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.
127 /* struct worker is defined in workqueue_internal.h */
130 spinlock_t lock; /* the pool lock */
131 int cpu; /* I: the associated cpu */
132 int id; /* I: pool ID */
133 unsigned int flags; /* X: flags */
135 struct list_head worklist; /* L: list of pending works */
136 int nr_workers; /* L: total number of workers */
138 /* nr_idle includes the ones off idle_list for rebinding */
139 int nr_idle; /* L: currently idle ones */
141 struct list_head idle_list; /* X: list of idle workers */
142 struct timer_list idle_timer; /* L: worker idle timeout */
143 struct timer_list mayday_timer; /* L: SOS timer for workers */
145 /* workers are chained either in busy_hash or idle_list */
146 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
147 /* L: hash of busy workers */
149 struct mutex manager_arb; /* manager arbitration */
150 struct mutex assoc_mutex; /* protect POOL_DISASSOCIATED */
151 struct ida worker_ida; /* L: for worker IDs */
153 struct workqueue_attrs *attrs; /* I: worker attributes */
154 struct hlist_node hash_node; /* R: unbound_pool_hash node */
155 int refcnt; /* refcnt for unbound pools */
158 * The current concurrency level. As it's likely to be accessed
159 * from other CPUs during try_to_wake_up(), put it in a separate
162 atomic_t nr_running ____cacheline_aligned_in_smp;
165 * Destruction of pool is sched-RCU protected to allow dereferences
166 * from get_work_pool().
169 } ____cacheline_aligned_in_smp;
172 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
173 * of work_struct->data are used for flags and the remaining high bits
174 * point to the pwq; thus, pwqs need to be aligned at two's power of the
175 * number of flag bits.
177 struct pool_workqueue {
178 struct worker_pool *pool; /* I: the associated pool */
179 struct workqueue_struct *wq; /* I: the owning workqueue */
180 int work_color; /* L: current color */
181 int flush_color; /* L: flushing color */
182 int nr_in_flight[WORK_NR_COLORS];
183 /* L: nr of in_flight works */
184 int nr_active; /* L: nr of active works */
185 int max_active; /* L: max active works */
186 struct list_head delayed_works; /* L: delayed works */
187 struct list_head pwqs_node; /* R: node on wq->pwqs */
188 struct list_head mayday_node; /* W: node on wq->maydays */
189 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
192 * Structure used to wait for workqueue flush.
195 struct list_head list; /* F: list of flushers */
196 int flush_color; /* F: flush color waiting for */
197 struct completion done; /* flush completion */
201 * The externally visible workqueue abstraction is an array of
202 * per-CPU workqueues:
204 struct workqueue_struct {
205 unsigned int flags; /* W: WQ_* flags */
206 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */
207 struct list_head pwqs; /* R: all pwqs of this wq */
208 struct list_head list; /* W: list of all workqueues */
210 struct mutex flush_mutex; /* protects wq flushing */
211 int work_color; /* F: current work color */
212 int flush_color; /* F: current flush color */
213 atomic_t nr_pwqs_to_flush; /* flush in progress */
214 struct wq_flusher *first_flusher; /* F: first flusher */
215 struct list_head flusher_queue; /* F: flush waiters */
216 struct list_head flusher_overflow; /* F: flush overflow list */
218 struct list_head maydays; /* W: pwqs requesting rescue */
219 struct worker *rescuer; /* I: rescue worker */
221 int nr_drainers; /* W: drain in progress */
222 int saved_max_active; /* W: saved pwq max_active */
223 #ifdef CONFIG_LOCKDEP
224 struct lockdep_map lockdep_map;
226 char name[]; /* I: workqueue name */
229 static struct kmem_cache *pwq_cache;
231 /* hash of all unbound pools keyed by pool->attrs */
232 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
234 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
236 struct workqueue_struct *system_wq __read_mostly;
237 EXPORT_SYMBOL_GPL(system_wq);
238 struct workqueue_struct *system_highpri_wq __read_mostly;
239 EXPORT_SYMBOL_GPL(system_highpri_wq);
240 struct workqueue_struct *system_long_wq __read_mostly;
241 EXPORT_SYMBOL_GPL(system_long_wq);
242 struct workqueue_struct *system_unbound_wq __read_mostly;
243 EXPORT_SYMBOL_GPL(system_unbound_wq);
244 struct workqueue_struct *system_freezable_wq __read_mostly;
245 EXPORT_SYMBOL_GPL(system_freezable_wq);
247 #define CREATE_TRACE_POINTS
248 #include <trace/events/workqueue.h>
250 #define assert_rcu_or_wq_lock() \
251 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
252 lockdep_is_held(&workqueue_lock), \
253 "sched RCU or workqueue lock should be held")
255 #define for_each_cpu_worker_pool(pool, cpu) \
256 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
257 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
260 #define for_each_busy_worker(worker, i, pool) \
261 hash_for_each(pool->busy_hash, i, worker, hentry)
264 * for_each_pool - iterate through all worker_pools in the system
265 * @pool: iteration cursor
266 * @id: integer used for iteration
268 * This must be called either with workqueue_lock held or sched RCU read
269 * locked. If the pool needs to be used beyond the locking in effect, the
270 * caller is responsible for guaranteeing that the pool stays online.
272 * The if/else clause exists only for the lockdep assertion and can be
275 #define for_each_pool(pool, id) \
276 idr_for_each_entry(&worker_pool_idr, pool, id) \
277 if (({ assert_rcu_or_wq_lock(); false; })) { } \
281 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
282 * @pwq: iteration cursor
283 * @wq: the target workqueue
285 * This must be called either with workqueue_lock held or sched RCU read
286 * locked. If the pwq needs to be used beyond the locking in effect, the
287 * caller is responsible for guaranteeing that the pwq stays online.
289 * The if/else clause exists only for the lockdep assertion and can be
292 #define for_each_pwq(pwq, wq) \
293 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
294 if (({ assert_rcu_or_wq_lock(); false; })) { } \
297 #ifdef CONFIG_DEBUG_OBJECTS_WORK
299 static struct debug_obj_descr work_debug_descr;
301 static void *work_debug_hint(void *addr)
303 return ((struct work_struct *) addr)->func;
307 * fixup_init is called when:
308 * - an active object is initialized
310 static int work_fixup_init(void *addr, enum debug_obj_state state)
312 struct work_struct *work = addr;
315 case ODEBUG_STATE_ACTIVE:
316 cancel_work_sync(work);
317 debug_object_init(work, &work_debug_descr);
325 * fixup_activate is called when:
326 * - an active object is activated
327 * - an unknown object is activated (might be a statically initialized object)
329 static int work_fixup_activate(void *addr, enum debug_obj_state state)
331 struct work_struct *work = addr;
335 case ODEBUG_STATE_NOTAVAILABLE:
337 * This is not really a fixup. The work struct was
338 * statically initialized. We just make sure that it
339 * is tracked in the object tracker.
341 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
342 debug_object_init(work, &work_debug_descr);
343 debug_object_activate(work, &work_debug_descr);
349 case ODEBUG_STATE_ACTIVE:
358 * fixup_free is called when:
359 * - an active object is freed
361 static int work_fixup_free(void *addr, enum debug_obj_state state)
363 struct work_struct *work = addr;
366 case ODEBUG_STATE_ACTIVE:
367 cancel_work_sync(work);
368 debug_object_free(work, &work_debug_descr);
375 static struct debug_obj_descr work_debug_descr = {
376 .name = "work_struct",
377 .debug_hint = work_debug_hint,
378 .fixup_init = work_fixup_init,
379 .fixup_activate = work_fixup_activate,
380 .fixup_free = work_fixup_free,
383 static inline void debug_work_activate(struct work_struct *work)
385 debug_object_activate(work, &work_debug_descr);
388 static inline void debug_work_deactivate(struct work_struct *work)
390 debug_object_deactivate(work, &work_debug_descr);
393 void __init_work(struct work_struct *work, int onstack)
396 debug_object_init_on_stack(work, &work_debug_descr);
398 debug_object_init(work, &work_debug_descr);
400 EXPORT_SYMBOL_GPL(__init_work);
402 void destroy_work_on_stack(struct work_struct *work)
404 debug_object_free(work, &work_debug_descr);
406 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
409 static inline void debug_work_activate(struct work_struct *work) { }
410 static inline void debug_work_deactivate(struct work_struct *work) { }
413 /* Serializes the accesses to the list of workqueues. */
414 static DEFINE_SPINLOCK(workqueue_lock);
415 static LIST_HEAD(workqueues);
416 static bool workqueue_freezing; /* W: have wqs started freezing? */
419 * The CPU and unbound standard worker pools. The unbound ones have
420 * POOL_DISASSOCIATED set, and their workers have WORKER_UNBOUND set.
422 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
426 * idr of all pools. Modifications are protected by workqueue_lock. Read
427 * accesses are protected by sched-RCU protected.
429 static DEFINE_IDR(worker_pool_idr);
431 static int worker_thread(void *__worker);
433 /* allocate ID and assign it to @pool */
434 static int worker_pool_assign_id(struct worker_pool *pool)
439 if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL))
442 spin_lock_irq(&workqueue_lock);
443 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
444 spin_unlock_irq(&workqueue_lock);
445 } while (ret == -EAGAIN);
451 * first_pwq - return the first pool_workqueue of the specified workqueue
452 * @wq: the target workqueue
454 * This must be called either with workqueue_lock held or sched RCU read
455 * locked. If the pwq needs to be used beyond the locking in effect, the
456 * caller is responsible for guaranteeing that the pwq stays online.
458 static struct pool_workqueue *first_pwq(struct workqueue_struct *wq)
460 assert_rcu_or_wq_lock();
461 return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue,
465 static unsigned int work_color_to_flags(int color)
467 return color << WORK_STRUCT_COLOR_SHIFT;
470 static int get_work_color(struct work_struct *work)
472 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
473 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
476 static int work_next_color(int color)
478 return (color + 1) % WORK_NR_COLORS;
482 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
483 * contain the pointer to the queued pwq. Once execution starts, the flag
484 * is cleared and the high bits contain OFFQ flags and pool ID.
486 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
487 * and clear_work_data() can be used to set the pwq, pool or clear
488 * work->data. These functions should only be called while the work is
489 * owned - ie. while the PENDING bit is set.
491 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
492 * corresponding to a work. Pool is available once the work has been
493 * queued anywhere after initialization until it is sync canceled. pwq is
494 * available only while the work item is queued.
496 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
497 * canceled. While being canceled, a work item may have its PENDING set
498 * but stay off timer and worklist for arbitrarily long and nobody should
499 * try to steal the PENDING bit.
501 static inline void set_work_data(struct work_struct *work, unsigned long data,
504 WARN_ON_ONCE(!work_pending(work));
505 atomic_long_set(&work->data, data | flags | work_static(work));
508 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
509 unsigned long extra_flags)
511 set_work_data(work, (unsigned long)pwq,
512 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
515 static void set_work_pool_and_keep_pending(struct work_struct *work,
518 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
519 WORK_STRUCT_PENDING);
522 static void set_work_pool_and_clear_pending(struct work_struct *work,
526 * The following wmb is paired with the implied mb in
527 * test_and_set_bit(PENDING) and ensures all updates to @work made
528 * here are visible to and precede any updates by the next PENDING
532 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
535 static void clear_work_data(struct work_struct *work)
537 smp_wmb(); /* see set_work_pool_and_clear_pending() */
538 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
541 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
543 unsigned long data = atomic_long_read(&work->data);
545 if (data & WORK_STRUCT_PWQ)
546 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
552 * get_work_pool - return the worker_pool a given work was associated with
553 * @work: the work item of interest
555 * Return the worker_pool @work was last associated with. %NULL if none.
557 * Pools are created and destroyed under workqueue_lock, and allows read
558 * access under sched-RCU read lock. As such, this function should be
559 * called under workqueue_lock or with preemption disabled.
561 * All fields of the returned pool are accessible as long as the above
562 * mentioned locking is in effect. If the returned pool needs to be used
563 * beyond the critical section, the caller is responsible for ensuring the
564 * returned pool is and stays online.
566 static struct worker_pool *get_work_pool(struct work_struct *work)
568 unsigned long data = atomic_long_read(&work->data);
571 assert_rcu_or_wq_lock();
573 if (data & WORK_STRUCT_PWQ)
574 return ((struct pool_workqueue *)
575 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
577 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
578 if (pool_id == WORK_OFFQ_POOL_NONE)
581 return idr_find(&worker_pool_idr, pool_id);
585 * get_work_pool_id - return the worker pool ID a given work is associated with
586 * @work: the work item of interest
588 * Return the worker_pool ID @work was last associated with.
589 * %WORK_OFFQ_POOL_NONE if none.
591 static int get_work_pool_id(struct work_struct *work)
593 unsigned long data = atomic_long_read(&work->data);
595 if (data & WORK_STRUCT_PWQ)
596 return ((struct pool_workqueue *)
597 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
599 return data >> WORK_OFFQ_POOL_SHIFT;
602 static void mark_work_canceling(struct work_struct *work)
604 unsigned long pool_id = get_work_pool_id(work);
606 pool_id <<= WORK_OFFQ_POOL_SHIFT;
607 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
610 static bool work_is_canceling(struct work_struct *work)
612 unsigned long data = atomic_long_read(&work->data);
614 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
618 * Policy functions. These define the policies on how the global worker
619 * pools are managed. Unless noted otherwise, these functions assume that
620 * they're being called with pool->lock held.
623 static bool __need_more_worker(struct worker_pool *pool)
625 return !atomic_read(&pool->nr_running);
629 * Need to wake up a worker? Called from anything but currently
632 * Note that, because unbound workers never contribute to nr_running, this
633 * function will always return %true for unbound pools as long as the
634 * worklist isn't empty.
636 static bool need_more_worker(struct worker_pool *pool)
638 return !list_empty(&pool->worklist) && __need_more_worker(pool);
641 /* Can I start working? Called from busy but !running workers. */
642 static bool may_start_working(struct worker_pool *pool)
644 return pool->nr_idle;
647 /* Do I need to keep working? Called from currently running workers. */
648 static bool keep_working(struct worker_pool *pool)
650 return !list_empty(&pool->worklist) &&
651 atomic_read(&pool->nr_running) <= 1;
654 /* Do we need a new worker? Called from manager. */
655 static bool need_to_create_worker(struct worker_pool *pool)
657 return need_more_worker(pool) && !may_start_working(pool);
660 /* Do I need to be the manager? */
661 static bool need_to_manage_workers(struct worker_pool *pool)
663 return need_to_create_worker(pool) ||
664 (pool->flags & POOL_MANAGE_WORKERS);
667 /* Do we have too many workers and should some go away? */
668 static bool too_many_workers(struct worker_pool *pool)
670 bool managing = mutex_is_locked(&pool->manager_arb);
671 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
672 int nr_busy = pool->nr_workers - nr_idle;
675 * nr_idle and idle_list may disagree if idle rebinding is in
676 * progress. Never return %true if idle_list is empty.
678 if (list_empty(&pool->idle_list))
681 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
688 /* Return the first worker. Safe with preemption disabled */
689 static struct worker *first_worker(struct worker_pool *pool)
691 if (unlikely(list_empty(&pool->idle_list)))
694 return list_first_entry(&pool->idle_list, struct worker, entry);
698 * wake_up_worker - wake up an idle worker
699 * @pool: worker pool to wake worker from
701 * Wake up the first idle worker of @pool.
704 * spin_lock_irq(pool->lock).
706 static void wake_up_worker(struct worker_pool *pool)
708 struct worker *worker = first_worker(pool);
711 wake_up_process(worker->task);
715 * wq_worker_waking_up - a worker is waking up
716 * @task: task waking up
717 * @cpu: CPU @task is waking up to
719 * This function is called during try_to_wake_up() when a worker is
723 * spin_lock_irq(rq->lock)
725 void wq_worker_waking_up(struct task_struct *task, int cpu)
727 struct worker *worker = kthread_data(task);
729 if (!(worker->flags & WORKER_NOT_RUNNING)) {
730 WARN_ON_ONCE(worker->pool->cpu != cpu);
731 atomic_inc(&worker->pool->nr_running);
736 * wq_worker_sleeping - a worker is going to sleep
737 * @task: task going to sleep
738 * @cpu: CPU in question, must be the current CPU number
740 * This function is called during schedule() when a busy worker is
741 * going to sleep. Worker on the same cpu can be woken up by
742 * returning pointer to its task.
745 * spin_lock_irq(rq->lock)
748 * Worker task on @cpu to wake up, %NULL if none.
750 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
752 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
753 struct worker_pool *pool;
756 * Rescuers, which may not have all the fields set up like normal
757 * workers, also reach here, let's not access anything before
758 * checking NOT_RUNNING.
760 if (worker->flags & WORKER_NOT_RUNNING)
765 /* this can only happen on the local cpu */
766 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
770 * The counterpart of the following dec_and_test, implied mb,
771 * worklist not empty test sequence is in insert_work().
772 * Please read comment there.
774 * NOT_RUNNING is clear. This means that we're bound to and
775 * running on the local cpu w/ rq lock held and preemption
776 * disabled, which in turn means that none else could be
777 * manipulating idle_list, so dereferencing idle_list without pool
780 if (atomic_dec_and_test(&pool->nr_running) &&
781 !list_empty(&pool->worklist))
782 to_wakeup = first_worker(pool);
783 return to_wakeup ? to_wakeup->task : NULL;
787 * worker_set_flags - set worker flags and adjust nr_running accordingly
789 * @flags: flags to set
790 * @wakeup: wakeup an idle worker if necessary
792 * Set @flags in @worker->flags and adjust nr_running accordingly. If
793 * nr_running becomes zero and @wakeup is %true, an idle worker is
797 * spin_lock_irq(pool->lock)
799 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
802 struct worker_pool *pool = worker->pool;
804 WARN_ON_ONCE(worker->task != current);
807 * If transitioning into NOT_RUNNING, adjust nr_running and
808 * wake up an idle worker as necessary if requested by
811 if ((flags & WORKER_NOT_RUNNING) &&
812 !(worker->flags & WORKER_NOT_RUNNING)) {
814 if (atomic_dec_and_test(&pool->nr_running) &&
815 !list_empty(&pool->worklist))
816 wake_up_worker(pool);
818 atomic_dec(&pool->nr_running);
821 worker->flags |= flags;
825 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
827 * @flags: flags to clear
829 * Clear @flags in @worker->flags and adjust nr_running accordingly.
832 * spin_lock_irq(pool->lock)
834 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
836 struct worker_pool *pool = worker->pool;
837 unsigned int oflags = worker->flags;
839 WARN_ON_ONCE(worker->task != current);
841 worker->flags &= ~flags;
844 * If transitioning out of NOT_RUNNING, increment nr_running. Note
845 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
846 * of multiple flags, not a single flag.
848 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
849 if (!(worker->flags & WORKER_NOT_RUNNING))
850 atomic_inc(&pool->nr_running);
854 * find_worker_executing_work - find worker which is executing a work
855 * @pool: pool of interest
856 * @work: work to find worker for
858 * Find a worker which is executing @work on @pool by searching
859 * @pool->busy_hash which is keyed by the address of @work. For a worker
860 * to match, its current execution should match the address of @work and
861 * its work function. This is to avoid unwanted dependency between
862 * unrelated work executions through a work item being recycled while still
865 * This is a bit tricky. A work item may be freed once its execution
866 * starts and nothing prevents the freed area from being recycled for
867 * another work item. If the same work item address ends up being reused
868 * before the original execution finishes, workqueue will identify the
869 * recycled work item as currently executing and make it wait until the
870 * current execution finishes, introducing an unwanted dependency.
872 * This function checks the work item address, work function and workqueue
873 * to avoid false positives. Note that this isn't complete as one may
874 * construct a work function which can introduce dependency onto itself
875 * through a recycled work item. Well, if somebody wants to shoot oneself
876 * in the foot that badly, there's only so much we can do, and if such
877 * deadlock actually occurs, it should be easy to locate the culprit work
881 * spin_lock_irq(pool->lock).
884 * Pointer to worker which is executing @work if found, NULL
887 static struct worker *find_worker_executing_work(struct worker_pool *pool,
888 struct work_struct *work)
890 struct worker *worker;
892 hash_for_each_possible(pool->busy_hash, worker, hentry,
894 if (worker->current_work == work &&
895 worker->current_func == work->func)
902 * move_linked_works - move linked works to a list
903 * @work: start of series of works to be scheduled
904 * @head: target list to append @work to
905 * @nextp: out paramter for nested worklist walking
907 * Schedule linked works starting from @work to @head. Work series to
908 * be scheduled starts at @work and includes any consecutive work with
909 * WORK_STRUCT_LINKED set in its predecessor.
911 * If @nextp is not NULL, it's updated to point to the next work of
912 * the last scheduled work. This allows move_linked_works() to be
913 * nested inside outer list_for_each_entry_safe().
916 * spin_lock_irq(pool->lock).
918 static void move_linked_works(struct work_struct *work, struct list_head *head,
919 struct work_struct **nextp)
921 struct work_struct *n;
924 * Linked worklist will always end before the end of the list,
925 * use NULL for list head.
927 list_for_each_entry_safe_from(work, n, NULL, entry) {
928 list_move_tail(&work->entry, head);
929 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
934 * If we're already inside safe list traversal and have moved
935 * multiple works to the scheduled queue, the next position
936 * needs to be updated.
942 static void pwq_activate_delayed_work(struct work_struct *work)
944 struct pool_workqueue *pwq = get_work_pwq(work);
946 trace_workqueue_activate_work(work);
947 move_linked_works(work, &pwq->pool->worklist, NULL);
948 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
952 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
954 struct work_struct *work = list_first_entry(&pwq->delayed_works,
955 struct work_struct, entry);
957 pwq_activate_delayed_work(work);
961 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
962 * @pwq: pwq of interest
963 * @color: color of work which left the queue
965 * A work either has completed or is removed from pending queue,
966 * decrement nr_in_flight of its pwq and handle workqueue flushing.
969 * spin_lock_irq(pool->lock).
971 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
973 /* ignore uncolored works */
974 if (color == WORK_NO_COLOR)
977 pwq->nr_in_flight[color]--;
980 if (!list_empty(&pwq->delayed_works)) {
981 /* one down, submit a delayed one */
982 if (pwq->nr_active < pwq->max_active)
983 pwq_activate_first_delayed(pwq);
986 /* is flush in progress and are we at the flushing tip? */
987 if (likely(pwq->flush_color != color))
990 /* are there still in-flight works? */
991 if (pwq->nr_in_flight[color])
994 /* this pwq is done, clear flush_color */
995 pwq->flush_color = -1;
998 * If this was the last pwq, wake up the first flusher. It
999 * will handle the rest.
1001 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1002 complete(&pwq->wq->first_flusher->done);
1006 * try_to_grab_pending - steal work item from worklist and disable irq
1007 * @work: work item to steal
1008 * @is_dwork: @work is a delayed_work
1009 * @flags: place to store irq state
1011 * Try to grab PENDING bit of @work. This function can handle @work in any
1012 * stable state - idle, on timer or on worklist. Return values are
1014 * 1 if @work was pending and we successfully stole PENDING
1015 * 0 if @work was idle and we claimed PENDING
1016 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1017 * -ENOENT if someone else is canceling @work, this state may persist
1018 * for arbitrarily long
1020 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1021 * interrupted while holding PENDING and @work off queue, irq must be
1022 * disabled on entry. This, combined with delayed_work->timer being
1023 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1025 * On successful return, >= 0, irq is disabled and the caller is
1026 * responsible for releasing it using local_irq_restore(*@flags).
1028 * This function is safe to call from any context including IRQ handler.
1030 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1031 unsigned long *flags)
1033 struct worker_pool *pool;
1034 struct pool_workqueue *pwq;
1036 local_irq_save(*flags);
1038 /* try to steal the timer if it exists */
1040 struct delayed_work *dwork = to_delayed_work(work);
1043 * dwork->timer is irqsafe. If del_timer() fails, it's
1044 * guaranteed that the timer is not queued anywhere and not
1045 * running on the local CPU.
1047 if (likely(del_timer(&dwork->timer)))
1051 /* try to claim PENDING the normal way */
1052 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1056 * The queueing is in progress, or it is already queued. Try to
1057 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1059 pool = get_work_pool(work);
1063 spin_lock(&pool->lock);
1065 * work->data is guaranteed to point to pwq only while the work
1066 * item is queued on pwq->wq, and both updating work->data to point
1067 * to pwq on queueing and to pool on dequeueing are done under
1068 * pwq->pool->lock. This in turn guarantees that, if work->data
1069 * points to pwq which is associated with a locked pool, the work
1070 * item is currently queued on that pool.
1072 pwq = get_work_pwq(work);
1073 if (pwq && pwq->pool == pool) {
1074 debug_work_deactivate(work);
1077 * A delayed work item cannot be grabbed directly because
1078 * it might have linked NO_COLOR work items which, if left
1079 * on the delayed_list, will confuse pwq->nr_active
1080 * management later on and cause stall. Make sure the work
1081 * item is activated before grabbing.
1083 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1084 pwq_activate_delayed_work(work);
1086 list_del_init(&work->entry);
1087 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1089 /* work->data points to pwq iff queued, point to pool */
1090 set_work_pool_and_keep_pending(work, pool->id);
1092 spin_unlock(&pool->lock);
1095 spin_unlock(&pool->lock);
1097 local_irq_restore(*flags);
1098 if (work_is_canceling(work))
1105 * insert_work - insert a work into a pool
1106 * @pwq: pwq @work belongs to
1107 * @work: work to insert
1108 * @head: insertion point
1109 * @extra_flags: extra WORK_STRUCT_* flags to set
1111 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1112 * work_struct flags.
1115 * spin_lock_irq(pool->lock).
1117 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1118 struct list_head *head, unsigned int extra_flags)
1120 struct worker_pool *pool = pwq->pool;
1122 /* we own @work, set data and link */
1123 set_work_pwq(work, pwq, extra_flags);
1124 list_add_tail(&work->entry, head);
1127 * Ensure either worker_sched_deactivated() sees the above
1128 * list_add_tail() or we see zero nr_running to avoid workers
1129 * lying around lazily while there are works to be processed.
1133 if (__need_more_worker(pool))
1134 wake_up_worker(pool);
1138 * Test whether @work is being queued from another work executing on the
1141 static bool is_chained_work(struct workqueue_struct *wq)
1143 struct worker *worker;
1145 worker = current_wq_worker();
1147 * Return %true iff I'm a worker execuing a work item on @wq. If
1148 * I'm @worker, it's safe to dereference it without locking.
1150 return worker && worker->current_pwq->wq == wq;
1153 static void __queue_work(int cpu, struct workqueue_struct *wq,
1154 struct work_struct *work)
1156 struct pool_workqueue *pwq;
1157 struct list_head *worklist;
1158 unsigned int work_flags;
1159 unsigned int req_cpu = cpu;
1162 * While a work item is PENDING && off queue, a task trying to
1163 * steal the PENDING will busy-loop waiting for it to either get
1164 * queued or lose PENDING. Grabbing PENDING and queueing should
1165 * happen with IRQ disabled.
1167 WARN_ON_ONCE(!irqs_disabled());
1169 debug_work_activate(work);
1171 /* if dying, only works from the same workqueue are allowed */
1172 if (unlikely(wq->flags & WQ_DRAINING) &&
1173 WARN_ON_ONCE(!is_chained_work(wq)))
1176 /* determine the pwq to use */
1177 if (!(wq->flags & WQ_UNBOUND)) {
1178 struct worker_pool *last_pool;
1180 if (cpu == WORK_CPU_UNBOUND)
1181 cpu = raw_smp_processor_id();
1184 * It's multi cpu. If @work was previously on a different
1185 * cpu, it might still be running there, in which case the
1186 * work needs to be queued on that cpu to guarantee
1189 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1190 last_pool = get_work_pool(work);
1192 if (last_pool && last_pool != pwq->pool) {
1193 struct worker *worker;
1195 spin_lock(&last_pool->lock);
1197 worker = find_worker_executing_work(last_pool, work);
1199 if (worker && worker->current_pwq->wq == wq) {
1200 pwq = per_cpu_ptr(wq->cpu_pwqs, last_pool->cpu);
1202 /* meh... not running there, queue here */
1203 spin_unlock(&last_pool->lock);
1204 spin_lock(&pwq->pool->lock);
1207 spin_lock(&pwq->pool->lock);
1210 pwq = first_pwq(wq);
1211 spin_lock(&pwq->pool->lock);
1214 /* pwq determined, queue */
1215 trace_workqueue_queue_work(req_cpu, pwq, work);
1217 if (WARN_ON(!list_empty(&work->entry))) {
1218 spin_unlock(&pwq->pool->lock);
1222 pwq->nr_in_flight[pwq->work_color]++;
1223 work_flags = work_color_to_flags(pwq->work_color);
1225 if (likely(pwq->nr_active < pwq->max_active)) {
1226 trace_workqueue_activate_work(work);
1228 worklist = &pwq->pool->worklist;
1230 work_flags |= WORK_STRUCT_DELAYED;
1231 worklist = &pwq->delayed_works;
1234 insert_work(pwq, work, worklist, work_flags);
1236 spin_unlock(&pwq->pool->lock);
1240 * queue_work_on - queue work on specific cpu
1241 * @cpu: CPU number to execute work on
1242 * @wq: workqueue to use
1243 * @work: work to queue
1245 * Returns %false if @work was already on a queue, %true otherwise.
1247 * We queue the work to a specific CPU, the caller must ensure it
1250 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1251 struct work_struct *work)
1254 unsigned long flags;
1256 local_irq_save(flags);
1258 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1259 __queue_work(cpu, wq, work);
1263 local_irq_restore(flags);
1266 EXPORT_SYMBOL_GPL(queue_work_on);
1269 * queue_work - queue work on a workqueue
1270 * @wq: workqueue to use
1271 * @work: work to queue
1273 * Returns %false if @work was already on a queue, %true otherwise.
1275 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1276 * it can be processed by another CPU.
1278 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1280 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1282 EXPORT_SYMBOL_GPL(queue_work);
1284 void delayed_work_timer_fn(unsigned long __data)
1286 struct delayed_work *dwork = (struct delayed_work *)__data;
1288 /* should have been called from irqsafe timer with irq already off */
1289 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1291 EXPORT_SYMBOL(delayed_work_timer_fn);
1293 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1294 struct delayed_work *dwork, unsigned long delay)
1296 struct timer_list *timer = &dwork->timer;
1297 struct work_struct *work = &dwork->work;
1299 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1300 timer->data != (unsigned long)dwork);
1301 WARN_ON_ONCE(timer_pending(timer));
1302 WARN_ON_ONCE(!list_empty(&work->entry));
1305 * If @delay is 0, queue @dwork->work immediately. This is for
1306 * both optimization and correctness. The earliest @timer can
1307 * expire is on the closest next tick and delayed_work users depend
1308 * on that there's no such delay when @delay is 0.
1311 __queue_work(cpu, wq, &dwork->work);
1315 timer_stats_timer_set_start_info(&dwork->timer);
1319 timer->expires = jiffies + delay;
1321 if (unlikely(cpu != WORK_CPU_UNBOUND))
1322 add_timer_on(timer, cpu);
1328 * queue_delayed_work_on - queue work on specific CPU after delay
1329 * @cpu: CPU number to execute work on
1330 * @wq: workqueue to use
1331 * @dwork: work to queue
1332 * @delay: number of jiffies to wait before queueing
1334 * Returns %false if @work was already on a queue, %true otherwise. If
1335 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1338 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1339 struct delayed_work *dwork, unsigned long delay)
1341 struct work_struct *work = &dwork->work;
1343 unsigned long flags;
1345 /* read the comment in __queue_work() */
1346 local_irq_save(flags);
1348 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1349 __queue_delayed_work(cpu, wq, dwork, delay);
1353 local_irq_restore(flags);
1356 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1359 * queue_delayed_work - queue work on a workqueue after delay
1360 * @wq: workqueue to use
1361 * @dwork: delayable work to queue
1362 * @delay: number of jiffies to wait before queueing
1364 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1366 bool queue_delayed_work(struct workqueue_struct *wq,
1367 struct delayed_work *dwork, unsigned long delay)
1369 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1371 EXPORT_SYMBOL_GPL(queue_delayed_work);
1374 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1375 * @cpu: CPU number to execute work on
1376 * @wq: workqueue to use
1377 * @dwork: work to queue
1378 * @delay: number of jiffies to wait before queueing
1380 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1381 * modify @dwork's timer so that it expires after @delay. If @delay is
1382 * zero, @work is guaranteed to be scheduled immediately regardless of its
1385 * Returns %false if @dwork was idle and queued, %true if @dwork was
1386 * pending and its timer was modified.
1388 * This function is safe to call from any context including IRQ handler.
1389 * See try_to_grab_pending() for details.
1391 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1392 struct delayed_work *dwork, unsigned long delay)
1394 unsigned long flags;
1398 ret = try_to_grab_pending(&dwork->work, true, &flags);
1399 } while (unlikely(ret == -EAGAIN));
1401 if (likely(ret >= 0)) {
1402 __queue_delayed_work(cpu, wq, dwork, delay);
1403 local_irq_restore(flags);
1406 /* -ENOENT from try_to_grab_pending() becomes %true */
1409 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1412 * mod_delayed_work - modify delay of or queue a delayed work
1413 * @wq: workqueue to use
1414 * @dwork: work to queue
1415 * @delay: number of jiffies to wait before queueing
1417 * mod_delayed_work_on() on local CPU.
1419 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1420 unsigned long delay)
1422 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1424 EXPORT_SYMBOL_GPL(mod_delayed_work);
1427 * worker_enter_idle - enter idle state
1428 * @worker: worker which is entering idle state
1430 * @worker is entering idle state. Update stats and idle timer if
1434 * spin_lock_irq(pool->lock).
1436 static void worker_enter_idle(struct worker *worker)
1438 struct worker_pool *pool = worker->pool;
1440 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1441 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1442 (worker->hentry.next || worker->hentry.pprev)))
1445 /* can't use worker_set_flags(), also called from start_worker() */
1446 worker->flags |= WORKER_IDLE;
1448 worker->last_active = jiffies;
1450 /* idle_list is LIFO */
1451 list_add(&worker->entry, &pool->idle_list);
1453 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1454 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1457 * Sanity check nr_running. Because wq_unbind_fn() releases
1458 * pool->lock between setting %WORKER_UNBOUND and zapping
1459 * nr_running, the warning may trigger spuriously. Check iff
1460 * unbind is not in progress.
1462 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1463 pool->nr_workers == pool->nr_idle &&
1464 atomic_read(&pool->nr_running));
1468 * worker_leave_idle - leave idle state
1469 * @worker: worker which is leaving idle state
1471 * @worker is leaving idle state. Update stats.
1474 * spin_lock_irq(pool->lock).
1476 static void worker_leave_idle(struct worker *worker)
1478 struct worker_pool *pool = worker->pool;
1480 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1482 worker_clr_flags(worker, WORKER_IDLE);
1484 list_del_init(&worker->entry);
1488 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1489 * @pool: target worker_pool
1491 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1493 * Works which are scheduled while the cpu is online must at least be
1494 * scheduled to a worker which is bound to the cpu so that if they are
1495 * flushed from cpu callbacks while cpu is going down, they are
1496 * guaranteed to execute on the cpu.
1498 * This function is to be used by unbound workers and rescuers to bind
1499 * themselves to the target cpu and may race with cpu going down or
1500 * coming online. kthread_bind() can't be used because it may put the
1501 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1502 * verbatim as it's best effort and blocking and pool may be
1503 * [dis]associated in the meantime.
1505 * This function tries set_cpus_allowed() and locks pool and verifies the
1506 * binding against %POOL_DISASSOCIATED which is set during
1507 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1508 * enters idle state or fetches works without dropping lock, it can
1509 * guarantee the scheduling requirement described in the first paragraph.
1512 * Might sleep. Called without any lock but returns with pool->lock
1516 * %true if the associated pool is online (@worker is successfully
1517 * bound), %false if offline.
1519 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1520 __acquires(&pool->lock)
1524 * The following call may fail, succeed or succeed
1525 * without actually migrating the task to the cpu if
1526 * it races with cpu hotunplug operation. Verify
1527 * against POOL_DISASSOCIATED.
1529 if (!(pool->flags & POOL_DISASSOCIATED))
1530 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1532 spin_lock_irq(&pool->lock);
1533 if (pool->flags & POOL_DISASSOCIATED)
1535 if (task_cpu(current) == pool->cpu &&
1536 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1538 spin_unlock_irq(&pool->lock);
1541 * We've raced with CPU hot[un]plug. Give it a breather
1542 * and retry migration. cond_resched() is required here;
1543 * otherwise, we might deadlock against cpu_stop trying to
1544 * bring down the CPU on non-preemptive kernel.
1552 * Rebind an idle @worker to its CPU. worker_thread() will test
1553 * list_empty(@worker->entry) before leaving idle and call this function.
1555 static void idle_worker_rebind(struct worker *worker)
1557 /* CPU may go down again inbetween, clear UNBOUND only on success */
1558 if (worker_maybe_bind_and_lock(worker->pool))
1559 worker_clr_flags(worker, WORKER_UNBOUND);
1561 /* rebind complete, become available again */
1562 list_add(&worker->entry, &worker->pool->idle_list);
1563 spin_unlock_irq(&worker->pool->lock);
1567 * Function for @worker->rebind.work used to rebind unbound busy workers to
1568 * the associated cpu which is coming back online. This is scheduled by
1569 * cpu up but can race with other cpu hotplug operations and may be
1570 * executed twice without intervening cpu down.
1572 static void busy_worker_rebind_fn(struct work_struct *work)
1574 struct worker *worker = container_of(work, struct worker, rebind_work);
1576 if (worker_maybe_bind_and_lock(worker->pool))
1577 worker_clr_flags(worker, WORKER_UNBOUND);
1579 spin_unlock_irq(&worker->pool->lock);
1583 * rebind_workers - rebind all workers of a pool to the associated CPU
1584 * @pool: pool of interest
1586 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1587 * is different for idle and busy ones.
1589 * Idle ones will be removed from the idle_list and woken up. They will
1590 * add themselves back after completing rebind. This ensures that the
1591 * idle_list doesn't contain any unbound workers when re-bound busy workers
1592 * try to perform local wake-ups for concurrency management.
1594 * Busy workers can rebind after they finish their current work items.
1595 * Queueing the rebind work item at the head of the scheduled list is
1596 * enough. Note that nr_running will be properly bumped as busy workers
1599 * On return, all non-manager workers are scheduled for rebind - see
1600 * manage_workers() for the manager special case. Any idle worker
1601 * including the manager will not appear on @idle_list until rebind is
1602 * complete, making local wake-ups safe.
1604 static void rebind_workers(struct worker_pool *pool)
1606 struct worker *worker, *n;
1609 lockdep_assert_held(&pool->assoc_mutex);
1610 lockdep_assert_held(&pool->lock);
1612 /* dequeue and kick idle ones */
1613 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1615 * idle workers should be off @pool->idle_list until rebind
1616 * is complete to avoid receiving premature local wake-ups.
1618 list_del_init(&worker->entry);
1621 * worker_thread() will see the above dequeuing and call
1622 * idle_worker_rebind().
1624 wake_up_process(worker->task);
1627 /* rebind busy workers */
1628 for_each_busy_worker(worker, i, pool) {
1629 struct work_struct *rebind_work = &worker->rebind_work;
1630 struct workqueue_struct *wq;
1632 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1633 work_data_bits(rebind_work)))
1636 debug_work_activate(rebind_work);
1639 * wq doesn't really matter but let's keep @worker->pool
1640 * and @pwq->pool consistent for sanity.
1642 if (worker->pool->attrs->nice < 0)
1643 wq = system_highpri_wq;
1647 insert_work(per_cpu_ptr(wq->cpu_pwqs, pool->cpu), rebind_work,
1648 worker->scheduled.next,
1649 work_color_to_flags(WORK_NO_COLOR));
1653 static struct worker *alloc_worker(void)
1655 struct worker *worker;
1657 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1659 INIT_LIST_HEAD(&worker->entry);
1660 INIT_LIST_HEAD(&worker->scheduled);
1661 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1662 /* on creation a worker is in !idle && prep state */
1663 worker->flags = WORKER_PREP;
1669 * create_worker - create a new workqueue worker
1670 * @pool: pool the new worker will belong to
1672 * Create a new worker which is bound to @pool. The returned worker
1673 * can be started by calling start_worker() or destroyed using
1677 * Might sleep. Does GFP_KERNEL allocations.
1680 * Pointer to the newly created worker.
1682 static struct worker *create_worker(struct worker_pool *pool)
1684 const char *pri = pool->attrs->nice < 0 ? "H" : "";
1685 struct worker *worker = NULL;
1688 spin_lock_irq(&pool->lock);
1689 while (ida_get_new(&pool->worker_ida, &id)) {
1690 spin_unlock_irq(&pool->lock);
1691 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1693 spin_lock_irq(&pool->lock);
1695 spin_unlock_irq(&pool->lock);
1697 worker = alloc_worker();
1701 worker->pool = pool;
1705 worker->task = kthread_create_on_node(worker_thread,
1706 worker, cpu_to_node(pool->cpu),
1707 "kworker/%d:%d%s", pool->cpu, id, pri);
1709 worker->task = kthread_create(worker_thread, worker,
1712 if (IS_ERR(worker->task))
1715 set_user_nice(worker->task, pool->attrs->nice);
1716 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1719 * %PF_THREAD_BOUND is used to prevent userland from meddling with
1720 * cpumask of workqueue workers. This is an abuse. We need
1721 * %PF_NO_SETAFFINITY.
1723 worker->task->flags |= PF_THREAD_BOUND;
1726 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1727 * remains stable across this function. See the comments above the
1728 * flag definition for details.
1730 if (pool->flags & POOL_DISASSOCIATED)
1731 worker->flags |= WORKER_UNBOUND;
1736 spin_lock_irq(&pool->lock);
1737 ida_remove(&pool->worker_ida, id);
1738 spin_unlock_irq(&pool->lock);
1745 * start_worker - start a newly created worker
1746 * @worker: worker to start
1748 * Make the pool aware of @worker and start it.
1751 * spin_lock_irq(pool->lock).
1753 static void start_worker(struct worker *worker)
1755 worker->flags |= WORKER_STARTED;
1756 worker->pool->nr_workers++;
1757 worker_enter_idle(worker);
1758 wake_up_process(worker->task);
1762 * destroy_worker - destroy a workqueue worker
1763 * @worker: worker to be destroyed
1765 * Destroy @worker and adjust @pool stats accordingly.
1768 * spin_lock_irq(pool->lock) which is released and regrabbed.
1770 static void destroy_worker(struct worker *worker)
1772 struct worker_pool *pool = worker->pool;
1773 int id = worker->id;
1775 /* sanity check frenzy */
1776 if (WARN_ON(worker->current_work) ||
1777 WARN_ON(!list_empty(&worker->scheduled)))
1780 if (worker->flags & WORKER_STARTED)
1782 if (worker->flags & WORKER_IDLE)
1785 list_del_init(&worker->entry);
1786 worker->flags |= WORKER_DIE;
1788 spin_unlock_irq(&pool->lock);
1790 kthread_stop(worker->task);
1793 spin_lock_irq(&pool->lock);
1794 ida_remove(&pool->worker_ida, id);
1797 static void idle_worker_timeout(unsigned long __pool)
1799 struct worker_pool *pool = (void *)__pool;
1801 spin_lock_irq(&pool->lock);
1803 if (too_many_workers(pool)) {
1804 struct worker *worker;
1805 unsigned long expires;
1807 /* idle_list is kept in LIFO order, check the last one */
1808 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1809 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1811 if (time_before(jiffies, expires))
1812 mod_timer(&pool->idle_timer, expires);
1814 /* it's been idle for too long, wake up manager */
1815 pool->flags |= POOL_MANAGE_WORKERS;
1816 wake_up_worker(pool);
1820 spin_unlock_irq(&pool->lock);
1823 static void send_mayday(struct work_struct *work)
1825 struct pool_workqueue *pwq = get_work_pwq(work);
1826 struct workqueue_struct *wq = pwq->wq;
1828 lockdep_assert_held(&workqueue_lock);
1830 if (!(wq->flags & WQ_RESCUER))
1833 /* mayday mayday mayday */
1834 if (list_empty(&pwq->mayday_node)) {
1835 list_add_tail(&pwq->mayday_node, &wq->maydays);
1836 wake_up_process(wq->rescuer->task);
1840 static void pool_mayday_timeout(unsigned long __pool)
1842 struct worker_pool *pool = (void *)__pool;
1843 struct work_struct *work;
1845 spin_lock_irq(&workqueue_lock); /* for wq->maydays */
1846 spin_lock(&pool->lock);
1848 if (need_to_create_worker(pool)) {
1850 * We've been trying to create a new worker but
1851 * haven't been successful. We might be hitting an
1852 * allocation deadlock. Send distress signals to
1855 list_for_each_entry(work, &pool->worklist, entry)
1859 spin_unlock(&pool->lock);
1860 spin_unlock_irq(&workqueue_lock);
1862 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1866 * maybe_create_worker - create a new worker if necessary
1867 * @pool: pool to create a new worker for
1869 * Create a new worker for @pool if necessary. @pool is guaranteed to
1870 * have at least one idle worker on return from this function. If
1871 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1872 * sent to all rescuers with works scheduled on @pool to resolve
1873 * possible allocation deadlock.
1875 * On return, need_to_create_worker() is guaranteed to be false and
1876 * may_start_working() true.
1879 * spin_lock_irq(pool->lock) which may be released and regrabbed
1880 * multiple times. Does GFP_KERNEL allocations. Called only from
1884 * false if no action was taken and pool->lock stayed locked, true
1887 static bool maybe_create_worker(struct worker_pool *pool)
1888 __releases(&pool->lock)
1889 __acquires(&pool->lock)
1891 if (!need_to_create_worker(pool))
1894 spin_unlock_irq(&pool->lock);
1896 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1897 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1900 struct worker *worker;
1902 worker = create_worker(pool);
1904 del_timer_sync(&pool->mayday_timer);
1905 spin_lock_irq(&pool->lock);
1906 start_worker(worker);
1907 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1912 if (!need_to_create_worker(pool))
1915 __set_current_state(TASK_INTERRUPTIBLE);
1916 schedule_timeout(CREATE_COOLDOWN);
1918 if (!need_to_create_worker(pool))
1922 del_timer_sync(&pool->mayday_timer);
1923 spin_lock_irq(&pool->lock);
1924 if (need_to_create_worker(pool))
1930 * maybe_destroy_worker - destroy workers which have been idle for a while
1931 * @pool: pool to destroy workers for
1933 * Destroy @pool workers which have been idle for longer than
1934 * IDLE_WORKER_TIMEOUT.
1937 * spin_lock_irq(pool->lock) which may be released and regrabbed
1938 * multiple times. Called only from manager.
1941 * false if no action was taken and pool->lock stayed locked, true
1944 static bool maybe_destroy_workers(struct worker_pool *pool)
1948 while (too_many_workers(pool)) {
1949 struct worker *worker;
1950 unsigned long expires;
1952 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1953 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1955 if (time_before(jiffies, expires)) {
1956 mod_timer(&pool->idle_timer, expires);
1960 destroy_worker(worker);
1968 * manage_workers - manage worker pool
1971 * Assume the manager role and manage the worker pool @worker belongs
1972 * to. At any given time, there can be only zero or one manager per
1973 * pool. The exclusion is handled automatically by this function.
1975 * The caller can safely start processing works on false return. On
1976 * true return, it's guaranteed that need_to_create_worker() is false
1977 * and may_start_working() is true.
1980 * spin_lock_irq(pool->lock) which may be released and regrabbed
1981 * multiple times. Does GFP_KERNEL allocations.
1984 * spin_lock_irq(pool->lock) which may be released and regrabbed
1985 * multiple times. Does GFP_KERNEL allocations.
1987 static bool manage_workers(struct worker *worker)
1989 struct worker_pool *pool = worker->pool;
1992 if (!mutex_trylock(&pool->manager_arb))
1996 * To simplify both worker management and CPU hotplug, hold off
1997 * management while hotplug is in progress. CPU hotplug path can't
1998 * grab @pool->manager_arb to achieve this because that can lead to
1999 * idle worker depletion (all become busy thinking someone else is
2000 * managing) which in turn can result in deadlock under extreme
2001 * circumstances. Use @pool->assoc_mutex to synchronize manager
2002 * against CPU hotplug.
2004 * assoc_mutex would always be free unless CPU hotplug is in
2005 * progress. trylock first without dropping @pool->lock.
2007 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2008 spin_unlock_irq(&pool->lock);
2009 mutex_lock(&pool->assoc_mutex);
2011 * CPU hotplug could have happened while we were waiting
2012 * for assoc_mutex. Hotplug itself can't handle us
2013 * because manager isn't either on idle or busy list, and
2014 * @pool's state and ours could have deviated.
2016 * As hotplug is now excluded via assoc_mutex, we can
2017 * simply try to bind. It will succeed or fail depending
2018 * on @pool's current state. Try it and adjust
2019 * %WORKER_UNBOUND accordingly.
2021 if (worker_maybe_bind_and_lock(pool))
2022 worker->flags &= ~WORKER_UNBOUND;
2024 worker->flags |= WORKER_UNBOUND;
2029 pool->flags &= ~POOL_MANAGE_WORKERS;
2032 * Destroy and then create so that may_start_working() is true
2035 ret |= maybe_destroy_workers(pool);
2036 ret |= maybe_create_worker(pool);
2038 mutex_unlock(&pool->assoc_mutex);
2039 mutex_unlock(&pool->manager_arb);
2044 * process_one_work - process single work
2046 * @work: work to process
2048 * Process @work. This function contains all the logics necessary to
2049 * process a single work including synchronization against and
2050 * interaction with other workers on the same cpu, queueing and
2051 * flushing. As long as context requirement is met, any worker can
2052 * call this function to process a work.
2055 * spin_lock_irq(pool->lock) which is released and regrabbed.
2057 static void process_one_work(struct worker *worker, struct work_struct *work)
2058 __releases(&pool->lock)
2059 __acquires(&pool->lock)
2061 struct pool_workqueue *pwq = get_work_pwq(work);
2062 struct worker_pool *pool = worker->pool;
2063 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2065 struct worker *collision;
2066 #ifdef CONFIG_LOCKDEP
2068 * It is permissible to free the struct work_struct from
2069 * inside the function that is called from it, this we need to
2070 * take into account for lockdep too. To avoid bogus "held
2071 * lock freed" warnings as well as problems when looking into
2072 * work->lockdep_map, make a copy and use that here.
2074 struct lockdep_map lockdep_map;
2076 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2079 * Ensure we're on the correct CPU. DISASSOCIATED test is
2080 * necessary to avoid spurious warnings from rescuers servicing the
2081 * unbound or a disassociated pool.
2083 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2084 !(pool->flags & POOL_DISASSOCIATED) &&
2085 raw_smp_processor_id() != pool->cpu);
2088 * A single work shouldn't be executed concurrently by
2089 * multiple workers on a single cpu. Check whether anyone is
2090 * already processing the work. If so, defer the work to the
2091 * currently executing one.
2093 collision = find_worker_executing_work(pool, work);
2094 if (unlikely(collision)) {
2095 move_linked_works(work, &collision->scheduled, NULL);
2099 /* claim and dequeue */
2100 debug_work_deactivate(work);
2101 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2102 worker->current_work = work;
2103 worker->current_func = work->func;
2104 worker->current_pwq = pwq;
2105 work_color = get_work_color(work);
2107 list_del_init(&work->entry);
2110 * CPU intensive works don't participate in concurrency
2111 * management. They're the scheduler's responsibility.
2113 if (unlikely(cpu_intensive))
2114 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2117 * Unbound pool isn't concurrency managed and work items should be
2118 * executed ASAP. Wake up another worker if necessary.
2120 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2121 wake_up_worker(pool);
2124 * Record the last pool and clear PENDING which should be the last
2125 * update to @work. Also, do this inside @pool->lock so that
2126 * PENDING and queued state changes happen together while IRQ is
2129 set_work_pool_and_clear_pending(work, pool->id);
2131 spin_unlock_irq(&pool->lock);
2133 lock_map_acquire_read(&pwq->wq->lockdep_map);
2134 lock_map_acquire(&lockdep_map);
2135 trace_workqueue_execute_start(work);
2136 worker->current_func(work);
2138 * While we must be careful to not use "work" after this, the trace
2139 * point will only record its address.
2141 trace_workqueue_execute_end(work);
2142 lock_map_release(&lockdep_map);
2143 lock_map_release(&pwq->wq->lockdep_map);
2145 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2146 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2147 " last function: %pf\n",
2148 current->comm, preempt_count(), task_pid_nr(current),
2149 worker->current_func);
2150 debug_show_held_locks(current);
2154 spin_lock_irq(&pool->lock);
2156 /* clear cpu intensive status */
2157 if (unlikely(cpu_intensive))
2158 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2160 /* we're done with it, release */
2161 hash_del(&worker->hentry);
2162 worker->current_work = NULL;
2163 worker->current_func = NULL;
2164 worker->current_pwq = NULL;
2165 pwq_dec_nr_in_flight(pwq, work_color);
2169 * process_scheduled_works - process scheduled works
2172 * Process all scheduled works. Please note that the scheduled list
2173 * may change while processing a work, so this function repeatedly
2174 * fetches a work from the top and executes it.
2177 * spin_lock_irq(pool->lock) which may be released and regrabbed
2180 static void process_scheduled_works(struct worker *worker)
2182 while (!list_empty(&worker->scheduled)) {
2183 struct work_struct *work = list_first_entry(&worker->scheduled,
2184 struct work_struct, entry);
2185 process_one_work(worker, work);
2190 * worker_thread - the worker thread function
2193 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2194 * of these per each cpu. These workers process all works regardless of
2195 * their specific target workqueue. The only exception is works which
2196 * belong to workqueues with a rescuer which will be explained in
2199 static int worker_thread(void *__worker)
2201 struct worker *worker = __worker;
2202 struct worker_pool *pool = worker->pool;
2204 /* tell the scheduler that this is a workqueue worker */
2205 worker->task->flags |= PF_WQ_WORKER;
2207 spin_lock_irq(&pool->lock);
2209 /* we are off idle list if destruction or rebind is requested */
2210 if (unlikely(list_empty(&worker->entry))) {
2211 spin_unlock_irq(&pool->lock);
2213 /* if DIE is set, destruction is requested */
2214 if (worker->flags & WORKER_DIE) {
2215 worker->task->flags &= ~PF_WQ_WORKER;
2219 /* otherwise, rebind */
2220 idle_worker_rebind(worker);
2224 worker_leave_idle(worker);
2226 /* no more worker necessary? */
2227 if (!need_more_worker(pool))
2230 /* do we need to manage? */
2231 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2235 * ->scheduled list can only be filled while a worker is
2236 * preparing to process a work or actually processing it.
2237 * Make sure nobody diddled with it while I was sleeping.
2239 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2242 * When control reaches this point, we're guaranteed to have
2243 * at least one idle worker or that someone else has already
2244 * assumed the manager role.
2246 worker_clr_flags(worker, WORKER_PREP);
2249 struct work_struct *work =
2250 list_first_entry(&pool->worklist,
2251 struct work_struct, entry);
2253 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2254 /* optimization path, not strictly necessary */
2255 process_one_work(worker, work);
2256 if (unlikely(!list_empty(&worker->scheduled)))
2257 process_scheduled_works(worker);
2259 move_linked_works(work, &worker->scheduled, NULL);
2260 process_scheduled_works(worker);
2262 } while (keep_working(pool));
2264 worker_set_flags(worker, WORKER_PREP, false);
2266 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2270 * pool->lock is held and there's no work to process and no need to
2271 * manage, sleep. Workers are woken up only while holding
2272 * pool->lock or from local cpu, so setting the current state
2273 * before releasing pool->lock is enough to prevent losing any
2276 worker_enter_idle(worker);
2277 __set_current_state(TASK_INTERRUPTIBLE);
2278 spin_unlock_irq(&pool->lock);
2284 * rescuer_thread - the rescuer thread function
2287 * Workqueue rescuer thread function. There's one rescuer for each
2288 * workqueue which has WQ_RESCUER set.
2290 * Regular work processing on a pool may block trying to create a new
2291 * worker which uses GFP_KERNEL allocation which has slight chance of
2292 * developing into deadlock if some works currently on the same queue
2293 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2294 * the problem rescuer solves.
2296 * When such condition is possible, the pool summons rescuers of all
2297 * workqueues which have works queued on the pool and let them process
2298 * those works so that forward progress can be guaranteed.
2300 * This should happen rarely.
2302 static int rescuer_thread(void *__rescuer)
2304 struct worker *rescuer = __rescuer;
2305 struct workqueue_struct *wq = rescuer->rescue_wq;
2306 struct list_head *scheduled = &rescuer->scheduled;
2308 set_user_nice(current, RESCUER_NICE_LEVEL);
2311 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2312 * doesn't participate in concurrency management.
2314 rescuer->task->flags |= PF_WQ_WORKER;
2316 set_current_state(TASK_INTERRUPTIBLE);
2318 if (kthread_should_stop()) {
2319 __set_current_state(TASK_RUNNING);
2320 rescuer->task->flags &= ~PF_WQ_WORKER;
2324 /* see whether any pwq is asking for help */
2325 spin_lock_irq(&workqueue_lock);
2327 while (!list_empty(&wq->maydays)) {
2328 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2329 struct pool_workqueue, mayday_node);
2330 struct worker_pool *pool = pwq->pool;
2331 struct work_struct *work, *n;
2333 __set_current_state(TASK_RUNNING);
2334 list_del_init(&pwq->mayday_node);
2336 spin_unlock_irq(&workqueue_lock);
2338 /* migrate to the target cpu if possible */
2339 worker_maybe_bind_and_lock(pool);
2340 rescuer->pool = pool;
2343 * Slurp in all works issued via this workqueue and
2346 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2347 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2348 if (get_work_pwq(work) == pwq)
2349 move_linked_works(work, scheduled, &n);
2351 process_scheduled_works(rescuer);
2354 * Leave this pool. If keep_working() is %true, notify a
2355 * regular worker; otherwise, we end up with 0 concurrency
2356 * and stalling the execution.
2358 if (keep_working(pool))
2359 wake_up_worker(pool);
2361 rescuer->pool = NULL;
2362 spin_unlock(&pool->lock);
2363 spin_lock(&workqueue_lock);
2366 spin_unlock_irq(&workqueue_lock);
2368 /* rescuers should never participate in concurrency management */
2369 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2375 struct work_struct work;
2376 struct completion done;
2379 static void wq_barrier_func(struct work_struct *work)
2381 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2382 complete(&barr->done);
2386 * insert_wq_barrier - insert a barrier work
2387 * @pwq: pwq to insert barrier into
2388 * @barr: wq_barrier to insert
2389 * @target: target work to attach @barr to
2390 * @worker: worker currently executing @target, NULL if @target is not executing
2392 * @barr is linked to @target such that @barr is completed only after
2393 * @target finishes execution. Please note that the ordering
2394 * guarantee is observed only with respect to @target and on the local
2397 * Currently, a queued barrier can't be canceled. This is because
2398 * try_to_grab_pending() can't determine whether the work to be
2399 * grabbed is at the head of the queue and thus can't clear LINKED
2400 * flag of the previous work while there must be a valid next work
2401 * after a work with LINKED flag set.
2403 * Note that when @worker is non-NULL, @target may be modified
2404 * underneath us, so we can't reliably determine pwq from @target.
2407 * spin_lock_irq(pool->lock).
2409 static void insert_wq_barrier(struct pool_workqueue *pwq,
2410 struct wq_barrier *barr,
2411 struct work_struct *target, struct worker *worker)
2413 struct list_head *head;
2414 unsigned int linked = 0;
2417 * debugobject calls are safe here even with pool->lock locked
2418 * as we know for sure that this will not trigger any of the
2419 * checks and call back into the fixup functions where we
2422 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2423 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2424 init_completion(&barr->done);
2427 * If @target is currently being executed, schedule the
2428 * barrier to the worker; otherwise, put it after @target.
2431 head = worker->scheduled.next;
2433 unsigned long *bits = work_data_bits(target);
2435 head = target->entry.next;
2436 /* there can already be other linked works, inherit and set */
2437 linked = *bits & WORK_STRUCT_LINKED;
2438 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2441 debug_work_activate(&barr->work);
2442 insert_work(pwq, &barr->work, head,
2443 work_color_to_flags(WORK_NO_COLOR) | linked);
2447 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2448 * @wq: workqueue being flushed
2449 * @flush_color: new flush color, < 0 for no-op
2450 * @work_color: new work color, < 0 for no-op
2452 * Prepare pwqs for workqueue flushing.
2454 * If @flush_color is non-negative, flush_color on all pwqs should be
2455 * -1. If no pwq has in-flight commands at the specified color, all
2456 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2457 * has in flight commands, its pwq->flush_color is set to
2458 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2459 * wakeup logic is armed and %true is returned.
2461 * The caller should have initialized @wq->first_flusher prior to
2462 * calling this function with non-negative @flush_color. If
2463 * @flush_color is negative, no flush color update is done and %false
2466 * If @work_color is non-negative, all pwqs should have the same
2467 * work_color which is previous to @work_color and all will be
2468 * advanced to @work_color.
2471 * mutex_lock(wq->flush_mutex).
2474 * %true if @flush_color >= 0 and there's something to flush. %false
2477 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2478 int flush_color, int work_color)
2481 struct pool_workqueue *pwq;
2483 if (flush_color >= 0) {
2484 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2485 atomic_set(&wq->nr_pwqs_to_flush, 1);
2488 local_irq_disable();
2490 for_each_pwq(pwq, wq) {
2491 struct worker_pool *pool = pwq->pool;
2493 spin_lock(&pool->lock);
2495 if (flush_color >= 0) {
2496 WARN_ON_ONCE(pwq->flush_color != -1);
2498 if (pwq->nr_in_flight[flush_color]) {
2499 pwq->flush_color = flush_color;
2500 atomic_inc(&wq->nr_pwqs_to_flush);
2505 if (work_color >= 0) {
2506 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2507 pwq->work_color = work_color;
2510 spin_unlock(&pool->lock);
2515 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2516 complete(&wq->first_flusher->done);
2522 * flush_workqueue - ensure that any scheduled work has run to completion.
2523 * @wq: workqueue to flush
2525 * Forces execution of the workqueue and blocks until its completion.
2526 * This is typically used in driver shutdown handlers.
2528 * We sleep until all works which were queued on entry have been handled,
2529 * but we are not livelocked by new incoming ones.
2531 void flush_workqueue(struct workqueue_struct *wq)
2533 struct wq_flusher this_flusher = {
2534 .list = LIST_HEAD_INIT(this_flusher.list),
2536 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2540 lock_map_acquire(&wq->lockdep_map);
2541 lock_map_release(&wq->lockdep_map);
2543 mutex_lock(&wq->flush_mutex);
2546 * Start-to-wait phase
2548 next_color = work_next_color(wq->work_color);
2550 if (next_color != wq->flush_color) {
2552 * Color space is not full. The current work_color
2553 * becomes our flush_color and work_color is advanced
2556 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2557 this_flusher.flush_color = wq->work_color;
2558 wq->work_color = next_color;
2560 if (!wq->first_flusher) {
2561 /* no flush in progress, become the first flusher */
2562 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2564 wq->first_flusher = &this_flusher;
2566 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2568 /* nothing to flush, done */
2569 wq->flush_color = next_color;
2570 wq->first_flusher = NULL;
2575 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2576 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2577 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2581 * Oops, color space is full, wait on overflow queue.
2582 * The next flush completion will assign us
2583 * flush_color and transfer to flusher_queue.
2585 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2588 mutex_unlock(&wq->flush_mutex);
2590 wait_for_completion(&this_flusher.done);
2593 * Wake-up-and-cascade phase
2595 * First flushers are responsible for cascading flushes and
2596 * handling overflow. Non-first flushers can simply return.
2598 if (wq->first_flusher != &this_flusher)
2601 mutex_lock(&wq->flush_mutex);
2603 /* we might have raced, check again with mutex held */
2604 if (wq->first_flusher != &this_flusher)
2607 wq->first_flusher = NULL;
2609 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2610 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2613 struct wq_flusher *next, *tmp;
2615 /* complete all the flushers sharing the current flush color */
2616 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2617 if (next->flush_color != wq->flush_color)
2619 list_del_init(&next->list);
2620 complete(&next->done);
2623 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2624 wq->flush_color != work_next_color(wq->work_color));
2626 /* this flush_color is finished, advance by one */
2627 wq->flush_color = work_next_color(wq->flush_color);
2629 /* one color has been freed, handle overflow queue */
2630 if (!list_empty(&wq->flusher_overflow)) {
2632 * Assign the same color to all overflowed
2633 * flushers, advance work_color and append to
2634 * flusher_queue. This is the start-to-wait
2635 * phase for these overflowed flushers.
2637 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2638 tmp->flush_color = wq->work_color;
2640 wq->work_color = work_next_color(wq->work_color);
2642 list_splice_tail_init(&wq->flusher_overflow,
2643 &wq->flusher_queue);
2644 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2647 if (list_empty(&wq->flusher_queue)) {
2648 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2653 * Need to flush more colors. Make the next flusher
2654 * the new first flusher and arm pwqs.
2656 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2657 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2659 list_del_init(&next->list);
2660 wq->first_flusher = next;
2662 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2666 * Meh... this color is already done, clear first
2667 * flusher and repeat cascading.
2669 wq->first_flusher = NULL;
2673 mutex_unlock(&wq->flush_mutex);
2675 EXPORT_SYMBOL_GPL(flush_workqueue);
2678 * drain_workqueue - drain a workqueue
2679 * @wq: workqueue to drain
2681 * Wait until the workqueue becomes empty. While draining is in progress,
2682 * only chain queueing is allowed. IOW, only currently pending or running
2683 * work items on @wq can queue further work items on it. @wq is flushed
2684 * repeatedly until it becomes empty. The number of flushing is detemined
2685 * by the depth of chaining and should be relatively short. Whine if it
2688 void drain_workqueue(struct workqueue_struct *wq)
2690 unsigned int flush_cnt = 0;
2691 struct pool_workqueue *pwq;
2694 * __queue_work() needs to test whether there are drainers, is much
2695 * hotter than drain_workqueue() and already looks at @wq->flags.
2696 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2698 spin_lock_irq(&workqueue_lock);
2699 if (!wq->nr_drainers++)
2700 wq->flags |= WQ_DRAINING;
2701 spin_unlock_irq(&workqueue_lock);
2703 flush_workqueue(wq);
2705 local_irq_disable();
2707 for_each_pwq(pwq, wq) {
2710 spin_lock(&pwq->pool->lock);
2711 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2712 spin_unlock(&pwq->pool->lock);
2717 if (++flush_cnt == 10 ||
2718 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2719 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2720 wq->name, flush_cnt);
2726 spin_lock(&workqueue_lock);
2727 if (!--wq->nr_drainers)
2728 wq->flags &= ~WQ_DRAINING;
2729 spin_unlock(&workqueue_lock);
2733 EXPORT_SYMBOL_GPL(drain_workqueue);
2735 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2737 struct worker *worker = NULL;
2738 struct worker_pool *pool;
2739 struct pool_workqueue *pwq;
2743 local_irq_disable();
2744 pool = get_work_pool(work);
2750 spin_lock(&pool->lock);
2751 /* see the comment in try_to_grab_pending() with the same code */
2752 pwq = get_work_pwq(work);
2754 if (unlikely(pwq->pool != pool))
2757 worker = find_worker_executing_work(pool, work);
2760 pwq = worker->current_pwq;
2763 insert_wq_barrier(pwq, barr, work, worker);
2764 spin_unlock_irq(&pool->lock);
2767 * If @max_active is 1 or rescuer is in use, flushing another work
2768 * item on the same workqueue may lead to deadlock. Make sure the
2769 * flusher is not running on the same workqueue by verifying write
2772 if (pwq->wq->saved_max_active == 1 || pwq->wq->flags & WQ_RESCUER)
2773 lock_map_acquire(&pwq->wq->lockdep_map);
2775 lock_map_acquire_read(&pwq->wq->lockdep_map);
2776 lock_map_release(&pwq->wq->lockdep_map);
2780 spin_unlock_irq(&pool->lock);
2785 * flush_work - wait for a work to finish executing the last queueing instance
2786 * @work: the work to flush
2788 * Wait until @work has finished execution. @work is guaranteed to be idle
2789 * on return if it hasn't been requeued since flush started.
2792 * %true if flush_work() waited for the work to finish execution,
2793 * %false if it was already idle.
2795 bool flush_work(struct work_struct *work)
2797 struct wq_barrier barr;
2799 lock_map_acquire(&work->lockdep_map);
2800 lock_map_release(&work->lockdep_map);
2802 if (start_flush_work(work, &barr)) {
2803 wait_for_completion(&barr.done);
2804 destroy_work_on_stack(&barr.work);
2810 EXPORT_SYMBOL_GPL(flush_work);
2812 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2814 unsigned long flags;
2818 ret = try_to_grab_pending(work, is_dwork, &flags);
2820 * If someone else is canceling, wait for the same event it
2821 * would be waiting for before retrying.
2823 if (unlikely(ret == -ENOENT))
2825 } while (unlikely(ret < 0));
2827 /* tell other tasks trying to grab @work to back off */
2828 mark_work_canceling(work);
2829 local_irq_restore(flags);
2832 clear_work_data(work);
2837 * cancel_work_sync - cancel a work and wait for it to finish
2838 * @work: the work to cancel
2840 * Cancel @work and wait for its execution to finish. This function
2841 * can be used even if the work re-queues itself or migrates to
2842 * another workqueue. On return from this function, @work is
2843 * guaranteed to be not pending or executing on any CPU.
2845 * cancel_work_sync(&delayed_work->work) must not be used for
2846 * delayed_work's. Use cancel_delayed_work_sync() instead.
2848 * The caller must ensure that the workqueue on which @work was last
2849 * queued can't be destroyed before this function returns.
2852 * %true if @work was pending, %false otherwise.
2854 bool cancel_work_sync(struct work_struct *work)
2856 return __cancel_work_timer(work, false);
2858 EXPORT_SYMBOL_GPL(cancel_work_sync);
2861 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2862 * @dwork: the delayed work to flush
2864 * Delayed timer is cancelled and the pending work is queued for
2865 * immediate execution. Like flush_work(), this function only
2866 * considers the last queueing instance of @dwork.
2869 * %true if flush_work() waited for the work to finish execution,
2870 * %false if it was already idle.
2872 bool flush_delayed_work(struct delayed_work *dwork)
2874 local_irq_disable();
2875 if (del_timer_sync(&dwork->timer))
2876 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2878 return flush_work(&dwork->work);
2880 EXPORT_SYMBOL(flush_delayed_work);
2883 * cancel_delayed_work - cancel a delayed work
2884 * @dwork: delayed_work to cancel
2886 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2887 * and canceled; %false if wasn't pending. Note that the work callback
2888 * function may still be running on return, unless it returns %true and the
2889 * work doesn't re-arm itself. Explicitly flush or use
2890 * cancel_delayed_work_sync() to wait on it.
2892 * This function is safe to call from any context including IRQ handler.
2894 bool cancel_delayed_work(struct delayed_work *dwork)
2896 unsigned long flags;
2900 ret = try_to_grab_pending(&dwork->work, true, &flags);
2901 } while (unlikely(ret == -EAGAIN));
2903 if (unlikely(ret < 0))
2906 set_work_pool_and_clear_pending(&dwork->work,
2907 get_work_pool_id(&dwork->work));
2908 local_irq_restore(flags);
2911 EXPORT_SYMBOL(cancel_delayed_work);
2914 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2915 * @dwork: the delayed work cancel
2917 * This is cancel_work_sync() for delayed works.
2920 * %true if @dwork was pending, %false otherwise.
2922 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2924 return __cancel_work_timer(&dwork->work, true);
2926 EXPORT_SYMBOL(cancel_delayed_work_sync);
2929 * schedule_work_on - put work task on a specific cpu
2930 * @cpu: cpu to put the work task on
2931 * @work: job to be done
2933 * This puts a job on a specific cpu
2935 bool schedule_work_on(int cpu, struct work_struct *work)
2937 return queue_work_on(cpu, system_wq, work);
2939 EXPORT_SYMBOL(schedule_work_on);
2942 * schedule_work - put work task in global workqueue
2943 * @work: job to be done
2945 * Returns %false if @work was already on the kernel-global workqueue and
2948 * This puts a job in the kernel-global workqueue if it was not already
2949 * queued and leaves it in the same position on the kernel-global
2950 * workqueue otherwise.
2952 bool schedule_work(struct work_struct *work)
2954 return queue_work(system_wq, work);
2956 EXPORT_SYMBOL(schedule_work);
2959 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2961 * @dwork: job to be done
2962 * @delay: number of jiffies to wait
2964 * After waiting for a given time this puts a job in the kernel-global
2965 * workqueue on the specified CPU.
2967 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
2968 unsigned long delay)
2970 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2972 EXPORT_SYMBOL(schedule_delayed_work_on);
2975 * schedule_delayed_work - put work task in global workqueue after delay
2976 * @dwork: job to be done
2977 * @delay: number of jiffies to wait or 0 for immediate execution
2979 * After waiting for a given time this puts a job in the kernel-global
2982 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
2984 return queue_delayed_work(system_wq, dwork, delay);
2986 EXPORT_SYMBOL(schedule_delayed_work);
2989 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2990 * @func: the function to call
2992 * schedule_on_each_cpu() executes @func on each online CPU using the
2993 * system workqueue and blocks until all CPUs have completed.
2994 * schedule_on_each_cpu() is very slow.
2997 * 0 on success, -errno on failure.
2999 int schedule_on_each_cpu(work_func_t func)
3002 struct work_struct __percpu *works;
3004 works = alloc_percpu(struct work_struct);
3010 for_each_online_cpu(cpu) {
3011 struct work_struct *work = per_cpu_ptr(works, cpu);
3013 INIT_WORK(work, func);
3014 schedule_work_on(cpu, work);
3017 for_each_online_cpu(cpu)
3018 flush_work(per_cpu_ptr(works, cpu));
3026 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3028 * Forces execution of the kernel-global workqueue and blocks until its
3031 * Think twice before calling this function! It's very easy to get into
3032 * trouble if you don't take great care. Either of the following situations
3033 * will lead to deadlock:
3035 * One of the work items currently on the workqueue needs to acquire
3036 * a lock held by your code or its caller.
3038 * Your code is running in the context of a work routine.
3040 * They will be detected by lockdep when they occur, but the first might not
3041 * occur very often. It depends on what work items are on the workqueue and
3042 * what locks they need, which you have no control over.
3044 * In most situations flushing the entire workqueue is overkill; you merely
3045 * need to know that a particular work item isn't queued and isn't running.
3046 * In such cases you should use cancel_delayed_work_sync() or
3047 * cancel_work_sync() instead.
3049 void flush_scheduled_work(void)
3051 flush_workqueue(system_wq);
3053 EXPORT_SYMBOL(flush_scheduled_work);
3056 * execute_in_process_context - reliably execute the routine with user context
3057 * @fn: the function to execute
3058 * @ew: guaranteed storage for the execute work structure (must
3059 * be available when the work executes)
3061 * Executes the function immediately if process context is available,
3062 * otherwise schedules the function for delayed execution.
3064 * Returns: 0 - function was executed
3065 * 1 - function was scheduled for execution
3067 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3069 if (!in_interrupt()) {
3074 INIT_WORK(&ew->work, fn);
3075 schedule_work(&ew->work);
3079 EXPORT_SYMBOL_GPL(execute_in_process_context);
3081 int keventd_up(void)
3083 return system_wq != NULL;
3087 * free_workqueue_attrs - free a workqueue_attrs
3088 * @attrs: workqueue_attrs to free
3090 * Undo alloc_workqueue_attrs().
3092 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3095 free_cpumask_var(attrs->cpumask);
3101 * alloc_workqueue_attrs - allocate a workqueue_attrs
3102 * @gfp_mask: allocation mask to use
3104 * Allocate a new workqueue_attrs, initialize with default settings and
3105 * return it. Returns NULL on failure.
3107 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3109 struct workqueue_attrs *attrs;
3111 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3114 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3117 cpumask_setall(attrs->cpumask);
3120 free_workqueue_attrs(attrs);
3124 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3125 const struct workqueue_attrs *from)
3127 to->nice = from->nice;
3128 cpumask_copy(to->cpumask, from->cpumask);
3132 * Hacky implementation of jhash of bitmaps which only considers the
3133 * specified number of bits. We probably want a proper implementation in
3134 * include/linux/jhash.h.
3136 static u32 jhash_bitmap(const unsigned long *bitmap, int bits, u32 hash)
3138 int nr_longs = bits / BITS_PER_LONG;
3139 int nr_leftover = bits % BITS_PER_LONG;
3140 unsigned long leftover = 0;
3143 hash = jhash(bitmap, nr_longs * sizeof(long), hash);
3145 bitmap_copy(&leftover, bitmap + nr_longs, nr_leftover);
3146 hash = jhash(&leftover, sizeof(long), hash);
3151 /* hash value of the content of @attr */
3152 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3156 hash = jhash_1word(attrs->nice, hash);
3157 hash = jhash_bitmap(cpumask_bits(attrs->cpumask), nr_cpu_ids, hash);
3161 /* content equality test */
3162 static bool wqattrs_equal(const struct workqueue_attrs *a,
3163 const struct workqueue_attrs *b)
3165 if (a->nice != b->nice)
3167 if (!cpumask_equal(a->cpumask, b->cpumask))
3173 * init_worker_pool - initialize a newly zalloc'd worker_pool
3174 * @pool: worker_pool to initialize
3176 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3177 * Returns 0 on success, -errno on failure. Even on failure, all fields
3178 * inside @pool proper are initialized and put_unbound_pool() can be called
3179 * on @pool safely to release it.
3181 static int init_worker_pool(struct worker_pool *pool)
3183 spin_lock_init(&pool->lock);
3186 pool->flags |= POOL_DISASSOCIATED;
3187 INIT_LIST_HEAD(&pool->worklist);
3188 INIT_LIST_HEAD(&pool->idle_list);
3189 hash_init(pool->busy_hash);
3191 init_timer_deferrable(&pool->idle_timer);
3192 pool->idle_timer.function = idle_worker_timeout;
3193 pool->idle_timer.data = (unsigned long)pool;
3195 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3196 (unsigned long)pool);
3198 mutex_init(&pool->manager_arb);
3199 mutex_init(&pool->assoc_mutex);
3200 ida_init(&pool->worker_ida);
3202 INIT_HLIST_NODE(&pool->hash_node);
3205 /* shouldn't fail above this point */
3206 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3212 static void rcu_free_pool(struct rcu_head *rcu)
3214 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3216 ida_destroy(&pool->worker_ida);
3217 free_workqueue_attrs(pool->attrs);
3222 * put_unbound_pool - put a worker_pool
3223 * @pool: worker_pool to put
3225 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3228 static void put_unbound_pool(struct worker_pool *pool)
3230 struct worker *worker;
3232 spin_lock_irq(&workqueue_lock);
3233 if (--pool->refcnt) {
3234 spin_unlock_irq(&workqueue_lock);
3239 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3240 WARN_ON(!list_empty(&pool->worklist))) {
3241 spin_unlock_irq(&workqueue_lock);
3245 /* release id and unhash */
3247 idr_remove(&worker_pool_idr, pool->id);
3248 hash_del(&pool->hash_node);
3250 spin_unlock_irq(&workqueue_lock);
3252 /* lock out manager and destroy all workers */
3253 mutex_lock(&pool->manager_arb);
3254 spin_lock_irq(&pool->lock);
3256 while ((worker = first_worker(pool)))
3257 destroy_worker(worker);
3258 WARN_ON(pool->nr_workers || pool->nr_idle);
3260 spin_unlock_irq(&pool->lock);
3261 mutex_unlock(&pool->manager_arb);
3263 /* shut down the timers */
3264 del_timer_sync(&pool->idle_timer);
3265 del_timer_sync(&pool->mayday_timer);
3267 /* sched-RCU protected to allow dereferences from get_work_pool() */
3268 call_rcu_sched(&pool->rcu, rcu_free_pool);
3272 * get_unbound_pool - get a worker_pool with the specified attributes
3273 * @attrs: the attributes of the worker_pool to get
3275 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3276 * reference count and return it. If there already is a matching
3277 * worker_pool, it will be used; otherwise, this function attempts to
3278 * create a new one. On failure, returns NULL.
3280 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3282 static DEFINE_MUTEX(create_mutex);
3283 u32 hash = wqattrs_hash(attrs);
3284 struct worker_pool *pool;
3285 struct worker *worker;
3287 mutex_lock(&create_mutex);
3289 /* do we already have a matching pool? */
3290 spin_lock_irq(&workqueue_lock);
3291 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3292 if (wqattrs_equal(pool->attrs, attrs)) {
3297 spin_unlock_irq(&workqueue_lock);
3299 /* nope, create a new one */
3300 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3301 if (!pool || init_worker_pool(pool) < 0)
3304 copy_workqueue_attrs(pool->attrs, attrs);
3306 if (worker_pool_assign_id(pool) < 0)
3309 /* create and start the initial worker */
3310 worker = create_worker(pool);
3314 spin_lock_irq(&pool->lock);
3315 start_worker(worker);
3316 spin_unlock_irq(&pool->lock);
3319 spin_lock_irq(&workqueue_lock);
3320 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3322 spin_unlock_irq(&workqueue_lock);
3323 mutex_unlock(&create_mutex);
3326 mutex_unlock(&create_mutex);
3328 put_unbound_pool(pool);
3332 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3334 bool highpri = wq->flags & WQ_HIGHPRI;
3337 if (!(wq->flags & WQ_UNBOUND)) {
3338 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3342 for_each_possible_cpu(cpu) {
3343 struct pool_workqueue *pwq =
3344 per_cpu_ptr(wq->cpu_pwqs, cpu);
3345 struct worker_pool *cpu_pools =
3346 per_cpu(cpu_worker_pools, cpu);
3348 pwq->pool = &cpu_pools[highpri];
3349 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs);
3352 struct pool_workqueue *pwq;
3354 pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL);
3358 pwq->pool = get_unbound_pool(unbound_std_wq_attrs[highpri]);
3360 kmem_cache_free(pwq_cache, pwq);
3364 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs);
3370 static void free_pwqs(struct workqueue_struct *wq)
3372 if (!(wq->flags & WQ_UNBOUND))
3373 free_percpu(wq->cpu_pwqs);
3374 else if (!list_empty(&wq->pwqs))
3375 kmem_cache_free(pwq_cache, list_first_entry(&wq->pwqs,
3376 struct pool_workqueue, pwqs_node));
3379 static int wq_clamp_max_active(int max_active, unsigned int flags,
3382 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3384 if (max_active < 1 || max_active > lim)
3385 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3386 max_active, name, 1, lim);
3388 return clamp_val(max_active, 1, lim);
3391 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3394 struct lock_class_key *key,
3395 const char *lock_name, ...)
3397 va_list args, args1;
3398 struct workqueue_struct *wq;
3399 struct pool_workqueue *pwq;
3402 /* determine namelen, allocate wq and format name */
3403 va_start(args, lock_name);
3404 va_copy(args1, args);
3405 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3407 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3411 vsnprintf(wq->name, namelen, fmt, args1);
3416 * Workqueues which may be used during memory reclaim should
3417 * have a rescuer to guarantee forward progress.
3419 if (flags & WQ_MEM_RECLAIM)
3420 flags |= WQ_RESCUER;
3422 max_active = max_active ?: WQ_DFL_ACTIVE;
3423 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3427 wq->saved_max_active = max_active;
3428 mutex_init(&wq->flush_mutex);
3429 atomic_set(&wq->nr_pwqs_to_flush, 0);
3430 INIT_LIST_HEAD(&wq->pwqs);
3431 INIT_LIST_HEAD(&wq->flusher_queue);
3432 INIT_LIST_HEAD(&wq->flusher_overflow);
3433 INIT_LIST_HEAD(&wq->maydays);
3435 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3436 INIT_LIST_HEAD(&wq->list);
3438 if (alloc_and_link_pwqs(wq) < 0)
3441 local_irq_disable();
3442 for_each_pwq(pwq, wq) {
3443 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3445 pwq->flush_color = -1;
3446 pwq->max_active = max_active;
3447 INIT_LIST_HEAD(&pwq->delayed_works);
3448 INIT_LIST_HEAD(&pwq->mayday_node);
3452 if (flags & WQ_RESCUER) {
3453 struct worker *rescuer;
3455 wq->rescuer = rescuer = alloc_worker();
3459 rescuer->rescue_wq = wq;
3460 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3462 if (IS_ERR(rescuer->task))
3465 rescuer->task->flags |= PF_THREAD_BOUND;
3466 wake_up_process(rescuer->task);
3470 * workqueue_lock protects global freeze state and workqueues
3471 * list. Grab it, set max_active accordingly and add the new
3472 * workqueue to workqueues list.
3474 spin_lock_irq(&workqueue_lock);
3476 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3477 for_each_pwq(pwq, wq)
3478 pwq->max_active = 0;
3480 list_add(&wq->list, &workqueues);
3482 spin_unlock_irq(&workqueue_lock);
3493 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3496 * destroy_workqueue - safely terminate a workqueue
3497 * @wq: target workqueue
3499 * Safely destroy a workqueue. All work currently pending will be done first.
3501 void destroy_workqueue(struct workqueue_struct *wq)
3503 struct pool_workqueue *pwq;
3505 /* drain it before proceeding with destruction */
3506 drain_workqueue(wq);
3508 spin_lock_irq(&workqueue_lock);
3511 for_each_pwq(pwq, wq) {
3514 for (i = 0; i < WORK_NR_COLORS; i++) {
3515 if (WARN_ON(pwq->nr_in_flight[i])) {
3516 spin_unlock_irq(&workqueue_lock);
3521 if (WARN_ON(pwq->nr_active) ||
3522 WARN_ON(!list_empty(&pwq->delayed_works))) {
3523 spin_unlock_irq(&workqueue_lock);
3529 * wq list is used to freeze wq, remove from list after
3530 * flushing is complete in case freeze races us.
3532 list_del(&wq->list);
3534 spin_unlock_irq(&workqueue_lock);
3536 if (wq->flags & WQ_RESCUER) {
3537 kthread_stop(wq->rescuer->task);
3542 * We're the sole accessor of @wq at this point. Directly access
3543 * the first pwq and put its pool.
3545 if (wq->flags & WQ_UNBOUND) {
3546 pwq = list_first_entry(&wq->pwqs, struct pool_workqueue,
3548 put_unbound_pool(pwq->pool);
3553 EXPORT_SYMBOL_GPL(destroy_workqueue);
3556 * pwq_set_max_active - adjust max_active of a pwq
3557 * @pwq: target pool_workqueue
3558 * @max_active: new max_active value.
3560 * Set @pwq->max_active to @max_active and activate delayed works if
3564 * spin_lock_irq(pool->lock).
3566 static void pwq_set_max_active(struct pool_workqueue *pwq, int max_active)
3568 pwq->max_active = max_active;
3570 while (!list_empty(&pwq->delayed_works) &&
3571 pwq->nr_active < pwq->max_active)
3572 pwq_activate_first_delayed(pwq);
3576 * workqueue_set_max_active - adjust max_active of a workqueue
3577 * @wq: target workqueue
3578 * @max_active: new max_active value.
3580 * Set max_active of @wq to @max_active.
3583 * Don't call from IRQ context.
3585 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3587 struct pool_workqueue *pwq;
3589 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3591 spin_lock_irq(&workqueue_lock);
3593 wq->saved_max_active = max_active;
3595 for_each_pwq(pwq, wq) {
3596 struct worker_pool *pool = pwq->pool;
3598 spin_lock(&pool->lock);
3600 if (!(wq->flags & WQ_FREEZABLE) ||
3601 !(pool->flags & POOL_FREEZING))
3602 pwq_set_max_active(pwq, max_active);
3604 spin_unlock(&pool->lock);
3607 spin_unlock_irq(&workqueue_lock);
3609 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3612 * workqueue_congested - test whether a workqueue is congested
3613 * @cpu: CPU in question
3614 * @wq: target workqueue
3616 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3617 * no synchronization around this function and the test result is
3618 * unreliable and only useful as advisory hints or for debugging.
3621 * %true if congested, %false otherwise.
3623 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
3625 struct pool_workqueue *pwq;
3630 if (!(wq->flags & WQ_UNBOUND))
3631 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
3633 pwq = first_pwq(wq);
3635 ret = !list_empty(&pwq->delayed_works);
3640 EXPORT_SYMBOL_GPL(workqueue_congested);
3643 * work_busy - test whether a work is currently pending or running
3644 * @work: the work to be tested
3646 * Test whether @work is currently pending or running. There is no
3647 * synchronization around this function and the test result is
3648 * unreliable and only useful as advisory hints or for debugging.
3651 * OR'd bitmask of WORK_BUSY_* bits.
3653 unsigned int work_busy(struct work_struct *work)
3655 struct worker_pool *pool;
3656 unsigned long flags;
3657 unsigned int ret = 0;
3659 if (work_pending(work))
3660 ret |= WORK_BUSY_PENDING;
3662 local_irq_save(flags);
3663 pool = get_work_pool(work);
3665 spin_lock(&pool->lock);
3666 if (find_worker_executing_work(pool, work))
3667 ret |= WORK_BUSY_RUNNING;
3668 spin_unlock(&pool->lock);
3670 local_irq_restore(flags);
3674 EXPORT_SYMBOL_GPL(work_busy);
3679 * There are two challenges in supporting CPU hotplug. Firstly, there
3680 * are a lot of assumptions on strong associations among work, pwq and
3681 * pool which make migrating pending and scheduled works very
3682 * difficult to implement without impacting hot paths. Secondly,
3683 * worker pools serve mix of short, long and very long running works making
3684 * blocked draining impractical.
3686 * This is solved by allowing the pools to be disassociated from the CPU
3687 * running as an unbound one and allowing it to be reattached later if the
3688 * cpu comes back online.
3691 static void wq_unbind_fn(struct work_struct *work)
3693 int cpu = smp_processor_id();
3694 struct worker_pool *pool;
3695 struct worker *worker;
3698 for_each_cpu_worker_pool(pool, cpu) {
3699 WARN_ON_ONCE(cpu != smp_processor_id());
3701 mutex_lock(&pool->assoc_mutex);
3702 spin_lock_irq(&pool->lock);
3705 * We've claimed all manager positions. Make all workers
3706 * unbound and set DISASSOCIATED. Before this, all workers
3707 * except for the ones which are still executing works from
3708 * before the last CPU down must be on the cpu. After
3709 * this, they may become diasporas.
3711 list_for_each_entry(worker, &pool->idle_list, entry)
3712 worker->flags |= WORKER_UNBOUND;
3714 for_each_busy_worker(worker, i, pool)
3715 worker->flags |= WORKER_UNBOUND;
3717 pool->flags |= POOL_DISASSOCIATED;
3719 spin_unlock_irq(&pool->lock);
3720 mutex_unlock(&pool->assoc_mutex);
3724 * Call schedule() so that we cross rq->lock and thus can guarantee
3725 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3726 * as scheduler callbacks may be invoked from other cpus.
3731 * Sched callbacks are disabled now. Zap nr_running. After this,
3732 * nr_running stays zero and need_more_worker() and keep_working()
3733 * are always true as long as the worklist is not empty. Pools on
3734 * @cpu now behave as unbound (in terms of concurrency management)
3735 * pools which are served by workers tied to the CPU.
3737 * On return from this function, the current worker would trigger
3738 * unbound chain execution of pending work items if other workers
3741 for_each_cpu_worker_pool(pool, cpu)
3742 atomic_set(&pool->nr_running, 0);
3746 * Workqueues should be brought up before normal priority CPU notifiers.
3747 * This will be registered high priority CPU notifier.
3749 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3750 unsigned long action,
3753 int cpu = (unsigned long)hcpu;
3754 struct worker_pool *pool;
3756 switch (action & ~CPU_TASKS_FROZEN) {
3757 case CPU_UP_PREPARE:
3758 for_each_cpu_worker_pool(pool, cpu) {
3759 struct worker *worker;
3761 if (pool->nr_workers)
3764 worker = create_worker(pool);
3768 spin_lock_irq(&pool->lock);
3769 start_worker(worker);
3770 spin_unlock_irq(&pool->lock);
3774 case CPU_DOWN_FAILED:
3776 for_each_cpu_worker_pool(pool, cpu) {
3777 mutex_lock(&pool->assoc_mutex);
3778 spin_lock_irq(&pool->lock);
3780 pool->flags &= ~POOL_DISASSOCIATED;
3781 rebind_workers(pool);
3783 spin_unlock_irq(&pool->lock);
3784 mutex_unlock(&pool->assoc_mutex);
3792 * Workqueues should be brought down after normal priority CPU notifiers.
3793 * This will be registered as low priority CPU notifier.
3795 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3796 unsigned long action,
3799 int cpu = (unsigned long)hcpu;
3800 struct work_struct unbind_work;
3802 switch (action & ~CPU_TASKS_FROZEN) {
3803 case CPU_DOWN_PREPARE:
3804 /* unbinding should happen on the local CPU */
3805 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
3806 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3807 flush_work(&unbind_work);
3815 struct work_for_cpu {
3816 struct work_struct work;
3822 static void work_for_cpu_fn(struct work_struct *work)
3824 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3826 wfc->ret = wfc->fn(wfc->arg);
3830 * work_on_cpu - run a function in user context on a particular cpu
3831 * @cpu: the cpu to run on
3832 * @fn: the function to run
3833 * @arg: the function arg
3835 * This will return the value @fn returns.
3836 * It is up to the caller to ensure that the cpu doesn't go offline.
3837 * The caller must not hold any locks which would prevent @fn from completing.
3839 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
3841 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3843 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3844 schedule_work_on(cpu, &wfc.work);
3845 flush_work(&wfc.work);
3848 EXPORT_SYMBOL_GPL(work_on_cpu);
3849 #endif /* CONFIG_SMP */
3851 #ifdef CONFIG_FREEZER
3854 * freeze_workqueues_begin - begin freezing workqueues
3856 * Start freezing workqueues. After this function returns, all freezable
3857 * workqueues will queue new works to their frozen_works list instead of
3861 * Grabs and releases workqueue_lock and pool->lock's.
3863 void freeze_workqueues_begin(void)
3865 struct worker_pool *pool;
3866 struct workqueue_struct *wq;
3867 struct pool_workqueue *pwq;
3870 spin_lock_irq(&workqueue_lock);
3872 WARN_ON_ONCE(workqueue_freezing);
3873 workqueue_freezing = true;
3876 for_each_pool(pool, id) {
3877 spin_lock(&pool->lock);
3878 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
3879 pool->flags |= POOL_FREEZING;
3880 spin_unlock(&pool->lock);
3883 /* suppress further executions by setting max_active to zero */
3884 list_for_each_entry(wq, &workqueues, list) {
3885 if (!(wq->flags & WQ_FREEZABLE))
3888 for_each_pwq(pwq, wq) {
3889 spin_lock(&pwq->pool->lock);
3890 pwq->max_active = 0;
3891 spin_unlock(&pwq->pool->lock);
3895 spin_unlock_irq(&workqueue_lock);
3899 * freeze_workqueues_busy - are freezable workqueues still busy?
3901 * Check whether freezing is complete. This function must be called
3902 * between freeze_workqueues_begin() and thaw_workqueues().
3905 * Grabs and releases workqueue_lock.
3908 * %true if some freezable workqueues are still busy. %false if freezing
3911 bool freeze_workqueues_busy(void)
3914 struct workqueue_struct *wq;
3915 struct pool_workqueue *pwq;
3917 spin_lock_irq(&workqueue_lock);
3919 WARN_ON_ONCE(!workqueue_freezing);
3921 list_for_each_entry(wq, &workqueues, list) {
3922 if (!(wq->flags & WQ_FREEZABLE))
3925 * nr_active is monotonically decreasing. It's safe
3926 * to peek without lock.
3928 for_each_pwq(pwq, wq) {
3929 WARN_ON_ONCE(pwq->nr_active < 0);
3930 if (pwq->nr_active) {
3937 spin_unlock_irq(&workqueue_lock);
3942 * thaw_workqueues - thaw workqueues
3944 * Thaw workqueues. Normal queueing is restored and all collected
3945 * frozen works are transferred to their respective pool worklists.
3948 * Grabs and releases workqueue_lock and pool->lock's.
3950 void thaw_workqueues(void)
3952 struct workqueue_struct *wq;
3953 struct pool_workqueue *pwq;
3954 struct worker_pool *pool;
3957 spin_lock_irq(&workqueue_lock);
3959 if (!workqueue_freezing)
3962 /* clear FREEZING */
3963 for_each_pool(pool, id) {
3964 spin_lock(&pool->lock);
3965 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
3966 pool->flags &= ~POOL_FREEZING;
3967 spin_unlock(&pool->lock);
3970 /* restore max_active and repopulate worklist */
3971 list_for_each_entry(wq, &workqueues, list) {
3972 if (!(wq->flags & WQ_FREEZABLE))
3975 for_each_pwq(pwq, wq) {
3976 spin_lock(&pwq->pool->lock);
3977 pwq_set_max_active(pwq, wq->saved_max_active);
3978 spin_unlock(&pwq->pool->lock);
3983 for_each_pool(pool, id) {
3984 spin_lock(&pool->lock);
3985 wake_up_worker(pool);
3986 spin_unlock(&pool->lock);
3989 workqueue_freezing = false;
3991 spin_unlock_irq(&workqueue_lock);
3993 #endif /* CONFIG_FREEZER */
3995 static int __init init_workqueues(void)
3997 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4000 /* make sure we have enough bits for OFFQ pool ID */
4001 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4002 WORK_CPU_END * NR_STD_WORKER_POOLS);
4004 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4006 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4008 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4009 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4011 /* initialize CPU pools */
4012 for_each_possible_cpu(cpu) {
4013 struct worker_pool *pool;
4016 for_each_cpu_worker_pool(pool, cpu) {
4017 BUG_ON(init_worker_pool(pool));
4019 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4020 pool->attrs->nice = std_nice[i++];
4023 BUG_ON(worker_pool_assign_id(pool));
4027 /* create the initial worker */
4028 for_each_online_cpu(cpu) {
4029 struct worker_pool *pool;
4031 for_each_cpu_worker_pool(pool, cpu) {
4032 struct worker *worker;
4034 pool->flags &= ~POOL_DISASSOCIATED;
4036 worker = create_worker(pool);
4038 spin_lock_irq(&pool->lock);
4039 start_worker(worker);
4040 spin_unlock_irq(&pool->lock);
4044 /* create default unbound wq attrs */
4045 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4046 struct workqueue_attrs *attrs;
4048 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4050 attrs->nice = std_nice[i];
4051 cpumask_setall(attrs->cpumask);
4053 unbound_std_wq_attrs[i] = attrs;
4056 system_wq = alloc_workqueue("events", 0, 0);
4057 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4058 system_long_wq = alloc_workqueue("events_long", 0, 0);
4059 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4060 WQ_UNBOUND_MAX_ACTIVE);
4061 system_freezable_wq = alloc_workqueue("events_freezable",
4063 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4064 !system_unbound_wq || !system_freezable_wq);
4067 early_initcall(init_workqueues);