush_work(struct work_struct *work)
2875 struct wq_barrier barr;
2877 lock_map_acquire(&work->lockdep_map);
2878 lock_map_release(&work->lockdep_map);
2880 if (start_flush_work(work, &barr, true)) {
2881 wait_for_completion(&barr.done);
2882 destroy_work_on_stack(&barr.work);
2887 EXPORT_SYMBOL_GPL(flush_work);
2889 static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2891 struct wq_barrier barr;
2892 struct worker *worker;
2894 spin_lock_irq(&gcwq->lock);
2896 worker = find_worker_executing_work(gcwq, work);
2897 if (unlikely(worker))
2898 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2900 spin_unlock_irq(&gcwq->lock);
2902 if (unlikely(worker)) {
2903 wait_for_completion(&barr.done);
2904 destroy_work_on_stack(&barr.work);
2910 static bool wait_on_work(struct work_struct *work)
2917 lock_map_acquire(&work->lockdep_map);
2918 lock_map_release(&work->lockdep_map);
2920 for_each_gcwq_cpu(cpu)
2921 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2926 * flush_work_sync - wait until a work has finished execution
2927 * @work: the work to flush
2929 * Wait until @work has finished execution. On return, it's
2930 * guaranteed that all queueing instances of @work which happened
2931 * before this function is called are finished. In other words, if
2932 * @work hasn't been requeued since this function was called, @work is
2933 * guaranteed to be idle on return.
2936 * %true if flush_work_sync() waited for the work to finish execution,
2937 * %false if it was already idle.
2939 bool flush_work_sync(struct work_struct *work)
2941 struct wq_barrier barr;
2942 bool pending, waited;
2944 /* we'll wait for executions separately, queue barr only if pending */
2945 pending = start_flush_work(work, &barr, false);
2947 /* wait for executions to finish */
2948 waited = wait_on_work(work);
2950 /* wait for the pending one */
2952 wait_for_completion(&barr.done);
2953 destroy_work_on_stack(&barr.work);
2956 return pending || waited;
2958 EXPORT_SYMBOL_GPL(flush_work_sync);
2960 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2962 unsigned long flags;
2966 ret = try_to_grab_pending(work, is_dwork, &flags);
2968 * If someone else is canceling, wait for the same event it
2969 * would be waiting for before retrying.
2971 if (unlikely(ret == -ENOENT))
2973 } while (unlikely(ret < 0));
2975 /* tell other tasks trying to grab @work to back off */
2976 mark_work_canceling(work);
2977 local_irq_restore(flags);
2980 clear_work_data(work);
2985 * cancel_work_sync - cancel a work and wait for it to finish
2986 * @work: the work to cancel
2988 * Cancel @work and wait for its execution to finish. This function
2989 * can be used even if the work re-queues itself or migrates to
2990 * another workqueue. On return from this function, @work is
2991 * guaranteed to be not pending or executing on any CPU.
2993 * cancel_work_sync(&delayed_work->work) must not be used for
2994 * delayed_work's. Use cancel_delayed_work_sync() instead.
2996 * The caller must ensure that the workqueue on which @work was last
2997 * queued can't be destroyed before this function returns.
3000 * %true if @work was pending, %false otherwise.
3002 bool cancel_work_sync(struct work_struct *work)
3004 return __cancel_work_timer(work, false);
3006 EXPORT_SYMBOL_GPL(cancel_work_sync);
3009 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3010 * @dwork: the delayed work to flush
3012 * Delayed timer is cancelled and the pending work is queued for
3013 * immediate execution. Like flush_work(), this function only
3014 * considers the last queueing instance of @dwork.
3017 * %true if flush_work() waited for the work to finish execution,
3018 * %false if it was already idle.
3020 bool flush_delayed_work(struct delayed_work *dwork)
3022 local_irq_disable();
3023 if (del_timer_sync(&dwork->timer))
3024 __queue_work(dwork->cpu,
3025 get_work_cwq(&dwork->work)->wq, &dwork->work);
3027 return flush_work(&dwork->work);
3029 EXPORT_SYMBOL(flush_delayed_work);
3032 * flush_delayed_work_sync - wait for a dwork to finish
3033 * @dwork: the delayed work to flush
3035 * Delayed timer is cancelled and the pending work is queued for
3036 * execution immediately. Other than timer handling, its behavior
3037 * is identical to flush_work_sync().
3040 * %true if flush_work_sync() waited for the work to finish execution,
3041 * %false if it was already idle.
3043 bool flush_delayed_work_sync(struct delayed_work *dwork)
3045 local_irq_disable();
3046 if (del_timer_sync(&dwork->timer))
3047 __queue_work(dwork->cpu,
3048 get_work_cwq(&dwork->work)->wq, &dwork->work);
3050 return flush_work_sync(&dwork->work);
3052 EXPORT_SYMBOL(flush_delayed_work_sync);
3055 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3056 * @dwork: the delayed work cancel
3058 * This is cancel_work_sync() for delayed works.
3061 * %true if @dwork was pending, %false otherwise.
3063 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3065 return __cancel_work_timer(&dwork->work, true);
3067 EXPORT_SYMBOL(cancel_delayed_work_sync);
3070 * schedule_work_on - put work task on a specific cpu
3071 * @cpu: cpu to put the work task on
3072 * @work: job to be done
3074 * This puts a job on a specific cpu
3076 bool schedule_work_on(int cpu, struct work_struct *work)
3078 return queue_work_on(cpu, system_wq, work);
3080 EXPORT_SYMBOL(schedule_work_on);
3083 * schedule_work - put work task in global workqueue
3084 * @work: job to be done
3086 * Returns %false if @work was already on the kernel-global workqueue and
3089 * This puts a job in the kernel-global workqueue if it was not already
3090 * queued and leaves it in the same position on the kernel-global
3091 * workqueue otherwise.
3093 bool schedule_work(struct work_struct *work)
3095 return queue_work(system_wq, work);
3097 EXPORT_SYMBOL(schedule_work);
3100 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3102 * @dwork: job to be done
3103 * @delay: number of jiffies to wait
3105 * After waiting for a given time this puts a job in the kernel-global
3106 * workqueue on the specified CPU.
3108 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3109 unsigned long delay)
3111 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3113 EXPORT_SYMBOL(schedule_delayed_work_on);
3116 * schedule_delayed_work - put work task in global workqueue after delay
3117 * @dwork: job to be done
3118 * @delay: number of jiffies to wait or 0 for immediate execution
3120 * After waiting for a given time this puts a job in the kernel-global
3123 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3125 return queue_delayed_work(system_wq, dwork, delay);
3127 EXPORT_SYMBOL(schedule_delayed_work);
3130 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3131 * @func: the function to call
3133 * schedule_on_each_cpu() executes @func on each online CPU using the
3134 * system workqueue and blocks until all CPUs have completed.
3135 * schedule_on_each_cpu() is very slow.
3138 * 0 on success, -errno on failure.
3140 int schedule_on_each_cpu(work_func_t func)
3143 struct work_struct __percpu *works;
3145 works = alloc_percpu(struct work_struct);
3151 for_each_online_cpu(cpu) {
3152 struct work_struct *work = per_cpu_ptr(works, cpu);
3154 INIT_WORK(work, func);
3155 schedule_work_on(cpu, work);
3158 for_each_online_cpu(cpu)
3159 flush_work(per_cpu_ptr(works, cpu));
3167 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3169 * Forces execution of the kernel-global workqueue and blocks until its
3172 * Think twice before calling this function! It's very easy to get into
3173 * trouble if you don't take great care. Either of the following situations
3174 * will lead to deadlock:
3176 * One of the work items currently on the workqueue needs to acquire
3177 * a lock held by your code or its caller.
3179 * Your code is running in the context of a work routine.
3181 * They will be detected by lockdep when they occur, but the first might not
3182 * occur very often. It depends on what work items are on the workqueue and
3183 * what locks they need, which you have no control over.
3185 * In most situations flushing the entire workqueue is overkill; you merely
3186 * need to know that a particular work item isn't queued and isn't running.
3187 * In such cases you should use cancel_delayed_work_sync() or
3188 * cancel_work_sync() instead.
3190 void flush_scheduled_work(void)
3192 flush_workqueue(system_wq);
3194 EXPORT_SYMBOL(flush_scheduled_work);
3197 * execute_in_process_context - reliably execute the routine with user context
3198 * @fn: the function to execute
3199 * @ew: guaranteed storage for the execute work structure (must
3200 * be available when the work executes)
3202 * Executes the function immediately if process context is available,
3203 * otherwise schedules the function for delayed execution.
3205 * Returns: 0 - function was executed
3206 * 1 - function was scheduled for execution
3208 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3210 if (!in_interrupt()) {
3215 INIT_WORK(&ew->work, fn);
3216 schedule_work(&ew->work);
3220 EXPORT_SYMBOL_GPL(execute_in_process_context);
3222 int keventd_up(void)
3224 return system_wq != NULL;
3227 static int alloc_cwqs(struct workqueue_struct *wq)
3230 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3231 * Make sure that the alignment isn't lower than that of
3232 * unsigned long long.
3234 const size_t size = sizeof(struct cpu_workqueue_struct);
3235 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3236 __alignof__(unsigned long long));
3238 if (!(wq->flags & WQ_UNBOUND))
3239 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3244 * Allocate enough room to align cwq and put an extra
3245 * pointer at the end pointing back to the originally
3246 * allocated pointer which will be used for free.
3248 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3250 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3251 *(void **)(wq->cpu_wq.single + 1) = ptr;
3255 /* just in case, make sure it's actually aligned */
3256 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3257 return wq->cpu_wq.v ? 0 : -ENOMEM;
3260 static void free_cwqs(struct workqueue_struct *wq)
3262 if (!(wq->flags & WQ_UNBOUND))
3263 free_percpu(wq->cpu_wq.pcpu);
3264 else if (wq->cpu_wq.single) {
3265 /* the pointer to free is stored right after the cwq */
3266 kfree(*(void **)(wq->cpu_wq.single + 1));
3270 static int wq_clamp_max_active(int max_active, unsigned int flags,
3273 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3275 if (max_active < 1 || max_active > lim)
3276 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3277 max_active, name, 1, lim);
3279 return clamp_val(max_active, 1, lim);
3282 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3285 struct lock_class_key *key,
3286 const char *lock_name, ...)
3288 va_list args, args1;
3289 struct workqueue_struct *wq;
3293 /* determine namelen, allocate wq and format name */
3294 va_start(args, lock_name);
3295 va_copy(args1, args);
3296 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3298 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3302 vsnprintf(wq->name, namelen, fmt, args1);
3307 * Workqueues which may be used during memory reclaim should
3308 * have a rescuer to guarantee forward progress.
3310 if (flags & WQ_MEM_RECLAIM)
3311 flags |= WQ_RESCUER;
3313 max_active = max_active ?: WQ_DFL_ACTIVE;
3314 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3318 wq->saved_max_active = max_active;
3319 mutex_init(&wq->flush_mutex);
3320 atomic_set(&wq->nr_cwqs_to_flush, 0);
3321 INIT_LIST_HEAD(&wq->flusher_queue);
3322 INIT_LIST_HEAD(&wq->flusher_overflow);
3324 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3325 INIT_LIST_HEAD(&wq->list);
3327 if (alloc_cwqs(wq) < 0)
3330 for_each_cwq_cpu(cpu, wq) {
3331 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3332 struct global_cwq *gcwq = get_gcwq(cpu);
3333 int pool_idx = (bool)(flags & WQ_HIGHPRI);
3335 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3336 cwq->pool = &gcwq->pools[pool_idx];
3338 cwq->flush_color = -1;
3339 cwq->max_active = max_active;
3340 INIT_LIST_HEAD(&cwq->delayed_works);
3343 if (flags & WQ_RESCUER) {
3344 struct worker *rescuer;
3346 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3349 wq->rescuer = rescuer = alloc_worker();
3353 rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3355 if (IS_ERR(rescuer->task))
3358 rescuer->task->flags |= PF_THREAD_BOUND;
3359 wake_up_process(rescuer->task);
3363 * workqueue_lock protects global freeze state and workqueues
3364 * list. Grab it, set max_active accordingly and add the new
3365 * workqueue to workqueues list.
3367 spin_lock(&workqueue_lock);
3369 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3370 for_each_cwq_cpu(cpu, wq)
3371 get_cwq(cpu, wq)->max_active = 0;
3373 list_add(&wq->list, &workqueues);
3375 spin_unlock(&workqueue_lock);
3381 free_mayday_mask(wq->mayday_mask);
3387 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3390 * destroy_workqueue - safely terminate a workqueue
3391 * @wq: target workqueue
3393 * Safely destroy a workqueue. All work currently pending will be done first.
3395 void destroy_workqueue(struct workqueue_struct *wq)
3399 /* drain it before proceeding with destruction */
3400 drain_workqueue(wq);
3403 * wq list is used to freeze wq, remove from list after
3404 * flushing is complete in case freeze races us.
3406 spin_lock(&workqueue_lock);
3407 list_del(&wq->list);
3408 spin_unlock(&workqueue_lock);
3411 for_each_cwq_cpu(cpu, wq) {
3412 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3415 for (i = 0; i < WORK_NR_COLORS; i++)
3416 BUG_ON(cwq->nr_in_flight[i]);
3417 BUG_ON(cwq->nr_active);
3418 BUG_ON(!list_empty(&cwq->delayed_works));
3421 if (wq->flags & WQ_RESCUER) {
3422 kthread_stop(wq->rescuer->task);
3423 free_mayday_mask(wq->mayday_mask);
3430 EXPORT_SYMBOL_GPL(destroy_workqueue);
3433 * workqueue_set_max_active - adjust max_active of a workqueue
3434 * @wq: target workqueue
3435 * @max_active: new max_active value.
3437 * Set max_active of @wq to @max_active.
3440 * Don't call from IRQ context.
3442 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3446 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3448 spin_lock(&workqueue_lock);
3450 wq->saved_max_active = max_active;
3452 for_each_cwq_cpu(cpu, wq) {
3453 struct global_cwq *gcwq = get_gcwq(cpu);
3455 spin_lock_irq(&gcwq->lock);
3457 if (!(wq->flags & WQ_FREEZABLE) ||
3458 !(gcwq->flags & GCWQ_FREEZING))
3459 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3461 spin_unlock_irq(&gcwq->lock);
3464 spin_unlock(&workqueue_lock);
3466 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3469 * workqueue_congested - test whether a workqueue is congested
3470 * @cpu: CPU in question
3471 * @wq: target workqueue
3473 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3474 * no synchronization around this function and the test result is
3475 * unreliable and only useful as advisory hints or for debugging.
3478 * %true if congested, %false otherwise.
3480 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3482 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3484 return !list_empty(&cwq->delayed_works);
3486 EXPORT_SYMBOL_GPL(workqueue_congested);
3489 * work_cpu - return the last known associated cpu for @work
3490 * @work: the work of interest
3493 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3495 unsigned int work_cpu(struct work_struct *work)
3497 struct global_cwq *gcwq = get_work_gcwq(work);
3499 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3501 EXPORT_SYMBOL_GPL(work_cpu);
3504 * work_busy - test whether a work is currently pending or running
3505 * @work: the work to be tested
3507 * Test whether @work is currently pending or running. There is no
3508 * synchronization around this function and the test result is
3509 * unreliable and only useful as advisory hints or for debugging.
3510 * Especially for reentrant wqs, the pending state might hide the
3514 * OR'd bitmask of WORK_BUSY_* bits.
3516 unsigned int work_busy(struct work_struct *work)
3518 struct global_cwq *gcwq = get_work_gcwq(work);
3519 unsigned long flags;
3520 unsigned int ret = 0;
3525 spin_lock_irqsave(&gcwq->lock, flags);
3527 if (work_pending(work))
3528 ret |= WORK_BUSY_PENDING;
3529 if (find_worker_executing_work(gcwq, work))
3530 ret |= WORK_BUSY_RUNNING;
3532 spin_unlock_irqrestore(&gcwq->lock, flags);
3536 EXPORT_SYMBOL_GPL(work_busy);
3541 * There are two challenges in supporting CPU hotplug. Firstly, there
3542 * are a lot of assumptions on strong associations among work, cwq and
3543 * gcwq which make migrating pending and scheduled works very
3544 * difficult to implement without impacting hot paths. Secondly,
3545 * gcwqs serve mix of short, long and very long running works making
3546 * blocked draining impractical.
3548 * This is solved by allowing a gcwq to be disassociated from the CPU
3549 * running as an unbound one and allowing it to be reattached later if the
3550 * cpu comes back online.
3553 /* claim manager positions of all pools */
3554 static void gcwq_claim_management_and_lock(struct global_cwq *gcwq)
3556 struct worker_pool *pool;
3558 for_each_worker_pool(pool, gcwq)
3559 mutex_lock_nested(&pool->manager_mutex, pool - gcwq->pools);
3560 spin_lock_irq(&gcwq->lock);
3563 /* release manager positions */
3564 static void gcwq_release_management_and_unlock(struct global_cwq *gcwq)
3566 struct worker_pool *pool;
3568 spin_unlock_irq(&gcwq->lock);
3569 for_each_worker_pool(pool, gcwq)
3570 mutex_unlock(&pool->manager_mutex);
3573 static void gcwq_unbind_fn(struct work_struct *work)
3575 struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3576 struct worker_pool *pool;
3577 struct worker *worker;
3578 struct hlist_node *pos;
3581 BUG_ON(gcwq->cpu != smp_processor_id());
3583 gcwq_claim_management_and_lock(gcwq);
3586 * We've claimed all manager positions. Make all workers unbound
3587 * and set DISASSOCIATED. Before this, all workers except for the
3588 * ones which are still executing works from before the last CPU
3589 * down must be on the cpu. After this, they may become diasporas.
3591 for_each_worker_pool(pool, gcwq)
3592 list_for_each_entry(worker, &pool->idle_list, entry)
3593 worker->flags |= WORKER_UNBOUND;
3595 for_each_busy_worker(worker, i, pos, gcwq)
3596 worker->flags |= WORKER_UNBOUND;
3598 gcwq->flags |= GCWQ_DISASSOCIATED;
3600 gcwq_release_management_and_unlock(gcwq);
3603 * Call schedule() so that we cross rq->lock and thus can guarantee
3604 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3605 * as scheduler callbacks may be invoked from other cpus.
3610 * Sched callbacks are disabled now. Zap nr_running. After this,
3611 * nr_running stays zero and need_more_worker() and keep_working()
3612 * are always true as long as the worklist is not empty. @gcwq now
3613 * behaves as unbound (in terms of concurrency management) gcwq
3614 * which is served by workers tied to the CPU.
3616 * On return from this function, the current worker would trigger
3617 * unbound chain execution of pending work items if other workers
3620 for_each_worker_pool(pool, gcwq)
3621 atomic_set(get_pool_nr_running(pool), 0);
3625 * Workqueues should be brought up before normal priority CPU notifiers.
3626 * This will be registered high priority CPU notifier.
3628 static int __devinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3629 unsigned long action,
3632 unsigned int cpu = (unsigned long)hcpu;
3633 struct global_cwq *gcwq = get_gcwq(cpu);
3634 struct worker_pool *pool;
3636 switch (action & ~CPU_TASKS_FROZEN) {
3637 case CPU_UP_PREPARE:
3638 for_each_worker_pool(pool, gcwq) {
3639 struct worker *worker;
3641 if (pool->nr_workers)
3644 worker = create_worker(pool);
3648 spin_lock_irq(&gcwq->lock);
3649 start_worker(worker);
3650 spin_unlock_irq(&gcwq->lock);
3654 case CPU_DOWN_FAILED:
3656 gcwq_claim_management_and_lock(gcwq);
3657 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3658 rebind_workers(gcwq);
3659 gcwq_release_management_and_unlock(gcwq);
3666 * Workqueues should be brought down after normal priority CPU notifiers.
3667 * This will be registered as low priority CPU notifier.
3669 static int __devinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3670 unsigned long action,
3673 unsigned int cpu = (unsigned long)hcpu;
3674 struct work_struct unbind_work;
3676 switch (action & ~CPU_TASKS_FROZEN) {
3677 case CPU_DOWN_PREPARE:
3678 /* unbinding should happen on the local CPU */
3679 INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3680 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3681 flush_work(&unbind_work);
3689 struct work_for_cpu {
3690 struct completion completion;
3696 static int do_work_for_cpu(void *_wfc)
3698 struct work_for_cpu *wfc = _wfc;
3699 wfc->ret = wfc->fn(wfc->arg);
3700 complete(&wfc->completion);
3705 * work_on_cpu - run a function in user context on a particular cpu
3706 * @cpu: the cpu to run on
3707 * @fn: the function to run
3708 * @arg: the function arg
3710 * This will return the value @fn returns.
3711 * It is up to the caller to ensure that the cpu doesn't go offline.
3712 * The caller must not hold any locks which would prevent @fn from completing.
3714 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3716 struct task_struct *sub_thread;
3717 struct work_for_cpu wfc = {
3718 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3723 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3724 if (IS_ERR(sub_thread))
3725 return PTR_ERR(sub_thread);
3726 kthread_bind(sub_thread, cpu);
3727 wake_up_process(sub_thread);
3728 wait_for_completion(&wfc.completion);
3731 EXPORT_SYMBOL_GPL(work_on_cpu);
3732 #endif /* CONFIG_SMP */
3734 #ifdef CONFIG_FREEZER
3737 * freeze_workqueues_begin - begin freezing workqueues
3739 * Start freezing workqueues. After this function returns, all freezable
3740 * workqueues will queue new works to their frozen_works list instead of
3744 * Grabs and releases workqueue_lock and gcwq->lock's.
3746 void freeze_workqueues_begin(void)
3750 spin_lock(&workqueue_lock);
3752 BUG_ON(workqueue_freezing);
3753 workqueue_freezing = true;
3755 for_each_gcwq_cpu(cpu) {
3756 struct global_cwq *gcwq = get_gcwq(cpu);
3757 struct workqueue_struct *wq;
3759 spin_lock_irq(&gcwq->lock);
3761 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3762 gcwq->flags |= GCWQ_FREEZING;
3764 list_for_each_entry(wq, &workqueues, list) {
3765 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3767 if (cwq && wq->flags & WQ_FREEZABLE)
3768 cwq->max_active = 0;
3771 spin_unlock_irq(&gcwq->lock);
3774 spin_unlock(&workqueue_lock);
3778 * freeze_workqueues_busy - are freezable workqueues still busy?
3780 * Check whether freezing is complete. This function must be called
3781 * between freeze_workqueues_begin() and thaw_workqueues().
3784 * Grabs and releases workqueue_lock.
3787 * %true if some freezable workqueues are still busy. %false if freezing
3790 bool freeze_workqueues_busy(void)
3795 spin_lock(&workqueue_lock);
3797 BUG_ON(!workqueue_freezing);
3799 for_each_gcwq_cpu(cpu) {
3800 struct workqueue_struct *wq;
3802 * nr_active is monotonically decreasing. It's safe
3803 * to peek without lock.
3805 list_for_each_entry(wq, &workqueues, list) {
3806 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3808 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3811 BUG_ON(cwq->nr_active < 0);
3812 if (cwq->nr_active) {
3819 spin_unlock(&workqueue_lock);
3824 * thaw_workqueues - thaw workqueues
3826 * Thaw workqueues. Normal queueing is restored and all collected
3827 * frozen works are transferred to their respective gcwq worklists.
3830 * Grabs and releases workqueue_lock and gcwq->lock's.
3832 void thaw_workqueues(void)
3836 spin_lock(&workqueue_lock);
3838 if (!workqueue_freezing)
3841 for_each_gcwq_cpu(cpu) {
3842 struct global_cwq *gcwq = get_gcwq(cpu);
3843 struct worker_pool *pool;
3844 struct workqueue_struct *wq;
3846 spin_lock_irq(&gcwq->lock);
3848 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3849 gcwq->flags &= ~GCWQ_FREEZING;
3851 list_for_each_entry(wq, &workqueues, list) {
3852 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3854 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3857 /* restore max_active and repopulate worklist */
3858 cwq->max_active = wq->saved_max_active;
3860 while (!list_empty(&cwq->delayed_works) &&
3861 cwq->nr_active < cwq->max_active)
3862 cwq_activate_first_delayed(cwq);
3865 for_each_worker_pool(pool, gcwq)
3866 wake_up_worker(pool);
3868 spin_unlock_irq(&gcwq->lock);
3871 workqueue_freezing = false;
3873 spin_unlock(&workqueue_lock);
3875 #endif /* CONFIG_FREEZER */
3877 static int __init init_workqueues(void)
3882 /* make sure we have enough bits for OFFQ CPU number */
3883 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
3886 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3887 cpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3889 /* initialize gcwqs */
3890 for_each_gcwq_cpu(cpu) {
3891 struct global_cwq *gcwq = get_gcwq(cpu);
3892 struct worker_pool *pool;
3894 spin_lock_init(&gcwq->lock);
3896 gcwq->flags |= GCWQ_DISASSOCIATED;
3898 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3899 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3901 for_each_worker_pool(pool, gcwq) {
3903 INIT_LIST_HEAD(&pool->worklist);
3904 INIT_LIST_HEAD(&pool->idle_list);
3906 init_timer_deferrable(&pool->idle_timer);
3907 pool->idle_timer.function = idle_worker_timeout;
3908 pool->idle_timer.data = (unsigned long)pool;
3910 setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3911 (unsigned long)pool);
3913 mutex_init(&pool->manager_mutex);
3914 ida_init(&pool->worker_ida);
3917 init_waitqueue_head(&gcwq->rebind_hold);
3920 /* create the initial worker */
3921 for_each_online_gcwq_cpu(cpu) {
3922 struct global_cwq *gcwq = get_gcwq(cpu);
3923 struct worker_pool *pool;
3925 if (cpu != WORK_CPU_UNBOUND)
3926 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3928 for_each_worker_pool(pool, gcwq) {
3929 struct worker *worker;
3931 worker = create_worker(pool);
3933 spin_lock_irq(&gcwq->lock);
3934 start_worker(worker);
3935 spin_unlock_irq(&gcwq->lock);
3939 system_wq = alloc_workqueue("events", 0, 0);
3940 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3941 system_long_wq = alloc_workqueue("events_long", 0, 0);
3942 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3943 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3944 WQ_UNBOUND_MAX_ACTIVE);
3945 system_freezable_wq = alloc_workqueue("events_freezable",
3947 system_nrt_freezable_wq = alloc_workqueue("events_nrt_freezable",
3948 WQ_NON_REENTRANT | WQ_FREEZABLE, 0);
3949 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3950 !system_nrt_wq || !system_unbound_wq || !system_freezable_wq ||
3951 !system_nrt_freezable_wq);
3954 early_initcall(init_workqueues);