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/jhash.h>
45 #include <linux/hashtable.h>
46 #include <linux/rculist.h>
48 #include "workqueue_internal.h"
54 * A bound pool is either associated or disassociated with its CPU.
55 * While associated (!DISASSOCIATED), all workers are bound to the
56 * CPU and none has %WORKER_UNBOUND set and concurrency management
59 * While DISASSOCIATED, the cpu may be offline and all workers have
60 * %WORKER_UNBOUND set and concurrency management disabled, and may
61 * be executing on any CPU. The pool behaves as an unbound one.
63 * Note that DISASSOCIATED can be flipped only while holding
64 * assoc_mutex to avoid changing binding state while
65 * create_worker() is in progress.
67 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
68 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
69 POOL_FREEZING = 1 << 3, /* freeze in progress */
72 WORKER_STARTED = 1 << 0, /* started */
73 WORKER_DIE = 1 << 1, /* die die die */
74 WORKER_IDLE = 1 << 2, /* is idle */
75 WORKER_PREP = 1 << 3, /* preparing to run works */
76 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
77 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND |
82 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
84 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
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: pool->lock protected. Access with pool->lock held.
115 * X: During normal operation, modification requires pool->lock and should
116 * be done only from local cpu. Either disabling preemption on local
117 * cpu or grabbing pool->lock is enough for read access. If
118 * POOL_DISASSOCIATED is set, it's identical to L.
120 * F: wq->flush_mutex protected.
122 * W: workqueue_lock protected.
124 * R: workqueue_lock protected for writes. Sched-RCU protected for reads.
126 * FR: wq->flush_mutex and workqueue_lock protected for writes. Sched-RCU
127 * protected for reads.
130 /* struct worker is defined in workqueue_internal.h */
133 spinlock_t lock; /* the pool lock */
134 int cpu; /* I: the associated cpu */
135 int id; /* I: pool ID */
136 unsigned int flags; /* X: flags */
138 struct list_head worklist; /* L: list of pending works */
139 int nr_workers; /* L: total number of workers */
141 /* nr_idle includes the ones off idle_list for rebinding */
142 int nr_idle; /* L: currently idle ones */
144 struct list_head idle_list; /* X: list of idle workers */
145 struct timer_list idle_timer; /* L: worker idle timeout */
146 struct timer_list mayday_timer; /* L: SOS timer for workers */
148 /* workers are chained either in busy_hash or idle_list */
149 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
150 /* L: hash of busy workers */
152 struct mutex manager_arb; /* manager arbitration */
153 struct mutex assoc_mutex; /* protect POOL_DISASSOCIATED */
154 struct ida worker_ida; /* L: for worker IDs */
156 struct workqueue_attrs *attrs; /* I: worker attributes */
157 struct hlist_node hash_node; /* R: unbound_pool_hash node */
158 int refcnt; /* refcnt for unbound pools */
161 * The current concurrency level. As it's likely to be accessed
162 * from other CPUs during try_to_wake_up(), put it in a separate
165 atomic_t nr_running ____cacheline_aligned_in_smp;
168 * Destruction of pool is sched-RCU protected to allow dereferences
169 * from get_work_pool().
172 } ____cacheline_aligned_in_smp;
175 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
176 * of work_struct->data are used for flags and the remaining high bits
177 * point to the pwq; thus, pwqs need to be aligned at two's power of the
178 * number of flag bits.
180 struct pool_workqueue {
181 struct worker_pool *pool; /* I: the associated pool */
182 struct workqueue_struct *wq; /* I: the owning workqueue */
183 int work_color; /* L: current color */
184 int flush_color; /* L: flushing color */
185 int refcnt; /* L: reference count */
186 int nr_in_flight[WORK_NR_COLORS];
187 /* L: nr of in_flight works */
188 int nr_active; /* L: nr of active works */
189 int max_active; /* L: max active works */
190 struct list_head delayed_works; /* L: delayed works */
191 struct list_head pwqs_node; /* FR: node on wq->pwqs */
192 struct list_head mayday_node; /* W: node on wq->maydays */
195 * Release of unbound pwq is punted to system_wq. See put_pwq()
196 * and pwq_unbound_release_workfn() for details. pool_workqueue
197 * itself is also sched-RCU protected so that the first pwq can be
198 * determined without grabbing workqueue_lock.
200 struct work_struct unbound_release_work;
202 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
205 * Structure used to wait for workqueue flush.
208 struct list_head list; /* F: list of flushers */
209 int flush_color; /* F: flush color waiting for */
210 struct completion done; /* flush completion */
214 * The externally visible workqueue abstraction is an array of
215 * per-CPU workqueues:
217 struct workqueue_struct {
218 unsigned int flags; /* W: WQ_* flags */
219 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */
220 struct list_head pwqs; /* FR: all pwqs of this wq */
221 struct list_head list; /* W: list of all workqueues */
223 struct mutex flush_mutex; /* protects wq flushing */
224 int work_color; /* F: current work color */
225 int flush_color; /* F: current flush color */
226 atomic_t nr_pwqs_to_flush; /* flush in progress */
227 struct wq_flusher *first_flusher; /* F: first flusher */
228 struct list_head flusher_queue; /* F: flush waiters */
229 struct list_head flusher_overflow; /* F: flush overflow list */
231 struct list_head maydays; /* W: pwqs requesting rescue */
232 struct worker *rescuer; /* I: rescue worker */
234 int nr_drainers; /* W: drain in progress */
235 int saved_max_active; /* W: saved pwq max_active */
236 #ifdef CONFIG_LOCKDEP
237 struct lockdep_map lockdep_map;
239 char name[]; /* I: workqueue name */
242 static struct kmem_cache *pwq_cache;
244 /* hash of all unbound pools keyed by pool->attrs */
245 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
247 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
249 struct workqueue_struct *system_wq __read_mostly;
250 EXPORT_SYMBOL_GPL(system_wq);
251 struct workqueue_struct *system_highpri_wq __read_mostly;
252 EXPORT_SYMBOL_GPL(system_highpri_wq);
253 struct workqueue_struct *system_long_wq __read_mostly;
254 EXPORT_SYMBOL_GPL(system_long_wq);
255 struct workqueue_struct *system_unbound_wq __read_mostly;
256 EXPORT_SYMBOL_GPL(system_unbound_wq);
257 struct workqueue_struct *system_freezable_wq __read_mostly;
258 EXPORT_SYMBOL_GPL(system_freezable_wq);
260 #define CREATE_TRACE_POINTS
261 #include <trace/events/workqueue.h>
263 #define assert_rcu_or_wq_lock() \
264 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
265 lockdep_is_held(&workqueue_lock), \
266 "sched RCU or workqueue lock should be held")
268 #define for_each_cpu_worker_pool(pool, cpu) \
269 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
270 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
273 #define for_each_busy_worker(worker, i, pool) \
274 hash_for_each(pool->busy_hash, i, worker, hentry)
277 * for_each_pool - iterate through all worker_pools in the system
278 * @pool: iteration cursor
279 * @id: integer used for iteration
281 * This must be called either with workqueue_lock held or sched RCU read
282 * locked. If the pool needs to be used beyond the locking in effect, the
283 * caller is responsible for guaranteeing that the pool stays online.
285 * The if/else clause exists only for the lockdep assertion and can be
288 #define for_each_pool(pool, id) \
289 idr_for_each_entry(&worker_pool_idr, pool, id) \
290 if (({ assert_rcu_or_wq_lock(); false; })) { } \
294 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
295 * @pwq: iteration cursor
296 * @wq: the target workqueue
298 * This must be called either with workqueue_lock held or sched RCU read
299 * locked. If the pwq needs to be used beyond the locking in effect, the
300 * caller is responsible for guaranteeing that the pwq stays online.
302 * The if/else clause exists only for the lockdep assertion and can be
305 #define for_each_pwq(pwq, wq) \
306 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
307 if (({ assert_rcu_or_wq_lock(); false; })) { } \
310 #ifdef CONFIG_DEBUG_OBJECTS_WORK
312 static struct debug_obj_descr work_debug_descr;
314 static void *work_debug_hint(void *addr)
316 return ((struct work_struct *) addr)->func;
320 * fixup_init is called when:
321 * - an active object is initialized
323 static int work_fixup_init(void *addr, enum debug_obj_state state)
325 struct work_struct *work = addr;
328 case ODEBUG_STATE_ACTIVE:
329 cancel_work_sync(work);
330 debug_object_init(work, &work_debug_descr);
338 * fixup_activate is called when:
339 * - an active object is activated
340 * - an unknown object is activated (might be a statically initialized object)
342 static int work_fixup_activate(void *addr, enum debug_obj_state state)
344 struct work_struct *work = addr;
348 case ODEBUG_STATE_NOTAVAILABLE:
350 * This is not really a fixup. The work struct was
351 * statically initialized. We just make sure that it
352 * is tracked in the object tracker.
354 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
355 debug_object_init(work, &work_debug_descr);
356 debug_object_activate(work, &work_debug_descr);
362 case ODEBUG_STATE_ACTIVE:
371 * fixup_free is called when:
372 * - an active object is freed
374 static int work_fixup_free(void *addr, enum debug_obj_state state)
376 struct work_struct *work = addr;
379 case ODEBUG_STATE_ACTIVE:
380 cancel_work_sync(work);
381 debug_object_free(work, &work_debug_descr);
388 static struct debug_obj_descr work_debug_descr = {
389 .name = "work_struct",
390 .debug_hint = work_debug_hint,
391 .fixup_init = work_fixup_init,
392 .fixup_activate = work_fixup_activate,
393 .fixup_free = work_fixup_free,
396 static inline void debug_work_activate(struct work_struct *work)
398 debug_object_activate(work, &work_debug_descr);
401 static inline void debug_work_deactivate(struct work_struct *work)
403 debug_object_deactivate(work, &work_debug_descr);
406 void __init_work(struct work_struct *work, int onstack)
409 debug_object_init_on_stack(work, &work_debug_descr);
411 debug_object_init(work, &work_debug_descr);
413 EXPORT_SYMBOL_GPL(__init_work);
415 void destroy_work_on_stack(struct work_struct *work)
417 debug_object_free(work, &work_debug_descr);
419 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
422 static inline void debug_work_activate(struct work_struct *work) { }
423 static inline void debug_work_deactivate(struct work_struct *work) { }
426 /* Serializes the accesses to the list of workqueues. */
427 static DEFINE_SPINLOCK(workqueue_lock);
428 static LIST_HEAD(workqueues);
429 static bool workqueue_freezing; /* W: have wqs started freezing? */
432 * The CPU and unbound standard worker pools. The unbound ones have
433 * POOL_DISASSOCIATED set, and their workers have WORKER_UNBOUND set.
435 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
439 * idr of all pools. Modifications are protected by workqueue_lock. Read
440 * accesses are protected by sched-RCU protected.
442 static DEFINE_IDR(worker_pool_idr);
444 static int worker_thread(void *__worker);
446 /* allocate ID and assign it to @pool */
447 static int worker_pool_assign_id(struct worker_pool *pool)
452 if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL))
455 spin_lock_irq(&workqueue_lock);
456 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
457 spin_unlock_irq(&workqueue_lock);
458 } while (ret == -EAGAIN);
464 * first_pwq - return the first pool_workqueue of the specified workqueue
465 * @wq: the target workqueue
467 * This must be called either with workqueue_lock held or sched RCU read
468 * locked. If the pwq needs to be used beyond the locking in effect, the
469 * caller is responsible for guaranteeing that the pwq stays online.
471 static struct pool_workqueue *first_pwq(struct workqueue_struct *wq)
473 assert_rcu_or_wq_lock();
474 return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue,
478 static unsigned int work_color_to_flags(int color)
480 return color << WORK_STRUCT_COLOR_SHIFT;
483 static int get_work_color(struct work_struct *work)
485 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
486 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
489 static int work_next_color(int color)
491 return (color + 1) % WORK_NR_COLORS;
495 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
496 * contain the pointer to the queued pwq. Once execution starts, the flag
497 * is cleared and the high bits contain OFFQ flags and pool ID.
499 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
500 * and clear_work_data() can be used to set the pwq, pool or clear
501 * work->data. These functions should only be called while the work is
502 * owned - ie. while the PENDING bit is set.
504 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
505 * corresponding to a work. Pool is available once the work has been
506 * queued anywhere after initialization until it is sync canceled. pwq is
507 * available only while the work item is queued.
509 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
510 * canceled. While being canceled, a work item may have its PENDING set
511 * but stay off timer and worklist for arbitrarily long and nobody should
512 * try to steal the PENDING bit.
514 static inline void set_work_data(struct work_struct *work, unsigned long data,
517 WARN_ON_ONCE(!work_pending(work));
518 atomic_long_set(&work->data, data | flags | work_static(work));
521 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
522 unsigned long extra_flags)
524 set_work_data(work, (unsigned long)pwq,
525 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
528 static void set_work_pool_and_keep_pending(struct work_struct *work,
531 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
532 WORK_STRUCT_PENDING);
535 static void set_work_pool_and_clear_pending(struct work_struct *work,
539 * The following wmb is paired with the implied mb in
540 * test_and_set_bit(PENDING) and ensures all updates to @work made
541 * here are visible to and precede any updates by the next PENDING
545 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
548 static void clear_work_data(struct work_struct *work)
550 smp_wmb(); /* see set_work_pool_and_clear_pending() */
551 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
554 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
556 unsigned long data = atomic_long_read(&work->data);
558 if (data & WORK_STRUCT_PWQ)
559 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
565 * get_work_pool - return the worker_pool a given work was associated with
566 * @work: the work item of interest
568 * Return the worker_pool @work was last associated with. %NULL if none.
570 * Pools are created and destroyed under workqueue_lock, and allows read
571 * access under sched-RCU read lock. As such, this function should be
572 * called under workqueue_lock or with preemption disabled.
574 * All fields of the returned pool are accessible as long as the above
575 * mentioned locking is in effect. If the returned pool needs to be used
576 * beyond the critical section, the caller is responsible for ensuring the
577 * returned pool is and stays online.
579 static struct worker_pool *get_work_pool(struct work_struct *work)
581 unsigned long data = atomic_long_read(&work->data);
584 assert_rcu_or_wq_lock();
586 if (data & WORK_STRUCT_PWQ)
587 return ((struct pool_workqueue *)
588 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
590 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
591 if (pool_id == WORK_OFFQ_POOL_NONE)
594 return idr_find(&worker_pool_idr, pool_id);
598 * get_work_pool_id - return the worker pool ID a given work is associated with
599 * @work: the work item of interest
601 * Return the worker_pool ID @work was last associated with.
602 * %WORK_OFFQ_POOL_NONE if none.
604 static int get_work_pool_id(struct work_struct *work)
606 unsigned long data = atomic_long_read(&work->data);
608 if (data & WORK_STRUCT_PWQ)
609 return ((struct pool_workqueue *)
610 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
612 return data >> WORK_OFFQ_POOL_SHIFT;
615 static void mark_work_canceling(struct work_struct *work)
617 unsigned long pool_id = get_work_pool_id(work);
619 pool_id <<= WORK_OFFQ_POOL_SHIFT;
620 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
623 static bool work_is_canceling(struct work_struct *work)
625 unsigned long data = atomic_long_read(&work->data);
627 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
631 * Policy functions. These define the policies on how the global worker
632 * pools are managed. Unless noted otherwise, these functions assume that
633 * they're being called with pool->lock held.
636 static bool __need_more_worker(struct worker_pool *pool)
638 return !atomic_read(&pool->nr_running);
642 * Need to wake up a worker? Called from anything but currently
645 * Note that, because unbound workers never contribute to nr_running, this
646 * function will always return %true for unbound pools as long as the
647 * worklist isn't empty.
649 static bool need_more_worker(struct worker_pool *pool)
651 return !list_empty(&pool->worklist) && __need_more_worker(pool);
654 /* Can I start working? Called from busy but !running workers. */
655 static bool may_start_working(struct worker_pool *pool)
657 return pool->nr_idle;
660 /* Do I need to keep working? Called from currently running workers. */
661 static bool keep_working(struct worker_pool *pool)
663 return !list_empty(&pool->worklist) &&
664 atomic_read(&pool->nr_running) <= 1;
667 /* Do we need a new worker? Called from manager. */
668 static bool need_to_create_worker(struct worker_pool *pool)
670 return need_more_worker(pool) && !may_start_working(pool);
673 /* Do I need to be the manager? */
674 static bool need_to_manage_workers(struct worker_pool *pool)
676 return need_to_create_worker(pool) ||
677 (pool->flags & POOL_MANAGE_WORKERS);
680 /* Do we have too many workers and should some go away? */
681 static bool too_many_workers(struct worker_pool *pool)
683 bool managing = mutex_is_locked(&pool->manager_arb);
684 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
685 int nr_busy = pool->nr_workers - nr_idle;
688 * nr_idle and idle_list may disagree if idle rebinding is in
689 * progress. Never return %true if idle_list is empty.
691 if (list_empty(&pool->idle_list))
694 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
701 /* Return the first worker. Safe with preemption disabled */
702 static struct worker *first_worker(struct worker_pool *pool)
704 if (unlikely(list_empty(&pool->idle_list)))
707 return list_first_entry(&pool->idle_list, struct worker, entry);
711 * wake_up_worker - wake up an idle worker
712 * @pool: worker pool to wake worker from
714 * Wake up the first idle worker of @pool.
717 * spin_lock_irq(pool->lock).
719 static void wake_up_worker(struct worker_pool *pool)
721 struct worker *worker = first_worker(pool);
724 wake_up_process(worker->task);
728 * wq_worker_waking_up - a worker is waking up
729 * @task: task waking up
730 * @cpu: CPU @task is waking up to
732 * This function is called during try_to_wake_up() when a worker is
736 * spin_lock_irq(rq->lock)
738 void wq_worker_waking_up(struct task_struct *task, int cpu)
740 struct worker *worker = kthread_data(task);
742 if (!(worker->flags & WORKER_NOT_RUNNING)) {
743 WARN_ON_ONCE(worker->pool->cpu != cpu);
744 atomic_inc(&worker->pool->nr_running);
749 * wq_worker_sleeping - a worker is going to sleep
750 * @task: task going to sleep
751 * @cpu: CPU in question, must be the current CPU number
753 * This function is called during schedule() when a busy worker is
754 * going to sleep. Worker on the same cpu can be woken up by
755 * returning pointer to its task.
758 * spin_lock_irq(rq->lock)
761 * Worker task on @cpu to wake up, %NULL if none.
763 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
765 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
766 struct worker_pool *pool;
769 * Rescuers, which may not have all the fields set up like normal
770 * workers, also reach here, let's not access anything before
771 * checking NOT_RUNNING.
773 if (worker->flags & WORKER_NOT_RUNNING)
778 /* this can only happen on the local cpu */
779 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
783 * The counterpart of the following dec_and_test, implied mb,
784 * worklist not empty test sequence is in insert_work().
785 * Please read comment there.
787 * NOT_RUNNING is clear. This means that we're bound to and
788 * running on the local cpu w/ rq lock held and preemption
789 * disabled, which in turn means that none else could be
790 * manipulating idle_list, so dereferencing idle_list without pool
793 if (atomic_dec_and_test(&pool->nr_running) &&
794 !list_empty(&pool->worklist))
795 to_wakeup = first_worker(pool);
796 return to_wakeup ? to_wakeup->task : NULL;
800 * worker_set_flags - set worker flags and adjust nr_running accordingly
802 * @flags: flags to set
803 * @wakeup: wakeup an idle worker if necessary
805 * Set @flags in @worker->flags and adjust nr_running accordingly. If
806 * nr_running becomes zero and @wakeup is %true, an idle worker is
810 * spin_lock_irq(pool->lock)
812 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
815 struct worker_pool *pool = worker->pool;
817 WARN_ON_ONCE(worker->task != current);
820 * If transitioning into NOT_RUNNING, adjust nr_running and
821 * wake up an idle worker as necessary if requested by
824 if ((flags & WORKER_NOT_RUNNING) &&
825 !(worker->flags & WORKER_NOT_RUNNING)) {
827 if (atomic_dec_and_test(&pool->nr_running) &&
828 !list_empty(&pool->worklist))
829 wake_up_worker(pool);
831 atomic_dec(&pool->nr_running);
834 worker->flags |= flags;
838 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
840 * @flags: flags to clear
842 * Clear @flags in @worker->flags and adjust nr_running accordingly.
845 * spin_lock_irq(pool->lock)
847 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
849 struct worker_pool *pool = worker->pool;
850 unsigned int oflags = worker->flags;
852 WARN_ON_ONCE(worker->task != current);
854 worker->flags &= ~flags;
857 * If transitioning out of NOT_RUNNING, increment nr_running. Note
858 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
859 * of multiple flags, not a single flag.
861 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
862 if (!(worker->flags & WORKER_NOT_RUNNING))
863 atomic_inc(&pool->nr_running);
867 * find_worker_executing_work - find worker which is executing a work
868 * @pool: pool of interest
869 * @work: work to find worker for
871 * Find a worker which is executing @work on @pool by searching
872 * @pool->busy_hash which is keyed by the address of @work. For a worker
873 * to match, its current execution should match the address of @work and
874 * its work function. This is to avoid unwanted dependency between
875 * unrelated work executions through a work item being recycled while still
878 * This is a bit tricky. A work item may be freed once its execution
879 * starts and nothing prevents the freed area from being recycled for
880 * another work item. If the same work item address ends up being reused
881 * before the original execution finishes, workqueue will identify the
882 * recycled work item as currently executing and make it wait until the
883 * current execution finishes, introducing an unwanted dependency.
885 * This function checks the work item address, work function and workqueue
886 * to avoid false positives. Note that this isn't complete as one may
887 * construct a work function which can introduce dependency onto itself
888 * through a recycled work item. Well, if somebody wants to shoot oneself
889 * in the foot that badly, there's only so much we can do, and if such
890 * deadlock actually occurs, it should be easy to locate the culprit work
894 * spin_lock_irq(pool->lock).
897 * Pointer to worker which is executing @work if found, NULL
900 static struct worker *find_worker_executing_work(struct worker_pool *pool,
901 struct work_struct *work)
903 struct worker *worker;
905 hash_for_each_possible(pool->busy_hash, worker, hentry,
907 if (worker->current_work == work &&
908 worker->current_func == work->func)
915 * move_linked_works - move linked works to a list
916 * @work: start of series of works to be scheduled
917 * @head: target list to append @work to
918 * @nextp: out paramter for nested worklist walking
920 * Schedule linked works starting from @work to @head. Work series to
921 * be scheduled starts at @work and includes any consecutive work with
922 * WORK_STRUCT_LINKED set in its predecessor.
924 * If @nextp is not NULL, it's updated to point to the next work of
925 * the last scheduled work. This allows move_linked_works() to be
926 * nested inside outer list_for_each_entry_safe().
929 * spin_lock_irq(pool->lock).
931 static void move_linked_works(struct work_struct *work, struct list_head *head,
932 struct work_struct **nextp)
934 struct work_struct *n;
937 * Linked worklist will always end before the end of the list,
938 * use NULL for list head.
940 list_for_each_entry_safe_from(work, n, NULL, entry) {
941 list_move_tail(&work->entry, head);
942 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
947 * If we're already inside safe list traversal and have moved
948 * multiple works to the scheduled queue, the next position
949 * needs to be updated.
956 * get_pwq - get an extra reference on the specified pool_workqueue
957 * @pwq: pool_workqueue to get
959 * Obtain an extra reference on @pwq. The caller should guarantee that
960 * @pwq has positive refcnt and be holding the matching pool->lock.
962 static void get_pwq(struct pool_workqueue *pwq)
964 lockdep_assert_held(&pwq->pool->lock);
965 WARN_ON_ONCE(pwq->refcnt <= 0);
970 * put_pwq - put a pool_workqueue reference
971 * @pwq: pool_workqueue to put
973 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
974 * destruction. The caller should be holding the matching pool->lock.
976 static void put_pwq(struct pool_workqueue *pwq)
978 lockdep_assert_held(&pwq->pool->lock);
979 if (likely(--pwq->refcnt))
981 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
984 * @pwq can't be released under pool->lock, bounce to
985 * pwq_unbound_release_workfn(). This never recurses on the same
986 * pool->lock as this path is taken only for unbound workqueues and
987 * the release work item is scheduled on a per-cpu workqueue. To
988 * avoid lockdep warning, unbound pool->locks are given lockdep
989 * subclass of 1 in get_unbound_pool().
991 schedule_work(&pwq->unbound_release_work);
994 static void pwq_activate_delayed_work(struct work_struct *work)
996 struct pool_workqueue *pwq = get_work_pwq(work);
998 trace_workqueue_activate_work(work);
999 move_linked_works(work, &pwq->pool->worklist, NULL);
1000 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1004 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1006 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1007 struct work_struct, entry);
1009 pwq_activate_delayed_work(work);
1013 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1014 * @pwq: pwq of interest
1015 * @color: color of work which left the queue
1017 * A work either has completed or is removed from pending queue,
1018 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1021 * spin_lock_irq(pool->lock).
1023 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1025 /* uncolored work items don't participate in flushing or nr_active */
1026 if (color == WORK_NO_COLOR)
1029 pwq->nr_in_flight[color]--;
1032 if (!list_empty(&pwq->delayed_works)) {
1033 /* one down, submit a delayed one */
1034 if (pwq->nr_active < pwq->max_active)
1035 pwq_activate_first_delayed(pwq);
1038 /* is flush in progress and are we at the flushing tip? */
1039 if (likely(pwq->flush_color != color))
1042 /* are there still in-flight works? */
1043 if (pwq->nr_in_flight[color])
1046 /* this pwq is done, clear flush_color */
1047 pwq->flush_color = -1;
1050 * If this was the last pwq, wake up the first flusher. It
1051 * will handle the rest.
1053 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1054 complete(&pwq->wq->first_flusher->done);
1060 * try_to_grab_pending - steal work item from worklist and disable irq
1061 * @work: work item to steal
1062 * @is_dwork: @work is a delayed_work
1063 * @flags: place to store irq state
1065 * Try to grab PENDING bit of @work. This function can handle @work in any
1066 * stable state - idle, on timer or on worklist. Return values are
1068 * 1 if @work was pending and we successfully stole PENDING
1069 * 0 if @work was idle and we claimed PENDING
1070 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1071 * -ENOENT if someone else is canceling @work, this state may persist
1072 * for arbitrarily long
1074 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1075 * interrupted while holding PENDING and @work off queue, irq must be
1076 * disabled on entry. This, combined with delayed_work->timer being
1077 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1079 * On successful return, >= 0, irq is disabled and the caller is
1080 * responsible for releasing it using local_irq_restore(*@flags).
1082 * This function is safe to call from any context including IRQ handler.
1084 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1085 unsigned long *flags)
1087 struct worker_pool *pool;
1088 struct pool_workqueue *pwq;
1090 local_irq_save(*flags);
1092 /* try to steal the timer if it exists */
1094 struct delayed_work *dwork = to_delayed_work(work);
1097 * dwork->timer is irqsafe. If del_timer() fails, it's
1098 * guaranteed that the timer is not queued anywhere and not
1099 * running on the local CPU.
1101 if (likely(del_timer(&dwork->timer)))
1105 /* try to claim PENDING the normal way */
1106 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1110 * The queueing is in progress, or it is already queued. Try to
1111 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1113 pool = get_work_pool(work);
1117 spin_lock(&pool->lock);
1119 * work->data is guaranteed to point to pwq only while the work
1120 * item is queued on pwq->wq, and both updating work->data to point
1121 * to pwq on queueing and to pool on dequeueing are done under
1122 * pwq->pool->lock. This in turn guarantees that, if work->data
1123 * points to pwq which is associated with a locked pool, the work
1124 * item is currently queued on that pool.
1126 pwq = get_work_pwq(work);
1127 if (pwq && pwq->pool == pool) {
1128 debug_work_deactivate(work);
1131 * A delayed work item cannot be grabbed directly because
1132 * it might have linked NO_COLOR work items which, if left
1133 * on the delayed_list, will confuse pwq->nr_active
1134 * management later on and cause stall. Make sure the work
1135 * item is activated before grabbing.
1137 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1138 pwq_activate_delayed_work(work);
1140 list_del_init(&work->entry);
1141 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1143 /* work->data points to pwq iff queued, point to pool */
1144 set_work_pool_and_keep_pending(work, pool->id);
1146 spin_unlock(&pool->lock);
1149 spin_unlock(&pool->lock);
1151 local_irq_restore(*flags);
1152 if (work_is_canceling(work))
1159 * insert_work - insert a work into a pool
1160 * @pwq: pwq @work belongs to
1161 * @work: work to insert
1162 * @head: insertion point
1163 * @extra_flags: extra WORK_STRUCT_* flags to set
1165 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1166 * work_struct flags.
1169 * spin_lock_irq(pool->lock).
1171 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1172 struct list_head *head, unsigned int extra_flags)
1174 struct worker_pool *pool = pwq->pool;
1176 /* we own @work, set data and link */
1177 set_work_pwq(work, pwq, extra_flags);
1178 list_add_tail(&work->entry, head);
1182 * Ensure either worker_sched_deactivated() sees the above
1183 * list_add_tail() or we see zero nr_running to avoid workers
1184 * lying around lazily while there are works to be processed.
1188 if (__need_more_worker(pool))
1189 wake_up_worker(pool);
1193 * Test whether @work is being queued from another work executing on the
1196 static bool is_chained_work(struct workqueue_struct *wq)
1198 struct worker *worker;
1200 worker = current_wq_worker();
1202 * Return %true iff I'm a worker execuing a work item on @wq. If
1203 * I'm @worker, it's safe to dereference it without locking.
1205 return worker && worker->current_pwq->wq == wq;
1208 static void __queue_work(int cpu, struct workqueue_struct *wq,
1209 struct work_struct *work)
1211 struct pool_workqueue *pwq;
1212 struct list_head *worklist;
1213 unsigned int work_flags;
1214 unsigned int req_cpu = cpu;
1217 * While a work item is PENDING && off queue, a task trying to
1218 * steal the PENDING will busy-loop waiting for it to either get
1219 * queued or lose PENDING. Grabbing PENDING and queueing should
1220 * happen with IRQ disabled.
1222 WARN_ON_ONCE(!irqs_disabled());
1224 debug_work_activate(work);
1226 /* if dying, only works from the same workqueue are allowed */
1227 if (unlikely(wq->flags & WQ_DRAINING) &&
1228 WARN_ON_ONCE(!is_chained_work(wq)))
1231 /* determine the pwq to use */
1232 if (!(wq->flags & WQ_UNBOUND)) {
1233 struct worker_pool *last_pool;
1235 if (cpu == WORK_CPU_UNBOUND)
1236 cpu = raw_smp_processor_id();
1239 * It's multi cpu. If @work was previously on a different
1240 * cpu, it might still be running there, in which case the
1241 * work needs to be queued on that cpu to guarantee
1244 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1245 last_pool = get_work_pool(work);
1247 if (last_pool && last_pool != pwq->pool) {
1248 struct worker *worker;
1250 spin_lock(&last_pool->lock);
1252 worker = find_worker_executing_work(last_pool, work);
1254 if (worker && worker->current_pwq->wq == wq) {
1255 pwq = per_cpu_ptr(wq->cpu_pwqs, last_pool->cpu);
1257 /* meh... not running there, queue here */
1258 spin_unlock(&last_pool->lock);
1259 spin_lock(&pwq->pool->lock);
1262 spin_lock(&pwq->pool->lock);
1265 pwq = first_pwq(wq);
1266 spin_lock(&pwq->pool->lock);
1269 /* pwq determined, queue */
1270 trace_workqueue_queue_work(req_cpu, pwq, work);
1272 if (WARN_ON(!list_empty(&work->entry))) {
1273 spin_unlock(&pwq->pool->lock);
1277 pwq->nr_in_flight[pwq->work_color]++;
1278 work_flags = work_color_to_flags(pwq->work_color);
1280 if (likely(pwq->nr_active < pwq->max_active)) {
1281 trace_workqueue_activate_work(work);
1283 worklist = &pwq->pool->worklist;
1285 work_flags |= WORK_STRUCT_DELAYED;
1286 worklist = &pwq->delayed_works;
1289 insert_work(pwq, work, worklist, work_flags);
1291 spin_unlock(&pwq->pool->lock);
1295 * queue_work_on - queue work on specific cpu
1296 * @cpu: CPU number to execute work on
1297 * @wq: workqueue to use
1298 * @work: work to queue
1300 * Returns %false if @work was already on a queue, %true otherwise.
1302 * We queue the work to a specific CPU, the caller must ensure it
1305 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1306 struct work_struct *work)
1309 unsigned long flags;
1311 local_irq_save(flags);
1313 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1314 __queue_work(cpu, wq, work);
1318 local_irq_restore(flags);
1321 EXPORT_SYMBOL_GPL(queue_work_on);
1324 * queue_work - queue work on a workqueue
1325 * @wq: workqueue to use
1326 * @work: work to queue
1328 * Returns %false if @work was already on a queue, %true otherwise.
1330 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1331 * it can be processed by another CPU.
1333 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1335 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1337 EXPORT_SYMBOL_GPL(queue_work);
1339 void delayed_work_timer_fn(unsigned long __data)
1341 struct delayed_work *dwork = (struct delayed_work *)__data;
1343 /* should have been called from irqsafe timer with irq already off */
1344 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1346 EXPORT_SYMBOL(delayed_work_timer_fn);
1348 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1349 struct delayed_work *dwork, unsigned long delay)
1351 struct timer_list *timer = &dwork->timer;
1352 struct work_struct *work = &dwork->work;
1354 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1355 timer->data != (unsigned long)dwork);
1356 WARN_ON_ONCE(timer_pending(timer));
1357 WARN_ON_ONCE(!list_empty(&work->entry));
1360 * If @delay is 0, queue @dwork->work immediately. This is for
1361 * both optimization and correctness. The earliest @timer can
1362 * expire is on the closest next tick and delayed_work users depend
1363 * on that there's no such delay when @delay is 0.
1366 __queue_work(cpu, wq, &dwork->work);
1370 timer_stats_timer_set_start_info(&dwork->timer);
1374 timer->expires = jiffies + delay;
1376 if (unlikely(cpu != WORK_CPU_UNBOUND))
1377 add_timer_on(timer, cpu);
1383 * queue_delayed_work_on - queue work on specific CPU after delay
1384 * @cpu: CPU number to execute work on
1385 * @wq: workqueue to use
1386 * @dwork: work to queue
1387 * @delay: number of jiffies to wait before queueing
1389 * Returns %false if @work was already on a queue, %true otherwise. If
1390 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1393 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1394 struct delayed_work *dwork, unsigned long delay)
1396 struct work_struct *work = &dwork->work;
1398 unsigned long flags;
1400 /* read the comment in __queue_work() */
1401 local_irq_save(flags);
1403 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1404 __queue_delayed_work(cpu, wq, dwork, delay);
1408 local_irq_restore(flags);
1411 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1414 * queue_delayed_work - queue work on a workqueue after delay
1415 * @wq: workqueue to use
1416 * @dwork: delayable work to queue
1417 * @delay: number of jiffies to wait before queueing
1419 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1421 bool queue_delayed_work(struct workqueue_struct *wq,
1422 struct delayed_work *dwork, unsigned long delay)
1424 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1426 EXPORT_SYMBOL_GPL(queue_delayed_work);
1429 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1430 * @cpu: CPU number to execute work on
1431 * @wq: workqueue to use
1432 * @dwork: work to queue
1433 * @delay: number of jiffies to wait before queueing
1435 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1436 * modify @dwork's timer so that it expires after @delay. If @delay is
1437 * zero, @work is guaranteed to be scheduled immediately regardless of its
1440 * Returns %false if @dwork was idle and queued, %true if @dwork was
1441 * pending and its timer was modified.
1443 * This function is safe to call from any context including IRQ handler.
1444 * See try_to_grab_pending() for details.
1446 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1447 struct delayed_work *dwork, unsigned long delay)
1449 unsigned long flags;
1453 ret = try_to_grab_pending(&dwork->work, true, &flags);
1454 } while (unlikely(ret == -EAGAIN));
1456 if (likely(ret >= 0)) {
1457 __queue_delayed_work(cpu, wq, dwork, delay);
1458 local_irq_restore(flags);
1461 /* -ENOENT from try_to_grab_pending() becomes %true */
1464 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1467 * mod_delayed_work - modify delay of or queue a delayed work
1468 * @wq: workqueue to use
1469 * @dwork: work to queue
1470 * @delay: number of jiffies to wait before queueing
1472 * mod_delayed_work_on() on local CPU.
1474 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1475 unsigned long delay)
1477 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1479 EXPORT_SYMBOL_GPL(mod_delayed_work);
1482 * worker_enter_idle - enter idle state
1483 * @worker: worker which is entering idle state
1485 * @worker is entering idle state. Update stats and idle timer if
1489 * spin_lock_irq(pool->lock).
1491 static void worker_enter_idle(struct worker *worker)
1493 struct worker_pool *pool = worker->pool;
1495 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1496 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1497 (worker->hentry.next || worker->hentry.pprev)))
1500 /* can't use worker_set_flags(), also called from start_worker() */
1501 worker->flags |= WORKER_IDLE;
1503 worker->last_active = jiffies;
1505 /* idle_list is LIFO */
1506 list_add(&worker->entry, &pool->idle_list);
1508 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1509 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1512 * Sanity check nr_running. Because wq_unbind_fn() releases
1513 * pool->lock between setting %WORKER_UNBOUND and zapping
1514 * nr_running, the warning may trigger spuriously. Check iff
1515 * unbind is not in progress.
1517 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1518 pool->nr_workers == pool->nr_idle &&
1519 atomic_read(&pool->nr_running));
1523 * worker_leave_idle - leave idle state
1524 * @worker: worker which is leaving idle state
1526 * @worker is leaving idle state. Update stats.
1529 * spin_lock_irq(pool->lock).
1531 static void worker_leave_idle(struct worker *worker)
1533 struct worker_pool *pool = worker->pool;
1535 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1537 worker_clr_flags(worker, WORKER_IDLE);
1539 list_del_init(&worker->entry);
1543 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1544 * @pool: target worker_pool
1546 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1548 * Works which are scheduled while the cpu is online must at least be
1549 * scheduled to a worker which is bound to the cpu so that if they are
1550 * flushed from cpu callbacks while cpu is going down, they are
1551 * guaranteed to execute on the cpu.
1553 * This function is to be used by unbound workers and rescuers to bind
1554 * themselves to the target cpu and may race with cpu going down or
1555 * coming online. kthread_bind() can't be used because it may put the
1556 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1557 * verbatim as it's best effort and blocking and pool may be
1558 * [dis]associated in the meantime.
1560 * This function tries set_cpus_allowed() and locks pool and verifies the
1561 * binding against %POOL_DISASSOCIATED which is set during
1562 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1563 * enters idle state or fetches works without dropping lock, it can
1564 * guarantee the scheduling requirement described in the first paragraph.
1567 * Might sleep. Called without any lock but returns with pool->lock
1571 * %true if the associated pool is online (@worker is successfully
1572 * bound), %false if offline.
1574 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1575 __acquires(&pool->lock)
1579 * The following call may fail, succeed or succeed
1580 * without actually migrating the task to the cpu if
1581 * it races with cpu hotunplug operation. Verify
1582 * against POOL_DISASSOCIATED.
1584 if (!(pool->flags & POOL_DISASSOCIATED))
1585 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1587 spin_lock_irq(&pool->lock);
1588 if (pool->flags & POOL_DISASSOCIATED)
1590 if (task_cpu(current) == pool->cpu &&
1591 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1593 spin_unlock_irq(&pool->lock);
1596 * We've raced with CPU hot[un]plug. Give it a breather
1597 * and retry migration. cond_resched() is required here;
1598 * otherwise, we might deadlock against cpu_stop trying to
1599 * bring down the CPU on non-preemptive kernel.
1607 * Rebind an idle @worker to its CPU. worker_thread() will test
1608 * list_empty(@worker->entry) before leaving idle and call this function.
1610 static void idle_worker_rebind(struct worker *worker)
1612 /* CPU may go down again inbetween, clear UNBOUND only on success */
1613 if (worker_maybe_bind_and_lock(worker->pool))
1614 worker_clr_flags(worker, WORKER_UNBOUND);
1616 /* rebind complete, become available again */
1617 list_add(&worker->entry, &worker->pool->idle_list);
1618 spin_unlock_irq(&worker->pool->lock);
1622 * Function for @worker->rebind.work used to rebind unbound busy workers to
1623 * the associated cpu which is coming back online. This is scheduled by
1624 * cpu up but can race with other cpu hotplug operations and may be
1625 * executed twice without intervening cpu down.
1627 static void busy_worker_rebind_fn(struct work_struct *work)
1629 struct worker *worker = container_of(work, struct worker, rebind_work);
1631 if (worker_maybe_bind_and_lock(worker->pool))
1632 worker_clr_flags(worker, WORKER_UNBOUND);
1634 spin_unlock_irq(&worker->pool->lock);
1638 * rebind_workers - rebind all workers of a pool to the associated CPU
1639 * @pool: pool of interest
1641 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1642 * is different for idle and busy ones.
1644 * Idle ones will be removed from the idle_list and woken up. They will
1645 * add themselves back after completing rebind. This ensures that the
1646 * idle_list doesn't contain any unbound workers when re-bound busy workers
1647 * try to perform local wake-ups for concurrency management.
1649 * Busy workers can rebind after they finish their current work items.
1650 * Queueing the rebind work item at the head of the scheduled list is
1651 * enough. Note that nr_running will be properly bumped as busy workers
1654 * On return, all non-manager workers are scheduled for rebind - see
1655 * manage_workers() for the manager special case. Any idle worker
1656 * including the manager will not appear on @idle_list until rebind is
1657 * complete, making local wake-ups safe.
1659 static void rebind_workers(struct worker_pool *pool)
1661 struct worker *worker, *n;
1664 lockdep_assert_held(&pool->assoc_mutex);
1665 lockdep_assert_held(&pool->lock);
1667 /* dequeue and kick idle ones */
1668 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1670 * idle workers should be off @pool->idle_list until rebind
1671 * is complete to avoid receiving premature local wake-ups.
1673 list_del_init(&worker->entry);
1676 * worker_thread() will see the above dequeuing and call
1677 * idle_worker_rebind().
1679 wake_up_process(worker->task);
1682 /* rebind busy workers */
1683 for_each_busy_worker(worker, i, pool) {
1684 struct work_struct *rebind_work = &worker->rebind_work;
1685 struct workqueue_struct *wq;
1687 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1688 work_data_bits(rebind_work)))
1691 debug_work_activate(rebind_work);
1694 * wq doesn't really matter but let's keep @worker->pool
1695 * and @pwq->pool consistent for sanity.
1697 if (worker->pool->attrs->nice < 0)
1698 wq = system_highpri_wq;
1702 insert_work(per_cpu_ptr(wq->cpu_pwqs, pool->cpu), rebind_work,
1703 worker->scheduled.next,
1704 work_color_to_flags(WORK_NO_COLOR));
1708 static struct worker *alloc_worker(void)
1710 struct worker *worker;
1712 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1714 INIT_LIST_HEAD(&worker->entry);
1715 INIT_LIST_HEAD(&worker->scheduled);
1716 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1717 /* on creation a worker is in !idle && prep state */
1718 worker->flags = WORKER_PREP;
1724 * create_worker - create a new workqueue worker
1725 * @pool: pool the new worker will belong to
1727 * Create a new worker which is bound to @pool. The returned worker
1728 * can be started by calling start_worker() or destroyed using
1732 * Might sleep. Does GFP_KERNEL allocations.
1735 * Pointer to the newly created worker.
1737 static struct worker *create_worker(struct worker_pool *pool)
1739 const char *pri = pool->attrs->nice < 0 ? "H" : "";
1740 struct worker *worker = NULL;
1743 spin_lock_irq(&pool->lock);
1744 while (ida_get_new(&pool->worker_ida, &id)) {
1745 spin_unlock_irq(&pool->lock);
1746 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1748 spin_lock_irq(&pool->lock);
1750 spin_unlock_irq(&pool->lock);
1752 worker = alloc_worker();
1756 worker->pool = pool;
1760 worker->task = kthread_create_on_node(worker_thread,
1761 worker, cpu_to_node(pool->cpu),
1762 "kworker/%d:%d%s", pool->cpu, id, pri);
1764 worker->task = kthread_create(worker_thread, worker,
1767 if (IS_ERR(worker->task))
1770 set_user_nice(worker->task, pool->attrs->nice);
1771 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1774 * %PF_THREAD_BOUND is used to prevent userland from meddling with
1775 * cpumask of workqueue workers. This is an abuse. We need
1776 * %PF_NO_SETAFFINITY.
1778 worker->task->flags |= PF_THREAD_BOUND;
1781 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1782 * remains stable across this function. See the comments above the
1783 * flag definition for details.
1785 if (pool->flags & POOL_DISASSOCIATED)
1786 worker->flags |= WORKER_UNBOUND;
1791 spin_lock_irq(&pool->lock);
1792 ida_remove(&pool->worker_ida, id);
1793 spin_unlock_irq(&pool->lock);
1800 * start_worker - start a newly created worker
1801 * @worker: worker to start
1803 * Make the pool aware of @worker and start it.
1806 * spin_lock_irq(pool->lock).
1808 static void start_worker(struct worker *worker)
1810 worker->flags |= WORKER_STARTED;
1811 worker->pool->nr_workers++;
1812 worker_enter_idle(worker);
1813 wake_up_process(worker->task);
1817 * destroy_worker - destroy a workqueue worker
1818 * @worker: worker to be destroyed
1820 * Destroy @worker and adjust @pool stats accordingly.
1823 * spin_lock_irq(pool->lock) which is released and regrabbed.
1825 static void destroy_worker(struct worker *worker)
1827 struct worker_pool *pool = worker->pool;
1828 int id = worker->id;
1830 /* sanity check frenzy */
1831 if (WARN_ON(worker->current_work) ||
1832 WARN_ON(!list_empty(&worker->scheduled)))
1835 if (worker->flags & WORKER_STARTED)
1837 if (worker->flags & WORKER_IDLE)
1840 list_del_init(&worker->entry);
1841 worker->flags |= WORKER_DIE;
1843 spin_unlock_irq(&pool->lock);
1845 kthread_stop(worker->task);
1848 spin_lock_irq(&pool->lock);
1849 ida_remove(&pool->worker_ida, id);
1852 static void idle_worker_timeout(unsigned long __pool)
1854 struct worker_pool *pool = (void *)__pool;
1856 spin_lock_irq(&pool->lock);
1858 if (too_many_workers(pool)) {
1859 struct worker *worker;
1860 unsigned long expires;
1862 /* idle_list is kept in LIFO order, check the last one */
1863 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1864 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1866 if (time_before(jiffies, expires))
1867 mod_timer(&pool->idle_timer, expires);
1869 /* it's been idle for too long, wake up manager */
1870 pool->flags |= POOL_MANAGE_WORKERS;
1871 wake_up_worker(pool);
1875 spin_unlock_irq(&pool->lock);
1878 static void send_mayday(struct work_struct *work)
1880 struct pool_workqueue *pwq = get_work_pwq(work);
1881 struct workqueue_struct *wq = pwq->wq;
1883 lockdep_assert_held(&workqueue_lock);
1888 /* mayday mayday mayday */
1889 if (list_empty(&pwq->mayday_node)) {
1890 list_add_tail(&pwq->mayday_node, &wq->maydays);
1891 wake_up_process(wq->rescuer->task);
1895 static void pool_mayday_timeout(unsigned long __pool)
1897 struct worker_pool *pool = (void *)__pool;
1898 struct work_struct *work;
1900 spin_lock_irq(&workqueue_lock); /* for wq->maydays */
1901 spin_lock(&pool->lock);
1903 if (need_to_create_worker(pool)) {
1905 * We've been trying to create a new worker but
1906 * haven't been successful. We might be hitting an
1907 * allocation deadlock. Send distress signals to
1910 list_for_each_entry(work, &pool->worklist, entry)
1914 spin_unlock(&pool->lock);
1915 spin_unlock_irq(&workqueue_lock);
1917 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1921 * maybe_create_worker - create a new worker if necessary
1922 * @pool: pool to create a new worker for
1924 * Create a new worker for @pool if necessary. @pool is guaranteed to
1925 * have at least one idle worker on return from this function. If
1926 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1927 * sent to all rescuers with works scheduled on @pool to resolve
1928 * possible allocation deadlock.
1930 * On return, need_to_create_worker() is guaranteed to be false and
1931 * may_start_working() true.
1934 * spin_lock_irq(pool->lock) which may be released and regrabbed
1935 * multiple times. Does GFP_KERNEL allocations. Called only from
1939 * false if no action was taken and pool->lock stayed locked, true
1942 static bool maybe_create_worker(struct worker_pool *pool)
1943 __releases(&pool->lock)
1944 __acquires(&pool->lock)
1946 if (!need_to_create_worker(pool))
1949 spin_unlock_irq(&pool->lock);
1951 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1952 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1955 struct worker *worker;
1957 worker = create_worker(pool);
1959 del_timer_sync(&pool->mayday_timer);
1960 spin_lock_irq(&pool->lock);
1961 start_worker(worker);
1962 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1967 if (!need_to_create_worker(pool))
1970 __set_current_state(TASK_INTERRUPTIBLE);
1971 schedule_timeout(CREATE_COOLDOWN);
1973 if (!need_to_create_worker(pool))
1977 del_timer_sync(&pool->mayday_timer);
1978 spin_lock_irq(&pool->lock);
1979 if (need_to_create_worker(pool))
1985 * maybe_destroy_worker - destroy workers which have been idle for a while
1986 * @pool: pool to destroy workers for
1988 * Destroy @pool workers which have been idle for longer than
1989 * IDLE_WORKER_TIMEOUT.
1992 * spin_lock_irq(pool->lock) which may be released and regrabbed
1993 * multiple times. Called only from manager.
1996 * false if no action was taken and pool->lock stayed locked, true
1999 static bool maybe_destroy_workers(struct worker_pool *pool)
2003 while (too_many_workers(pool)) {
2004 struct worker *worker;
2005 unsigned long expires;
2007 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2008 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2010 if (time_before(jiffies, expires)) {
2011 mod_timer(&pool->idle_timer, expires);
2015 destroy_worker(worker);
2023 * manage_workers - manage worker pool
2026 * Assume the manager role and manage the worker pool @worker belongs
2027 * to. At any given time, there can be only zero or one manager per
2028 * pool. The exclusion is handled automatically by this function.
2030 * The caller can safely start processing works on false return. On
2031 * true return, it's guaranteed that need_to_create_worker() is false
2032 * and may_start_working() is true.
2035 * spin_lock_irq(pool->lock) which may be released and regrabbed
2036 * multiple times. Does GFP_KERNEL allocations.
2039 * spin_lock_irq(pool->lock) which may be released and regrabbed
2040 * multiple times. Does GFP_KERNEL allocations.
2042 static bool manage_workers(struct worker *worker)
2044 struct worker_pool *pool = worker->pool;
2047 if (!mutex_trylock(&pool->manager_arb))
2051 * To simplify both worker management and CPU hotplug, hold off
2052 * management while hotplug is in progress. CPU hotplug path can't
2053 * grab @pool->manager_arb to achieve this because that can lead to
2054 * idle worker depletion (all become busy thinking someone else is
2055 * managing) which in turn can result in deadlock under extreme
2056 * circumstances. Use @pool->assoc_mutex to synchronize manager
2057 * against CPU hotplug.
2059 * assoc_mutex would always be free unless CPU hotplug is in
2060 * progress. trylock first without dropping @pool->lock.
2062 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2063 spin_unlock_irq(&pool->lock);
2064 mutex_lock(&pool->assoc_mutex);
2066 * CPU hotplug could have happened while we were waiting
2067 * for assoc_mutex. Hotplug itself can't handle us
2068 * because manager isn't either on idle or busy list, and
2069 * @pool's state and ours could have deviated.
2071 * As hotplug is now excluded via assoc_mutex, we can
2072 * simply try to bind. It will succeed or fail depending
2073 * on @pool's current state. Try it and adjust
2074 * %WORKER_UNBOUND accordingly.
2076 if (worker_maybe_bind_and_lock(pool))
2077 worker->flags &= ~WORKER_UNBOUND;
2079 worker->flags |= WORKER_UNBOUND;
2084 pool->flags &= ~POOL_MANAGE_WORKERS;
2087 * Destroy and then create so that may_start_working() is true
2090 ret |= maybe_destroy_workers(pool);
2091 ret |= maybe_create_worker(pool);
2093 mutex_unlock(&pool->assoc_mutex);
2094 mutex_unlock(&pool->manager_arb);
2099 * process_one_work - process single work
2101 * @work: work to process
2103 * Process @work. This function contains all the logics necessary to
2104 * process a single work including synchronization against and
2105 * interaction with other workers on the same cpu, queueing and
2106 * flushing. As long as context requirement is met, any worker can
2107 * call this function to process a work.
2110 * spin_lock_irq(pool->lock) which is released and regrabbed.
2112 static void process_one_work(struct worker *worker, struct work_struct *work)
2113 __releases(&pool->lock)
2114 __acquires(&pool->lock)
2116 struct pool_workqueue *pwq = get_work_pwq(work);
2117 struct worker_pool *pool = worker->pool;
2118 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2120 struct worker *collision;
2121 #ifdef CONFIG_LOCKDEP
2123 * It is permissible to free the struct work_struct from
2124 * inside the function that is called from it, this we need to
2125 * take into account for lockdep too. To avoid bogus "held
2126 * lock freed" warnings as well as problems when looking into
2127 * work->lockdep_map, make a copy and use that here.
2129 struct lockdep_map lockdep_map;
2131 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2134 * Ensure we're on the correct CPU. DISASSOCIATED test is
2135 * necessary to avoid spurious warnings from rescuers servicing the
2136 * unbound or a disassociated pool.
2138 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2139 !(pool->flags & POOL_DISASSOCIATED) &&
2140 raw_smp_processor_id() != pool->cpu);
2143 * A single work shouldn't be executed concurrently by
2144 * multiple workers on a single cpu. Check whether anyone is
2145 * already processing the work. If so, defer the work to the
2146 * currently executing one.
2148 collision = find_worker_executing_work(pool, work);
2149 if (unlikely(collision)) {
2150 move_linked_works(work, &collision->scheduled, NULL);
2154 /* claim and dequeue */
2155 debug_work_deactivate(work);
2156 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2157 worker->current_work = work;
2158 worker->current_func = work->func;
2159 worker->current_pwq = pwq;
2160 work_color = get_work_color(work);
2162 list_del_init(&work->entry);
2165 * CPU intensive works don't participate in concurrency
2166 * management. They're the scheduler's responsibility.
2168 if (unlikely(cpu_intensive))
2169 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2172 * Unbound pool isn't concurrency managed and work items should be
2173 * executed ASAP. Wake up another worker if necessary.
2175 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2176 wake_up_worker(pool);
2179 * Record the last pool and clear PENDING which should be the last
2180 * update to @work. Also, do this inside @pool->lock so that
2181 * PENDING and queued state changes happen together while IRQ is
2184 set_work_pool_and_clear_pending(work, pool->id);
2186 spin_unlock_irq(&pool->lock);
2188 lock_map_acquire_read(&pwq->wq->lockdep_map);
2189 lock_map_acquire(&lockdep_map);
2190 trace_workqueue_execute_start(work);
2191 worker->current_func(work);
2193 * While we must be careful to not use "work" after this, the trace
2194 * point will only record its address.
2196 trace_workqueue_execute_end(work);
2197 lock_map_release(&lockdep_map);
2198 lock_map_release(&pwq->wq->lockdep_map);
2200 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2201 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2202 " last function: %pf\n",
2203 current->comm, preempt_count(), task_pid_nr(current),
2204 worker->current_func);
2205 debug_show_held_locks(current);
2209 spin_lock_irq(&pool->lock);
2211 /* clear cpu intensive status */
2212 if (unlikely(cpu_intensive))
2213 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2215 /* we're done with it, release */
2216 hash_del(&worker->hentry);
2217 worker->current_work = NULL;
2218 worker->current_func = NULL;
2219 worker->current_pwq = NULL;
2220 pwq_dec_nr_in_flight(pwq, work_color);
2224 * process_scheduled_works - process scheduled works
2227 * Process all scheduled works. Please note that the scheduled list
2228 * may change while processing a work, so this function repeatedly
2229 * fetches a work from the top and executes it.
2232 * spin_lock_irq(pool->lock) which may be released and regrabbed
2235 static void process_scheduled_works(struct worker *worker)
2237 while (!list_empty(&worker->scheduled)) {
2238 struct work_struct *work = list_first_entry(&worker->scheduled,
2239 struct work_struct, entry);
2240 process_one_work(worker, work);
2245 * worker_thread - the worker thread function
2248 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2249 * of these per each cpu. These workers process all works regardless of
2250 * their specific target workqueue. The only exception is works which
2251 * belong to workqueues with a rescuer which will be explained in
2254 static int worker_thread(void *__worker)
2256 struct worker *worker = __worker;
2257 struct worker_pool *pool = worker->pool;
2259 /* tell the scheduler that this is a workqueue worker */
2260 worker->task->flags |= PF_WQ_WORKER;
2262 spin_lock_irq(&pool->lock);
2264 /* we are off idle list if destruction or rebind is requested */
2265 if (unlikely(list_empty(&worker->entry))) {
2266 spin_unlock_irq(&pool->lock);
2268 /* if DIE is set, destruction is requested */
2269 if (worker->flags & WORKER_DIE) {
2270 worker->task->flags &= ~PF_WQ_WORKER;
2274 /* otherwise, rebind */
2275 idle_worker_rebind(worker);
2279 worker_leave_idle(worker);
2281 /* no more worker necessary? */
2282 if (!need_more_worker(pool))
2285 /* do we need to manage? */
2286 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2290 * ->scheduled list can only be filled while a worker is
2291 * preparing to process a work or actually processing it.
2292 * Make sure nobody diddled with it while I was sleeping.
2294 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2297 * When control reaches this point, we're guaranteed to have
2298 * at least one idle worker or that someone else has already
2299 * assumed the manager role.
2301 worker_clr_flags(worker, WORKER_PREP);
2304 struct work_struct *work =
2305 list_first_entry(&pool->worklist,
2306 struct work_struct, entry);
2308 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2309 /* optimization path, not strictly necessary */
2310 process_one_work(worker, work);
2311 if (unlikely(!list_empty(&worker->scheduled)))
2312 process_scheduled_works(worker);
2314 move_linked_works(work, &worker->scheduled, NULL);
2315 process_scheduled_works(worker);
2317 } while (keep_working(pool));
2319 worker_set_flags(worker, WORKER_PREP, false);
2321 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2325 * pool->lock is held and there's no work to process and no need to
2326 * manage, sleep. Workers are woken up only while holding
2327 * pool->lock or from local cpu, so setting the current state
2328 * before releasing pool->lock is enough to prevent losing any
2331 worker_enter_idle(worker);
2332 __set_current_state(TASK_INTERRUPTIBLE);
2333 spin_unlock_irq(&pool->lock);
2339 * rescuer_thread - the rescuer thread function
2342 * Workqueue rescuer thread function. There's one rescuer for each
2343 * workqueue which has WQ_MEM_RECLAIM set.
2345 * Regular work processing on a pool may block trying to create a new
2346 * worker which uses GFP_KERNEL allocation which has slight chance of
2347 * developing into deadlock if some works currently on the same queue
2348 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2349 * the problem rescuer solves.
2351 * When such condition is possible, the pool summons rescuers of all
2352 * workqueues which have works queued on the pool and let them process
2353 * those works so that forward progress can be guaranteed.
2355 * This should happen rarely.
2357 static int rescuer_thread(void *__rescuer)
2359 struct worker *rescuer = __rescuer;
2360 struct workqueue_struct *wq = rescuer->rescue_wq;
2361 struct list_head *scheduled = &rescuer->scheduled;
2363 set_user_nice(current, RESCUER_NICE_LEVEL);
2366 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2367 * doesn't participate in concurrency management.
2369 rescuer->task->flags |= PF_WQ_WORKER;
2371 set_current_state(TASK_INTERRUPTIBLE);
2373 if (kthread_should_stop()) {
2374 __set_current_state(TASK_RUNNING);
2375 rescuer->task->flags &= ~PF_WQ_WORKER;
2379 /* see whether any pwq is asking for help */
2380 spin_lock_irq(&workqueue_lock);
2382 while (!list_empty(&wq->maydays)) {
2383 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2384 struct pool_workqueue, mayday_node);
2385 struct worker_pool *pool = pwq->pool;
2386 struct work_struct *work, *n;
2388 __set_current_state(TASK_RUNNING);
2389 list_del_init(&pwq->mayday_node);
2391 spin_unlock_irq(&workqueue_lock);
2393 /* migrate to the target cpu if possible */
2394 worker_maybe_bind_and_lock(pool);
2395 rescuer->pool = pool;
2398 * Slurp in all works issued via this workqueue and
2401 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2402 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2403 if (get_work_pwq(work) == pwq)
2404 move_linked_works(work, scheduled, &n);
2406 process_scheduled_works(rescuer);
2409 * Leave this pool. If keep_working() is %true, notify a
2410 * regular worker; otherwise, we end up with 0 concurrency
2411 * and stalling the execution.
2413 if (keep_working(pool))
2414 wake_up_worker(pool);
2416 rescuer->pool = NULL;
2417 spin_unlock(&pool->lock);
2418 spin_lock(&workqueue_lock);
2421 spin_unlock_irq(&workqueue_lock);
2423 /* rescuers should never participate in concurrency management */
2424 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2430 struct work_struct work;
2431 struct completion done;
2434 static void wq_barrier_func(struct work_struct *work)
2436 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2437 complete(&barr->done);
2441 * insert_wq_barrier - insert a barrier work
2442 * @pwq: pwq to insert barrier into
2443 * @barr: wq_barrier to insert
2444 * @target: target work to attach @barr to
2445 * @worker: worker currently executing @target, NULL if @target is not executing
2447 * @barr is linked to @target such that @barr is completed only after
2448 * @target finishes execution. Please note that the ordering
2449 * guarantee is observed only with respect to @target and on the local
2452 * Currently, a queued barrier can't be canceled. This is because
2453 * try_to_grab_pending() can't determine whether the work to be
2454 * grabbed is at the head of the queue and thus can't clear LINKED
2455 * flag of the previous work while there must be a valid next work
2456 * after a work with LINKED flag set.
2458 * Note that when @worker is non-NULL, @target may be modified
2459 * underneath us, so we can't reliably determine pwq from @target.
2462 * spin_lock_irq(pool->lock).
2464 static void insert_wq_barrier(struct pool_workqueue *pwq,
2465 struct wq_barrier *barr,
2466 struct work_struct *target, struct worker *worker)
2468 struct list_head *head;
2469 unsigned int linked = 0;
2472 * debugobject calls are safe here even with pool->lock locked
2473 * as we know for sure that this will not trigger any of the
2474 * checks and call back into the fixup functions where we
2477 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2478 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2479 init_completion(&barr->done);
2482 * If @target is currently being executed, schedule the
2483 * barrier to the worker; otherwise, put it after @target.
2486 head = worker->scheduled.next;
2488 unsigned long *bits = work_data_bits(target);
2490 head = target->entry.next;
2491 /* there can already be other linked works, inherit and set */
2492 linked = *bits & WORK_STRUCT_LINKED;
2493 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2496 debug_work_activate(&barr->work);
2497 insert_work(pwq, &barr->work, head,
2498 work_color_to_flags(WORK_NO_COLOR) | linked);
2502 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2503 * @wq: workqueue being flushed
2504 * @flush_color: new flush color, < 0 for no-op
2505 * @work_color: new work color, < 0 for no-op
2507 * Prepare pwqs for workqueue flushing.
2509 * If @flush_color is non-negative, flush_color on all pwqs should be
2510 * -1. If no pwq has in-flight commands at the specified color, all
2511 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2512 * has in flight commands, its pwq->flush_color is set to
2513 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2514 * wakeup logic is armed and %true is returned.
2516 * The caller should have initialized @wq->first_flusher prior to
2517 * calling this function with non-negative @flush_color. If
2518 * @flush_color is negative, no flush color update is done and %false
2521 * If @work_color is non-negative, all pwqs should have the same
2522 * work_color which is previous to @work_color and all will be
2523 * advanced to @work_color.
2526 * mutex_lock(wq->flush_mutex).
2529 * %true if @flush_color >= 0 and there's something to flush. %false
2532 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2533 int flush_color, int work_color)
2536 struct pool_workqueue *pwq;
2538 if (flush_color >= 0) {
2539 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2540 atomic_set(&wq->nr_pwqs_to_flush, 1);
2543 local_irq_disable();
2545 for_each_pwq(pwq, wq) {
2546 struct worker_pool *pool = pwq->pool;
2548 spin_lock(&pool->lock);
2550 if (flush_color >= 0) {
2551 WARN_ON_ONCE(pwq->flush_color != -1);
2553 if (pwq->nr_in_flight[flush_color]) {
2554 pwq->flush_color = flush_color;
2555 atomic_inc(&wq->nr_pwqs_to_flush);
2560 if (work_color >= 0) {
2561 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2562 pwq->work_color = work_color;
2565 spin_unlock(&pool->lock);
2570 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2571 complete(&wq->first_flusher->done);
2577 * flush_workqueue - ensure that any scheduled work has run to completion.
2578 * @wq: workqueue to flush
2580 * Forces execution of the workqueue and blocks until its completion.
2581 * This is typically used in driver shutdown handlers.
2583 * We sleep until all works which were queued on entry have been handled,
2584 * but we are not livelocked by new incoming ones.
2586 void flush_workqueue(struct workqueue_struct *wq)
2588 struct wq_flusher this_flusher = {
2589 .list = LIST_HEAD_INIT(this_flusher.list),
2591 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2595 lock_map_acquire(&wq->lockdep_map);
2596 lock_map_release(&wq->lockdep_map);
2598 mutex_lock(&wq->flush_mutex);
2601 * Start-to-wait phase
2603 next_color = work_next_color(wq->work_color);
2605 if (next_color != wq->flush_color) {
2607 * Color space is not full. The current work_color
2608 * becomes our flush_color and work_color is advanced
2611 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2612 this_flusher.flush_color = wq->work_color;
2613 wq->work_color = next_color;
2615 if (!wq->first_flusher) {
2616 /* no flush in progress, become the first flusher */
2617 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2619 wq->first_flusher = &this_flusher;
2621 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2623 /* nothing to flush, done */
2624 wq->flush_color = next_color;
2625 wq->first_flusher = NULL;
2630 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2631 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2632 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2636 * Oops, color space is full, wait on overflow queue.
2637 * The next flush completion will assign us
2638 * flush_color and transfer to flusher_queue.
2640 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2643 mutex_unlock(&wq->flush_mutex);
2645 wait_for_completion(&this_flusher.done);
2648 * Wake-up-and-cascade phase
2650 * First flushers are responsible for cascading flushes and
2651 * handling overflow. Non-first flushers can simply return.
2653 if (wq->first_flusher != &this_flusher)
2656 mutex_lock(&wq->flush_mutex);
2658 /* we might have raced, check again with mutex held */
2659 if (wq->first_flusher != &this_flusher)
2662 wq->first_flusher = NULL;
2664 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2665 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2668 struct wq_flusher *next, *tmp;
2670 /* complete all the flushers sharing the current flush color */
2671 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2672 if (next->flush_color != wq->flush_color)
2674 list_del_init(&next->list);
2675 complete(&next->done);
2678 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2679 wq->flush_color != work_next_color(wq->work_color));
2681 /* this flush_color is finished, advance by one */
2682 wq->flush_color = work_next_color(wq->flush_color);
2684 /* one color has been freed, handle overflow queue */
2685 if (!list_empty(&wq->flusher_overflow)) {
2687 * Assign the same color to all overflowed
2688 * flushers, advance work_color and append to
2689 * flusher_queue. This is the start-to-wait
2690 * phase for these overflowed flushers.
2692 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2693 tmp->flush_color = wq->work_color;
2695 wq->work_color = work_next_color(wq->work_color);
2697 list_splice_tail_init(&wq->flusher_overflow,
2698 &wq->flusher_queue);
2699 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2702 if (list_empty(&wq->flusher_queue)) {
2703 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2708 * Need to flush more colors. Make the next flusher
2709 * the new first flusher and arm pwqs.
2711 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2712 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2714 list_del_init(&next->list);
2715 wq->first_flusher = next;
2717 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2721 * Meh... this color is already done, clear first
2722 * flusher and repeat cascading.
2724 wq->first_flusher = NULL;
2728 mutex_unlock(&wq->flush_mutex);
2730 EXPORT_SYMBOL_GPL(flush_workqueue);
2733 * drain_workqueue - drain a workqueue
2734 * @wq: workqueue to drain
2736 * Wait until the workqueue becomes empty. While draining is in progress,
2737 * only chain queueing is allowed. IOW, only currently pending or running
2738 * work items on @wq can queue further work items on it. @wq is flushed
2739 * repeatedly until it becomes empty. The number of flushing is detemined
2740 * by the depth of chaining and should be relatively short. Whine if it
2743 void drain_workqueue(struct workqueue_struct *wq)
2745 unsigned int flush_cnt = 0;
2746 struct pool_workqueue *pwq;
2749 * __queue_work() needs to test whether there are drainers, is much
2750 * hotter than drain_workqueue() and already looks at @wq->flags.
2751 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2753 spin_lock_irq(&workqueue_lock);
2754 if (!wq->nr_drainers++)
2755 wq->flags |= WQ_DRAINING;
2756 spin_unlock_irq(&workqueue_lock);
2758 flush_workqueue(wq);
2760 local_irq_disable();
2762 for_each_pwq(pwq, wq) {
2765 spin_lock(&pwq->pool->lock);
2766 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2767 spin_unlock(&pwq->pool->lock);
2772 if (++flush_cnt == 10 ||
2773 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2774 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2775 wq->name, flush_cnt);
2781 spin_lock(&workqueue_lock);
2782 if (!--wq->nr_drainers)
2783 wq->flags &= ~WQ_DRAINING;
2784 spin_unlock(&workqueue_lock);
2788 EXPORT_SYMBOL_GPL(drain_workqueue);
2790 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2792 struct worker *worker = NULL;
2793 struct worker_pool *pool;
2794 struct pool_workqueue *pwq;
2798 local_irq_disable();
2799 pool = get_work_pool(work);
2805 spin_lock(&pool->lock);
2806 /* see the comment in try_to_grab_pending() with the same code */
2807 pwq = get_work_pwq(work);
2809 if (unlikely(pwq->pool != pool))
2812 worker = find_worker_executing_work(pool, work);
2815 pwq = worker->current_pwq;
2818 insert_wq_barrier(pwq, barr, work, worker);
2819 spin_unlock_irq(&pool->lock);
2822 * If @max_active is 1 or rescuer is in use, flushing another work
2823 * item on the same workqueue may lead to deadlock. Make sure the
2824 * flusher is not running on the same workqueue by verifying write
2827 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2828 lock_map_acquire(&pwq->wq->lockdep_map);
2830 lock_map_acquire_read(&pwq->wq->lockdep_map);
2831 lock_map_release(&pwq->wq->lockdep_map);
2835 spin_unlock_irq(&pool->lock);
2840 * flush_work - wait for a work to finish executing the last queueing instance
2841 * @work: the work to flush
2843 * Wait until @work has finished execution. @work is guaranteed to be idle
2844 * on return if it hasn't been requeued since flush started.
2847 * %true if flush_work() waited for the work to finish execution,
2848 * %false if it was already idle.
2850 bool flush_work(struct work_struct *work)
2852 struct wq_barrier barr;
2854 lock_map_acquire(&work->lockdep_map);
2855 lock_map_release(&work->lockdep_map);
2857 if (start_flush_work(work, &barr)) {
2858 wait_for_completion(&barr.done);
2859 destroy_work_on_stack(&barr.work);
2865 EXPORT_SYMBOL_GPL(flush_work);
2867 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2869 unsigned long flags;
2873 ret = try_to_grab_pending(work, is_dwork, &flags);
2875 * If someone else is canceling, wait for the same event it
2876 * would be waiting for before retrying.
2878 if (unlikely(ret == -ENOENT))
2880 } while (unlikely(ret < 0));
2882 /* tell other tasks trying to grab @work to back off */
2883 mark_work_canceling(work);
2884 local_irq_restore(flags);
2887 clear_work_data(work);
2892 * cancel_work_sync - cancel a work and wait for it to finish
2893 * @work: the work to cancel
2895 * Cancel @work and wait for its execution to finish. This function
2896 * can be used even if the work re-queues itself or migrates to
2897 * another workqueue. On return from this function, @work is
2898 * guaranteed to be not pending or executing on any CPU.
2900 * cancel_work_sync(&delayed_work->work) must not be used for
2901 * delayed_work's. Use cancel_delayed_work_sync() instead.
2903 * The caller must ensure that the workqueue on which @work was last
2904 * queued can't be destroyed before this function returns.
2907 * %true if @work was pending, %false otherwise.
2909 bool cancel_work_sync(struct work_struct *work)
2911 return __cancel_work_timer(work, false);
2913 EXPORT_SYMBOL_GPL(cancel_work_sync);
2916 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2917 * @dwork: the delayed work to flush
2919 * Delayed timer is cancelled and the pending work is queued for
2920 * immediate execution. Like flush_work(), this function only
2921 * considers the last queueing instance of @dwork.
2924 * %true if flush_work() waited for the work to finish execution,
2925 * %false if it was already idle.
2927 bool flush_delayed_work(struct delayed_work *dwork)
2929 local_irq_disable();
2930 if (del_timer_sync(&dwork->timer))
2931 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2933 return flush_work(&dwork->work);
2935 EXPORT_SYMBOL(flush_delayed_work);
2938 * cancel_delayed_work - cancel a delayed work
2939 * @dwork: delayed_work to cancel
2941 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2942 * and canceled; %false if wasn't pending. Note that the work callback
2943 * function may still be running on return, unless it returns %true and the
2944 * work doesn't re-arm itself. Explicitly flush or use
2945 * cancel_delayed_work_sync() to wait on it.
2947 * This function is safe to call from any context including IRQ handler.
2949 bool cancel_delayed_work(struct delayed_work *dwork)
2951 unsigned long flags;
2955 ret = try_to_grab_pending(&dwork->work, true, &flags);
2956 } while (unlikely(ret == -EAGAIN));
2958 if (unlikely(ret < 0))
2961 set_work_pool_and_clear_pending(&dwork->work,
2962 get_work_pool_id(&dwork->work));
2963 local_irq_restore(flags);
2966 EXPORT_SYMBOL(cancel_delayed_work);
2969 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2970 * @dwork: the delayed work cancel
2972 * This is cancel_work_sync() for delayed works.
2975 * %true if @dwork was pending, %false otherwise.
2977 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2979 return __cancel_work_timer(&dwork->work, true);
2981 EXPORT_SYMBOL(cancel_delayed_work_sync);
2984 * schedule_work_on - put work task on a specific cpu
2985 * @cpu: cpu to put the work task on
2986 * @work: job to be done
2988 * This puts a job on a specific cpu
2990 bool schedule_work_on(int cpu, struct work_struct *work)
2992 return queue_work_on(cpu, system_wq, work);
2994 EXPORT_SYMBOL(schedule_work_on);
2997 * schedule_work - put work task in global workqueue
2998 * @work: job to be done
3000 * Returns %false if @work was already on the kernel-global workqueue and
3003 * This puts a job in the kernel-global workqueue if it was not already
3004 * queued and leaves it in the same position on the kernel-global
3005 * workqueue otherwise.
3007 bool schedule_work(struct work_struct *work)
3009 return queue_work(system_wq, work);
3011 EXPORT_SYMBOL(schedule_work);
3014 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3016 * @dwork: job to be done
3017 * @delay: number of jiffies to wait
3019 * After waiting for a given time this puts a job in the kernel-global
3020 * workqueue on the specified CPU.
3022 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3023 unsigned long delay)
3025 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3027 EXPORT_SYMBOL(schedule_delayed_work_on);
3030 * schedule_delayed_work - put work task in global workqueue after delay
3031 * @dwork: job to be done
3032 * @delay: number of jiffies to wait or 0 for immediate execution
3034 * After waiting for a given time this puts a job in the kernel-global
3037 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3039 return queue_delayed_work(system_wq, dwork, delay);
3041 EXPORT_SYMBOL(schedule_delayed_work);
3044 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3045 * @func: the function to call
3047 * schedule_on_each_cpu() executes @func on each online CPU using the
3048 * system workqueue and blocks until all CPUs have completed.
3049 * schedule_on_each_cpu() is very slow.
3052 * 0 on success, -errno on failure.
3054 int schedule_on_each_cpu(work_func_t func)
3057 struct work_struct __percpu *works;
3059 works = alloc_percpu(struct work_struct);
3065 for_each_online_cpu(cpu) {
3066 struct work_struct *work = per_cpu_ptr(works, cpu);
3068 INIT_WORK(work, func);
3069 schedule_work_on(cpu, work);
3072 for_each_online_cpu(cpu)
3073 flush_work(per_cpu_ptr(works, cpu));
3081 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3083 * Forces execution of the kernel-global workqueue and blocks until its
3086 * Think twice before calling this function! It's very easy to get into
3087 * trouble if you don't take great care. Either of the following situations
3088 * will lead to deadlock:
3090 * One of the work items currently on the workqueue needs to acquire
3091 * a lock held by your code or its caller.
3093 * Your code is running in the context of a work routine.
3095 * They will be detected by lockdep when they occur, but the first might not
3096 * occur very often. It depends on what work items are on the workqueue and
3097 * what locks they need, which you have no control over.
3099 * In most situations flushing the entire workqueue is overkill; you merely
3100 * need to know that a particular work item isn't queued and isn't running.
3101 * In such cases you should use cancel_delayed_work_sync() or
3102 * cancel_work_sync() instead.
3104 void flush_scheduled_work(void)
3106 flush_workqueue(system_wq);
3108 EXPORT_SYMBOL(flush_scheduled_work);
3111 * execute_in_process_context - reliably execute the routine with user context
3112 * @fn: the function to execute
3113 * @ew: guaranteed storage for the execute work structure (must
3114 * be available when the work executes)
3116 * Executes the function immediately if process context is available,
3117 * otherwise schedules the function for delayed execution.
3119 * Returns: 0 - function was executed
3120 * 1 - function was scheduled for execution
3122 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3124 if (!in_interrupt()) {
3129 INIT_WORK(&ew->work, fn);
3130 schedule_work(&ew->work);
3134 EXPORT_SYMBOL_GPL(execute_in_process_context);
3136 int keventd_up(void)
3138 return system_wq != NULL;
3142 * free_workqueue_attrs - free a workqueue_attrs
3143 * @attrs: workqueue_attrs to free
3145 * Undo alloc_workqueue_attrs().
3147 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3150 free_cpumask_var(attrs->cpumask);
3156 * alloc_workqueue_attrs - allocate a workqueue_attrs
3157 * @gfp_mask: allocation mask to use
3159 * Allocate a new workqueue_attrs, initialize with default settings and
3160 * return it. Returns NULL on failure.
3162 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3164 struct workqueue_attrs *attrs;
3166 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3169 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3172 cpumask_setall(attrs->cpumask);
3175 free_workqueue_attrs(attrs);
3179 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3180 const struct workqueue_attrs *from)
3182 to->nice = from->nice;
3183 cpumask_copy(to->cpumask, from->cpumask);
3187 * Hacky implementation of jhash of bitmaps which only considers the
3188 * specified number of bits. We probably want a proper implementation in
3189 * include/linux/jhash.h.
3191 static u32 jhash_bitmap(const unsigned long *bitmap, int bits, u32 hash)
3193 int nr_longs = bits / BITS_PER_LONG;
3194 int nr_leftover = bits % BITS_PER_LONG;
3195 unsigned long leftover = 0;
3198 hash = jhash(bitmap, nr_longs * sizeof(long), hash);
3200 bitmap_copy(&leftover, bitmap + nr_longs, nr_leftover);
3201 hash = jhash(&leftover, sizeof(long), hash);
3206 /* hash value of the content of @attr */
3207 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3211 hash = jhash_1word(attrs->nice, hash);
3212 hash = jhash_bitmap(cpumask_bits(attrs->cpumask), nr_cpu_ids, hash);
3216 /* content equality test */
3217 static bool wqattrs_equal(const struct workqueue_attrs *a,
3218 const struct workqueue_attrs *b)
3220 if (a->nice != b->nice)
3222 if (!cpumask_equal(a->cpumask, b->cpumask))
3228 * init_worker_pool - initialize a newly zalloc'd worker_pool
3229 * @pool: worker_pool to initialize
3231 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3232 * Returns 0 on success, -errno on failure. Even on failure, all fields
3233 * inside @pool proper are initialized and put_unbound_pool() can be called
3234 * on @pool safely to release it.
3236 static int init_worker_pool(struct worker_pool *pool)
3238 spin_lock_init(&pool->lock);
3241 pool->flags |= POOL_DISASSOCIATED;
3242 INIT_LIST_HEAD(&pool->worklist);
3243 INIT_LIST_HEAD(&pool->idle_list);
3244 hash_init(pool->busy_hash);
3246 init_timer_deferrable(&pool->idle_timer);
3247 pool->idle_timer.function = idle_worker_timeout;
3248 pool->idle_timer.data = (unsigned long)pool;
3250 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3251 (unsigned long)pool);
3253 mutex_init(&pool->manager_arb);
3254 mutex_init(&pool->assoc_mutex);
3255 ida_init(&pool->worker_ida);
3257 INIT_HLIST_NODE(&pool->hash_node);
3260 /* shouldn't fail above this point */
3261 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3267 static void rcu_free_pool(struct rcu_head *rcu)
3269 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3271 ida_destroy(&pool->worker_ida);
3272 free_workqueue_attrs(pool->attrs);
3277 * put_unbound_pool - put a worker_pool
3278 * @pool: worker_pool to put
3280 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3283 static void put_unbound_pool(struct worker_pool *pool)
3285 struct worker *worker;
3287 spin_lock_irq(&workqueue_lock);
3288 if (--pool->refcnt) {
3289 spin_unlock_irq(&workqueue_lock);
3294 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3295 WARN_ON(!list_empty(&pool->worklist))) {
3296 spin_unlock_irq(&workqueue_lock);
3300 /* release id and unhash */
3302 idr_remove(&worker_pool_idr, pool->id);
3303 hash_del(&pool->hash_node);
3305 spin_unlock_irq(&workqueue_lock);
3307 /* lock out manager and destroy all workers */
3308 mutex_lock(&pool->manager_arb);
3309 spin_lock_irq(&pool->lock);
3311 while ((worker = first_worker(pool)))
3312 destroy_worker(worker);
3313 WARN_ON(pool->nr_workers || pool->nr_idle);
3315 spin_unlock_irq(&pool->lock);
3316 mutex_unlock(&pool->manager_arb);
3318 /* shut down the timers */
3319 del_timer_sync(&pool->idle_timer);
3320 del_timer_sync(&pool->mayday_timer);
3322 /* sched-RCU protected to allow dereferences from get_work_pool() */
3323 call_rcu_sched(&pool->rcu, rcu_free_pool);
3327 * get_unbound_pool - get a worker_pool with the specified attributes
3328 * @attrs: the attributes of the worker_pool to get
3330 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3331 * reference count and return it. If there already is a matching
3332 * worker_pool, it will be used; otherwise, this function attempts to
3333 * create a new one. On failure, returns NULL.
3335 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3337 static DEFINE_MUTEX(create_mutex);
3338 u32 hash = wqattrs_hash(attrs);
3339 struct worker_pool *pool;
3340 struct worker *worker;
3342 mutex_lock(&create_mutex);
3344 /* do we already have a matching pool? */
3345 spin_lock_irq(&workqueue_lock);
3346 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3347 if (wqattrs_equal(pool->attrs, attrs)) {
3352 spin_unlock_irq(&workqueue_lock);
3354 /* nope, create a new one */
3355 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3356 if (!pool || init_worker_pool(pool) < 0)
3359 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3360 copy_workqueue_attrs(pool->attrs, attrs);
3362 if (worker_pool_assign_id(pool) < 0)
3365 /* create and start the initial worker */
3366 worker = create_worker(pool);
3370 spin_lock_irq(&pool->lock);
3371 start_worker(worker);
3372 spin_unlock_irq(&pool->lock);
3375 spin_lock_irq(&workqueue_lock);
3376 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3378 spin_unlock_irq(&workqueue_lock);
3379 mutex_unlock(&create_mutex);
3382 mutex_unlock(&create_mutex);
3384 put_unbound_pool(pool);
3388 static void rcu_free_pwq(struct rcu_head *rcu)
3390 kmem_cache_free(pwq_cache,
3391 container_of(rcu, struct pool_workqueue, rcu));
3395 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3396 * and needs to be destroyed.
3398 static void pwq_unbound_release_workfn(struct work_struct *work)
3400 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3401 unbound_release_work);
3402 struct workqueue_struct *wq = pwq->wq;
3403 struct worker_pool *pool = pwq->pool;
3405 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3409 * Unlink @pwq. Synchronization against flush_mutex isn't strictly
3410 * necessary on release but do it anyway. It's easier to verify
3411 * and consistent with the linking path.
3413 mutex_lock(&wq->flush_mutex);
3414 spin_lock_irq(&workqueue_lock);
3415 list_del_rcu(&pwq->pwqs_node);
3416 spin_unlock_irq(&workqueue_lock);
3417 mutex_unlock(&wq->flush_mutex);
3419 put_unbound_pool(pool);
3420 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3423 * If we're the last pwq going away, @wq is already dead and no one
3424 * is gonna access it anymore. Free it.
3426 if (list_empty(&wq->pwqs))
3430 static void init_and_link_pwq(struct pool_workqueue *pwq,
3431 struct workqueue_struct *wq,
3432 struct worker_pool *pool)
3434 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3438 pwq->flush_color = -1;
3440 pwq->max_active = wq->saved_max_active;
3441 INIT_LIST_HEAD(&pwq->delayed_works);
3442 INIT_LIST_HEAD(&pwq->mayday_node);
3443 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3446 * Link @pwq and set the matching work_color. This is synchronized
3447 * with flush_mutex to avoid confusing flush_workqueue().
3449 mutex_lock(&wq->flush_mutex);
3450 spin_lock_irq(&workqueue_lock);
3452 pwq->work_color = wq->work_color;
3453 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs);
3455 spin_unlock_irq(&workqueue_lock);
3456 mutex_unlock(&wq->flush_mutex);
3459 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3461 bool highpri = wq->flags & WQ_HIGHPRI;
3464 if (!(wq->flags & WQ_UNBOUND)) {
3465 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3469 for_each_possible_cpu(cpu) {
3470 struct pool_workqueue *pwq =
3471 per_cpu_ptr(wq->cpu_pwqs, cpu);
3472 struct worker_pool *cpu_pools =
3473 per_cpu(cpu_worker_pools, cpu);
3475 init_and_link_pwq(pwq, wq, &cpu_pools[highpri]);
3478 struct pool_workqueue *pwq;
3479 struct worker_pool *pool;
3481 pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL);
3485 pool = get_unbound_pool(unbound_std_wq_attrs[highpri]);
3487 kmem_cache_free(pwq_cache, pwq);
3491 init_and_link_pwq(pwq, wq, pool);
3497 static int wq_clamp_max_active(int max_active, unsigned int flags,
3500 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3502 if (max_active < 1 || max_active > lim)
3503 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3504 max_active, name, 1, lim);
3506 return clamp_val(max_active, 1, lim);
3509 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3512 struct lock_class_key *key,
3513 const char *lock_name, ...)
3515 va_list args, args1;
3516 struct workqueue_struct *wq;
3517 struct pool_workqueue *pwq;
3520 /* determine namelen, allocate wq and format name */
3521 va_start(args, lock_name);
3522 va_copy(args1, args);
3523 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3525 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3529 vsnprintf(wq->name, namelen, fmt, args1);
3533 max_active = max_active ?: WQ_DFL_ACTIVE;
3534 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3538 wq->saved_max_active = max_active;
3539 mutex_init(&wq->flush_mutex);
3540 atomic_set(&wq->nr_pwqs_to_flush, 0);
3541 INIT_LIST_HEAD(&wq->pwqs);
3542 INIT_LIST_HEAD(&wq->flusher_queue);
3543 INIT_LIST_HEAD(&wq->flusher_overflow);
3544 INIT_LIST_HEAD(&wq->maydays);
3546 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3547 INIT_LIST_HEAD(&wq->list);
3549 if (alloc_and_link_pwqs(wq) < 0)
3553 * Workqueues which may be used during memory reclaim should
3554 * have a rescuer to guarantee forward progress.
3556 if (flags & WQ_MEM_RECLAIM) {
3557 struct worker *rescuer;
3559 rescuer = alloc_worker();
3563 rescuer->rescue_wq = wq;
3564 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3566 if (IS_ERR(rescuer->task)) {
3571 wq->rescuer = rescuer;
3572 rescuer->task->flags |= PF_THREAD_BOUND;
3573 wake_up_process(rescuer->task);
3577 * workqueue_lock protects global freeze state and workqueues
3578 * list. Grab it, set max_active accordingly and add the new
3579 * workqueue to workqueues list.
3581 spin_lock_irq(&workqueue_lock);
3583 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3584 for_each_pwq(pwq, wq)
3585 pwq->max_active = 0;
3587 list_add(&wq->list, &workqueues);
3589 spin_unlock_irq(&workqueue_lock);
3597 destroy_workqueue(wq);
3600 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3603 * destroy_workqueue - safely terminate a workqueue
3604 * @wq: target workqueue
3606 * Safely destroy a workqueue. All work currently pending will be done first.
3608 void destroy_workqueue(struct workqueue_struct *wq)
3610 struct pool_workqueue *pwq;
3612 /* drain it before proceeding with destruction */
3613 drain_workqueue(wq);
3615 spin_lock_irq(&workqueue_lock);
3618 for_each_pwq(pwq, wq) {
3621 for (i = 0; i < WORK_NR_COLORS; i++) {
3622 if (WARN_ON(pwq->nr_in_flight[i])) {
3623 spin_unlock_irq(&workqueue_lock);
3628 if (WARN_ON(pwq->refcnt > 1) ||
3629 WARN_ON(pwq->nr_active) ||
3630 WARN_ON(!list_empty(&pwq->delayed_works))) {
3631 spin_unlock_irq(&workqueue_lock);
3637 * wq list is used to freeze wq, remove from list after
3638 * flushing is complete in case freeze races us.
3640 list_del_init(&wq->list);
3642 spin_unlock_irq(&workqueue_lock);
3645 kthread_stop(wq->rescuer->task);
3650 if (!(wq->flags & WQ_UNBOUND)) {
3652 * The base ref is never dropped on per-cpu pwqs. Directly
3653 * free the pwqs and wq.
3655 free_percpu(wq->cpu_pwqs);
3659 * We're the sole accessor of @wq at this point. Directly
3660 * access the first pwq and put the base ref. As both pwqs
3661 * and pools are sched-RCU protected, the lock operations
3662 * are safe. @wq will be freed when the last pwq is
3665 pwq = list_first_entry(&wq->pwqs, struct pool_workqueue,
3667 spin_lock_irq(&pwq->pool->lock);
3669 spin_unlock_irq(&pwq->pool->lock);
3672 EXPORT_SYMBOL_GPL(destroy_workqueue);
3675 * pwq_set_max_active - adjust max_active of a pwq
3676 * @pwq: target pool_workqueue
3677 * @max_active: new max_active value.
3679 * Set @pwq->max_active to @max_active and activate delayed works if
3683 * spin_lock_irq(pool->lock).
3685 static void pwq_set_max_active(struct pool_workqueue *pwq, int max_active)
3687 pwq->max_active = max_active;
3689 while (!list_empty(&pwq->delayed_works) &&
3690 pwq->nr_active < pwq->max_active)
3691 pwq_activate_first_delayed(pwq);
3695 * workqueue_set_max_active - adjust max_active of a workqueue
3696 * @wq: target workqueue
3697 * @max_active: new max_active value.
3699 * Set max_active of @wq to @max_active.
3702 * Don't call from IRQ context.
3704 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3706 struct pool_workqueue *pwq;
3708 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3710 spin_lock_irq(&workqueue_lock);
3712 wq->saved_max_active = max_active;
3714 for_each_pwq(pwq, wq) {
3715 struct worker_pool *pool = pwq->pool;
3717 spin_lock(&pool->lock);
3719 if (!(wq->flags & WQ_FREEZABLE) ||
3720 !(pool->flags & POOL_FREEZING))
3721 pwq_set_max_active(pwq, max_active);
3723 spin_unlock(&pool->lock);
3726 spin_unlock_irq(&workqueue_lock);
3728 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3731 * workqueue_congested - test whether a workqueue is congested
3732 * @cpu: CPU in question
3733 * @wq: target workqueue
3735 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3736 * no synchronization around this function and the test result is
3737 * unreliable and only useful as advisory hints or for debugging.
3740 * %true if congested, %false otherwise.
3742 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
3744 struct pool_workqueue *pwq;
3749 if (!(wq->flags & WQ_UNBOUND))
3750 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
3752 pwq = first_pwq(wq);
3754 ret = !list_empty(&pwq->delayed_works);
3759 EXPORT_SYMBOL_GPL(workqueue_congested);
3762 * work_busy - test whether a work is currently pending or running
3763 * @work: the work to be tested
3765 * Test whether @work is currently pending or running. There is no
3766 * synchronization around this function and the test result is
3767 * unreliable and only useful as advisory hints or for debugging.
3770 * OR'd bitmask of WORK_BUSY_* bits.
3772 unsigned int work_busy(struct work_struct *work)
3774 struct worker_pool *pool;
3775 unsigned long flags;
3776 unsigned int ret = 0;
3778 if (work_pending(work))
3779 ret |= WORK_BUSY_PENDING;
3781 local_irq_save(flags);
3782 pool = get_work_pool(work);
3784 spin_lock(&pool->lock);
3785 if (find_worker_executing_work(pool, work))
3786 ret |= WORK_BUSY_RUNNING;
3787 spin_unlock(&pool->lock);
3789 local_irq_restore(flags);
3793 EXPORT_SYMBOL_GPL(work_busy);
3798 * There are two challenges in supporting CPU hotplug. Firstly, there
3799 * are a lot of assumptions on strong associations among work, pwq and
3800 * pool which make migrating pending and scheduled works very
3801 * difficult to implement without impacting hot paths. Secondly,
3802 * worker pools serve mix of short, long and very long running works making
3803 * blocked draining impractical.
3805 * This is solved by allowing the pools to be disassociated from the CPU
3806 * running as an unbound one and allowing it to be reattached later if the
3807 * cpu comes back online.
3810 static void wq_unbind_fn(struct work_struct *work)
3812 int cpu = smp_processor_id();
3813 struct worker_pool *pool;
3814 struct worker *worker;
3817 for_each_cpu_worker_pool(pool, cpu) {
3818 WARN_ON_ONCE(cpu != smp_processor_id());
3820 mutex_lock(&pool->assoc_mutex);
3821 spin_lock_irq(&pool->lock);
3824 * We've claimed all manager positions. Make all workers
3825 * unbound and set DISASSOCIATED. Before this, all workers
3826 * except for the ones which are still executing works from
3827 * before the last CPU down must be on the cpu. After
3828 * this, they may become diasporas.
3830 list_for_each_entry(worker, &pool->idle_list, entry)
3831 worker->flags |= WORKER_UNBOUND;
3833 for_each_busy_worker(worker, i, pool)
3834 worker->flags |= WORKER_UNBOUND;
3836 pool->flags |= POOL_DISASSOCIATED;
3838 spin_unlock_irq(&pool->lock);
3839 mutex_unlock(&pool->assoc_mutex);
3843 * Call schedule() so that we cross rq->lock and thus can guarantee
3844 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3845 * as scheduler callbacks may be invoked from other cpus.
3850 * Sched callbacks are disabled now. Zap nr_running. After this,
3851 * nr_running stays zero and need_more_worker() and keep_working()
3852 * are always true as long as the worklist is not empty. Pools on
3853 * @cpu now behave as unbound (in terms of concurrency management)
3854 * pools which are served by workers tied to the CPU.
3856 * On return from this function, the current worker would trigger
3857 * unbound chain execution of pending work items if other workers
3860 for_each_cpu_worker_pool(pool, cpu)
3861 atomic_set(&pool->nr_running, 0);
3865 * Workqueues should be brought up before normal priority CPU notifiers.
3866 * This will be registered high priority CPU notifier.
3868 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3869 unsigned long action,
3872 int cpu = (unsigned long)hcpu;
3873 struct worker_pool *pool;
3875 switch (action & ~CPU_TASKS_FROZEN) {
3876 case CPU_UP_PREPARE:
3877 for_each_cpu_worker_pool(pool, cpu) {
3878 struct worker *worker;
3880 if (pool->nr_workers)
3883 worker = create_worker(pool);
3887 spin_lock_irq(&pool->lock);
3888 start_worker(worker);
3889 spin_unlock_irq(&pool->lock);
3893 case CPU_DOWN_FAILED:
3895 for_each_cpu_worker_pool(pool, cpu) {
3896 mutex_lock(&pool->assoc_mutex);
3897 spin_lock_irq(&pool->lock);
3899 pool->flags &= ~POOL_DISASSOCIATED;
3900 rebind_workers(pool);
3902 spin_unlock_irq(&pool->lock);
3903 mutex_unlock(&pool->assoc_mutex);
3911 * Workqueues should be brought down after normal priority CPU notifiers.
3912 * This will be registered as low priority CPU notifier.
3914 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3915 unsigned long action,
3918 int cpu = (unsigned long)hcpu;
3919 struct work_struct unbind_work;
3921 switch (action & ~CPU_TASKS_FROZEN) {
3922 case CPU_DOWN_PREPARE:
3923 /* unbinding should happen on the local CPU */
3924 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
3925 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3926 flush_work(&unbind_work);
3934 struct work_for_cpu {
3935 struct work_struct work;
3941 static void work_for_cpu_fn(struct work_struct *work)
3943 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3945 wfc->ret = wfc->fn(wfc->arg);
3949 * work_on_cpu - run a function in user context on a particular cpu
3950 * @cpu: the cpu to run on
3951 * @fn: the function to run
3952 * @arg: the function arg
3954 * This will return the value @fn returns.
3955 * It is up to the caller to ensure that the cpu doesn't go offline.
3956 * The caller must not hold any locks which would prevent @fn from completing.
3958 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
3960 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3962 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3963 schedule_work_on(cpu, &wfc.work);
3964 flush_work(&wfc.work);
3967 EXPORT_SYMBOL_GPL(work_on_cpu);
3968 #endif /* CONFIG_SMP */
3970 #ifdef CONFIG_FREEZER
3973 * freeze_workqueues_begin - begin freezing workqueues
3975 * Start freezing workqueues. After this function returns, all freezable
3976 * workqueues will queue new works to their frozen_works list instead of
3980 * Grabs and releases workqueue_lock and pool->lock's.
3982 void freeze_workqueues_begin(void)
3984 struct worker_pool *pool;
3985 struct workqueue_struct *wq;
3986 struct pool_workqueue *pwq;
3989 spin_lock_irq(&workqueue_lock);
3991 WARN_ON_ONCE(workqueue_freezing);
3992 workqueue_freezing = true;
3995 for_each_pool(pool, id) {
3996 spin_lock(&pool->lock);
3997 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
3998 pool->flags |= POOL_FREEZING;
3999 spin_unlock(&pool->lock);
4002 /* suppress further executions by setting max_active to zero */
4003 list_for_each_entry(wq, &workqueues, list) {
4004 if (!(wq->flags & WQ_FREEZABLE))
4007 for_each_pwq(pwq, wq) {
4008 spin_lock(&pwq->pool->lock);
4009 pwq->max_active = 0;
4010 spin_unlock(&pwq->pool->lock);
4014 spin_unlock_irq(&workqueue_lock);
4018 * freeze_workqueues_busy - are freezable workqueues still busy?
4020 * Check whether freezing is complete. This function must be called
4021 * between freeze_workqueues_begin() and thaw_workqueues().
4024 * Grabs and releases workqueue_lock.
4027 * %true if some freezable workqueues are still busy. %false if freezing
4030 bool freeze_workqueues_busy(void)
4033 struct workqueue_struct *wq;
4034 struct pool_workqueue *pwq;
4036 spin_lock_irq(&workqueue_lock);
4038 WARN_ON_ONCE(!workqueue_freezing);
4040 list_for_each_entry(wq, &workqueues, list) {
4041 if (!(wq->flags & WQ_FREEZABLE))
4044 * nr_active is monotonically decreasing. It's safe
4045 * to peek without lock.
4047 for_each_pwq(pwq, wq) {
4048 WARN_ON_ONCE(pwq->nr_active < 0);
4049 if (pwq->nr_active) {
4056 spin_unlock_irq(&workqueue_lock);
4061 * thaw_workqueues - thaw workqueues
4063 * Thaw workqueues. Normal queueing is restored and all collected
4064 * frozen works are transferred to their respective pool worklists.
4067 * Grabs and releases workqueue_lock and pool->lock's.
4069 void thaw_workqueues(void)
4071 struct workqueue_struct *wq;
4072 struct pool_workqueue *pwq;
4073 struct worker_pool *pool;
4076 spin_lock_irq(&workqueue_lock);
4078 if (!workqueue_freezing)
4081 /* clear FREEZING */
4082 for_each_pool(pool, id) {
4083 spin_lock(&pool->lock);
4084 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4085 pool->flags &= ~POOL_FREEZING;
4086 spin_unlock(&pool->lock);
4089 /* restore max_active and repopulate worklist */
4090 list_for_each_entry(wq, &workqueues, list) {
4091 if (!(wq->flags & WQ_FREEZABLE))
4094 for_each_pwq(pwq, wq) {
4095 spin_lock(&pwq->pool->lock);
4096 pwq_set_max_active(pwq, wq->saved_max_active);
4097 spin_unlock(&pwq->pool->lock);
4102 for_each_pool(pool, id) {
4103 spin_lock(&pool->lock);
4104 wake_up_worker(pool);
4105 spin_unlock(&pool->lock);
4108 workqueue_freezing = false;
4110 spin_unlock_irq(&workqueue_lock);
4112 #endif /* CONFIG_FREEZER */
4114 static int __init init_workqueues(void)
4116 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4119 /* make sure we have enough bits for OFFQ pool ID */
4120 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4121 WORK_CPU_END * NR_STD_WORKER_POOLS);
4123 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4125 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4127 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4128 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4130 /* initialize CPU pools */
4131 for_each_possible_cpu(cpu) {
4132 struct worker_pool *pool;
4135 for_each_cpu_worker_pool(pool, cpu) {
4136 BUG_ON(init_worker_pool(pool));
4138 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4139 pool->attrs->nice = std_nice[i++];
4142 BUG_ON(worker_pool_assign_id(pool));
4146 /* create the initial worker */
4147 for_each_online_cpu(cpu) {
4148 struct worker_pool *pool;
4150 for_each_cpu_worker_pool(pool, cpu) {
4151 struct worker *worker;
4153 pool->flags &= ~POOL_DISASSOCIATED;
4155 worker = create_worker(pool);
4157 spin_lock_irq(&pool->lock);
4158 start_worker(worker);
4159 spin_unlock_irq(&pool->lock);
4163 /* create default unbound wq attrs */
4164 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4165 struct workqueue_attrs *attrs;
4167 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4169 attrs->nice = std_nice[i];
4170 cpumask_setall(attrs->cpumask);
4172 unbound_std_wq_attrs[i] = attrs;
4175 system_wq = alloc_workqueue("events", 0, 0);
4176 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4177 system_long_wq = alloc_workqueue("events_long", 0, 0);
4178 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4179 WQ_UNBOUND_MAX_ACTIVE);
4180 system_freezable_wq = alloc_workqueue("events_freezable",
4182 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4183 !system_unbound_wq || !system_freezable_wq);
4186 early_initcall(init_workqueues);