2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/hashtable.h>
46 #include "workqueue_sched.h"
52 * A bound gcwq is either associated or disassociated with its CPU.
53 * While associated (!DISASSOCIATED), all workers are bound to the
54 * CPU and none has %WORKER_UNBOUND set and concurrency management
57 * While DISASSOCIATED, the cpu may be offline and all workers have
58 * %WORKER_UNBOUND set and concurrency management disabled, and may
59 * be executing on any CPU. The gcwq behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex of all pools on the gcwq to avoid changing binding
63 * state while create_worker() is in progress.
65 GCWQ_DISASSOCIATED = 1 << 0, /* cpu can't serve workers */
66 GCWQ_FREEZING = 1 << 1, /* freeze in progress */
69 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
70 POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */
73 WORKER_STARTED = 1 << 0, /* started */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
80 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND |
83 NR_WORKER_POOLS = 2, /* # worker pools per gcwq */
85 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
87 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
88 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
90 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
91 /* call for help after 10ms
93 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
94 CREATE_COOLDOWN = HZ, /* time to breath after fail */
97 * Rescue workers are used only on emergencies and shared by
100 RESCUER_NICE_LEVEL = -20,
101 HIGHPRI_NICE_LEVEL = -20,
105 * Structure fields follow one of the following exclusion rules.
107 * I: Modifiable by initialization/destruction paths and read-only for
110 * P: Preemption protected. Disabling preemption is enough and should
111 * only be modified and accessed from the local cpu.
113 * L: gcwq->lock protected. Access with gcwq->lock held.
115 * X: During normal operation, modification requires gcwq->lock and
116 * should be done only from local cpu. Either disabling preemption
117 * on local cpu or grabbing gcwq->lock is enough for read access.
118 * If GCWQ_DISASSOCIATED is set, it's identical to L.
120 * F: wq->flush_mutex protected.
122 * W: workqueue_lock protected.
129 * The poor guys doing the actual heavy lifting. All on-duty workers
130 * are either serving the manager role, on idle list or on busy hash.
133 /* on idle list while idle, on busy hash table while busy */
135 struct list_head entry; /* L: while idle */
136 struct hlist_node hentry; /* L: while busy */
139 struct work_struct *current_work; /* L: work being processed */
140 work_func_t current_func; /* L: current_work's fn */
141 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
142 struct list_head scheduled; /* L: scheduled works */
143 struct task_struct *task; /* I: worker task */
144 struct worker_pool *pool; /* I: the associated pool */
145 /* 64 bytes boundary on 64bit, 32 on 32bit */
146 unsigned long last_active; /* L: last active timestamp */
147 unsigned int flags; /* X: flags */
148 int id; /* I: worker id */
150 /* for rebinding worker to CPU */
151 struct work_struct rebind_work; /* L: for busy worker */
155 struct global_cwq *gcwq; /* I: the owning gcwq */
156 unsigned int flags; /* X: flags */
158 struct list_head worklist; /* L: list of pending works */
159 int nr_workers; /* L: total number of workers */
161 /* nr_idle includes the ones off idle_list for rebinding */
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 assoc_mutex; /* protect GCWQ_DISASSOCIATED */
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 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
184 /* L: hash of busy workers */
186 struct worker_pool pools[NR_WORKER_POOLS];
187 /* normal and highpri pools */
188 } ____cacheline_aligned_in_smp;
191 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
192 * work_struct->data are used for flags and thus cwqs need to be
193 * aligned at two's power of the number of flag bits.
195 struct cpu_workqueue_struct {
196 struct worker_pool *pool; /* I: the associated pool */
197 struct workqueue_struct *wq; /* I: the owning workqueue */
198 int work_color; /* L: current color */
199 int flush_color; /* L: flushing color */
200 int nr_in_flight[WORK_NR_COLORS];
201 /* L: nr of in_flight works */
202 int nr_active; /* L: nr of active works */
203 int max_active; /* L: max active works */
204 struct list_head delayed_works; /* L: delayed works */
208 * Structure used to wait for workqueue flush.
211 struct list_head list; /* F: list of flushers */
212 int flush_color; /* F: flush color waiting for */
213 struct completion done; /* flush completion */
217 * All cpumasks are assumed to be always set on UP and thus can't be
218 * used to determine whether there's something to be done.
221 typedef cpumask_var_t mayday_mask_t;
222 #define mayday_test_and_set_cpu(cpu, mask) \
223 cpumask_test_and_set_cpu((cpu), (mask))
224 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
225 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
226 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
227 #define free_mayday_mask(mask) free_cpumask_var((mask))
229 typedef unsigned long mayday_mask_t;
230 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
231 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
232 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
233 #define alloc_mayday_mask(maskp, gfp) true
234 #define free_mayday_mask(mask) do { } while (0)
238 * The externally visible workqueue abstraction is an array of
239 * per-CPU workqueues:
241 struct workqueue_struct {
242 unsigned int flags; /* W: WQ_* flags */
244 struct cpu_workqueue_struct __percpu *pcpu;
245 struct cpu_workqueue_struct *single;
247 } cpu_wq; /* I: cwq's */
248 struct list_head list; /* W: list of all workqueues */
250 struct mutex flush_mutex; /* protects wq flushing */
251 int work_color; /* F: current work color */
252 int flush_color; /* F: current flush color */
253 atomic_t nr_cwqs_to_flush; /* flush in progress */
254 struct wq_flusher *first_flusher; /* F: first flusher */
255 struct list_head flusher_queue; /* F: flush waiters */
256 struct list_head flusher_overflow; /* F: flush overflow list */
258 mayday_mask_t mayday_mask; /* cpus requesting rescue */
259 struct worker *rescuer; /* I: rescue worker */
261 int nr_drainers; /* W: drain in progress */
262 int saved_max_active; /* W: saved cwq max_active */
263 #ifdef CONFIG_LOCKDEP
264 struct lockdep_map lockdep_map;
266 char name[]; /* I: workqueue name */
269 struct workqueue_struct *system_wq __read_mostly;
270 EXPORT_SYMBOL_GPL(system_wq);
271 struct workqueue_struct *system_highpri_wq __read_mostly;
272 EXPORT_SYMBOL_GPL(system_highpri_wq);
273 struct workqueue_struct *system_long_wq __read_mostly;
274 EXPORT_SYMBOL_GPL(system_long_wq);
275 struct workqueue_struct *system_unbound_wq __read_mostly;
276 EXPORT_SYMBOL_GPL(system_unbound_wq);
277 struct workqueue_struct *system_freezable_wq __read_mostly;
278 EXPORT_SYMBOL_GPL(system_freezable_wq);
280 #define CREATE_TRACE_POINTS
281 #include <trace/events/workqueue.h>
283 #define for_each_worker_pool(pool, gcwq) \
284 for ((pool) = &(gcwq)->pools[0]; \
285 (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
287 #define for_each_busy_worker(worker, i, pos, gcwq) \
288 hash_for_each(gcwq->busy_hash, i, pos, worker, hentry)
290 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
293 if (cpu < nr_cpu_ids) {
295 cpu = cpumask_next(cpu, mask);
296 if (cpu < nr_cpu_ids)
300 return WORK_CPU_UNBOUND;
302 return WORK_CPU_NONE;
305 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
306 struct workqueue_struct *wq)
308 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
314 * An extra gcwq is defined for an invalid cpu number
315 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
316 * specific CPU. The following iterators are similar to
317 * for_each_*_cpu() iterators but also considers the unbound gcwq.
319 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
320 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
321 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
322 * WORK_CPU_UNBOUND for unbound workqueues
324 #define for_each_gcwq_cpu(cpu) \
325 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
326 (cpu) < WORK_CPU_NONE; \
327 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
329 #define for_each_online_gcwq_cpu(cpu) \
330 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
331 (cpu) < WORK_CPU_NONE; \
332 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
334 #define for_each_cwq_cpu(cpu, wq) \
335 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
336 (cpu) < WORK_CPU_NONE; \
337 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
339 #ifdef CONFIG_DEBUG_OBJECTS_WORK
341 static struct debug_obj_descr work_debug_descr;
343 static void *work_debug_hint(void *addr)
345 return ((struct work_struct *) addr)->func;
349 * fixup_init is called when:
350 * - an active object is initialized
352 static int work_fixup_init(void *addr, enum debug_obj_state state)
354 struct work_struct *work = addr;
357 case ODEBUG_STATE_ACTIVE:
358 cancel_work_sync(work);
359 debug_object_init(work, &work_debug_descr);
367 * fixup_activate is called when:
368 * - an active object is activated
369 * - an unknown object is activated (might be a statically initialized object)
371 static int work_fixup_activate(void *addr, enum debug_obj_state state)
373 struct work_struct *work = addr;
377 case ODEBUG_STATE_NOTAVAILABLE:
379 * This is not really a fixup. The work struct was
380 * statically initialized. We just make sure that it
381 * is tracked in the object tracker.
383 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
384 debug_object_init(work, &work_debug_descr);
385 debug_object_activate(work, &work_debug_descr);
391 case ODEBUG_STATE_ACTIVE:
400 * fixup_free is called when:
401 * - an active object is freed
403 static int work_fixup_free(void *addr, enum debug_obj_state state)
405 struct work_struct *work = addr;
408 case ODEBUG_STATE_ACTIVE:
409 cancel_work_sync(work);
410 debug_object_free(work, &work_debug_descr);
417 static struct debug_obj_descr work_debug_descr = {
418 .name = "work_struct",
419 .debug_hint = work_debug_hint,
420 .fixup_init = work_fixup_init,
421 .fixup_activate = work_fixup_activate,
422 .fixup_free = work_fixup_free,
425 static inline void debug_work_activate(struct work_struct *work)
427 debug_object_activate(work, &work_debug_descr);
430 static inline void debug_work_deactivate(struct work_struct *work)
432 debug_object_deactivate(work, &work_debug_descr);
435 void __init_work(struct work_struct *work, int onstack)
438 debug_object_init_on_stack(work, &work_debug_descr);
440 debug_object_init(work, &work_debug_descr);
442 EXPORT_SYMBOL_GPL(__init_work);
444 void destroy_work_on_stack(struct work_struct *work)
446 debug_object_free(work, &work_debug_descr);
448 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
451 static inline void debug_work_activate(struct work_struct *work) { }
452 static inline void debug_work_deactivate(struct work_struct *work) { }
455 /* Serializes the accesses to the list of workqueues. */
456 static DEFINE_SPINLOCK(workqueue_lock);
457 static LIST_HEAD(workqueues);
458 static bool workqueue_freezing; /* W: have wqs started freezing? */
461 * The almighty global cpu workqueues. nr_running is the only field
462 * which is expected to be used frequently by other cpus via
463 * try_to_wake_up(). Put it in a separate cacheline.
465 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
466 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]);
469 * Global cpu workqueue and nr_running counter for unbound gcwq. The
470 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
471 * workers have WORKER_UNBOUND set.
473 static struct global_cwq unbound_global_cwq;
474 static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = {
475 [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
478 static int worker_thread(void *__worker);
480 static int worker_pool_pri(struct worker_pool *pool)
482 return pool - pool->gcwq->pools;
485 static struct global_cwq *get_gcwq(unsigned int cpu)
487 if (cpu != WORK_CPU_UNBOUND)
488 return &per_cpu(global_cwq, cpu);
490 return &unbound_global_cwq;
493 static atomic_t *get_pool_nr_running(struct worker_pool *pool)
495 int cpu = pool->gcwq->cpu;
496 int idx = worker_pool_pri(pool);
498 if (cpu != WORK_CPU_UNBOUND)
499 return &per_cpu(pool_nr_running, cpu)[idx];
501 return &unbound_pool_nr_running[idx];
504 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
505 struct workqueue_struct *wq)
507 if (!(wq->flags & WQ_UNBOUND)) {
508 if (likely(cpu < nr_cpu_ids))
509 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
510 } else if (likely(cpu == WORK_CPU_UNBOUND))
511 return wq->cpu_wq.single;
515 static unsigned int work_color_to_flags(int color)
517 return color << WORK_STRUCT_COLOR_SHIFT;
520 static int get_work_color(struct work_struct *work)
522 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
523 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
526 static int work_next_color(int color)
528 return (color + 1) % WORK_NR_COLORS;
532 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
533 * contain the pointer to the queued cwq. Once execution starts, the flag
534 * is cleared and the high bits contain OFFQ flags and CPU number.
536 * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
537 * and clear_work_data() can be used to set the cwq, cpu or clear
538 * work->data. These functions should only be called while the work is
539 * owned - ie. while the PENDING bit is set.
541 * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
542 * a work. gcwq is available once the work has been queued anywhere after
543 * initialization until it is sync canceled. cwq is available only while
544 * the work item is queued.
546 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
547 * canceled. While being canceled, a work item may have its PENDING set
548 * but stay off timer and worklist for arbitrarily long and nobody should
549 * try to steal the PENDING bit.
551 static inline void set_work_data(struct work_struct *work, unsigned long data,
554 BUG_ON(!work_pending(work));
555 atomic_long_set(&work->data, data | flags | work_static(work));
558 static void set_work_cwq(struct work_struct *work,
559 struct cpu_workqueue_struct *cwq,
560 unsigned long extra_flags)
562 set_work_data(work, (unsigned long)cwq,
563 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
566 static void set_work_cpu_and_clear_pending(struct work_struct *work,
570 * The following wmb is paired with the implied mb in
571 * test_and_set_bit(PENDING) and ensures all updates to @work made
572 * here are visible to and precede any updates by the next PENDING
576 set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0);
579 static void clear_work_data(struct work_struct *work)
581 smp_wmb(); /* see set_work_cpu_and_clear_pending() */
582 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
585 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
587 unsigned long data = atomic_long_read(&work->data);
589 if (data & WORK_STRUCT_CWQ)
590 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
595 static struct global_cwq *get_work_gcwq(struct work_struct *work)
597 unsigned long data = atomic_long_read(&work->data);
600 if (data & WORK_STRUCT_CWQ)
601 return ((struct cpu_workqueue_struct *)
602 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq;
604 cpu = data >> WORK_OFFQ_CPU_SHIFT;
605 if (cpu == WORK_CPU_NONE)
608 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
609 return get_gcwq(cpu);
612 static void mark_work_canceling(struct work_struct *work)
614 struct global_cwq *gcwq = get_work_gcwq(work);
615 unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE;
617 set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) | WORK_OFFQ_CANCELING,
618 WORK_STRUCT_PENDING);
621 static bool work_is_canceling(struct work_struct *work)
623 unsigned long data = atomic_long_read(&work->data);
625 return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
629 * Policy functions. These define the policies on how the global worker
630 * pools are managed. Unless noted otherwise, these functions assume that
631 * they're being called with gcwq->lock held.
634 static bool __need_more_worker(struct worker_pool *pool)
636 return !atomic_read(get_pool_nr_running(pool));
640 * Need to wake up a worker? Called from anything but currently
643 * Note that, because unbound workers never contribute to nr_running, this
644 * function will always return %true for unbound gcwq as long as the
645 * worklist isn't empty.
647 static bool need_more_worker(struct worker_pool *pool)
649 return !list_empty(&pool->worklist) && __need_more_worker(pool);
652 /* Can I start working? Called from busy but !running workers. */
653 static bool may_start_working(struct worker_pool *pool)
655 return pool->nr_idle;
658 /* Do I need to keep working? Called from currently running workers. */
659 static bool keep_working(struct worker_pool *pool)
661 atomic_t *nr_running = get_pool_nr_running(pool);
663 return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
666 /* Do we need a new worker? Called from manager. */
667 static bool need_to_create_worker(struct worker_pool *pool)
669 return need_more_worker(pool) && !may_start_working(pool);
672 /* Do I need to be the manager? */
673 static bool need_to_manage_workers(struct worker_pool *pool)
675 return need_to_create_worker(pool) ||
676 (pool->flags & POOL_MANAGE_WORKERS);
679 /* Do we have too many workers and should some go away? */
680 static bool too_many_workers(struct worker_pool *pool)
682 bool managing = pool->flags & POOL_MANAGING_WORKERS;
683 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
684 int nr_busy = pool->nr_workers - nr_idle;
687 * nr_idle and idle_list may disagree if idle rebinding is in
688 * progress. Never return %true if idle_list is empty.
690 if (list_empty(&pool->idle_list))
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 WARN_ON_ONCE(worker->pool->gcwq->cpu != cpu);
743 atomic_inc(get_pool_nr_running(worker->pool));
748 * wq_worker_sleeping - a worker is going to sleep
749 * @task: task going to sleep
750 * @cpu: CPU in question, must be the current CPU number
752 * This function is called during schedule() when a busy worker is
753 * going to sleep. Worker on the same cpu can be woken up by
754 * returning pointer to its task.
757 * spin_lock_irq(rq->lock)
760 * Worker task on @cpu to wake up, %NULL if none.
762 struct task_struct *wq_worker_sleeping(struct task_struct *task,
765 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
766 struct worker_pool *pool = worker->pool;
767 atomic_t *nr_running = get_pool_nr_running(pool);
769 if (worker->flags & WORKER_NOT_RUNNING)
772 /* this can only happen on the local cpu */
773 BUG_ON(cpu != raw_smp_processor_id());
776 * The counterpart of the following dec_and_test, implied mb,
777 * worklist not empty test sequence is in insert_work().
778 * Please read comment there.
780 * NOT_RUNNING is clear. This means that we're bound to and
781 * running on the local cpu w/ rq lock held and preemption
782 * disabled, which in turn means that none else could be
783 * manipulating idle_list, so dereferencing idle_list without gcwq
786 if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
787 to_wakeup = first_worker(pool);
788 return to_wakeup ? to_wakeup->task : NULL;
792 * worker_set_flags - set worker flags and adjust nr_running accordingly
794 * @flags: flags to set
795 * @wakeup: wakeup an idle worker if necessary
797 * Set @flags in @worker->flags and adjust nr_running accordingly. If
798 * nr_running becomes zero and @wakeup is %true, an idle worker is
802 * spin_lock_irq(gcwq->lock)
804 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
807 struct worker_pool *pool = worker->pool;
809 WARN_ON_ONCE(worker->task != current);
812 * If transitioning into NOT_RUNNING, adjust nr_running and
813 * wake up an idle worker as necessary if requested by
816 if ((flags & WORKER_NOT_RUNNING) &&
817 !(worker->flags & WORKER_NOT_RUNNING)) {
818 atomic_t *nr_running = get_pool_nr_running(pool);
821 if (atomic_dec_and_test(nr_running) &&
822 !list_empty(&pool->worklist))
823 wake_up_worker(pool);
825 atomic_dec(nr_running);
828 worker->flags |= flags;
832 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
834 * @flags: flags to clear
836 * Clear @flags in @worker->flags and adjust nr_running accordingly.
839 * spin_lock_irq(gcwq->lock)
841 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
843 struct worker_pool *pool = worker->pool;
844 unsigned int oflags = worker->flags;
846 WARN_ON_ONCE(worker->task != current);
848 worker->flags &= ~flags;
851 * If transitioning out of NOT_RUNNING, increment nr_running. Note
852 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
853 * of multiple flags, not a single flag.
855 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
856 if (!(worker->flags & WORKER_NOT_RUNNING))
857 atomic_inc(get_pool_nr_running(pool));
861 * find_worker_executing_work - find worker which is executing a work
862 * @gcwq: gcwq of interest
863 * @work: work to find worker for
865 * Find a worker which is executing @work on @gcwq by searching
866 * @gcwq->busy_hash which is keyed by the address of @work. For a worker
867 * to match, its current execution should match the address of @work and
868 * its work function. This is to avoid unwanted dependency between
869 * unrelated work executions through a work item being recycled while still
872 * This is a bit tricky. A work item may be freed once its execution
873 * starts and nothing prevents the freed area from being recycled for
874 * another work item. If the same work item address ends up being reused
875 * before the original execution finishes, workqueue will identify the
876 * recycled work item as currently executing and make it wait until the
877 * current execution finishes, introducing an unwanted dependency.
879 * This function checks the work item address, work function and workqueue
880 * to avoid false positives. Note that this isn't complete as one may
881 * construct a work function which can introduce dependency onto itself
882 * through a recycled work item. Well, if somebody wants to shoot oneself
883 * in the foot that badly, there's only so much we can do, and if such
884 * deadlock actually occurs, it should be easy to locate the culprit work
888 * spin_lock_irq(gcwq->lock).
891 * Pointer to worker which is executing @work if found, NULL
894 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
895 struct work_struct *work)
897 struct worker *worker;
898 struct hlist_node *tmp;
900 hash_for_each_possible(gcwq->busy_hash, worker, tmp, hentry,
902 if (worker->current_work == work &&
903 worker->current_func == work->func)
910 * move_linked_works - move linked works to a list
911 * @work: start of series of works to be scheduled
912 * @head: target list to append @work to
913 * @nextp: out paramter for nested worklist walking
915 * Schedule linked works starting from @work to @head. Work series to
916 * be scheduled starts at @work and includes any consecutive work with
917 * WORK_STRUCT_LINKED set in its predecessor.
919 * If @nextp is not NULL, it's updated to point to the next work of
920 * the last scheduled work. This allows move_linked_works() to be
921 * nested inside outer list_for_each_entry_safe().
924 * spin_lock_irq(gcwq->lock).
926 static void move_linked_works(struct work_struct *work, struct list_head *head,
927 struct work_struct **nextp)
929 struct work_struct *n;
932 * Linked worklist will always end before the end of the list,
933 * use NULL for list head.
935 list_for_each_entry_safe_from(work, n, NULL, entry) {
936 list_move_tail(&work->entry, head);
937 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
942 * If we're already inside safe list traversal and have moved
943 * multiple works to the scheduled queue, the next position
944 * needs to be updated.
950 static void cwq_activate_delayed_work(struct work_struct *work)
952 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
954 trace_workqueue_activate_work(work);
955 move_linked_works(work, &cwq->pool->worklist, NULL);
956 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
960 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
962 struct work_struct *work = list_first_entry(&cwq->delayed_works,
963 struct work_struct, entry);
965 cwq_activate_delayed_work(work);
969 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
970 * @cwq: cwq of interest
971 * @color: color of work which left the queue
973 * A work either has completed or is removed from pending queue,
974 * decrement nr_in_flight of its cwq and handle workqueue flushing.
977 * spin_lock_irq(gcwq->lock).
979 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
981 /* ignore uncolored works */
982 if (color == WORK_NO_COLOR)
985 cwq->nr_in_flight[color]--;
988 if (!list_empty(&cwq->delayed_works)) {
989 /* one down, submit a delayed one */
990 if (cwq->nr_active < cwq->max_active)
991 cwq_activate_first_delayed(cwq);
994 /* is flush in progress and are we at the flushing tip? */
995 if (likely(cwq->flush_color != color))
998 /* are there still in-flight works? */
999 if (cwq->nr_in_flight[color])
1002 /* this cwq is done, clear flush_color */
1003 cwq->flush_color = -1;
1006 * If this was the last cwq, wake up the first flusher. It
1007 * will handle the rest.
1009 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1010 complete(&cwq->wq->first_flusher->done);
1014 * try_to_grab_pending - steal work item from worklist and disable irq
1015 * @work: work item to steal
1016 * @is_dwork: @work is a delayed_work
1017 * @flags: place to store irq state
1019 * Try to grab PENDING bit of @work. This function can handle @work in any
1020 * stable state - idle, on timer or on worklist. Return values are
1022 * 1 if @work was pending and we successfully stole PENDING
1023 * 0 if @work was idle and we claimed PENDING
1024 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1025 * -ENOENT if someone else is canceling @work, this state may persist
1026 * for arbitrarily long
1028 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1029 * interrupted while holding PENDING and @work off queue, irq must be
1030 * disabled on entry. This, combined with delayed_work->timer being
1031 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1033 * On successful return, >= 0, irq is disabled and the caller is
1034 * responsible for releasing it using local_irq_restore(*@flags).
1036 * This function is safe to call from any context including IRQ handler.
1038 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1039 unsigned long *flags)
1041 struct global_cwq *gcwq;
1043 local_irq_save(*flags);
1045 /* try to steal the timer if it exists */
1047 struct delayed_work *dwork = to_delayed_work(work);
1050 * dwork->timer is irqsafe. If del_timer() fails, it's
1051 * guaranteed that the timer is not queued anywhere and not
1052 * running on the local CPU.
1054 if (likely(del_timer(&dwork->timer)))
1058 /* try to claim PENDING the normal way */
1059 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1063 * The queueing is in progress, or it is already queued. Try to
1064 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1066 gcwq = get_work_gcwq(work);
1070 spin_lock(&gcwq->lock);
1071 if (!list_empty(&work->entry)) {
1073 * This work is queued, but perhaps we locked the wrong gcwq.
1074 * In that case we must see the new value after rmb(), see
1075 * insert_work()->wmb().
1078 if (gcwq == get_work_gcwq(work)) {
1079 debug_work_deactivate(work);
1082 * A delayed work item cannot be grabbed directly
1083 * because it might have linked NO_COLOR work items
1084 * which, if left on the delayed_list, will confuse
1085 * cwq->nr_active management later on and cause
1086 * stall. Make sure the work item is activated
1089 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1090 cwq_activate_delayed_work(work);
1092 list_del_init(&work->entry);
1093 cwq_dec_nr_in_flight(get_work_cwq(work),
1094 get_work_color(work));
1096 spin_unlock(&gcwq->lock);
1100 spin_unlock(&gcwq->lock);
1102 local_irq_restore(*flags);
1103 if (work_is_canceling(work))
1110 * insert_work - insert a work into gcwq
1111 * @cwq: cwq @work belongs to
1112 * @work: work to insert
1113 * @head: insertion point
1114 * @extra_flags: extra WORK_STRUCT_* flags to set
1116 * Insert @work which belongs to @cwq into @gcwq after @head.
1117 * @extra_flags is or'd to work_struct flags.
1120 * spin_lock_irq(gcwq->lock).
1122 static void insert_work(struct cpu_workqueue_struct *cwq,
1123 struct work_struct *work, struct list_head *head,
1124 unsigned int extra_flags)
1126 struct worker_pool *pool = cwq->pool;
1128 /* we own @work, set data and link */
1129 set_work_cwq(work, cwq, extra_flags);
1132 * Ensure that we get the right work->data if we see the
1133 * result of list_add() below, see try_to_grab_pending().
1137 list_add_tail(&work->entry, head);
1140 * Ensure either worker_sched_deactivated() sees the above
1141 * list_add_tail() or we see zero nr_running to avoid workers
1142 * lying around lazily while there are works to be processed.
1146 if (__need_more_worker(pool))
1147 wake_up_worker(pool);
1151 * Test whether @work is being queued from another work executing on the
1152 * same workqueue. This is rather expensive and should only be used from
1155 static bool is_chained_work(struct workqueue_struct *wq)
1157 unsigned long flags;
1160 for_each_gcwq_cpu(cpu) {
1161 struct global_cwq *gcwq = get_gcwq(cpu);
1162 struct worker *worker;
1163 struct hlist_node *pos;
1166 spin_lock_irqsave(&gcwq->lock, flags);
1167 for_each_busy_worker(worker, i, pos, gcwq) {
1168 if (worker->task != current)
1170 spin_unlock_irqrestore(&gcwq->lock, flags);
1172 * I'm @worker, no locking necessary. See if @work
1173 * is headed to the same workqueue.
1175 return worker->current_cwq->wq == wq;
1177 spin_unlock_irqrestore(&gcwq->lock, flags);
1182 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1183 struct work_struct *work)
1185 struct global_cwq *gcwq;
1186 struct cpu_workqueue_struct *cwq;
1187 struct list_head *worklist;
1188 unsigned int work_flags;
1189 unsigned int req_cpu = cpu;
1192 * While a work item is PENDING && off queue, a task trying to
1193 * steal the PENDING will busy-loop waiting for it to either get
1194 * queued or lose PENDING. Grabbing PENDING and queueing should
1195 * happen with IRQ disabled.
1197 WARN_ON_ONCE(!irqs_disabled());
1199 debug_work_activate(work);
1201 /* if dying, only works from the same workqueue are allowed */
1202 if (unlikely(wq->flags & WQ_DRAINING) &&
1203 WARN_ON_ONCE(!is_chained_work(wq)))
1206 /* determine gcwq to use */
1207 if (!(wq->flags & WQ_UNBOUND)) {
1208 struct global_cwq *last_gcwq;
1210 if (cpu == WORK_CPU_UNBOUND)
1211 cpu = raw_smp_processor_id();
1214 * It's multi cpu. If @work was previously on a different
1215 * cpu, it might still be running there, in which case the
1216 * work needs to be queued on that cpu to guarantee
1219 gcwq = get_gcwq(cpu);
1220 last_gcwq = get_work_gcwq(work);
1222 if (last_gcwq && last_gcwq != gcwq) {
1223 struct worker *worker;
1225 spin_lock(&last_gcwq->lock);
1227 worker = find_worker_executing_work(last_gcwq, work);
1229 if (worker && worker->current_cwq->wq == wq)
1232 /* meh... not running there, queue here */
1233 spin_unlock(&last_gcwq->lock);
1234 spin_lock(&gcwq->lock);
1237 spin_lock(&gcwq->lock);
1240 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1241 spin_lock(&gcwq->lock);
1244 /* gcwq determined, get cwq and queue */
1245 cwq = get_cwq(gcwq->cpu, wq);
1246 trace_workqueue_queue_work(req_cpu, cwq, work);
1248 if (WARN_ON(!list_empty(&work->entry))) {
1249 spin_unlock(&gcwq->lock);
1253 cwq->nr_in_flight[cwq->work_color]++;
1254 work_flags = work_color_to_flags(cwq->work_color);
1256 if (likely(cwq->nr_active < cwq->max_active)) {
1257 trace_workqueue_activate_work(work);
1259 worklist = &cwq->pool->worklist;
1261 work_flags |= WORK_STRUCT_DELAYED;
1262 worklist = &cwq->delayed_works;
1265 insert_work(cwq, work, worklist, work_flags);
1267 spin_unlock(&gcwq->lock);
1271 * queue_work_on - queue work on specific cpu
1272 * @cpu: CPU number to execute work on
1273 * @wq: workqueue to use
1274 * @work: work to queue
1276 * Returns %false if @work was already on a queue, %true otherwise.
1278 * We queue the work to a specific CPU, the caller must ensure it
1281 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1282 struct work_struct *work)
1285 unsigned long flags;
1287 local_irq_save(flags);
1289 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1290 __queue_work(cpu, wq, work);
1294 local_irq_restore(flags);
1297 EXPORT_SYMBOL_GPL(queue_work_on);
1300 * queue_work - queue work on a workqueue
1301 * @wq: workqueue to use
1302 * @work: work to queue
1304 * Returns %false if @work was already on a queue, %true otherwise.
1306 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1307 * it can be processed by another CPU.
1309 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1311 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1313 EXPORT_SYMBOL_GPL(queue_work);
1315 void delayed_work_timer_fn(unsigned long __data)
1317 struct delayed_work *dwork = (struct delayed_work *)__data;
1318 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1320 /* should have been called from irqsafe timer with irq already off */
1321 __queue_work(dwork->cpu, cwq->wq, &dwork->work);
1323 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1325 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1326 struct delayed_work *dwork, unsigned long delay)
1328 struct timer_list *timer = &dwork->timer;
1329 struct work_struct *work = &dwork->work;
1332 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1333 timer->data != (unsigned long)dwork);
1334 WARN_ON_ONCE(timer_pending(timer));
1335 WARN_ON_ONCE(!list_empty(&work->entry));
1338 * If @delay is 0, queue @dwork->work immediately. This is for
1339 * both optimization and correctness. The earliest @timer can
1340 * expire is on the closest next tick and delayed_work users depend
1341 * on that there's no such delay when @delay is 0.
1344 __queue_work(cpu, wq, &dwork->work);
1348 timer_stats_timer_set_start_info(&dwork->timer);
1351 * This stores cwq for the moment, for the timer_fn. Note that the
1352 * work's gcwq is preserved to allow reentrance detection for
1355 if (!(wq->flags & WQ_UNBOUND)) {
1356 struct global_cwq *gcwq = get_work_gcwq(work);
1359 * If we cannot get the last gcwq from @work directly,
1360 * select the last CPU such that it avoids unnecessarily
1361 * triggering non-reentrancy check in __queue_work().
1366 if (lcpu == WORK_CPU_UNBOUND)
1367 lcpu = raw_smp_processor_id();
1369 lcpu = WORK_CPU_UNBOUND;
1372 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1375 timer->expires = jiffies + delay;
1377 if (unlikely(cpu != WORK_CPU_UNBOUND))
1378 add_timer_on(timer, cpu);
1384 * queue_delayed_work_on - queue work on specific CPU after delay
1385 * @cpu: CPU number to execute work on
1386 * @wq: workqueue to use
1387 * @dwork: work to queue
1388 * @delay: number of jiffies to wait before queueing
1390 * Returns %false if @work was already on a queue, %true otherwise. If
1391 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1394 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1395 struct delayed_work *dwork, unsigned long delay)
1397 struct work_struct *work = &dwork->work;
1399 unsigned long flags;
1401 /* read the comment in __queue_work() */
1402 local_irq_save(flags);
1404 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1405 __queue_delayed_work(cpu, wq, dwork, delay);
1409 local_irq_restore(flags);
1412 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1415 * queue_delayed_work - queue work on a workqueue after delay
1416 * @wq: workqueue to use
1417 * @dwork: delayable work to queue
1418 * @delay: number of jiffies to wait before queueing
1420 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1422 bool queue_delayed_work(struct workqueue_struct *wq,
1423 struct delayed_work *dwork, unsigned long delay)
1425 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1427 EXPORT_SYMBOL_GPL(queue_delayed_work);
1430 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1431 * @cpu: CPU number to execute work on
1432 * @wq: workqueue to use
1433 * @dwork: work to queue
1434 * @delay: number of jiffies to wait before queueing
1436 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1437 * modify @dwork's timer so that it expires after @delay. If @delay is
1438 * zero, @work is guaranteed to be scheduled immediately regardless of its
1441 * Returns %false if @dwork was idle and queued, %true if @dwork was
1442 * pending and its timer was modified.
1444 * This function is safe to call from any context including IRQ handler.
1445 * See try_to_grab_pending() for details.
1447 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1448 struct delayed_work *dwork, unsigned long delay)
1450 unsigned long flags;
1454 ret = try_to_grab_pending(&dwork->work, true, &flags);
1455 } while (unlikely(ret == -EAGAIN));
1457 if (likely(ret >= 0)) {
1458 __queue_delayed_work(cpu, wq, dwork, delay);
1459 local_irq_restore(flags);
1462 /* -ENOENT from try_to_grab_pending() becomes %true */
1465 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1468 * mod_delayed_work - modify delay of or queue a delayed work
1469 * @wq: workqueue to use
1470 * @dwork: work to queue
1471 * @delay: number of jiffies to wait before queueing
1473 * mod_delayed_work_on() on local CPU.
1475 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1476 unsigned long delay)
1478 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1480 EXPORT_SYMBOL_GPL(mod_delayed_work);
1483 * worker_enter_idle - enter idle state
1484 * @worker: worker which is entering idle state
1486 * @worker is entering idle state. Update stats and idle timer if
1490 * spin_lock_irq(gcwq->lock).
1492 static void worker_enter_idle(struct worker *worker)
1494 struct worker_pool *pool = worker->pool;
1495 struct global_cwq *gcwq = pool->gcwq;
1497 BUG_ON(worker->flags & WORKER_IDLE);
1498 BUG_ON(!list_empty(&worker->entry) &&
1499 (worker->hentry.next || worker->hentry.pprev));
1501 /* can't use worker_set_flags(), also called from start_worker() */
1502 worker->flags |= WORKER_IDLE;
1504 worker->last_active = jiffies;
1506 /* idle_list is LIFO */
1507 list_add(&worker->entry, &pool->idle_list);
1509 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1510 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1513 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1514 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1515 * nr_running, the warning may trigger spuriously. Check iff
1516 * unbind is not in progress.
1518 WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
1519 pool->nr_workers == pool->nr_idle &&
1520 atomic_read(get_pool_nr_running(pool)));
1524 * worker_leave_idle - leave idle state
1525 * @worker: worker which is leaving idle state
1527 * @worker is leaving idle state. Update stats.
1530 * spin_lock_irq(gcwq->lock).
1532 static void worker_leave_idle(struct worker *worker)
1534 struct worker_pool *pool = worker->pool;
1536 BUG_ON(!(worker->flags & WORKER_IDLE));
1537 worker_clr_flags(worker, WORKER_IDLE);
1539 list_del_init(&worker->entry);
1543 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1546 * Works which are scheduled while the cpu is online must at least be
1547 * scheduled to a worker which is bound to the cpu so that if they are
1548 * flushed from cpu callbacks while cpu is going down, they are
1549 * guaranteed to execute on the cpu.
1551 * This function is to be used by rogue workers and rescuers to bind
1552 * themselves to the target cpu and may race with cpu going down or
1553 * coming online. kthread_bind() can't be used because it may put the
1554 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1555 * verbatim as it's best effort and blocking and gcwq may be
1556 * [dis]associated in the meantime.
1558 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1559 * binding against %GCWQ_DISASSOCIATED which is set during
1560 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1561 * enters idle state or fetches works without dropping lock, it can
1562 * guarantee the scheduling requirement described in the first paragraph.
1565 * Might sleep. Called without any lock but returns with gcwq->lock
1569 * %true if the associated gcwq is online (@worker is successfully
1570 * bound), %false if offline.
1572 static bool worker_maybe_bind_and_lock(struct worker *worker)
1573 __acquires(&gcwq->lock)
1575 struct global_cwq *gcwq = worker->pool->gcwq;
1576 struct task_struct *task = worker->task;
1580 * The following call may fail, succeed or succeed
1581 * without actually migrating the task to the cpu if
1582 * it races with cpu hotunplug operation. Verify
1583 * against GCWQ_DISASSOCIATED.
1585 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1586 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1588 spin_lock_irq(&gcwq->lock);
1589 if (gcwq->flags & GCWQ_DISASSOCIATED)
1591 if (task_cpu(task) == gcwq->cpu &&
1592 cpumask_equal(¤t->cpus_allowed,
1593 get_cpu_mask(gcwq->cpu)))
1595 spin_unlock_irq(&gcwq->lock);
1598 * We've raced with CPU hot[un]plug. Give it a breather
1599 * and retry migration. cond_resched() is required here;
1600 * otherwise, we might deadlock against cpu_stop trying to
1601 * bring down the CPU on non-preemptive kernel.
1609 * Rebind an idle @worker to its CPU. worker_thread() will test
1610 * list_empty(@worker->entry) before leaving idle and call this function.
1612 static void idle_worker_rebind(struct worker *worker)
1614 struct global_cwq *gcwq = worker->pool->gcwq;
1616 /* CPU may go down again inbetween, clear UNBOUND only on success */
1617 if (worker_maybe_bind_and_lock(worker))
1618 worker_clr_flags(worker, WORKER_UNBOUND);
1620 /* rebind complete, become available again */
1621 list_add(&worker->entry, &worker->pool->idle_list);
1622 spin_unlock_irq(&gcwq->lock);
1626 * Function for @worker->rebind.work used to rebind unbound busy workers to
1627 * the associated cpu which is coming back online. This is scheduled by
1628 * cpu up but can race with other cpu hotplug operations and may be
1629 * executed twice without intervening cpu down.
1631 static void busy_worker_rebind_fn(struct work_struct *work)
1633 struct worker *worker = container_of(work, struct worker, rebind_work);
1634 struct global_cwq *gcwq = worker->pool->gcwq;
1636 if (worker_maybe_bind_and_lock(worker))
1637 worker_clr_flags(worker, WORKER_UNBOUND);
1639 spin_unlock_irq(&gcwq->lock);
1643 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1644 * @gcwq: gcwq of interest
1646 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1647 * is different for idle and busy ones.
1649 * Idle ones will be removed from the idle_list and woken up. They will
1650 * add themselves back after completing rebind. This ensures that the
1651 * idle_list doesn't contain any unbound workers when re-bound busy workers
1652 * try to perform local wake-ups for concurrency management.
1654 * Busy workers can rebind after they finish their current work items.
1655 * Queueing the rebind work item at the head of the scheduled list is
1656 * enough. Note that nr_running will be properly bumped as busy workers
1659 * On return, all non-manager workers are scheduled for rebind - see
1660 * manage_workers() for the manager special case. Any idle worker
1661 * including the manager will not appear on @idle_list until rebind is
1662 * complete, making local wake-ups safe.
1664 static void rebind_workers(struct global_cwq *gcwq)
1666 struct worker_pool *pool;
1667 struct worker *worker, *n;
1668 struct hlist_node *pos;
1671 lockdep_assert_held(&gcwq->lock);
1673 for_each_worker_pool(pool, gcwq)
1674 lockdep_assert_held(&pool->assoc_mutex);
1676 /* dequeue and kick idle ones */
1677 for_each_worker_pool(pool, gcwq) {
1678 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1680 * idle workers should be off @pool->idle_list
1681 * until rebind is complete to avoid receiving
1682 * premature local wake-ups.
1684 list_del_init(&worker->entry);
1687 * worker_thread() will see the above dequeuing
1688 * and call idle_worker_rebind().
1690 wake_up_process(worker->task);
1694 /* rebind busy workers */
1695 for_each_busy_worker(worker, i, pos, gcwq) {
1696 struct work_struct *rebind_work = &worker->rebind_work;
1697 struct workqueue_struct *wq;
1699 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1700 work_data_bits(rebind_work)))
1703 debug_work_activate(rebind_work);
1706 * wq doesn't really matter but let's keep @worker->pool
1707 * and @cwq->pool consistent for sanity.
1709 if (worker_pool_pri(worker->pool))
1710 wq = system_highpri_wq;
1714 insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
1715 worker->scheduled.next,
1716 work_color_to_flags(WORK_NO_COLOR));
1720 static struct worker *alloc_worker(void)
1722 struct worker *worker;
1724 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1726 INIT_LIST_HEAD(&worker->entry);
1727 INIT_LIST_HEAD(&worker->scheduled);
1728 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1729 /* on creation a worker is in !idle && prep state */
1730 worker->flags = WORKER_PREP;
1736 * create_worker - create a new workqueue worker
1737 * @pool: pool the new worker will belong to
1739 * Create a new worker which is bound to @pool. The returned worker
1740 * can be started by calling start_worker() or destroyed using
1744 * Might sleep. Does GFP_KERNEL allocations.
1747 * Pointer to the newly created worker.
1749 static struct worker *create_worker(struct worker_pool *pool)
1751 struct global_cwq *gcwq = pool->gcwq;
1752 const char *pri = worker_pool_pri(pool) ? "H" : "";
1753 struct worker *worker = NULL;
1756 spin_lock_irq(&gcwq->lock);
1757 while (ida_get_new(&pool->worker_ida, &id)) {
1758 spin_unlock_irq(&gcwq->lock);
1759 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1761 spin_lock_irq(&gcwq->lock);
1763 spin_unlock_irq(&gcwq->lock);
1765 worker = alloc_worker();
1769 worker->pool = pool;
1772 if (gcwq->cpu != WORK_CPU_UNBOUND)
1773 worker->task = kthread_create_on_node(worker_thread,
1774 worker, cpu_to_node(gcwq->cpu),
1775 "kworker/%u:%d%s", gcwq->cpu, id, pri);
1777 worker->task = kthread_create(worker_thread, worker,
1778 "kworker/u:%d%s", id, pri);
1779 if (IS_ERR(worker->task))
1782 if (worker_pool_pri(pool))
1783 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1786 * Determine CPU binding of the new worker depending on
1787 * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1788 * flag remains stable across this function. See the comments
1789 * above the flag definition for details.
1791 * As an unbound worker may later become a regular one if CPU comes
1792 * online, make sure every worker has %PF_THREAD_BOUND set.
1794 if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
1795 kthread_bind(worker->task, gcwq->cpu);
1797 worker->task->flags |= PF_THREAD_BOUND;
1798 worker->flags |= WORKER_UNBOUND;
1804 spin_lock_irq(&gcwq->lock);
1805 ida_remove(&pool->worker_ida, id);
1806 spin_unlock_irq(&gcwq->lock);
1813 * start_worker - start a newly created worker
1814 * @worker: worker to start
1816 * Make the gcwq aware of @worker and start it.
1819 * spin_lock_irq(gcwq->lock).
1821 static void start_worker(struct worker *worker)
1823 worker->flags |= WORKER_STARTED;
1824 worker->pool->nr_workers++;
1825 worker_enter_idle(worker);
1826 wake_up_process(worker->task);
1830 * destroy_worker - destroy a workqueue worker
1831 * @worker: worker to be destroyed
1833 * Destroy @worker and adjust @gcwq stats accordingly.
1836 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1838 static void destroy_worker(struct worker *worker)
1840 struct worker_pool *pool = worker->pool;
1841 struct global_cwq *gcwq = pool->gcwq;
1842 int id = worker->id;
1844 /* sanity check frenzy */
1845 BUG_ON(worker->current_work);
1846 BUG_ON(!list_empty(&worker->scheduled));
1848 if (worker->flags & WORKER_STARTED)
1850 if (worker->flags & WORKER_IDLE)
1853 list_del_init(&worker->entry);
1854 worker->flags |= WORKER_DIE;
1856 spin_unlock_irq(&gcwq->lock);
1858 kthread_stop(worker->task);
1861 spin_lock_irq(&gcwq->lock);
1862 ida_remove(&pool->worker_ida, id);
1865 static void idle_worker_timeout(unsigned long __pool)
1867 struct worker_pool *pool = (void *)__pool;
1868 struct global_cwq *gcwq = pool->gcwq;
1870 spin_lock_irq(&gcwq->lock);
1872 if (too_many_workers(pool)) {
1873 struct worker *worker;
1874 unsigned long expires;
1876 /* idle_list is kept in LIFO order, check the last one */
1877 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1878 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1880 if (time_before(jiffies, expires))
1881 mod_timer(&pool->idle_timer, expires);
1883 /* it's been idle for too long, wake up manager */
1884 pool->flags |= POOL_MANAGE_WORKERS;
1885 wake_up_worker(pool);
1889 spin_unlock_irq(&gcwq->lock);
1892 static bool send_mayday(struct work_struct *work)
1894 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1895 struct workqueue_struct *wq = cwq->wq;
1898 if (!(wq->flags & WQ_RESCUER))
1901 /* mayday mayday mayday */
1902 cpu = cwq->pool->gcwq->cpu;
1903 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1904 if (cpu == WORK_CPU_UNBOUND)
1906 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1907 wake_up_process(wq->rescuer->task);
1911 static void gcwq_mayday_timeout(unsigned long __pool)
1913 struct worker_pool *pool = (void *)__pool;
1914 struct global_cwq *gcwq = pool->gcwq;
1915 struct work_struct *work;
1917 spin_lock_irq(&gcwq->lock);
1919 if (need_to_create_worker(pool)) {
1921 * We've been trying to create a new worker but
1922 * haven't been successful. We might be hitting an
1923 * allocation deadlock. Send distress signals to
1926 list_for_each_entry(work, &pool->worklist, entry)
1930 spin_unlock_irq(&gcwq->lock);
1932 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1936 * maybe_create_worker - create a new worker if necessary
1937 * @pool: pool to create a new worker for
1939 * Create a new worker for @pool if necessary. @pool is guaranteed to
1940 * have at least one idle worker on return from this function. If
1941 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1942 * sent to all rescuers with works scheduled on @pool to resolve
1943 * possible allocation deadlock.
1945 * On return, need_to_create_worker() is guaranteed to be false and
1946 * may_start_working() true.
1949 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1950 * multiple times. Does GFP_KERNEL allocations. Called only from
1954 * false if no action was taken and gcwq->lock stayed locked, true
1957 static bool maybe_create_worker(struct worker_pool *pool)
1958 __releases(&gcwq->lock)
1959 __acquires(&gcwq->lock)
1961 struct global_cwq *gcwq = pool->gcwq;
1963 if (!need_to_create_worker(pool))
1966 spin_unlock_irq(&gcwq->lock);
1968 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1969 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1972 struct worker *worker;
1974 worker = create_worker(pool);
1976 del_timer_sync(&pool->mayday_timer);
1977 spin_lock_irq(&gcwq->lock);
1978 start_worker(worker);
1979 BUG_ON(need_to_create_worker(pool));
1983 if (!need_to_create_worker(pool))
1986 __set_current_state(TASK_INTERRUPTIBLE);
1987 schedule_timeout(CREATE_COOLDOWN);
1989 if (!need_to_create_worker(pool))
1993 del_timer_sync(&pool->mayday_timer);
1994 spin_lock_irq(&gcwq->lock);
1995 if (need_to_create_worker(pool))
2001 * maybe_destroy_worker - destroy workers which have been idle for a while
2002 * @pool: pool to destroy workers for
2004 * Destroy @pool workers which have been idle for longer than
2005 * IDLE_WORKER_TIMEOUT.
2008 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2009 * multiple times. Called only from manager.
2012 * false if no action was taken and gcwq->lock stayed locked, true
2015 static bool maybe_destroy_workers(struct worker_pool *pool)
2019 while (too_many_workers(pool)) {
2020 struct worker *worker;
2021 unsigned long expires;
2023 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2024 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2026 if (time_before(jiffies, expires)) {
2027 mod_timer(&pool->idle_timer, expires);
2031 destroy_worker(worker);
2039 * manage_workers - manage worker pool
2042 * Assume the manager role and manage gcwq worker pool @worker belongs
2043 * to. At any given time, there can be only zero or one manager per
2044 * gcwq. The exclusion is handled automatically by this function.
2046 * The caller can safely start processing works on false return. On
2047 * true return, it's guaranteed that need_to_create_worker() is false
2048 * and may_start_working() is true.
2051 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2052 * multiple times. Does GFP_KERNEL allocations.
2055 * false if no action was taken and gcwq->lock stayed locked, true if
2056 * some action was taken.
2058 static bool manage_workers(struct worker *worker)
2060 struct worker_pool *pool = worker->pool;
2063 if (pool->flags & POOL_MANAGING_WORKERS)
2066 pool->flags |= POOL_MANAGING_WORKERS;
2069 * To simplify both worker management and CPU hotplug, hold off
2070 * management while hotplug is in progress. CPU hotplug path can't
2071 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2072 * lead to idle worker depletion (all become busy thinking someone
2073 * else is managing) which in turn can result in deadlock under
2074 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2075 * manager against CPU hotplug.
2077 * assoc_mutex would always be free unless CPU hotplug is in
2078 * progress. trylock first without dropping @gcwq->lock.
2080 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2081 spin_unlock_irq(&pool->gcwq->lock);
2082 mutex_lock(&pool->assoc_mutex);
2084 * CPU hotplug could have happened while we were waiting
2085 * for assoc_mutex. Hotplug itself can't handle us
2086 * because manager isn't either on idle or busy list, and
2087 * @gcwq's state and ours could have deviated.
2089 * As hotplug is now excluded via assoc_mutex, we can
2090 * simply try to bind. It will succeed or fail depending
2091 * on @gcwq's current state. Try it and adjust
2092 * %WORKER_UNBOUND accordingly.
2094 if (worker_maybe_bind_and_lock(worker))
2095 worker->flags &= ~WORKER_UNBOUND;
2097 worker->flags |= WORKER_UNBOUND;
2102 pool->flags &= ~POOL_MANAGE_WORKERS;
2105 * Destroy and then create so that may_start_working() is true
2108 ret |= maybe_destroy_workers(pool);
2109 ret |= maybe_create_worker(pool);
2111 pool->flags &= ~POOL_MANAGING_WORKERS;
2112 mutex_unlock(&pool->assoc_mutex);
2117 * process_one_work - process single work
2119 * @work: work to process
2121 * Process @work. This function contains all the logics necessary to
2122 * process a single work including synchronization against and
2123 * interaction with other workers on the same cpu, queueing and
2124 * flushing. As long as context requirement is met, any worker can
2125 * call this function to process a work.
2128 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2130 static void process_one_work(struct worker *worker, struct work_struct *work)
2131 __releases(&gcwq->lock)
2132 __acquires(&gcwq->lock)
2134 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2135 struct worker_pool *pool = worker->pool;
2136 struct global_cwq *gcwq = pool->gcwq;
2137 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2139 struct worker *collision;
2140 #ifdef CONFIG_LOCKDEP
2142 * It is permissible to free the struct work_struct from
2143 * inside the function that is called from it, this we need to
2144 * take into account for lockdep too. To avoid bogus "held
2145 * lock freed" warnings as well as problems when looking into
2146 * work->lockdep_map, make a copy and use that here.
2148 struct lockdep_map lockdep_map;
2150 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2153 * Ensure we're on the correct CPU. DISASSOCIATED test is
2154 * necessary to avoid spurious warnings from rescuers servicing the
2155 * unbound or a disassociated gcwq.
2157 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2158 !(gcwq->flags & GCWQ_DISASSOCIATED) &&
2159 raw_smp_processor_id() != gcwq->cpu);
2162 * A single work shouldn't be executed concurrently by
2163 * multiple workers on a single cpu. Check whether anyone is
2164 * already processing the work. If so, defer the work to the
2165 * currently executing one.
2167 collision = find_worker_executing_work(gcwq, work);
2168 if (unlikely(collision)) {
2169 move_linked_works(work, &collision->scheduled, NULL);
2173 /* claim and dequeue */
2174 debug_work_deactivate(work);
2175 hash_add(gcwq->busy_hash, &worker->hentry, (unsigned long)work);
2176 worker->current_work = work;
2177 worker->current_func = work->func;
2178 worker->current_cwq = cwq;
2179 work_color = get_work_color(work);
2181 list_del_init(&work->entry);
2184 * CPU intensive works don't participate in concurrency
2185 * management. They're the scheduler's responsibility.
2187 if (unlikely(cpu_intensive))
2188 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2191 * Unbound gcwq isn't concurrency managed and work items should be
2192 * executed ASAP. Wake up another worker if necessary.
2194 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2195 wake_up_worker(pool);
2198 * Record the last CPU and clear PENDING which should be the last
2199 * update to @work. Also, do this inside @gcwq->lock so that
2200 * PENDING and queued state changes happen together while IRQ is
2203 set_work_cpu_and_clear_pending(work, gcwq->cpu);
2205 spin_unlock_irq(&gcwq->lock);
2207 lock_map_acquire_read(&cwq->wq->lockdep_map);
2208 lock_map_acquire(&lockdep_map);
2209 trace_workqueue_execute_start(work);
2210 worker->current_func(work);
2212 * While we must be careful to not use "work" after this, the trace
2213 * point will only record its address.
2215 trace_workqueue_execute_end(work);
2216 lock_map_release(&lockdep_map);
2217 lock_map_release(&cwq->wq->lockdep_map);
2219 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2220 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2221 " last function: %pf\n",
2222 current->comm, preempt_count(), task_pid_nr(current),
2223 worker->current_func);
2224 debug_show_held_locks(current);
2228 spin_lock_irq(&gcwq->lock);
2230 /* clear cpu intensive status */
2231 if (unlikely(cpu_intensive))
2232 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2234 /* we're done with it, release */
2235 hash_del(&worker->hentry);
2236 worker->current_work = NULL;
2237 worker->current_func = NULL;
2238 worker->current_cwq = NULL;
2239 cwq_dec_nr_in_flight(cwq, work_color);
2243 * process_scheduled_works - process scheduled works
2246 * Process all scheduled works. Please note that the scheduled list
2247 * may change while processing a work, so this function repeatedly
2248 * fetches a work from the top and executes it.
2251 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2254 static void process_scheduled_works(struct worker *worker)
2256 while (!list_empty(&worker->scheduled)) {
2257 struct work_struct *work = list_first_entry(&worker->scheduled,
2258 struct work_struct, entry);
2259 process_one_work(worker, work);
2264 * worker_thread - the worker thread function
2267 * The gcwq worker thread function. There's a single dynamic pool of
2268 * these per each cpu. These workers process all works regardless of
2269 * their specific target workqueue. The only exception is works which
2270 * belong to workqueues with a rescuer which will be explained in
2273 static int worker_thread(void *__worker)
2275 struct worker *worker = __worker;
2276 struct worker_pool *pool = worker->pool;
2277 struct global_cwq *gcwq = pool->gcwq;
2279 /* tell the scheduler that this is a workqueue worker */
2280 worker->task->flags |= PF_WQ_WORKER;
2282 spin_lock_irq(&gcwq->lock);
2284 /* we are off idle list if destruction or rebind is requested */
2285 if (unlikely(list_empty(&worker->entry))) {
2286 spin_unlock_irq(&gcwq->lock);
2288 /* if DIE is set, destruction is requested */
2289 if (worker->flags & WORKER_DIE) {
2290 worker->task->flags &= ~PF_WQ_WORKER;
2294 /* otherwise, rebind */
2295 idle_worker_rebind(worker);
2299 worker_leave_idle(worker);
2301 /* no more worker necessary? */
2302 if (!need_more_worker(pool))
2305 /* do we need to manage? */
2306 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2310 * ->scheduled list can only be filled while a worker is
2311 * preparing to process a work or actually processing it.
2312 * Make sure nobody diddled with it while I was sleeping.
2314 BUG_ON(!list_empty(&worker->scheduled));
2317 * When control reaches this point, we're guaranteed to have
2318 * at least one idle worker or that someone else has already
2319 * assumed the manager role.
2321 worker_clr_flags(worker, WORKER_PREP);
2324 struct work_struct *work =
2325 list_first_entry(&pool->worklist,
2326 struct work_struct, entry);
2328 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2329 /* optimization path, not strictly necessary */
2330 process_one_work(worker, work);
2331 if (unlikely(!list_empty(&worker->scheduled)))
2332 process_scheduled_works(worker);
2334 move_linked_works(work, &worker->scheduled, NULL);
2335 process_scheduled_works(worker);
2337 } while (keep_working(pool));
2339 worker_set_flags(worker, WORKER_PREP, false);
2341 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2345 * gcwq->lock is held and there's no work to process and no
2346 * need to manage, sleep. Workers are woken up only while
2347 * holding gcwq->lock or from local cpu, so setting the
2348 * current state before releasing gcwq->lock is enough to
2349 * prevent losing any event.
2351 worker_enter_idle(worker);
2352 __set_current_state(TASK_INTERRUPTIBLE);
2353 spin_unlock_irq(&gcwq->lock);
2359 * rescuer_thread - the rescuer thread function
2360 * @__wq: the associated workqueue
2362 * Workqueue rescuer thread function. There's one rescuer for each
2363 * workqueue which has WQ_RESCUER set.
2365 * Regular work processing on a gcwq may block trying to create a new
2366 * worker which uses GFP_KERNEL allocation which has slight chance of
2367 * developing into deadlock if some works currently on the same queue
2368 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2369 * the problem rescuer solves.
2371 * When such condition is possible, the gcwq summons rescuers of all
2372 * workqueues which have works queued on the gcwq and let them process
2373 * those works so that forward progress can be guaranteed.
2375 * This should happen rarely.
2377 static int rescuer_thread(void *__wq)
2379 struct workqueue_struct *wq = __wq;
2380 struct worker *rescuer = wq->rescuer;
2381 struct list_head *scheduled = &rescuer->scheduled;
2382 bool is_unbound = wq->flags & WQ_UNBOUND;
2385 set_user_nice(current, RESCUER_NICE_LEVEL);
2387 set_current_state(TASK_INTERRUPTIBLE);
2389 if (kthread_should_stop()) {
2390 __set_current_state(TASK_RUNNING);
2395 * See whether any cpu is asking for help. Unbounded
2396 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2398 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2399 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2400 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2401 struct worker_pool *pool = cwq->pool;
2402 struct global_cwq *gcwq = pool->gcwq;
2403 struct work_struct *work, *n;
2405 __set_current_state(TASK_RUNNING);
2406 mayday_clear_cpu(cpu, wq->mayday_mask);
2408 /* migrate to the target cpu if possible */
2409 rescuer->pool = pool;
2410 worker_maybe_bind_and_lock(rescuer);
2413 * Slurp in all works issued via this workqueue and
2416 BUG_ON(!list_empty(&rescuer->scheduled));
2417 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2418 if (get_work_cwq(work) == cwq)
2419 move_linked_works(work, scheduled, &n);
2421 process_scheduled_works(rescuer);
2424 * Leave this gcwq. If keep_working() is %true, notify a
2425 * regular worker; otherwise, we end up with 0 concurrency
2426 * and stalling the execution.
2428 if (keep_working(pool))
2429 wake_up_worker(pool);
2431 spin_unlock_irq(&gcwq->lock);
2439 struct work_struct work;
2440 struct completion done;
2443 static void wq_barrier_func(struct work_struct *work)
2445 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2446 complete(&barr->done);
2450 * insert_wq_barrier - insert a barrier work
2451 * @cwq: cwq to insert barrier into
2452 * @barr: wq_barrier to insert
2453 * @target: target work to attach @barr to
2454 * @worker: worker currently executing @target, NULL if @target is not executing
2456 * @barr is linked to @target such that @barr is completed only after
2457 * @target finishes execution. Please note that the ordering
2458 * guarantee is observed only with respect to @target and on the local
2461 * Currently, a queued barrier can't be canceled. This is because
2462 * try_to_grab_pending() can't determine whether the work to be
2463 * grabbed is at the head of the queue and thus can't clear LINKED
2464 * flag of the previous work while there must be a valid next work
2465 * after a work with LINKED flag set.
2467 * Note that when @worker is non-NULL, @target may be modified
2468 * underneath us, so we can't reliably determine cwq from @target.
2471 * spin_lock_irq(gcwq->lock).
2473 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2474 struct wq_barrier *barr,
2475 struct work_struct *target, struct worker *worker)
2477 struct list_head *head;
2478 unsigned int linked = 0;
2481 * debugobject calls are safe here even with gcwq->lock locked
2482 * as we know for sure that this will not trigger any of the
2483 * checks and call back into the fixup functions where we
2486 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2487 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2488 init_completion(&barr->done);
2491 * If @target is currently being executed, schedule the
2492 * barrier to the worker; otherwise, put it after @target.
2495 head = worker->scheduled.next;
2497 unsigned long *bits = work_data_bits(target);
2499 head = target->entry.next;
2500 /* there can already be other linked works, inherit and set */
2501 linked = *bits & WORK_STRUCT_LINKED;
2502 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2505 debug_work_activate(&barr->work);
2506 insert_work(cwq, &barr->work, head,
2507 work_color_to_flags(WORK_NO_COLOR) | linked);
2511 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2512 * @wq: workqueue being flushed
2513 * @flush_color: new flush color, < 0 for no-op
2514 * @work_color: new work color, < 0 for no-op
2516 * Prepare cwqs for workqueue flushing.
2518 * If @flush_color is non-negative, flush_color on all cwqs should be
2519 * -1. If no cwq has in-flight commands at the specified color, all
2520 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2521 * has in flight commands, its cwq->flush_color is set to
2522 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2523 * wakeup logic is armed and %true is returned.
2525 * The caller should have initialized @wq->first_flusher prior to
2526 * calling this function with non-negative @flush_color. If
2527 * @flush_color is negative, no flush color update is done and %false
2530 * If @work_color is non-negative, all cwqs should have the same
2531 * work_color which is previous to @work_color and all will be
2532 * advanced to @work_color.
2535 * mutex_lock(wq->flush_mutex).
2538 * %true if @flush_color >= 0 and there's something to flush. %false
2541 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2542 int flush_color, int work_color)
2547 if (flush_color >= 0) {
2548 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2549 atomic_set(&wq->nr_cwqs_to_flush, 1);
2552 for_each_cwq_cpu(cpu, wq) {
2553 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2554 struct global_cwq *gcwq = cwq->pool->gcwq;
2556 spin_lock_irq(&gcwq->lock);
2558 if (flush_color >= 0) {
2559 BUG_ON(cwq->flush_color != -1);
2561 if (cwq->nr_in_flight[flush_color]) {
2562 cwq->flush_color = flush_color;
2563 atomic_inc(&wq->nr_cwqs_to_flush);
2568 if (work_color >= 0) {
2569 BUG_ON(work_color != work_next_color(cwq->work_color));
2570 cwq->work_color = work_color;
2573 spin_unlock_irq(&gcwq->lock);
2576 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2577 complete(&wq->first_flusher->done);
2583 * flush_workqueue - ensure that any scheduled work has run to completion.
2584 * @wq: workqueue to flush
2586 * Forces execution of the workqueue and blocks until its completion.
2587 * This is typically used in driver shutdown handlers.
2589 * We sleep until all works which were queued on entry have been handled,
2590 * but we are not livelocked by new incoming ones.
2592 void flush_workqueue(struct workqueue_struct *wq)
2594 struct wq_flusher this_flusher = {
2595 .list = LIST_HEAD_INIT(this_flusher.list),
2597 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2601 lock_map_acquire(&wq->lockdep_map);
2602 lock_map_release(&wq->lockdep_map);
2604 mutex_lock(&wq->flush_mutex);
2607 * Start-to-wait phase
2609 next_color = work_next_color(wq->work_color);
2611 if (next_color != wq->flush_color) {
2613 * Color space is not full. The current work_color
2614 * becomes our flush_color and work_color is advanced
2617 BUG_ON(!list_empty(&wq->flusher_overflow));
2618 this_flusher.flush_color = wq->work_color;
2619 wq->work_color = next_color;
2621 if (!wq->first_flusher) {
2622 /* no flush in progress, become the first flusher */
2623 BUG_ON(wq->flush_color != this_flusher.flush_color);
2625 wq->first_flusher = &this_flusher;
2627 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2629 /* nothing to flush, done */
2630 wq->flush_color = next_color;
2631 wq->first_flusher = NULL;
2636 BUG_ON(wq->flush_color == this_flusher.flush_color);
2637 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2638 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2642 * Oops, color space is full, wait on overflow queue.
2643 * The next flush completion will assign us
2644 * flush_color and transfer to flusher_queue.
2646 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2649 mutex_unlock(&wq->flush_mutex);
2651 wait_for_completion(&this_flusher.done);
2654 * Wake-up-and-cascade phase
2656 * First flushers are responsible for cascading flushes and
2657 * handling overflow. Non-first flushers can simply return.
2659 if (wq->first_flusher != &this_flusher)
2662 mutex_lock(&wq->flush_mutex);
2664 /* we might have raced, check again with mutex held */
2665 if (wq->first_flusher != &this_flusher)
2668 wq->first_flusher = NULL;
2670 BUG_ON(!list_empty(&this_flusher.list));
2671 BUG_ON(wq->flush_color != this_flusher.flush_color);
2674 struct wq_flusher *next, *tmp;
2676 /* complete all the flushers sharing the current flush color */
2677 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2678 if (next->flush_color != wq->flush_color)
2680 list_del_init(&next->list);
2681 complete(&next->done);
2684 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2685 wq->flush_color != work_next_color(wq->work_color));
2687 /* this flush_color is finished, advance by one */
2688 wq->flush_color = work_next_color(wq->flush_color);
2690 /* one color has been freed, handle overflow queue */
2691 if (!list_empty(&wq->flusher_overflow)) {
2693 * Assign the same color to all overflowed
2694 * flushers, advance work_color and append to
2695 * flusher_queue. This is the start-to-wait
2696 * phase for these overflowed flushers.
2698 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2699 tmp->flush_color = wq->work_color;
2701 wq->work_color = work_next_color(wq->work_color);
2703 list_splice_tail_init(&wq->flusher_overflow,
2704 &wq->flusher_queue);
2705 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2708 if (list_empty(&wq->flusher_queue)) {
2709 BUG_ON(wq->flush_color != wq->work_color);
2714 * Need to flush more colors. Make the next flusher
2715 * the new first flusher and arm cwqs.
2717 BUG_ON(wq->flush_color == wq->work_color);
2718 BUG_ON(wq->flush_color != next->flush_color);
2720 list_del_init(&next->list);
2721 wq->first_flusher = next;
2723 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2727 * Meh... this color is already done, clear first
2728 * flusher and repeat cascading.
2730 wq->first_flusher = NULL;
2734 mutex_unlock(&wq->flush_mutex);
2736 EXPORT_SYMBOL_GPL(flush_workqueue);
2739 * drain_workqueue - drain a workqueue
2740 * @wq: workqueue to drain
2742 * Wait until the workqueue becomes empty. While draining is in progress,
2743 * only chain queueing is allowed. IOW, only currently pending or running
2744 * work items on @wq can queue further work items on it. @wq is flushed
2745 * repeatedly until it becomes empty. The number of flushing is detemined
2746 * by the depth of chaining and should be relatively short. Whine if it
2749 void drain_workqueue(struct workqueue_struct *wq)
2751 unsigned int flush_cnt = 0;
2755 * __queue_work() needs to test whether there are drainers, is much
2756 * hotter than drain_workqueue() and already looks at @wq->flags.
2757 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2759 spin_lock(&workqueue_lock);
2760 if (!wq->nr_drainers++)
2761 wq->flags |= WQ_DRAINING;
2762 spin_unlock(&workqueue_lock);
2764 flush_workqueue(wq);
2766 for_each_cwq_cpu(cpu, wq) {
2767 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2770 spin_lock_irq(&cwq->pool->gcwq->lock);
2771 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2772 spin_unlock_irq(&cwq->pool->gcwq->lock);
2777 if (++flush_cnt == 10 ||
2778 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2779 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2780 wq->name, flush_cnt);
2784 spin_lock(&workqueue_lock);
2785 if (!--wq->nr_drainers)
2786 wq->flags &= ~WQ_DRAINING;
2787 spin_unlock(&workqueue_lock);
2789 EXPORT_SYMBOL_GPL(drain_workqueue);
2791 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2793 struct worker *worker = NULL;
2794 struct global_cwq *gcwq;
2795 struct cpu_workqueue_struct *cwq;
2798 gcwq = get_work_gcwq(work);
2802 spin_lock_irq(&gcwq->lock);
2803 if (!list_empty(&work->entry)) {
2805 * See the comment near try_to_grab_pending()->smp_rmb().
2806 * If it was re-queued to a different gcwq under us, we
2807 * are not going to wait.
2810 cwq = get_work_cwq(work);
2811 if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
2814 worker = find_worker_executing_work(gcwq, work);
2817 cwq = worker->current_cwq;
2820 insert_wq_barrier(cwq, barr, work, worker);
2821 spin_unlock_irq(&gcwq->lock);
2824 * If @max_active is 1 or rescuer is in use, flushing another work
2825 * item on the same workqueue may lead to deadlock. Make sure the
2826 * flusher is not running on the same workqueue by verifying write
2829 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2830 lock_map_acquire(&cwq->wq->lockdep_map);
2832 lock_map_acquire_read(&cwq->wq->lockdep_map);
2833 lock_map_release(&cwq->wq->lockdep_map);
2837 spin_unlock_irq(&gcwq->lock);
2842 * flush_work - wait for a work to finish executing the last queueing instance
2843 * @work: the work to flush
2845 * Wait until @work has finished execution. @work is guaranteed to be idle
2846 * on return if it hasn't been requeued since flush started.
2849 * %true if flush_work() waited for the work to finish execution,
2850 * %false if it was already idle.
2852 bool flush_work(struct work_struct *work)
2854 struct wq_barrier barr;
2856 lock_map_acquire(&work->lockdep_map);
2857 lock_map_release(&work->lockdep_map);
2859 if (start_flush_work(work, &barr)) {
2860 wait_for_completion(&barr.done);
2861 destroy_work_on_stack(&barr.work);
2867 EXPORT_SYMBOL_GPL(flush_work);
2869 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2871 unsigned long flags;
2875 ret = try_to_grab_pending(work, is_dwork, &flags);
2877 * If someone else is canceling, wait for the same event it
2878 * would be waiting for before retrying.
2880 if (unlikely(ret == -ENOENT))
2882 } while (unlikely(ret < 0));
2884 /* tell other tasks trying to grab @work to back off */
2885 mark_work_canceling(work);
2886 local_irq_restore(flags);
2889 clear_work_data(work);
2894 * cancel_work_sync - cancel a work and wait for it to finish
2895 * @work: the work to cancel
2897 * Cancel @work and wait for its execution to finish. This function
2898 * can be used even if the work re-queues itself or migrates to
2899 * another workqueue. On return from this function, @work is
2900 * guaranteed to be not pending or executing on any CPU.
2902 * cancel_work_sync(&delayed_work->work) must not be used for
2903 * delayed_work's. Use cancel_delayed_work_sync() instead.
2905 * The caller must ensure that the workqueue on which @work was last
2906 * queued can't be destroyed before this function returns.
2909 * %true if @work was pending, %false otherwise.
2911 bool cancel_work_sync(struct work_struct *work)
2913 return __cancel_work_timer(work, false);
2915 EXPORT_SYMBOL_GPL(cancel_work_sync);
2918 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2919 * @dwork: the delayed work to flush
2921 * Delayed timer is cancelled and the pending work is queued for
2922 * immediate execution. Like flush_work(), this function only
2923 * considers the last queueing instance of @dwork.
2926 * %true if flush_work() waited for the work to finish execution,
2927 * %false if it was already idle.
2929 bool flush_delayed_work(struct delayed_work *dwork)
2931 local_irq_disable();
2932 if (del_timer_sync(&dwork->timer))
2933 __queue_work(dwork->cpu,
2934 get_work_cwq(&dwork->work)->wq, &dwork->work);
2936 return flush_work(&dwork->work);
2938 EXPORT_SYMBOL(flush_delayed_work);
2941 * cancel_delayed_work - cancel a delayed work
2942 * @dwork: delayed_work to cancel
2944 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2945 * and canceled; %false if wasn't pending. Note that the work callback
2946 * function may still be running on return, unless it returns %true and the
2947 * work doesn't re-arm itself. Explicitly flush or use
2948 * cancel_delayed_work_sync() to wait on it.
2950 * This function is safe to call from any context including IRQ handler.
2952 bool cancel_delayed_work(struct delayed_work *dwork)
2954 unsigned long flags;
2958 ret = try_to_grab_pending(&dwork->work, true, &flags);
2959 } while (unlikely(ret == -EAGAIN));
2961 if (unlikely(ret < 0))
2964 set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work));
2965 local_irq_restore(flags);
2968 EXPORT_SYMBOL(cancel_delayed_work);
2971 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2972 * @dwork: the delayed work cancel
2974 * This is cancel_work_sync() for delayed works.
2977 * %true if @dwork was pending, %false otherwise.
2979 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2981 return __cancel_work_timer(&dwork->work, true);
2983 EXPORT_SYMBOL(cancel_delayed_work_sync);
2986 * schedule_work_on - put work task on a specific cpu
2987 * @cpu: cpu to put the work task on
2988 * @work: job to be done
2990 * This puts a job on a specific cpu
2992 bool schedule_work_on(int cpu, struct work_struct *work)
2994 return queue_work_on(cpu, system_wq, work);
2996 EXPORT_SYMBOL(schedule_work_on);
2999 * schedule_work - put work task in global workqueue
3000 * @work: job to be done
3002 * Returns %false if @work was already on the kernel-global workqueue and
3005 * This puts a job in the kernel-global workqueue if it was not already
3006 * queued and leaves it in the same position on the kernel-global
3007 * workqueue otherwise.
3009 bool schedule_work(struct work_struct *work)
3011 return queue_work(system_wq, work);
3013 EXPORT_SYMBOL(schedule_work);
3016 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3018 * @dwork: job to be done
3019 * @delay: number of jiffies to wait
3021 * After waiting for a given time this puts a job in the kernel-global
3022 * workqueue on the specified CPU.
3024 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3025 unsigned long delay)
3027 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3029 EXPORT_SYMBOL(schedule_delayed_work_on);
3032 * schedule_delayed_work - put work task in global workqueue after delay
3033 * @dwork: job to be done
3034 * @delay: number of jiffies to wait or 0 for immediate execution
3036 * After waiting for a given time this puts a job in the kernel-global
3039 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3041 return queue_delayed_work(system_wq, dwork, delay);
3043 EXPORT_SYMBOL(schedule_delayed_work);
3046 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3047 * @func: the function to call
3049 * schedule_on_each_cpu() executes @func on each online CPU using the
3050 * system workqueue and blocks until all CPUs have completed.
3051 * schedule_on_each_cpu() is very slow.
3054 * 0 on success, -errno on failure.
3056 int schedule_on_each_cpu(work_func_t func)
3059 struct work_struct __percpu *works;
3061 works = alloc_percpu(struct work_struct);
3067 for_each_online_cpu(cpu) {
3068 struct work_struct *work = per_cpu_ptr(works, cpu);
3070 INIT_WORK(work, func);
3071 schedule_work_on(cpu, work);
3074 for_each_online_cpu(cpu)
3075 flush_work(per_cpu_ptr(works, cpu));
3083 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3085 * Forces execution of the kernel-global workqueue and blocks until its
3088 * Think twice before calling this function! It's very easy to get into
3089 * trouble if you don't take great care. Either of the following situations
3090 * will lead to deadlock:
3092 * One of the work items currently on the workqueue needs to acquire
3093 * a lock held by your code or its caller.
3095 * Your code is running in the context of a work routine.
3097 * They will be detected by lockdep when they occur, but the first might not
3098 * occur very often. It depends on what work items are on the workqueue and
3099 * what locks they need, which you have no control over.
3101 * In most situations flushing the entire workqueue is overkill; you merely
3102 * need to know that a particular work item isn't queued and isn't running.
3103 * In such cases you should use cancel_delayed_work_sync() or
3104 * cancel_work_sync() instead.
3106 void flush_scheduled_work(void)
3108 flush_workqueue(system_wq);
3110 EXPORT_SYMBOL(flush_scheduled_work);
3113 * execute_in_process_context - reliably execute the routine with user context
3114 * @fn: the function to execute
3115 * @ew: guaranteed storage for the execute work structure (must
3116 * be available when the work executes)
3118 * Executes the function immediately if process context is available,
3119 * otherwise schedules the function for delayed execution.
3121 * Returns: 0 - function was executed
3122 * 1 - function was scheduled for execution
3124 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3126 if (!in_interrupt()) {
3131 INIT_WORK(&ew->work, fn);
3132 schedule_work(&ew->work);
3136 EXPORT_SYMBOL_GPL(execute_in_process_context);
3138 int keventd_up(void)
3140 return system_wq != NULL;
3143 static int alloc_cwqs(struct workqueue_struct *wq)
3146 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3147 * Make sure that the alignment isn't lower than that of
3148 * unsigned long long.
3150 const size_t size = sizeof(struct cpu_workqueue_struct);
3151 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3152 __alignof__(unsigned long long));
3154 if (!(wq->flags & WQ_UNBOUND))
3155 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3160 * Allocate enough room to align cwq and put an extra
3161 * pointer at the end pointing back to the originally
3162 * allocated pointer which will be used for free.
3164 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3166 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3167 *(void **)(wq->cpu_wq.single + 1) = ptr;
3171 /* just in case, make sure it's actually aligned */
3172 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3173 return wq->cpu_wq.v ? 0 : -ENOMEM;
3176 static void free_cwqs(struct workqueue_struct *wq)
3178 if (!(wq->flags & WQ_UNBOUND))
3179 free_percpu(wq->cpu_wq.pcpu);
3180 else if (wq->cpu_wq.single) {
3181 /* the pointer to free is stored right after the cwq */
3182 kfree(*(void **)(wq->cpu_wq.single + 1));
3186 static int wq_clamp_max_active(int max_active, unsigned int flags,
3189 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3191 if (max_active < 1 || max_active > lim)
3192 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3193 max_active, name, 1, lim);
3195 return clamp_val(max_active, 1, lim);
3198 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3201 struct lock_class_key *key,
3202 const char *lock_name, ...)
3204 va_list args, args1;
3205 struct workqueue_struct *wq;
3209 /* determine namelen, allocate wq and format name */
3210 va_start(args, lock_name);
3211 va_copy(args1, args);
3212 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3214 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3218 vsnprintf(wq->name, namelen, fmt, args1);
3223 * Workqueues which may be used during memory reclaim should
3224 * have a rescuer to guarantee forward progress.
3226 if (flags & WQ_MEM_RECLAIM)
3227 flags |= WQ_RESCUER;
3229 max_active = max_active ?: WQ_DFL_ACTIVE;
3230 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3234 wq->saved_max_active = max_active;
3235 mutex_init(&wq->flush_mutex);
3236 atomic_set(&wq->nr_cwqs_to_flush, 0);
3237 INIT_LIST_HEAD(&wq->flusher_queue);
3238 INIT_LIST_HEAD(&wq->flusher_overflow);
3240 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3241 INIT_LIST_HEAD(&wq->list);
3243 if (alloc_cwqs(wq) < 0)
3246 for_each_cwq_cpu(cpu, wq) {
3247 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3248 struct global_cwq *gcwq = get_gcwq(cpu);
3249 int pool_idx = (bool)(flags & WQ_HIGHPRI);
3251 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3252 cwq->pool = &gcwq->pools[pool_idx];
3254 cwq->flush_color = -1;
3255 cwq->max_active = max_active;
3256 INIT_LIST_HEAD(&cwq->delayed_works);
3259 if (flags & WQ_RESCUER) {
3260 struct worker *rescuer;
3262 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3265 wq->rescuer = rescuer = alloc_worker();
3269 rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3271 if (IS_ERR(rescuer->task))
3274 rescuer->task->flags |= PF_THREAD_BOUND;
3275 wake_up_process(rescuer->task);
3279 * workqueue_lock protects global freeze state and workqueues
3280 * list. Grab it, set max_active accordingly and add the new
3281 * workqueue to workqueues list.
3283 spin_lock(&workqueue_lock);
3285 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3286 for_each_cwq_cpu(cpu, wq)
3287 get_cwq(cpu, wq)->max_active = 0;
3289 list_add(&wq->list, &workqueues);
3291 spin_unlock(&workqueue_lock);
3297 free_mayday_mask(wq->mayday_mask);
3303 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3306 * destroy_workqueue - safely terminate a workqueue
3307 * @wq: target workqueue
3309 * Safely destroy a workqueue. All work currently pending will be done first.
3311 void destroy_workqueue(struct workqueue_struct *wq)
3315 /* drain it before proceeding with destruction */
3316 drain_workqueue(wq);
3319 * wq list is used to freeze wq, remove from list after
3320 * flushing is complete in case freeze races us.
3322 spin_lock(&workqueue_lock);
3323 list_del(&wq->list);
3324 spin_unlock(&workqueue_lock);
3327 for_each_cwq_cpu(cpu, wq) {
3328 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3331 for (i = 0; i < WORK_NR_COLORS; i++)
3332 BUG_ON(cwq->nr_in_flight[i]);
3333 BUG_ON(cwq->nr_active);
3334 BUG_ON(!list_empty(&cwq->delayed_works));
3337 if (wq->flags & WQ_RESCUER) {
3338 kthread_stop(wq->rescuer->task);
3339 free_mayday_mask(wq->mayday_mask);
3346 EXPORT_SYMBOL_GPL(destroy_workqueue);
3349 * cwq_set_max_active - adjust max_active of a cwq
3350 * @cwq: target cpu_workqueue_struct
3351 * @max_active: new max_active value.
3353 * Set @cwq->max_active to @max_active and activate delayed works if
3357 * spin_lock_irq(gcwq->lock).
3359 static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
3361 cwq->max_active = max_active;
3363 while (!list_empty(&cwq->delayed_works) &&
3364 cwq->nr_active < cwq->max_active)
3365 cwq_activate_first_delayed(cwq);
3369 * workqueue_set_max_active - adjust max_active of a workqueue
3370 * @wq: target workqueue
3371 * @max_active: new max_active value.
3373 * Set max_active of @wq to @max_active.
3376 * Don't call from IRQ context.
3378 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3382 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3384 spin_lock(&workqueue_lock);
3386 wq->saved_max_active = max_active;
3388 for_each_cwq_cpu(cpu, wq) {
3389 struct global_cwq *gcwq = get_gcwq(cpu);
3391 spin_lock_irq(&gcwq->lock);
3393 if (!(wq->flags & WQ_FREEZABLE) ||
3394 !(gcwq->flags & GCWQ_FREEZING))
3395 cwq_set_max_active(get_cwq(gcwq->cpu, wq), max_active);
3397 spin_unlock_irq(&gcwq->lock);
3400 spin_unlock(&workqueue_lock);
3402 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3405 * workqueue_congested - test whether a workqueue is congested
3406 * @cpu: CPU in question
3407 * @wq: target workqueue
3409 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3410 * no synchronization around this function and the test result is
3411 * unreliable and only useful as advisory hints or for debugging.
3414 * %true if congested, %false otherwise.
3416 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3418 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3420 return !list_empty(&cwq->delayed_works);
3422 EXPORT_SYMBOL_GPL(workqueue_congested);
3425 * work_cpu - return the last known associated cpu for @work
3426 * @work: the work of interest
3429 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3431 unsigned int work_cpu(struct work_struct *work)
3433 struct global_cwq *gcwq = get_work_gcwq(work);
3435 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3437 EXPORT_SYMBOL_GPL(work_cpu);
3440 * work_busy - test whether a work is currently pending or running
3441 * @work: the work to be tested
3443 * Test whether @work is currently pending or running. There is no
3444 * synchronization around this function and the test result is
3445 * unreliable and only useful as advisory hints or for debugging.
3446 * Especially for reentrant wqs, the pending state might hide the
3450 * OR'd bitmask of WORK_BUSY_* bits.
3452 unsigned int work_busy(struct work_struct *work)
3454 struct global_cwq *gcwq = get_work_gcwq(work);
3455 unsigned long flags;
3456 unsigned int ret = 0;
3461 spin_lock_irqsave(&gcwq->lock, flags);
3463 if (work_pending(work))
3464 ret |= WORK_BUSY_PENDING;
3465 if (find_worker_executing_work(gcwq, work))
3466 ret |= WORK_BUSY_RUNNING;
3468 spin_unlock_irqrestore(&gcwq->lock, flags);
3472 EXPORT_SYMBOL_GPL(work_busy);
3477 * There are two challenges in supporting CPU hotplug. Firstly, there
3478 * are a lot of assumptions on strong associations among work, cwq and
3479 * gcwq which make migrating pending and scheduled works very
3480 * difficult to implement without impacting hot paths. Secondly,
3481 * gcwqs serve mix of short, long and very long running works making
3482 * blocked draining impractical.
3484 * This is solved by allowing a gcwq to be disassociated from the CPU
3485 * running as an unbound one and allowing it to be reattached later if the
3486 * cpu comes back online.
3489 /* claim manager positions of all pools */
3490 static void gcwq_claim_assoc_and_lock(struct global_cwq *gcwq)
3492 struct worker_pool *pool;
3494 for_each_worker_pool(pool, gcwq)
3495 mutex_lock_nested(&pool->assoc_mutex, pool - gcwq->pools);
3496 spin_lock_irq(&gcwq->lock);
3499 /* release manager positions */
3500 static void gcwq_release_assoc_and_unlock(struct global_cwq *gcwq)
3502 struct worker_pool *pool;
3504 spin_unlock_irq(&gcwq->lock);
3505 for_each_worker_pool(pool, gcwq)
3506 mutex_unlock(&pool->assoc_mutex);
3509 static void gcwq_unbind_fn(struct work_struct *work)
3511 struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3512 struct worker_pool *pool;
3513 struct worker *worker;
3514 struct hlist_node *pos;
3517 BUG_ON(gcwq->cpu != smp_processor_id());
3519 gcwq_claim_assoc_and_lock(gcwq);
3522 * We've claimed all manager positions. Make all workers unbound
3523 * and set DISASSOCIATED. Before this, all workers except for the
3524 * ones which are still executing works from before the last CPU
3525 * down must be on the cpu. After this, they may become diasporas.
3527 for_each_worker_pool(pool, gcwq)
3528 list_for_each_entry(worker, &pool->idle_list, entry)
3529 worker->flags |= WORKER_UNBOUND;
3531 for_each_busy_worker(worker, i, pos, gcwq)
3532 worker->flags |= WORKER_UNBOUND;
3534 gcwq->flags |= GCWQ_DISASSOCIATED;
3536 gcwq_release_assoc_and_unlock(gcwq);
3539 * Call schedule() so that we cross rq->lock and thus can guarantee
3540 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3541 * as scheduler callbacks may be invoked from other cpus.
3546 * Sched callbacks are disabled now. Zap nr_running. After this,
3547 * nr_running stays zero and need_more_worker() and keep_working()
3548 * are always true as long as the worklist is not empty. @gcwq now
3549 * behaves as unbound (in terms of concurrency management) gcwq
3550 * which is served by workers tied to the CPU.
3552 * On return from this function, the current worker would trigger
3553 * unbound chain execution of pending work items if other workers
3556 for_each_worker_pool(pool, gcwq)
3557 atomic_set(get_pool_nr_running(pool), 0);
3561 * Workqueues should be brought up before normal priority CPU notifiers.
3562 * This will be registered high priority CPU notifier.
3564 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3565 unsigned long action,
3568 unsigned int cpu = (unsigned long)hcpu;
3569 struct global_cwq *gcwq = get_gcwq(cpu);
3570 struct worker_pool *pool;
3572 switch (action & ~CPU_TASKS_FROZEN) {
3573 case CPU_UP_PREPARE:
3574 for_each_worker_pool(pool, gcwq) {
3575 struct worker *worker;
3577 if (pool->nr_workers)
3580 worker = create_worker(pool);
3584 spin_lock_irq(&gcwq->lock);
3585 start_worker(worker);
3586 spin_unlock_irq(&gcwq->lock);
3590 case CPU_DOWN_FAILED:
3592 gcwq_claim_assoc_and_lock(gcwq);
3593 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3594 rebind_workers(gcwq);
3595 gcwq_release_assoc_and_unlock(gcwq);
3602 * Workqueues should be brought down after normal priority CPU notifiers.
3603 * This will be registered as low priority CPU notifier.
3605 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3606 unsigned long action,
3609 unsigned int cpu = (unsigned long)hcpu;
3610 struct work_struct unbind_work;
3612 switch (action & ~CPU_TASKS_FROZEN) {
3613 case CPU_DOWN_PREPARE:
3614 /* unbinding should happen on the local CPU */
3615 INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3616 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3617 flush_work(&unbind_work);
3625 struct work_for_cpu {
3626 struct work_struct work;
3632 static void work_for_cpu_fn(struct work_struct *work)
3634 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3636 wfc->ret = wfc->fn(wfc->arg);
3640 * work_on_cpu - run a function in user context on a particular cpu
3641 * @cpu: the cpu to run on
3642 * @fn: the function to run
3643 * @arg: the function arg
3645 * This will return the value @fn returns.
3646 * It is up to the caller to ensure that the cpu doesn't go offline.
3647 * The caller must not hold any locks which would prevent @fn from completing.
3649 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3651 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3653 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3654 schedule_work_on(cpu, &wfc.work);
3655 flush_work(&wfc.work);
3658 EXPORT_SYMBOL_GPL(work_on_cpu);
3659 #endif /* CONFIG_SMP */
3661 #ifdef CONFIG_FREEZER
3664 * freeze_workqueues_begin - begin freezing workqueues
3666 * Start freezing workqueues. After this function returns, all freezable
3667 * workqueues will queue new works to their frozen_works list instead of
3671 * Grabs and releases workqueue_lock and gcwq->lock's.
3673 void freeze_workqueues_begin(void)
3677 spin_lock(&workqueue_lock);
3679 BUG_ON(workqueue_freezing);
3680 workqueue_freezing = true;
3682 for_each_gcwq_cpu(cpu) {
3683 struct global_cwq *gcwq = get_gcwq(cpu);
3684 struct workqueue_struct *wq;
3686 spin_lock_irq(&gcwq->lock);
3688 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3689 gcwq->flags |= GCWQ_FREEZING;
3691 list_for_each_entry(wq, &workqueues, list) {
3692 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3694 if (cwq && wq->flags & WQ_FREEZABLE)
3695 cwq->max_active = 0;
3698 spin_unlock_irq(&gcwq->lock);
3701 spin_unlock(&workqueue_lock);
3705 * freeze_workqueues_busy - are freezable workqueues still busy?
3707 * Check whether freezing is complete. This function must be called
3708 * between freeze_workqueues_begin() and thaw_workqueues().
3711 * Grabs and releases workqueue_lock.
3714 * %true if some freezable workqueues are still busy. %false if freezing
3717 bool freeze_workqueues_busy(void)
3722 spin_lock(&workqueue_lock);
3724 BUG_ON(!workqueue_freezing);
3726 for_each_gcwq_cpu(cpu) {
3727 struct workqueue_struct *wq;
3729 * nr_active is monotonically decreasing. It's safe
3730 * to peek without lock.
3732 list_for_each_entry(wq, &workqueues, list) {
3733 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3735 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3738 BUG_ON(cwq->nr_active < 0);
3739 if (cwq->nr_active) {
3746 spin_unlock(&workqueue_lock);
3751 * thaw_workqueues - thaw workqueues
3753 * Thaw workqueues. Normal queueing is restored and all collected
3754 * frozen works are transferred to their respective gcwq worklists.
3757 * Grabs and releases workqueue_lock and gcwq->lock's.
3759 void thaw_workqueues(void)
3763 spin_lock(&workqueue_lock);
3765 if (!workqueue_freezing)
3768 for_each_gcwq_cpu(cpu) {
3769 struct global_cwq *gcwq = get_gcwq(cpu);
3770 struct worker_pool *pool;
3771 struct workqueue_struct *wq;
3773 spin_lock_irq(&gcwq->lock);
3775 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3776 gcwq->flags &= ~GCWQ_FREEZING;
3778 list_for_each_entry(wq, &workqueues, list) {
3779 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3781 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3784 /* restore max_active and repopulate worklist */
3785 cwq_set_max_active(cwq, wq->saved_max_active);
3788 for_each_worker_pool(pool, gcwq)
3789 wake_up_worker(pool);
3791 spin_unlock_irq(&gcwq->lock);
3794 workqueue_freezing = false;
3796 spin_unlock(&workqueue_lock);
3798 #endif /* CONFIG_FREEZER */
3800 static int __init init_workqueues(void)
3804 /* make sure we have enough bits for OFFQ CPU number */
3805 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
3808 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3809 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3811 /* initialize gcwqs */
3812 for_each_gcwq_cpu(cpu) {
3813 struct global_cwq *gcwq = get_gcwq(cpu);
3814 struct worker_pool *pool;
3816 spin_lock_init(&gcwq->lock);
3818 gcwq->flags |= GCWQ_DISASSOCIATED;
3820 hash_init(gcwq->busy_hash);
3822 for_each_worker_pool(pool, gcwq) {
3824 INIT_LIST_HEAD(&pool->worklist);
3825 INIT_LIST_HEAD(&pool->idle_list);
3827 init_timer_deferrable(&pool->idle_timer);
3828 pool->idle_timer.function = idle_worker_timeout;
3829 pool->idle_timer.data = (unsigned long)pool;
3831 setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3832 (unsigned long)pool);
3834 mutex_init(&pool->assoc_mutex);
3835 ida_init(&pool->worker_ida);
3839 /* create the initial worker */
3840 for_each_online_gcwq_cpu(cpu) {
3841 struct global_cwq *gcwq = get_gcwq(cpu);
3842 struct worker_pool *pool;
3844 if (cpu != WORK_CPU_UNBOUND)
3845 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3847 for_each_worker_pool(pool, gcwq) {
3848 struct worker *worker;
3850 worker = create_worker(pool);
3852 spin_lock_irq(&gcwq->lock);
3853 start_worker(worker);
3854 spin_unlock_irq(&gcwq->lock);
3858 system_wq = alloc_workqueue("events", 0, 0);
3859 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3860 system_long_wq = alloc_workqueue("events_long", 0, 0);
3861 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3862 WQ_UNBOUND_MAX_ACTIVE);
3863 system_freezable_wq = alloc_workqueue("events_freezable",
3865 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3866 !system_unbound_wq || !system_freezable_wq);
3869 early_initcall(init_workqueues);