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 * WQ: wq_mutex protected.
124 * WR: wq_mutex protected for writes. Sched-RCU protected for reads.
126 * PW: pwq_lock protected.
128 * W: workqueue_lock protected.
130 * FR: wq->flush_mutex and pwq_lock protected for writes. Sched-RCU
131 * protected for reads.
134 /* struct worker is defined in workqueue_internal.h */
137 spinlock_t lock; /* the pool lock */
138 int cpu; /* I: the associated cpu */
139 int id; /* I: pool ID */
140 unsigned int flags; /* X: flags */
142 struct list_head worklist; /* L: list of pending works */
143 int nr_workers; /* L: total number of workers */
145 /* nr_idle includes the ones off idle_list for rebinding */
146 int nr_idle; /* L: currently idle ones */
148 struct list_head idle_list; /* X: list of idle workers */
149 struct timer_list idle_timer; /* L: worker idle timeout */
150 struct timer_list mayday_timer; /* L: SOS timer for workers */
152 /* a workers is either on busy_hash or idle_list, or the manager */
153 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
154 /* L: hash of busy workers */
156 /* see manage_workers() for details on the two manager mutexes */
157 struct mutex manager_arb; /* manager arbitration */
158 struct mutex manager_mutex; /* manager exclusion */
159 struct ida worker_ida; /* L: for worker IDs */
161 struct workqueue_attrs *attrs; /* I: worker attributes */
162 struct hlist_node hash_node; /* WQ: unbound_pool_hash node */
163 int refcnt; /* WQ: refcnt for unbound pools */
166 * The current concurrency level. As it's likely to be accessed
167 * from other CPUs during try_to_wake_up(), put it in a separate
170 atomic_t nr_running ____cacheline_aligned_in_smp;
173 * Destruction of pool is sched-RCU protected to allow dereferences
174 * from get_work_pool().
177 } ____cacheline_aligned_in_smp;
180 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
181 * of work_struct->data are used for flags and the remaining high bits
182 * point to the pwq; thus, pwqs need to be aligned at two's power of the
183 * number of flag bits.
185 struct pool_workqueue {
186 struct worker_pool *pool; /* I: the associated pool */
187 struct workqueue_struct *wq; /* I: the owning workqueue */
188 int work_color; /* L: current color */
189 int flush_color; /* L: flushing color */
190 int refcnt; /* L: reference count */
191 int nr_in_flight[WORK_NR_COLORS];
192 /* L: nr of in_flight works */
193 int nr_active; /* L: nr of active works */
194 int max_active; /* L: max active works */
195 struct list_head delayed_works; /* L: delayed works */
196 struct list_head pwqs_node; /* FR: node on wq->pwqs */
197 struct list_head mayday_node; /* W: node on wq->maydays */
200 * Release of unbound pwq is punted to system_wq. See put_pwq()
201 * and pwq_unbound_release_workfn() for details. pool_workqueue
202 * itself is also sched-RCU protected so that the first pwq can be
203 * determined without grabbing pwq_lock.
205 struct work_struct unbound_release_work;
207 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
210 * Structure used to wait for workqueue flush.
213 struct list_head list; /* F: list of flushers */
214 int flush_color; /* F: flush color waiting for */
215 struct completion done; /* flush completion */
221 * The externally visible workqueue. It relays the issued work items to
222 * the appropriate worker_pool through its pool_workqueues.
224 struct workqueue_struct {
225 unsigned int flags; /* WQ: WQ_* flags */
226 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */
227 struct list_head pwqs; /* FR: all pwqs of this wq */
228 struct list_head list; /* WQ: list of all workqueues */
230 struct mutex flush_mutex; /* protects wq flushing */
231 int work_color; /* F: current work color */
232 int flush_color; /* F: current flush color */
233 atomic_t nr_pwqs_to_flush; /* flush in progress */
234 struct wq_flusher *first_flusher; /* F: first flusher */
235 struct list_head flusher_queue; /* F: flush waiters */
236 struct list_head flusher_overflow; /* F: flush overflow list */
238 struct list_head maydays; /* W: pwqs requesting rescue */
239 struct worker *rescuer; /* I: rescue worker */
241 int nr_drainers; /* WQ: drain in progress */
242 int saved_max_active; /* PW: saved pwq max_active */
245 struct wq_device *wq_dev; /* I: for sysfs interface */
247 #ifdef CONFIG_LOCKDEP
248 struct lockdep_map lockdep_map;
250 char name[]; /* I: workqueue name */
253 static struct kmem_cache *pwq_cache;
255 static DEFINE_MUTEX(wq_mutex); /* protects workqueues and pools */
256 static DEFINE_SPINLOCK(pwq_lock); /* protects pool_workqueues */
257 static DEFINE_SPINLOCK(workqueue_lock);
259 static LIST_HEAD(workqueues); /* WQ: list of all workqueues */
260 static bool workqueue_freezing; /* WQ: have wqs started freezing? */
262 /* the per-cpu worker pools */
263 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
266 static DEFINE_IDR(worker_pool_idr); /* WR: idr of all pools */
268 /* WQ: hash of all unbound pools keyed by pool->attrs */
269 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
271 /* I: attributes used when instantiating standard unbound pools on demand */
272 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
274 struct workqueue_struct *system_wq __read_mostly;
275 EXPORT_SYMBOL_GPL(system_wq);
276 struct workqueue_struct *system_highpri_wq __read_mostly;
277 EXPORT_SYMBOL_GPL(system_highpri_wq);
278 struct workqueue_struct *system_long_wq __read_mostly;
279 EXPORT_SYMBOL_GPL(system_long_wq);
280 struct workqueue_struct *system_unbound_wq __read_mostly;
281 EXPORT_SYMBOL_GPL(system_unbound_wq);
282 struct workqueue_struct *system_freezable_wq __read_mostly;
283 EXPORT_SYMBOL_GPL(system_freezable_wq);
285 static int worker_thread(void *__worker);
286 static void copy_workqueue_attrs(struct workqueue_attrs *to,
287 const struct workqueue_attrs *from);
289 #define CREATE_TRACE_POINTS
290 #include <trace/events/workqueue.h>
292 #define assert_rcu_or_wq_mutex() \
293 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
294 lockdep_is_held(&wq_mutex), \
295 "sched RCU or wq_mutex should be held")
297 #define assert_rcu_or_pwq_lock() \
298 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
299 lockdep_is_held(&pwq_lock), \
300 "sched RCU or pwq_lock should be held")
302 #define for_each_cpu_worker_pool(pool, cpu) \
303 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
304 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
307 #define for_each_busy_worker(worker, i, pool) \
308 hash_for_each(pool->busy_hash, i, worker, hentry)
311 * for_each_pool - iterate through all worker_pools in the system
312 * @pool: iteration cursor
313 * @pi: integer used for iteration
315 * This must be called either with wq_mutex held or sched RCU read locked.
316 * If the pool needs to be used beyond the locking in effect, the caller is
317 * responsible for guaranteeing that the pool stays online.
319 * The if/else clause exists only for the lockdep assertion and can be
322 #define for_each_pool(pool, pi) \
323 idr_for_each_entry(&worker_pool_idr, pool, pi) \
324 if (({ assert_rcu_or_wq_mutex(); false; })) { } \
328 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
329 * @pwq: iteration cursor
330 * @wq: the target workqueue
332 * This must be called either with pwq_lock held or sched RCU read locked.
333 * If the pwq needs to be used beyond the locking in effect, the caller is
334 * responsible for guaranteeing that the pwq stays online.
336 * The if/else clause exists only for the lockdep assertion and can be
339 #define for_each_pwq(pwq, wq) \
340 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
341 if (({ assert_rcu_or_pwq_lock(); false; })) { } \
344 #ifdef CONFIG_DEBUG_OBJECTS_WORK
346 static struct debug_obj_descr work_debug_descr;
348 static void *work_debug_hint(void *addr)
350 return ((struct work_struct *) addr)->func;
354 * fixup_init is called when:
355 * - an active object is initialized
357 static int work_fixup_init(void *addr, enum debug_obj_state state)
359 struct work_struct *work = addr;
362 case ODEBUG_STATE_ACTIVE:
363 cancel_work_sync(work);
364 debug_object_init(work, &work_debug_descr);
372 * fixup_activate is called when:
373 * - an active object is activated
374 * - an unknown object is activated (might be a statically initialized object)
376 static int work_fixup_activate(void *addr, enum debug_obj_state state)
378 struct work_struct *work = addr;
382 case ODEBUG_STATE_NOTAVAILABLE:
384 * This is not really a fixup. The work struct was
385 * statically initialized. We just make sure that it
386 * is tracked in the object tracker.
388 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
389 debug_object_init(work, &work_debug_descr);
390 debug_object_activate(work, &work_debug_descr);
396 case ODEBUG_STATE_ACTIVE:
405 * fixup_free is called when:
406 * - an active object is freed
408 static int work_fixup_free(void *addr, enum debug_obj_state state)
410 struct work_struct *work = addr;
413 case ODEBUG_STATE_ACTIVE:
414 cancel_work_sync(work);
415 debug_object_free(work, &work_debug_descr);
422 static struct debug_obj_descr work_debug_descr = {
423 .name = "work_struct",
424 .debug_hint = work_debug_hint,
425 .fixup_init = work_fixup_init,
426 .fixup_activate = work_fixup_activate,
427 .fixup_free = work_fixup_free,
430 static inline void debug_work_activate(struct work_struct *work)
432 debug_object_activate(work, &work_debug_descr);
435 static inline void debug_work_deactivate(struct work_struct *work)
437 debug_object_deactivate(work, &work_debug_descr);
440 void __init_work(struct work_struct *work, int onstack)
443 debug_object_init_on_stack(work, &work_debug_descr);
445 debug_object_init(work, &work_debug_descr);
447 EXPORT_SYMBOL_GPL(__init_work);
449 void destroy_work_on_stack(struct work_struct *work)
451 debug_object_free(work, &work_debug_descr);
453 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
456 static inline void debug_work_activate(struct work_struct *work) { }
457 static inline void debug_work_deactivate(struct work_struct *work) { }
460 /* allocate ID and assign it to @pool */
461 static int worker_pool_assign_id(struct worker_pool *pool)
465 lockdep_assert_held(&wq_mutex);
468 if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL))
470 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
471 } while (ret == -EAGAIN);
477 * first_pwq - return the first pool_workqueue of the specified workqueue
478 * @wq: the target workqueue
480 * This must be called either with pwq_lock held or sched RCU read locked.
481 * If the pwq needs to be used beyond the locking in effect, the caller is
482 * responsible for guaranteeing that the pwq stays online.
484 static struct pool_workqueue *first_pwq(struct workqueue_struct *wq)
486 assert_rcu_or_pwq_lock();
487 return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue,
491 static unsigned int work_color_to_flags(int color)
493 return color << WORK_STRUCT_COLOR_SHIFT;
496 static int get_work_color(struct work_struct *work)
498 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
499 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
502 static int work_next_color(int color)
504 return (color + 1) % WORK_NR_COLORS;
508 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
509 * contain the pointer to the queued pwq. Once execution starts, the flag
510 * is cleared and the high bits contain OFFQ flags and pool ID.
512 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
513 * and clear_work_data() can be used to set the pwq, pool or clear
514 * work->data. These functions should only be called while the work is
515 * owned - ie. while the PENDING bit is set.
517 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
518 * corresponding to a work. Pool is available once the work has been
519 * queued anywhere after initialization until it is sync canceled. pwq is
520 * available only while the work item is queued.
522 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
523 * canceled. While being canceled, a work item may have its PENDING set
524 * but stay off timer and worklist for arbitrarily long and nobody should
525 * try to steal the PENDING bit.
527 static inline void set_work_data(struct work_struct *work, unsigned long data,
530 WARN_ON_ONCE(!work_pending(work));
531 atomic_long_set(&work->data, data | flags | work_static(work));
534 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
535 unsigned long extra_flags)
537 set_work_data(work, (unsigned long)pwq,
538 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
541 static void set_work_pool_and_keep_pending(struct work_struct *work,
544 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
545 WORK_STRUCT_PENDING);
548 static void set_work_pool_and_clear_pending(struct work_struct *work,
552 * The following wmb is paired with the implied mb in
553 * test_and_set_bit(PENDING) and ensures all updates to @work made
554 * here are visible to and precede any updates by the next PENDING
558 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
561 static void clear_work_data(struct work_struct *work)
563 smp_wmb(); /* see set_work_pool_and_clear_pending() */
564 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
567 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
569 unsigned long data = atomic_long_read(&work->data);
571 if (data & WORK_STRUCT_PWQ)
572 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
578 * get_work_pool - return the worker_pool a given work was associated with
579 * @work: the work item of interest
581 * Return the worker_pool @work was last associated with. %NULL if none.
583 * Pools are created and destroyed under wq_mutex, and allows read access
584 * under sched-RCU read lock. As such, this function should be called
585 * under wq_mutex or with preemption disabled.
587 * All fields of the returned pool are accessible as long as the above
588 * mentioned locking is in effect. If the returned pool needs to be used
589 * beyond the critical section, the caller is responsible for ensuring the
590 * returned pool is and stays online.
592 static struct worker_pool *get_work_pool(struct work_struct *work)
594 unsigned long data = atomic_long_read(&work->data);
597 assert_rcu_or_wq_mutex();
599 if (data & WORK_STRUCT_PWQ)
600 return ((struct pool_workqueue *)
601 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
603 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
604 if (pool_id == WORK_OFFQ_POOL_NONE)
607 return idr_find(&worker_pool_idr, pool_id);
611 * get_work_pool_id - return the worker pool ID a given work is associated with
612 * @work: the work item of interest
614 * Return the worker_pool ID @work was last associated with.
615 * %WORK_OFFQ_POOL_NONE if none.
617 static int get_work_pool_id(struct work_struct *work)
619 unsigned long data = atomic_long_read(&work->data);
621 if (data & WORK_STRUCT_PWQ)
622 return ((struct pool_workqueue *)
623 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
625 return data >> WORK_OFFQ_POOL_SHIFT;
628 static void mark_work_canceling(struct work_struct *work)
630 unsigned long pool_id = get_work_pool_id(work);
632 pool_id <<= WORK_OFFQ_POOL_SHIFT;
633 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
636 static bool work_is_canceling(struct work_struct *work)
638 unsigned long data = atomic_long_read(&work->data);
640 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
644 * Policy functions. These define the policies on how the global worker
645 * pools are managed. Unless noted otherwise, these functions assume that
646 * they're being called with pool->lock held.
649 static bool __need_more_worker(struct worker_pool *pool)
651 return !atomic_read(&pool->nr_running);
655 * Need to wake up a worker? Called from anything but currently
658 * Note that, because unbound workers never contribute to nr_running, this
659 * function will always return %true for unbound pools as long as the
660 * worklist isn't empty.
662 static bool need_more_worker(struct worker_pool *pool)
664 return !list_empty(&pool->worklist) && __need_more_worker(pool);
667 /* Can I start working? Called from busy but !running workers. */
668 static bool may_start_working(struct worker_pool *pool)
670 return pool->nr_idle;
673 /* Do I need to keep working? Called from currently running workers. */
674 static bool keep_working(struct worker_pool *pool)
676 return !list_empty(&pool->worklist) &&
677 atomic_read(&pool->nr_running) <= 1;
680 /* Do we need a new worker? Called from manager. */
681 static bool need_to_create_worker(struct worker_pool *pool)
683 return need_more_worker(pool) && !may_start_working(pool);
686 /* Do I need to be the manager? */
687 static bool need_to_manage_workers(struct worker_pool *pool)
689 return need_to_create_worker(pool) ||
690 (pool->flags & POOL_MANAGE_WORKERS);
693 /* Do we have too many workers and should some go away? */
694 static bool too_many_workers(struct worker_pool *pool)
696 bool managing = mutex_is_locked(&pool->manager_arb);
697 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
698 int nr_busy = pool->nr_workers - nr_idle;
701 * nr_idle and idle_list may disagree if idle rebinding is in
702 * progress. Never return %true if idle_list is empty.
704 if (list_empty(&pool->idle_list))
707 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
714 /* Return the first worker. Safe with preemption disabled */
715 static struct worker *first_worker(struct worker_pool *pool)
717 if (unlikely(list_empty(&pool->idle_list)))
720 return list_first_entry(&pool->idle_list, struct worker, entry);
724 * wake_up_worker - wake up an idle worker
725 * @pool: worker pool to wake worker from
727 * Wake up the first idle worker of @pool.
730 * spin_lock_irq(pool->lock).
732 static void wake_up_worker(struct worker_pool *pool)
734 struct worker *worker = first_worker(pool);
737 wake_up_process(worker->task);
741 * wq_worker_waking_up - a worker is waking up
742 * @task: task waking up
743 * @cpu: CPU @task is waking up to
745 * This function is called during try_to_wake_up() when a worker is
749 * spin_lock_irq(rq->lock)
751 void wq_worker_waking_up(struct task_struct *task, int cpu)
753 struct worker *worker = kthread_data(task);
755 if (!(worker->flags & WORKER_NOT_RUNNING)) {
756 WARN_ON_ONCE(worker->pool->cpu != cpu);
757 atomic_inc(&worker->pool->nr_running);
762 * wq_worker_sleeping - a worker is going to sleep
763 * @task: task going to sleep
764 * @cpu: CPU in question, must be the current CPU number
766 * This function is called during schedule() when a busy worker is
767 * going to sleep. Worker on the same cpu can be woken up by
768 * returning pointer to its task.
771 * spin_lock_irq(rq->lock)
774 * Worker task on @cpu to wake up, %NULL if none.
776 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
778 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
779 struct worker_pool *pool;
782 * Rescuers, which may not have all the fields set up like normal
783 * workers, also reach here, let's not access anything before
784 * checking NOT_RUNNING.
786 if (worker->flags & WORKER_NOT_RUNNING)
791 /* this can only happen on the local cpu */
792 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
796 * The counterpart of the following dec_and_test, implied mb,
797 * worklist not empty test sequence is in insert_work().
798 * Please read comment there.
800 * NOT_RUNNING is clear. This means that we're bound to and
801 * running on the local cpu w/ rq lock held and preemption
802 * disabled, which in turn means that none else could be
803 * manipulating idle_list, so dereferencing idle_list without pool
806 if (atomic_dec_and_test(&pool->nr_running) &&
807 !list_empty(&pool->worklist))
808 to_wakeup = first_worker(pool);
809 return to_wakeup ? to_wakeup->task : NULL;
813 * worker_set_flags - set worker flags and adjust nr_running accordingly
815 * @flags: flags to set
816 * @wakeup: wakeup an idle worker if necessary
818 * Set @flags in @worker->flags and adjust nr_running accordingly. If
819 * nr_running becomes zero and @wakeup is %true, an idle worker is
823 * spin_lock_irq(pool->lock)
825 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
828 struct worker_pool *pool = worker->pool;
830 WARN_ON_ONCE(worker->task != current);
833 * If transitioning into NOT_RUNNING, adjust nr_running and
834 * wake up an idle worker as necessary if requested by
837 if ((flags & WORKER_NOT_RUNNING) &&
838 !(worker->flags & WORKER_NOT_RUNNING)) {
840 if (atomic_dec_and_test(&pool->nr_running) &&
841 !list_empty(&pool->worklist))
842 wake_up_worker(pool);
844 atomic_dec(&pool->nr_running);
847 worker->flags |= flags;
851 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
853 * @flags: flags to clear
855 * Clear @flags in @worker->flags and adjust nr_running accordingly.
858 * spin_lock_irq(pool->lock)
860 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
862 struct worker_pool *pool = worker->pool;
863 unsigned int oflags = worker->flags;
865 WARN_ON_ONCE(worker->task != current);
867 worker->flags &= ~flags;
870 * If transitioning out of NOT_RUNNING, increment nr_running. Note
871 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
872 * of multiple flags, not a single flag.
874 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
875 if (!(worker->flags & WORKER_NOT_RUNNING))
876 atomic_inc(&pool->nr_running);
880 * find_worker_executing_work - find worker which is executing a work
881 * @pool: pool of interest
882 * @work: work to find worker for
884 * Find a worker which is executing @work on @pool by searching
885 * @pool->busy_hash which is keyed by the address of @work. For a worker
886 * to match, its current execution should match the address of @work and
887 * its work function. This is to avoid unwanted dependency between
888 * unrelated work executions through a work item being recycled while still
891 * This is a bit tricky. A work item may be freed once its execution
892 * starts and nothing prevents the freed area from being recycled for
893 * another work item. If the same work item address ends up being reused
894 * before the original execution finishes, workqueue will identify the
895 * recycled work item as currently executing and make it wait until the
896 * current execution finishes, introducing an unwanted dependency.
898 * This function checks the work item address and work function to avoid
899 * false positives. Note that this isn't complete as one may construct a
900 * work function which can introduce dependency onto itself through a
901 * recycled work item. Well, if somebody wants to shoot oneself in the
902 * foot that badly, there's only so much we can do, and if such deadlock
903 * actually occurs, it should be easy to locate the culprit work function.
906 * spin_lock_irq(pool->lock).
909 * Pointer to worker which is executing @work if found, NULL
912 static struct worker *find_worker_executing_work(struct worker_pool *pool,
913 struct work_struct *work)
915 struct worker *worker;
917 hash_for_each_possible(pool->busy_hash, worker, hentry,
919 if (worker->current_work == work &&
920 worker->current_func == work->func)
927 * move_linked_works - move linked works to a list
928 * @work: start of series of works to be scheduled
929 * @head: target list to append @work to
930 * @nextp: out paramter for nested worklist walking
932 * Schedule linked works starting from @work to @head. Work series to
933 * be scheduled starts at @work and includes any consecutive work with
934 * WORK_STRUCT_LINKED set in its predecessor.
936 * If @nextp is not NULL, it's updated to point to the next work of
937 * the last scheduled work. This allows move_linked_works() to be
938 * nested inside outer list_for_each_entry_safe().
941 * spin_lock_irq(pool->lock).
943 static void move_linked_works(struct work_struct *work, struct list_head *head,
944 struct work_struct **nextp)
946 struct work_struct *n;
949 * Linked worklist will always end before the end of the list,
950 * use NULL for list head.
952 list_for_each_entry_safe_from(work, n, NULL, entry) {
953 list_move_tail(&work->entry, head);
954 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
959 * If we're already inside safe list traversal and have moved
960 * multiple works to the scheduled queue, the next position
961 * needs to be updated.
968 * get_pwq - get an extra reference on the specified pool_workqueue
969 * @pwq: pool_workqueue to get
971 * Obtain an extra reference on @pwq. The caller should guarantee that
972 * @pwq has positive refcnt and be holding the matching pool->lock.
974 static void get_pwq(struct pool_workqueue *pwq)
976 lockdep_assert_held(&pwq->pool->lock);
977 WARN_ON_ONCE(pwq->refcnt <= 0);
982 * put_pwq - put a pool_workqueue reference
983 * @pwq: pool_workqueue to put
985 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
986 * destruction. The caller should be holding the matching pool->lock.
988 static void put_pwq(struct pool_workqueue *pwq)
990 lockdep_assert_held(&pwq->pool->lock);
991 if (likely(--pwq->refcnt))
993 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
996 * @pwq can't be released under pool->lock, bounce to
997 * pwq_unbound_release_workfn(). This never recurses on the same
998 * pool->lock as this path is taken only for unbound workqueues and
999 * the release work item is scheduled on a per-cpu workqueue. To
1000 * avoid lockdep warning, unbound pool->locks are given lockdep
1001 * subclass of 1 in get_unbound_pool().
1003 schedule_work(&pwq->unbound_release_work);
1006 static void pwq_activate_delayed_work(struct work_struct *work)
1008 struct pool_workqueue *pwq = get_work_pwq(work);
1010 trace_workqueue_activate_work(work);
1011 move_linked_works(work, &pwq->pool->worklist, NULL);
1012 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1016 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1018 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1019 struct work_struct, entry);
1021 pwq_activate_delayed_work(work);
1025 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1026 * @pwq: pwq of interest
1027 * @color: color of work which left the queue
1029 * A work either has completed or is removed from pending queue,
1030 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1033 * spin_lock_irq(pool->lock).
1035 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1037 /* uncolored work items don't participate in flushing or nr_active */
1038 if (color == WORK_NO_COLOR)
1041 pwq->nr_in_flight[color]--;
1044 if (!list_empty(&pwq->delayed_works)) {
1045 /* one down, submit a delayed one */
1046 if (pwq->nr_active < pwq->max_active)
1047 pwq_activate_first_delayed(pwq);
1050 /* is flush in progress and are we at the flushing tip? */
1051 if (likely(pwq->flush_color != color))
1054 /* are there still in-flight works? */
1055 if (pwq->nr_in_flight[color])
1058 /* this pwq is done, clear flush_color */
1059 pwq->flush_color = -1;
1062 * If this was the last pwq, wake up the first flusher. It
1063 * will handle the rest.
1065 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1066 complete(&pwq->wq->first_flusher->done);
1072 * try_to_grab_pending - steal work item from worklist and disable irq
1073 * @work: work item to steal
1074 * @is_dwork: @work is a delayed_work
1075 * @flags: place to store irq state
1077 * Try to grab PENDING bit of @work. This function can handle @work in any
1078 * stable state - idle, on timer or on worklist. Return values are
1080 * 1 if @work was pending and we successfully stole PENDING
1081 * 0 if @work was idle and we claimed PENDING
1082 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1083 * -ENOENT if someone else is canceling @work, this state may persist
1084 * for arbitrarily long
1086 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1087 * interrupted while holding PENDING and @work off queue, irq must be
1088 * disabled on entry. This, combined with delayed_work->timer being
1089 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1091 * On successful return, >= 0, irq is disabled and the caller is
1092 * responsible for releasing it using local_irq_restore(*@flags).
1094 * This function is safe to call from any context including IRQ handler.
1096 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1097 unsigned long *flags)
1099 struct worker_pool *pool;
1100 struct pool_workqueue *pwq;
1102 local_irq_save(*flags);
1104 /* try to steal the timer if it exists */
1106 struct delayed_work *dwork = to_delayed_work(work);
1109 * dwork->timer is irqsafe. If del_timer() fails, it's
1110 * guaranteed that the timer is not queued anywhere and not
1111 * running on the local CPU.
1113 if (likely(del_timer(&dwork->timer)))
1117 /* try to claim PENDING the normal way */
1118 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1122 * The queueing is in progress, or it is already queued. Try to
1123 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1125 pool = get_work_pool(work);
1129 spin_lock(&pool->lock);
1131 * work->data is guaranteed to point to pwq only while the work
1132 * item is queued on pwq->wq, and both updating work->data to point
1133 * to pwq on queueing and to pool on dequeueing are done under
1134 * pwq->pool->lock. This in turn guarantees that, if work->data
1135 * points to pwq which is associated with a locked pool, the work
1136 * item is currently queued on that pool.
1138 pwq = get_work_pwq(work);
1139 if (pwq && pwq->pool == pool) {
1140 debug_work_deactivate(work);
1143 * A delayed work item cannot be grabbed directly because
1144 * it might have linked NO_COLOR work items which, if left
1145 * on the delayed_list, will confuse pwq->nr_active
1146 * management later on and cause stall. Make sure the work
1147 * item is activated before grabbing.
1149 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1150 pwq_activate_delayed_work(work);
1152 list_del_init(&work->entry);
1153 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1155 /* work->data points to pwq iff queued, point to pool */
1156 set_work_pool_and_keep_pending(work, pool->id);
1158 spin_unlock(&pool->lock);
1161 spin_unlock(&pool->lock);
1163 local_irq_restore(*flags);
1164 if (work_is_canceling(work))
1171 * insert_work - insert a work into a pool
1172 * @pwq: pwq @work belongs to
1173 * @work: work to insert
1174 * @head: insertion point
1175 * @extra_flags: extra WORK_STRUCT_* flags to set
1177 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1178 * work_struct flags.
1181 * spin_lock_irq(pool->lock).
1183 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1184 struct list_head *head, unsigned int extra_flags)
1186 struct worker_pool *pool = pwq->pool;
1188 /* we own @work, set data and link */
1189 set_work_pwq(work, pwq, extra_flags);
1190 list_add_tail(&work->entry, head);
1194 * Ensure either wq_worker_sleeping() sees the above
1195 * list_add_tail() or we see zero nr_running to avoid workers lying
1196 * around lazily while there are works to be processed.
1200 if (__need_more_worker(pool))
1201 wake_up_worker(pool);
1205 * Test whether @work is being queued from another work executing on the
1208 static bool is_chained_work(struct workqueue_struct *wq)
1210 struct worker *worker;
1212 worker = current_wq_worker();
1214 * Return %true iff I'm a worker execuing a work item on @wq. If
1215 * I'm @worker, it's safe to dereference it without locking.
1217 return worker && worker->current_pwq->wq == wq;
1220 static void __queue_work(int cpu, struct workqueue_struct *wq,
1221 struct work_struct *work)
1223 struct pool_workqueue *pwq;
1224 struct worker_pool *last_pool;
1225 struct list_head *worklist;
1226 unsigned int work_flags;
1227 unsigned int req_cpu = cpu;
1230 * While a work item is PENDING && off queue, a task trying to
1231 * steal the PENDING will busy-loop waiting for it to either get
1232 * queued or lose PENDING. Grabbing PENDING and queueing should
1233 * happen with IRQ disabled.
1235 WARN_ON_ONCE(!irqs_disabled());
1237 debug_work_activate(work);
1239 /* if dying, only works from the same workqueue are allowed */
1240 if (unlikely(wq->flags & __WQ_DRAINING) &&
1241 WARN_ON_ONCE(!is_chained_work(wq)))
1244 /* pwq which will be used unless @work is executing elsewhere */
1245 if (!(wq->flags & WQ_UNBOUND)) {
1246 if (cpu == WORK_CPU_UNBOUND)
1247 cpu = raw_smp_processor_id();
1248 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1250 pwq = first_pwq(wq);
1254 * If @work was previously on a different pool, it might still be
1255 * running there, in which case the work needs to be queued on that
1256 * pool to guarantee non-reentrancy.
1258 last_pool = get_work_pool(work);
1259 if (last_pool && last_pool != pwq->pool) {
1260 struct worker *worker;
1262 spin_lock(&last_pool->lock);
1264 worker = find_worker_executing_work(last_pool, work);
1266 if (worker && worker->current_pwq->wq == wq) {
1267 pwq = worker->current_pwq;
1269 /* meh... not running there, queue here */
1270 spin_unlock(&last_pool->lock);
1271 spin_lock(&pwq->pool->lock);
1274 spin_lock(&pwq->pool->lock);
1278 * pwq is determined and locked. For unbound pools, we could have
1279 * raced with pwq release and it could already be dead. If its
1280 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1281 * without another pwq replacing it as the first pwq or while a
1282 * work item is executing on it, so the retying is guaranteed to
1283 * make forward-progress.
1285 if (unlikely(!pwq->refcnt)) {
1286 if (wq->flags & WQ_UNBOUND) {
1287 spin_unlock(&pwq->pool->lock);
1292 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1296 /* pwq determined, queue */
1297 trace_workqueue_queue_work(req_cpu, pwq, work);
1299 if (WARN_ON(!list_empty(&work->entry))) {
1300 spin_unlock(&pwq->pool->lock);
1304 pwq->nr_in_flight[pwq->work_color]++;
1305 work_flags = work_color_to_flags(pwq->work_color);
1307 if (likely(pwq->nr_active < pwq->max_active)) {
1308 trace_workqueue_activate_work(work);
1310 worklist = &pwq->pool->worklist;
1312 work_flags |= WORK_STRUCT_DELAYED;
1313 worklist = &pwq->delayed_works;
1316 insert_work(pwq, work, worklist, work_flags);
1318 spin_unlock(&pwq->pool->lock);
1322 * queue_work_on - queue work on specific cpu
1323 * @cpu: CPU number to execute work on
1324 * @wq: workqueue to use
1325 * @work: work to queue
1327 * Returns %false if @work was already on a queue, %true otherwise.
1329 * We queue the work to a specific CPU, the caller must ensure it
1332 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1333 struct work_struct *work)
1336 unsigned long flags;
1338 local_irq_save(flags);
1340 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1341 __queue_work(cpu, wq, work);
1345 local_irq_restore(flags);
1348 EXPORT_SYMBOL_GPL(queue_work_on);
1350 void delayed_work_timer_fn(unsigned long __data)
1352 struct delayed_work *dwork = (struct delayed_work *)__data;
1354 /* should have been called from irqsafe timer with irq already off */
1355 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1357 EXPORT_SYMBOL(delayed_work_timer_fn);
1359 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1360 struct delayed_work *dwork, unsigned long delay)
1362 struct timer_list *timer = &dwork->timer;
1363 struct work_struct *work = &dwork->work;
1365 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1366 timer->data != (unsigned long)dwork);
1367 WARN_ON_ONCE(timer_pending(timer));
1368 WARN_ON_ONCE(!list_empty(&work->entry));
1371 * If @delay is 0, queue @dwork->work immediately. This is for
1372 * both optimization and correctness. The earliest @timer can
1373 * expire is on the closest next tick and delayed_work users depend
1374 * on that there's no such delay when @delay is 0.
1377 __queue_work(cpu, wq, &dwork->work);
1381 timer_stats_timer_set_start_info(&dwork->timer);
1385 timer->expires = jiffies + delay;
1387 if (unlikely(cpu != WORK_CPU_UNBOUND))
1388 add_timer_on(timer, cpu);
1394 * queue_delayed_work_on - queue work on specific CPU after delay
1395 * @cpu: CPU number to execute work on
1396 * @wq: workqueue to use
1397 * @dwork: work to queue
1398 * @delay: number of jiffies to wait before queueing
1400 * Returns %false if @work was already on a queue, %true otherwise. If
1401 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1404 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1405 struct delayed_work *dwork, unsigned long delay)
1407 struct work_struct *work = &dwork->work;
1409 unsigned long flags;
1411 /* read the comment in __queue_work() */
1412 local_irq_save(flags);
1414 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1415 __queue_delayed_work(cpu, wq, dwork, delay);
1419 local_irq_restore(flags);
1422 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1425 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1426 * @cpu: CPU number to execute work on
1427 * @wq: workqueue to use
1428 * @dwork: work to queue
1429 * @delay: number of jiffies to wait before queueing
1431 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1432 * modify @dwork's timer so that it expires after @delay. If @delay is
1433 * zero, @work is guaranteed to be scheduled immediately regardless of its
1436 * Returns %false if @dwork was idle and queued, %true if @dwork was
1437 * pending and its timer was modified.
1439 * This function is safe to call from any context including IRQ handler.
1440 * See try_to_grab_pending() for details.
1442 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1443 struct delayed_work *dwork, unsigned long delay)
1445 unsigned long flags;
1449 ret = try_to_grab_pending(&dwork->work, true, &flags);
1450 } while (unlikely(ret == -EAGAIN));
1452 if (likely(ret >= 0)) {
1453 __queue_delayed_work(cpu, wq, dwork, delay);
1454 local_irq_restore(flags);
1457 /* -ENOENT from try_to_grab_pending() becomes %true */
1460 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1463 * worker_enter_idle - enter idle state
1464 * @worker: worker which is entering idle state
1466 * @worker is entering idle state. Update stats and idle timer if
1470 * spin_lock_irq(pool->lock).
1472 static void worker_enter_idle(struct worker *worker)
1474 struct worker_pool *pool = worker->pool;
1476 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1477 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1478 (worker->hentry.next || worker->hentry.pprev)))
1481 /* can't use worker_set_flags(), also called from start_worker() */
1482 worker->flags |= WORKER_IDLE;
1484 worker->last_active = jiffies;
1486 /* idle_list is LIFO */
1487 list_add(&worker->entry, &pool->idle_list);
1489 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1490 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1493 * Sanity check nr_running. Because wq_unbind_fn() releases
1494 * pool->lock between setting %WORKER_UNBOUND and zapping
1495 * nr_running, the warning may trigger spuriously. Check iff
1496 * unbind is not in progress.
1498 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1499 pool->nr_workers == pool->nr_idle &&
1500 atomic_read(&pool->nr_running));
1504 * worker_leave_idle - leave idle state
1505 * @worker: worker which is leaving idle state
1507 * @worker is leaving idle state. Update stats.
1510 * spin_lock_irq(pool->lock).
1512 static void worker_leave_idle(struct worker *worker)
1514 struct worker_pool *pool = worker->pool;
1516 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1518 worker_clr_flags(worker, WORKER_IDLE);
1520 list_del_init(&worker->entry);
1524 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1525 * @pool: target worker_pool
1527 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1529 * Works which are scheduled while the cpu is online must at least be
1530 * scheduled to a worker which is bound to the cpu so that if they are
1531 * flushed from cpu callbacks while cpu is going down, they are
1532 * guaranteed to execute on the cpu.
1534 * This function is to be used by unbound workers and rescuers to bind
1535 * themselves to the target cpu and may race with cpu going down or
1536 * coming online. kthread_bind() can't be used because it may put the
1537 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1538 * verbatim as it's best effort and blocking and pool may be
1539 * [dis]associated in the meantime.
1541 * This function tries set_cpus_allowed() and locks pool and verifies the
1542 * binding against %POOL_DISASSOCIATED which is set during
1543 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1544 * enters idle state or fetches works without dropping lock, it can
1545 * guarantee the scheduling requirement described in the first paragraph.
1548 * Might sleep. Called without any lock but returns with pool->lock
1552 * %true if the associated pool is online (@worker is successfully
1553 * bound), %false if offline.
1555 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1556 __acquires(&pool->lock)
1560 * The following call may fail, succeed or succeed
1561 * without actually migrating the task to the cpu if
1562 * it races with cpu hotunplug operation. Verify
1563 * against POOL_DISASSOCIATED.
1565 if (!(pool->flags & POOL_DISASSOCIATED))
1566 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1568 spin_lock_irq(&pool->lock);
1569 if (pool->flags & POOL_DISASSOCIATED)
1571 if (task_cpu(current) == pool->cpu &&
1572 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1574 spin_unlock_irq(&pool->lock);
1577 * We've raced with CPU hot[un]plug. Give it a breather
1578 * and retry migration. cond_resched() is required here;
1579 * otherwise, we might deadlock against cpu_stop trying to
1580 * bring down the CPU on non-preemptive kernel.
1588 * Rebind an idle @worker to its CPU. worker_thread() will test
1589 * list_empty(@worker->entry) before leaving idle and call this function.
1591 static void idle_worker_rebind(struct worker *worker)
1593 /* CPU may go down again inbetween, clear UNBOUND only on success */
1594 if (worker_maybe_bind_and_lock(worker->pool))
1595 worker_clr_flags(worker, WORKER_UNBOUND);
1597 /* rebind complete, become available again */
1598 list_add(&worker->entry, &worker->pool->idle_list);
1599 spin_unlock_irq(&worker->pool->lock);
1603 * Function for @worker->rebind.work used to rebind unbound busy workers to
1604 * the associated cpu which is coming back online. This is scheduled by
1605 * cpu up but can race with other cpu hotplug operations and may be
1606 * executed twice without intervening cpu down.
1608 static void busy_worker_rebind_fn(struct work_struct *work)
1610 struct worker *worker = container_of(work, struct worker, rebind_work);
1612 if (worker_maybe_bind_and_lock(worker->pool))
1613 worker_clr_flags(worker, WORKER_UNBOUND);
1615 spin_unlock_irq(&worker->pool->lock);
1619 * rebind_workers - rebind all workers of a pool to the associated CPU
1620 * @pool: pool of interest
1622 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1623 * is different for idle and busy ones.
1625 * Idle ones will be removed from the idle_list and woken up. They will
1626 * add themselves back after completing rebind. This ensures that the
1627 * idle_list doesn't contain any unbound workers when re-bound busy workers
1628 * try to perform local wake-ups for concurrency management.
1630 * Busy workers can rebind after they finish their current work items.
1631 * Queueing the rebind work item at the head of the scheduled list is
1632 * enough. Note that nr_running will be properly bumped as busy workers
1635 * On return, all non-manager workers are scheduled for rebind - see
1636 * manage_workers() for the manager special case. Any idle worker
1637 * including the manager will not appear on @idle_list until rebind is
1638 * complete, making local wake-ups safe.
1640 static void rebind_workers(struct worker_pool *pool)
1642 struct worker *worker, *n;
1645 lockdep_assert_held(&pool->manager_mutex);
1646 lockdep_assert_held(&pool->lock);
1648 /* dequeue and kick idle ones */
1649 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1651 * idle workers should be off @pool->idle_list until rebind
1652 * is complete to avoid receiving premature local wake-ups.
1654 list_del_init(&worker->entry);
1657 * worker_thread() will see the above dequeuing and call
1658 * idle_worker_rebind().
1660 wake_up_process(worker->task);
1663 /* rebind busy workers */
1664 for_each_busy_worker(worker, i, pool) {
1665 struct work_struct *rebind_work = &worker->rebind_work;
1666 struct workqueue_struct *wq;
1668 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1669 work_data_bits(rebind_work)))
1672 debug_work_activate(rebind_work);
1675 * wq doesn't really matter but let's keep @worker->pool
1676 * and @pwq->pool consistent for sanity.
1678 if (worker->pool->attrs->nice < 0)
1679 wq = system_highpri_wq;
1683 insert_work(per_cpu_ptr(wq->cpu_pwqs, pool->cpu), rebind_work,
1684 worker->scheduled.next,
1685 work_color_to_flags(WORK_NO_COLOR));
1689 static struct worker *alloc_worker(void)
1691 struct worker *worker;
1693 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1695 INIT_LIST_HEAD(&worker->entry);
1696 INIT_LIST_HEAD(&worker->scheduled);
1697 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1698 /* on creation a worker is in !idle && prep state */
1699 worker->flags = WORKER_PREP;
1705 * create_worker - create a new workqueue worker
1706 * @pool: pool the new worker will belong to
1708 * Create a new worker which is bound to @pool. The returned worker
1709 * can be started by calling start_worker() or destroyed using
1713 * Might sleep. Does GFP_KERNEL allocations.
1716 * Pointer to the newly created worker.
1718 static struct worker *create_worker(struct worker_pool *pool)
1720 const char *pri = pool->attrs->nice < 0 ? "H" : "";
1721 struct worker *worker = NULL;
1724 lockdep_assert_held(&pool->manager_mutex);
1726 spin_lock_irq(&pool->lock);
1727 while (ida_get_new(&pool->worker_ida, &id)) {
1728 spin_unlock_irq(&pool->lock);
1729 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1731 spin_lock_irq(&pool->lock);
1733 spin_unlock_irq(&pool->lock);
1735 worker = alloc_worker();
1739 worker->pool = pool;
1743 worker->task = kthread_create_on_node(worker_thread,
1744 worker, cpu_to_node(pool->cpu),
1745 "kworker/%d:%d%s", pool->cpu, id, pri);
1747 worker->task = kthread_create(worker_thread, worker,
1750 if (IS_ERR(worker->task))
1754 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1755 * online CPUs. It'll be re-applied when any of the CPUs come up.
1757 set_user_nice(worker->task, pool->attrs->nice);
1758 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1761 * %PF_THREAD_BOUND is used to prevent userland from meddling with
1762 * cpumask of workqueue workers. This is an abuse. We need
1763 * %PF_NO_SETAFFINITY.
1765 worker->task->flags |= PF_THREAD_BOUND;
1768 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1769 * remains stable across this function. See the comments above the
1770 * flag definition for details.
1772 if (pool->flags & POOL_DISASSOCIATED)
1773 worker->flags |= WORKER_UNBOUND;
1778 spin_lock_irq(&pool->lock);
1779 ida_remove(&pool->worker_ida, id);
1780 spin_unlock_irq(&pool->lock);
1787 * start_worker - start a newly created worker
1788 * @worker: worker to start
1790 * Make the pool aware of @worker and start it.
1793 * spin_lock_irq(pool->lock).
1795 static void start_worker(struct worker *worker)
1797 worker->flags |= WORKER_STARTED;
1798 worker->pool->nr_workers++;
1799 worker_enter_idle(worker);
1800 wake_up_process(worker->task);
1804 * create_and_start_worker - create and start a worker for a pool
1805 * @pool: the target pool
1807 * Grab the managership of @pool and create and start a new worker for it.
1809 static int create_and_start_worker(struct worker_pool *pool)
1811 struct worker *worker;
1813 mutex_lock(&pool->manager_mutex);
1815 worker = create_worker(pool);
1817 spin_lock_irq(&pool->lock);
1818 start_worker(worker);
1819 spin_unlock_irq(&pool->lock);
1822 mutex_unlock(&pool->manager_mutex);
1824 return worker ? 0 : -ENOMEM;
1828 * destroy_worker - destroy a workqueue worker
1829 * @worker: worker to be destroyed
1831 * Destroy @worker and adjust @pool stats accordingly.
1834 * spin_lock_irq(pool->lock) which is released and regrabbed.
1836 static void destroy_worker(struct worker *worker)
1838 struct worker_pool *pool = worker->pool;
1839 int id = worker->id;
1841 lockdep_assert_held(&pool->manager_mutex);
1842 lockdep_assert_held(&pool->lock);
1844 /* sanity check frenzy */
1845 if (WARN_ON(worker->current_work) ||
1846 WARN_ON(!list_empty(&worker->scheduled)))
1849 if (worker->flags & WORKER_STARTED)
1851 if (worker->flags & WORKER_IDLE)
1854 list_del_init(&worker->entry);
1855 worker->flags |= WORKER_DIE;
1857 spin_unlock_irq(&pool->lock);
1859 kthread_stop(worker->task);
1862 spin_lock_irq(&pool->lock);
1863 ida_remove(&pool->worker_ida, id);
1866 static void idle_worker_timeout(unsigned long __pool)
1868 struct worker_pool *pool = (void *)__pool;
1870 spin_lock_irq(&pool->lock);
1872 if (too_many_workers(pool)) {
1873 struct worker *worker;
1874 unsigned long expires;
1876 /* idle_list is kept in LIFO order, check the last one */
1877 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1878 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1880 if (time_before(jiffies, expires))
1881 mod_timer(&pool->idle_timer, expires);
1883 /* it's been idle for too long, wake up manager */
1884 pool->flags |= POOL_MANAGE_WORKERS;
1885 wake_up_worker(pool);
1889 spin_unlock_irq(&pool->lock);
1892 static void send_mayday(struct work_struct *work)
1894 struct pool_workqueue *pwq = get_work_pwq(work);
1895 struct workqueue_struct *wq = pwq->wq;
1897 lockdep_assert_held(&workqueue_lock);
1902 /* mayday mayday mayday */
1903 if (list_empty(&pwq->mayday_node)) {
1904 list_add_tail(&pwq->mayday_node, &wq->maydays);
1905 wake_up_process(wq->rescuer->task);
1909 static void pool_mayday_timeout(unsigned long __pool)
1911 struct worker_pool *pool = (void *)__pool;
1912 struct work_struct *work;
1914 spin_lock_irq(&workqueue_lock); /* for wq->maydays */
1915 spin_lock(&pool->lock);
1917 if (need_to_create_worker(pool)) {
1919 * We've been trying to create a new worker but
1920 * haven't been successful. We might be hitting an
1921 * allocation deadlock. Send distress signals to
1924 list_for_each_entry(work, &pool->worklist, entry)
1928 spin_unlock(&pool->lock);
1929 spin_unlock_irq(&workqueue_lock);
1931 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1935 * maybe_create_worker - create a new worker if necessary
1936 * @pool: pool to create a new worker for
1938 * Create a new worker for @pool if necessary. @pool is guaranteed to
1939 * have at least one idle worker on return from this function. If
1940 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1941 * sent to all rescuers with works scheduled on @pool to resolve
1942 * possible allocation deadlock.
1944 * On return, need_to_create_worker() is guaranteed to be %false and
1945 * may_start_working() %true.
1948 * spin_lock_irq(pool->lock) which may be released and regrabbed
1949 * multiple times. Does GFP_KERNEL allocations. Called only from
1953 * %false if no action was taken and pool->lock stayed locked, %true
1956 static bool maybe_create_worker(struct worker_pool *pool)
1957 __releases(&pool->lock)
1958 __acquires(&pool->lock)
1960 if (!need_to_create_worker(pool))
1963 spin_unlock_irq(&pool->lock);
1965 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1966 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1969 struct worker *worker;
1971 worker = create_worker(pool);
1973 del_timer_sync(&pool->mayday_timer);
1974 spin_lock_irq(&pool->lock);
1975 start_worker(worker);
1976 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1981 if (!need_to_create_worker(pool))
1984 __set_current_state(TASK_INTERRUPTIBLE);
1985 schedule_timeout(CREATE_COOLDOWN);
1987 if (!need_to_create_worker(pool))
1991 del_timer_sync(&pool->mayday_timer);
1992 spin_lock_irq(&pool->lock);
1993 if (need_to_create_worker(pool))
1999 * maybe_destroy_worker - destroy workers which have been idle for a while
2000 * @pool: pool to destroy workers for
2002 * Destroy @pool workers which have been idle for longer than
2003 * IDLE_WORKER_TIMEOUT.
2006 * spin_lock_irq(pool->lock) which may be released and regrabbed
2007 * multiple times. Called only from manager.
2010 * %false if no action was taken and pool->lock stayed locked, %true
2013 static bool maybe_destroy_workers(struct worker_pool *pool)
2017 while (too_many_workers(pool)) {
2018 struct worker *worker;
2019 unsigned long expires;
2021 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2022 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2024 if (time_before(jiffies, expires)) {
2025 mod_timer(&pool->idle_timer, expires);
2029 destroy_worker(worker);
2037 * manage_workers - manage worker pool
2040 * Assume the manager role and manage the worker pool @worker belongs
2041 * to. At any given time, there can be only zero or one manager per
2042 * pool. The exclusion is handled automatically by this function.
2044 * The caller can safely start processing works on false return. On
2045 * true return, it's guaranteed that need_to_create_worker() is false
2046 * and may_start_working() is true.
2049 * spin_lock_irq(pool->lock) which may be released and regrabbed
2050 * multiple times. Does GFP_KERNEL allocations.
2053 * spin_lock_irq(pool->lock) which may be released and regrabbed
2054 * multiple times. Does GFP_KERNEL allocations.
2056 static bool manage_workers(struct worker *worker)
2058 struct worker_pool *pool = worker->pool;
2062 * Managership is governed by two mutexes - manager_arb and
2063 * manager_mutex. manager_arb handles arbitration of manager role.
2064 * Anyone who successfully grabs manager_arb wins the arbitration
2065 * and becomes the manager. mutex_trylock() on pool->manager_arb
2066 * failure while holding pool->lock reliably indicates that someone
2067 * else is managing the pool and the worker which failed trylock
2068 * can proceed to executing work items. This means that anyone
2069 * grabbing manager_arb is responsible for actually performing
2070 * manager duties. If manager_arb is grabbed and released without
2071 * actual management, the pool may stall indefinitely.
2073 * manager_mutex is used for exclusion of actual management
2074 * operations. The holder of manager_mutex can be sure that none
2075 * of management operations, including creation and destruction of
2076 * workers, won't take place until the mutex is released. Because
2077 * manager_mutex doesn't interfere with manager role arbitration,
2078 * it is guaranteed that the pool's management, while may be
2079 * delayed, won't be disturbed by someone else grabbing
2082 if (!mutex_trylock(&pool->manager_arb))
2086 * With manager arbitration won, manager_mutex would be free in
2087 * most cases. trylock first without dropping @pool->lock.
2089 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2090 spin_unlock_irq(&pool->lock);
2091 mutex_lock(&pool->manager_mutex);
2093 * CPU hotplug could have happened while we were waiting
2094 * for assoc_mutex. Hotplug itself can't handle us
2095 * because manager isn't either on idle or busy list, and
2096 * @pool's state and ours could have deviated.
2098 * As hotplug is now excluded via manager_mutex, we can
2099 * simply try to bind. It will succeed or fail depending
2100 * on @pool's current state. Try it and adjust
2101 * %WORKER_UNBOUND accordingly.
2103 if (worker_maybe_bind_and_lock(pool))
2104 worker->flags &= ~WORKER_UNBOUND;
2106 worker->flags |= WORKER_UNBOUND;
2111 pool->flags &= ~POOL_MANAGE_WORKERS;
2114 * Destroy and then create so that may_start_working() is true
2117 ret |= maybe_destroy_workers(pool);
2118 ret |= maybe_create_worker(pool);
2120 mutex_unlock(&pool->manager_mutex);
2121 mutex_unlock(&pool->manager_arb);
2126 * process_one_work - process single work
2128 * @work: work to process
2130 * Process @work. This function contains all the logics necessary to
2131 * process a single work including synchronization against and
2132 * interaction with other workers on the same cpu, queueing and
2133 * flushing. As long as context requirement is met, any worker can
2134 * call this function to process a work.
2137 * spin_lock_irq(pool->lock) which is released and regrabbed.
2139 static void process_one_work(struct worker *worker, struct work_struct *work)
2140 __releases(&pool->lock)
2141 __acquires(&pool->lock)
2143 struct pool_workqueue *pwq = get_work_pwq(work);
2144 struct worker_pool *pool = worker->pool;
2145 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2147 struct worker *collision;
2148 #ifdef CONFIG_LOCKDEP
2150 * It is permissible to free the struct work_struct from
2151 * inside the function that is called from it, this we need to
2152 * take into account for lockdep too. To avoid bogus "held
2153 * lock freed" warnings as well as problems when looking into
2154 * work->lockdep_map, make a copy and use that here.
2156 struct lockdep_map lockdep_map;
2158 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2161 * Ensure we're on the correct CPU. DISASSOCIATED test is
2162 * necessary to avoid spurious warnings from rescuers servicing the
2163 * unbound or a disassociated pool.
2165 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2166 !(pool->flags & POOL_DISASSOCIATED) &&
2167 raw_smp_processor_id() != pool->cpu);
2170 * A single work shouldn't be executed concurrently by
2171 * multiple workers on a single cpu. Check whether anyone is
2172 * already processing the work. If so, defer the work to the
2173 * currently executing one.
2175 collision = find_worker_executing_work(pool, work);
2176 if (unlikely(collision)) {
2177 move_linked_works(work, &collision->scheduled, NULL);
2181 /* claim and dequeue */
2182 debug_work_deactivate(work);
2183 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2184 worker->current_work = work;
2185 worker->current_func = work->func;
2186 worker->current_pwq = pwq;
2187 work_color = get_work_color(work);
2189 list_del_init(&work->entry);
2192 * CPU intensive works don't participate in concurrency
2193 * management. They're the scheduler's responsibility.
2195 if (unlikely(cpu_intensive))
2196 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2199 * Unbound pool isn't concurrency managed and work items should be
2200 * executed ASAP. Wake up another worker if necessary.
2202 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2203 wake_up_worker(pool);
2206 * Record the last pool and clear PENDING which should be the last
2207 * update to @work. Also, do this inside @pool->lock so that
2208 * PENDING and queued state changes happen together while IRQ is
2211 set_work_pool_and_clear_pending(work, pool->id);
2213 spin_unlock_irq(&pool->lock);
2215 lock_map_acquire_read(&pwq->wq->lockdep_map);
2216 lock_map_acquire(&lockdep_map);
2217 trace_workqueue_execute_start(work);
2218 worker->current_func(work);
2220 * While we must be careful to not use "work" after this, the trace
2221 * point will only record its address.
2223 trace_workqueue_execute_end(work);
2224 lock_map_release(&lockdep_map);
2225 lock_map_release(&pwq->wq->lockdep_map);
2227 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2228 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2229 " last function: %pf\n",
2230 current->comm, preempt_count(), task_pid_nr(current),
2231 worker->current_func);
2232 debug_show_held_locks(current);
2236 spin_lock_irq(&pool->lock);
2238 /* clear cpu intensive status */
2239 if (unlikely(cpu_intensive))
2240 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2242 /* we're done with it, release */
2243 hash_del(&worker->hentry);
2244 worker->current_work = NULL;
2245 worker->current_func = NULL;
2246 worker->current_pwq = NULL;
2247 pwq_dec_nr_in_flight(pwq, work_color);
2251 * process_scheduled_works - process scheduled works
2254 * Process all scheduled works. Please note that the scheduled list
2255 * may change while processing a work, so this function repeatedly
2256 * fetches a work from the top and executes it.
2259 * spin_lock_irq(pool->lock) which may be released and regrabbed
2262 static void process_scheduled_works(struct worker *worker)
2264 while (!list_empty(&worker->scheduled)) {
2265 struct work_struct *work = list_first_entry(&worker->scheduled,
2266 struct work_struct, entry);
2267 process_one_work(worker, work);
2272 * worker_thread - the worker thread function
2275 * The worker thread function. All workers belong to a worker_pool -
2276 * either a per-cpu one or dynamic unbound one. These workers process all
2277 * work items regardless of their specific target workqueue. The only
2278 * exception is work items which belong to workqueues with a rescuer which
2279 * will be explained in rescuer_thread().
2281 static int worker_thread(void *__worker)
2283 struct worker *worker = __worker;
2284 struct worker_pool *pool = worker->pool;
2286 /* tell the scheduler that this is a workqueue worker */
2287 worker->task->flags |= PF_WQ_WORKER;
2289 spin_lock_irq(&pool->lock);
2291 /* we are off idle list if destruction or rebind is requested */
2292 if (unlikely(list_empty(&worker->entry))) {
2293 spin_unlock_irq(&pool->lock);
2295 /* if DIE is set, destruction is requested */
2296 if (worker->flags & WORKER_DIE) {
2297 worker->task->flags &= ~PF_WQ_WORKER;
2301 /* otherwise, rebind */
2302 idle_worker_rebind(worker);
2306 worker_leave_idle(worker);
2308 /* no more worker necessary? */
2309 if (!need_more_worker(pool))
2312 /* do we need to manage? */
2313 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2317 * ->scheduled list can only be filled while a worker is
2318 * preparing to process a work or actually processing it.
2319 * Make sure nobody diddled with it while I was sleeping.
2321 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2324 * When control reaches this point, we're guaranteed to have
2325 * at least one idle worker or that someone else has already
2326 * assumed the manager role.
2328 worker_clr_flags(worker, WORKER_PREP);
2331 struct work_struct *work =
2332 list_first_entry(&pool->worklist,
2333 struct work_struct, entry);
2335 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2336 /* optimization path, not strictly necessary */
2337 process_one_work(worker, work);
2338 if (unlikely(!list_empty(&worker->scheduled)))
2339 process_scheduled_works(worker);
2341 move_linked_works(work, &worker->scheduled, NULL);
2342 process_scheduled_works(worker);
2344 } while (keep_working(pool));
2346 worker_set_flags(worker, WORKER_PREP, false);
2348 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2352 * pool->lock is held and there's no work to process and no need to
2353 * manage, sleep. Workers are woken up only while holding
2354 * pool->lock or from local cpu, so setting the current state
2355 * before releasing pool->lock is enough to prevent losing any
2358 worker_enter_idle(worker);
2359 __set_current_state(TASK_INTERRUPTIBLE);
2360 spin_unlock_irq(&pool->lock);
2366 * rescuer_thread - the rescuer thread function
2369 * Workqueue rescuer thread function. There's one rescuer for each
2370 * workqueue which has WQ_MEM_RECLAIM set.
2372 * Regular work processing on a pool may block trying to create a new
2373 * worker which uses GFP_KERNEL allocation which has slight chance of
2374 * developing into deadlock if some works currently on the same queue
2375 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2376 * the problem rescuer solves.
2378 * When such condition is possible, the pool summons rescuers of all
2379 * workqueues which have works queued on the pool and let them process
2380 * those works so that forward progress can be guaranteed.
2382 * This should happen rarely.
2384 static int rescuer_thread(void *__rescuer)
2386 struct worker *rescuer = __rescuer;
2387 struct workqueue_struct *wq = rescuer->rescue_wq;
2388 struct list_head *scheduled = &rescuer->scheduled;
2390 set_user_nice(current, RESCUER_NICE_LEVEL);
2393 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2394 * doesn't participate in concurrency management.
2396 rescuer->task->flags |= PF_WQ_WORKER;
2398 set_current_state(TASK_INTERRUPTIBLE);
2400 if (kthread_should_stop()) {
2401 __set_current_state(TASK_RUNNING);
2402 rescuer->task->flags &= ~PF_WQ_WORKER;
2406 /* see whether any pwq is asking for help */
2407 spin_lock_irq(&workqueue_lock);
2409 while (!list_empty(&wq->maydays)) {
2410 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2411 struct pool_workqueue, mayday_node);
2412 struct worker_pool *pool = pwq->pool;
2413 struct work_struct *work, *n;
2415 __set_current_state(TASK_RUNNING);
2416 list_del_init(&pwq->mayday_node);
2418 spin_unlock_irq(&workqueue_lock);
2420 /* migrate to the target cpu if possible */
2421 worker_maybe_bind_and_lock(pool);
2422 rescuer->pool = pool;
2425 * Slurp in all works issued via this workqueue and
2428 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2429 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2430 if (get_work_pwq(work) == pwq)
2431 move_linked_works(work, scheduled, &n);
2433 process_scheduled_works(rescuer);
2436 * Leave this pool. If keep_working() is %true, notify a
2437 * regular worker; otherwise, we end up with 0 concurrency
2438 * and stalling the execution.
2440 if (keep_working(pool))
2441 wake_up_worker(pool);
2443 rescuer->pool = NULL;
2444 spin_unlock(&pool->lock);
2445 spin_lock(&workqueue_lock);
2448 spin_unlock_irq(&workqueue_lock);
2450 /* rescuers should never participate in concurrency management */
2451 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2457 struct work_struct work;
2458 struct completion done;
2461 static void wq_barrier_func(struct work_struct *work)
2463 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2464 complete(&barr->done);
2468 * insert_wq_barrier - insert a barrier work
2469 * @pwq: pwq to insert barrier into
2470 * @barr: wq_barrier to insert
2471 * @target: target work to attach @barr to
2472 * @worker: worker currently executing @target, NULL if @target is not executing
2474 * @barr is linked to @target such that @barr is completed only after
2475 * @target finishes execution. Please note that the ordering
2476 * guarantee is observed only with respect to @target and on the local
2479 * Currently, a queued barrier can't be canceled. This is because
2480 * try_to_grab_pending() can't determine whether the work to be
2481 * grabbed is at the head of the queue and thus can't clear LINKED
2482 * flag of the previous work while there must be a valid next work
2483 * after a work with LINKED flag set.
2485 * Note that when @worker is non-NULL, @target may be modified
2486 * underneath us, so we can't reliably determine pwq from @target.
2489 * spin_lock_irq(pool->lock).
2491 static void insert_wq_barrier(struct pool_workqueue *pwq,
2492 struct wq_barrier *barr,
2493 struct work_struct *target, struct worker *worker)
2495 struct list_head *head;
2496 unsigned int linked = 0;
2499 * debugobject calls are safe here even with pool->lock locked
2500 * as we know for sure that this will not trigger any of the
2501 * checks and call back into the fixup functions where we
2504 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2505 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2506 init_completion(&barr->done);
2509 * If @target is currently being executed, schedule the
2510 * barrier to the worker; otherwise, put it after @target.
2513 head = worker->scheduled.next;
2515 unsigned long *bits = work_data_bits(target);
2517 head = target->entry.next;
2518 /* there can already be other linked works, inherit and set */
2519 linked = *bits & WORK_STRUCT_LINKED;
2520 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2523 debug_work_activate(&barr->work);
2524 insert_work(pwq, &barr->work, head,
2525 work_color_to_flags(WORK_NO_COLOR) | linked);
2529 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2530 * @wq: workqueue being flushed
2531 * @flush_color: new flush color, < 0 for no-op
2532 * @work_color: new work color, < 0 for no-op
2534 * Prepare pwqs for workqueue flushing.
2536 * If @flush_color is non-negative, flush_color on all pwqs should be
2537 * -1. If no pwq has in-flight commands at the specified color, all
2538 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2539 * has in flight commands, its pwq->flush_color is set to
2540 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2541 * wakeup logic is armed and %true is returned.
2543 * The caller should have initialized @wq->first_flusher prior to
2544 * calling this function with non-negative @flush_color. If
2545 * @flush_color is negative, no flush color update is done and %false
2548 * If @work_color is non-negative, all pwqs should have the same
2549 * work_color which is previous to @work_color and all will be
2550 * advanced to @work_color.
2553 * mutex_lock(wq->flush_mutex).
2556 * %true if @flush_color >= 0 and there's something to flush. %false
2559 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2560 int flush_color, int work_color)
2563 struct pool_workqueue *pwq;
2565 if (flush_color >= 0) {
2566 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2567 atomic_set(&wq->nr_pwqs_to_flush, 1);
2570 local_irq_disable();
2572 for_each_pwq(pwq, wq) {
2573 struct worker_pool *pool = pwq->pool;
2575 spin_lock(&pool->lock);
2577 if (flush_color >= 0) {
2578 WARN_ON_ONCE(pwq->flush_color != -1);
2580 if (pwq->nr_in_flight[flush_color]) {
2581 pwq->flush_color = flush_color;
2582 atomic_inc(&wq->nr_pwqs_to_flush);
2587 if (work_color >= 0) {
2588 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2589 pwq->work_color = work_color;
2592 spin_unlock(&pool->lock);
2597 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2598 complete(&wq->first_flusher->done);
2604 * flush_workqueue - ensure that any scheduled work has run to completion.
2605 * @wq: workqueue to flush
2607 * This function sleeps until all work items which were queued on entry
2608 * have finished execution, but it is not livelocked by new incoming ones.
2610 void flush_workqueue(struct workqueue_struct *wq)
2612 struct wq_flusher this_flusher = {
2613 .list = LIST_HEAD_INIT(this_flusher.list),
2615 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2619 lock_map_acquire(&wq->lockdep_map);
2620 lock_map_release(&wq->lockdep_map);
2622 mutex_lock(&wq->flush_mutex);
2625 * Start-to-wait phase
2627 next_color = work_next_color(wq->work_color);
2629 if (next_color != wq->flush_color) {
2631 * Color space is not full. The current work_color
2632 * becomes our flush_color and work_color is advanced
2635 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2636 this_flusher.flush_color = wq->work_color;
2637 wq->work_color = next_color;
2639 if (!wq->first_flusher) {
2640 /* no flush in progress, become the first flusher */
2641 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2643 wq->first_flusher = &this_flusher;
2645 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2647 /* nothing to flush, done */
2648 wq->flush_color = next_color;
2649 wq->first_flusher = NULL;
2654 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2655 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2656 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2660 * Oops, color space is full, wait on overflow queue.
2661 * The next flush completion will assign us
2662 * flush_color and transfer to flusher_queue.
2664 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2667 mutex_unlock(&wq->flush_mutex);
2669 wait_for_completion(&this_flusher.done);
2672 * Wake-up-and-cascade phase
2674 * First flushers are responsible for cascading flushes and
2675 * handling overflow. Non-first flushers can simply return.
2677 if (wq->first_flusher != &this_flusher)
2680 mutex_lock(&wq->flush_mutex);
2682 /* we might have raced, check again with mutex held */
2683 if (wq->first_flusher != &this_flusher)
2686 wq->first_flusher = NULL;
2688 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2689 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2692 struct wq_flusher *next, *tmp;
2694 /* complete all the flushers sharing the current flush color */
2695 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2696 if (next->flush_color != wq->flush_color)
2698 list_del_init(&next->list);
2699 complete(&next->done);
2702 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2703 wq->flush_color != work_next_color(wq->work_color));
2705 /* this flush_color is finished, advance by one */
2706 wq->flush_color = work_next_color(wq->flush_color);
2708 /* one color has been freed, handle overflow queue */
2709 if (!list_empty(&wq->flusher_overflow)) {
2711 * Assign the same color to all overflowed
2712 * flushers, advance work_color and append to
2713 * flusher_queue. This is the start-to-wait
2714 * phase for these overflowed flushers.
2716 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2717 tmp->flush_color = wq->work_color;
2719 wq->work_color = work_next_color(wq->work_color);
2721 list_splice_tail_init(&wq->flusher_overflow,
2722 &wq->flusher_queue);
2723 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2726 if (list_empty(&wq->flusher_queue)) {
2727 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2732 * Need to flush more colors. Make the next flusher
2733 * the new first flusher and arm pwqs.
2735 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2736 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2738 list_del_init(&next->list);
2739 wq->first_flusher = next;
2741 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2745 * Meh... this color is already done, clear first
2746 * flusher and repeat cascading.
2748 wq->first_flusher = NULL;
2752 mutex_unlock(&wq->flush_mutex);
2754 EXPORT_SYMBOL_GPL(flush_workqueue);
2757 * drain_workqueue - drain a workqueue
2758 * @wq: workqueue to drain
2760 * Wait until the workqueue becomes empty. While draining is in progress,
2761 * only chain queueing is allowed. IOW, only currently pending or running
2762 * work items on @wq can queue further work items on it. @wq is flushed
2763 * repeatedly until it becomes empty. The number of flushing is detemined
2764 * by the depth of chaining and should be relatively short. Whine if it
2767 void drain_workqueue(struct workqueue_struct *wq)
2769 unsigned int flush_cnt = 0;
2770 struct pool_workqueue *pwq;
2773 * __queue_work() needs to test whether there are drainers, is much
2774 * hotter than drain_workqueue() and already looks at @wq->flags.
2775 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2777 mutex_lock(&wq_mutex);
2778 if (!wq->nr_drainers++)
2779 wq->flags |= __WQ_DRAINING;
2780 mutex_unlock(&wq_mutex);
2782 flush_workqueue(wq);
2784 local_irq_disable();
2786 for_each_pwq(pwq, wq) {
2789 spin_lock(&pwq->pool->lock);
2790 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2791 spin_unlock(&pwq->pool->lock);
2796 if (++flush_cnt == 10 ||
2797 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2798 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2799 wq->name, flush_cnt);
2807 mutex_lock(&wq_mutex);
2808 if (!--wq->nr_drainers)
2809 wq->flags &= ~__WQ_DRAINING;
2810 mutex_unlock(&wq_mutex);
2812 EXPORT_SYMBOL_GPL(drain_workqueue);
2814 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2816 struct worker *worker = NULL;
2817 struct worker_pool *pool;
2818 struct pool_workqueue *pwq;
2822 local_irq_disable();
2823 pool = get_work_pool(work);
2829 spin_lock(&pool->lock);
2830 /* see the comment in try_to_grab_pending() with the same code */
2831 pwq = get_work_pwq(work);
2833 if (unlikely(pwq->pool != pool))
2836 worker = find_worker_executing_work(pool, work);
2839 pwq = worker->current_pwq;
2842 insert_wq_barrier(pwq, barr, work, worker);
2843 spin_unlock_irq(&pool->lock);
2846 * If @max_active is 1 or rescuer is in use, flushing another work
2847 * item on the same workqueue may lead to deadlock. Make sure the
2848 * flusher is not running on the same workqueue by verifying write
2851 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2852 lock_map_acquire(&pwq->wq->lockdep_map);
2854 lock_map_acquire_read(&pwq->wq->lockdep_map);
2855 lock_map_release(&pwq->wq->lockdep_map);
2859 spin_unlock_irq(&pool->lock);
2864 * flush_work - wait for a work to finish executing the last queueing instance
2865 * @work: the work to flush
2867 * Wait until @work has finished execution. @work is guaranteed to be idle
2868 * on return if it hasn't been requeued since flush started.
2871 * %true if flush_work() waited for the work to finish execution,
2872 * %false if it was already idle.
2874 bool flush_work(struct work_struct *work)
2876 struct wq_barrier barr;
2878 lock_map_acquire(&work->lockdep_map);
2879 lock_map_release(&work->lockdep_map);
2881 if (start_flush_work(work, &barr)) {
2882 wait_for_completion(&barr.done);
2883 destroy_work_on_stack(&barr.work);
2889 EXPORT_SYMBOL_GPL(flush_work);
2891 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2893 unsigned long flags;
2897 ret = try_to_grab_pending(work, is_dwork, &flags);
2899 * If someone else is canceling, wait for the same event it
2900 * would be waiting for before retrying.
2902 if (unlikely(ret == -ENOENT))
2904 } while (unlikely(ret < 0));
2906 /* tell other tasks trying to grab @work to back off */
2907 mark_work_canceling(work);
2908 local_irq_restore(flags);
2911 clear_work_data(work);
2916 * cancel_work_sync - cancel a work and wait for it to finish
2917 * @work: the work to cancel
2919 * Cancel @work and wait for its execution to finish. This function
2920 * can be used even if the work re-queues itself or migrates to
2921 * another workqueue. On return from this function, @work is
2922 * guaranteed to be not pending or executing on any CPU.
2924 * cancel_work_sync(&delayed_work->work) must not be used for
2925 * delayed_work's. Use cancel_delayed_work_sync() instead.
2927 * The caller must ensure that the workqueue on which @work was last
2928 * queued can't be destroyed before this function returns.
2931 * %true if @work was pending, %false otherwise.
2933 bool cancel_work_sync(struct work_struct *work)
2935 return __cancel_work_timer(work, false);
2937 EXPORT_SYMBOL_GPL(cancel_work_sync);
2940 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2941 * @dwork: the delayed work to flush
2943 * Delayed timer is cancelled and the pending work is queued for
2944 * immediate execution. Like flush_work(), this function only
2945 * considers the last queueing instance of @dwork.
2948 * %true if flush_work() waited for the work to finish execution,
2949 * %false if it was already idle.
2951 bool flush_delayed_work(struct delayed_work *dwork)
2953 local_irq_disable();
2954 if (del_timer_sync(&dwork->timer))
2955 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2957 return flush_work(&dwork->work);
2959 EXPORT_SYMBOL(flush_delayed_work);
2962 * cancel_delayed_work - cancel a delayed work
2963 * @dwork: delayed_work to cancel
2965 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2966 * and canceled; %false if wasn't pending. Note that the work callback
2967 * function may still be running on return, unless it returns %true and the
2968 * work doesn't re-arm itself. Explicitly flush or use
2969 * cancel_delayed_work_sync() to wait on it.
2971 * This function is safe to call from any context including IRQ handler.
2973 bool cancel_delayed_work(struct delayed_work *dwork)
2975 unsigned long flags;
2979 ret = try_to_grab_pending(&dwork->work, true, &flags);
2980 } while (unlikely(ret == -EAGAIN));
2982 if (unlikely(ret < 0))
2985 set_work_pool_and_clear_pending(&dwork->work,
2986 get_work_pool_id(&dwork->work));
2987 local_irq_restore(flags);
2990 EXPORT_SYMBOL(cancel_delayed_work);
2993 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2994 * @dwork: the delayed work cancel
2996 * This is cancel_work_sync() for delayed works.
2999 * %true if @dwork was pending, %false otherwise.
3001 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3003 return __cancel_work_timer(&dwork->work, true);
3005 EXPORT_SYMBOL(cancel_delayed_work_sync);
3008 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3009 * @func: the function to call
3011 * schedule_on_each_cpu() executes @func on each online CPU using the
3012 * system workqueue and blocks until all CPUs have completed.
3013 * schedule_on_each_cpu() is very slow.
3016 * 0 on success, -errno on failure.
3018 int schedule_on_each_cpu(work_func_t func)
3021 struct work_struct __percpu *works;
3023 works = alloc_percpu(struct work_struct);
3029 for_each_online_cpu(cpu) {
3030 struct work_struct *work = per_cpu_ptr(works, cpu);
3032 INIT_WORK(work, func);
3033 schedule_work_on(cpu, work);
3036 for_each_online_cpu(cpu)
3037 flush_work(per_cpu_ptr(works, cpu));
3045 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3047 * Forces execution of the kernel-global workqueue and blocks until its
3050 * Think twice before calling this function! It's very easy to get into
3051 * trouble if you don't take great care. Either of the following situations
3052 * will lead to deadlock:
3054 * One of the work items currently on the workqueue needs to acquire
3055 * a lock held by your code or its caller.
3057 * Your code is running in the context of a work routine.
3059 * They will be detected by lockdep when they occur, but the first might not
3060 * occur very often. It depends on what work items are on the workqueue and
3061 * what locks they need, which you have no control over.
3063 * In most situations flushing the entire workqueue is overkill; you merely
3064 * need to know that a particular work item isn't queued and isn't running.
3065 * In such cases you should use cancel_delayed_work_sync() or
3066 * cancel_work_sync() instead.
3068 void flush_scheduled_work(void)
3070 flush_workqueue(system_wq);
3072 EXPORT_SYMBOL(flush_scheduled_work);
3075 * execute_in_process_context - reliably execute the routine with user context
3076 * @fn: the function to execute
3077 * @ew: guaranteed storage for the execute work structure (must
3078 * be available when the work executes)
3080 * Executes the function immediately if process context is available,
3081 * otherwise schedules the function for delayed execution.
3083 * Returns: 0 - function was executed
3084 * 1 - function was scheduled for execution
3086 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3088 if (!in_interrupt()) {
3093 INIT_WORK(&ew->work, fn);
3094 schedule_work(&ew->work);
3098 EXPORT_SYMBOL_GPL(execute_in_process_context);
3102 * Workqueues with WQ_SYSFS flag set is visible to userland via
3103 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3104 * following attributes.
3106 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3107 * max_active RW int : maximum number of in-flight work items
3109 * Unbound workqueues have the following extra attributes.
3111 * id RO int : the associated pool ID
3112 * nice RW int : nice value of the workers
3113 * cpumask RW mask : bitmask of allowed CPUs for the workers
3116 struct workqueue_struct *wq;
3120 static struct workqueue_struct *dev_to_wq(struct device *dev)
3122 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3127 static ssize_t wq_per_cpu_show(struct device *dev,
3128 struct device_attribute *attr, char *buf)
3130 struct workqueue_struct *wq = dev_to_wq(dev);
3132 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3135 static ssize_t wq_max_active_show(struct device *dev,
3136 struct device_attribute *attr, char *buf)
3138 struct workqueue_struct *wq = dev_to_wq(dev);
3140 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3143 static ssize_t wq_max_active_store(struct device *dev,
3144 struct device_attribute *attr,
3145 const char *buf, size_t count)
3147 struct workqueue_struct *wq = dev_to_wq(dev);
3150 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3153 workqueue_set_max_active(wq, val);
3157 static struct device_attribute wq_sysfs_attrs[] = {
3158 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3159 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3163 static ssize_t wq_pool_id_show(struct device *dev,
3164 struct device_attribute *attr, char *buf)
3166 struct workqueue_struct *wq = dev_to_wq(dev);
3167 struct worker_pool *pool;
3170 rcu_read_lock_sched();
3171 pool = first_pwq(wq)->pool;
3172 written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id);
3173 rcu_read_unlock_sched();
3178 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3181 struct workqueue_struct *wq = dev_to_wq(dev);
3184 rcu_read_lock_sched();
3185 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3186 first_pwq(wq)->pool->attrs->nice);
3187 rcu_read_unlock_sched();
3192 /* prepare workqueue_attrs for sysfs store operations */
3193 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3195 struct workqueue_attrs *attrs;
3197 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3201 rcu_read_lock_sched();
3202 copy_workqueue_attrs(attrs, first_pwq(wq)->pool->attrs);
3203 rcu_read_unlock_sched();
3207 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3208 const char *buf, size_t count)
3210 struct workqueue_struct *wq = dev_to_wq(dev);
3211 struct workqueue_attrs *attrs;
3214 attrs = wq_sysfs_prep_attrs(wq);
3218 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3219 attrs->nice >= -20 && attrs->nice <= 19)
3220 ret = apply_workqueue_attrs(wq, attrs);
3224 free_workqueue_attrs(attrs);
3225 return ret ?: count;
3228 static ssize_t wq_cpumask_show(struct device *dev,
3229 struct device_attribute *attr, char *buf)
3231 struct workqueue_struct *wq = dev_to_wq(dev);
3234 rcu_read_lock_sched();
3235 written = cpumask_scnprintf(buf, PAGE_SIZE,
3236 first_pwq(wq)->pool->attrs->cpumask);
3237 rcu_read_unlock_sched();
3239 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3243 static ssize_t wq_cpumask_store(struct device *dev,
3244 struct device_attribute *attr,
3245 const char *buf, size_t count)
3247 struct workqueue_struct *wq = dev_to_wq(dev);
3248 struct workqueue_attrs *attrs;
3251 attrs = wq_sysfs_prep_attrs(wq);
3255 ret = cpumask_parse(buf, attrs->cpumask);
3257 ret = apply_workqueue_attrs(wq, attrs);
3259 free_workqueue_attrs(attrs);
3260 return ret ?: count;
3263 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3264 __ATTR(pool_id, 0444, wq_pool_id_show, NULL),
3265 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3266 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3270 static struct bus_type wq_subsys = {
3271 .name = "workqueue",
3272 .dev_attrs = wq_sysfs_attrs,
3275 static int __init wq_sysfs_init(void)
3277 return subsys_virtual_register(&wq_subsys, NULL);
3279 core_initcall(wq_sysfs_init);
3281 static void wq_device_release(struct device *dev)
3283 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3289 * workqueue_sysfs_register - make a workqueue visible in sysfs
3290 * @wq: the workqueue to register
3292 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3293 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3294 * which is the preferred method.
3296 * Workqueue user should use this function directly iff it wants to apply
3297 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3298 * apply_workqueue_attrs() may race against userland updating the
3301 * Returns 0 on success, -errno on failure.
3303 int workqueue_sysfs_register(struct workqueue_struct *wq)
3305 struct wq_device *wq_dev;
3309 * Adjusting max_active or creating new pwqs by applyting
3310 * attributes breaks ordering guarantee. Disallow exposing ordered
3313 if (WARN_ON(wq->flags & __WQ_ORDERED))
3316 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3321 wq_dev->dev.bus = &wq_subsys;
3322 wq_dev->dev.init_name = wq->name;
3323 wq_dev->dev.release = wq_device_release;
3326 * unbound_attrs are created separately. Suppress uevent until
3327 * everything is ready.
3329 dev_set_uevent_suppress(&wq_dev->dev, true);
3331 ret = device_register(&wq_dev->dev);
3338 if (wq->flags & WQ_UNBOUND) {
3339 struct device_attribute *attr;
3341 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3342 ret = device_create_file(&wq_dev->dev, attr);
3344 device_unregister(&wq_dev->dev);
3351 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3356 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3357 * @wq: the workqueue to unregister
3359 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3361 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3363 struct wq_device *wq_dev = wq->wq_dev;
3369 device_unregister(&wq_dev->dev);
3371 #else /* CONFIG_SYSFS */
3372 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3373 #endif /* CONFIG_SYSFS */
3376 * free_workqueue_attrs - free a workqueue_attrs
3377 * @attrs: workqueue_attrs to free
3379 * Undo alloc_workqueue_attrs().
3381 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3384 free_cpumask_var(attrs->cpumask);
3390 * alloc_workqueue_attrs - allocate a workqueue_attrs
3391 * @gfp_mask: allocation mask to use
3393 * Allocate a new workqueue_attrs, initialize with default settings and
3394 * return it. Returns NULL on failure.
3396 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3398 struct workqueue_attrs *attrs;
3400 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3403 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3406 cpumask_setall(attrs->cpumask);
3409 free_workqueue_attrs(attrs);
3413 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3414 const struct workqueue_attrs *from)
3416 to->nice = from->nice;
3417 cpumask_copy(to->cpumask, from->cpumask);
3421 * Hacky implementation of jhash of bitmaps which only considers the
3422 * specified number of bits. We probably want a proper implementation in
3423 * include/linux/jhash.h.
3425 static u32 jhash_bitmap(const unsigned long *bitmap, int bits, u32 hash)
3427 int nr_longs = bits / BITS_PER_LONG;
3428 int nr_leftover = bits % BITS_PER_LONG;
3429 unsigned long leftover = 0;
3432 hash = jhash(bitmap, nr_longs * sizeof(long), hash);
3434 bitmap_copy(&leftover, bitmap + nr_longs, nr_leftover);
3435 hash = jhash(&leftover, sizeof(long), hash);
3440 /* hash value of the content of @attr */
3441 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3445 hash = jhash_1word(attrs->nice, hash);
3446 hash = jhash_bitmap(cpumask_bits(attrs->cpumask), nr_cpu_ids, hash);
3450 /* content equality test */
3451 static bool wqattrs_equal(const struct workqueue_attrs *a,
3452 const struct workqueue_attrs *b)
3454 if (a->nice != b->nice)
3456 if (!cpumask_equal(a->cpumask, b->cpumask))
3462 * init_worker_pool - initialize a newly zalloc'd worker_pool
3463 * @pool: worker_pool to initialize
3465 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3466 * Returns 0 on success, -errno on failure. Even on failure, all fields
3467 * inside @pool proper are initialized and put_unbound_pool() can be called
3468 * on @pool safely to release it.
3470 static int init_worker_pool(struct worker_pool *pool)
3472 spin_lock_init(&pool->lock);
3475 pool->flags |= POOL_DISASSOCIATED;
3476 INIT_LIST_HEAD(&pool->worklist);
3477 INIT_LIST_HEAD(&pool->idle_list);
3478 hash_init(pool->busy_hash);
3480 init_timer_deferrable(&pool->idle_timer);
3481 pool->idle_timer.function = idle_worker_timeout;
3482 pool->idle_timer.data = (unsigned long)pool;
3484 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3485 (unsigned long)pool);
3487 mutex_init(&pool->manager_arb);
3488 mutex_init(&pool->manager_mutex);
3489 ida_init(&pool->worker_ida);
3491 INIT_HLIST_NODE(&pool->hash_node);
3494 /* shouldn't fail above this point */
3495 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3501 static void rcu_free_pool(struct rcu_head *rcu)
3503 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3505 ida_destroy(&pool->worker_ida);
3506 free_workqueue_attrs(pool->attrs);
3511 * put_unbound_pool - put a worker_pool
3512 * @pool: worker_pool to put
3514 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3515 * safe manner. get_unbound_pool() calls this function on its failure path
3516 * and this function should be able to release pools which went through,
3517 * successfully or not, init_worker_pool().
3519 static void put_unbound_pool(struct worker_pool *pool)
3521 struct worker *worker;
3523 mutex_lock(&wq_mutex);
3524 if (--pool->refcnt) {
3525 mutex_unlock(&wq_mutex);
3530 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3531 WARN_ON(!list_empty(&pool->worklist))) {
3532 mutex_unlock(&wq_mutex);
3536 /* release id and unhash */
3538 idr_remove(&worker_pool_idr, pool->id);
3539 hash_del(&pool->hash_node);
3541 mutex_unlock(&wq_mutex);
3544 * Become the manager and destroy all workers. Grabbing
3545 * manager_arb prevents @pool's workers from blocking on
3548 mutex_lock(&pool->manager_arb);
3549 mutex_lock(&pool->manager_mutex);
3550 spin_lock_irq(&pool->lock);
3552 while ((worker = first_worker(pool)))
3553 destroy_worker(worker);
3554 WARN_ON(pool->nr_workers || pool->nr_idle);
3556 spin_unlock_irq(&pool->lock);
3557 mutex_unlock(&pool->manager_mutex);
3558 mutex_unlock(&pool->manager_arb);
3560 /* shut down the timers */
3561 del_timer_sync(&pool->idle_timer);
3562 del_timer_sync(&pool->mayday_timer);
3564 /* sched-RCU protected to allow dereferences from get_work_pool() */
3565 call_rcu_sched(&pool->rcu, rcu_free_pool);
3569 * get_unbound_pool - get a worker_pool with the specified attributes
3570 * @attrs: the attributes of the worker_pool to get
3572 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3573 * reference count and return it. If there already is a matching
3574 * worker_pool, it will be used; otherwise, this function attempts to
3575 * create a new one. On failure, returns NULL.
3577 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3579 u32 hash = wqattrs_hash(attrs);
3580 struct worker_pool *pool;
3582 mutex_lock(&wq_mutex);
3584 /* do we already have a matching pool? */
3585 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3586 if (wqattrs_equal(pool->attrs, attrs)) {
3592 /* nope, create a new one */
3593 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3594 if (!pool || init_worker_pool(pool) < 0)
3597 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3598 copy_workqueue_attrs(pool->attrs, attrs);
3600 if (worker_pool_assign_id(pool) < 0)
3603 /* create and start the initial worker */
3604 if (create_and_start_worker(pool) < 0)
3608 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3610 mutex_unlock(&wq_mutex);
3613 mutex_unlock(&wq_mutex);
3615 put_unbound_pool(pool);
3619 static void rcu_free_pwq(struct rcu_head *rcu)
3621 kmem_cache_free(pwq_cache,
3622 container_of(rcu, struct pool_workqueue, rcu));
3626 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3627 * and needs to be destroyed.
3629 static void pwq_unbound_release_workfn(struct work_struct *work)
3631 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3632 unbound_release_work);
3633 struct workqueue_struct *wq = pwq->wq;
3634 struct worker_pool *pool = pwq->pool;
3636 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3640 * Unlink @pwq. Synchronization against flush_mutex isn't strictly
3641 * necessary on release but do it anyway. It's easier to verify
3642 * and consistent with the linking path.
3644 mutex_lock(&wq->flush_mutex);
3645 spin_lock_irq(&pwq_lock);
3646 list_del_rcu(&pwq->pwqs_node);
3647 spin_unlock_irq(&pwq_lock);
3648 mutex_unlock(&wq->flush_mutex);
3650 put_unbound_pool(pool);
3651 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3654 * If we're the last pwq going away, @wq is already dead and no one
3655 * is gonna access it anymore. Free it.
3657 if (list_empty(&wq->pwqs))
3662 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3663 * @pwq: target pool_workqueue
3665 * If @pwq isn't freezing, set @pwq->max_active to the associated
3666 * workqueue's saved_max_active and activate delayed work items
3667 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3669 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3671 struct workqueue_struct *wq = pwq->wq;
3672 bool freezable = wq->flags & WQ_FREEZABLE;
3674 /* for @wq->saved_max_active */
3675 lockdep_assert_held(&pwq_lock);
3677 /* fast exit for non-freezable wqs */
3678 if (!freezable && pwq->max_active == wq->saved_max_active)
3681 spin_lock(&pwq->pool->lock);
3683 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3684 pwq->max_active = wq->saved_max_active;
3686 while (!list_empty(&pwq->delayed_works) &&
3687 pwq->nr_active < pwq->max_active)
3688 pwq_activate_first_delayed(pwq);
3690 pwq->max_active = 0;
3693 spin_unlock(&pwq->pool->lock);
3696 static void init_and_link_pwq(struct pool_workqueue *pwq,
3697 struct workqueue_struct *wq,
3698 struct worker_pool *pool,
3699 struct pool_workqueue **p_last_pwq)
3701 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3705 pwq->flush_color = -1;
3707 INIT_LIST_HEAD(&pwq->delayed_works);
3708 INIT_LIST_HEAD(&pwq->mayday_node);
3709 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3711 mutex_lock(&wq->flush_mutex);
3712 spin_lock_irq(&pwq_lock);
3715 * Set the matching work_color. This is synchronized with
3716 * flush_mutex to avoid confusing flush_workqueue().
3719 *p_last_pwq = first_pwq(wq);
3720 pwq->work_color = wq->work_color;
3722 /* sync max_active to the current setting */
3723 pwq_adjust_max_active(pwq);
3726 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3728 spin_unlock_irq(&pwq_lock);
3729 mutex_unlock(&wq->flush_mutex);
3733 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3734 * @wq: the target workqueue
3735 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3737 * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the
3738 * current attributes, a new pwq is created and made the first pwq which
3739 * will serve all new work items. Older pwqs are released as in-flight
3740 * work items finish. Note that a work item which repeatedly requeues
3741 * itself back-to-back will stay on its current pwq.
3743 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3746 int apply_workqueue_attrs(struct workqueue_struct *wq,
3747 const struct workqueue_attrs *attrs)
3749 struct pool_workqueue *pwq, *last_pwq;
3750 struct worker_pool *pool;
3752 /* only unbound workqueues can change attributes */
3753 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3756 /* creating multiple pwqs breaks ordering guarantee */
3757 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3760 pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL);
3764 pool = get_unbound_pool(attrs);
3766 kmem_cache_free(pwq_cache, pwq);
3770 init_and_link_pwq(pwq, wq, pool, &last_pwq);
3772 spin_lock_irq(&last_pwq->pool->lock);
3774 spin_unlock_irq(&last_pwq->pool->lock);
3780 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3782 bool highpri = wq->flags & WQ_HIGHPRI;
3785 if (!(wq->flags & WQ_UNBOUND)) {
3786 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3790 for_each_possible_cpu(cpu) {
3791 struct pool_workqueue *pwq =
3792 per_cpu_ptr(wq->cpu_pwqs, cpu);
3793 struct worker_pool *cpu_pools =
3794 per_cpu(cpu_worker_pools, cpu);
3796 init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL);
3800 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3804 static int wq_clamp_max_active(int max_active, unsigned int flags,
3807 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3809 if (max_active < 1 || max_active > lim)
3810 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3811 max_active, name, 1, lim);
3813 return clamp_val(max_active, 1, lim);
3816 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3819 struct lock_class_key *key,
3820 const char *lock_name, ...)
3822 va_list args, args1;
3823 struct workqueue_struct *wq;
3824 struct pool_workqueue *pwq;
3827 /* determine namelen, allocate wq and format name */
3828 va_start(args, lock_name);
3829 va_copy(args1, args);
3830 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3832 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3836 vsnprintf(wq->name, namelen, fmt, args1);
3840 max_active = max_active ?: WQ_DFL_ACTIVE;
3841 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3845 wq->saved_max_active = max_active;
3846 mutex_init(&wq->flush_mutex);
3847 atomic_set(&wq->nr_pwqs_to_flush, 0);
3848 INIT_LIST_HEAD(&wq->pwqs);
3849 INIT_LIST_HEAD(&wq->flusher_queue);
3850 INIT_LIST_HEAD(&wq->flusher_overflow);
3851 INIT_LIST_HEAD(&wq->maydays);
3853 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3854 INIT_LIST_HEAD(&wq->list);
3856 if (alloc_and_link_pwqs(wq) < 0)
3860 * Workqueues which may be used during memory reclaim should
3861 * have a rescuer to guarantee forward progress.
3863 if (flags & WQ_MEM_RECLAIM) {
3864 struct worker *rescuer;
3866 rescuer = alloc_worker();
3870 rescuer->rescue_wq = wq;
3871 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3873 if (IS_ERR(rescuer->task)) {
3878 wq->rescuer = rescuer;
3879 rescuer->task->flags |= PF_THREAD_BOUND;
3880 wake_up_process(rescuer->task);
3883 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3887 * wq_mutex protects global freeze state and workqueues list. Grab
3888 * it, adjust max_active and add the new @wq to workqueues list.
3890 mutex_lock(&wq_mutex);
3892 spin_lock_irq(&pwq_lock);
3893 for_each_pwq(pwq, wq)
3894 pwq_adjust_max_active(pwq);
3895 spin_unlock_irq(&pwq_lock);
3897 list_add(&wq->list, &workqueues);
3899 mutex_unlock(&wq_mutex);
3907 destroy_workqueue(wq);
3910 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3913 * destroy_workqueue - safely terminate a workqueue
3914 * @wq: target workqueue
3916 * Safely destroy a workqueue. All work currently pending will be done first.
3918 void destroy_workqueue(struct workqueue_struct *wq)
3920 struct pool_workqueue *pwq;
3922 /* drain it before proceeding with destruction */
3923 drain_workqueue(wq);
3926 spin_lock_irq(&pwq_lock);
3927 for_each_pwq(pwq, wq) {
3930 for (i = 0; i < WORK_NR_COLORS; i++) {
3931 if (WARN_ON(pwq->nr_in_flight[i])) {
3932 spin_unlock_irq(&pwq_lock);
3937 if (WARN_ON(pwq->refcnt > 1) ||
3938 WARN_ON(pwq->nr_active) ||
3939 WARN_ON(!list_empty(&pwq->delayed_works))) {
3940 spin_unlock_irq(&pwq_lock);
3944 spin_unlock_irq(&pwq_lock);
3947 * wq list is used to freeze wq, remove from list after
3948 * flushing is complete in case freeze races us.
3950 mutex_lock(&wq_mutex);
3951 list_del_init(&wq->list);
3952 mutex_unlock(&wq_mutex);
3954 workqueue_sysfs_unregister(wq);
3957 kthread_stop(wq->rescuer->task);
3962 if (!(wq->flags & WQ_UNBOUND)) {
3964 * The base ref is never dropped on per-cpu pwqs. Directly
3965 * free the pwqs and wq.
3967 free_percpu(wq->cpu_pwqs);
3971 * We're the sole accessor of @wq at this point. Directly
3972 * access the first pwq and put the base ref. As both pwqs
3973 * and pools are sched-RCU protected, the lock operations
3974 * are safe. @wq will be freed when the last pwq is
3977 pwq = list_first_entry(&wq->pwqs, struct pool_workqueue,
3979 spin_lock_irq(&pwq->pool->lock);
3981 spin_unlock_irq(&pwq->pool->lock);
3984 EXPORT_SYMBOL_GPL(destroy_workqueue);
3987 * workqueue_set_max_active - adjust max_active of a workqueue
3988 * @wq: target workqueue
3989 * @max_active: new max_active value.
3991 * Set max_active of @wq to @max_active.
3994 * Don't call from IRQ context.
3996 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3998 struct pool_workqueue *pwq;
4000 /* disallow meddling with max_active for ordered workqueues */
4001 if (WARN_ON(wq->flags & __WQ_ORDERED))
4004 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4006 spin_lock_irq(&pwq_lock);
4008 wq->saved_max_active = max_active;
4010 for_each_pwq(pwq, wq)
4011 pwq_adjust_max_active(pwq);
4013 spin_unlock_irq(&pwq_lock);
4015 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4018 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4020 * Determine whether %current is a workqueue rescuer. Can be used from
4021 * work functions to determine whether it's being run off the rescuer task.
4023 bool current_is_workqueue_rescuer(void)
4025 struct worker *worker = current_wq_worker();
4027 return worker && worker == worker->current_pwq->wq->rescuer;
4031 * workqueue_congested - test whether a workqueue is congested
4032 * @cpu: CPU in question
4033 * @wq: target workqueue
4035 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4036 * no synchronization around this function and the test result is
4037 * unreliable and only useful as advisory hints or for debugging.
4040 * %true if congested, %false otherwise.
4042 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4044 struct pool_workqueue *pwq;
4049 if (!(wq->flags & WQ_UNBOUND))
4050 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4052 pwq = first_pwq(wq);
4054 ret = !list_empty(&pwq->delayed_works);
4059 EXPORT_SYMBOL_GPL(workqueue_congested);
4062 * work_busy - test whether a work is currently pending or running
4063 * @work: the work to be tested
4065 * Test whether @work is currently pending or running. There is no
4066 * synchronization around this function and the test result is
4067 * unreliable and only useful as advisory hints or for debugging.
4070 * OR'd bitmask of WORK_BUSY_* bits.
4072 unsigned int work_busy(struct work_struct *work)
4074 struct worker_pool *pool;
4075 unsigned long flags;
4076 unsigned int ret = 0;
4078 if (work_pending(work))
4079 ret |= WORK_BUSY_PENDING;
4081 local_irq_save(flags);
4082 pool = get_work_pool(work);
4084 spin_lock(&pool->lock);
4085 if (find_worker_executing_work(pool, work))
4086 ret |= WORK_BUSY_RUNNING;
4087 spin_unlock(&pool->lock);
4089 local_irq_restore(flags);
4093 EXPORT_SYMBOL_GPL(work_busy);
4098 * There are two challenges in supporting CPU hotplug. Firstly, there
4099 * are a lot of assumptions on strong associations among work, pwq and
4100 * pool which make migrating pending and scheduled works very
4101 * difficult to implement without impacting hot paths. Secondly,
4102 * worker pools serve mix of short, long and very long running works making
4103 * blocked draining impractical.
4105 * This is solved by allowing the pools to be disassociated from the CPU
4106 * running as an unbound one and allowing it to be reattached later if the
4107 * cpu comes back online.
4110 static void wq_unbind_fn(struct work_struct *work)
4112 int cpu = smp_processor_id();
4113 struct worker_pool *pool;
4114 struct worker *worker;
4117 for_each_cpu_worker_pool(pool, cpu) {
4118 WARN_ON_ONCE(cpu != smp_processor_id());
4120 mutex_lock(&pool->manager_mutex);
4121 spin_lock_irq(&pool->lock);
4124 * We've blocked all manager operations. Make all workers
4125 * unbound and set DISASSOCIATED. Before this, all workers
4126 * except for the ones which are still executing works from
4127 * before the last CPU down must be on the cpu. After
4128 * this, they may become diasporas.
4130 list_for_each_entry(worker, &pool->idle_list, entry)
4131 worker->flags |= WORKER_UNBOUND;
4133 for_each_busy_worker(worker, i, pool)
4134 worker->flags |= WORKER_UNBOUND;
4136 pool->flags |= POOL_DISASSOCIATED;
4138 spin_unlock_irq(&pool->lock);
4139 mutex_unlock(&pool->manager_mutex);
4143 * Call schedule() so that we cross rq->lock and thus can guarantee
4144 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
4145 * as scheduler callbacks may be invoked from other cpus.
4150 * Sched callbacks are disabled now. Zap nr_running. After this,
4151 * nr_running stays zero and need_more_worker() and keep_working()
4152 * are always true as long as the worklist is not empty. Pools on
4153 * @cpu now behave as unbound (in terms of concurrency management)
4154 * pools which are served by workers tied to the CPU.
4156 * On return from this function, the current worker would trigger
4157 * unbound chain execution of pending work items if other workers
4160 for_each_cpu_worker_pool(pool, cpu)
4161 atomic_set(&pool->nr_running, 0);
4165 * Workqueues should be brought up before normal priority CPU notifiers.
4166 * This will be registered high priority CPU notifier.
4168 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
4169 unsigned long action,
4172 int cpu = (unsigned long)hcpu;
4173 struct worker_pool *pool;
4175 switch (action & ~CPU_TASKS_FROZEN) {
4176 case CPU_UP_PREPARE:
4177 for_each_cpu_worker_pool(pool, cpu) {
4178 if (pool->nr_workers)
4180 if (create_and_start_worker(pool) < 0)
4185 case CPU_DOWN_FAILED:
4187 for_each_cpu_worker_pool(pool, cpu) {
4188 mutex_lock(&pool->manager_mutex);
4189 spin_lock_irq(&pool->lock);
4191 pool->flags &= ~POOL_DISASSOCIATED;
4192 rebind_workers(pool);
4194 spin_unlock_irq(&pool->lock);
4195 mutex_unlock(&pool->manager_mutex);
4203 * Workqueues should be brought down after normal priority CPU notifiers.
4204 * This will be registered as low priority CPU notifier.
4206 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
4207 unsigned long action,
4210 int cpu = (unsigned long)hcpu;
4211 struct work_struct unbind_work;
4213 switch (action & ~CPU_TASKS_FROZEN) {
4214 case CPU_DOWN_PREPARE:
4215 /* unbinding should happen on the local CPU */
4216 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4217 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4218 flush_work(&unbind_work);
4226 struct work_for_cpu {
4227 struct work_struct work;
4233 static void work_for_cpu_fn(struct work_struct *work)
4235 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4237 wfc->ret = wfc->fn(wfc->arg);
4241 * work_on_cpu - run a function in user context on a particular cpu
4242 * @cpu: the cpu to run on
4243 * @fn: the function to run
4244 * @arg: the function arg
4246 * This will return the value @fn returns.
4247 * It is up to the caller to ensure that the cpu doesn't go offline.
4248 * The caller must not hold any locks which would prevent @fn from completing.
4250 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4252 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4254 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4255 schedule_work_on(cpu, &wfc.work);
4256 flush_work(&wfc.work);
4259 EXPORT_SYMBOL_GPL(work_on_cpu);
4260 #endif /* CONFIG_SMP */
4262 #ifdef CONFIG_FREEZER
4265 * freeze_workqueues_begin - begin freezing workqueues
4267 * Start freezing workqueues. After this function returns, all freezable
4268 * workqueues will queue new works to their delayed_works list instead of
4272 * Grabs and releases wq_mutex, pwq_lock and pool->lock's.
4274 void freeze_workqueues_begin(void)
4276 struct worker_pool *pool;
4277 struct workqueue_struct *wq;
4278 struct pool_workqueue *pwq;
4281 mutex_lock(&wq_mutex);
4283 WARN_ON_ONCE(workqueue_freezing);
4284 workqueue_freezing = true;
4287 for_each_pool(pool, pi) {
4288 spin_lock_irq(&pool->lock);
4289 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4290 pool->flags |= POOL_FREEZING;
4291 spin_unlock_irq(&pool->lock);
4294 /* suppress further executions by setting max_active to zero */
4295 spin_lock_irq(&pwq_lock);
4296 list_for_each_entry(wq, &workqueues, list) {
4297 for_each_pwq(pwq, wq)
4298 pwq_adjust_max_active(pwq);
4300 spin_unlock_irq(&pwq_lock);
4302 mutex_unlock(&wq_mutex);
4306 * freeze_workqueues_busy - are freezable workqueues still busy?
4308 * Check whether freezing is complete. This function must be called
4309 * between freeze_workqueues_begin() and thaw_workqueues().
4312 * Grabs and releases wq_mutex.
4315 * %true if some freezable workqueues are still busy. %false if freezing
4318 bool freeze_workqueues_busy(void)
4321 struct workqueue_struct *wq;
4322 struct pool_workqueue *pwq;
4324 mutex_lock(&wq_mutex);
4326 WARN_ON_ONCE(!workqueue_freezing);
4328 list_for_each_entry(wq, &workqueues, list) {
4329 if (!(wq->flags & WQ_FREEZABLE))
4332 * nr_active is monotonically decreasing. It's safe
4333 * to peek without lock.
4336 for_each_pwq(pwq, wq) {
4337 WARN_ON_ONCE(pwq->nr_active < 0);
4338 if (pwq->nr_active) {
4347 mutex_unlock(&wq_mutex);
4352 * thaw_workqueues - thaw workqueues
4354 * Thaw workqueues. Normal queueing is restored and all collected
4355 * frozen works are transferred to their respective pool worklists.
4358 * Grabs and releases wq_mutex, pwq_lock and pool->lock's.
4360 void thaw_workqueues(void)
4362 struct workqueue_struct *wq;
4363 struct pool_workqueue *pwq;
4364 struct worker_pool *pool;
4367 mutex_lock(&wq_mutex);
4369 if (!workqueue_freezing)
4372 /* clear FREEZING */
4373 for_each_pool(pool, pi) {
4374 spin_lock_irq(&pool->lock);
4375 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4376 pool->flags &= ~POOL_FREEZING;
4377 spin_unlock_irq(&pool->lock);
4380 /* restore max_active and repopulate worklist */
4381 spin_lock_irq(&pwq_lock);
4382 list_for_each_entry(wq, &workqueues, list) {
4383 for_each_pwq(pwq, wq)
4384 pwq_adjust_max_active(pwq);
4386 spin_unlock_irq(&pwq_lock);
4389 for_each_pool(pool, pi) {
4390 spin_lock_irq(&pool->lock);
4391 wake_up_worker(pool);
4392 spin_unlock_irq(&pool->lock);
4395 workqueue_freezing = false;
4397 mutex_unlock(&wq_mutex);
4399 #endif /* CONFIG_FREEZER */
4401 static int __init init_workqueues(void)
4403 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4406 /* make sure we have enough bits for OFFQ pool ID */
4407 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4408 WORK_CPU_END * NR_STD_WORKER_POOLS);
4410 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4412 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4414 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4415 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4417 /* initialize CPU pools */
4418 for_each_possible_cpu(cpu) {
4419 struct worker_pool *pool;
4422 for_each_cpu_worker_pool(pool, cpu) {
4423 BUG_ON(init_worker_pool(pool));
4425 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4426 pool->attrs->nice = std_nice[i++];
4429 mutex_lock(&wq_mutex);
4430 BUG_ON(worker_pool_assign_id(pool));
4431 mutex_unlock(&wq_mutex);
4435 /* create the initial worker */
4436 for_each_online_cpu(cpu) {
4437 struct worker_pool *pool;
4439 for_each_cpu_worker_pool(pool, cpu) {
4440 pool->flags &= ~POOL_DISASSOCIATED;
4441 BUG_ON(create_and_start_worker(pool) < 0);
4445 /* create default unbound wq attrs */
4446 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4447 struct workqueue_attrs *attrs;
4449 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4451 attrs->nice = std_nice[i];
4452 cpumask_setall(attrs->cpumask);
4454 unbound_std_wq_attrs[i] = attrs;
4457 system_wq = alloc_workqueue("events", 0, 0);
4458 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4459 system_long_wq = alloc_workqueue("events_long", 0, 0);
4460 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4461 WQ_UNBOUND_MAX_ACTIVE);
4462 system_freezable_wq = alloc_workqueue("events_freezable",
4464 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4465 !system_unbound_wq || !system_freezable_wq);
4468 early_initcall(init_workqueues);