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 */
78 WORKER_REBOUND = 1 << 8, /* worker was rebound */
80 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
81 WORKER_UNBOUND | WORKER_REBOUND,
83 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
85 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
86 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
88 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
89 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
92 /* call for help after 10ms
94 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
95 CREATE_COOLDOWN = HZ, /* time to breath after fail */
98 * Rescue workers are used only on emergencies and shared by
101 RESCUER_NICE_LEVEL = -20,
102 HIGHPRI_NICE_LEVEL = -20,
106 * Structure fields follow one of the following exclusion rules.
108 * I: Modifiable by initialization/destruction paths and read-only for
111 * P: Preemption protected. Disabling preemption is enough and should
112 * only be modified and accessed from the local cpu.
114 * L: pool->lock protected. Access with pool->lock held.
116 * X: During normal operation, modification requires pool->lock and should
117 * be done only from local cpu. Either disabling preemption on local
118 * cpu or grabbing pool->lock is enough for read access. If
119 * POOL_DISASSOCIATED is set, it's identical to L.
121 * MG: pool->manager_mutex and pool->lock protected. Writes require both
122 * locks. Reads can happen under either lock.
124 * PL: wq_pool_mutex protected.
126 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
128 * WQ: wq->mutex protected.
130 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
132 * MD: wq_mayday_lock protected.
135 /* struct worker is defined in workqueue_internal.h */
138 spinlock_t lock; /* the pool lock */
139 int cpu; /* I: the associated cpu */
140 int id; /* I: pool ID */
141 unsigned int flags; /* X: flags */
143 struct list_head worklist; /* L: list of pending works */
144 int nr_workers; /* L: total number of workers */
146 /* nr_idle includes the ones off idle_list for rebinding */
147 int nr_idle; /* L: currently idle ones */
149 struct list_head idle_list; /* X: list of idle workers */
150 struct timer_list idle_timer; /* L: worker idle timeout */
151 struct timer_list mayday_timer; /* L: SOS timer for workers */
153 /* a workers is either on busy_hash or idle_list, or the manager */
154 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
155 /* L: hash of busy workers */
157 /* see manage_workers() for details on the two manager mutexes */
158 struct mutex manager_arb; /* manager arbitration */
159 struct mutex manager_mutex; /* manager exclusion */
160 struct idr worker_idr; /* MG: worker IDs and iteration */
162 struct workqueue_attrs *attrs; /* I: worker attributes */
163 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
164 int refcnt; /* PL: refcnt for unbound pools */
167 * The current concurrency level. As it's likely to be accessed
168 * from other CPUs during try_to_wake_up(), put it in a separate
171 atomic_t nr_running ____cacheline_aligned_in_smp;
174 * Destruction of pool is sched-RCU protected to allow dereferences
175 * from get_work_pool().
178 } ____cacheline_aligned_in_smp;
181 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
182 * of work_struct->data are used for flags and the remaining high bits
183 * point to the pwq; thus, pwqs need to be aligned at two's power of the
184 * number of flag bits.
186 struct pool_workqueue {
187 struct worker_pool *pool; /* I: the associated pool */
188 struct workqueue_struct *wq; /* I: the owning workqueue */
189 int work_color; /* L: current color */
190 int flush_color; /* L: flushing color */
191 int refcnt; /* L: reference count */
192 int nr_in_flight[WORK_NR_COLORS];
193 /* L: nr of in_flight works */
194 int nr_active; /* L: nr of active works */
195 int max_active; /* L: max active works */
196 struct list_head delayed_works; /* L: delayed works */
197 struct list_head pwqs_node; /* WR: node on wq->pwqs */
198 struct list_head mayday_node; /* MD: node on wq->maydays */
201 * Release of unbound pwq is punted to system_wq. See put_pwq()
202 * and pwq_unbound_release_workfn() for details. pool_workqueue
203 * itself is also sched-RCU protected so that the first pwq can be
204 * determined without grabbing wq->mutex.
206 struct work_struct unbound_release_work;
208 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
211 * Structure used to wait for workqueue flush.
214 struct list_head list; /* WQ: list of flushers */
215 int flush_color; /* WQ: flush color waiting for */
216 struct completion done; /* flush completion */
222 * The externally visible workqueue. It relays the issued work items to
223 * the appropriate worker_pool through its pool_workqueues.
225 struct workqueue_struct {
226 unsigned int flags; /* WQ: WQ_* flags */
227 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */
228 struct list_head pwqs; /* WR: all pwqs of this wq */
229 struct list_head list; /* PL: list of all workqueues */
231 struct mutex mutex; /* protects this wq */
232 int work_color; /* WQ: current work color */
233 int flush_color; /* WQ: current flush color */
234 atomic_t nr_pwqs_to_flush; /* flush in progress */
235 struct wq_flusher *first_flusher; /* WQ: first flusher */
236 struct list_head flusher_queue; /* WQ: flush waiters */
237 struct list_head flusher_overflow; /* WQ: flush overflow list */
239 struct list_head maydays; /* MD: pwqs requesting rescue */
240 struct worker *rescuer; /* I: rescue worker */
242 int nr_drainers; /* WQ: drain in progress */
243 int saved_max_active; /* WQ: saved pwq max_active */
246 struct wq_device *wq_dev; /* I: for sysfs interface */
248 #ifdef CONFIG_LOCKDEP
249 struct lockdep_map lockdep_map;
251 char name[]; /* I: workqueue name */
254 static struct kmem_cache *pwq_cache;
256 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
257 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
259 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
260 static bool workqueue_freezing; /* PL: 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); /* PR: idr of all pools */
268 /* PL: 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_pool_mutex() \
293 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
294 lockdep_is_held(&wq_pool_mutex), \
295 "sched RCU or wq_pool_mutex should be held")
297 #define assert_rcu_or_wq_mutex(wq) \
298 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
299 lockdep_is_held(&wq->mutex), \
300 "sched RCU or wq->mutex should be held")
302 #ifdef CONFIG_LOCKDEP
303 #define assert_manager_or_pool_lock(pool) \
304 WARN_ONCE(debug_locks && \
305 !lockdep_is_held(&(pool)->manager_mutex) && \
306 !lockdep_is_held(&(pool)->lock), \
307 "pool->manager_mutex or ->lock should be held")
309 #define assert_manager_or_pool_lock(pool) do { } while (0)
312 #define for_each_cpu_worker_pool(pool, cpu) \
313 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
314 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
318 * for_each_pool - iterate through all worker_pools in the system
319 * @pool: iteration cursor
320 * @pi: integer used for iteration
322 * This must be called either with wq_pool_mutex held or sched RCU read
323 * locked. If the pool needs to be used beyond the locking in effect, the
324 * caller is responsible for guaranteeing that the pool stays online.
326 * The if/else clause exists only for the lockdep assertion and can be
329 #define for_each_pool(pool, pi) \
330 idr_for_each_entry(&worker_pool_idr, pool, pi) \
331 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
335 * for_each_pool_worker - iterate through all workers of a worker_pool
336 * @worker: iteration cursor
337 * @wi: integer used for iteration
338 * @pool: worker_pool to iterate workers of
340 * This must be called with either @pool->manager_mutex or ->lock held.
342 * The if/else clause exists only for the lockdep assertion and can be
345 #define for_each_pool_worker(worker, wi, pool) \
346 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
347 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
351 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
352 * @pwq: iteration cursor
353 * @wq: the target workqueue
355 * This must be called either with wq->mutex held or sched RCU read locked.
356 * If the pwq needs to be used beyond the locking in effect, the caller is
357 * responsible for guaranteeing that the pwq stays online.
359 * The if/else clause exists only for the lockdep assertion and can be
362 #define for_each_pwq(pwq, wq) \
363 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
364 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
367 #ifdef CONFIG_DEBUG_OBJECTS_WORK
369 static struct debug_obj_descr work_debug_descr;
371 static void *work_debug_hint(void *addr)
373 return ((struct work_struct *) addr)->func;
377 * fixup_init is called when:
378 * - an active object is initialized
380 static int work_fixup_init(void *addr, enum debug_obj_state state)
382 struct work_struct *work = addr;
385 case ODEBUG_STATE_ACTIVE:
386 cancel_work_sync(work);
387 debug_object_init(work, &work_debug_descr);
395 * fixup_activate is called when:
396 * - an active object is activated
397 * - an unknown object is activated (might be a statically initialized object)
399 static int work_fixup_activate(void *addr, enum debug_obj_state state)
401 struct work_struct *work = addr;
405 case ODEBUG_STATE_NOTAVAILABLE:
407 * This is not really a fixup. The work struct was
408 * statically initialized. We just make sure that it
409 * is tracked in the object tracker.
411 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
412 debug_object_init(work, &work_debug_descr);
413 debug_object_activate(work, &work_debug_descr);
419 case ODEBUG_STATE_ACTIVE:
428 * fixup_free is called when:
429 * - an active object is freed
431 static int work_fixup_free(void *addr, enum debug_obj_state state)
433 struct work_struct *work = addr;
436 case ODEBUG_STATE_ACTIVE:
437 cancel_work_sync(work);
438 debug_object_free(work, &work_debug_descr);
445 static struct debug_obj_descr work_debug_descr = {
446 .name = "work_struct",
447 .debug_hint = work_debug_hint,
448 .fixup_init = work_fixup_init,
449 .fixup_activate = work_fixup_activate,
450 .fixup_free = work_fixup_free,
453 static inline void debug_work_activate(struct work_struct *work)
455 debug_object_activate(work, &work_debug_descr);
458 static inline void debug_work_deactivate(struct work_struct *work)
460 debug_object_deactivate(work, &work_debug_descr);
463 void __init_work(struct work_struct *work, int onstack)
466 debug_object_init_on_stack(work, &work_debug_descr);
468 debug_object_init(work, &work_debug_descr);
470 EXPORT_SYMBOL_GPL(__init_work);
472 void destroy_work_on_stack(struct work_struct *work)
474 debug_object_free(work, &work_debug_descr);
476 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
479 static inline void debug_work_activate(struct work_struct *work) { }
480 static inline void debug_work_deactivate(struct work_struct *work) { }
483 /* allocate ID and assign it to @pool */
484 static int worker_pool_assign_id(struct worker_pool *pool)
488 lockdep_assert_held(&wq_pool_mutex);
491 if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL))
493 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
494 } while (ret == -EAGAIN);
500 * first_pwq - return the first pool_workqueue of the specified workqueue
501 * @wq: the target workqueue
503 * This must be called either with wq->mutex held or sched RCU read locked.
504 * If the pwq needs to be used beyond the locking in effect, the caller is
505 * responsible for guaranteeing that the pwq stays online.
507 static struct pool_workqueue *first_pwq(struct workqueue_struct *wq)
509 assert_rcu_or_wq_mutex(wq);
510 return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue,
514 static unsigned int work_color_to_flags(int color)
516 return color << WORK_STRUCT_COLOR_SHIFT;
519 static int get_work_color(struct work_struct *work)
521 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
522 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
525 static int work_next_color(int color)
527 return (color + 1) % WORK_NR_COLORS;
531 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
532 * contain the pointer to the queued pwq. Once execution starts, the flag
533 * is cleared and the high bits contain OFFQ flags and pool ID.
535 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
536 * and clear_work_data() can be used to set the pwq, pool or clear
537 * work->data. These functions should only be called while the work is
538 * owned - ie. while the PENDING bit is set.
540 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
541 * corresponding to a work. Pool is available once the work has been
542 * queued anywhere after initialization until it is sync canceled. pwq is
543 * available only while the work item is queued.
545 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
546 * canceled. While being canceled, a work item may have its PENDING set
547 * but stay off timer and worklist for arbitrarily long and nobody should
548 * try to steal the PENDING bit.
550 static inline void set_work_data(struct work_struct *work, unsigned long data,
553 WARN_ON_ONCE(!work_pending(work));
554 atomic_long_set(&work->data, data | flags | work_static(work));
557 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
558 unsigned long extra_flags)
560 set_work_data(work, (unsigned long)pwq,
561 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
564 static void set_work_pool_and_keep_pending(struct work_struct *work,
567 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
568 WORK_STRUCT_PENDING);
571 static void set_work_pool_and_clear_pending(struct work_struct *work,
575 * The following wmb is paired with the implied mb in
576 * test_and_set_bit(PENDING) and ensures all updates to @work made
577 * here are visible to and precede any updates by the next PENDING
581 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
584 static void clear_work_data(struct work_struct *work)
586 smp_wmb(); /* see set_work_pool_and_clear_pending() */
587 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
590 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
592 unsigned long data = atomic_long_read(&work->data);
594 if (data & WORK_STRUCT_PWQ)
595 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
601 * get_work_pool - return the worker_pool a given work was associated with
602 * @work: the work item of interest
604 * Return the worker_pool @work was last associated with. %NULL if none.
606 * Pools are created and destroyed under wq_pool_mutex, and allows read
607 * access under sched-RCU read lock. As such, this function should be
608 * called under wq_pool_mutex or with preemption disabled.
610 * All fields of the returned pool are accessible as long as the above
611 * mentioned locking is in effect. If the returned pool needs to be used
612 * beyond the critical section, the caller is responsible for ensuring the
613 * returned pool is and stays online.
615 static struct worker_pool *get_work_pool(struct work_struct *work)
617 unsigned long data = atomic_long_read(&work->data);
620 assert_rcu_or_pool_mutex();
622 if (data & WORK_STRUCT_PWQ)
623 return ((struct pool_workqueue *)
624 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
626 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
627 if (pool_id == WORK_OFFQ_POOL_NONE)
630 return idr_find(&worker_pool_idr, pool_id);
634 * get_work_pool_id - return the worker pool ID a given work is associated with
635 * @work: the work item of interest
637 * Return the worker_pool ID @work was last associated with.
638 * %WORK_OFFQ_POOL_NONE if none.
640 static int get_work_pool_id(struct work_struct *work)
642 unsigned long data = atomic_long_read(&work->data);
644 if (data & WORK_STRUCT_PWQ)
645 return ((struct pool_workqueue *)
646 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
648 return data >> WORK_OFFQ_POOL_SHIFT;
651 static void mark_work_canceling(struct work_struct *work)
653 unsigned long pool_id = get_work_pool_id(work);
655 pool_id <<= WORK_OFFQ_POOL_SHIFT;
656 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
659 static bool work_is_canceling(struct work_struct *work)
661 unsigned long data = atomic_long_read(&work->data);
663 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
667 * Policy functions. These define the policies on how the global worker
668 * pools are managed. Unless noted otherwise, these functions assume that
669 * they're being called with pool->lock held.
672 static bool __need_more_worker(struct worker_pool *pool)
674 return !atomic_read(&pool->nr_running);
678 * Need to wake up a worker? Called from anything but currently
681 * Note that, because unbound workers never contribute to nr_running, this
682 * function will always return %true for unbound pools as long as the
683 * worklist isn't empty.
685 static bool need_more_worker(struct worker_pool *pool)
687 return !list_empty(&pool->worklist) && __need_more_worker(pool);
690 /* Can I start working? Called from busy but !running workers. */
691 static bool may_start_working(struct worker_pool *pool)
693 return pool->nr_idle;
696 /* Do I need to keep working? Called from currently running workers. */
697 static bool keep_working(struct worker_pool *pool)
699 return !list_empty(&pool->worklist) &&
700 atomic_read(&pool->nr_running) <= 1;
703 /* Do we need a new worker? Called from manager. */
704 static bool need_to_create_worker(struct worker_pool *pool)
706 return need_more_worker(pool) && !may_start_working(pool);
709 /* Do I need to be the manager? */
710 static bool need_to_manage_workers(struct worker_pool *pool)
712 return need_to_create_worker(pool) ||
713 (pool->flags & POOL_MANAGE_WORKERS);
716 /* Do we have too many workers and should some go away? */
717 static bool too_many_workers(struct worker_pool *pool)
719 bool managing = mutex_is_locked(&pool->manager_arb);
720 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
721 int nr_busy = pool->nr_workers - nr_idle;
724 * nr_idle and idle_list may disagree if idle rebinding is in
725 * progress. Never return %true if idle_list is empty.
727 if (list_empty(&pool->idle_list))
730 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
737 /* Return the first worker. Safe with preemption disabled */
738 static struct worker *first_worker(struct worker_pool *pool)
740 if (unlikely(list_empty(&pool->idle_list)))
743 return list_first_entry(&pool->idle_list, struct worker, entry);
747 * wake_up_worker - wake up an idle worker
748 * @pool: worker pool to wake worker from
750 * Wake up the first idle worker of @pool.
753 * spin_lock_irq(pool->lock).
755 static void wake_up_worker(struct worker_pool *pool)
757 struct worker *worker = first_worker(pool);
760 wake_up_process(worker->task);
764 * wq_worker_waking_up - a worker is waking up
765 * @task: task waking up
766 * @cpu: CPU @task is waking up to
768 * This function is called during try_to_wake_up() when a worker is
772 * spin_lock_irq(rq->lock)
774 void wq_worker_waking_up(struct task_struct *task, int cpu)
776 struct worker *worker = kthread_data(task);
778 if (!(worker->flags & WORKER_NOT_RUNNING)) {
779 WARN_ON_ONCE(worker->pool->cpu != cpu);
780 atomic_inc(&worker->pool->nr_running);
785 * wq_worker_sleeping - a worker is going to sleep
786 * @task: task going to sleep
787 * @cpu: CPU in question, must be the current CPU number
789 * This function is called during schedule() when a busy worker is
790 * going to sleep. Worker on the same cpu can be woken up by
791 * returning pointer to its task.
794 * spin_lock_irq(rq->lock)
797 * Worker task on @cpu to wake up, %NULL if none.
799 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
801 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
802 struct worker_pool *pool;
805 * Rescuers, which may not have all the fields set up like normal
806 * workers, also reach here, let's not access anything before
807 * checking NOT_RUNNING.
809 if (worker->flags & WORKER_NOT_RUNNING)
814 /* this can only happen on the local cpu */
815 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
819 * The counterpart of the following dec_and_test, implied mb,
820 * worklist not empty test sequence is in insert_work().
821 * Please read comment there.
823 * NOT_RUNNING is clear. This means that we're bound to and
824 * running on the local cpu w/ rq lock held and preemption
825 * disabled, which in turn means that none else could be
826 * manipulating idle_list, so dereferencing idle_list without pool
829 if (atomic_dec_and_test(&pool->nr_running) &&
830 !list_empty(&pool->worklist))
831 to_wakeup = first_worker(pool);
832 return to_wakeup ? to_wakeup->task : NULL;
836 * worker_set_flags - set worker flags and adjust nr_running accordingly
838 * @flags: flags to set
839 * @wakeup: wakeup an idle worker if necessary
841 * Set @flags in @worker->flags and adjust nr_running accordingly. If
842 * nr_running becomes zero and @wakeup is %true, an idle worker is
846 * spin_lock_irq(pool->lock)
848 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
851 struct worker_pool *pool = worker->pool;
853 WARN_ON_ONCE(worker->task != current);
856 * If transitioning into NOT_RUNNING, adjust nr_running and
857 * wake up an idle worker as necessary if requested by
860 if ((flags & WORKER_NOT_RUNNING) &&
861 !(worker->flags & WORKER_NOT_RUNNING)) {
863 if (atomic_dec_and_test(&pool->nr_running) &&
864 !list_empty(&pool->worklist))
865 wake_up_worker(pool);
867 atomic_dec(&pool->nr_running);
870 worker->flags |= flags;
874 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
876 * @flags: flags to clear
878 * Clear @flags in @worker->flags and adjust nr_running accordingly.
881 * spin_lock_irq(pool->lock)
883 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
885 struct worker_pool *pool = worker->pool;
886 unsigned int oflags = worker->flags;
888 WARN_ON_ONCE(worker->task != current);
890 worker->flags &= ~flags;
893 * If transitioning out of NOT_RUNNING, increment nr_running. Note
894 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
895 * of multiple flags, not a single flag.
897 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
898 if (!(worker->flags & WORKER_NOT_RUNNING))
899 atomic_inc(&pool->nr_running);
903 * find_worker_executing_work - find worker which is executing a work
904 * @pool: pool of interest
905 * @work: work to find worker for
907 * Find a worker which is executing @work on @pool by searching
908 * @pool->busy_hash which is keyed by the address of @work. For a worker
909 * to match, its current execution should match the address of @work and
910 * its work function. This is to avoid unwanted dependency between
911 * unrelated work executions through a work item being recycled while still
914 * This is a bit tricky. A work item may be freed once its execution
915 * starts and nothing prevents the freed area from being recycled for
916 * another work item. If the same work item address ends up being reused
917 * before the original execution finishes, workqueue will identify the
918 * recycled work item as currently executing and make it wait until the
919 * current execution finishes, introducing an unwanted dependency.
921 * This function checks the work item address and work function to avoid
922 * false positives. Note that this isn't complete as one may construct a
923 * work function which can introduce dependency onto itself through a
924 * recycled work item. Well, if somebody wants to shoot oneself in the
925 * foot that badly, there's only so much we can do, and if such deadlock
926 * actually occurs, it should be easy to locate the culprit work function.
929 * spin_lock_irq(pool->lock).
932 * Pointer to worker which is executing @work if found, NULL
935 static struct worker *find_worker_executing_work(struct worker_pool *pool,
936 struct work_struct *work)
938 struct worker *worker;
940 hash_for_each_possible(pool->busy_hash, worker, hentry,
942 if (worker->current_work == work &&
943 worker->current_func == work->func)
950 * move_linked_works - move linked works to a list
951 * @work: start of series of works to be scheduled
952 * @head: target list to append @work to
953 * @nextp: out paramter for nested worklist walking
955 * Schedule linked works starting from @work to @head. Work series to
956 * be scheduled starts at @work and includes any consecutive work with
957 * WORK_STRUCT_LINKED set in its predecessor.
959 * If @nextp is not NULL, it's updated to point to the next work of
960 * the last scheduled work. This allows move_linked_works() to be
961 * nested inside outer list_for_each_entry_safe().
964 * spin_lock_irq(pool->lock).
966 static void move_linked_works(struct work_struct *work, struct list_head *head,
967 struct work_struct **nextp)
969 struct work_struct *n;
972 * Linked worklist will always end before the end of the list,
973 * use NULL for list head.
975 list_for_each_entry_safe_from(work, n, NULL, entry) {
976 list_move_tail(&work->entry, head);
977 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
982 * If we're already inside safe list traversal and have moved
983 * multiple works to the scheduled queue, the next position
984 * needs to be updated.
991 * get_pwq - get an extra reference on the specified pool_workqueue
992 * @pwq: pool_workqueue to get
994 * Obtain an extra reference on @pwq. The caller should guarantee that
995 * @pwq has positive refcnt and be holding the matching pool->lock.
997 static void get_pwq(struct pool_workqueue *pwq)
999 lockdep_assert_held(&pwq->pool->lock);
1000 WARN_ON_ONCE(pwq->refcnt <= 0);
1005 * put_pwq - put a pool_workqueue reference
1006 * @pwq: pool_workqueue to put
1008 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1009 * destruction. The caller should be holding the matching pool->lock.
1011 static void put_pwq(struct pool_workqueue *pwq)
1013 lockdep_assert_held(&pwq->pool->lock);
1014 if (likely(--pwq->refcnt))
1016 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1019 * @pwq can't be released under pool->lock, bounce to
1020 * pwq_unbound_release_workfn(). This never recurses on the same
1021 * pool->lock as this path is taken only for unbound workqueues and
1022 * the release work item is scheduled on a per-cpu workqueue. To
1023 * avoid lockdep warning, unbound pool->locks are given lockdep
1024 * subclass of 1 in get_unbound_pool().
1026 schedule_work(&pwq->unbound_release_work);
1029 static void pwq_activate_delayed_work(struct work_struct *work)
1031 struct pool_workqueue *pwq = get_work_pwq(work);
1033 trace_workqueue_activate_work(work);
1034 move_linked_works(work, &pwq->pool->worklist, NULL);
1035 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1039 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1041 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1042 struct work_struct, entry);
1044 pwq_activate_delayed_work(work);
1048 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1049 * @pwq: pwq of interest
1050 * @color: color of work which left the queue
1052 * A work either has completed or is removed from pending queue,
1053 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1056 * spin_lock_irq(pool->lock).
1058 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1060 /* uncolored work items don't participate in flushing or nr_active */
1061 if (color == WORK_NO_COLOR)
1064 pwq->nr_in_flight[color]--;
1067 if (!list_empty(&pwq->delayed_works)) {
1068 /* one down, submit a delayed one */
1069 if (pwq->nr_active < pwq->max_active)
1070 pwq_activate_first_delayed(pwq);
1073 /* is flush in progress and are we at the flushing tip? */
1074 if (likely(pwq->flush_color != color))
1077 /* are there still in-flight works? */
1078 if (pwq->nr_in_flight[color])
1081 /* this pwq is done, clear flush_color */
1082 pwq->flush_color = -1;
1085 * If this was the last pwq, wake up the first flusher. It
1086 * will handle the rest.
1088 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1089 complete(&pwq->wq->first_flusher->done);
1095 * try_to_grab_pending - steal work item from worklist and disable irq
1096 * @work: work item to steal
1097 * @is_dwork: @work is a delayed_work
1098 * @flags: place to store irq state
1100 * Try to grab PENDING bit of @work. This function can handle @work in any
1101 * stable state - idle, on timer or on worklist. Return values are
1103 * 1 if @work was pending and we successfully stole PENDING
1104 * 0 if @work was idle and we claimed PENDING
1105 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1106 * -ENOENT if someone else is canceling @work, this state may persist
1107 * for arbitrarily long
1109 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1110 * interrupted while holding PENDING and @work off queue, irq must be
1111 * disabled on entry. This, combined with delayed_work->timer being
1112 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1114 * On successful return, >= 0, irq is disabled and the caller is
1115 * responsible for releasing it using local_irq_restore(*@flags).
1117 * This function is safe to call from any context including IRQ handler.
1119 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1120 unsigned long *flags)
1122 struct worker_pool *pool;
1123 struct pool_workqueue *pwq;
1125 local_irq_save(*flags);
1127 /* try to steal the timer if it exists */
1129 struct delayed_work *dwork = to_delayed_work(work);
1132 * dwork->timer is irqsafe. If del_timer() fails, it's
1133 * guaranteed that the timer is not queued anywhere and not
1134 * running on the local CPU.
1136 if (likely(del_timer(&dwork->timer)))
1140 /* try to claim PENDING the normal way */
1141 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1145 * The queueing is in progress, or it is already queued. Try to
1146 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1148 pool = get_work_pool(work);
1152 spin_lock(&pool->lock);
1154 * work->data is guaranteed to point to pwq only while the work
1155 * item is queued on pwq->wq, and both updating work->data to point
1156 * to pwq on queueing and to pool on dequeueing are done under
1157 * pwq->pool->lock. This in turn guarantees that, if work->data
1158 * points to pwq which is associated with a locked pool, the work
1159 * item is currently queued on that pool.
1161 pwq = get_work_pwq(work);
1162 if (pwq && pwq->pool == pool) {
1163 debug_work_deactivate(work);
1166 * A delayed work item cannot be grabbed directly because
1167 * it might have linked NO_COLOR work items which, if left
1168 * on the delayed_list, will confuse pwq->nr_active
1169 * management later on and cause stall. Make sure the work
1170 * item is activated before grabbing.
1172 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1173 pwq_activate_delayed_work(work);
1175 list_del_init(&work->entry);
1176 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1178 /* work->data points to pwq iff queued, point to pool */
1179 set_work_pool_and_keep_pending(work, pool->id);
1181 spin_unlock(&pool->lock);
1184 spin_unlock(&pool->lock);
1186 local_irq_restore(*flags);
1187 if (work_is_canceling(work))
1194 * insert_work - insert a work into a pool
1195 * @pwq: pwq @work belongs to
1196 * @work: work to insert
1197 * @head: insertion point
1198 * @extra_flags: extra WORK_STRUCT_* flags to set
1200 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1201 * work_struct flags.
1204 * spin_lock_irq(pool->lock).
1206 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1207 struct list_head *head, unsigned int extra_flags)
1209 struct worker_pool *pool = pwq->pool;
1211 /* we own @work, set data and link */
1212 set_work_pwq(work, pwq, extra_flags);
1213 list_add_tail(&work->entry, head);
1217 * Ensure either wq_worker_sleeping() sees the above
1218 * list_add_tail() or we see zero nr_running to avoid workers lying
1219 * around lazily while there are works to be processed.
1223 if (__need_more_worker(pool))
1224 wake_up_worker(pool);
1228 * Test whether @work is being queued from another work executing on the
1231 static bool is_chained_work(struct workqueue_struct *wq)
1233 struct worker *worker;
1235 worker = current_wq_worker();
1237 * Return %true iff I'm a worker execuing a work item on @wq. If
1238 * I'm @worker, it's safe to dereference it without locking.
1240 return worker && worker->current_pwq->wq == wq;
1243 static void __queue_work(int cpu, struct workqueue_struct *wq,
1244 struct work_struct *work)
1246 struct pool_workqueue *pwq;
1247 struct worker_pool *last_pool;
1248 struct list_head *worklist;
1249 unsigned int work_flags;
1250 unsigned int req_cpu = cpu;
1253 * While a work item is PENDING && off queue, a task trying to
1254 * steal the PENDING will busy-loop waiting for it to either get
1255 * queued or lose PENDING. Grabbing PENDING and queueing should
1256 * happen with IRQ disabled.
1258 WARN_ON_ONCE(!irqs_disabled());
1260 debug_work_activate(work);
1262 /* if dying, only works from the same workqueue are allowed */
1263 if (unlikely(wq->flags & __WQ_DRAINING) &&
1264 WARN_ON_ONCE(!is_chained_work(wq)))
1267 /* pwq which will be used unless @work is executing elsewhere */
1268 if (!(wq->flags & WQ_UNBOUND)) {
1269 if (cpu == WORK_CPU_UNBOUND)
1270 cpu = raw_smp_processor_id();
1271 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1273 pwq = first_pwq(wq);
1277 * If @work was previously on a different pool, it might still be
1278 * running there, in which case the work needs to be queued on that
1279 * pool to guarantee non-reentrancy.
1281 last_pool = get_work_pool(work);
1282 if (last_pool && last_pool != pwq->pool) {
1283 struct worker *worker;
1285 spin_lock(&last_pool->lock);
1287 worker = find_worker_executing_work(last_pool, work);
1289 if (worker && worker->current_pwq->wq == wq) {
1290 pwq = worker->current_pwq;
1292 /* meh... not running there, queue here */
1293 spin_unlock(&last_pool->lock);
1294 spin_lock(&pwq->pool->lock);
1297 spin_lock(&pwq->pool->lock);
1301 * pwq is determined and locked. For unbound pools, we could have
1302 * raced with pwq release and it could already be dead. If its
1303 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1304 * without another pwq replacing it as the first pwq or while a
1305 * work item is executing on it, so the retying is guaranteed to
1306 * make forward-progress.
1308 if (unlikely(!pwq->refcnt)) {
1309 if (wq->flags & WQ_UNBOUND) {
1310 spin_unlock(&pwq->pool->lock);
1315 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1319 /* pwq determined, queue */
1320 trace_workqueue_queue_work(req_cpu, pwq, work);
1322 if (WARN_ON(!list_empty(&work->entry))) {
1323 spin_unlock(&pwq->pool->lock);
1327 pwq->nr_in_flight[pwq->work_color]++;
1328 work_flags = work_color_to_flags(pwq->work_color);
1330 if (likely(pwq->nr_active < pwq->max_active)) {
1331 trace_workqueue_activate_work(work);
1333 worklist = &pwq->pool->worklist;
1335 work_flags |= WORK_STRUCT_DELAYED;
1336 worklist = &pwq->delayed_works;
1339 insert_work(pwq, work, worklist, work_flags);
1341 spin_unlock(&pwq->pool->lock);
1345 * queue_work_on - queue work on specific cpu
1346 * @cpu: CPU number to execute work on
1347 * @wq: workqueue to use
1348 * @work: work to queue
1350 * Returns %false if @work was already on a queue, %true otherwise.
1352 * We queue the work to a specific CPU, the caller must ensure it
1355 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1356 struct work_struct *work)
1359 unsigned long flags;
1361 local_irq_save(flags);
1363 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1364 __queue_work(cpu, wq, work);
1368 local_irq_restore(flags);
1371 EXPORT_SYMBOL_GPL(queue_work_on);
1373 void delayed_work_timer_fn(unsigned long __data)
1375 struct delayed_work *dwork = (struct delayed_work *)__data;
1377 /* should have been called from irqsafe timer with irq already off */
1378 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1380 EXPORT_SYMBOL(delayed_work_timer_fn);
1382 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1383 struct delayed_work *dwork, unsigned long delay)
1385 struct timer_list *timer = &dwork->timer;
1386 struct work_struct *work = &dwork->work;
1388 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1389 timer->data != (unsigned long)dwork);
1390 WARN_ON_ONCE(timer_pending(timer));
1391 WARN_ON_ONCE(!list_empty(&work->entry));
1394 * If @delay is 0, queue @dwork->work immediately. This is for
1395 * both optimization and correctness. The earliest @timer can
1396 * expire is on the closest next tick and delayed_work users depend
1397 * on that there's no such delay when @delay is 0.
1400 __queue_work(cpu, wq, &dwork->work);
1404 timer_stats_timer_set_start_info(&dwork->timer);
1408 timer->expires = jiffies + delay;
1410 if (unlikely(cpu != WORK_CPU_UNBOUND))
1411 add_timer_on(timer, cpu);
1417 * queue_delayed_work_on - queue work on specific CPU after delay
1418 * @cpu: CPU number to execute work on
1419 * @wq: workqueue to use
1420 * @dwork: work to queue
1421 * @delay: number of jiffies to wait before queueing
1423 * Returns %false if @work was already on a queue, %true otherwise. If
1424 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1427 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1428 struct delayed_work *dwork, unsigned long delay)
1430 struct work_struct *work = &dwork->work;
1432 unsigned long flags;
1434 /* read the comment in __queue_work() */
1435 local_irq_save(flags);
1437 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1438 __queue_delayed_work(cpu, wq, dwork, delay);
1442 local_irq_restore(flags);
1445 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1448 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1449 * @cpu: CPU number to execute work on
1450 * @wq: workqueue to use
1451 * @dwork: work to queue
1452 * @delay: number of jiffies to wait before queueing
1454 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1455 * modify @dwork's timer so that it expires after @delay. If @delay is
1456 * zero, @work is guaranteed to be scheduled immediately regardless of its
1459 * Returns %false if @dwork was idle and queued, %true if @dwork was
1460 * pending and its timer was modified.
1462 * This function is safe to call from any context including IRQ handler.
1463 * See try_to_grab_pending() for details.
1465 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1466 struct delayed_work *dwork, unsigned long delay)
1468 unsigned long flags;
1472 ret = try_to_grab_pending(&dwork->work, true, &flags);
1473 } while (unlikely(ret == -EAGAIN));
1475 if (likely(ret >= 0)) {
1476 __queue_delayed_work(cpu, wq, dwork, delay);
1477 local_irq_restore(flags);
1480 /* -ENOENT from try_to_grab_pending() becomes %true */
1483 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1486 * worker_enter_idle - enter idle state
1487 * @worker: worker which is entering idle state
1489 * @worker is entering idle state. Update stats and idle timer if
1493 * spin_lock_irq(pool->lock).
1495 static void worker_enter_idle(struct worker *worker)
1497 struct worker_pool *pool = worker->pool;
1499 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1500 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1501 (worker->hentry.next || worker->hentry.pprev)))
1504 /* can't use worker_set_flags(), also called from start_worker() */
1505 worker->flags |= WORKER_IDLE;
1507 worker->last_active = jiffies;
1509 /* idle_list is LIFO */
1510 list_add(&worker->entry, &pool->idle_list);
1512 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1513 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1516 * Sanity check nr_running. Because wq_unbind_fn() releases
1517 * pool->lock between setting %WORKER_UNBOUND and zapping
1518 * nr_running, the warning may trigger spuriously. Check iff
1519 * unbind is not in progress.
1521 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1522 pool->nr_workers == pool->nr_idle &&
1523 atomic_read(&pool->nr_running));
1527 * worker_leave_idle - leave idle state
1528 * @worker: worker which is leaving idle state
1530 * @worker is leaving idle state. Update stats.
1533 * spin_lock_irq(pool->lock).
1535 static void worker_leave_idle(struct worker *worker)
1537 struct worker_pool *pool = worker->pool;
1539 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1541 worker_clr_flags(worker, WORKER_IDLE);
1543 list_del_init(&worker->entry);
1547 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1548 * @pool: target worker_pool
1550 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1552 * Works which are scheduled while the cpu is online must at least be
1553 * scheduled to a worker which is bound to the cpu so that if they are
1554 * flushed from cpu callbacks while cpu is going down, they are
1555 * guaranteed to execute on the cpu.
1557 * This function is to be used by unbound workers and rescuers to bind
1558 * themselves to the target cpu and may race with cpu going down or
1559 * coming online. kthread_bind() can't be used because it may put the
1560 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1561 * verbatim as it's best effort and blocking and pool may be
1562 * [dis]associated in the meantime.
1564 * This function tries set_cpus_allowed() and locks pool and verifies the
1565 * binding against %POOL_DISASSOCIATED which is set during
1566 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1567 * enters idle state or fetches works without dropping lock, it can
1568 * guarantee the scheduling requirement described in the first paragraph.
1571 * Might sleep. Called without any lock but returns with pool->lock
1575 * %true if the associated pool is online (@worker is successfully
1576 * bound), %false if offline.
1578 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1579 __acquires(&pool->lock)
1583 * The following call may fail, succeed or succeed
1584 * without actually migrating the task to the cpu if
1585 * it races with cpu hotunplug operation. Verify
1586 * against POOL_DISASSOCIATED.
1588 if (!(pool->flags & POOL_DISASSOCIATED))
1589 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1591 spin_lock_irq(&pool->lock);
1592 if (pool->flags & POOL_DISASSOCIATED)
1594 if (task_cpu(current) == pool->cpu &&
1595 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1597 spin_unlock_irq(&pool->lock);
1600 * We've raced with CPU hot[un]plug. Give it a breather
1601 * and retry migration. cond_resched() is required here;
1602 * otherwise, we might deadlock against cpu_stop trying to
1603 * bring down the CPU on non-preemptive kernel.
1610 static struct worker *alloc_worker(void)
1612 struct worker *worker;
1614 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1616 INIT_LIST_HEAD(&worker->entry);
1617 INIT_LIST_HEAD(&worker->scheduled);
1618 /* on creation a worker is in !idle && prep state */
1619 worker->flags = WORKER_PREP;
1625 * create_worker - create a new workqueue worker
1626 * @pool: pool the new worker will belong to
1628 * Create a new worker which is bound to @pool. The returned worker
1629 * can be started by calling start_worker() or destroyed using
1633 * Might sleep. Does GFP_KERNEL allocations.
1636 * Pointer to the newly created worker.
1638 static struct worker *create_worker(struct worker_pool *pool)
1640 const char *pri = pool->attrs->nice < 0 ? "H" : "";
1641 struct worker *worker = NULL;
1644 lockdep_assert_held(&pool->manager_mutex);
1647 * ID is needed to determine kthread name. Allocate ID first
1648 * without installing the pointer.
1650 idr_preload(GFP_KERNEL);
1651 spin_lock_irq(&pool->lock);
1653 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1655 spin_unlock_irq(&pool->lock);
1660 worker = alloc_worker();
1664 worker->pool = pool;
1668 worker->task = kthread_create_on_node(worker_thread,
1669 worker, cpu_to_node(pool->cpu),
1670 "kworker/%d:%d%s", pool->cpu, id, pri);
1672 worker->task = kthread_create(worker_thread, worker,
1675 if (IS_ERR(worker->task))
1679 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1680 * online CPUs. It'll be re-applied when any of the CPUs come up.
1682 set_user_nice(worker->task, pool->attrs->nice);
1683 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1685 /* prevent userland from meddling with cpumask of workqueue workers */
1686 worker->task->flags |= PF_NO_SETAFFINITY;
1689 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1690 * remains stable across this function. See the comments above the
1691 * flag definition for details.
1693 if (pool->flags & POOL_DISASSOCIATED)
1694 worker->flags |= WORKER_UNBOUND;
1696 /* successful, commit the pointer to idr */
1697 spin_lock_irq(&pool->lock);
1698 idr_replace(&pool->worker_idr, worker, worker->id);
1699 spin_unlock_irq(&pool->lock);
1705 spin_lock_irq(&pool->lock);
1706 idr_remove(&pool->worker_idr, id);
1707 spin_unlock_irq(&pool->lock);
1714 * start_worker - start a newly created worker
1715 * @worker: worker to start
1717 * Make the pool aware of @worker and start it.
1720 * spin_lock_irq(pool->lock).
1722 static void start_worker(struct worker *worker)
1724 worker->flags |= WORKER_STARTED;
1725 worker->pool->nr_workers++;
1726 worker_enter_idle(worker);
1727 wake_up_process(worker->task);
1731 * create_and_start_worker - create and start a worker for a pool
1732 * @pool: the target pool
1734 * Grab the managership of @pool and create and start a new worker for it.
1736 static int create_and_start_worker(struct worker_pool *pool)
1738 struct worker *worker;
1740 mutex_lock(&pool->manager_mutex);
1742 worker = create_worker(pool);
1744 spin_lock_irq(&pool->lock);
1745 start_worker(worker);
1746 spin_unlock_irq(&pool->lock);
1749 mutex_unlock(&pool->manager_mutex);
1751 return worker ? 0 : -ENOMEM;
1755 * destroy_worker - destroy a workqueue worker
1756 * @worker: worker to be destroyed
1758 * Destroy @worker and adjust @pool stats accordingly.
1761 * spin_lock_irq(pool->lock) which is released and regrabbed.
1763 static void destroy_worker(struct worker *worker)
1765 struct worker_pool *pool = worker->pool;
1767 lockdep_assert_held(&pool->manager_mutex);
1768 lockdep_assert_held(&pool->lock);
1770 /* sanity check frenzy */
1771 if (WARN_ON(worker->current_work) ||
1772 WARN_ON(!list_empty(&worker->scheduled)))
1775 if (worker->flags & WORKER_STARTED)
1777 if (worker->flags & WORKER_IDLE)
1780 list_del_init(&worker->entry);
1781 worker->flags |= WORKER_DIE;
1783 idr_remove(&pool->worker_idr, worker->id);
1785 spin_unlock_irq(&pool->lock);
1787 kthread_stop(worker->task);
1790 spin_lock_irq(&pool->lock);
1793 static void idle_worker_timeout(unsigned long __pool)
1795 struct worker_pool *pool = (void *)__pool;
1797 spin_lock_irq(&pool->lock);
1799 if (too_many_workers(pool)) {
1800 struct worker *worker;
1801 unsigned long expires;
1803 /* idle_list is kept in LIFO order, check the last one */
1804 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1805 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1807 if (time_before(jiffies, expires))
1808 mod_timer(&pool->idle_timer, expires);
1810 /* it's been idle for too long, wake up manager */
1811 pool->flags |= POOL_MANAGE_WORKERS;
1812 wake_up_worker(pool);
1816 spin_unlock_irq(&pool->lock);
1819 static void send_mayday(struct work_struct *work)
1821 struct pool_workqueue *pwq = get_work_pwq(work);
1822 struct workqueue_struct *wq = pwq->wq;
1824 lockdep_assert_held(&wq_mayday_lock);
1829 /* mayday mayday mayday */
1830 if (list_empty(&pwq->mayday_node)) {
1831 list_add_tail(&pwq->mayday_node, &wq->maydays);
1832 wake_up_process(wq->rescuer->task);
1836 static void pool_mayday_timeout(unsigned long __pool)
1838 struct worker_pool *pool = (void *)__pool;
1839 struct work_struct *work;
1841 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1842 spin_lock(&pool->lock);
1844 if (need_to_create_worker(pool)) {
1846 * We've been trying to create a new worker but
1847 * haven't been successful. We might be hitting an
1848 * allocation deadlock. Send distress signals to
1851 list_for_each_entry(work, &pool->worklist, entry)
1855 spin_unlock(&pool->lock);
1856 spin_unlock_irq(&wq_mayday_lock);
1858 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1862 * maybe_create_worker - create a new worker if necessary
1863 * @pool: pool to create a new worker for
1865 * Create a new worker for @pool if necessary. @pool is guaranteed to
1866 * have at least one idle worker on return from this function. If
1867 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1868 * sent to all rescuers with works scheduled on @pool to resolve
1869 * possible allocation deadlock.
1871 * On return, need_to_create_worker() is guaranteed to be %false and
1872 * may_start_working() %true.
1875 * spin_lock_irq(pool->lock) which may be released and regrabbed
1876 * multiple times. Does GFP_KERNEL allocations. Called only from
1880 * %false if no action was taken and pool->lock stayed locked, %true
1883 static bool maybe_create_worker(struct worker_pool *pool)
1884 __releases(&pool->lock)
1885 __acquires(&pool->lock)
1887 if (!need_to_create_worker(pool))
1890 spin_unlock_irq(&pool->lock);
1892 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1893 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1896 struct worker *worker;
1898 worker = create_worker(pool);
1900 del_timer_sync(&pool->mayday_timer);
1901 spin_lock_irq(&pool->lock);
1902 start_worker(worker);
1903 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1908 if (!need_to_create_worker(pool))
1911 __set_current_state(TASK_INTERRUPTIBLE);
1912 schedule_timeout(CREATE_COOLDOWN);
1914 if (!need_to_create_worker(pool))
1918 del_timer_sync(&pool->mayday_timer);
1919 spin_lock_irq(&pool->lock);
1920 if (need_to_create_worker(pool))
1926 * maybe_destroy_worker - destroy workers which have been idle for a while
1927 * @pool: pool to destroy workers for
1929 * Destroy @pool workers which have been idle for longer than
1930 * IDLE_WORKER_TIMEOUT.
1933 * spin_lock_irq(pool->lock) which may be released and regrabbed
1934 * multiple times. Called only from manager.
1937 * %false if no action was taken and pool->lock stayed locked, %true
1940 static bool maybe_destroy_workers(struct worker_pool *pool)
1944 while (too_many_workers(pool)) {
1945 struct worker *worker;
1946 unsigned long expires;
1948 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1949 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1951 if (time_before(jiffies, expires)) {
1952 mod_timer(&pool->idle_timer, expires);
1956 destroy_worker(worker);
1964 * manage_workers - manage worker pool
1967 * Assume the manager role and manage the worker pool @worker belongs
1968 * to. At any given time, there can be only zero or one manager per
1969 * pool. The exclusion is handled automatically by this function.
1971 * The caller can safely start processing works on false return. On
1972 * true return, it's guaranteed that need_to_create_worker() is false
1973 * and may_start_working() is true.
1976 * spin_lock_irq(pool->lock) which may be released and regrabbed
1977 * multiple times. Does GFP_KERNEL allocations.
1980 * spin_lock_irq(pool->lock) which may be released and regrabbed
1981 * multiple times. Does GFP_KERNEL allocations.
1983 static bool manage_workers(struct worker *worker)
1985 struct worker_pool *pool = worker->pool;
1989 * Managership is governed by two mutexes - manager_arb and
1990 * manager_mutex. manager_arb handles arbitration of manager role.
1991 * Anyone who successfully grabs manager_arb wins the arbitration
1992 * and becomes the manager. mutex_trylock() on pool->manager_arb
1993 * failure while holding pool->lock reliably indicates that someone
1994 * else is managing the pool and the worker which failed trylock
1995 * can proceed to executing work items. This means that anyone
1996 * grabbing manager_arb is responsible for actually performing
1997 * manager duties. If manager_arb is grabbed and released without
1998 * actual management, the pool may stall indefinitely.
2000 * manager_mutex is used for exclusion of actual management
2001 * operations. The holder of manager_mutex can be sure that none
2002 * of management operations, including creation and destruction of
2003 * workers, won't take place until the mutex is released. Because
2004 * manager_mutex doesn't interfere with manager role arbitration,
2005 * it is guaranteed that the pool's management, while may be
2006 * delayed, won't be disturbed by someone else grabbing
2009 if (!mutex_trylock(&pool->manager_arb))
2013 * With manager arbitration won, manager_mutex would be free in
2014 * most cases. trylock first without dropping @pool->lock.
2016 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2017 spin_unlock_irq(&pool->lock);
2018 mutex_lock(&pool->manager_mutex);
2022 pool->flags &= ~POOL_MANAGE_WORKERS;
2025 * Destroy and then create so that may_start_working() is true
2028 ret |= maybe_destroy_workers(pool);
2029 ret |= maybe_create_worker(pool);
2031 mutex_unlock(&pool->manager_mutex);
2032 mutex_unlock(&pool->manager_arb);
2037 * process_one_work - process single work
2039 * @work: work to process
2041 * Process @work. This function contains all the logics necessary to
2042 * process a single work including synchronization against and
2043 * interaction with other workers on the same cpu, queueing and
2044 * flushing. As long as context requirement is met, any worker can
2045 * call this function to process a work.
2048 * spin_lock_irq(pool->lock) which is released and regrabbed.
2050 static void process_one_work(struct worker *worker, struct work_struct *work)
2051 __releases(&pool->lock)
2052 __acquires(&pool->lock)
2054 struct pool_workqueue *pwq = get_work_pwq(work);
2055 struct worker_pool *pool = worker->pool;
2056 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2058 struct worker *collision;
2059 #ifdef CONFIG_LOCKDEP
2061 * It is permissible to free the struct work_struct from
2062 * inside the function that is called from it, this we need to
2063 * take into account for lockdep too. To avoid bogus "held
2064 * lock freed" warnings as well as problems when looking into
2065 * work->lockdep_map, make a copy and use that here.
2067 struct lockdep_map lockdep_map;
2069 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2072 * Ensure we're on the correct CPU. DISASSOCIATED test is
2073 * necessary to avoid spurious warnings from rescuers servicing the
2074 * unbound or a disassociated pool.
2076 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2077 !(pool->flags & POOL_DISASSOCIATED) &&
2078 raw_smp_processor_id() != pool->cpu);
2081 * A single work shouldn't be executed concurrently by
2082 * multiple workers on a single cpu. Check whether anyone is
2083 * already processing the work. If so, defer the work to the
2084 * currently executing one.
2086 collision = find_worker_executing_work(pool, work);
2087 if (unlikely(collision)) {
2088 move_linked_works(work, &collision->scheduled, NULL);
2092 /* claim and dequeue */
2093 debug_work_deactivate(work);
2094 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2095 worker->current_work = work;
2096 worker->current_func = work->func;
2097 worker->current_pwq = pwq;
2098 work_color = get_work_color(work);
2100 list_del_init(&work->entry);
2103 * CPU intensive works don't participate in concurrency
2104 * management. They're the scheduler's responsibility.
2106 if (unlikely(cpu_intensive))
2107 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2110 * Unbound pool isn't concurrency managed and work items should be
2111 * executed ASAP. Wake up another worker if necessary.
2113 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2114 wake_up_worker(pool);
2117 * Record the last pool and clear PENDING which should be the last
2118 * update to @work. Also, do this inside @pool->lock so that
2119 * PENDING and queued state changes happen together while IRQ is
2122 set_work_pool_and_clear_pending(work, pool->id);
2124 spin_unlock_irq(&pool->lock);
2126 lock_map_acquire_read(&pwq->wq->lockdep_map);
2127 lock_map_acquire(&lockdep_map);
2128 trace_workqueue_execute_start(work);
2129 worker->current_func(work);
2131 * While we must be careful to not use "work" after this, the trace
2132 * point will only record its address.
2134 trace_workqueue_execute_end(work);
2135 lock_map_release(&lockdep_map);
2136 lock_map_release(&pwq->wq->lockdep_map);
2138 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2139 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2140 " last function: %pf\n",
2141 current->comm, preempt_count(), task_pid_nr(current),
2142 worker->current_func);
2143 debug_show_held_locks(current);
2147 spin_lock_irq(&pool->lock);
2149 /* clear cpu intensive status */
2150 if (unlikely(cpu_intensive))
2151 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2153 /* we're done with it, release */
2154 hash_del(&worker->hentry);
2155 worker->current_work = NULL;
2156 worker->current_func = NULL;
2157 worker->current_pwq = NULL;
2158 pwq_dec_nr_in_flight(pwq, work_color);
2162 * process_scheduled_works - process scheduled works
2165 * Process all scheduled works. Please note that the scheduled list
2166 * may change while processing a work, so this function repeatedly
2167 * fetches a work from the top and executes it.
2170 * spin_lock_irq(pool->lock) which may be released and regrabbed
2173 static void process_scheduled_works(struct worker *worker)
2175 while (!list_empty(&worker->scheduled)) {
2176 struct work_struct *work = list_first_entry(&worker->scheduled,
2177 struct work_struct, entry);
2178 process_one_work(worker, work);
2183 * worker_thread - the worker thread function
2186 * The worker thread function. All workers belong to a worker_pool -
2187 * either a per-cpu one or dynamic unbound one. These workers process all
2188 * work items regardless of their specific target workqueue. The only
2189 * exception is work items which belong to workqueues with a rescuer which
2190 * will be explained in rescuer_thread().
2192 static int worker_thread(void *__worker)
2194 struct worker *worker = __worker;
2195 struct worker_pool *pool = worker->pool;
2197 /* tell the scheduler that this is a workqueue worker */
2198 worker->task->flags |= PF_WQ_WORKER;
2200 spin_lock_irq(&pool->lock);
2202 /* am I supposed to die? */
2203 if (unlikely(worker->flags & WORKER_DIE)) {
2204 spin_unlock_irq(&pool->lock);
2205 WARN_ON_ONCE(!list_empty(&worker->entry));
2206 worker->task->flags &= ~PF_WQ_WORKER;
2210 worker_leave_idle(worker);
2212 /* no more worker necessary? */
2213 if (!need_more_worker(pool))
2216 /* do we need to manage? */
2217 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2221 * ->scheduled list can only be filled while a worker is
2222 * preparing to process a work or actually processing it.
2223 * Make sure nobody diddled with it while I was sleeping.
2225 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2228 * Finish PREP stage. We're guaranteed to have at least one idle
2229 * worker or that someone else has already assumed the manager
2230 * role. This is where @worker starts participating in concurrency
2231 * management if applicable and concurrency management is restored
2232 * after being rebound. See rebind_workers() for details.
2234 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2237 struct work_struct *work =
2238 list_first_entry(&pool->worklist,
2239 struct work_struct, entry);
2241 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2242 /* optimization path, not strictly necessary */
2243 process_one_work(worker, work);
2244 if (unlikely(!list_empty(&worker->scheduled)))
2245 process_scheduled_works(worker);
2247 move_linked_works(work, &worker->scheduled, NULL);
2248 process_scheduled_works(worker);
2250 } while (keep_working(pool));
2252 worker_set_flags(worker, WORKER_PREP, false);
2254 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2258 * pool->lock is held and there's no work to process and no need to
2259 * manage, sleep. Workers are woken up only while holding
2260 * pool->lock or from local cpu, so setting the current state
2261 * before releasing pool->lock is enough to prevent losing any
2264 worker_enter_idle(worker);
2265 __set_current_state(TASK_INTERRUPTIBLE);
2266 spin_unlock_irq(&pool->lock);
2272 * rescuer_thread - the rescuer thread function
2275 * Workqueue rescuer thread function. There's one rescuer for each
2276 * workqueue which has WQ_MEM_RECLAIM set.
2278 * Regular work processing on a pool may block trying to create a new
2279 * worker which uses GFP_KERNEL allocation which has slight chance of
2280 * developing into deadlock if some works currently on the same queue
2281 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2282 * the problem rescuer solves.
2284 * When such condition is possible, the pool summons rescuers of all
2285 * workqueues which have works queued on the pool and let them process
2286 * those works so that forward progress can be guaranteed.
2288 * This should happen rarely.
2290 static int rescuer_thread(void *__rescuer)
2292 struct worker *rescuer = __rescuer;
2293 struct workqueue_struct *wq = rescuer->rescue_wq;
2294 struct list_head *scheduled = &rescuer->scheduled;
2296 set_user_nice(current, RESCUER_NICE_LEVEL);
2299 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2300 * doesn't participate in concurrency management.
2302 rescuer->task->flags |= PF_WQ_WORKER;
2304 set_current_state(TASK_INTERRUPTIBLE);
2306 if (kthread_should_stop()) {
2307 __set_current_state(TASK_RUNNING);
2308 rescuer->task->flags &= ~PF_WQ_WORKER;
2312 /* see whether any pwq is asking for help */
2313 spin_lock_irq(&wq_mayday_lock);
2315 while (!list_empty(&wq->maydays)) {
2316 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2317 struct pool_workqueue, mayday_node);
2318 struct worker_pool *pool = pwq->pool;
2319 struct work_struct *work, *n;
2321 __set_current_state(TASK_RUNNING);
2322 list_del_init(&pwq->mayday_node);
2324 spin_unlock_irq(&wq_mayday_lock);
2326 /* migrate to the target cpu if possible */
2327 worker_maybe_bind_and_lock(pool);
2328 rescuer->pool = pool;
2331 * Slurp in all works issued via this workqueue and
2334 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2335 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2336 if (get_work_pwq(work) == pwq)
2337 move_linked_works(work, scheduled, &n);
2339 process_scheduled_works(rescuer);
2342 * Leave this pool. If keep_working() is %true, notify a
2343 * regular worker; otherwise, we end up with 0 concurrency
2344 * and stalling the execution.
2346 if (keep_working(pool))
2347 wake_up_worker(pool);
2349 rescuer->pool = NULL;
2350 spin_unlock(&pool->lock);
2351 spin_lock(&wq_mayday_lock);
2354 spin_unlock_irq(&wq_mayday_lock);
2356 /* rescuers should never participate in concurrency management */
2357 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2363 struct work_struct work;
2364 struct completion done;
2367 static void wq_barrier_func(struct work_struct *work)
2369 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2370 complete(&barr->done);
2374 * insert_wq_barrier - insert a barrier work
2375 * @pwq: pwq to insert barrier into
2376 * @barr: wq_barrier to insert
2377 * @target: target work to attach @barr to
2378 * @worker: worker currently executing @target, NULL if @target is not executing
2380 * @barr is linked to @target such that @barr is completed only after
2381 * @target finishes execution. Please note that the ordering
2382 * guarantee is observed only with respect to @target and on the local
2385 * Currently, a queued barrier can't be canceled. This is because
2386 * try_to_grab_pending() can't determine whether the work to be
2387 * grabbed is at the head of the queue and thus can't clear LINKED
2388 * flag of the previous work while there must be a valid next work
2389 * after a work with LINKED flag set.
2391 * Note that when @worker is non-NULL, @target may be modified
2392 * underneath us, so we can't reliably determine pwq from @target.
2395 * spin_lock_irq(pool->lock).
2397 static void insert_wq_barrier(struct pool_workqueue *pwq,
2398 struct wq_barrier *barr,
2399 struct work_struct *target, struct worker *worker)
2401 struct list_head *head;
2402 unsigned int linked = 0;
2405 * debugobject calls are safe here even with pool->lock locked
2406 * as we know for sure that this will not trigger any of the
2407 * checks and call back into the fixup functions where we
2410 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2411 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2412 init_completion(&barr->done);
2415 * If @target is currently being executed, schedule the
2416 * barrier to the worker; otherwise, put it after @target.
2419 head = worker->scheduled.next;
2421 unsigned long *bits = work_data_bits(target);
2423 head = target->entry.next;
2424 /* there can already be other linked works, inherit and set */
2425 linked = *bits & WORK_STRUCT_LINKED;
2426 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2429 debug_work_activate(&barr->work);
2430 insert_work(pwq, &barr->work, head,
2431 work_color_to_flags(WORK_NO_COLOR) | linked);
2435 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2436 * @wq: workqueue being flushed
2437 * @flush_color: new flush color, < 0 for no-op
2438 * @work_color: new work color, < 0 for no-op
2440 * Prepare pwqs for workqueue flushing.
2442 * If @flush_color is non-negative, flush_color on all pwqs should be
2443 * -1. If no pwq has in-flight commands at the specified color, all
2444 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2445 * has in flight commands, its pwq->flush_color is set to
2446 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2447 * wakeup logic is armed and %true is returned.
2449 * The caller should have initialized @wq->first_flusher prior to
2450 * calling this function with non-negative @flush_color. If
2451 * @flush_color is negative, no flush color update is done and %false
2454 * If @work_color is non-negative, all pwqs should have the same
2455 * work_color which is previous to @work_color and all will be
2456 * advanced to @work_color.
2459 * mutex_lock(wq->mutex).
2462 * %true if @flush_color >= 0 and there's something to flush. %false
2465 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2466 int flush_color, int work_color)
2469 struct pool_workqueue *pwq;
2471 if (flush_color >= 0) {
2472 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2473 atomic_set(&wq->nr_pwqs_to_flush, 1);
2476 for_each_pwq(pwq, wq) {
2477 struct worker_pool *pool = pwq->pool;
2479 spin_lock_irq(&pool->lock);
2481 if (flush_color >= 0) {
2482 WARN_ON_ONCE(pwq->flush_color != -1);
2484 if (pwq->nr_in_flight[flush_color]) {
2485 pwq->flush_color = flush_color;
2486 atomic_inc(&wq->nr_pwqs_to_flush);
2491 if (work_color >= 0) {
2492 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2493 pwq->work_color = work_color;
2496 spin_unlock_irq(&pool->lock);
2499 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2500 complete(&wq->first_flusher->done);
2506 * flush_workqueue - ensure that any scheduled work has run to completion.
2507 * @wq: workqueue to flush
2509 * This function sleeps until all work items which were queued on entry
2510 * have finished execution, but it is not livelocked by new incoming ones.
2512 void flush_workqueue(struct workqueue_struct *wq)
2514 struct wq_flusher this_flusher = {
2515 .list = LIST_HEAD_INIT(this_flusher.list),
2517 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2521 lock_map_acquire(&wq->lockdep_map);
2522 lock_map_release(&wq->lockdep_map);
2524 mutex_lock(&wq->mutex);
2527 * Start-to-wait phase
2529 next_color = work_next_color(wq->work_color);
2531 if (next_color != wq->flush_color) {
2533 * Color space is not full. The current work_color
2534 * becomes our flush_color and work_color is advanced
2537 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2538 this_flusher.flush_color = wq->work_color;
2539 wq->work_color = next_color;
2541 if (!wq->first_flusher) {
2542 /* no flush in progress, become the first flusher */
2543 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2545 wq->first_flusher = &this_flusher;
2547 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2549 /* nothing to flush, done */
2550 wq->flush_color = next_color;
2551 wq->first_flusher = NULL;
2556 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2557 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2558 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2562 * Oops, color space is full, wait on overflow queue.
2563 * The next flush completion will assign us
2564 * flush_color and transfer to flusher_queue.
2566 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2569 mutex_unlock(&wq->mutex);
2571 wait_for_completion(&this_flusher.done);
2574 * Wake-up-and-cascade phase
2576 * First flushers are responsible for cascading flushes and
2577 * handling overflow. Non-first flushers can simply return.
2579 if (wq->first_flusher != &this_flusher)
2582 mutex_lock(&wq->mutex);
2584 /* we might have raced, check again with mutex held */
2585 if (wq->first_flusher != &this_flusher)
2588 wq->first_flusher = NULL;
2590 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2591 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2594 struct wq_flusher *next, *tmp;
2596 /* complete all the flushers sharing the current flush color */
2597 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2598 if (next->flush_color != wq->flush_color)
2600 list_del_init(&next->list);
2601 complete(&next->done);
2604 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2605 wq->flush_color != work_next_color(wq->work_color));
2607 /* this flush_color is finished, advance by one */
2608 wq->flush_color = work_next_color(wq->flush_color);
2610 /* one color has been freed, handle overflow queue */
2611 if (!list_empty(&wq->flusher_overflow)) {
2613 * Assign the same color to all overflowed
2614 * flushers, advance work_color and append to
2615 * flusher_queue. This is the start-to-wait
2616 * phase for these overflowed flushers.
2618 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2619 tmp->flush_color = wq->work_color;
2621 wq->work_color = work_next_color(wq->work_color);
2623 list_splice_tail_init(&wq->flusher_overflow,
2624 &wq->flusher_queue);
2625 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2628 if (list_empty(&wq->flusher_queue)) {
2629 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2634 * Need to flush more colors. Make the next flusher
2635 * the new first flusher and arm pwqs.
2637 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2638 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2640 list_del_init(&next->list);
2641 wq->first_flusher = next;
2643 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2647 * Meh... this color is already done, clear first
2648 * flusher and repeat cascading.
2650 wq->first_flusher = NULL;
2654 mutex_unlock(&wq->mutex);
2656 EXPORT_SYMBOL_GPL(flush_workqueue);
2659 * drain_workqueue - drain a workqueue
2660 * @wq: workqueue to drain
2662 * Wait until the workqueue becomes empty. While draining is in progress,
2663 * only chain queueing is allowed. IOW, only currently pending or running
2664 * work items on @wq can queue further work items on it. @wq is flushed
2665 * repeatedly until it becomes empty. The number of flushing is detemined
2666 * by the depth of chaining and should be relatively short. Whine if it
2669 void drain_workqueue(struct workqueue_struct *wq)
2671 unsigned int flush_cnt = 0;
2672 struct pool_workqueue *pwq;
2675 * __queue_work() needs to test whether there are drainers, is much
2676 * hotter than drain_workqueue() and already looks at @wq->flags.
2677 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2679 mutex_lock(&wq->mutex);
2680 if (!wq->nr_drainers++)
2681 wq->flags |= __WQ_DRAINING;
2682 mutex_unlock(&wq->mutex);
2684 flush_workqueue(wq);
2686 mutex_lock(&wq->mutex);
2688 for_each_pwq(pwq, wq) {
2691 spin_lock_irq(&pwq->pool->lock);
2692 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2693 spin_unlock_irq(&pwq->pool->lock);
2698 if (++flush_cnt == 10 ||
2699 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2700 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2701 wq->name, flush_cnt);
2703 mutex_unlock(&wq->mutex);
2707 if (!--wq->nr_drainers)
2708 wq->flags &= ~__WQ_DRAINING;
2709 mutex_unlock(&wq->mutex);
2711 EXPORT_SYMBOL_GPL(drain_workqueue);
2713 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2715 struct worker *worker = NULL;
2716 struct worker_pool *pool;
2717 struct pool_workqueue *pwq;
2721 local_irq_disable();
2722 pool = get_work_pool(work);
2728 spin_lock(&pool->lock);
2729 /* see the comment in try_to_grab_pending() with the same code */
2730 pwq = get_work_pwq(work);
2732 if (unlikely(pwq->pool != pool))
2735 worker = find_worker_executing_work(pool, work);
2738 pwq = worker->current_pwq;
2741 insert_wq_barrier(pwq, barr, work, worker);
2742 spin_unlock_irq(&pool->lock);
2745 * If @max_active is 1 or rescuer is in use, flushing another work
2746 * item on the same workqueue may lead to deadlock. Make sure the
2747 * flusher is not running on the same workqueue by verifying write
2750 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2751 lock_map_acquire(&pwq->wq->lockdep_map);
2753 lock_map_acquire_read(&pwq->wq->lockdep_map);
2754 lock_map_release(&pwq->wq->lockdep_map);
2758 spin_unlock_irq(&pool->lock);
2763 * flush_work - wait for a work to finish executing the last queueing instance
2764 * @work: the work to flush
2766 * Wait until @work has finished execution. @work is guaranteed to be idle
2767 * on return if it hasn't been requeued since flush started.
2770 * %true if flush_work() waited for the work to finish execution,
2771 * %false if it was already idle.
2773 bool flush_work(struct work_struct *work)
2775 struct wq_barrier barr;
2777 lock_map_acquire(&work->lockdep_map);
2778 lock_map_release(&work->lockdep_map);
2780 if (start_flush_work(work, &barr)) {
2781 wait_for_completion(&barr.done);
2782 destroy_work_on_stack(&barr.work);
2788 EXPORT_SYMBOL_GPL(flush_work);
2790 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2792 unsigned long flags;
2796 ret = try_to_grab_pending(work, is_dwork, &flags);
2798 * If someone else is canceling, wait for the same event it
2799 * would be waiting for before retrying.
2801 if (unlikely(ret == -ENOENT))
2803 } while (unlikely(ret < 0));
2805 /* tell other tasks trying to grab @work to back off */
2806 mark_work_canceling(work);
2807 local_irq_restore(flags);
2810 clear_work_data(work);
2815 * cancel_work_sync - cancel a work and wait for it to finish
2816 * @work: the work to cancel
2818 * Cancel @work and wait for its execution to finish. This function
2819 * can be used even if the work re-queues itself or migrates to
2820 * another workqueue. On return from this function, @work is
2821 * guaranteed to be not pending or executing on any CPU.
2823 * cancel_work_sync(&delayed_work->work) must not be used for
2824 * delayed_work's. Use cancel_delayed_work_sync() instead.
2826 * The caller must ensure that the workqueue on which @work was last
2827 * queued can't be destroyed before this function returns.
2830 * %true if @work was pending, %false otherwise.
2832 bool cancel_work_sync(struct work_struct *work)
2834 return __cancel_work_timer(work, false);
2836 EXPORT_SYMBOL_GPL(cancel_work_sync);
2839 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2840 * @dwork: the delayed work to flush
2842 * Delayed timer is cancelled and the pending work is queued for
2843 * immediate execution. Like flush_work(), this function only
2844 * considers the last queueing instance of @dwork.
2847 * %true if flush_work() waited for the work to finish execution,
2848 * %false if it was already idle.
2850 bool flush_delayed_work(struct delayed_work *dwork)
2852 local_irq_disable();
2853 if (del_timer_sync(&dwork->timer))
2854 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2856 return flush_work(&dwork->work);
2858 EXPORT_SYMBOL(flush_delayed_work);
2861 * cancel_delayed_work - cancel a delayed work
2862 * @dwork: delayed_work to cancel
2864 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2865 * and canceled; %false if wasn't pending. Note that the work callback
2866 * function may still be running on return, unless it returns %true and the
2867 * work doesn't re-arm itself. Explicitly flush or use
2868 * cancel_delayed_work_sync() to wait on it.
2870 * This function is safe to call from any context including IRQ handler.
2872 bool cancel_delayed_work(struct delayed_work *dwork)
2874 unsigned long flags;
2878 ret = try_to_grab_pending(&dwork->work, true, &flags);
2879 } while (unlikely(ret == -EAGAIN));
2881 if (unlikely(ret < 0))
2884 set_work_pool_and_clear_pending(&dwork->work,
2885 get_work_pool_id(&dwork->work));
2886 local_irq_restore(flags);
2889 EXPORT_SYMBOL(cancel_delayed_work);
2892 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2893 * @dwork: the delayed work cancel
2895 * This is cancel_work_sync() for delayed works.
2898 * %true if @dwork was pending, %false otherwise.
2900 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2902 return __cancel_work_timer(&dwork->work, true);
2904 EXPORT_SYMBOL(cancel_delayed_work_sync);
2907 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2908 * @func: the function to call
2910 * schedule_on_each_cpu() executes @func on each online CPU using the
2911 * system workqueue and blocks until all CPUs have completed.
2912 * schedule_on_each_cpu() is very slow.
2915 * 0 on success, -errno on failure.
2917 int schedule_on_each_cpu(work_func_t func)
2920 struct work_struct __percpu *works;
2922 works = alloc_percpu(struct work_struct);
2928 for_each_online_cpu(cpu) {
2929 struct work_struct *work = per_cpu_ptr(works, cpu);
2931 INIT_WORK(work, func);
2932 schedule_work_on(cpu, work);
2935 for_each_online_cpu(cpu)
2936 flush_work(per_cpu_ptr(works, cpu));
2944 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2946 * Forces execution of the kernel-global workqueue and blocks until its
2949 * Think twice before calling this function! It's very easy to get into
2950 * trouble if you don't take great care. Either of the following situations
2951 * will lead to deadlock:
2953 * One of the work items currently on the workqueue needs to acquire
2954 * a lock held by your code or its caller.
2956 * Your code is running in the context of a work routine.
2958 * They will be detected by lockdep when they occur, but the first might not
2959 * occur very often. It depends on what work items are on the workqueue and
2960 * what locks they need, which you have no control over.
2962 * In most situations flushing the entire workqueue is overkill; you merely
2963 * need to know that a particular work item isn't queued and isn't running.
2964 * In such cases you should use cancel_delayed_work_sync() or
2965 * cancel_work_sync() instead.
2967 void flush_scheduled_work(void)
2969 flush_workqueue(system_wq);
2971 EXPORT_SYMBOL(flush_scheduled_work);
2974 * execute_in_process_context - reliably execute the routine with user context
2975 * @fn: the function to execute
2976 * @ew: guaranteed storage for the execute work structure (must
2977 * be available when the work executes)
2979 * Executes the function immediately if process context is available,
2980 * otherwise schedules the function for delayed execution.
2982 * Returns: 0 - function was executed
2983 * 1 - function was scheduled for execution
2985 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2987 if (!in_interrupt()) {
2992 INIT_WORK(&ew->work, fn);
2993 schedule_work(&ew->work);
2997 EXPORT_SYMBOL_GPL(execute_in_process_context);
3001 * Workqueues with WQ_SYSFS flag set is visible to userland via
3002 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3003 * following attributes.
3005 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3006 * max_active RW int : maximum number of in-flight work items
3008 * Unbound workqueues have the following extra attributes.
3010 * id RO int : the associated pool ID
3011 * nice RW int : nice value of the workers
3012 * cpumask RW mask : bitmask of allowed CPUs for the workers
3015 struct workqueue_struct *wq;
3019 static struct workqueue_struct *dev_to_wq(struct device *dev)
3021 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3026 static ssize_t wq_per_cpu_show(struct device *dev,
3027 struct device_attribute *attr, char *buf)
3029 struct workqueue_struct *wq = dev_to_wq(dev);
3031 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3034 static ssize_t wq_max_active_show(struct device *dev,
3035 struct device_attribute *attr, char *buf)
3037 struct workqueue_struct *wq = dev_to_wq(dev);
3039 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3042 static ssize_t wq_max_active_store(struct device *dev,
3043 struct device_attribute *attr,
3044 const char *buf, size_t count)
3046 struct workqueue_struct *wq = dev_to_wq(dev);
3049 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3052 workqueue_set_max_active(wq, val);
3056 static struct device_attribute wq_sysfs_attrs[] = {
3057 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3058 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3062 static ssize_t wq_pool_id_show(struct device *dev,
3063 struct device_attribute *attr, char *buf)
3065 struct workqueue_struct *wq = dev_to_wq(dev);
3066 struct worker_pool *pool;
3069 rcu_read_lock_sched();
3070 pool = first_pwq(wq)->pool;
3071 written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id);
3072 rcu_read_unlock_sched();
3077 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3080 struct workqueue_struct *wq = dev_to_wq(dev);
3083 rcu_read_lock_sched();
3084 written = scnprintf(buf, PAGE_SIZE, "%d\n",
3085 first_pwq(wq)->pool->attrs->nice);
3086 rcu_read_unlock_sched();
3091 /* prepare workqueue_attrs for sysfs store operations */
3092 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3094 struct workqueue_attrs *attrs;
3096 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3100 rcu_read_lock_sched();
3101 copy_workqueue_attrs(attrs, first_pwq(wq)->pool->attrs);
3102 rcu_read_unlock_sched();
3106 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3107 const char *buf, size_t count)
3109 struct workqueue_struct *wq = dev_to_wq(dev);
3110 struct workqueue_attrs *attrs;
3113 attrs = wq_sysfs_prep_attrs(wq);
3117 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3118 attrs->nice >= -20 && attrs->nice <= 19)
3119 ret = apply_workqueue_attrs(wq, attrs);
3123 free_workqueue_attrs(attrs);
3124 return ret ?: count;
3127 static ssize_t wq_cpumask_show(struct device *dev,
3128 struct device_attribute *attr, char *buf)
3130 struct workqueue_struct *wq = dev_to_wq(dev);
3133 rcu_read_lock_sched();
3134 written = cpumask_scnprintf(buf, PAGE_SIZE,
3135 first_pwq(wq)->pool->attrs->cpumask);
3136 rcu_read_unlock_sched();
3138 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3142 static ssize_t wq_cpumask_store(struct device *dev,
3143 struct device_attribute *attr,
3144 const char *buf, size_t count)
3146 struct workqueue_struct *wq = dev_to_wq(dev);
3147 struct workqueue_attrs *attrs;
3150 attrs = wq_sysfs_prep_attrs(wq);
3154 ret = cpumask_parse(buf, attrs->cpumask);
3156 ret = apply_workqueue_attrs(wq, attrs);
3158 free_workqueue_attrs(attrs);
3159 return ret ?: count;
3162 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3163 __ATTR(pool_id, 0444, wq_pool_id_show, NULL),
3164 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3165 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3169 static struct bus_type wq_subsys = {
3170 .name = "workqueue",
3171 .dev_attrs = wq_sysfs_attrs,
3174 static int __init wq_sysfs_init(void)
3176 return subsys_virtual_register(&wq_subsys, NULL);
3178 core_initcall(wq_sysfs_init);
3180 static void wq_device_release(struct device *dev)
3182 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3188 * workqueue_sysfs_register - make a workqueue visible in sysfs
3189 * @wq: the workqueue to register
3191 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3192 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3193 * which is the preferred method.
3195 * Workqueue user should use this function directly iff it wants to apply
3196 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3197 * apply_workqueue_attrs() may race against userland updating the
3200 * Returns 0 on success, -errno on failure.
3202 int workqueue_sysfs_register(struct workqueue_struct *wq)
3204 struct wq_device *wq_dev;
3208 * Adjusting max_active or creating new pwqs by applyting
3209 * attributes breaks ordering guarantee. Disallow exposing ordered
3212 if (WARN_ON(wq->flags & __WQ_ORDERED))
3215 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3220 wq_dev->dev.bus = &wq_subsys;
3221 wq_dev->dev.init_name = wq->name;
3222 wq_dev->dev.release = wq_device_release;
3225 * unbound_attrs are created separately. Suppress uevent until
3226 * everything is ready.
3228 dev_set_uevent_suppress(&wq_dev->dev, true);
3230 ret = device_register(&wq_dev->dev);
3237 if (wq->flags & WQ_UNBOUND) {
3238 struct device_attribute *attr;
3240 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3241 ret = device_create_file(&wq_dev->dev, attr);
3243 device_unregister(&wq_dev->dev);
3250 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3255 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3256 * @wq: the workqueue to unregister
3258 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3260 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3262 struct wq_device *wq_dev = wq->wq_dev;
3268 device_unregister(&wq_dev->dev);
3270 #else /* CONFIG_SYSFS */
3271 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3272 #endif /* CONFIG_SYSFS */
3275 * free_workqueue_attrs - free a workqueue_attrs
3276 * @attrs: workqueue_attrs to free
3278 * Undo alloc_workqueue_attrs().
3280 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3283 free_cpumask_var(attrs->cpumask);
3289 * alloc_workqueue_attrs - allocate a workqueue_attrs
3290 * @gfp_mask: allocation mask to use
3292 * Allocate a new workqueue_attrs, initialize with default settings and
3293 * return it. Returns NULL on failure.
3295 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3297 struct workqueue_attrs *attrs;
3299 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3302 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3305 cpumask_setall(attrs->cpumask);
3308 free_workqueue_attrs(attrs);
3312 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3313 const struct workqueue_attrs *from)
3315 to->nice = from->nice;
3316 cpumask_copy(to->cpumask, from->cpumask);
3320 * Hacky implementation of jhash of bitmaps which only considers the
3321 * specified number of bits. We probably want a proper implementation in
3322 * include/linux/jhash.h.
3324 static u32 jhash_bitmap(const unsigned long *bitmap, int bits, u32 hash)
3326 int nr_longs = bits / BITS_PER_LONG;
3327 int nr_leftover = bits % BITS_PER_LONG;
3328 unsigned long leftover = 0;
3331 hash = jhash(bitmap, nr_longs * sizeof(long), hash);
3333 bitmap_copy(&leftover, bitmap + nr_longs, nr_leftover);
3334 hash = jhash(&leftover, sizeof(long), hash);
3339 /* hash value of the content of @attr */
3340 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3344 hash = jhash_1word(attrs->nice, hash);
3345 hash = jhash_bitmap(cpumask_bits(attrs->cpumask), nr_cpu_ids, hash);
3349 /* content equality test */
3350 static bool wqattrs_equal(const struct workqueue_attrs *a,
3351 const struct workqueue_attrs *b)
3353 if (a->nice != b->nice)
3355 if (!cpumask_equal(a->cpumask, b->cpumask))
3361 * init_worker_pool - initialize a newly zalloc'd worker_pool
3362 * @pool: worker_pool to initialize
3364 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3365 * Returns 0 on success, -errno on failure. Even on failure, all fields
3366 * inside @pool proper are initialized and put_unbound_pool() can be called
3367 * on @pool safely to release it.
3369 static int init_worker_pool(struct worker_pool *pool)
3371 spin_lock_init(&pool->lock);
3374 pool->flags |= POOL_DISASSOCIATED;
3375 INIT_LIST_HEAD(&pool->worklist);
3376 INIT_LIST_HEAD(&pool->idle_list);
3377 hash_init(pool->busy_hash);
3379 init_timer_deferrable(&pool->idle_timer);
3380 pool->idle_timer.function = idle_worker_timeout;
3381 pool->idle_timer.data = (unsigned long)pool;
3383 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3384 (unsigned long)pool);
3386 mutex_init(&pool->manager_arb);
3387 mutex_init(&pool->manager_mutex);
3388 idr_init(&pool->worker_idr);
3390 INIT_HLIST_NODE(&pool->hash_node);
3393 /* shouldn't fail above this point */
3394 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3400 static void rcu_free_pool(struct rcu_head *rcu)
3402 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3404 idr_destroy(&pool->worker_idr);
3405 free_workqueue_attrs(pool->attrs);
3410 * put_unbound_pool - put a worker_pool
3411 * @pool: worker_pool to put
3413 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3414 * safe manner. get_unbound_pool() calls this function on its failure path
3415 * and this function should be able to release pools which went through,
3416 * successfully or not, init_worker_pool().
3418 static void put_unbound_pool(struct worker_pool *pool)
3420 struct worker *worker;
3422 mutex_lock(&wq_pool_mutex);
3423 if (--pool->refcnt) {
3424 mutex_unlock(&wq_pool_mutex);
3429 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3430 WARN_ON(!list_empty(&pool->worklist))) {
3431 mutex_unlock(&wq_pool_mutex);
3435 /* release id and unhash */
3437 idr_remove(&worker_pool_idr, pool->id);
3438 hash_del(&pool->hash_node);
3440 mutex_unlock(&wq_pool_mutex);
3443 * Become the manager and destroy all workers. Grabbing
3444 * manager_arb prevents @pool's workers from blocking on
3447 mutex_lock(&pool->manager_arb);
3448 mutex_lock(&pool->manager_mutex);
3449 spin_lock_irq(&pool->lock);
3451 while ((worker = first_worker(pool)))
3452 destroy_worker(worker);
3453 WARN_ON(pool->nr_workers || pool->nr_idle);
3455 spin_unlock_irq(&pool->lock);
3456 mutex_unlock(&pool->manager_mutex);
3457 mutex_unlock(&pool->manager_arb);
3459 /* shut down the timers */
3460 del_timer_sync(&pool->idle_timer);
3461 del_timer_sync(&pool->mayday_timer);
3463 /* sched-RCU protected to allow dereferences from get_work_pool() */
3464 call_rcu_sched(&pool->rcu, rcu_free_pool);
3468 * get_unbound_pool - get a worker_pool with the specified attributes
3469 * @attrs: the attributes of the worker_pool to get
3471 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3472 * reference count and return it. If there already is a matching
3473 * worker_pool, it will be used; otherwise, this function attempts to
3474 * create a new one. On failure, returns NULL.
3476 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3478 u32 hash = wqattrs_hash(attrs);
3479 struct worker_pool *pool;
3481 mutex_lock(&wq_pool_mutex);
3483 /* do we already have a matching pool? */
3484 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3485 if (wqattrs_equal(pool->attrs, attrs)) {
3491 /* nope, create a new one */
3492 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3493 if (!pool || init_worker_pool(pool) < 0)
3496 if (workqueue_freezing)
3497 pool->flags |= POOL_FREEZING;
3499 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3500 copy_workqueue_attrs(pool->attrs, attrs);
3502 if (worker_pool_assign_id(pool) < 0)
3505 /* create and start the initial worker */
3506 if (create_and_start_worker(pool) < 0)
3510 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3512 mutex_unlock(&wq_pool_mutex);
3515 mutex_unlock(&wq_pool_mutex);
3517 put_unbound_pool(pool);
3521 static void rcu_free_pwq(struct rcu_head *rcu)
3523 kmem_cache_free(pwq_cache,
3524 container_of(rcu, struct pool_workqueue, rcu));
3528 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3529 * and needs to be destroyed.
3531 static void pwq_unbound_release_workfn(struct work_struct *work)
3533 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3534 unbound_release_work);
3535 struct workqueue_struct *wq = pwq->wq;
3536 struct worker_pool *pool = pwq->pool;
3539 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3543 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3544 * necessary on release but do it anyway. It's easier to verify
3545 * and consistent with the linking path.
3547 mutex_lock(&wq->mutex);
3548 list_del_rcu(&pwq->pwqs_node);
3549 is_last = list_empty(&wq->pwqs);
3550 mutex_unlock(&wq->mutex);
3552 put_unbound_pool(pool);
3553 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3556 * If we're the last pwq going away, @wq is already dead and no one
3557 * is gonna access it anymore. Free it.
3564 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3565 * @pwq: target pool_workqueue
3567 * If @pwq isn't freezing, set @pwq->max_active to the associated
3568 * workqueue's saved_max_active and activate delayed work items
3569 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3571 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3573 struct workqueue_struct *wq = pwq->wq;
3574 bool freezable = wq->flags & WQ_FREEZABLE;
3576 /* for @wq->saved_max_active */
3577 lockdep_assert_held(&wq->mutex);
3579 /* fast exit for non-freezable wqs */
3580 if (!freezable && pwq->max_active == wq->saved_max_active)
3583 spin_lock_irq(&pwq->pool->lock);
3585 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3586 pwq->max_active = wq->saved_max_active;
3588 while (!list_empty(&pwq->delayed_works) &&
3589 pwq->nr_active < pwq->max_active)
3590 pwq_activate_first_delayed(pwq);
3593 * Need to kick a worker after thawed or an unbound wq's
3594 * max_active is bumped. It's a slow path. Do it always.
3596 wake_up_worker(pwq->pool);
3598 pwq->max_active = 0;
3601 spin_unlock_irq(&pwq->pool->lock);
3604 static void init_and_link_pwq(struct pool_workqueue *pwq,
3605 struct workqueue_struct *wq,
3606 struct worker_pool *pool,
3607 struct pool_workqueue **p_last_pwq)
3609 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3613 pwq->flush_color = -1;
3615 INIT_LIST_HEAD(&pwq->delayed_works);
3616 INIT_LIST_HEAD(&pwq->mayday_node);
3617 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3619 mutex_lock(&wq->mutex);
3622 * Set the matching work_color. This is synchronized with
3623 * wq->mutex to avoid confusing flush_workqueue().
3626 *p_last_pwq = first_pwq(wq);
3627 pwq->work_color = wq->work_color;
3629 /* sync max_active to the current setting */
3630 pwq_adjust_max_active(pwq);
3633 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3635 mutex_unlock(&wq->mutex);
3639 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3640 * @wq: the target workqueue
3641 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3643 * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the
3644 * current attributes, a new pwq is created and made the first pwq which
3645 * will serve all new work items. Older pwqs are released as in-flight
3646 * work items finish. Note that a work item which repeatedly requeues
3647 * itself back-to-back will stay on its current pwq.
3649 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3652 int apply_workqueue_attrs(struct workqueue_struct *wq,
3653 const struct workqueue_attrs *attrs)
3655 struct pool_workqueue *pwq, *last_pwq;
3656 struct worker_pool *pool;
3658 /* only unbound workqueues can change attributes */
3659 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3662 /* creating multiple pwqs breaks ordering guarantee */
3663 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3666 pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL);
3670 pool = get_unbound_pool(attrs);
3672 kmem_cache_free(pwq_cache, pwq);
3676 init_and_link_pwq(pwq, wq, pool, &last_pwq);
3678 spin_lock_irq(&last_pwq->pool->lock);
3680 spin_unlock_irq(&last_pwq->pool->lock);
3686 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3688 bool highpri = wq->flags & WQ_HIGHPRI;
3691 if (!(wq->flags & WQ_UNBOUND)) {
3692 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3696 for_each_possible_cpu(cpu) {
3697 struct pool_workqueue *pwq =
3698 per_cpu_ptr(wq->cpu_pwqs, cpu);
3699 struct worker_pool *cpu_pools =
3700 per_cpu(cpu_worker_pools, cpu);
3702 init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL);
3706 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3710 static int wq_clamp_max_active(int max_active, unsigned int flags,
3713 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3715 if (max_active < 1 || max_active > lim)
3716 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3717 max_active, name, 1, lim);
3719 return clamp_val(max_active, 1, lim);
3722 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3725 struct lock_class_key *key,
3726 const char *lock_name, ...)
3728 va_list args, args1;
3729 struct workqueue_struct *wq;
3730 struct pool_workqueue *pwq;
3733 /* determine namelen, allocate wq and format name */
3734 va_start(args, lock_name);
3735 va_copy(args1, args);
3736 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3738 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3742 vsnprintf(wq->name, namelen, fmt, args1);
3746 max_active = max_active ?: WQ_DFL_ACTIVE;
3747 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3751 wq->saved_max_active = max_active;
3752 mutex_init(&wq->mutex);
3753 atomic_set(&wq->nr_pwqs_to_flush, 0);
3754 INIT_LIST_HEAD(&wq->pwqs);
3755 INIT_LIST_HEAD(&wq->flusher_queue);
3756 INIT_LIST_HEAD(&wq->flusher_overflow);
3757 INIT_LIST_HEAD(&wq->maydays);
3759 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3760 INIT_LIST_HEAD(&wq->list);
3762 if (alloc_and_link_pwqs(wq) < 0)
3766 * Workqueues which may be used during memory reclaim should
3767 * have a rescuer to guarantee forward progress.
3769 if (flags & WQ_MEM_RECLAIM) {
3770 struct worker *rescuer;
3772 rescuer = alloc_worker();
3776 rescuer->rescue_wq = wq;
3777 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3779 if (IS_ERR(rescuer->task)) {
3784 wq->rescuer = rescuer;
3785 rescuer->task->flags |= PF_NO_SETAFFINITY;
3786 wake_up_process(rescuer->task);
3789 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3793 * wq_pool_mutex protects global freeze state and workqueues list.
3794 * Grab it, adjust max_active and add the new @wq to workqueues
3797 mutex_lock(&wq_pool_mutex);
3799 mutex_lock(&wq->mutex);
3800 for_each_pwq(pwq, wq)
3801 pwq_adjust_max_active(pwq);
3802 mutex_unlock(&wq->mutex);
3804 list_add(&wq->list, &workqueues);
3806 mutex_unlock(&wq_pool_mutex);
3814 destroy_workqueue(wq);
3817 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3820 * destroy_workqueue - safely terminate a workqueue
3821 * @wq: target workqueue
3823 * Safely destroy a workqueue. All work currently pending will be done first.
3825 void destroy_workqueue(struct workqueue_struct *wq)
3827 struct pool_workqueue *pwq;
3829 /* drain it before proceeding with destruction */
3830 drain_workqueue(wq);
3833 mutex_lock(&wq->mutex);
3834 for_each_pwq(pwq, wq) {
3837 for (i = 0; i < WORK_NR_COLORS; i++) {
3838 if (WARN_ON(pwq->nr_in_flight[i])) {
3839 mutex_unlock(&wq->mutex);
3844 if (WARN_ON(pwq->refcnt > 1) ||
3845 WARN_ON(pwq->nr_active) ||
3846 WARN_ON(!list_empty(&pwq->delayed_works))) {
3847 mutex_unlock(&wq->mutex);
3851 mutex_unlock(&wq->mutex);
3854 * wq list is used to freeze wq, remove from list after
3855 * flushing is complete in case freeze races us.
3857 mutex_lock(&wq_pool_mutex);
3858 list_del_init(&wq->list);
3859 mutex_unlock(&wq_pool_mutex);
3861 workqueue_sysfs_unregister(wq);
3864 kthread_stop(wq->rescuer->task);
3869 if (!(wq->flags & WQ_UNBOUND)) {
3871 * The base ref is never dropped on per-cpu pwqs. Directly
3872 * free the pwqs and wq.
3874 free_percpu(wq->cpu_pwqs);
3878 * We're the sole accessor of @wq at this point. Directly
3879 * access the first pwq and put the base ref. As both pwqs
3880 * and pools are sched-RCU protected, the lock operations
3881 * are safe. @wq will be freed when the last pwq is
3884 pwq = list_first_entry(&wq->pwqs, struct pool_workqueue,
3886 spin_lock_irq(&pwq->pool->lock);
3888 spin_unlock_irq(&pwq->pool->lock);
3891 EXPORT_SYMBOL_GPL(destroy_workqueue);
3894 * workqueue_set_max_active - adjust max_active of a workqueue
3895 * @wq: target workqueue
3896 * @max_active: new max_active value.
3898 * Set max_active of @wq to @max_active.
3901 * Don't call from IRQ context.
3903 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3905 struct pool_workqueue *pwq;
3907 /* disallow meddling with max_active for ordered workqueues */
3908 if (WARN_ON(wq->flags & __WQ_ORDERED))
3911 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3913 mutex_lock(&wq->mutex);
3915 wq->saved_max_active = max_active;
3917 for_each_pwq(pwq, wq)
3918 pwq_adjust_max_active(pwq);
3920 mutex_unlock(&wq->mutex);
3922 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3925 * current_is_workqueue_rescuer - is %current workqueue rescuer?
3927 * Determine whether %current is a workqueue rescuer. Can be used from
3928 * work functions to determine whether it's being run off the rescuer task.
3930 bool current_is_workqueue_rescuer(void)
3932 struct worker *worker = current_wq_worker();
3934 return worker && worker->rescue_wq;
3938 * workqueue_congested - test whether a workqueue is congested
3939 * @cpu: CPU in question
3940 * @wq: target workqueue
3942 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3943 * no synchronization around this function and the test result is
3944 * unreliable and only useful as advisory hints or for debugging.
3947 * %true if congested, %false otherwise.
3949 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
3951 struct pool_workqueue *pwq;
3954 rcu_read_lock_sched();
3956 if (!(wq->flags & WQ_UNBOUND))
3957 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
3959 pwq = first_pwq(wq);
3961 ret = !list_empty(&pwq->delayed_works);
3962 rcu_read_unlock_sched();
3966 EXPORT_SYMBOL_GPL(workqueue_congested);
3969 * work_busy - test whether a work is currently pending or running
3970 * @work: the work to be tested
3972 * Test whether @work is currently pending or running. There is no
3973 * synchronization around this function and the test result is
3974 * unreliable and only useful as advisory hints or for debugging.
3977 * OR'd bitmask of WORK_BUSY_* bits.
3979 unsigned int work_busy(struct work_struct *work)
3981 struct worker_pool *pool;
3982 unsigned long flags;
3983 unsigned int ret = 0;
3985 if (work_pending(work))
3986 ret |= WORK_BUSY_PENDING;
3988 local_irq_save(flags);
3989 pool = get_work_pool(work);
3991 spin_lock(&pool->lock);
3992 if (find_worker_executing_work(pool, work))
3993 ret |= WORK_BUSY_RUNNING;
3994 spin_unlock(&pool->lock);
3996 local_irq_restore(flags);
4000 EXPORT_SYMBOL_GPL(work_busy);
4005 * There are two challenges in supporting CPU hotplug. Firstly, there
4006 * are a lot of assumptions on strong associations among work, pwq and
4007 * pool which make migrating pending and scheduled works very
4008 * difficult to implement without impacting hot paths. Secondly,
4009 * worker pools serve mix of short, long and very long running works making
4010 * blocked draining impractical.
4012 * This is solved by allowing the pools to be disassociated from the CPU
4013 * running as an unbound one and allowing it to be reattached later if the
4014 * cpu comes back online.
4017 static void wq_unbind_fn(struct work_struct *work)
4019 int cpu = smp_processor_id();
4020 struct worker_pool *pool;
4021 struct worker *worker;
4024 for_each_cpu_worker_pool(pool, cpu) {
4025 WARN_ON_ONCE(cpu != smp_processor_id());
4027 mutex_lock(&pool->manager_mutex);
4028 spin_lock_irq(&pool->lock);
4031 * We've blocked all manager operations. Make all workers
4032 * unbound and set DISASSOCIATED. Before this, all workers
4033 * except for the ones which are still executing works from
4034 * before the last CPU down must be on the cpu. After
4035 * this, they may become diasporas.
4037 for_each_pool_worker(worker, wi, pool)
4038 worker->flags |= WORKER_UNBOUND;
4040 pool->flags |= POOL_DISASSOCIATED;
4042 spin_unlock_irq(&pool->lock);
4043 mutex_unlock(&pool->manager_mutex);
4047 * Call schedule() so that we cross rq->lock and thus can guarantee
4048 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
4049 * as scheduler callbacks may be invoked from other cpus.
4054 * Sched callbacks are disabled now. Zap nr_running. After this,
4055 * nr_running stays zero and need_more_worker() and keep_working()
4056 * are always true as long as the worklist is not empty. Pools on
4057 * @cpu now behave as unbound (in terms of concurrency management)
4058 * pools which are served by workers tied to the CPU.
4060 * On return from this function, the current worker would trigger
4061 * unbound chain execution of pending work items if other workers
4064 for_each_cpu_worker_pool(pool, cpu)
4065 atomic_set(&pool->nr_running, 0);
4069 * rebind_workers - rebind all workers of a pool to the associated CPU
4070 * @pool: pool of interest
4072 * @pool->cpu is coming online. Rebind all workers to the CPU.
4074 static void rebind_workers(struct worker_pool *pool)
4076 struct worker *worker;
4079 lockdep_assert_held(&pool->manager_mutex);
4082 * Restore CPU affinity of all workers. As all idle workers should
4083 * be on the run-queue of the associated CPU before any local
4084 * wake-ups for concurrency management happen, restore CPU affinty
4085 * of all workers first and then clear UNBOUND. As we're called
4086 * from CPU_ONLINE, the following shouldn't fail.
4088 for_each_pool_worker(worker, wi, pool)
4089 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4090 pool->attrs->cpumask) < 0);
4092 spin_lock_irq(&pool->lock);
4094 for_each_pool_worker(worker, wi, pool) {
4095 unsigned int worker_flags = worker->flags;
4098 * A bound idle worker should actually be on the runqueue
4099 * of the associated CPU for local wake-ups targeting it to
4100 * work. Kick all idle workers so that they migrate to the
4101 * associated CPU. Doing this in the same loop as
4102 * replacing UNBOUND with REBOUND is safe as no worker will
4103 * be bound before @pool->lock is released.
4105 if (worker_flags & WORKER_IDLE)
4106 wake_up_process(worker->task);
4109 * We want to clear UNBOUND but can't directly call
4110 * worker_clr_flags() or adjust nr_running. Atomically
4111 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4112 * @worker will clear REBOUND using worker_clr_flags() when
4113 * it initiates the next execution cycle thus restoring
4114 * concurrency management. Note that when or whether
4115 * @worker clears REBOUND doesn't affect correctness.
4117 * ACCESS_ONCE() is necessary because @worker->flags may be
4118 * tested without holding any lock in
4119 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4120 * fail incorrectly leading to premature concurrency
4121 * management operations.
4123 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4124 worker_flags |= WORKER_REBOUND;
4125 worker_flags &= ~WORKER_UNBOUND;
4126 ACCESS_ONCE(worker->flags) = worker_flags;
4129 spin_unlock_irq(&pool->lock);
4133 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4134 * @pool: unbound pool of interest
4135 * @cpu: the CPU which is coming up
4137 * An unbound pool may end up with a cpumask which doesn't have any online
4138 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4139 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4140 * online CPU before, cpus_allowed of all its workers should be restored.
4142 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4144 static cpumask_t cpumask;
4145 struct worker *worker;
4148 lockdep_assert_held(&pool->manager_mutex);
4150 /* is @cpu allowed for @pool? */
4151 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4154 /* is @cpu the only online CPU? */
4155 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4156 if (cpumask_weight(&cpumask) != 1)
4159 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4160 for_each_pool_worker(worker, wi, pool)
4161 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4162 pool->attrs->cpumask) < 0);
4166 * Workqueues should be brought up before normal priority CPU notifiers.
4167 * This will be registered high priority CPU notifier.
4169 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
4170 unsigned long action,
4173 int cpu = (unsigned long)hcpu;
4174 struct worker_pool *pool;
4177 switch (action & ~CPU_TASKS_FROZEN) {
4178 case CPU_UP_PREPARE:
4179 for_each_cpu_worker_pool(pool, cpu) {
4180 if (pool->nr_workers)
4182 if (create_and_start_worker(pool) < 0)
4187 case CPU_DOWN_FAILED:
4189 mutex_lock(&wq_pool_mutex);
4191 for_each_pool(pool, pi) {
4192 mutex_lock(&pool->manager_mutex);
4194 if (pool->cpu == cpu) {
4195 spin_lock_irq(&pool->lock);
4196 pool->flags &= ~POOL_DISASSOCIATED;
4197 spin_unlock_irq(&pool->lock);
4199 rebind_workers(pool);
4200 } else if (pool->cpu < 0) {
4201 restore_unbound_workers_cpumask(pool, cpu);
4204 mutex_unlock(&pool->manager_mutex);
4207 mutex_unlock(&wq_pool_mutex);
4214 * Workqueues should be brought down after normal priority CPU notifiers.
4215 * This will be registered as low priority CPU notifier.
4217 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
4218 unsigned long action,
4221 int cpu = (unsigned long)hcpu;
4222 struct work_struct unbind_work;
4224 switch (action & ~CPU_TASKS_FROZEN) {
4225 case CPU_DOWN_PREPARE:
4226 /* unbinding should happen on the local CPU */
4227 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4228 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4229 flush_work(&unbind_work);
4237 struct work_for_cpu {
4238 struct work_struct work;
4244 static void work_for_cpu_fn(struct work_struct *work)
4246 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4248 wfc->ret = wfc->fn(wfc->arg);
4252 * work_on_cpu - run a function in user context on a particular cpu
4253 * @cpu: the cpu to run on
4254 * @fn: the function to run
4255 * @arg: the function arg
4257 * This will return the value @fn returns.
4258 * It is up to the caller to ensure that the cpu doesn't go offline.
4259 * The caller must not hold any locks which would prevent @fn from completing.
4261 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4263 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4265 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4266 schedule_work_on(cpu, &wfc.work);
4267 flush_work(&wfc.work);
4270 EXPORT_SYMBOL_GPL(work_on_cpu);
4271 #endif /* CONFIG_SMP */
4273 #ifdef CONFIG_FREEZER
4276 * freeze_workqueues_begin - begin freezing workqueues
4278 * Start freezing workqueues. After this function returns, all freezable
4279 * workqueues will queue new works to their delayed_works list instead of
4283 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4285 void freeze_workqueues_begin(void)
4287 struct worker_pool *pool;
4288 struct workqueue_struct *wq;
4289 struct pool_workqueue *pwq;
4292 mutex_lock(&wq_pool_mutex);
4294 WARN_ON_ONCE(workqueue_freezing);
4295 workqueue_freezing = true;
4298 for_each_pool(pool, pi) {
4299 spin_lock_irq(&pool->lock);
4300 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4301 pool->flags |= POOL_FREEZING;
4302 spin_unlock_irq(&pool->lock);
4305 list_for_each_entry(wq, &workqueues, list) {
4306 mutex_lock(&wq->mutex);
4307 for_each_pwq(pwq, wq)
4308 pwq_adjust_max_active(pwq);
4309 mutex_unlock(&wq->mutex);
4312 mutex_unlock(&wq_pool_mutex);
4316 * freeze_workqueues_busy - are freezable workqueues still busy?
4318 * Check whether freezing is complete. This function must be called
4319 * between freeze_workqueues_begin() and thaw_workqueues().
4322 * Grabs and releases wq_pool_mutex.
4325 * %true if some freezable workqueues are still busy. %false if freezing
4328 bool freeze_workqueues_busy(void)
4331 struct workqueue_struct *wq;
4332 struct pool_workqueue *pwq;
4334 mutex_lock(&wq_pool_mutex);
4336 WARN_ON_ONCE(!workqueue_freezing);
4338 list_for_each_entry(wq, &workqueues, list) {
4339 if (!(wq->flags & WQ_FREEZABLE))
4342 * nr_active is monotonically decreasing. It's safe
4343 * to peek without lock.
4345 rcu_read_lock_sched();
4346 for_each_pwq(pwq, wq) {
4347 WARN_ON_ONCE(pwq->nr_active < 0);
4348 if (pwq->nr_active) {
4350 rcu_read_unlock_sched();
4354 rcu_read_unlock_sched();
4357 mutex_unlock(&wq_pool_mutex);
4362 * thaw_workqueues - thaw workqueues
4364 * Thaw workqueues. Normal queueing is restored and all collected
4365 * frozen works are transferred to their respective pool worklists.
4368 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4370 void thaw_workqueues(void)
4372 struct workqueue_struct *wq;
4373 struct pool_workqueue *pwq;
4374 struct worker_pool *pool;
4377 mutex_lock(&wq_pool_mutex);
4379 if (!workqueue_freezing)
4382 /* clear FREEZING */
4383 for_each_pool(pool, pi) {
4384 spin_lock_irq(&pool->lock);
4385 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4386 pool->flags &= ~POOL_FREEZING;
4387 spin_unlock_irq(&pool->lock);
4390 /* restore max_active and repopulate worklist */
4391 list_for_each_entry(wq, &workqueues, list) {
4392 mutex_lock(&wq->mutex);
4393 for_each_pwq(pwq, wq)
4394 pwq_adjust_max_active(pwq);
4395 mutex_unlock(&wq->mutex);
4398 workqueue_freezing = false;
4400 mutex_unlock(&wq_pool_mutex);
4402 #endif /* CONFIG_FREEZER */
4404 static int __init init_workqueues(void)
4406 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4409 /* make sure we have enough bits for OFFQ pool ID */
4410 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4411 WORK_CPU_END * NR_STD_WORKER_POOLS);
4413 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4415 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4417 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4418 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4420 /* initialize CPU pools */
4421 for_each_possible_cpu(cpu) {
4422 struct worker_pool *pool;
4425 for_each_cpu_worker_pool(pool, cpu) {
4426 BUG_ON(init_worker_pool(pool));
4428 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4429 pool->attrs->nice = std_nice[i++];
4432 mutex_lock(&wq_pool_mutex);
4433 BUG_ON(worker_pool_assign_id(pool));
4434 mutex_unlock(&wq_pool_mutex);
4438 /* create the initial worker */
4439 for_each_online_cpu(cpu) {
4440 struct worker_pool *pool;
4442 for_each_cpu_worker_pool(pool, cpu) {
4443 pool->flags &= ~POOL_DISASSOCIATED;
4444 BUG_ON(create_and_start_worker(pool) < 0);
4448 /* create default unbound wq attrs */
4449 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4450 struct workqueue_attrs *attrs;
4452 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4454 attrs->nice = std_nice[i];
4455 cpumask_setall(attrs->cpumask);
4457 unbound_std_wq_attrs[i] = attrs;
4460 system_wq = alloc_workqueue("events", 0, 0);
4461 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4462 system_long_wq = alloc_workqueue("events_long", 0, 0);
4463 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4464 WQ_UNBOUND_MAX_ACTIVE);
4465 system_freezable_wq = alloc_workqueue("events_freezable",
4467 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4468 !system_unbound_wq || !system_freezable_wq);
4471 early_initcall(init_workqueues);