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 should be flipped only while holding
64 * manager_mutex to avoid changing binding state while
65 * create_worker() is in progress.
67 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
68 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
69 POOL_FREEZING = 1 << 3, /* freeze in progress */
72 WORKER_STARTED = 1 << 0, /* started */
73 WORKER_DIE = 1 << 1, /* die die die */
74 WORKER_IDLE = 1 << 2, /* is idle */
75 WORKER_PREP = 1 << 3, /* preparing to run works */
76 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
77 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND |
82 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
84 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
85 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
87 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
88 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
90 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
91 /* call for help after 10ms
93 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
94 CREATE_COOLDOWN = HZ, /* time to breath after fail */
97 * Rescue workers are used only on emergencies and shared by
100 RESCUER_NICE_LEVEL = -20,
101 HIGHPRI_NICE_LEVEL = -20,
105 * Structure fields follow one of the following exclusion rules.
107 * I: Modifiable by initialization/destruction paths and read-only for
110 * P: Preemption protected. Disabling preemption is enough and should
111 * only be modified and accessed from the local cpu.
113 * L: pool->lock protected. Access with pool->lock held.
115 * X: During normal operation, modification requires pool->lock and should
116 * be done only from local cpu. Either disabling preemption on local
117 * cpu or grabbing pool->lock is enough for read access. If
118 * POOL_DISASSOCIATED is set, it's identical to L.
120 * F: wq->flush_mutex protected.
122 * W: workqueue_lock protected.
124 * R: workqueue_lock protected for writes. Sched-RCU protected for reads.
126 * FR: wq->flush_mutex and workqueue_lock protected for writes. Sched-RCU
127 * protected for reads.
130 /* struct worker is defined in workqueue_internal.h */
133 spinlock_t lock; /* the pool lock */
134 int cpu; /* I: the associated cpu */
135 int id; /* I: pool ID */
136 unsigned int flags; /* X: flags */
138 struct list_head worklist; /* L: list of pending works */
139 int nr_workers; /* L: total number of workers */
141 /* nr_idle includes the ones off idle_list for rebinding */
142 int nr_idle; /* L: currently idle ones */
144 struct list_head idle_list; /* X: list of idle workers */
145 struct timer_list idle_timer; /* L: worker idle timeout */
146 struct timer_list mayday_timer; /* L: SOS timer for workers */
148 /* a workers is either on busy_hash or idle_list, or the manager */
149 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
150 /* L: hash of busy workers */
152 /* see manage_workers() for details on the two manager mutexes */
153 struct mutex manager_arb; /* manager arbitration */
154 struct mutex manager_mutex; /* manager exclusion */
155 struct ida worker_ida; /* L: for worker IDs */
157 struct workqueue_attrs *attrs; /* I: worker attributes */
158 struct hlist_node hash_node; /* W: unbound_pool_hash node */
159 int refcnt; /* W: refcnt for unbound pools */
162 * The current concurrency level. As it's likely to be accessed
163 * from other CPUs during try_to_wake_up(), put it in a separate
166 atomic_t nr_running ____cacheline_aligned_in_smp;
169 * Destruction of pool is sched-RCU protected to allow dereferences
170 * from get_work_pool().
173 } ____cacheline_aligned_in_smp;
176 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
177 * of work_struct->data are used for flags and the remaining high bits
178 * point to the pwq; thus, pwqs need to be aligned at two's power of the
179 * number of flag bits.
181 struct pool_workqueue {
182 struct worker_pool *pool; /* I: the associated pool */
183 struct workqueue_struct *wq; /* I: the owning workqueue */
184 int work_color; /* L: current color */
185 int flush_color; /* L: flushing color */
186 int refcnt; /* L: reference count */
187 int nr_in_flight[WORK_NR_COLORS];
188 /* L: nr of in_flight works */
189 int nr_active; /* L: nr of active works */
190 int max_active; /* L: max active works */
191 struct list_head delayed_works; /* L: delayed works */
192 struct list_head pwqs_node; /* FR: node on wq->pwqs */
193 struct list_head mayday_node; /* W: node on wq->maydays */
196 * Release of unbound pwq is punted to system_wq. See put_pwq()
197 * and pwq_unbound_release_workfn() for details. pool_workqueue
198 * itself is also sched-RCU protected so that the first pwq can be
199 * determined without grabbing workqueue_lock.
201 struct work_struct unbound_release_work;
203 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
206 * Structure used to wait for workqueue flush.
209 struct list_head list; /* F: list of flushers */
210 int flush_color; /* F: flush color waiting for */
211 struct completion done; /* flush completion */
217 * The externally visible workqueue. It relays the issued work items to
218 * the appropriate worker_pool through its pool_workqueues.
220 struct workqueue_struct {
221 unsigned int flags; /* W: WQ_* flags */
222 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */
223 struct list_head pwqs; /* FR: all pwqs of this wq */
224 struct list_head list; /* W: list of all workqueues */
226 struct mutex flush_mutex; /* protects wq flushing */
227 int work_color; /* F: current work color */
228 int flush_color; /* F: current flush color */
229 atomic_t nr_pwqs_to_flush; /* flush in progress */
230 struct wq_flusher *first_flusher; /* F: first flusher */
231 struct list_head flusher_queue; /* F: flush waiters */
232 struct list_head flusher_overflow; /* F: flush overflow list */
234 struct list_head maydays; /* W: pwqs requesting rescue */
235 struct worker *rescuer; /* I: rescue worker */
237 int nr_drainers; /* W: drain in progress */
238 int saved_max_active; /* W: saved pwq max_active */
241 struct wq_device *wq_dev; /* I: for sysfs interface */
243 #ifdef CONFIG_LOCKDEP
244 struct lockdep_map lockdep_map;
246 char name[]; /* I: workqueue name */
249 static struct kmem_cache *pwq_cache;
251 /* W: hash of all unbound pools keyed by pool->attrs */
252 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
254 /* I: attributes used when instantiating standard unbound pools on demand */
255 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
257 struct workqueue_struct *system_wq __read_mostly;
258 EXPORT_SYMBOL_GPL(system_wq);
259 struct workqueue_struct *system_highpri_wq __read_mostly;
260 EXPORT_SYMBOL_GPL(system_highpri_wq);
261 struct workqueue_struct *system_long_wq __read_mostly;
262 EXPORT_SYMBOL_GPL(system_long_wq);
263 struct workqueue_struct *system_unbound_wq __read_mostly;
264 EXPORT_SYMBOL_GPL(system_unbound_wq);
265 struct workqueue_struct *system_freezable_wq __read_mostly;
266 EXPORT_SYMBOL_GPL(system_freezable_wq);
268 #define CREATE_TRACE_POINTS
269 #include <trace/events/workqueue.h>
271 #define assert_rcu_or_wq_lock() \
272 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
273 lockdep_is_held(&workqueue_lock), \
274 "sched RCU or workqueue lock should be held")
276 #define for_each_cpu_worker_pool(pool, cpu) \
277 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
278 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
281 #define for_each_busy_worker(worker, i, pool) \
282 hash_for_each(pool->busy_hash, i, worker, hentry)
285 * for_each_pool - iterate through all worker_pools in the system
286 * @pool: iteration cursor
287 * @pi: integer used for iteration
289 * This must be called either with workqueue_lock held or sched RCU read
290 * locked. If the pool needs to be used beyond the locking in effect, the
291 * caller is responsible for guaranteeing that the pool stays online.
293 * The if/else clause exists only for the lockdep assertion and can be
296 #define for_each_pool(pool, pi) \
297 idr_for_each_entry(&worker_pool_idr, pool, pi) \
298 if (({ assert_rcu_or_wq_lock(); false; })) { } \
302 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
303 * @pwq: iteration cursor
304 * @wq: the target workqueue
306 * This must be called either with workqueue_lock held or sched RCU read
307 * locked. If the pwq needs to be used beyond the locking in effect, the
308 * caller is responsible for guaranteeing that the pwq stays online.
310 * The if/else clause exists only for the lockdep assertion and can be
313 #define for_each_pwq(pwq, wq) \
314 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
315 if (({ assert_rcu_or_wq_lock(); false; })) { } \
318 #ifdef CONFIG_DEBUG_OBJECTS_WORK
320 static struct debug_obj_descr work_debug_descr;
322 static void *work_debug_hint(void *addr)
324 return ((struct work_struct *) addr)->func;
328 * fixup_init is called when:
329 * - an active object is initialized
331 static int work_fixup_init(void *addr, enum debug_obj_state state)
333 struct work_struct *work = addr;
336 case ODEBUG_STATE_ACTIVE:
337 cancel_work_sync(work);
338 debug_object_init(work, &work_debug_descr);
346 * fixup_activate is called when:
347 * - an active object is activated
348 * - an unknown object is activated (might be a statically initialized object)
350 static int work_fixup_activate(void *addr, enum debug_obj_state state)
352 struct work_struct *work = addr;
356 case ODEBUG_STATE_NOTAVAILABLE:
358 * This is not really a fixup. The work struct was
359 * statically initialized. We just make sure that it
360 * is tracked in the object tracker.
362 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
363 debug_object_init(work, &work_debug_descr);
364 debug_object_activate(work, &work_debug_descr);
370 case ODEBUG_STATE_ACTIVE:
379 * fixup_free is called when:
380 * - an active object is freed
382 static int work_fixup_free(void *addr, enum debug_obj_state state)
384 struct work_struct *work = addr;
387 case ODEBUG_STATE_ACTIVE:
388 cancel_work_sync(work);
389 debug_object_free(work, &work_debug_descr);
396 static struct debug_obj_descr work_debug_descr = {
397 .name = "work_struct",
398 .debug_hint = work_debug_hint,
399 .fixup_init = work_fixup_init,
400 .fixup_activate = work_fixup_activate,
401 .fixup_free = work_fixup_free,
404 static inline void debug_work_activate(struct work_struct *work)
406 debug_object_activate(work, &work_debug_descr);
409 static inline void debug_work_deactivate(struct work_struct *work)
411 debug_object_deactivate(work, &work_debug_descr);
414 void __init_work(struct work_struct *work, int onstack)
417 debug_object_init_on_stack(work, &work_debug_descr);
419 debug_object_init(work, &work_debug_descr);
421 EXPORT_SYMBOL_GPL(__init_work);
423 void destroy_work_on_stack(struct work_struct *work)
425 debug_object_free(work, &work_debug_descr);
427 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
430 static inline void debug_work_activate(struct work_struct *work) { }
431 static inline void debug_work_deactivate(struct work_struct *work) { }
434 /* Serializes the accesses to the list of workqueues. */
435 static DEFINE_SPINLOCK(workqueue_lock);
436 static LIST_HEAD(workqueues);
437 static bool workqueue_freezing; /* W: have wqs started freezing? */
439 /* the per-cpu worker pools */
440 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
444 * R: idr of all pools. Modifications are protected by workqueue_lock.
445 * Read accesses are protected by sched-RCU protected.
447 static DEFINE_IDR(worker_pool_idr);
449 static int worker_thread(void *__worker);
450 static void copy_workqueue_attrs(struct workqueue_attrs *to,
451 const struct workqueue_attrs *from);
453 /* allocate ID and assign it to @pool */
454 static int worker_pool_assign_id(struct worker_pool *pool)
459 if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL))
462 spin_lock_irq(&workqueue_lock);
463 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
464 spin_unlock_irq(&workqueue_lock);
465 } while (ret == -EAGAIN);
471 * first_pwq - return the first pool_workqueue of the specified workqueue
472 * @wq: the target workqueue
474 * This must be called either with workqueue_lock held or sched RCU read
475 * locked. If the pwq needs to be used beyond the locking in effect, the
476 * caller is responsible for guaranteeing that the pwq stays online.
478 static struct pool_workqueue *first_pwq(struct workqueue_struct *wq)
480 assert_rcu_or_wq_lock();
481 return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue,
485 static unsigned int work_color_to_flags(int color)
487 return color << WORK_STRUCT_COLOR_SHIFT;
490 static int get_work_color(struct work_struct *work)
492 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
493 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
496 static int work_next_color(int color)
498 return (color + 1) % WORK_NR_COLORS;
502 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
503 * contain the pointer to the queued pwq. Once execution starts, the flag
504 * is cleared and the high bits contain OFFQ flags and pool ID.
506 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
507 * and clear_work_data() can be used to set the pwq, pool or clear
508 * work->data. These functions should only be called while the work is
509 * owned - ie. while the PENDING bit is set.
511 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
512 * corresponding to a work. Pool is available once the work has been
513 * queued anywhere after initialization until it is sync canceled. pwq is
514 * available only while the work item is queued.
516 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
517 * canceled. While being canceled, a work item may have its PENDING set
518 * but stay off timer and worklist for arbitrarily long and nobody should
519 * try to steal the PENDING bit.
521 static inline void set_work_data(struct work_struct *work, unsigned long data,
524 WARN_ON_ONCE(!work_pending(work));
525 atomic_long_set(&work->data, data | flags | work_static(work));
528 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
529 unsigned long extra_flags)
531 set_work_data(work, (unsigned long)pwq,
532 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
535 static void set_work_pool_and_keep_pending(struct work_struct *work,
538 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
539 WORK_STRUCT_PENDING);
542 static void set_work_pool_and_clear_pending(struct work_struct *work,
546 * The following wmb is paired with the implied mb in
547 * test_and_set_bit(PENDING) and ensures all updates to @work made
548 * here are visible to and precede any updates by the next PENDING
552 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
555 static void clear_work_data(struct work_struct *work)
557 smp_wmb(); /* see set_work_pool_and_clear_pending() */
558 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
561 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
563 unsigned long data = atomic_long_read(&work->data);
565 if (data & WORK_STRUCT_PWQ)
566 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
572 * get_work_pool - return the worker_pool a given work was associated with
573 * @work: the work item of interest
575 * Return the worker_pool @work was last associated with. %NULL if none.
577 * Pools are created and destroyed under workqueue_lock, and allows read
578 * access under sched-RCU read lock. As such, this function should be
579 * called under workqueue_lock or with preemption disabled.
581 * All fields of the returned pool are accessible as long as the above
582 * mentioned locking is in effect. If the returned pool needs to be used
583 * beyond the critical section, the caller is responsible for ensuring the
584 * returned pool is and stays online.
586 static struct worker_pool *get_work_pool(struct work_struct *work)
588 unsigned long data = atomic_long_read(&work->data);
591 assert_rcu_or_wq_lock();
593 if (data & WORK_STRUCT_PWQ)
594 return ((struct pool_workqueue *)
595 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
597 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
598 if (pool_id == WORK_OFFQ_POOL_NONE)
601 return idr_find(&worker_pool_idr, pool_id);
605 * get_work_pool_id - return the worker pool ID a given work is associated with
606 * @work: the work item of interest
608 * Return the worker_pool ID @work was last associated with.
609 * %WORK_OFFQ_POOL_NONE if none.
611 static int get_work_pool_id(struct work_struct *work)
613 unsigned long data = atomic_long_read(&work->data);
615 if (data & WORK_STRUCT_PWQ)
616 return ((struct pool_workqueue *)
617 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
619 return data >> WORK_OFFQ_POOL_SHIFT;
622 static void mark_work_canceling(struct work_struct *work)
624 unsigned long pool_id = get_work_pool_id(work);
626 pool_id <<= WORK_OFFQ_POOL_SHIFT;
627 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
630 static bool work_is_canceling(struct work_struct *work)
632 unsigned long data = atomic_long_read(&work->data);
634 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
638 * Policy functions. These define the policies on how the global worker
639 * pools are managed. Unless noted otherwise, these functions assume that
640 * they're being called with pool->lock held.
643 static bool __need_more_worker(struct worker_pool *pool)
645 return !atomic_read(&pool->nr_running);
649 * Need to wake up a worker? Called from anything but currently
652 * Note that, because unbound workers never contribute to nr_running, this
653 * function will always return %true for unbound pools as long as the
654 * worklist isn't empty.
656 static bool need_more_worker(struct worker_pool *pool)
658 return !list_empty(&pool->worklist) && __need_more_worker(pool);
661 /* Can I start working? Called from busy but !running workers. */
662 static bool may_start_working(struct worker_pool *pool)
664 return pool->nr_idle;
667 /* Do I need to keep working? Called from currently running workers. */
668 static bool keep_working(struct worker_pool *pool)
670 return !list_empty(&pool->worklist) &&
671 atomic_read(&pool->nr_running) <= 1;
674 /* Do we need a new worker? Called from manager. */
675 static bool need_to_create_worker(struct worker_pool *pool)
677 return need_more_worker(pool) && !may_start_working(pool);
680 /* Do I need to be the manager? */
681 static bool need_to_manage_workers(struct worker_pool *pool)
683 return need_to_create_worker(pool) ||
684 (pool->flags & POOL_MANAGE_WORKERS);
687 /* Do we have too many workers and should some go away? */
688 static bool too_many_workers(struct worker_pool *pool)
690 bool managing = mutex_is_locked(&pool->manager_arb);
691 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
692 int nr_busy = pool->nr_workers - nr_idle;
695 * nr_idle and idle_list may disagree if idle rebinding is in
696 * progress. Never return %true if idle_list is empty.
698 if (list_empty(&pool->idle_list))
701 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
708 /* Return the first worker. Safe with preemption disabled */
709 static struct worker *first_worker(struct worker_pool *pool)
711 if (unlikely(list_empty(&pool->idle_list)))
714 return list_first_entry(&pool->idle_list, struct worker, entry);
718 * wake_up_worker - wake up an idle worker
719 * @pool: worker pool to wake worker from
721 * Wake up the first idle worker of @pool.
724 * spin_lock_irq(pool->lock).
726 static void wake_up_worker(struct worker_pool *pool)
728 struct worker *worker = first_worker(pool);
731 wake_up_process(worker->task);
735 * wq_worker_waking_up - a worker is waking up
736 * @task: task waking up
737 * @cpu: CPU @task is waking up to
739 * This function is called during try_to_wake_up() when a worker is
743 * spin_lock_irq(rq->lock)
745 void wq_worker_waking_up(struct task_struct *task, int cpu)
747 struct worker *worker = kthread_data(task);
749 if (!(worker->flags & WORKER_NOT_RUNNING)) {
750 WARN_ON_ONCE(worker->pool->cpu != cpu);
751 atomic_inc(&worker->pool->nr_running);
756 * wq_worker_sleeping - a worker is going to sleep
757 * @task: task going to sleep
758 * @cpu: CPU in question, must be the current CPU number
760 * This function is called during schedule() when a busy worker is
761 * going to sleep. Worker on the same cpu can be woken up by
762 * returning pointer to its task.
765 * spin_lock_irq(rq->lock)
768 * Worker task on @cpu to wake up, %NULL if none.
770 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
772 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
773 struct worker_pool *pool;
776 * Rescuers, which may not have all the fields set up like normal
777 * workers, also reach here, let's not access anything before
778 * checking NOT_RUNNING.
780 if (worker->flags & WORKER_NOT_RUNNING)
785 /* this can only happen on the local cpu */
786 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
790 * The counterpart of the following dec_and_test, implied mb,
791 * worklist not empty test sequence is in insert_work().
792 * Please read comment there.
794 * NOT_RUNNING is clear. This means that we're bound to and
795 * running on the local cpu w/ rq lock held and preemption
796 * disabled, which in turn means that none else could be
797 * manipulating idle_list, so dereferencing idle_list without pool
800 if (atomic_dec_and_test(&pool->nr_running) &&
801 !list_empty(&pool->worklist))
802 to_wakeup = first_worker(pool);
803 return to_wakeup ? to_wakeup->task : NULL;
807 * worker_set_flags - set worker flags and adjust nr_running accordingly
809 * @flags: flags to set
810 * @wakeup: wakeup an idle worker if necessary
812 * Set @flags in @worker->flags and adjust nr_running accordingly. If
813 * nr_running becomes zero and @wakeup is %true, an idle worker is
817 * spin_lock_irq(pool->lock)
819 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
822 struct worker_pool *pool = worker->pool;
824 WARN_ON_ONCE(worker->task != current);
827 * If transitioning into NOT_RUNNING, adjust nr_running and
828 * wake up an idle worker as necessary if requested by
831 if ((flags & WORKER_NOT_RUNNING) &&
832 !(worker->flags & WORKER_NOT_RUNNING)) {
834 if (atomic_dec_and_test(&pool->nr_running) &&
835 !list_empty(&pool->worklist))
836 wake_up_worker(pool);
838 atomic_dec(&pool->nr_running);
841 worker->flags |= flags;
845 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
847 * @flags: flags to clear
849 * Clear @flags in @worker->flags and adjust nr_running accordingly.
852 * spin_lock_irq(pool->lock)
854 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
856 struct worker_pool *pool = worker->pool;
857 unsigned int oflags = worker->flags;
859 WARN_ON_ONCE(worker->task != current);
861 worker->flags &= ~flags;
864 * If transitioning out of NOT_RUNNING, increment nr_running. Note
865 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
866 * of multiple flags, not a single flag.
868 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
869 if (!(worker->flags & WORKER_NOT_RUNNING))
870 atomic_inc(&pool->nr_running);
874 * find_worker_executing_work - find worker which is executing a work
875 * @pool: pool of interest
876 * @work: work to find worker for
878 * Find a worker which is executing @work on @pool by searching
879 * @pool->busy_hash which is keyed by the address of @work. For a worker
880 * to match, its current execution should match the address of @work and
881 * its work function. This is to avoid unwanted dependency between
882 * unrelated work executions through a work item being recycled while still
885 * This is a bit tricky. A work item may be freed once its execution
886 * starts and nothing prevents the freed area from being recycled for
887 * another work item. If the same work item address ends up being reused
888 * before the original execution finishes, workqueue will identify the
889 * recycled work item as currently executing and make it wait until the
890 * current execution finishes, introducing an unwanted dependency.
892 * This function checks the work item address and work function to avoid
893 * false positives. Note that this isn't complete as one may construct a
894 * work function which can introduce dependency onto itself through a
895 * recycled work item. Well, if somebody wants to shoot oneself in the
896 * foot that badly, there's only so much we can do, and if such deadlock
897 * actually occurs, it should be easy to locate the culprit work function.
900 * spin_lock_irq(pool->lock).
903 * Pointer to worker which is executing @work if found, NULL
906 static struct worker *find_worker_executing_work(struct worker_pool *pool,
907 struct work_struct *work)
909 struct worker *worker;
911 hash_for_each_possible(pool->busy_hash, worker, hentry,
913 if (worker->current_work == work &&
914 worker->current_func == work->func)
921 * move_linked_works - move linked works to a list
922 * @work: start of series of works to be scheduled
923 * @head: target list to append @work to
924 * @nextp: out paramter for nested worklist walking
926 * Schedule linked works starting from @work to @head. Work series to
927 * be scheduled starts at @work and includes any consecutive work with
928 * WORK_STRUCT_LINKED set in its predecessor.
930 * If @nextp is not NULL, it's updated to point to the next work of
931 * the last scheduled work. This allows move_linked_works() to be
932 * nested inside outer list_for_each_entry_safe().
935 * spin_lock_irq(pool->lock).
937 static void move_linked_works(struct work_struct *work, struct list_head *head,
938 struct work_struct **nextp)
940 struct work_struct *n;
943 * Linked worklist will always end before the end of the list,
944 * use NULL for list head.
946 list_for_each_entry_safe_from(work, n, NULL, entry) {
947 list_move_tail(&work->entry, head);
948 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
953 * If we're already inside safe list traversal and have moved
954 * multiple works to the scheduled queue, the next position
955 * needs to be updated.
962 * get_pwq - get an extra reference on the specified pool_workqueue
963 * @pwq: pool_workqueue to get
965 * Obtain an extra reference on @pwq. The caller should guarantee that
966 * @pwq has positive refcnt and be holding the matching pool->lock.
968 static void get_pwq(struct pool_workqueue *pwq)
970 lockdep_assert_held(&pwq->pool->lock);
971 WARN_ON_ONCE(pwq->refcnt <= 0);
976 * put_pwq - put a pool_workqueue reference
977 * @pwq: pool_workqueue to put
979 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
980 * destruction. The caller should be holding the matching pool->lock.
982 static void put_pwq(struct pool_workqueue *pwq)
984 lockdep_assert_held(&pwq->pool->lock);
985 if (likely(--pwq->refcnt))
987 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
990 * @pwq can't be released under pool->lock, bounce to
991 * pwq_unbound_release_workfn(). This never recurses on the same
992 * pool->lock as this path is taken only for unbound workqueues and
993 * the release work item is scheduled on a per-cpu workqueue. To
994 * avoid lockdep warning, unbound pool->locks are given lockdep
995 * subclass of 1 in get_unbound_pool().
997 schedule_work(&pwq->unbound_release_work);
1000 static void pwq_activate_delayed_work(struct work_struct *work)
1002 struct pool_workqueue *pwq = get_work_pwq(work);
1004 trace_workqueue_activate_work(work);
1005 move_linked_works(work, &pwq->pool->worklist, NULL);
1006 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1010 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1012 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1013 struct work_struct, entry);
1015 pwq_activate_delayed_work(work);
1019 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1020 * @pwq: pwq of interest
1021 * @color: color of work which left the queue
1023 * A work either has completed or is removed from pending queue,
1024 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1027 * spin_lock_irq(pool->lock).
1029 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1031 /* uncolored work items don't participate in flushing or nr_active */
1032 if (color == WORK_NO_COLOR)
1035 pwq->nr_in_flight[color]--;
1038 if (!list_empty(&pwq->delayed_works)) {
1039 /* one down, submit a delayed one */
1040 if (pwq->nr_active < pwq->max_active)
1041 pwq_activate_first_delayed(pwq);
1044 /* is flush in progress and are we at the flushing tip? */
1045 if (likely(pwq->flush_color != color))
1048 /* are there still in-flight works? */
1049 if (pwq->nr_in_flight[color])
1052 /* this pwq is done, clear flush_color */
1053 pwq->flush_color = -1;
1056 * If this was the last pwq, wake up the first flusher. It
1057 * will handle the rest.
1059 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1060 complete(&pwq->wq->first_flusher->done);
1066 * try_to_grab_pending - steal work item from worklist and disable irq
1067 * @work: work item to steal
1068 * @is_dwork: @work is a delayed_work
1069 * @flags: place to store irq state
1071 * Try to grab PENDING bit of @work. This function can handle @work in any
1072 * stable state - idle, on timer or on worklist. Return values are
1074 * 1 if @work was pending and we successfully stole PENDING
1075 * 0 if @work was idle and we claimed PENDING
1076 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1077 * -ENOENT if someone else is canceling @work, this state may persist
1078 * for arbitrarily long
1080 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1081 * interrupted while holding PENDING and @work off queue, irq must be
1082 * disabled on entry. This, combined with delayed_work->timer being
1083 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1085 * On successful return, >= 0, irq is disabled and the caller is
1086 * responsible for releasing it using local_irq_restore(*@flags).
1088 * This function is safe to call from any context including IRQ handler.
1090 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1091 unsigned long *flags)
1093 struct worker_pool *pool;
1094 struct pool_workqueue *pwq;
1096 local_irq_save(*flags);
1098 /* try to steal the timer if it exists */
1100 struct delayed_work *dwork = to_delayed_work(work);
1103 * dwork->timer is irqsafe. If del_timer() fails, it's
1104 * guaranteed that the timer is not queued anywhere and not
1105 * running on the local CPU.
1107 if (likely(del_timer(&dwork->timer)))
1111 /* try to claim PENDING the normal way */
1112 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1116 * The queueing is in progress, or it is already queued. Try to
1117 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1119 pool = get_work_pool(work);
1123 spin_lock(&pool->lock);
1125 * work->data is guaranteed to point to pwq only while the work
1126 * item is queued on pwq->wq, and both updating work->data to point
1127 * to pwq on queueing and to pool on dequeueing are done under
1128 * pwq->pool->lock. This in turn guarantees that, if work->data
1129 * points to pwq which is associated with a locked pool, the work
1130 * item is currently queued on that pool.
1132 pwq = get_work_pwq(work);
1133 if (pwq && pwq->pool == pool) {
1134 debug_work_deactivate(work);
1137 * A delayed work item cannot be grabbed directly because
1138 * it might have linked NO_COLOR work items which, if left
1139 * on the delayed_list, will confuse pwq->nr_active
1140 * management later on and cause stall. Make sure the work
1141 * item is activated before grabbing.
1143 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1144 pwq_activate_delayed_work(work);
1146 list_del_init(&work->entry);
1147 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1149 /* work->data points to pwq iff queued, point to pool */
1150 set_work_pool_and_keep_pending(work, pool->id);
1152 spin_unlock(&pool->lock);
1155 spin_unlock(&pool->lock);
1157 local_irq_restore(*flags);
1158 if (work_is_canceling(work))
1165 * insert_work - insert a work into a pool
1166 * @pwq: pwq @work belongs to
1167 * @work: work to insert
1168 * @head: insertion point
1169 * @extra_flags: extra WORK_STRUCT_* flags to set
1171 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1172 * work_struct flags.
1175 * spin_lock_irq(pool->lock).
1177 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1178 struct list_head *head, unsigned int extra_flags)
1180 struct worker_pool *pool = pwq->pool;
1182 /* we own @work, set data and link */
1183 set_work_pwq(work, pwq, extra_flags);
1184 list_add_tail(&work->entry, head);
1188 * Ensure either wq_worker_sleeping() sees the above
1189 * list_add_tail() or we see zero nr_running to avoid workers lying
1190 * around lazily while there are works to be processed.
1194 if (__need_more_worker(pool))
1195 wake_up_worker(pool);
1199 * Test whether @work is being queued from another work executing on the
1202 static bool is_chained_work(struct workqueue_struct *wq)
1204 struct worker *worker;
1206 worker = current_wq_worker();
1208 * Return %true iff I'm a worker execuing a work item on @wq. If
1209 * I'm @worker, it's safe to dereference it without locking.
1211 return worker && worker->current_pwq->wq == wq;
1214 static void __queue_work(int cpu, struct workqueue_struct *wq,
1215 struct work_struct *work)
1217 struct pool_workqueue *pwq;
1218 struct worker_pool *last_pool;
1219 struct list_head *worklist;
1220 unsigned int work_flags;
1221 unsigned int req_cpu = cpu;
1224 * While a work item is PENDING && off queue, a task trying to
1225 * steal the PENDING will busy-loop waiting for it to either get
1226 * queued or lose PENDING. Grabbing PENDING and queueing should
1227 * happen with IRQ disabled.
1229 WARN_ON_ONCE(!irqs_disabled());
1231 debug_work_activate(work);
1233 /* if dying, only works from the same workqueue are allowed */
1234 if (unlikely(wq->flags & __WQ_DRAINING) &&
1235 WARN_ON_ONCE(!is_chained_work(wq)))
1238 /* pwq which will be used unless @work is executing elsewhere */
1239 if (!(wq->flags & WQ_UNBOUND)) {
1240 if (cpu == WORK_CPU_UNBOUND)
1241 cpu = raw_smp_processor_id();
1242 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1244 pwq = first_pwq(wq);
1248 * If @work was previously on a different pool, it might still be
1249 * running there, in which case the work needs to be queued on that
1250 * pool to guarantee non-reentrancy.
1252 last_pool = get_work_pool(work);
1253 if (last_pool && last_pool != pwq->pool) {
1254 struct worker *worker;
1256 spin_lock(&last_pool->lock);
1258 worker = find_worker_executing_work(last_pool, work);
1260 if (worker && worker->current_pwq->wq == wq) {
1261 pwq = worker->current_pwq;
1263 /* meh... not running there, queue here */
1264 spin_unlock(&last_pool->lock);
1265 spin_lock(&pwq->pool->lock);
1268 spin_lock(&pwq->pool->lock);
1272 * pwq is determined and locked. For unbound pools, we could have
1273 * raced with pwq release and it could already be dead. If its
1274 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1275 * without another pwq replacing it as the first pwq or while a
1276 * work item is executing on it, so the retying is guaranteed to
1277 * make forward-progress.
1279 if (unlikely(!pwq->refcnt)) {
1280 if (wq->flags & WQ_UNBOUND) {
1281 spin_unlock(&pwq->pool->lock);
1286 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1290 /* pwq determined, queue */
1291 trace_workqueue_queue_work(req_cpu, pwq, work);
1293 if (WARN_ON(!list_empty(&work->entry))) {
1294 spin_unlock(&pwq->pool->lock);
1298 pwq->nr_in_flight[pwq->work_color]++;
1299 work_flags = work_color_to_flags(pwq->work_color);
1301 if (likely(pwq->nr_active < pwq->max_active)) {
1302 trace_workqueue_activate_work(work);
1304 worklist = &pwq->pool->worklist;
1306 work_flags |= WORK_STRUCT_DELAYED;
1307 worklist = &pwq->delayed_works;
1310 insert_work(pwq, work, worklist, work_flags);
1312 spin_unlock(&pwq->pool->lock);
1316 * queue_work_on - queue work on specific cpu
1317 * @cpu: CPU number to execute work on
1318 * @wq: workqueue to use
1319 * @work: work to queue
1321 * Returns %false if @work was already on a queue, %true otherwise.
1323 * We queue the work to a specific CPU, the caller must ensure it
1326 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1327 struct work_struct *work)
1330 unsigned long flags;
1332 local_irq_save(flags);
1334 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1335 __queue_work(cpu, wq, work);
1339 local_irq_restore(flags);
1342 EXPORT_SYMBOL_GPL(queue_work_on);
1344 void delayed_work_timer_fn(unsigned long __data)
1346 struct delayed_work *dwork = (struct delayed_work *)__data;
1348 /* should have been called from irqsafe timer with irq already off */
1349 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1351 EXPORT_SYMBOL(delayed_work_timer_fn);
1353 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1354 struct delayed_work *dwork, unsigned long delay)
1356 struct timer_list *timer = &dwork->timer;
1357 struct work_struct *work = &dwork->work;
1359 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1360 timer->data != (unsigned long)dwork);
1361 WARN_ON_ONCE(timer_pending(timer));
1362 WARN_ON_ONCE(!list_empty(&work->entry));
1365 * If @delay is 0, queue @dwork->work immediately. This is for
1366 * both optimization and correctness. The earliest @timer can
1367 * expire is on the closest next tick and delayed_work users depend
1368 * on that there's no such delay when @delay is 0.
1371 __queue_work(cpu, wq, &dwork->work);
1375 timer_stats_timer_set_start_info(&dwork->timer);
1379 timer->expires = jiffies + delay;
1381 if (unlikely(cpu != WORK_CPU_UNBOUND))
1382 add_timer_on(timer, cpu);
1388 * queue_delayed_work_on - queue work on specific CPU after delay
1389 * @cpu: CPU number to execute work on
1390 * @wq: workqueue to use
1391 * @dwork: work to queue
1392 * @delay: number of jiffies to wait before queueing
1394 * Returns %false if @work was already on a queue, %true otherwise. If
1395 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1398 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1399 struct delayed_work *dwork, unsigned long delay)
1401 struct work_struct *work = &dwork->work;
1403 unsigned long flags;
1405 /* read the comment in __queue_work() */
1406 local_irq_save(flags);
1408 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1409 __queue_delayed_work(cpu, wq, dwork, delay);
1413 local_irq_restore(flags);
1416 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1419 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1420 * @cpu: CPU number to execute work on
1421 * @wq: workqueue to use
1422 * @dwork: work to queue
1423 * @delay: number of jiffies to wait before queueing
1425 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1426 * modify @dwork's timer so that it expires after @delay. If @delay is
1427 * zero, @work is guaranteed to be scheduled immediately regardless of its
1430 * Returns %false if @dwork was idle and queued, %true if @dwork was
1431 * pending and its timer was modified.
1433 * This function is safe to call from any context including IRQ handler.
1434 * See try_to_grab_pending() for details.
1436 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1437 struct delayed_work *dwork, unsigned long delay)
1439 unsigned long flags;
1443 ret = try_to_grab_pending(&dwork->work, true, &flags);
1444 } while (unlikely(ret == -EAGAIN));
1446 if (likely(ret >= 0)) {
1447 __queue_delayed_work(cpu, wq, dwork, delay);
1448 local_irq_restore(flags);
1451 /* -ENOENT from try_to_grab_pending() becomes %true */
1454 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1457 * worker_enter_idle - enter idle state
1458 * @worker: worker which is entering idle state
1460 * @worker is entering idle state. Update stats and idle timer if
1464 * spin_lock_irq(pool->lock).
1466 static void worker_enter_idle(struct worker *worker)
1468 struct worker_pool *pool = worker->pool;
1470 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1471 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1472 (worker->hentry.next || worker->hentry.pprev)))
1475 /* can't use worker_set_flags(), also called from start_worker() */
1476 worker->flags |= WORKER_IDLE;
1478 worker->last_active = jiffies;
1480 /* idle_list is LIFO */
1481 list_add(&worker->entry, &pool->idle_list);
1483 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1484 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1487 * Sanity check nr_running. Because wq_unbind_fn() releases
1488 * pool->lock between setting %WORKER_UNBOUND and zapping
1489 * nr_running, the warning may trigger spuriously. Check iff
1490 * unbind is not in progress.
1492 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1493 pool->nr_workers == pool->nr_idle &&
1494 atomic_read(&pool->nr_running));
1498 * worker_leave_idle - leave idle state
1499 * @worker: worker which is leaving idle state
1501 * @worker is leaving idle state. Update stats.
1504 * spin_lock_irq(pool->lock).
1506 static void worker_leave_idle(struct worker *worker)
1508 struct worker_pool *pool = worker->pool;
1510 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1512 worker_clr_flags(worker, WORKER_IDLE);
1514 list_del_init(&worker->entry);
1518 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1519 * @pool: target worker_pool
1521 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1523 * Works which are scheduled while the cpu is online must at least be
1524 * scheduled to a worker which is bound to the cpu so that if they are
1525 * flushed from cpu callbacks while cpu is going down, they are
1526 * guaranteed to execute on the cpu.
1528 * This function is to be used by unbound workers and rescuers to bind
1529 * themselves to the target cpu and may race with cpu going down or
1530 * coming online. kthread_bind() can't be used because it may put the
1531 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1532 * verbatim as it's best effort and blocking and pool may be
1533 * [dis]associated in the meantime.
1535 * This function tries set_cpus_allowed() and locks pool and verifies the
1536 * binding against %POOL_DISASSOCIATED which is set during
1537 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1538 * enters idle state or fetches works without dropping lock, it can
1539 * guarantee the scheduling requirement described in the first paragraph.
1542 * Might sleep. Called without any lock but returns with pool->lock
1546 * %true if the associated pool is online (@worker is successfully
1547 * bound), %false if offline.
1549 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1550 __acquires(&pool->lock)
1554 * The following call may fail, succeed or succeed
1555 * without actually migrating the task to the cpu if
1556 * it races with cpu hotunplug operation. Verify
1557 * against POOL_DISASSOCIATED.
1559 if (!(pool->flags & POOL_DISASSOCIATED))
1560 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1562 spin_lock_irq(&pool->lock);
1563 if (pool->flags & POOL_DISASSOCIATED)
1565 if (task_cpu(current) == pool->cpu &&
1566 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1568 spin_unlock_irq(&pool->lock);
1571 * We've raced with CPU hot[un]plug. Give it a breather
1572 * and retry migration. cond_resched() is required here;
1573 * otherwise, we might deadlock against cpu_stop trying to
1574 * bring down the CPU on non-preemptive kernel.
1582 * Rebind an idle @worker to its CPU. worker_thread() will test
1583 * list_empty(@worker->entry) before leaving idle and call this function.
1585 static void idle_worker_rebind(struct worker *worker)
1587 /* CPU may go down again inbetween, clear UNBOUND only on success */
1588 if (worker_maybe_bind_and_lock(worker->pool))
1589 worker_clr_flags(worker, WORKER_UNBOUND);
1591 /* rebind complete, become available again */
1592 list_add(&worker->entry, &worker->pool->idle_list);
1593 spin_unlock_irq(&worker->pool->lock);
1597 * Function for @worker->rebind.work used to rebind unbound busy workers to
1598 * the associated cpu which is coming back online. This is scheduled by
1599 * cpu up but can race with other cpu hotplug operations and may be
1600 * executed twice without intervening cpu down.
1602 static void busy_worker_rebind_fn(struct work_struct *work)
1604 struct worker *worker = container_of(work, struct worker, rebind_work);
1606 if (worker_maybe_bind_and_lock(worker->pool))
1607 worker_clr_flags(worker, WORKER_UNBOUND);
1609 spin_unlock_irq(&worker->pool->lock);
1613 * rebind_workers - rebind all workers of a pool to the associated CPU
1614 * @pool: pool of interest
1616 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1617 * is different for idle and busy ones.
1619 * Idle ones will be removed from the idle_list and woken up. They will
1620 * add themselves back after completing rebind. This ensures that the
1621 * idle_list doesn't contain any unbound workers when re-bound busy workers
1622 * try to perform local wake-ups for concurrency management.
1624 * Busy workers can rebind after they finish their current work items.
1625 * Queueing the rebind work item at the head of the scheduled list is
1626 * enough. Note that nr_running will be properly bumped as busy workers
1629 * On return, all non-manager workers are scheduled for rebind - see
1630 * manage_workers() for the manager special case. Any idle worker
1631 * including the manager will not appear on @idle_list until rebind is
1632 * complete, making local wake-ups safe.
1634 static void rebind_workers(struct worker_pool *pool)
1636 struct worker *worker, *n;
1639 lockdep_assert_held(&pool->manager_mutex);
1640 lockdep_assert_held(&pool->lock);
1642 /* dequeue and kick idle ones */
1643 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1645 * idle workers should be off @pool->idle_list until rebind
1646 * is complete to avoid receiving premature local wake-ups.
1648 list_del_init(&worker->entry);
1651 * worker_thread() will see the above dequeuing and call
1652 * idle_worker_rebind().
1654 wake_up_process(worker->task);
1657 /* rebind busy workers */
1658 for_each_busy_worker(worker, i, pool) {
1659 struct work_struct *rebind_work = &worker->rebind_work;
1660 struct workqueue_struct *wq;
1662 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1663 work_data_bits(rebind_work)))
1666 debug_work_activate(rebind_work);
1669 * wq doesn't really matter but let's keep @worker->pool
1670 * and @pwq->pool consistent for sanity.
1672 if (worker->pool->attrs->nice < 0)
1673 wq = system_highpri_wq;
1677 insert_work(per_cpu_ptr(wq->cpu_pwqs, pool->cpu), rebind_work,
1678 worker->scheduled.next,
1679 work_color_to_flags(WORK_NO_COLOR));
1683 static struct worker *alloc_worker(void)
1685 struct worker *worker;
1687 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1689 INIT_LIST_HEAD(&worker->entry);
1690 INIT_LIST_HEAD(&worker->scheduled);
1691 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1692 /* on creation a worker is in !idle && prep state */
1693 worker->flags = WORKER_PREP;
1699 * create_worker - create a new workqueue worker
1700 * @pool: pool the new worker will belong to
1702 * Create a new worker which is bound to @pool. The returned worker
1703 * can be started by calling start_worker() or destroyed using
1707 * Might sleep. Does GFP_KERNEL allocations.
1710 * Pointer to the newly created worker.
1712 static struct worker *create_worker(struct worker_pool *pool)
1714 const char *pri = pool->attrs->nice < 0 ? "H" : "";
1715 struct worker *worker = NULL;
1718 spin_lock_irq(&pool->lock);
1719 while (ida_get_new(&pool->worker_ida, &id)) {
1720 spin_unlock_irq(&pool->lock);
1721 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1723 spin_lock_irq(&pool->lock);
1725 spin_unlock_irq(&pool->lock);
1727 worker = alloc_worker();
1731 worker->pool = pool;
1735 worker->task = kthread_create_on_node(worker_thread,
1736 worker, cpu_to_node(pool->cpu),
1737 "kworker/%d:%d%s", pool->cpu, id, pri);
1739 worker->task = kthread_create(worker_thread, worker,
1742 if (IS_ERR(worker->task))
1746 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1747 * online CPUs. It'll be re-applied when any of the CPUs come up.
1749 set_user_nice(worker->task, pool->attrs->nice);
1750 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1753 * %PF_THREAD_BOUND is used to prevent userland from meddling with
1754 * cpumask of workqueue workers. This is an abuse. We need
1755 * %PF_NO_SETAFFINITY.
1757 worker->task->flags |= PF_THREAD_BOUND;
1760 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1761 * remains stable across this function. See the comments above the
1762 * flag definition for details.
1764 if (pool->flags & POOL_DISASSOCIATED)
1765 worker->flags |= WORKER_UNBOUND;
1770 spin_lock_irq(&pool->lock);
1771 ida_remove(&pool->worker_ida, id);
1772 spin_unlock_irq(&pool->lock);
1779 * start_worker - start a newly created worker
1780 * @worker: worker to start
1782 * Make the pool aware of @worker and start it.
1785 * spin_lock_irq(pool->lock).
1787 static void start_worker(struct worker *worker)
1789 worker->flags |= WORKER_STARTED;
1790 worker->pool->nr_workers++;
1791 worker_enter_idle(worker);
1792 wake_up_process(worker->task);
1796 * destroy_worker - destroy a workqueue worker
1797 * @worker: worker to be destroyed
1799 * Destroy @worker and adjust @pool stats accordingly.
1802 * spin_lock_irq(pool->lock) which is released and regrabbed.
1804 static void destroy_worker(struct worker *worker)
1806 struct worker_pool *pool = worker->pool;
1807 int id = worker->id;
1809 /* sanity check frenzy */
1810 if (WARN_ON(worker->current_work) ||
1811 WARN_ON(!list_empty(&worker->scheduled)))
1814 if (worker->flags & WORKER_STARTED)
1816 if (worker->flags & WORKER_IDLE)
1819 list_del_init(&worker->entry);
1820 worker->flags |= WORKER_DIE;
1822 spin_unlock_irq(&pool->lock);
1824 kthread_stop(worker->task);
1827 spin_lock_irq(&pool->lock);
1828 ida_remove(&pool->worker_ida, id);
1831 static void idle_worker_timeout(unsigned long __pool)
1833 struct worker_pool *pool = (void *)__pool;
1835 spin_lock_irq(&pool->lock);
1837 if (too_many_workers(pool)) {
1838 struct worker *worker;
1839 unsigned long expires;
1841 /* idle_list is kept in LIFO order, check the last one */
1842 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1843 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1845 if (time_before(jiffies, expires))
1846 mod_timer(&pool->idle_timer, expires);
1848 /* it's been idle for too long, wake up manager */
1849 pool->flags |= POOL_MANAGE_WORKERS;
1850 wake_up_worker(pool);
1854 spin_unlock_irq(&pool->lock);
1857 static void send_mayday(struct work_struct *work)
1859 struct pool_workqueue *pwq = get_work_pwq(work);
1860 struct workqueue_struct *wq = pwq->wq;
1862 lockdep_assert_held(&workqueue_lock);
1867 /* mayday mayday mayday */
1868 if (list_empty(&pwq->mayday_node)) {
1869 list_add_tail(&pwq->mayday_node, &wq->maydays);
1870 wake_up_process(wq->rescuer->task);
1874 static void pool_mayday_timeout(unsigned long __pool)
1876 struct worker_pool *pool = (void *)__pool;
1877 struct work_struct *work;
1879 spin_lock_irq(&workqueue_lock); /* for wq->maydays */
1880 spin_lock(&pool->lock);
1882 if (need_to_create_worker(pool)) {
1884 * We've been trying to create a new worker but
1885 * haven't been successful. We might be hitting an
1886 * allocation deadlock. Send distress signals to
1889 list_for_each_entry(work, &pool->worklist, entry)
1893 spin_unlock(&pool->lock);
1894 spin_unlock_irq(&workqueue_lock);
1896 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1900 * maybe_create_worker - create a new worker if necessary
1901 * @pool: pool to create a new worker for
1903 * Create a new worker for @pool if necessary. @pool is guaranteed to
1904 * have at least one idle worker on return from this function. If
1905 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1906 * sent to all rescuers with works scheduled on @pool to resolve
1907 * possible allocation deadlock.
1909 * On return, need_to_create_worker() is guaranteed to be %false and
1910 * may_start_working() %true.
1913 * spin_lock_irq(pool->lock) which may be released and regrabbed
1914 * multiple times. Does GFP_KERNEL allocations. Called only from
1918 * %false if no action was taken and pool->lock stayed locked, %true
1921 static bool maybe_create_worker(struct worker_pool *pool)
1922 __releases(&pool->lock)
1923 __acquires(&pool->lock)
1925 if (!need_to_create_worker(pool))
1928 spin_unlock_irq(&pool->lock);
1930 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1931 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1934 struct worker *worker;
1936 worker = create_worker(pool);
1938 del_timer_sync(&pool->mayday_timer);
1939 spin_lock_irq(&pool->lock);
1940 start_worker(worker);
1941 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1946 if (!need_to_create_worker(pool))
1949 __set_current_state(TASK_INTERRUPTIBLE);
1950 schedule_timeout(CREATE_COOLDOWN);
1952 if (!need_to_create_worker(pool))
1956 del_timer_sync(&pool->mayday_timer);
1957 spin_lock_irq(&pool->lock);
1958 if (need_to_create_worker(pool))
1964 * maybe_destroy_worker - destroy workers which have been idle for a while
1965 * @pool: pool to destroy workers for
1967 * Destroy @pool workers which have been idle for longer than
1968 * IDLE_WORKER_TIMEOUT.
1971 * spin_lock_irq(pool->lock) which may be released and regrabbed
1972 * multiple times. Called only from manager.
1975 * %false if no action was taken and pool->lock stayed locked, %true
1978 static bool maybe_destroy_workers(struct worker_pool *pool)
1982 while (too_many_workers(pool)) {
1983 struct worker *worker;
1984 unsigned long expires;
1986 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1987 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1989 if (time_before(jiffies, expires)) {
1990 mod_timer(&pool->idle_timer, expires);
1994 destroy_worker(worker);
2002 * manage_workers - manage worker pool
2005 * Assume the manager role and manage the worker pool @worker belongs
2006 * to. At any given time, there can be only zero or one manager per
2007 * pool. The exclusion is handled automatically by this function.
2009 * The caller can safely start processing works on false return. On
2010 * true return, it's guaranteed that need_to_create_worker() is false
2011 * and may_start_working() is true.
2014 * spin_lock_irq(pool->lock) which may be released and regrabbed
2015 * multiple times. Does GFP_KERNEL allocations.
2018 * spin_lock_irq(pool->lock) which may be released and regrabbed
2019 * multiple times. Does GFP_KERNEL allocations.
2021 static bool manage_workers(struct worker *worker)
2023 struct worker_pool *pool = worker->pool;
2027 * Managership is governed by two mutexes - manager_arb and
2028 * manager_mutex. manager_arb handles arbitration of manager role.
2029 * Anyone who successfully grabs manager_arb wins the arbitration
2030 * and becomes the manager. mutex_trylock() on pool->manager_arb
2031 * failure while holding pool->lock reliably indicates that someone
2032 * else is managing the pool and the worker which failed trylock
2033 * can proceed to executing work items. This means that anyone
2034 * grabbing manager_arb is responsible for actually performing
2035 * manager duties. If manager_arb is grabbed and released without
2036 * actual management, the pool may stall indefinitely.
2038 * manager_mutex is used for exclusion of actual management
2039 * operations. The holder of manager_mutex can be sure that none
2040 * of management operations, including creation and destruction of
2041 * workers, won't take place until the mutex is released. Because
2042 * manager_mutex doesn't interfere with manager role arbitration,
2043 * it is guaranteed that the pool's management, while may be
2044 * delayed, won't be disturbed by someone else grabbing
2047 if (!mutex_trylock(&pool->manager_arb))
2051 * With manager arbitration won, manager_mutex would be free in
2052 * most cases. trylock first without dropping @pool->lock.
2054 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2055 spin_unlock_irq(&pool->lock);
2056 mutex_lock(&pool->manager_mutex);
2058 * CPU hotplug could have happened while we were waiting
2059 * for assoc_mutex. Hotplug itself can't handle us
2060 * because manager isn't either on idle or busy list, and
2061 * @pool's state and ours could have deviated.
2063 * As hotplug is now excluded via manager_mutex, we can
2064 * simply try to bind. It will succeed or fail depending
2065 * on @pool's current state. Try it and adjust
2066 * %WORKER_UNBOUND accordingly.
2068 if (worker_maybe_bind_and_lock(pool))
2069 worker->flags &= ~WORKER_UNBOUND;
2071 worker->flags |= WORKER_UNBOUND;
2076 pool->flags &= ~POOL_MANAGE_WORKERS;
2079 * Destroy and then create so that may_start_working() is true
2082 ret |= maybe_destroy_workers(pool);
2083 ret |= maybe_create_worker(pool);
2085 mutex_unlock(&pool->manager_mutex);
2086 mutex_unlock(&pool->manager_arb);
2091 * process_one_work - process single work
2093 * @work: work to process
2095 * Process @work. This function contains all the logics necessary to
2096 * process a single work including synchronization against and
2097 * interaction with other workers on the same cpu, queueing and
2098 * flushing. As long as context requirement is met, any worker can
2099 * call this function to process a work.
2102 * spin_lock_irq(pool->lock) which is released and regrabbed.
2104 static void process_one_work(struct worker *worker, struct work_struct *work)
2105 __releases(&pool->lock)
2106 __acquires(&pool->lock)
2108 struct pool_workqueue *pwq = get_work_pwq(work);
2109 struct worker_pool *pool = worker->pool;
2110 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2112 struct worker *collision;
2113 #ifdef CONFIG_LOCKDEP
2115 * It is permissible to free the struct work_struct from
2116 * inside the function that is called from it, this we need to
2117 * take into account for lockdep too. To avoid bogus "held
2118 * lock freed" warnings as well as problems when looking into
2119 * work->lockdep_map, make a copy and use that here.
2121 struct lockdep_map lockdep_map;
2123 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2126 * Ensure we're on the correct CPU. DISASSOCIATED test is
2127 * necessary to avoid spurious warnings from rescuers servicing the
2128 * unbound or a disassociated pool.
2130 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2131 !(pool->flags & POOL_DISASSOCIATED) &&
2132 raw_smp_processor_id() != pool->cpu);
2135 * A single work shouldn't be executed concurrently by
2136 * multiple workers on a single cpu. Check whether anyone is
2137 * already processing the work. If so, defer the work to the
2138 * currently executing one.
2140 collision = find_worker_executing_work(pool, work);
2141 if (unlikely(collision)) {
2142 move_linked_works(work, &collision->scheduled, NULL);
2146 /* claim and dequeue */
2147 debug_work_deactivate(work);
2148 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2149 worker->current_work = work;
2150 worker->current_func = work->func;
2151 worker->current_pwq = pwq;
2152 work_color = get_work_color(work);
2154 list_del_init(&work->entry);
2157 * CPU intensive works don't participate in concurrency
2158 * management. They're the scheduler's responsibility.
2160 if (unlikely(cpu_intensive))
2161 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2164 * Unbound pool isn't concurrency managed and work items should be
2165 * executed ASAP. Wake up another worker if necessary.
2167 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2168 wake_up_worker(pool);
2171 * Record the last pool and clear PENDING which should be the last
2172 * update to @work. Also, do this inside @pool->lock so that
2173 * PENDING and queued state changes happen together while IRQ is
2176 set_work_pool_and_clear_pending(work, pool->id);
2178 spin_unlock_irq(&pool->lock);
2180 lock_map_acquire_read(&pwq->wq->lockdep_map);
2181 lock_map_acquire(&lockdep_map);
2182 trace_workqueue_execute_start(work);
2183 worker->current_func(work);
2185 * While we must be careful to not use "work" after this, the trace
2186 * point will only record its address.
2188 trace_workqueue_execute_end(work);
2189 lock_map_release(&lockdep_map);
2190 lock_map_release(&pwq->wq->lockdep_map);
2192 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2193 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2194 " last function: %pf\n",
2195 current->comm, preempt_count(), task_pid_nr(current),
2196 worker->current_func);
2197 debug_show_held_locks(current);
2201 spin_lock_irq(&pool->lock);
2203 /* clear cpu intensive status */
2204 if (unlikely(cpu_intensive))
2205 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2207 /* we're done with it, release */
2208 hash_del(&worker->hentry);
2209 worker->current_work = NULL;
2210 worker->current_func = NULL;
2211 worker->current_pwq = NULL;
2212 pwq_dec_nr_in_flight(pwq, work_color);
2216 * process_scheduled_works - process scheduled works
2219 * Process all scheduled works. Please note that the scheduled list
2220 * may change while processing a work, so this function repeatedly
2221 * fetches a work from the top and executes it.
2224 * spin_lock_irq(pool->lock) which may be released and regrabbed
2227 static void process_scheduled_works(struct worker *worker)
2229 while (!list_empty(&worker->scheduled)) {
2230 struct work_struct *work = list_first_entry(&worker->scheduled,
2231 struct work_struct, entry);
2232 process_one_work(worker, work);
2237 * worker_thread - the worker thread function
2240 * The worker thread function. All workers belong to a worker_pool -
2241 * either a per-cpu one or dynamic unbound one. These workers process all
2242 * work items regardless of their specific target workqueue. The only
2243 * exception is work items which belong to workqueues with a rescuer which
2244 * will be explained in rescuer_thread().
2246 static int worker_thread(void *__worker)
2248 struct worker *worker = __worker;
2249 struct worker_pool *pool = worker->pool;
2251 /* tell the scheduler that this is a workqueue worker */
2252 worker->task->flags |= PF_WQ_WORKER;
2254 spin_lock_irq(&pool->lock);
2256 /* we are off idle list if destruction or rebind is requested */
2257 if (unlikely(list_empty(&worker->entry))) {
2258 spin_unlock_irq(&pool->lock);
2260 /* if DIE is set, destruction is requested */
2261 if (worker->flags & WORKER_DIE) {
2262 worker->task->flags &= ~PF_WQ_WORKER;
2266 /* otherwise, rebind */
2267 idle_worker_rebind(worker);
2271 worker_leave_idle(worker);
2273 /* no more worker necessary? */
2274 if (!need_more_worker(pool))
2277 /* do we need to manage? */
2278 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2282 * ->scheduled list can only be filled while a worker is
2283 * preparing to process a work or actually processing it.
2284 * Make sure nobody diddled with it while I was sleeping.
2286 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2289 * When control reaches this point, we're guaranteed to have
2290 * at least one idle worker or that someone else has already
2291 * assumed the manager role.
2293 worker_clr_flags(worker, WORKER_PREP);
2296 struct work_struct *work =
2297 list_first_entry(&pool->worklist,
2298 struct work_struct, entry);
2300 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2301 /* optimization path, not strictly necessary */
2302 process_one_work(worker, work);
2303 if (unlikely(!list_empty(&worker->scheduled)))
2304 process_scheduled_works(worker);
2306 move_linked_works(work, &worker->scheduled, NULL);
2307 process_scheduled_works(worker);
2309 } while (keep_working(pool));
2311 worker_set_flags(worker, WORKER_PREP, false);
2313 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2317 * pool->lock is held and there's no work to process and no need to
2318 * manage, sleep. Workers are woken up only while holding
2319 * pool->lock or from local cpu, so setting the current state
2320 * before releasing pool->lock is enough to prevent losing any
2323 worker_enter_idle(worker);
2324 __set_current_state(TASK_INTERRUPTIBLE);
2325 spin_unlock_irq(&pool->lock);
2331 * rescuer_thread - the rescuer thread function
2334 * Workqueue rescuer thread function. There's one rescuer for each
2335 * workqueue which has WQ_MEM_RECLAIM set.
2337 * Regular work processing on a pool may block trying to create a new
2338 * worker which uses GFP_KERNEL allocation which has slight chance of
2339 * developing into deadlock if some works currently on the same queue
2340 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2341 * the problem rescuer solves.
2343 * When such condition is possible, the pool summons rescuers of all
2344 * workqueues which have works queued on the pool and let them process
2345 * those works so that forward progress can be guaranteed.
2347 * This should happen rarely.
2349 static int rescuer_thread(void *__rescuer)
2351 struct worker *rescuer = __rescuer;
2352 struct workqueue_struct *wq = rescuer->rescue_wq;
2353 struct list_head *scheduled = &rescuer->scheduled;
2355 set_user_nice(current, RESCUER_NICE_LEVEL);
2358 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2359 * doesn't participate in concurrency management.
2361 rescuer->task->flags |= PF_WQ_WORKER;
2363 set_current_state(TASK_INTERRUPTIBLE);
2365 if (kthread_should_stop()) {
2366 __set_current_state(TASK_RUNNING);
2367 rescuer->task->flags &= ~PF_WQ_WORKER;
2371 /* see whether any pwq is asking for help */
2372 spin_lock_irq(&workqueue_lock);
2374 while (!list_empty(&wq->maydays)) {
2375 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2376 struct pool_workqueue, mayday_node);
2377 struct worker_pool *pool = pwq->pool;
2378 struct work_struct *work, *n;
2380 __set_current_state(TASK_RUNNING);
2381 list_del_init(&pwq->mayday_node);
2383 spin_unlock_irq(&workqueue_lock);
2385 /* migrate to the target cpu if possible */
2386 worker_maybe_bind_and_lock(pool);
2387 rescuer->pool = pool;
2390 * Slurp in all works issued via this workqueue and
2393 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2394 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2395 if (get_work_pwq(work) == pwq)
2396 move_linked_works(work, scheduled, &n);
2398 process_scheduled_works(rescuer);
2401 * Leave this pool. If keep_working() is %true, notify a
2402 * regular worker; otherwise, we end up with 0 concurrency
2403 * and stalling the execution.
2405 if (keep_working(pool))
2406 wake_up_worker(pool);
2408 rescuer->pool = NULL;
2409 spin_unlock(&pool->lock);
2410 spin_lock(&workqueue_lock);
2413 spin_unlock_irq(&workqueue_lock);
2415 /* rescuers should never participate in concurrency management */
2416 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2422 struct work_struct work;
2423 struct completion done;
2426 static void wq_barrier_func(struct work_struct *work)
2428 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2429 complete(&barr->done);
2433 * insert_wq_barrier - insert a barrier work
2434 * @pwq: pwq to insert barrier into
2435 * @barr: wq_barrier to insert
2436 * @target: target work to attach @barr to
2437 * @worker: worker currently executing @target, NULL if @target is not executing
2439 * @barr is linked to @target such that @barr is completed only after
2440 * @target finishes execution. Please note that the ordering
2441 * guarantee is observed only with respect to @target and on the local
2444 * Currently, a queued barrier can't be canceled. This is because
2445 * try_to_grab_pending() can't determine whether the work to be
2446 * grabbed is at the head of the queue and thus can't clear LINKED
2447 * flag of the previous work while there must be a valid next work
2448 * after a work with LINKED flag set.
2450 * Note that when @worker is non-NULL, @target may be modified
2451 * underneath us, so we can't reliably determine pwq from @target.
2454 * spin_lock_irq(pool->lock).
2456 static void insert_wq_barrier(struct pool_workqueue *pwq,
2457 struct wq_barrier *barr,
2458 struct work_struct *target, struct worker *worker)
2460 struct list_head *head;
2461 unsigned int linked = 0;
2464 * debugobject calls are safe here even with pool->lock locked
2465 * as we know for sure that this will not trigger any of the
2466 * checks and call back into the fixup functions where we
2469 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2470 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2471 init_completion(&barr->done);
2474 * If @target is currently being executed, schedule the
2475 * barrier to the worker; otherwise, put it after @target.
2478 head = worker->scheduled.next;
2480 unsigned long *bits = work_data_bits(target);
2482 head = target->entry.next;
2483 /* there can already be other linked works, inherit and set */
2484 linked = *bits & WORK_STRUCT_LINKED;
2485 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2488 debug_work_activate(&barr->work);
2489 insert_work(pwq, &barr->work, head,
2490 work_color_to_flags(WORK_NO_COLOR) | linked);
2494 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2495 * @wq: workqueue being flushed
2496 * @flush_color: new flush color, < 0 for no-op
2497 * @work_color: new work color, < 0 for no-op
2499 * Prepare pwqs for workqueue flushing.
2501 * If @flush_color is non-negative, flush_color on all pwqs should be
2502 * -1. If no pwq has in-flight commands at the specified color, all
2503 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2504 * has in flight commands, its pwq->flush_color is set to
2505 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2506 * wakeup logic is armed and %true is returned.
2508 * The caller should have initialized @wq->first_flusher prior to
2509 * calling this function with non-negative @flush_color. If
2510 * @flush_color is negative, no flush color update is done and %false
2513 * If @work_color is non-negative, all pwqs should have the same
2514 * work_color which is previous to @work_color and all will be
2515 * advanced to @work_color.
2518 * mutex_lock(wq->flush_mutex).
2521 * %true if @flush_color >= 0 and there's something to flush. %false
2524 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2525 int flush_color, int work_color)
2528 struct pool_workqueue *pwq;
2530 if (flush_color >= 0) {
2531 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2532 atomic_set(&wq->nr_pwqs_to_flush, 1);
2535 local_irq_disable();
2537 for_each_pwq(pwq, wq) {
2538 struct worker_pool *pool = pwq->pool;
2540 spin_lock(&pool->lock);
2542 if (flush_color >= 0) {
2543 WARN_ON_ONCE(pwq->flush_color != -1);
2545 if (pwq->nr_in_flight[flush_color]) {
2546 pwq->flush_color = flush_color;
2547 atomic_inc(&wq->nr_pwqs_to_flush);
2552 if (work_color >= 0) {
2553 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2554 pwq->work_color = work_color;
2557 spin_unlock(&pool->lock);
2562 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2563 complete(&wq->first_flusher->done);
2569 * flush_workqueue - ensure that any scheduled work has run to completion.
2570 * @wq: workqueue to flush
2572 * This function sleeps until all work items which were queued on entry
2573 * have finished execution, but it is not livelocked by new incoming ones.
2575 void flush_workqueue(struct workqueue_struct *wq)
2577 struct wq_flusher this_flusher = {
2578 .list = LIST_HEAD_INIT(this_flusher.list),
2580 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2584 lock_map_acquire(&wq->lockdep_map);
2585 lock_map_release(&wq->lockdep_map);
2587 mutex_lock(&wq->flush_mutex);
2590 * Start-to-wait phase
2592 next_color = work_next_color(wq->work_color);
2594 if (next_color != wq->flush_color) {
2596 * Color space is not full. The current work_color
2597 * becomes our flush_color and work_color is advanced
2600 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2601 this_flusher.flush_color = wq->work_color;
2602 wq->work_color = next_color;
2604 if (!wq->first_flusher) {
2605 /* no flush in progress, become the first flusher */
2606 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2608 wq->first_flusher = &this_flusher;
2610 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2612 /* nothing to flush, done */
2613 wq->flush_color = next_color;
2614 wq->first_flusher = NULL;
2619 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2620 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2621 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2625 * Oops, color space is full, wait on overflow queue.
2626 * The next flush completion will assign us
2627 * flush_color and transfer to flusher_queue.
2629 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2632 mutex_unlock(&wq->flush_mutex);
2634 wait_for_completion(&this_flusher.done);
2637 * Wake-up-and-cascade phase
2639 * First flushers are responsible for cascading flushes and
2640 * handling overflow. Non-first flushers can simply return.
2642 if (wq->first_flusher != &this_flusher)
2645 mutex_lock(&wq->flush_mutex);
2647 /* we might have raced, check again with mutex held */
2648 if (wq->first_flusher != &this_flusher)
2651 wq->first_flusher = NULL;
2653 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2654 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2657 struct wq_flusher *next, *tmp;
2659 /* complete all the flushers sharing the current flush color */
2660 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2661 if (next->flush_color != wq->flush_color)
2663 list_del_init(&next->list);
2664 complete(&next->done);
2667 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2668 wq->flush_color != work_next_color(wq->work_color));
2670 /* this flush_color is finished, advance by one */
2671 wq->flush_color = work_next_color(wq->flush_color);
2673 /* one color has been freed, handle overflow queue */
2674 if (!list_empty(&wq->flusher_overflow)) {
2676 * Assign the same color to all overflowed
2677 * flushers, advance work_color and append to
2678 * flusher_queue. This is the start-to-wait
2679 * phase for these overflowed flushers.
2681 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2682 tmp->flush_color = wq->work_color;
2684 wq->work_color = work_next_color(wq->work_color);
2686 list_splice_tail_init(&wq->flusher_overflow,
2687 &wq->flusher_queue);
2688 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2691 if (list_empty(&wq->flusher_queue)) {
2692 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2697 * Need to flush more colors. Make the next flusher
2698 * the new first flusher and arm pwqs.
2700 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2701 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2703 list_del_init(&next->list);
2704 wq->first_flusher = next;
2706 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2710 * Meh... this color is already done, clear first
2711 * flusher and repeat cascading.
2713 wq->first_flusher = NULL;
2717 mutex_unlock(&wq->flush_mutex);
2719 EXPORT_SYMBOL_GPL(flush_workqueue);
2722 * drain_workqueue - drain a workqueue
2723 * @wq: workqueue to drain
2725 * Wait until the workqueue becomes empty. While draining is in progress,
2726 * only chain queueing is allowed. IOW, only currently pending or running
2727 * work items on @wq can queue further work items on it. @wq is flushed
2728 * repeatedly until it becomes empty. The number of flushing is detemined
2729 * by the depth of chaining and should be relatively short. Whine if it
2732 void drain_workqueue(struct workqueue_struct *wq)
2734 unsigned int flush_cnt = 0;
2735 struct pool_workqueue *pwq;
2738 * __queue_work() needs to test whether there are drainers, is much
2739 * hotter than drain_workqueue() and already looks at @wq->flags.
2740 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2742 spin_lock_irq(&workqueue_lock);
2743 if (!wq->nr_drainers++)
2744 wq->flags |= __WQ_DRAINING;
2745 spin_unlock_irq(&workqueue_lock);
2747 flush_workqueue(wq);
2749 local_irq_disable();
2751 for_each_pwq(pwq, wq) {
2754 spin_lock(&pwq->pool->lock);
2755 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2756 spin_unlock(&pwq->pool->lock);
2761 if (++flush_cnt == 10 ||
2762 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2763 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2764 wq->name, flush_cnt);
2770 spin_lock(&workqueue_lock);
2771 if (!--wq->nr_drainers)
2772 wq->flags &= ~__WQ_DRAINING;
2773 spin_unlock(&workqueue_lock);
2777 EXPORT_SYMBOL_GPL(drain_workqueue);
2779 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2781 struct worker *worker = NULL;
2782 struct worker_pool *pool;
2783 struct pool_workqueue *pwq;
2787 local_irq_disable();
2788 pool = get_work_pool(work);
2794 spin_lock(&pool->lock);
2795 /* see the comment in try_to_grab_pending() with the same code */
2796 pwq = get_work_pwq(work);
2798 if (unlikely(pwq->pool != pool))
2801 worker = find_worker_executing_work(pool, work);
2804 pwq = worker->current_pwq;
2807 insert_wq_barrier(pwq, barr, work, worker);
2808 spin_unlock_irq(&pool->lock);
2811 * If @max_active is 1 or rescuer is in use, flushing another work
2812 * item on the same workqueue may lead to deadlock. Make sure the
2813 * flusher is not running on the same workqueue by verifying write
2816 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2817 lock_map_acquire(&pwq->wq->lockdep_map);
2819 lock_map_acquire_read(&pwq->wq->lockdep_map);
2820 lock_map_release(&pwq->wq->lockdep_map);
2824 spin_unlock_irq(&pool->lock);
2829 * flush_work - wait for a work to finish executing the last queueing instance
2830 * @work: the work to flush
2832 * Wait until @work has finished execution. @work is guaranteed to be idle
2833 * on return if it hasn't been requeued since flush started.
2836 * %true if flush_work() waited for the work to finish execution,
2837 * %false if it was already idle.
2839 bool flush_work(struct work_struct *work)
2841 struct wq_barrier barr;
2843 lock_map_acquire(&work->lockdep_map);
2844 lock_map_release(&work->lockdep_map);
2846 if (start_flush_work(work, &barr)) {
2847 wait_for_completion(&barr.done);
2848 destroy_work_on_stack(&barr.work);
2854 EXPORT_SYMBOL_GPL(flush_work);
2856 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2858 unsigned long flags;
2862 ret = try_to_grab_pending(work, is_dwork, &flags);
2864 * If someone else is canceling, wait for the same event it
2865 * would be waiting for before retrying.
2867 if (unlikely(ret == -ENOENT))
2869 } while (unlikely(ret < 0));
2871 /* tell other tasks trying to grab @work to back off */
2872 mark_work_canceling(work);
2873 local_irq_restore(flags);
2876 clear_work_data(work);
2881 * cancel_work_sync - cancel a work and wait for it to finish
2882 * @work: the work to cancel
2884 * Cancel @work and wait for its execution to finish. This function
2885 * can be used even if the work re-queues itself or migrates to
2886 * another workqueue. On return from this function, @work is
2887 * guaranteed to be not pending or executing on any CPU.
2889 * cancel_work_sync(&delayed_work->work) must not be used for
2890 * delayed_work's. Use cancel_delayed_work_sync() instead.
2892 * The caller must ensure that the workqueue on which @work was last
2893 * queued can't be destroyed before this function returns.
2896 * %true if @work was pending, %false otherwise.
2898 bool cancel_work_sync(struct work_struct *work)
2900 return __cancel_work_timer(work, false);
2902 EXPORT_SYMBOL_GPL(cancel_work_sync);
2905 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2906 * @dwork: the delayed work to flush
2908 * Delayed timer is cancelled and the pending work is queued for
2909 * immediate execution. Like flush_work(), this function only
2910 * considers the last queueing instance of @dwork.
2913 * %true if flush_work() waited for the work to finish execution,
2914 * %false if it was already idle.
2916 bool flush_delayed_work(struct delayed_work *dwork)
2918 local_irq_disable();
2919 if (del_timer_sync(&dwork->timer))
2920 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2922 return flush_work(&dwork->work);
2924 EXPORT_SYMBOL(flush_delayed_work);
2927 * cancel_delayed_work - cancel a delayed work
2928 * @dwork: delayed_work to cancel
2930 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2931 * and canceled; %false if wasn't pending. Note that the work callback
2932 * function may still be running on return, unless it returns %true and the
2933 * work doesn't re-arm itself. Explicitly flush or use
2934 * cancel_delayed_work_sync() to wait on it.
2936 * This function is safe to call from any context including IRQ handler.
2938 bool cancel_delayed_work(struct delayed_work *dwork)
2940 unsigned long flags;
2944 ret = try_to_grab_pending(&dwork->work, true, &flags);
2945 } while (unlikely(ret == -EAGAIN));
2947 if (unlikely(ret < 0))
2950 set_work_pool_and_clear_pending(&dwork->work,
2951 get_work_pool_id(&dwork->work));
2952 local_irq_restore(flags);
2955 EXPORT_SYMBOL(cancel_delayed_work);
2958 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2959 * @dwork: the delayed work cancel
2961 * This is cancel_work_sync() for delayed works.
2964 * %true if @dwork was pending, %false otherwise.
2966 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2968 return __cancel_work_timer(&dwork->work, true);
2970 EXPORT_SYMBOL(cancel_delayed_work_sync);
2973 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2974 * @func: the function to call
2976 * schedule_on_each_cpu() executes @func on each online CPU using the
2977 * system workqueue and blocks until all CPUs have completed.
2978 * schedule_on_each_cpu() is very slow.
2981 * 0 on success, -errno on failure.
2983 int schedule_on_each_cpu(work_func_t func)
2986 struct work_struct __percpu *works;
2988 works = alloc_percpu(struct work_struct);
2994 for_each_online_cpu(cpu) {
2995 struct work_struct *work = per_cpu_ptr(works, cpu);
2997 INIT_WORK(work, func);
2998 schedule_work_on(cpu, work);
3001 for_each_online_cpu(cpu)
3002 flush_work(per_cpu_ptr(works, cpu));
3010 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3012 * Forces execution of the kernel-global workqueue and blocks until its
3015 * Think twice before calling this function! It's very easy to get into
3016 * trouble if you don't take great care. Either of the following situations
3017 * will lead to deadlock:
3019 * One of the work items currently on the workqueue needs to acquire
3020 * a lock held by your code or its caller.
3022 * Your code is running in the context of a work routine.
3024 * They will be detected by lockdep when they occur, but the first might not
3025 * occur very often. It depends on what work items are on the workqueue and
3026 * what locks they need, which you have no control over.
3028 * In most situations flushing the entire workqueue is overkill; you merely
3029 * need to know that a particular work item isn't queued and isn't running.
3030 * In such cases you should use cancel_delayed_work_sync() or
3031 * cancel_work_sync() instead.
3033 void flush_scheduled_work(void)
3035 flush_workqueue(system_wq);
3037 EXPORT_SYMBOL(flush_scheduled_work);
3040 * execute_in_process_context - reliably execute the routine with user context
3041 * @fn: the function to execute
3042 * @ew: guaranteed storage for the execute work structure (must
3043 * be available when the work executes)
3045 * Executes the function immediately if process context is available,
3046 * otherwise schedules the function for delayed execution.
3048 * Returns: 0 - function was executed
3049 * 1 - function was scheduled for execution
3051 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3053 if (!in_interrupt()) {
3058 INIT_WORK(&ew->work, fn);
3059 schedule_work(&ew->work);
3063 EXPORT_SYMBOL_GPL(execute_in_process_context);
3067 * Workqueues with WQ_SYSFS flag set is visible to userland via
3068 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3069 * following attributes.
3071 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3072 * max_active RW int : maximum number of in-flight work items
3074 * Unbound workqueues have the following extra attributes.
3076 * id RO int : the associated pool ID
3077 * nice RW int : nice value of the workers
3078 * cpumask RW mask : bitmask of allowed CPUs for the workers
3081 struct workqueue_struct *wq;
3085 static struct workqueue_struct *dev_to_wq(struct device *dev)
3087 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3092 static ssize_t wq_per_cpu_show(struct device *dev,
3093 struct device_attribute *attr, char *buf)
3095 struct workqueue_struct *wq = dev_to_wq(dev);
3097 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3100 static ssize_t wq_max_active_show(struct device *dev,
3101 struct device_attribute *attr, char *buf)
3103 struct workqueue_struct *wq = dev_to_wq(dev);
3105 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3108 static ssize_t wq_max_active_store(struct device *dev,
3109 struct device_attribute *attr,
3110 const char *buf, size_t count)
3112 struct workqueue_struct *wq = dev_to_wq(dev);
3115 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3118 workqueue_set_max_active(wq, val);
3122 static struct device_attribute wq_sysfs_attrs[] = {
3123 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3124 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3128 static ssize_t wq_pool_id_show(struct device *dev,
3129 struct device_attribute *attr, char *buf)
3131 struct workqueue_struct *wq = dev_to_wq(dev);
3132 struct worker_pool *pool;
3135 rcu_read_lock_sched();
3136 pool = first_pwq(wq)->pool;
3137 written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id);
3138 rcu_read_unlock_sched();
3143 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3146 struct workqueue_struct *wq = dev_to_wq(dev);
3149 rcu_read_lock_sched();
3150 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3151 first_pwq(wq)->pool->attrs->nice);
3152 rcu_read_unlock_sched();
3157 /* prepare workqueue_attrs for sysfs store operations */
3158 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3160 struct workqueue_attrs *attrs;
3162 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3166 rcu_read_lock_sched();
3167 copy_workqueue_attrs(attrs, first_pwq(wq)->pool->attrs);
3168 rcu_read_unlock_sched();
3172 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3173 const char *buf, size_t count)
3175 struct workqueue_struct *wq = dev_to_wq(dev);
3176 struct workqueue_attrs *attrs;
3179 attrs = wq_sysfs_prep_attrs(wq);
3183 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3184 attrs->nice >= -20 && attrs->nice <= 19)
3185 ret = apply_workqueue_attrs(wq, attrs);
3189 free_workqueue_attrs(attrs);
3190 return ret ?: count;
3193 static ssize_t wq_cpumask_show(struct device *dev,
3194 struct device_attribute *attr, char *buf)
3196 struct workqueue_struct *wq = dev_to_wq(dev);
3199 rcu_read_lock_sched();
3200 written = cpumask_scnprintf(buf, PAGE_SIZE,
3201 first_pwq(wq)->pool->attrs->cpumask);
3202 rcu_read_unlock_sched();
3204 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3208 static ssize_t wq_cpumask_store(struct device *dev,
3209 struct device_attribute *attr,
3210 const char *buf, size_t count)
3212 struct workqueue_struct *wq = dev_to_wq(dev);
3213 struct workqueue_attrs *attrs;
3216 attrs = wq_sysfs_prep_attrs(wq);
3220 ret = cpumask_parse(buf, attrs->cpumask);
3222 ret = apply_workqueue_attrs(wq, attrs);
3224 free_workqueue_attrs(attrs);
3225 return ret ?: count;
3228 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3229 __ATTR(pool_id, 0444, wq_pool_id_show, NULL),
3230 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3231 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3235 static struct bus_type wq_subsys = {
3236 .name = "workqueue",
3237 .dev_attrs = wq_sysfs_attrs,
3240 static int __init wq_sysfs_init(void)
3242 return subsys_virtual_register(&wq_subsys, NULL);
3244 core_initcall(wq_sysfs_init);
3246 static void wq_device_release(struct device *dev)
3248 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3254 * workqueue_sysfs_register - make a workqueue visible in sysfs
3255 * @wq: the workqueue to register
3257 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3258 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3259 * which is the preferred method.
3261 * Workqueue user should use this function directly iff it wants to apply
3262 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3263 * apply_workqueue_attrs() may race against userland updating the
3266 * Returns 0 on success, -errno on failure.
3268 int workqueue_sysfs_register(struct workqueue_struct *wq)
3270 struct wq_device *wq_dev;
3274 * Adjusting max_active or creating new pwqs by applyting
3275 * attributes breaks ordering guarantee. Disallow exposing ordered
3278 if (WARN_ON(wq->flags & __WQ_ORDERED))
3281 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3286 wq_dev->dev.bus = &wq_subsys;
3287 wq_dev->dev.init_name = wq->name;
3288 wq_dev->dev.release = wq_device_release;
3291 * unbound_attrs are created separately. Suppress uevent until
3292 * everything is ready.
3294 dev_set_uevent_suppress(&wq_dev->dev, true);
3296 ret = device_register(&wq_dev->dev);
3303 if (wq->flags & WQ_UNBOUND) {
3304 struct device_attribute *attr;
3306 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3307 ret = device_create_file(&wq_dev->dev, attr);
3309 device_unregister(&wq_dev->dev);
3316 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3321 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3322 * @wq: the workqueue to unregister
3324 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3326 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3328 struct wq_device *wq_dev = wq->wq_dev;
3334 device_unregister(&wq_dev->dev);
3336 #else /* CONFIG_SYSFS */
3337 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3338 #endif /* CONFIG_SYSFS */
3341 * free_workqueue_attrs - free a workqueue_attrs
3342 * @attrs: workqueue_attrs to free
3344 * Undo alloc_workqueue_attrs().
3346 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3349 free_cpumask_var(attrs->cpumask);
3355 * alloc_workqueue_attrs - allocate a workqueue_attrs
3356 * @gfp_mask: allocation mask to use
3358 * Allocate a new workqueue_attrs, initialize with default settings and
3359 * return it. Returns NULL on failure.
3361 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3363 struct workqueue_attrs *attrs;
3365 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3368 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3371 cpumask_setall(attrs->cpumask);
3374 free_workqueue_attrs(attrs);
3378 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3379 const struct workqueue_attrs *from)
3381 to->nice = from->nice;
3382 cpumask_copy(to->cpumask, from->cpumask);
3386 * Hacky implementation of jhash of bitmaps which only considers the
3387 * specified number of bits. We probably want a proper implementation in
3388 * include/linux/jhash.h.
3390 static u32 jhash_bitmap(const unsigned long *bitmap, int bits, u32 hash)
3392 int nr_longs = bits / BITS_PER_LONG;
3393 int nr_leftover = bits % BITS_PER_LONG;
3394 unsigned long leftover = 0;
3397 hash = jhash(bitmap, nr_longs * sizeof(long), hash);
3399 bitmap_copy(&leftover, bitmap + nr_longs, nr_leftover);
3400 hash = jhash(&leftover, sizeof(long), hash);
3405 /* hash value of the content of @attr */
3406 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3410 hash = jhash_1word(attrs->nice, hash);
3411 hash = jhash_bitmap(cpumask_bits(attrs->cpumask), nr_cpu_ids, hash);
3415 /* content equality test */
3416 static bool wqattrs_equal(const struct workqueue_attrs *a,
3417 const struct workqueue_attrs *b)
3419 if (a->nice != b->nice)
3421 if (!cpumask_equal(a->cpumask, b->cpumask))
3427 * init_worker_pool - initialize a newly zalloc'd worker_pool
3428 * @pool: worker_pool to initialize
3430 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3431 * Returns 0 on success, -errno on failure. Even on failure, all fields
3432 * inside @pool proper are initialized and put_unbound_pool() can be called
3433 * on @pool safely to release it.
3435 static int init_worker_pool(struct worker_pool *pool)
3437 spin_lock_init(&pool->lock);
3440 pool->flags |= POOL_DISASSOCIATED;
3441 INIT_LIST_HEAD(&pool->worklist);
3442 INIT_LIST_HEAD(&pool->idle_list);
3443 hash_init(pool->busy_hash);
3445 init_timer_deferrable(&pool->idle_timer);
3446 pool->idle_timer.function = idle_worker_timeout;
3447 pool->idle_timer.data = (unsigned long)pool;
3449 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3450 (unsigned long)pool);
3452 mutex_init(&pool->manager_arb);
3453 mutex_init(&pool->manager_mutex);
3454 ida_init(&pool->worker_ida);
3456 INIT_HLIST_NODE(&pool->hash_node);
3459 /* shouldn't fail above this point */
3460 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3466 static void rcu_free_pool(struct rcu_head *rcu)
3468 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3470 ida_destroy(&pool->worker_ida);
3471 free_workqueue_attrs(pool->attrs);
3476 * put_unbound_pool - put a worker_pool
3477 * @pool: worker_pool to put
3479 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3480 * safe manner. get_unbound_pool() calls this function on its failure path
3481 * and this function should be able to release pools which went through,
3482 * successfully or not, init_worker_pool().
3484 static void put_unbound_pool(struct worker_pool *pool)
3486 struct worker *worker;
3488 spin_lock_irq(&workqueue_lock);
3489 if (--pool->refcnt) {
3490 spin_unlock_irq(&workqueue_lock);
3495 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3496 WARN_ON(!list_empty(&pool->worklist))) {
3497 spin_unlock_irq(&workqueue_lock);
3501 /* release id and unhash */
3503 idr_remove(&worker_pool_idr, pool->id);
3504 hash_del(&pool->hash_node);
3506 spin_unlock_irq(&workqueue_lock);
3509 * Become the manager and destroy all workers. Grabbing
3510 * manager_arb prevents @pool's workers from blocking on
3513 mutex_lock(&pool->manager_arb);
3514 spin_lock_irq(&pool->lock);
3516 while ((worker = first_worker(pool)))
3517 destroy_worker(worker);
3518 WARN_ON(pool->nr_workers || pool->nr_idle);
3520 spin_unlock_irq(&pool->lock);
3521 mutex_unlock(&pool->manager_arb);
3523 /* shut down the timers */
3524 del_timer_sync(&pool->idle_timer);
3525 del_timer_sync(&pool->mayday_timer);
3527 /* sched-RCU protected to allow dereferences from get_work_pool() */
3528 call_rcu_sched(&pool->rcu, rcu_free_pool);
3532 * get_unbound_pool - get a worker_pool with the specified attributes
3533 * @attrs: the attributes of the worker_pool to get
3535 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3536 * reference count and return it. If there already is a matching
3537 * worker_pool, it will be used; otherwise, this function attempts to
3538 * create a new one. On failure, returns NULL.
3540 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3542 static DEFINE_MUTEX(create_mutex);
3543 u32 hash = wqattrs_hash(attrs);
3544 struct worker_pool *pool;
3545 struct worker *worker;
3547 mutex_lock(&create_mutex);
3549 /* do we already have a matching pool? */
3550 spin_lock_irq(&workqueue_lock);
3551 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3552 if (wqattrs_equal(pool->attrs, attrs)) {
3557 spin_unlock_irq(&workqueue_lock);
3559 /* nope, create a new one */
3560 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3561 if (!pool || init_worker_pool(pool) < 0)
3564 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3565 copy_workqueue_attrs(pool->attrs, attrs);
3567 if (worker_pool_assign_id(pool) < 0)
3570 /* create and start the initial worker */
3571 worker = create_worker(pool);
3575 spin_lock_irq(&pool->lock);
3576 start_worker(worker);
3577 spin_unlock_irq(&pool->lock);
3580 spin_lock_irq(&workqueue_lock);
3581 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3583 spin_unlock_irq(&workqueue_lock);
3584 mutex_unlock(&create_mutex);
3587 mutex_unlock(&create_mutex);
3589 put_unbound_pool(pool);
3593 static void rcu_free_pwq(struct rcu_head *rcu)
3595 kmem_cache_free(pwq_cache,
3596 container_of(rcu, struct pool_workqueue, rcu));
3600 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3601 * and needs to be destroyed.
3603 static void pwq_unbound_release_workfn(struct work_struct *work)
3605 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3606 unbound_release_work);
3607 struct workqueue_struct *wq = pwq->wq;
3608 struct worker_pool *pool = pwq->pool;
3610 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3614 * Unlink @pwq. Synchronization against flush_mutex isn't strictly
3615 * necessary on release but do it anyway. It's easier to verify
3616 * and consistent with the linking path.
3618 mutex_lock(&wq->flush_mutex);
3619 spin_lock_irq(&workqueue_lock);
3620 list_del_rcu(&pwq->pwqs_node);
3621 spin_unlock_irq(&workqueue_lock);
3622 mutex_unlock(&wq->flush_mutex);
3624 put_unbound_pool(pool);
3625 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3628 * If we're the last pwq going away, @wq is already dead and no one
3629 * is gonna access it anymore. Free it.
3631 if (list_empty(&wq->pwqs))
3636 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3637 * @pwq: target pool_workqueue
3639 * If @pwq isn't freezing, set @pwq->max_active to the associated
3640 * workqueue's saved_max_active and activate delayed work items
3641 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3643 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3645 struct workqueue_struct *wq = pwq->wq;
3646 bool freezable = wq->flags & WQ_FREEZABLE;
3648 /* for @wq->saved_max_active */
3649 lockdep_assert_held(&workqueue_lock);
3651 /* fast exit for non-freezable wqs */
3652 if (!freezable && pwq->max_active == wq->saved_max_active)
3655 spin_lock(&pwq->pool->lock);
3657 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3658 pwq->max_active = wq->saved_max_active;
3660 while (!list_empty(&pwq->delayed_works) &&
3661 pwq->nr_active < pwq->max_active)
3662 pwq_activate_first_delayed(pwq);
3664 pwq->max_active = 0;
3667 spin_unlock(&pwq->pool->lock);
3670 static void init_and_link_pwq(struct pool_workqueue *pwq,
3671 struct workqueue_struct *wq,
3672 struct worker_pool *pool,
3673 struct pool_workqueue **p_last_pwq)
3675 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3679 pwq->flush_color = -1;
3681 INIT_LIST_HEAD(&pwq->delayed_works);
3682 INIT_LIST_HEAD(&pwq->mayday_node);
3683 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3685 mutex_lock(&wq->flush_mutex);
3686 spin_lock_irq(&workqueue_lock);
3689 * Set the matching work_color. This is synchronized with
3690 * flush_mutex to avoid confusing flush_workqueue().
3693 *p_last_pwq = first_pwq(wq);
3694 pwq->work_color = wq->work_color;
3696 /* sync max_active to the current setting */
3697 pwq_adjust_max_active(pwq);
3700 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3702 spin_unlock_irq(&workqueue_lock);
3703 mutex_unlock(&wq->flush_mutex);
3707 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3708 * @wq: the target workqueue
3709 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3711 * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the
3712 * current attributes, a new pwq is created and made the first pwq which
3713 * will serve all new work items. Older pwqs are released as in-flight
3714 * work items finish. Note that a work item which repeatedly requeues
3715 * itself back-to-back will stay on its current pwq.
3717 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3720 int apply_workqueue_attrs(struct workqueue_struct *wq,
3721 const struct workqueue_attrs *attrs)
3723 struct pool_workqueue *pwq, *last_pwq;
3724 struct worker_pool *pool;
3726 /* only unbound workqueues can change attributes */
3727 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3730 /* creating multiple pwqs breaks ordering guarantee */
3731 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3734 pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL);
3738 pool = get_unbound_pool(attrs);
3740 kmem_cache_free(pwq_cache, pwq);
3744 init_and_link_pwq(pwq, wq, pool, &last_pwq);
3746 spin_lock_irq(&last_pwq->pool->lock);
3748 spin_unlock_irq(&last_pwq->pool->lock);
3754 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3756 bool highpri = wq->flags & WQ_HIGHPRI;
3759 if (!(wq->flags & WQ_UNBOUND)) {
3760 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3764 for_each_possible_cpu(cpu) {
3765 struct pool_workqueue *pwq =
3766 per_cpu_ptr(wq->cpu_pwqs, cpu);
3767 struct worker_pool *cpu_pools =
3768 per_cpu(cpu_worker_pools, cpu);
3770 init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL);
3774 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3778 static int wq_clamp_max_active(int max_active, unsigned int flags,
3781 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3783 if (max_active < 1 || max_active > lim)
3784 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3785 max_active, name, 1, lim);
3787 return clamp_val(max_active, 1, lim);
3790 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3793 struct lock_class_key *key,
3794 const char *lock_name, ...)
3796 va_list args, args1;
3797 struct workqueue_struct *wq;
3798 struct pool_workqueue *pwq;
3801 /* determine namelen, allocate wq and format name */
3802 va_start(args, lock_name);
3803 va_copy(args1, args);
3804 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3806 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3810 vsnprintf(wq->name, namelen, fmt, args1);
3814 max_active = max_active ?: WQ_DFL_ACTIVE;
3815 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3819 wq->saved_max_active = max_active;
3820 mutex_init(&wq->flush_mutex);
3821 atomic_set(&wq->nr_pwqs_to_flush, 0);
3822 INIT_LIST_HEAD(&wq->pwqs);
3823 INIT_LIST_HEAD(&wq->flusher_queue);
3824 INIT_LIST_HEAD(&wq->flusher_overflow);
3825 INIT_LIST_HEAD(&wq->maydays);
3827 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3828 INIT_LIST_HEAD(&wq->list);
3830 if (alloc_and_link_pwqs(wq) < 0)
3834 * Workqueues which may be used during memory reclaim should
3835 * have a rescuer to guarantee forward progress.
3837 if (flags & WQ_MEM_RECLAIM) {
3838 struct worker *rescuer;
3840 rescuer = alloc_worker();
3844 rescuer->rescue_wq = wq;
3845 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3847 if (IS_ERR(rescuer->task)) {
3852 wq->rescuer = rescuer;
3853 rescuer->task->flags |= PF_THREAD_BOUND;
3854 wake_up_process(rescuer->task);
3857 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3861 * workqueue_lock protects global freeze state and workqueues list.
3862 * Grab it, adjust max_active and add the new workqueue to
3865 spin_lock_irq(&workqueue_lock);
3867 for_each_pwq(pwq, wq)
3868 pwq_adjust_max_active(pwq);
3870 list_add(&wq->list, &workqueues);
3872 spin_unlock_irq(&workqueue_lock);
3880 destroy_workqueue(wq);
3883 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3886 * destroy_workqueue - safely terminate a workqueue
3887 * @wq: target workqueue
3889 * Safely destroy a workqueue. All work currently pending will be done first.
3891 void destroy_workqueue(struct workqueue_struct *wq)
3893 struct pool_workqueue *pwq;
3895 /* drain it before proceeding with destruction */
3896 drain_workqueue(wq);
3898 spin_lock_irq(&workqueue_lock);
3901 for_each_pwq(pwq, wq) {
3904 for (i = 0; i < WORK_NR_COLORS; i++) {
3905 if (WARN_ON(pwq->nr_in_flight[i])) {
3906 spin_unlock_irq(&workqueue_lock);
3911 if (WARN_ON(pwq->refcnt > 1) ||
3912 WARN_ON(pwq->nr_active) ||
3913 WARN_ON(!list_empty(&pwq->delayed_works))) {
3914 spin_unlock_irq(&workqueue_lock);
3920 * wq list is used to freeze wq, remove from list after
3921 * flushing is complete in case freeze races us.
3923 list_del_init(&wq->list);
3925 spin_unlock_irq(&workqueue_lock);
3927 workqueue_sysfs_unregister(wq);
3930 kthread_stop(wq->rescuer->task);
3935 if (!(wq->flags & WQ_UNBOUND)) {
3937 * The base ref is never dropped on per-cpu pwqs. Directly
3938 * free the pwqs and wq.
3940 free_percpu(wq->cpu_pwqs);
3944 * We're the sole accessor of @wq at this point. Directly
3945 * access the first pwq and put the base ref. As both pwqs
3946 * and pools are sched-RCU protected, the lock operations
3947 * are safe. @wq will be freed when the last pwq is
3950 pwq = list_first_entry(&wq->pwqs, struct pool_workqueue,
3952 spin_lock_irq(&pwq->pool->lock);
3954 spin_unlock_irq(&pwq->pool->lock);
3957 EXPORT_SYMBOL_GPL(destroy_workqueue);
3960 * workqueue_set_max_active - adjust max_active of a workqueue
3961 * @wq: target workqueue
3962 * @max_active: new max_active value.
3964 * Set max_active of @wq to @max_active.
3967 * Don't call from IRQ context.
3969 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3971 struct pool_workqueue *pwq;
3973 /* disallow meddling with max_active for ordered workqueues */
3974 if (WARN_ON(wq->flags & __WQ_ORDERED))
3977 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3979 spin_lock_irq(&workqueue_lock);
3981 wq->saved_max_active = max_active;
3983 for_each_pwq(pwq, wq)
3984 pwq_adjust_max_active(pwq);
3986 spin_unlock_irq(&workqueue_lock);
3988 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3991 * current_is_workqueue_rescuer - is %current workqueue rescuer?
3993 * Determine whether %current is a workqueue rescuer. Can be used from
3994 * work functions to determine whether it's being run off the rescuer task.
3996 bool current_is_workqueue_rescuer(void)
3998 struct worker *worker = current_wq_worker();
4000 return worker && worker == worker->current_pwq->wq->rescuer;
4004 * workqueue_congested - test whether a workqueue is congested
4005 * @cpu: CPU in question
4006 * @wq: target workqueue
4008 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4009 * no synchronization around this function and the test result is
4010 * unreliable and only useful as advisory hints or for debugging.
4013 * %true if congested, %false otherwise.
4015 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4017 struct pool_workqueue *pwq;
4022 if (!(wq->flags & WQ_UNBOUND))
4023 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4025 pwq = first_pwq(wq);
4027 ret = !list_empty(&pwq->delayed_works);
4032 EXPORT_SYMBOL_GPL(workqueue_congested);
4035 * work_busy - test whether a work is currently pending or running
4036 * @work: the work to be tested
4038 * Test whether @work is currently pending or running. There is no
4039 * synchronization around this function and the test result is
4040 * unreliable and only useful as advisory hints or for debugging.
4043 * OR'd bitmask of WORK_BUSY_* bits.
4045 unsigned int work_busy(struct work_struct *work)
4047 struct worker_pool *pool;
4048 unsigned long flags;
4049 unsigned int ret = 0;
4051 if (work_pending(work))
4052 ret |= WORK_BUSY_PENDING;
4054 local_irq_save(flags);
4055 pool = get_work_pool(work);
4057 spin_lock(&pool->lock);
4058 if (find_worker_executing_work(pool, work))
4059 ret |= WORK_BUSY_RUNNING;
4060 spin_unlock(&pool->lock);
4062 local_irq_restore(flags);
4066 EXPORT_SYMBOL_GPL(work_busy);
4071 * There are two challenges in supporting CPU hotplug. Firstly, there
4072 * are a lot of assumptions on strong associations among work, pwq and
4073 * pool which make migrating pending and scheduled works very
4074 * difficult to implement without impacting hot paths. Secondly,
4075 * worker pools serve mix of short, long and very long running works making
4076 * blocked draining impractical.
4078 * This is solved by allowing the pools to be disassociated from the CPU
4079 * running as an unbound one and allowing it to be reattached later if the
4080 * cpu comes back online.
4083 static void wq_unbind_fn(struct work_struct *work)
4085 int cpu = smp_processor_id();
4086 struct worker_pool *pool;
4087 struct worker *worker;
4090 for_each_cpu_worker_pool(pool, cpu) {
4091 WARN_ON_ONCE(cpu != smp_processor_id());
4093 mutex_lock(&pool->manager_mutex);
4094 spin_lock_irq(&pool->lock);
4097 * We've blocked all manager operations. Make all workers
4098 * unbound and set DISASSOCIATED. Before this, all workers
4099 * except for the ones which are still executing works from
4100 * before the last CPU down must be on the cpu. After
4101 * this, they may become diasporas.
4103 list_for_each_entry(worker, &pool->idle_list, entry)
4104 worker->flags |= WORKER_UNBOUND;
4106 for_each_busy_worker(worker, i, pool)
4107 worker->flags |= WORKER_UNBOUND;
4109 pool->flags |= POOL_DISASSOCIATED;
4111 spin_unlock_irq(&pool->lock);
4112 mutex_unlock(&pool->manager_mutex);
4116 * Call schedule() so that we cross rq->lock and thus can guarantee
4117 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
4118 * as scheduler callbacks may be invoked from other cpus.
4123 * Sched callbacks are disabled now. Zap nr_running. After this,
4124 * nr_running stays zero and need_more_worker() and keep_working()
4125 * are always true as long as the worklist is not empty. Pools on
4126 * @cpu now behave as unbound (in terms of concurrency management)
4127 * pools which are served by workers tied to the CPU.
4129 * On return from this function, the current worker would trigger
4130 * unbound chain execution of pending work items if other workers
4133 for_each_cpu_worker_pool(pool, cpu)
4134 atomic_set(&pool->nr_running, 0);
4138 * Workqueues should be brought up before normal priority CPU notifiers.
4139 * This will be registered high priority CPU notifier.
4141 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
4142 unsigned long action,
4145 int cpu = (unsigned long)hcpu;
4146 struct worker_pool *pool;
4148 switch (action & ~CPU_TASKS_FROZEN) {
4149 case CPU_UP_PREPARE:
4150 for_each_cpu_worker_pool(pool, cpu) {
4151 struct worker *worker;
4153 if (pool->nr_workers)
4156 worker = create_worker(pool);
4160 spin_lock_irq(&pool->lock);
4161 start_worker(worker);
4162 spin_unlock_irq(&pool->lock);
4166 case CPU_DOWN_FAILED:
4168 for_each_cpu_worker_pool(pool, cpu) {
4169 mutex_lock(&pool->manager_mutex);
4170 spin_lock_irq(&pool->lock);
4172 pool->flags &= ~POOL_DISASSOCIATED;
4173 rebind_workers(pool);
4175 spin_unlock_irq(&pool->lock);
4176 mutex_unlock(&pool->manager_mutex);
4184 * Workqueues should be brought down after normal priority CPU notifiers.
4185 * This will be registered as low priority CPU notifier.
4187 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
4188 unsigned long action,
4191 int cpu = (unsigned long)hcpu;
4192 struct work_struct unbind_work;
4194 switch (action & ~CPU_TASKS_FROZEN) {
4195 case CPU_DOWN_PREPARE:
4196 /* unbinding should happen on the local CPU */
4197 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4198 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4199 flush_work(&unbind_work);
4207 struct work_for_cpu {
4208 struct work_struct work;
4214 static void work_for_cpu_fn(struct work_struct *work)
4216 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4218 wfc->ret = wfc->fn(wfc->arg);
4222 * work_on_cpu - run a function in user context on a particular cpu
4223 * @cpu: the cpu to run on
4224 * @fn: the function to run
4225 * @arg: the function arg
4227 * This will return the value @fn returns.
4228 * It is up to the caller to ensure that the cpu doesn't go offline.
4229 * The caller must not hold any locks which would prevent @fn from completing.
4231 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4233 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4235 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4236 schedule_work_on(cpu, &wfc.work);
4237 flush_work(&wfc.work);
4240 EXPORT_SYMBOL_GPL(work_on_cpu);
4241 #endif /* CONFIG_SMP */
4243 #ifdef CONFIG_FREEZER
4246 * freeze_workqueues_begin - begin freezing workqueues
4248 * Start freezing workqueues. After this function returns, all freezable
4249 * workqueues will queue new works to their delayed_works list instead of
4253 * Grabs and releases workqueue_lock and pool->lock's.
4255 void freeze_workqueues_begin(void)
4257 struct worker_pool *pool;
4258 struct workqueue_struct *wq;
4259 struct pool_workqueue *pwq;
4262 spin_lock_irq(&workqueue_lock);
4264 WARN_ON_ONCE(workqueue_freezing);
4265 workqueue_freezing = true;
4268 for_each_pool(pool, pi) {
4269 spin_lock(&pool->lock);
4270 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4271 pool->flags |= POOL_FREEZING;
4272 spin_unlock(&pool->lock);
4275 /* suppress further executions by setting max_active to zero */
4276 list_for_each_entry(wq, &workqueues, list) {
4277 for_each_pwq(pwq, wq)
4278 pwq_adjust_max_active(pwq);
4281 spin_unlock_irq(&workqueue_lock);
4285 * freeze_workqueues_busy - are freezable workqueues still busy?
4287 * Check whether freezing is complete. This function must be called
4288 * between freeze_workqueues_begin() and thaw_workqueues().
4291 * Grabs and releases workqueue_lock.
4294 * %true if some freezable workqueues are still busy. %false if freezing
4297 bool freeze_workqueues_busy(void)
4300 struct workqueue_struct *wq;
4301 struct pool_workqueue *pwq;
4303 spin_lock_irq(&workqueue_lock);
4305 WARN_ON_ONCE(!workqueue_freezing);
4307 list_for_each_entry(wq, &workqueues, list) {
4308 if (!(wq->flags & WQ_FREEZABLE))
4311 * nr_active is monotonically decreasing. It's safe
4312 * to peek without lock.
4314 for_each_pwq(pwq, wq) {
4315 WARN_ON_ONCE(pwq->nr_active < 0);
4316 if (pwq->nr_active) {
4323 spin_unlock_irq(&workqueue_lock);
4328 * thaw_workqueues - thaw workqueues
4330 * Thaw workqueues. Normal queueing is restored and all collected
4331 * frozen works are transferred to their respective pool worklists.
4334 * Grabs and releases workqueue_lock and pool->lock's.
4336 void thaw_workqueues(void)
4338 struct workqueue_struct *wq;
4339 struct pool_workqueue *pwq;
4340 struct worker_pool *pool;
4343 spin_lock_irq(&workqueue_lock);
4345 if (!workqueue_freezing)
4348 /* clear FREEZING */
4349 for_each_pool(pool, pi) {
4350 spin_lock(&pool->lock);
4351 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4352 pool->flags &= ~POOL_FREEZING;
4353 spin_unlock(&pool->lock);
4356 /* restore max_active and repopulate worklist */
4357 list_for_each_entry(wq, &workqueues, list) {
4358 for_each_pwq(pwq, wq)
4359 pwq_adjust_max_active(pwq);
4363 for_each_pool(pool, pi) {
4364 spin_lock(&pool->lock);
4365 wake_up_worker(pool);
4366 spin_unlock(&pool->lock);
4369 workqueue_freezing = false;
4371 spin_unlock_irq(&workqueue_lock);
4373 #endif /* CONFIG_FREEZER */
4375 static int __init init_workqueues(void)
4377 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4380 /* make sure we have enough bits for OFFQ pool ID */
4381 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4382 WORK_CPU_END * NR_STD_WORKER_POOLS);
4384 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4386 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4388 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4389 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4391 /* initialize CPU pools */
4392 for_each_possible_cpu(cpu) {
4393 struct worker_pool *pool;
4396 for_each_cpu_worker_pool(pool, cpu) {
4397 BUG_ON(init_worker_pool(pool));
4399 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4400 pool->attrs->nice = std_nice[i++];
4403 BUG_ON(worker_pool_assign_id(pool));
4407 /* create the initial worker */
4408 for_each_online_cpu(cpu) {
4409 struct worker_pool *pool;
4411 for_each_cpu_worker_pool(pool, cpu) {
4412 struct worker *worker;
4414 pool->flags &= ~POOL_DISASSOCIATED;
4416 worker = create_worker(pool);
4418 spin_lock_irq(&pool->lock);
4419 start_worker(worker);
4420 spin_unlock_irq(&pool->lock);
4424 /* create default unbound wq attrs */
4425 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4426 struct workqueue_attrs *attrs;
4428 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4430 attrs->nice = std_nice[i];
4431 cpumask_setall(attrs->cpumask);
4433 unbound_std_wq_attrs[i] = attrs;
4436 system_wq = alloc_workqueue("events", 0, 0);
4437 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4438 system_long_wq = alloc_workqueue("events_long", 0, 0);
4439 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4440 WQ_UNBOUND_MAX_ACTIVE);
4441 system_freezable_wq = alloc_workqueue("events_freezable",
4443 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4444 !system_unbound_wq || !system_freezable_wq);
4447 early_initcall(init_workqueues);