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/hashtable.h>
46 #include "workqueue_internal.h"
52 * A bound pool is either associated or disassociated with its CPU.
53 * While associated (!DISASSOCIATED), all workers are bound to the
54 * CPU and none has %WORKER_UNBOUND set and concurrency management
57 * While DISASSOCIATED, the cpu may be offline and all workers have
58 * %WORKER_UNBOUND set and concurrency management disabled, and may
59 * be executing on any CPU. The pool behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex to avoid changing binding state while
63 * create_worker() is in progress.
65 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
66 POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */
67 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
68 POOL_FREEZING = 1 << 3, /* freeze in progress */
71 WORKER_STARTED = 1 << 0, /* started */
72 WORKER_DIE = 1 << 1, /* die die die */
73 WORKER_IDLE = 1 << 2, /* is idle */
74 WORKER_PREP = 1 << 3, /* preparing to run works */
75 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
76 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
78 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND |
81 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
83 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
85 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
86 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
88 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
89 /* call for help after 10ms
91 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
92 CREATE_COOLDOWN = HZ, /* time to breath after fail */
95 * Rescue workers are used only on emergencies and shared by
98 RESCUER_NICE_LEVEL = -20,
99 HIGHPRI_NICE_LEVEL = -20,
103 * Structure fields follow one of the following exclusion rules.
105 * I: Modifiable by initialization/destruction paths and read-only for
108 * P: Preemption protected. Disabling preemption is enough and should
109 * only be modified and accessed from the local cpu.
111 * L: pool->lock protected. Access with pool->lock held.
113 * X: During normal operation, modification requires pool->lock and should
114 * be done only from local cpu. Either disabling preemption on local
115 * cpu or grabbing pool->lock is enough for read access. If
116 * POOL_DISASSOCIATED is set, it's identical to L.
118 * F: wq->flush_mutex protected.
120 * W: workqueue_lock protected.
123 /* struct worker is defined in workqueue_internal.h */
126 struct global_cwq *gcwq; /* I: the owning gcwq */
127 spinlock_t lock; /* the pool lock */
128 unsigned int cpu; /* I: the associated cpu */
129 int id; /* I: pool ID */
130 unsigned int flags; /* X: flags */
132 struct list_head worklist; /* L: list of pending works */
133 int nr_workers; /* L: total number of workers */
135 /* nr_idle includes the ones off idle_list for rebinding */
136 int nr_idle; /* L: currently idle ones */
138 struct list_head idle_list; /* X: list of idle workers */
139 struct timer_list idle_timer; /* L: worker idle timeout */
140 struct timer_list mayday_timer; /* L: SOS timer for workers */
142 /* workers are chained either in busy_hash or idle_list */
143 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
144 /* L: hash of busy workers */
146 struct mutex assoc_mutex; /* protect POOL_DISASSOCIATED */
147 struct ida worker_ida; /* L: for worker IDs */
151 * Global per-cpu workqueue. There's one and only one for each cpu
152 * and all works are queued and processed here regardless of their
156 struct worker_pool pools[NR_STD_WORKER_POOLS];
157 /* normal and highpri pools */
158 } ____cacheline_aligned_in_smp;
161 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
162 * work_struct->data are used for flags and thus cwqs need to be
163 * aligned at two's power of the number of flag bits.
165 struct cpu_workqueue_struct {
166 struct worker_pool *pool; /* I: the associated pool */
167 struct workqueue_struct *wq; /* I: the owning workqueue */
168 int work_color; /* L: current color */
169 int flush_color; /* L: flushing color */
170 int nr_in_flight[WORK_NR_COLORS];
171 /* L: nr of in_flight works */
172 int nr_active; /* L: nr of active works */
173 int max_active; /* L: max active works */
174 struct list_head delayed_works; /* L: delayed works */
178 * Structure used to wait for workqueue flush.
181 struct list_head list; /* F: list of flushers */
182 int flush_color; /* F: flush color waiting for */
183 struct completion done; /* flush completion */
187 * All cpumasks are assumed to be always set on UP and thus can't be
188 * used to determine whether there's something to be done.
191 typedef cpumask_var_t mayday_mask_t;
192 #define mayday_test_and_set_cpu(cpu, mask) \
193 cpumask_test_and_set_cpu((cpu), (mask))
194 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
195 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
196 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
197 #define free_mayday_mask(mask) free_cpumask_var((mask))
199 typedef unsigned long mayday_mask_t;
200 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
201 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
202 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
203 #define alloc_mayday_mask(maskp, gfp) true
204 #define free_mayday_mask(mask) do { } while (0)
208 * The externally visible workqueue abstraction is an array of
209 * per-CPU workqueues:
211 struct workqueue_struct {
212 unsigned int flags; /* W: WQ_* flags */
214 struct cpu_workqueue_struct __percpu *pcpu;
215 struct cpu_workqueue_struct *single;
217 } cpu_wq; /* I: cwq's */
218 struct list_head list; /* W: list of all workqueues */
220 struct mutex flush_mutex; /* protects wq flushing */
221 int work_color; /* F: current work color */
222 int flush_color; /* F: current flush color */
223 atomic_t nr_cwqs_to_flush; /* flush in progress */
224 struct wq_flusher *first_flusher; /* F: first flusher */
225 struct list_head flusher_queue; /* F: flush waiters */
226 struct list_head flusher_overflow; /* F: flush overflow list */
228 mayday_mask_t mayday_mask; /* cpus requesting rescue */
229 struct worker *rescuer; /* I: rescue worker */
231 int nr_drainers; /* W: drain in progress */
232 int saved_max_active; /* W: saved cwq max_active */
233 #ifdef CONFIG_LOCKDEP
234 struct lockdep_map lockdep_map;
236 char name[]; /* I: workqueue name */
239 struct workqueue_struct *system_wq __read_mostly;
240 EXPORT_SYMBOL_GPL(system_wq);
241 struct workqueue_struct *system_highpri_wq __read_mostly;
242 EXPORT_SYMBOL_GPL(system_highpri_wq);
243 struct workqueue_struct *system_long_wq __read_mostly;
244 EXPORT_SYMBOL_GPL(system_long_wq);
245 struct workqueue_struct *system_unbound_wq __read_mostly;
246 EXPORT_SYMBOL_GPL(system_unbound_wq);
247 struct workqueue_struct *system_freezable_wq __read_mostly;
248 EXPORT_SYMBOL_GPL(system_freezable_wq);
250 #define CREATE_TRACE_POINTS
251 #include <trace/events/workqueue.h>
253 #define for_each_worker_pool(pool, gcwq) \
254 for ((pool) = &(gcwq)->pools[0]; \
255 (pool) < &(gcwq)->pools[NR_STD_WORKER_POOLS]; (pool)++)
257 #define for_each_busy_worker(worker, i, pos, pool) \
258 hash_for_each(pool->busy_hash, i, pos, worker, hentry)
260 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
263 if (cpu < nr_cpu_ids) {
265 cpu = cpumask_next(cpu, mask);
266 if (cpu < nr_cpu_ids)
270 return WORK_CPU_UNBOUND;
272 return WORK_CPU_NONE;
275 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
276 struct workqueue_struct *wq)
278 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
284 * An extra gcwq is defined for an invalid cpu number
285 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
286 * specific CPU. The following iterators are similar to
287 * for_each_*_cpu() iterators but also considers the unbound gcwq.
289 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
290 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
291 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
292 * WORK_CPU_UNBOUND for unbound workqueues
294 #define for_each_gcwq_cpu(cpu) \
295 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
296 (cpu) < WORK_CPU_NONE; \
297 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
299 #define for_each_online_gcwq_cpu(cpu) \
300 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
301 (cpu) < WORK_CPU_NONE; \
302 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
304 #define for_each_cwq_cpu(cpu, wq) \
305 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
306 (cpu) < WORK_CPU_NONE; \
307 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
309 #ifdef CONFIG_DEBUG_OBJECTS_WORK
311 static struct debug_obj_descr work_debug_descr;
313 static void *work_debug_hint(void *addr)
315 return ((struct work_struct *) addr)->func;
319 * fixup_init is called when:
320 * - an active object is initialized
322 static int work_fixup_init(void *addr, enum debug_obj_state state)
324 struct work_struct *work = addr;
327 case ODEBUG_STATE_ACTIVE:
328 cancel_work_sync(work);
329 debug_object_init(work, &work_debug_descr);
337 * fixup_activate is called when:
338 * - an active object is activated
339 * - an unknown object is activated (might be a statically initialized object)
341 static int work_fixup_activate(void *addr, enum debug_obj_state state)
343 struct work_struct *work = addr;
347 case ODEBUG_STATE_NOTAVAILABLE:
349 * This is not really a fixup. The work struct was
350 * statically initialized. We just make sure that it
351 * is tracked in the object tracker.
353 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
354 debug_object_init(work, &work_debug_descr);
355 debug_object_activate(work, &work_debug_descr);
361 case ODEBUG_STATE_ACTIVE:
370 * fixup_free is called when:
371 * - an active object is freed
373 static int work_fixup_free(void *addr, enum debug_obj_state state)
375 struct work_struct *work = addr;
378 case ODEBUG_STATE_ACTIVE:
379 cancel_work_sync(work);
380 debug_object_free(work, &work_debug_descr);
387 static struct debug_obj_descr work_debug_descr = {
388 .name = "work_struct",
389 .debug_hint = work_debug_hint,
390 .fixup_init = work_fixup_init,
391 .fixup_activate = work_fixup_activate,
392 .fixup_free = work_fixup_free,
395 static inline void debug_work_activate(struct work_struct *work)
397 debug_object_activate(work, &work_debug_descr);
400 static inline void debug_work_deactivate(struct work_struct *work)
402 debug_object_deactivate(work, &work_debug_descr);
405 void __init_work(struct work_struct *work, int onstack)
408 debug_object_init_on_stack(work, &work_debug_descr);
410 debug_object_init(work, &work_debug_descr);
412 EXPORT_SYMBOL_GPL(__init_work);
414 void destroy_work_on_stack(struct work_struct *work)
416 debug_object_free(work, &work_debug_descr);
418 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
421 static inline void debug_work_activate(struct work_struct *work) { }
422 static inline void debug_work_deactivate(struct work_struct *work) { }
425 /* Serializes the accesses to the list of workqueues. */
426 static DEFINE_SPINLOCK(workqueue_lock);
427 static LIST_HEAD(workqueues);
428 static bool workqueue_freezing; /* W: have wqs started freezing? */
431 * The almighty global cpu workqueues. nr_running is the only field
432 * which is expected to be used frequently by other cpus via
433 * try_to_wake_up(). Put it in a separate cacheline.
435 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
436 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_STD_WORKER_POOLS]);
439 * Global cpu workqueue and nr_running counter for unbound gcwq. The pools
440 * for online CPUs have POOL_DISASSOCIATED set, and all their workers have
441 * WORKER_UNBOUND set.
443 static struct global_cwq unbound_global_cwq;
444 static atomic_t unbound_pool_nr_running[NR_STD_WORKER_POOLS] = {
445 [0 ... NR_STD_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
448 /* idr of all pools */
449 static DEFINE_MUTEX(worker_pool_idr_mutex);
450 static DEFINE_IDR(worker_pool_idr);
452 static int worker_thread(void *__worker);
454 static int std_worker_pool_pri(struct worker_pool *pool)
456 return pool - pool->gcwq->pools;
459 static struct global_cwq *get_gcwq(unsigned int cpu)
461 if (cpu != WORK_CPU_UNBOUND)
462 return &per_cpu(global_cwq, cpu);
464 return &unbound_global_cwq;
467 /* allocate ID and assign it to @pool */
468 static int worker_pool_assign_id(struct worker_pool *pool)
472 mutex_lock(&worker_pool_idr_mutex);
473 idr_pre_get(&worker_pool_idr, GFP_KERNEL);
474 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
475 mutex_unlock(&worker_pool_idr_mutex);
481 * Lookup worker_pool by id. The idr currently is built during boot and
482 * never modified. Don't worry about locking for now.
484 static struct worker_pool *worker_pool_by_id(int pool_id)
486 return idr_find(&worker_pool_idr, pool_id);
489 static struct worker_pool *get_std_worker_pool(int cpu, bool highpri)
491 struct global_cwq *gcwq = get_gcwq(cpu);
493 return &gcwq->pools[highpri];
496 static atomic_t *get_pool_nr_running(struct worker_pool *pool)
499 int idx = std_worker_pool_pri(pool);
501 if (cpu != WORK_CPU_UNBOUND)
502 return &per_cpu(pool_nr_running, cpu)[idx];
504 return &unbound_pool_nr_running[idx];
507 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
508 struct workqueue_struct *wq)
510 if (!(wq->flags & WQ_UNBOUND)) {
511 if (likely(cpu < nr_cpu_ids))
512 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
513 } else if (likely(cpu == WORK_CPU_UNBOUND))
514 return wq->cpu_wq.single;
518 static unsigned int work_color_to_flags(int color)
520 return color << WORK_STRUCT_COLOR_SHIFT;
523 static int get_work_color(struct work_struct *work)
525 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
526 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
529 static int work_next_color(int color)
531 return (color + 1) % WORK_NR_COLORS;
535 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
536 * contain the pointer to the queued cwq. Once execution starts, the flag
537 * is cleared and the high bits contain OFFQ flags and pool ID.
539 * set_work_cwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
540 * and clear_work_data() can be used to set the cwq, pool or clear
541 * work->data. These functions should only be called while the work is
542 * owned - ie. while the PENDING bit is set.
544 * get_work_pool() and get_work_cwq() can be used to obtain the pool or cwq
545 * corresponding to a work. Pool is available once the work has been
546 * queued anywhere after initialization until it is sync canceled. cwq is
547 * available only while the work item is queued.
549 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
550 * canceled. While being canceled, a work item may have its PENDING set
551 * but stay off timer and worklist for arbitrarily long and nobody should
552 * try to steal the PENDING bit.
554 static inline void set_work_data(struct work_struct *work, unsigned long data,
557 BUG_ON(!work_pending(work));
558 atomic_long_set(&work->data, data | flags | work_static(work));
561 static void set_work_cwq(struct work_struct *work,
562 struct cpu_workqueue_struct *cwq,
563 unsigned long extra_flags)
565 set_work_data(work, (unsigned long)cwq,
566 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
569 static void set_work_pool_and_clear_pending(struct work_struct *work,
573 * The following wmb is paired with the implied mb in
574 * test_and_set_bit(PENDING) and ensures all updates to @work made
575 * here are visible to and precede any updates by the next PENDING
579 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
582 static void clear_work_data(struct work_struct *work)
584 smp_wmb(); /* see set_work_pool_and_clear_pending() */
585 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
588 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
590 unsigned long data = atomic_long_read(&work->data);
592 if (data & WORK_STRUCT_CWQ)
593 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
599 * get_work_pool - return the worker_pool a given work was associated with
600 * @work: the work item of interest
602 * Return the worker_pool @work was last associated with. %NULL if none.
604 static struct worker_pool *get_work_pool(struct work_struct *work)
606 unsigned long data = atomic_long_read(&work->data);
607 struct worker_pool *pool;
610 if (data & WORK_STRUCT_CWQ)
611 return ((struct cpu_workqueue_struct *)
612 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
614 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
615 if (pool_id == WORK_OFFQ_POOL_NONE)
618 pool = worker_pool_by_id(pool_id);
624 * get_work_pool_id - return the worker pool ID a given work is associated with
625 * @work: the work item of interest
627 * Return the worker_pool ID @work was last associated with.
628 * %WORK_OFFQ_POOL_NONE if none.
630 static int get_work_pool_id(struct work_struct *work)
632 struct worker_pool *pool = get_work_pool(work);
634 return pool ? pool->id : WORK_OFFQ_POOL_NONE;
637 static void mark_work_canceling(struct work_struct *work)
639 unsigned long pool_id = get_work_pool_id(work);
641 pool_id <<= WORK_OFFQ_POOL_SHIFT;
642 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
645 static bool work_is_canceling(struct work_struct *work)
647 unsigned long data = atomic_long_read(&work->data);
649 return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
653 * Policy functions. These define the policies on how the global worker
654 * pools are managed. Unless noted otherwise, these functions assume that
655 * they're being called with pool->lock held.
658 static bool __need_more_worker(struct worker_pool *pool)
660 return !atomic_read(get_pool_nr_running(pool));
664 * Need to wake up a worker? Called from anything but currently
667 * Note that, because unbound workers never contribute to nr_running, this
668 * function will always return %true for unbound gcwq as long as the
669 * worklist isn't empty.
671 static bool need_more_worker(struct worker_pool *pool)
673 return !list_empty(&pool->worklist) && __need_more_worker(pool);
676 /* Can I start working? Called from busy but !running workers. */
677 static bool may_start_working(struct worker_pool *pool)
679 return pool->nr_idle;
682 /* Do I need to keep working? Called from currently running workers. */
683 static bool keep_working(struct worker_pool *pool)
685 atomic_t *nr_running = get_pool_nr_running(pool);
687 return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
690 /* Do we need a new worker? Called from manager. */
691 static bool need_to_create_worker(struct worker_pool *pool)
693 return need_more_worker(pool) && !may_start_working(pool);
696 /* Do I need to be the manager? */
697 static bool need_to_manage_workers(struct worker_pool *pool)
699 return need_to_create_worker(pool) ||
700 (pool->flags & POOL_MANAGE_WORKERS);
703 /* Do we have too many workers and should some go away? */
704 static bool too_many_workers(struct worker_pool *pool)
706 bool managing = pool->flags & POOL_MANAGING_WORKERS;
707 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
708 int nr_busy = pool->nr_workers - nr_idle;
711 * nr_idle and idle_list may disagree if idle rebinding is in
712 * progress. Never return %true if idle_list is empty.
714 if (list_empty(&pool->idle_list))
717 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
724 /* Return the first worker. Safe with preemption disabled */
725 static struct worker *first_worker(struct worker_pool *pool)
727 if (unlikely(list_empty(&pool->idle_list)))
730 return list_first_entry(&pool->idle_list, struct worker, entry);
734 * wake_up_worker - wake up an idle worker
735 * @pool: worker pool to wake worker from
737 * Wake up the first idle worker of @pool.
740 * spin_lock_irq(pool->lock).
742 static void wake_up_worker(struct worker_pool *pool)
744 struct worker *worker = first_worker(pool);
747 wake_up_process(worker->task);
751 * wq_worker_waking_up - a worker is waking up
752 * @task: task waking up
753 * @cpu: CPU @task is waking up to
755 * This function is called during try_to_wake_up() when a worker is
759 * spin_lock_irq(rq->lock)
761 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
763 struct worker *worker = kthread_data(task);
765 if (!(worker->flags & WORKER_NOT_RUNNING)) {
766 WARN_ON_ONCE(worker->pool->cpu != cpu);
767 atomic_inc(get_pool_nr_running(worker->pool));
772 * wq_worker_sleeping - a worker is going to sleep
773 * @task: task going to sleep
774 * @cpu: CPU in question, must be the current CPU number
776 * This function is called during schedule() when a busy worker is
777 * going to sleep. Worker on the same cpu can be woken up by
778 * returning pointer to its task.
781 * spin_lock_irq(rq->lock)
784 * Worker task on @cpu to wake up, %NULL if none.
786 struct task_struct *wq_worker_sleeping(struct task_struct *task,
789 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
790 struct worker_pool *pool;
791 atomic_t *nr_running;
794 * Rescuers, which may not have all the fields set up like normal
795 * workers, also reach here, let's not access anything before
796 * checking NOT_RUNNING.
798 if (worker->flags & WORKER_NOT_RUNNING)
802 nr_running = get_pool_nr_running(pool);
804 /* this can only happen on the local cpu */
805 BUG_ON(cpu != raw_smp_processor_id());
808 * The counterpart of the following dec_and_test, implied mb,
809 * worklist not empty test sequence is in insert_work().
810 * Please read comment there.
812 * NOT_RUNNING is clear. This means that we're bound to and
813 * running on the local cpu w/ rq lock held and preemption
814 * disabled, which in turn means that none else could be
815 * manipulating idle_list, so dereferencing idle_list without pool
818 if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
819 to_wakeup = first_worker(pool);
820 return to_wakeup ? to_wakeup->task : NULL;
824 * worker_set_flags - set worker flags and adjust nr_running accordingly
826 * @flags: flags to set
827 * @wakeup: wakeup an idle worker if necessary
829 * Set @flags in @worker->flags and adjust nr_running accordingly. If
830 * nr_running becomes zero and @wakeup is %true, an idle worker is
834 * spin_lock_irq(pool->lock)
836 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
839 struct worker_pool *pool = worker->pool;
841 WARN_ON_ONCE(worker->task != current);
844 * If transitioning into NOT_RUNNING, adjust nr_running and
845 * wake up an idle worker as necessary if requested by
848 if ((flags & WORKER_NOT_RUNNING) &&
849 !(worker->flags & WORKER_NOT_RUNNING)) {
850 atomic_t *nr_running = get_pool_nr_running(pool);
853 if (atomic_dec_and_test(nr_running) &&
854 !list_empty(&pool->worklist))
855 wake_up_worker(pool);
857 atomic_dec(nr_running);
860 worker->flags |= flags;
864 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
866 * @flags: flags to clear
868 * Clear @flags in @worker->flags and adjust nr_running accordingly.
871 * spin_lock_irq(pool->lock)
873 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
875 struct worker_pool *pool = worker->pool;
876 unsigned int oflags = worker->flags;
878 WARN_ON_ONCE(worker->task != current);
880 worker->flags &= ~flags;
883 * If transitioning out of NOT_RUNNING, increment nr_running. Note
884 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
885 * of multiple flags, not a single flag.
887 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
888 if (!(worker->flags & WORKER_NOT_RUNNING))
889 atomic_inc(get_pool_nr_running(pool));
893 * find_worker_executing_work - find worker which is executing a work
894 * @pool: pool of interest
895 * @work: work to find worker for
897 * Find a worker which is executing @work on @pool by searching
898 * @pool->busy_hash which is keyed by the address of @work. For a worker
899 * to match, its current execution should match the address of @work and
900 * its work function. This is to avoid unwanted dependency between
901 * unrelated work executions through a work item being recycled while still
904 * This is a bit tricky. A work item may be freed once its execution
905 * starts and nothing prevents the freed area from being recycled for
906 * another work item. If the same work item address ends up being reused
907 * before the original execution finishes, workqueue will identify the
908 * recycled work item as currently executing and make it wait until the
909 * current execution finishes, introducing an unwanted dependency.
911 * This function checks the work item address, work function and workqueue
912 * to avoid false positives. Note that this isn't complete as one may
913 * construct a work function which can introduce dependency onto itself
914 * through a recycled work item. Well, if somebody wants to shoot oneself
915 * in the foot that badly, there's only so much we can do, and if such
916 * deadlock actually occurs, it should be easy to locate the culprit work
920 * spin_lock_irq(pool->lock).
923 * Pointer to worker which is executing @work if found, NULL
926 static struct worker *find_worker_executing_work(struct worker_pool *pool,
927 struct work_struct *work)
929 struct worker *worker;
930 struct hlist_node *tmp;
932 hash_for_each_possible(pool->busy_hash, worker, tmp, hentry,
934 if (worker->current_work == work &&
935 worker->current_func == work->func)
942 * move_linked_works - move linked works to a list
943 * @work: start of series of works to be scheduled
944 * @head: target list to append @work to
945 * @nextp: out paramter for nested worklist walking
947 * Schedule linked works starting from @work to @head. Work series to
948 * be scheduled starts at @work and includes any consecutive work with
949 * WORK_STRUCT_LINKED set in its predecessor.
951 * If @nextp is not NULL, it's updated to point to the next work of
952 * the last scheduled work. This allows move_linked_works() to be
953 * nested inside outer list_for_each_entry_safe().
956 * spin_lock_irq(pool->lock).
958 static void move_linked_works(struct work_struct *work, struct list_head *head,
959 struct work_struct **nextp)
961 struct work_struct *n;
964 * Linked worklist will always end before the end of the list,
965 * use NULL for list head.
967 list_for_each_entry_safe_from(work, n, NULL, entry) {
968 list_move_tail(&work->entry, head);
969 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
974 * If we're already inside safe list traversal and have moved
975 * multiple works to the scheduled queue, the next position
976 * needs to be updated.
982 static void cwq_activate_delayed_work(struct work_struct *work)
984 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
986 trace_workqueue_activate_work(work);
987 move_linked_works(work, &cwq->pool->worklist, NULL);
988 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
992 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
994 struct work_struct *work = list_first_entry(&cwq->delayed_works,
995 struct work_struct, entry);
997 cwq_activate_delayed_work(work);
1001 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1002 * @cwq: cwq of interest
1003 * @color: color of work which left the queue
1005 * A work either has completed or is removed from pending queue,
1006 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1009 * spin_lock_irq(pool->lock).
1011 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
1013 /* ignore uncolored works */
1014 if (color == WORK_NO_COLOR)
1017 cwq->nr_in_flight[color]--;
1020 if (!list_empty(&cwq->delayed_works)) {
1021 /* one down, submit a delayed one */
1022 if (cwq->nr_active < cwq->max_active)
1023 cwq_activate_first_delayed(cwq);
1026 /* is flush in progress and are we at the flushing tip? */
1027 if (likely(cwq->flush_color != color))
1030 /* are there still in-flight works? */
1031 if (cwq->nr_in_flight[color])
1034 /* this cwq is done, clear flush_color */
1035 cwq->flush_color = -1;
1038 * If this was the last cwq, wake up the first flusher. It
1039 * will handle the rest.
1041 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1042 complete(&cwq->wq->first_flusher->done);
1046 * try_to_grab_pending - steal work item from worklist and disable irq
1047 * @work: work item to steal
1048 * @is_dwork: @work is a delayed_work
1049 * @flags: place to store irq state
1051 * Try to grab PENDING bit of @work. This function can handle @work in any
1052 * stable state - idle, on timer or on worklist. Return values are
1054 * 1 if @work was pending and we successfully stole PENDING
1055 * 0 if @work was idle and we claimed PENDING
1056 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1057 * -ENOENT if someone else is canceling @work, this state may persist
1058 * for arbitrarily long
1060 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1061 * interrupted while holding PENDING and @work off queue, irq must be
1062 * disabled on entry. This, combined with delayed_work->timer being
1063 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1065 * On successful return, >= 0, irq is disabled and the caller is
1066 * responsible for releasing it using local_irq_restore(*@flags).
1068 * This function is safe to call from any context including IRQ handler.
1070 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1071 unsigned long *flags)
1073 struct worker_pool *pool;
1075 local_irq_save(*flags);
1077 /* try to steal the timer if it exists */
1079 struct delayed_work *dwork = to_delayed_work(work);
1082 * dwork->timer is irqsafe. If del_timer() fails, it's
1083 * guaranteed that the timer is not queued anywhere and not
1084 * running on the local CPU.
1086 if (likely(del_timer(&dwork->timer)))
1090 /* try to claim PENDING the normal way */
1091 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1095 * The queueing is in progress, or it is already queued. Try to
1096 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1098 pool = get_work_pool(work);
1102 spin_lock(&pool->lock);
1103 if (!list_empty(&work->entry)) {
1105 * This work is queued, but perhaps we locked the wrong
1106 * pool. In that case we must see the new value after
1107 * rmb(), see insert_work()->wmb().
1110 if (pool == get_work_pool(work)) {
1111 debug_work_deactivate(work);
1114 * A delayed work item cannot be grabbed directly
1115 * because it might have linked NO_COLOR work items
1116 * which, if left on the delayed_list, will confuse
1117 * cwq->nr_active management later on and cause
1118 * stall. Make sure the work item is activated
1121 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1122 cwq_activate_delayed_work(work);
1124 list_del_init(&work->entry);
1125 cwq_dec_nr_in_flight(get_work_cwq(work),
1126 get_work_color(work));
1128 spin_unlock(&pool->lock);
1132 spin_unlock(&pool->lock);
1134 local_irq_restore(*flags);
1135 if (work_is_canceling(work))
1142 * insert_work - insert a work into gcwq
1143 * @cwq: cwq @work belongs to
1144 * @work: work to insert
1145 * @head: insertion point
1146 * @extra_flags: extra WORK_STRUCT_* flags to set
1148 * Insert @work which belongs to @cwq into @gcwq after @head.
1149 * @extra_flags is or'd to work_struct flags.
1152 * spin_lock_irq(pool->lock).
1154 static void insert_work(struct cpu_workqueue_struct *cwq,
1155 struct work_struct *work, struct list_head *head,
1156 unsigned int extra_flags)
1158 struct worker_pool *pool = cwq->pool;
1160 /* we own @work, set data and link */
1161 set_work_cwq(work, cwq, extra_flags);
1164 * Ensure that we get the right work->data if we see the
1165 * result of list_add() below, see try_to_grab_pending().
1169 list_add_tail(&work->entry, head);
1172 * Ensure either worker_sched_deactivated() sees the above
1173 * list_add_tail() or we see zero nr_running to avoid workers
1174 * lying around lazily while there are works to be processed.
1178 if (__need_more_worker(pool))
1179 wake_up_worker(pool);
1183 * Test whether @work is being queued from another work executing on the
1184 * same workqueue. This is rather expensive and should only be used from
1187 static bool is_chained_work(struct workqueue_struct *wq)
1189 unsigned long flags;
1192 for_each_gcwq_cpu(cpu) {
1193 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
1194 struct worker_pool *pool = cwq->pool;
1195 struct worker *worker;
1196 struct hlist_node *pos;
1199 spin_lock_irqsave(&pool->lock, flags);
1200 for_each_busy_worker(worker, i, pos, pool) {
1201 if (worker->task != current)
1203 spin_unlock_irqrestore(&pool->lock, flags);
1205 * I'm @worker, no locking necessary. See if @work
1206 * is headed to the same workqueue.
1208 return worker->current_cwq->wq == wq;
1210 spin_unlock_irqrestore(&pool->lock, flags);
1215 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1216 struct work_struct *work)
1218 bool highpri = wq->flags & WQ_HIGHPRI;
1219 struct worker_pool *pool;
1220 struct cpu_workqueue_struct *cwq;
1221 struct list_head *worklist;
1222 unsigned int work_flags;
1223 unsigned int req_cpu = cpu;
1226 * While a work item is PENDING && off queue, a task trying to
1227 * steal the PENDING will busy-loop waiting for it to either get
1228 * queued or lose PENDING. Grabbing PENDING and queueing should
1229 * happen with IRQ disabled.
1231 WARN_ON_ONCE(!irqs_disabled());
1233 debug_work_activate(work);
1235 /* if dying, only works from the same workqueue are allowed */
1236 if (unlikely(wq->flags & WQ_DRAINING) &&
1237 WARN_ON_ONCE(!is_chained_work(wq)))
1240 /* determine pool to use */
1241 if (!(wq->flags & WQ_UNBOUND)) {
1242 struct worker_pool *last_pool;
1244 if (cpu == WORK_CPU_UNBOUND)
1245 cpu = raw_smp_processor_id();
1248 * It's multi cpu. If @work was previously on a different
1249 * cpu, it might still be running there, in which case the
1250 * work needs to be queued on that cpu to guarantee
1253 pool = get_std_worker_pool(cpu, highpri);
1254 last_pool = get_work_pool(work);
1256 if (last_pool && last_pool != pool) {
1257 struct worker *worker;
1259 spin_lock(&last_pool->lock);
1261 worker = find_worker_executing_work(last_pool, work);
1263 if (worker && worker->current_cwq->wq == wq)
1266 /* meh... not running there, queue here */
1267 spin_unlock(&last_pool->lock);
1268 spin_lock(&pool->lock);
1271 spin_lock(&pool->lock);
1274 pool = get_std_worker_pool(WORK_CPU_UNBOUND, highpri);
1275 spin_lock(&pool->lock);
1278 /* pool determined, get cwq and queue */
1279 cwq = get_cwq(pool->cpu, wq);
1280 trace_workqueue_queue_work(req_cpu, cwq, work);
1282 if (WARN_ON(!list_empty(&work->entry))) {
1283 spin_unlock(&pool->lock);
1287 cwq->nr_in_flight[cwq->work_color]++;
1288 work_flags = work_color_to_flags(cwq->work_color);
1290 if (likely(cwq->nr_active < cwq->max_active)) {
1291 trace_workqueue_activate_work(work);
1293 worklist = &cwq->pool->worklist;
1295 work_flags |= WORK_STRUCT_DELAYED;
1296 worklist = &cwq->delayed_works;
1299 insert_work(cwq, work, worklist, work_flags);
1301 spin_unlock(&pool->lock);
1305 * queue_work_on - queue work on specific cpu
1306 * @cpu: CPU number to execute work on
1307 * @wq: workqueue to use
1308 * @work: work to queue
1310 * Returns %false if @work was already on a queue, %true otherwise.
1312 * We queue the work to a specific CPU, the caller must ensure it
1315 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1316 struct work_struct *work)
1319 unsigned long flags;
1321 local_irq_save(flags);
1323 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1324 __queue_work(cpu, wq, work);
1328 local_irq_restore(flags);
1331 EXPORT_SYMBOL_GPL(queue_work_on);
1334 * queue_work - queue work on a workqueue
1335 * @wq: workqueue to use
1336 * @work: work to queue
1338 * Returns %false if @work was already on a queue, %true otherwise.
1340 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1341 * it can be processed by another CPU.
1343 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1345 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1347 EXPORT_SYMBOL_GPL(queue_work);
1349 void delayed_work_timer_fn(unsigned long __data)
1351 struct delayed_work *dwork = (struct delayed_work *)__data;
1352 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1354 /* should have been called from irqsafe timer with irq already off */
1355 __queue_work(dwork->cpu, cwq->wq, &dwork->work);
1357 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1359 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1360 struct delayed_work *dwork, unsigned long delay)
1362 struct timer_list *timer = &dwork->timer;
1363 struct work_struct *work = &dwork->work;
1366 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1367 timer->data != (unsigned long)dwork);
1368 WARN_ON_ONCE(timer_pending(timer));
1369 WARN_ON_ONCE(!list_empty(&work->entry));
1372 * If @delay is 0, queue @dwork->work immediately. This is for
1373 * both optimization and correctness. The earliest @timer can
1374 * expire is on the closest next tick and delayed_work users depend
1375 * on that there's no such delay when @delay is 0.
1378 __queue_work(cpu, wq, &dwork->work);
1382 timer_stats_timer_set_start_info(&dwork->timer);
1385 * This stores cwq for the moment, for the timer_fn. Note that the
1386 * work's pool is preserved to allow reentrance detection for
1389 if (!(wq->flags & WQ_UNBOUND)) {
1390 struct worker_pool *pool = get_work_pool(work);
1393 * If we cannot get the last pool from @work directly,
1394 * select the last CPU such that it avoids unnecessarily
1395 * triggering non-reentrancy check in __queue_work().
1400 if (lcpu == WORK_CPU_UNBOUND)
1401 lcpu = raw_smp_processor_id();
1403 lcpu = WORK_CPU_UNBOUND;
1406 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1409 timer->expires = jiffies + delay;
1411 if (unlikely(cpu != WORK_CPU_UNBOUND))
1412 add_timer_on(timer, cpu);
1418 * queue_delayed_work_on - queue work on specific CPU after delay
1419 * @cpu: CPU number to execute work on
1420 * @wq: workqueue to use
1421 * @dwork: work to queue
1422 * @delay: number of jiffies to wait before queueing
1424 * Returns %false if @work was already on a queue, %true otherwise. If
1425 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1428 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1429 struct delayed_work *dwork, unsigned long delay)
1431 struct work_struct *work = &dwork->work;
1433 unsigned long flags;
1435 /* read the comment in __queue_work() */
1436 local_irq_save(flags);
1438 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1439 __queue_delayed_work(cpu, wq, dwork, delay);
1443 local_irq_restore(flags);
1446 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1449 * queue_delayed_work - queue work on a workqueue after delay
1450 * @wq: workqueue to use
1451 * @dwork: delayable work to queue
1452 * @delay: number of jiffies to wait before queueing
1454 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1456 bool queue_delayed_work(struct workqueue_struct *wq,
1457 struct delayed_work *dwork, unsigned long delay)
1459 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1461 EXPORT_SYMBOL_GPL(queue_delayed_work);
1464 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1465 * @cpu: CPU number to execute work on
1466 * @wq: workqueue to use
1467 * @dwork: work to queue
1468 * @delay: number of jiffies to wait before queueing
1470 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1471 * modify @dwork's timer so that it expires after @delay. If @delay is
1472 * zero, @work is guaranteed to be scheduled immediately regardless of its
1475 * Returns %false if @dwork was idle and queued, %true if @dwork was
1476 * pending and its timer was modified.
1478 * This function is safe to call from any context including IRQ handler.
1479 * See try_to_grab_pending() for details.
1481 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1482 struct delayed_work *dwork, unsigned long delay)
1484 unsigned long flags;
1488 ret = try_to_grab_pending(&dwork->work, true, &flags);
1489 } while (unlikely(ret == -EAGAIN));
1491 if (likely(ret >= 0)) {
1492 __queue_delayed_work(cpu, wq, dwork, delay);
1493 local_irq_restore(flags);
1496 /* -ENOENT from try_to_grab_pending() becomes %true */
1499 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1502 * mod_delayed_work - modify delay of or queue a delayed work
1503 * @wq: workqueue to use
1504 * @dwork: work to queue
1505 * @delay: number of jiffies to wait before queueing
1507 * mod_delayed_work_on() on local CPU.
1509 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1510 unsigned long delay)
1512 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1514 EXPORT_SYMBOL_GPL(mod_delayed_work);
1517 * worker_enter_idle - enter idle state
1518 * @worker: worker which is entering idle state
1520 * @worker is entering idle state. Update stats and idle timer if
1524 * spin_lock_irq(pool->lock).
1526 static void worker_enter_idle(struct worker *worker)
1528 struct worker_pool *pool = worker->pool;
1530 BUG_ON(worker->flags & WORKER_IDLE);
1531 BUG_ON(!list_empty(&worker->entry) &&
1532 (worker->hentry.next || worker->hentry.pprev));
1534 /* can't use worker_set_flags(), also called from start_worker() */
1535 worker->flags |= WORKER_IDLE;
1537 worker->last_active = jiffies;
1539 /* idle_list is LIFO */
1540 list_add(&worker->entry, &pool->idle_list);
1542 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1543 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1546 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1547 * pool->lock between setting %WORKER_UNBOUND and zapping
1548 * nr_running, the warning may trigger spuriously. Check iff
1549 * unbind is not in progress.
1551 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1552 pool->nr_workers == pool->nr_idle &&
1553 atomic_read(get_pool_nr_running(pool)));
1557 * worker_leave_idle - leave idle state
1558 * @worker: worker which is leaving idle state
1560 * @worker is leaving idle state. Update stats.
1563 * spin_lock_irq(pool->lock).
1565 static void worker_leave_idle(struct worker *worker)
1567 struct worker_pool *pool = worker->pool;
1569 BUG_ON(!(worker->flags & WORKER_IDLE));
1570 worker_clr_flags(worker, WORKER_IDLE);
1572 list_del_init(&worker->entry);
1576 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1579 * Works which are scheduled while the cpu is online must at least be
1580 * scheduled to a worker which is bound to the cpu so that if they are
1581 * flushed from cpu callbacks while cpu is going down, they are
1582 * guaranteed to execute on the cpu.
1584 * This function is to be used by rogue workers and rescuers to bind
1585 * themselves to the target cpu and may race with cpu going down or
1586 * coming online. kthread_bind() can't be used because it may put the
1587 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1588 * verbatim as it's best effort and blocking and gcwq may be
1589 * [dis]associated in the meantime.
1591 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1592 * binding against %POOL_DISASSOCIATED which is set during
1593 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1594 * enters idle state or fetches works without dropping lock, it can
1595 * guarantee the scheduling requirement described in the first paragraph.
1598 * Might sleep. Called without any lock but returns with pool->lock
1602 * %true if the associated gcwq is online (@worker is successfully
1603 * bound), %false if offline.
1605 static bool worker_maybe_bind_and_lock(struct worker *worker)
1606 __acquires(&pool->lock)
1608 struct worker_pool *pool = worker->pool;
1609 struct task_struct *task = worker->task;
1613 * The following call may fail, succeed or succeed
1614 * without actually migrating the task to the cpu if
1615 * it races with cpu hotunplug operation. Verify
1616 * against POOL_DISASSOCIATED.
1618 if (!(pool->flags & POOL_DISASSOCIATED))
1619 set_cpus_allowed_ptr(task, get_cpu_mask(pool->cpu));
1621 spin_lock_irq(&pool->lock);
1622 if (pool->flags & POOL_DISASSOCIATED)
1624 if (task_cpu(task) == pool->cpu &&
1625 cpumask_equal(¤t->cpus_allowed,
1626 get_cpu_mask(pool->cpu)))
1628 spin_unlock_irq(&pool->lock);
1631 * We've raced with CPU hot[un]plug. Give it a breather
1632 * and retry migration. cond_resched() is required here;
1633 * otherwise, we might deadlock against cpu_stop trying to
1634 * bring down the CPU on non-preemptive kernel.
1642 * Rebind an idle @worker to its CPU. worker_thread() will test
1643 * list_empty(@worker->entry) before leaving idle and call this function.
1645 static void idle_worker_rebind(struct worker *worker)
1647 /* CPU may go down again inbetween, clear UNBOUND only on success */
1648 if (worker_maybe_bind_and_lock(worker))
1649 worker_clr_flags(worker, WORKER_UNBOUND);
1651 /* rebind complete, become available again */
1652 list_add(&worker->entry, &worker->pool->idle_list);
1653 spin_unlock_irq(&worker->pool->lock);
1657 * Function for @worker->rebind.work used to rebind unbound busy workers to
1658 * the associated cpu which is coming back online. This is scheduled by
1659 * cpu up but can race with other cpu hotplug operations and may be
1660 * executed twice without intervening cpu down.
1662 static void busy_worker_rebind_fn(struct work_struct *work)
1664 struct worker *worker = container_of(work, struct worker, rebind_work);
1666 if (worker_maybe_bind_and_lock(worker))
1667 worker_clr_flags(worker, WORKER_UNBOUND);
1669 spin_unlock_irq(&worker->pool->lock);
1673 * rebind_workers - rebind all workers of a pool to the associated CPU
1674 * @pool: pool of interest
1676 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1677 * is different for idle and busy ones.
1679 * Idle ones will be removed from the idle_list and woken up. They will
1680 * add themselves back after completing rebind. This ensures that the
1681 * idle_list doesn't contain any unbound workers when re-bound busy workers
1682 * try to perform local wake-ups for concurrency management.
1684 * Busy workers can rebind after they finish their current work items.
1685 * Queueing the rebind work item at the head of the scheduled list is
1686 * enough. Note that nr_running will be properly bumped as busy workers
1689 * On return, all non-manager workers are scheduled for rebind - see
1690 * manage_workers() for the manager special case. Any idle worker
1691 * including the manager will not appear on @idle_list until rebind is
1692 * complete, making local wake-ups safe.
1694 static void rebind_workers(struct worker_pool *pool)
1696 struct worker *worker, *n;
1697 struct hlist_node *pos;
1700 lockdep_assert_held(&pool->assoc_mutex);
1701 lockdep_assert_held(&pool->lock);
1703 /* dequeue and kick idle ones */
1704 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1706 * idle workers should be off @pool->idle_list until rebind
1707 * is complete to avoid receiving premature local wake-ups.
1709 list_del_init(&worker->entry);
1712 * worker_thread() will see the above dequeuing and call
1713 * idle_worker_rebind().
1715 wake_up_process(worker->task);
1718 /* rebind busy workers */
1719 for_each_busy_worker(worker, i, pos, pool) {
1720 struct work_struct *rebind_work = &worker->rebind_work;
1721 struct workqueue_struct *wq;
1723 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1724 work_data_bits(rebind_work)))
1727 debug_work_activate(rebind_work);
1730 * wq doesn't really matter but let's keep @worker->pool
1731 * and @cwq->pool consistent for sanity.
1733 if (std_worker_pool_pri(worker->pool))
1734 wq = system_highpri_wq;
1738 insert_work(get_cwq(pool->cpu, wq), rebind_work,
1739 worker->scheduled.next,
1740 work_color_to_flags(WORK_NO_COLOR));
1744 static struct worker *alloc_worker(void)
1746 struct worker *worker;
1748 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1750 INIT_LIST_HEAD(&worker->entry);
1751 INIT_LIST_HEAD(&worker->scheduled);
1752 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1753 /* on creation a worker is in !idle && prep state */
1754 worker->flags = WORKER_PREP;
1760 * create_worker - create a new workqueue worker
1761 * @pool: pool the new worker will belong to
1763 * Create a new worker which is bound to @pool. The returned worker
1764 * can be started by calling start_worker() or destroyed using
1768 * Might sleep. Does GFP_KERNEL allocations.
1771 * Pointer to the newly created worker.
1773 static struct worker *create_worker(struct worker_pool *pool)
1775 const char *pri = std_worker_pool_pri(pool) ? "H" : "";
1776 struct worker *worker = NULL;
1779 spin_lock_irq(&pool->lock);
1780 while (ida_get_new(&pool->worker_ida, &id)) {
1781 spin_unlock_irq(&pool->lock);
1782 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1784 spin_lock_irq(&pool->lock);
1786 spin_unlock_irq(&pool->lock);
1788 worker = alloc_worker();
1792 worker->pool = pool;
1795 if (pool->cpu != WORK_CPU_UNBOUND)
1796 worker->task = kthread_create_on_node(worker_thread,
1797 worker, cpu_to_node(pool->cpu),
1798 "kworker/%u:%d%s", pool->cpu, id, pri);
1800 worker->task = kthread_create(worker_thread, worker,
1801 "kworker/u:%d%s", id, pri);
1802 if (IS_ERR(worker->task))
1805 if (std_worker_pool_pri(pool))
1806 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1809 * Determine CPU binding of the new worker depending on
1810 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1811 * flag remains stable across this function. See the comments
1812 * above the flag definition for details.
1814 * As an unbound worker may later become a regular one if CPU comes
1815 * online, make sure every worker has %PF_THREAD_BOUND set.
1817 if (!(pool->flags & POOL_DISASSOCIATED)) {
1818 kthread_bind(worker->task, pool->cpu);
1820 worker->task->flags |= PF_THREAD_BOUND;
1821 worker->flags |= WORKER_UNBOUND;
1827 spin_lock_irq(&pool->lock);
1828 ida_remove(&pool->worker_ida, id);
1829 spin_unlock_irq(&pool->lock);
1836 * start_worker - start a newly created worker
1837 * @worker: worker to start
1839 * Make the gcwq aware of @worker and start it.
1842 * spin_lock_irq(pool->lock).
1844 static void start_worker(struct worker *worker)
1846 worker->flags |= WORKER_STARTED;
1847 worker->pool->nr_workers++;
1848 worker_enter_idle(worker);
1849 wake_up_process(worker->task);
1853 * destroy_worker - destroy a workqueue worker
1854 * @worker: worker to be destroyed
1856 * Destroy @worker and adjust @gcwq stats accordingly.
1859 * spin_lock_irq(pool->lock) which is released and regrabbed.
1861 static void destroy_worker(struct worker *worker)
1863 struct worker_pool *pool = worker->pool;
1864 int id = worker->id;
1866 /* sanity check frenzy */
1867 BUG_ON(worker->current_work);
1868 BUG_ON(!list_empty(&worker->scheduled));
1870 if (worker->flags & WORKER_STARTED)
1872 if (worker->flags & WORKER_IDLE)
1875 list_del_init(&worker->entry);
1876 worker->flags |= WORKER_DIE;
1878 spin_unlock_irq(&pool->lock);
1880 kthread_stop(worker->task);
1883 spin_lock_irq(&pool->lock);
1884 ida_remove(&pool->worker_ida, id);
1887 static void idle_worker_timeout(unsigned long __pool)
1889 struct worker_pool *pool = (void *)__pool;
1891 spin_lock_irq(&pool->lock);
1893 if (too_many_workers(pool)) {
1894 struct worker *worker;
1895 unsigned long expires;
1897 /* idle_list is kept in LIFO order, check the last one */
1898 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1899 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1901 if (time_before(jiffies, expires))
1902 mod_timer(&pool->idle_timer, expires);
1904 /* it's been idle for too long, wake up manager */
1905 pool->flags |= POOL_MANAGE_WORKERS;
1906 wake_up_worker(pool);
1910 spin_unlock_irq(&pool->lock);
1913 static bool send_mayday(struct work_struct *work)
1915 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1916 struct workqueue_struct *wq = cwq->wq;
1919 if (!(wq->flags & WQ_RESCUER))
1922 /* mayday mayday mayday */
1923 cpu = cwq->pool->cpu;
1924 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1925 if (cpu == WORK_CPU_UNBOUND)
1927 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1928 wake_up_process(wq->rescuer->task);
1932 static void gcwq_mayday_timeout(unsigned long __pool)
1934 struct worker_pool *pool = (void *)__pool;
1935 struct work_struct *work;
1937 spin_lock_irq(&pool->lock);
1939 if (need_to_create_worker(pool)) {
1941 * We've been trying to create a new worker but
1942 * haven't been successful. We might be hitting an
1943 * allocation deadlock. Send distress signals to
1946 list_for_each_entry(work, &pool->worklist, entry)
1950 spin_unlock_irq(&pool->lock);
1952 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1956 * maybe_create_worker - create a new worker if necessary
1957 * @pool: pool to create a new worker for
1959 * Create a new worker for @pool if necessary. @pool is guaranteed to
1960 * have at least one idle worker on return from this function. If
1961 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1962 * sent to all rescuers with works scheduled on @pool to resolve
1963 * possible allocation deadlock.
1965 * On return, need_to_create_worker() is guaranteed to be false and
1966 * may_start_working() true.
1969 * spin_lock_irq(pool->lock) which may be released and regrabbed
1970 * multiple times. Does GFP_KERNEL allocations. Called only from
1974 * false if no action was taken and pool->lock stayed locked, true
1977 static bool maybe_create_worker(struct worker_pool *pool)
1978 __releases(&pool->lock)
1979 __acquires(&pool->lock)
1981 if (!need_to_create_worker(pool))
1984 spin_unlock_irq(&pool->lock);
1986 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1987 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1990 struct worker *worker;
1992 worker = create_worker(pool);
1994 del_timer_sync(&pool->mayday_timer);
1995 spin_lock_irq(&pool->lock);
1996 start_worker(worker);
1997 BUG_ON(need_to_create_worker(pool));
2001 if (!need_to_create_worker(pool))
2004 __set_current_state(TASK_INTERRUPTIBLE);
2005 schedule_timeout(CREATE_COOLDOWN);
2007 if (!need_to_create_worker(pool))
2011 del_timer_sync(&pool->mayday_timer);
2012 spin_lock_irq(&pool->lock);
2013 if (need_to_create_worker(pool))
2019 * maybe_destroy_worker - destroy workers which have been idle for a while
2020 * @pool: pool to destroy workers for
2022 * Destroy @pool workers which have been idle for longer than
2023 * IDLE_WORKER_TIMEOUT.
2026 * spin_lock_irq(pool->lock) which may be released and regrabbed
2027 * multiple times. Called only from manager.
2030 * false if no action was taken and pool->lock stayed locked, true
2033 static bool maybe_destroy_workers(struct worker_pool *pool)
2037 while (too_many_workers(pool)) {
2038 struct worker *worker;
2039 unsigned long expires;
2041 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2042 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2044 if (time_before(jiffies, expires)) {
2045 mod_timer(&pool->idle_timer, expires);
2049 destroy_worker(worker);
2057 * manage_workers - manage worker pool
2060 * Assume the manager role and manage gcwq worker pool @worker belongs
2061 * to. At any given time, there can be only zero or one manager per
2062 * gcwq. The exclusion is handled automatically by this function.
2064 * The caller can safely start processing works on false return. On
2065 * true return, it's guaranteed that need_to_create_worker() is false
2066 * and may_start_working() is true.
2069 * spin_lock_irq(pool->lock) which may be released and regrabbed
2070 * multiple times. Does GFP_KERNEL allocations.
2073 * spin_lock_irq(pool->lock) which may be released and regrabbed
2074 * multiple times. Does GFP_KERNEL allocations.
2076 static bool manage_workers(struct worker *worker)
2078 struct worker_pool *pool = worker->pool;
2081 if (pool->flags & POOL_MANAGING_WORKERS)
2084 pool->flags |= POOL_MANAGING_WORKERS;
2087 * To simplify both worker management and CPU hotplug, hold off
2088 * management while hotplug is in progress. CPU hotplug path can't
2089 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2090 * lead to idle worker depletion (all become busy thinking someone
2091 * else is managing) which in turn can result in deadlock under
2092 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2093 * manager against CPU hotplug.
2095 * assoc_mutex would always be free unless CPU hotplug is in
2096 * progress. trylock first without dropping @pool->lock.
2098 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2099 spin_unlock_irq(&pool->lock);
2100 mutex_lock(&pool->assoc_mutex);
2102 * CPU hotplug could have happened while we were waiting
2103 * for assoc_mutex. Hotplug itself can't handle us
2104 * because manager isn't either on idle or busy list, and
2105 * @gcwq's state and ours could have deviated.
2107 * As hotplug is now excluded via assoc_mutex, we can
2108 * simply try to bind. It will succeed or fail depending
2109 * on @gcwq's current state. Try it and adjust
2110 * %WORKER_UNBOUND accordingly.
2112 if (worker_maybe_bind_and_lock(worker))
2113 worker->flags &= ~WORKER_UNBOUND;
2115 worker->flags |= WORKER_UNBOUND;
2120 pool->flags &= ~POOL_MANAGE_WORKERS;
2123 * Destroy and then create so that may_start_working() is true
2126 ret |= maybe_destroy_workers(pool);
2127 ret |= maybe_create_worker(pool);
2129 pool->flags &= ~POOL_MANAGING_WORKERS;
2130 mutex_unlock(&pool->assoc_mutex);
2135 * process_one_work - process single work
2137 * @work: work to process
2139 * Process @work. This function contains all the logics necessary to
2140 * process a single work including synchronization against and
2141 * interaction with other workers on the same cpu, queueing and
2142 * flushing. As long as context requirement is met, any worker can
2143 * call this function to process a work.
2146 * spin_lock_irq(pool->lock) which is released and regrabbed.
2148 static void process_one_work(struct worker *worker, struct work_struct *work)
2149 __releases(&pool->lock)
2150 __acquires(&pool->lock)
2152 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2153 struct worker_pool *pool = worker->pool;
2154 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2156 struct worker *collision;
2157 #ifdef CONFIG_LOCKDEP
2159 * It is permissible to free the struct work_struct from
2160 * inside the function that is called from it, this we need to
2161 * take into account for lockdep too. To avoid bogus "held
2162 * lock freed" warnings as well as problems when looking into
2163 * work->lockdep_map, make a copy and use that here.
2165 struct lockdep_map lockdep_map;
2167 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2170 * Ensure we're on the correct CPU. DISASSOCIATED test is
2171 * necessary to avoid spurious warnings from rescuers servicing the
2172 * unbound or a disassociated pool.
2174 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2175 !(pool->flags & POOL_DISASSOCIATED) &&
2176 raw_smp_processor_id() != pool->cpu);
2179 * A single work shouldn't be executed concurrently by
2180 * multiple workers on a single cpu. Check whether anyone is
2181 * already processing the work. If so, defer the work to the
2182 * currently executing one.
2184 collision = find_worker_executing_work(pool, work);
2185 if (unlikely(collision)) {
2186 move_linked_works(work, &collision->scheduled, NULL);
2190 /* claim and dequeue */
2191 debug_work_deactivate(work);
2192 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2193 worker->current_work = work;
2194 worker->current_func = work->func;
2195 worker->current_cwq = cwq;
2196 work_color = get_work_color(work);
2198 list_del_init(&work->entry);
2201 * CPU intensive works don't participate in concurrency
2202 * management. They're the scheduler's responsibility.
2204 if (unlikely(cpu_intensive))
2205 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2208 * Unbound pool isn't concurrency managed and work items should be
2209 * executed ASAP. Wake up another worker if necessary.
2211 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2212 wake_up_worker(pool);
2215 * Record the last pool and clear PENDING which should be the last
2216 * update to @work. Also, do this inside @pool->lock so that
2217 * PENDING and queued state changes happen together while IRQ is
2220 set_work_pool_and_clear_pending(work, pool->id);
2222 spin_unlock_irq(&pool->lock);
2224 lock_map_acquire_read(&cwq->wq->lockdep_map);
2225 lock_map_acquire(&lockdep_map);
2226 trace_workqueue_execute_start(work);
2227 worker->current_func(work);
2229 * While we must be careful to not use "work" after this, the trace
2230 * point will only record its address.
2232 trace_workqueue_execute_end(work);
2233 lock_map_release(&lockdep_map);
2234 lock_map_release(&cwq->wq->lockdep_map);
2236 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2237 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2238 " last function: %pf\n",
2239 current->comm, preempt_count(), task_pid_nr(current),
2240 worker->current_func);
2241 debug_show_held_locks(current);
2245 spin_lock_irq(&pool->lock);
2247 /* clear cpu intensive status */
2248 if (unlikely(cpu_intensive))
2249 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2251 /* we're done with it, release */
2252 hash_del(&worker->hentry);
2253 worker->current_work = NULL;
2254 worker->current_func = NULL;
2255 worker->current_cwq = NULL;
2256 cwq_dec_nr_in_flight(cwq, work_color);
2260 * process_scheduled_works - process scheduled works
2263 * Process all scheduled works. Please note that the scheduled list
2264 * may change while processing a work, so this function repeatedly
2265 * fetches a work from the top and executes it.
2268 * spin_lock_irq(pool->lock) which may be released and regrabbed
2271 static void process_scheduled_works(struct worker *worker)
2273 while (!list_empty(&worker->scheduled)) {
2274 struct work_struct *work = list_first_entry(&worker->scheduled,
2275 struct work_struct, entry);
2276 process_one_work(worker, work);
2281 * worker_thread - the worker thread function
2284 * The gcwq worker thread function. There's a single dynamic pool of
2285 * these per each cpu. These workers process all works regardless of
2286 * their specific target workqueue. The only exception is works which
2287 * belong to workqueues with a rescuer which will be explained in
2290 static int worker_thread(void *__worker)
2292 struct worker *worker = __worker;
2293 struct worker_pool *pool = worker->pool;
2295 /* tell the scheduler that this is a workqueue worker */
2296 worker->task->flags |= PF_WQ_WORKER;
2298 spin_lock_irq(&pool->lock);
2300 /* we are off idle list if destruction or rebind is requested */
2301 if (unlikely(list_empty(&worker->entry))) {
2302 spin_unlock_irq(&pool->lock);
2304 /* if DIE is set, destruction is requested */
2305 if (worker->flags & WORKER_DIE) {
2306 worker->task->flags &= ~PF_WQ_WORKER;
2310 /* otherwise, rebind */
2311 idle_worker_rebind(worker);
2315 worker_leave_idle(worker);
2317 /* no more worker necessary? */
2318 if (!need_more_worker(pool))
2321 /* do we need to manage? */
2322 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2326 * ->scheduled list can only be filled while a worker is
2327 * preparing to process a work or actually processing it.
2328 * Make sure nobody diddled with it while I was sleeping.
2330 BUG_ON(!list_empty(&worker->scheduled));
2333 * When control reaches this point, we're guaranteed to have
2334 * at least one idle worker or that someone else has already
2335 * assumed the manager role.
2337 worker_clr_flags(worker, WORKER_PREP);
2340 struct work_struct *work =
2341 list_first_entry(&pool->worklist,
2342 struct work_struct, entry);
2344 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2345 /* optimization path, not strictly necessary */
2346 process_one_work(worker, work);
2347 if (unlikely(!list_empty(&worker->scheduled)))
2348 process_scheduled_works(worker);
2350 move_linked_works(work, &worker->scheduled, NULL);
2351 process_scheduled_works(worker);
2353 } while (keep_working(pool));
2355 worker_set_flags(worker, WORKER_PREP, false);
2357 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2361 * pool->lock is held and there's no work to process and no need to
2362 * manage, sleep. Workers are woken up only while holding
2363 * pool->lock or from local cpu, so setting the current state
2364 * before releasing pool->lock is enough to prevent losing any
2367 worker_enter_idle(worker);
2368 __set_current_state(TASK_INTERRUPTIBLE);
2369 spin_unlock_irq(&pool->lock);
2375 * rescuer_thread - the rescuer thread function
2378 * Workqueue rescuer thread function. There's one rescuer for each
2379 * workqueue which has WQ_RESCUER set.
2381 * Regular work processing on a gcwq may block trying to create a new
2382 * worker which uses GFP_KERNEL allocation which has slight chance of
2383 * developing into deadlock if some works currently on the same queue
2384 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2385 * the problem rescuer solves.
2387 * When such condition is possible, the gcwq summons rescuers of all
2388 * workqueues which have works queued on the gcwq and let them process
2389 * those works so that forward progress can be guaranteed.
2391 * This should happen rarely.
2393 static int rescuer_thread(void *__rescuer)
2395 struct worker *rescuer = __rescuer;
2396 struct workqueue_struct *wq = rescuer->rescue_wq;
2397 struct list_head *scheduled = &rescuer->scheduled;
2398 bool is_unbound = wq->flags & WQ_UNBOUND;
2401 set_user_nice(current, RESCUER_NICE_LEVEL);
2404 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2405 * doesn't participate in concurrency management.
2407 rescuer->task->flags |= PF_WQ_WORKER;
2409 set_current_state(TASK_INTERRUPTIBLE);
2411 if (kthread_should_stop()) {
2412 __set_current_state(TASK_RUNNING);
2413 rescuer->task->flags &= ~PF_WQ_WORKER;
2418 * See whether any cpu is asking for help. Unbounded
2419 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2421 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2422 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2423 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2424 struct worker_pool *pool = cwq->pool;
2425 struct work_struct *work, *n;
2427 __set_current_state(TASK_RUNNING);
2428 mayday_clear_cpu(cpu, wq->mayday_mask);
2430 /* migrate to the target cpu if possible */
2431 rescuer->pool = pool;
2432 worker_maybe_bind_and_lock(rescuer);
2435 * Slurp in all works issued via this workqueue and
2438 BUG_ON(!list_empty(&rescuer->scheduled));
2439 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2440 if (get_work_cwq(work) == cwq)
2441 move_linked_works(work, scheduled, &n);
2443 process_scheduled_works(rescuer);
2446 * Leave this pool. If keep_working() is %true, notify a
2447 * regular worker; otherwise, we end up with 0 concurrency
2448 * and stalling the execution.
2450 if (keep_working(pool))
2451 wake_up_worker(pool);
2453 spin_unlock_irq(&pool->lock);
2456 /* rescuers should never participate in concurrency management */
2457 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2463 struct work_struct work;
2464 struct completion done;
2467 static void wq_barrier_func(struct work_struct *work)
2469 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2470 complete(&barr->done);
2474 * insert_wq_barrier - insert a barrier work
2475 * @cwq: cwq to insert barrier into
2476 * @barr: wq_barrier to insert
2477 * @target: target work to attach @barr to
2478 * @worker: worker currently executing @target, NULL if @target is not executing
2480 * @barr is linked to @target such that @barr is completed only after
2481 * @target finishes execution. Please note that the ordering
2482 * guarantee is observed only with respect to @target and on the local
2485 * Currently, a queued barrier can't be canceled. This is because
2486 * try_to_grab_pending() can't determine whether the work to be
2487 * grabbed is at the head of the queue and thus can't clear LINKED
2488 * flag of the previous work while there must be a valid next work
2489 * after a work with LINKED flag set.
2491 * Note that when @worker is non-NULL, @target may be modified
2492 * underneath us, so we can't reliably determine cwq from @target.
2495 * spin_lock_irq(pool->lock).
2497 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2498 struct wq_barrier *barr,
2499 struct work_struct *target, struct worker *worker)
2501 struct list_head *head;
2502 unsigned int linked = 0;
2505 * debugobject calls are safe here even with pool->lock locked
2506 * as we know for sure that this will not trigger any of the
2507 * checks and call back into the fixup functions where we
2510 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2511 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2512 init_completion(&barr->done);
2515 * If @target is currently being executed, schedule the
2516 * barrier to the worker; otherwise, put it after @target.
2519 head = worker->scheduled.next;
2521 unsigned long *bits = work_data_bits(target);
2523 head = target->entry.next;
2524 /* there can already be other linked works, inherit and set */
2525 linked = *bits & WORK_STRUCT_LINKED;
2526 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2529 debug_work_activate(&barr->work);
2530 insert_work(cwq, &barr->work, head,
2531 work_color_to_flags(WORK_NO_COLOR) | linked);
2535 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2536 * @wq: workqueue being flushed
2537 * @flush_color: new flush color, < 0 for no-op
2538 * @work_color: new work color, < 0 for no-op
2540 * Prepare cwqs for workqueue flushing.
2542 * If @flush_color is non-negative, flush_color on all cwqs should be
2543 * -1. If no cwq has in-flight commands at the specified color, all
2544 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2545 * has in flight commands, its cwq->flush_color is set to
2546 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2547 * wakeup logic is armed and %true is returned.
2549 * The caller should have initialized @wq->first_flusher prior to
2550 * calling this function with non-negative @flush_color. If
2551 * @flush_color is negative, no flush color update is done and %false
2554 * If @work_color is non-negative, all cwqs should have the same
2555 * work_color which is previous to @work_color and all will be
2556 * advanced to @work_color.
2559 * mutex_lock(wq->flush_mutex).
2562 * %true if @flush_color >= 0 and there's something to flush. %false
2565 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2566 int flush_color, int work_color)
2571 if (flush_color >= 0) {
2572 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2573 atomic_set(&wq->nr_cwqs_to_flush, 1);
2576 for_each_cwq_cpu(cpu, wq) {
2577 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2578 struct worker_pool *pool = cwq->pool;
2580 spin_lock_irq(&pool->lock);
2582 if (flush_color >= 0) {
2583 BUG_ON(cwq->flush_color != -1);
2585 if (cwq->nr_in_flight[flush_color]) {
2586 cwq->flush_color = flush_color;
2587 atomic_inc(&wq->nr_cwqs_to_flush);
2592 if (work_color >= 0) {
2593 BUG_ON(work_color != work_next_color(cwq->work_color));
2594 cwq->work_color = work_color;
2597 spin_unlock_irq(&pool->lock);
2600 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2601 complete(&wq->first_flusher->done);
2607 * flush_workqueue - ensure that any scheduled work has run to completion.
2608 * @wq: workqueue to flush
2610 * Forces execution of the workqueue and blocks until its completion.
2611 * This is typically used in driver shutdown handlers.
2613 * We sleep until all works which were queued on entry have been handled,
2614 * but we are not livelocked by new incoming ones.
2616 void flush_workqueue(struct workqueue_struct *wq)
2618 struct wq_flusher this_flusher = {
2619 .list = LIST_HEAD_INIT(this_flusher.list),
2621 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2625 lock_map_acquire(&wq->lockdep_map);
2626 lock_map_release(&wq->lockdep_map);
2628 mutex_lock(&wq->flush_mutex);
2631 * Start-to-wait phase
2633 next_color = work_next_color(wq->work_color);
2635 if (next_color != wq->flush_color) {
2637 * Color space is not full. The current work_color
2638 * becomes our flush_color and work_color is advanced
2641 BUG_ON(!list_empty(&wq->flusher_overflow));
2642 this_flusher.flush_color = wq->work_color;
2643 wq->work_color = next_color;
2645 if (!wq->first_flusher) {
2646 /* no flush in progress, become the first flusher */
2647 BUG_ON(wq->flush_color != this_flusher.flush_color);
2649 wq->first_flusher = &this_flusher;
2651 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2653 /* nothing to flush, done */
2654 wq->flush_color = next_color;
2655 wq->first_flusher = NULL;
2660 BUG_ON(wq->flush_color == this_flusher.flush_color);
2661 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2662 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2666 * Oops, color space is full, wait on overflow queue.
2667 * The next flush completion will assign us
2668 * flush_color and transfer to flusher_queue.
2670 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2673 mutex_unlock(&wq->flush_mutex);
2675 wait_for_completion(&this_flusher.done);
2678 * Wake-up-and-cascade phase
2680 * First flushers are responsible for cascading flushes and
2681 * handling overflow. Non-first flushers can simply return.
2683 if (wq->first_flusher != &this_flusher)
2686 mutex_lock(&wq->flush_mutex);
2688 /* we might have raced, check again with mutex held */
2689 if (wq->first_flusher != &this_flusher)
2692 wq->first_flusher = NULL;
2694 BUG_ON(!list_empty(&this_flusher.list));
2695 BUG_ON(wq->flush_color != this_flusher.flush_color);
2698 struct wq_flusher *next, *tmp;
2700 /* complete all the flushers sharing the current flush color */
2701 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2702 if (next->flush_color != wq->flush_color)
2704 list_del_init(&next->list);
2705 complete(&next->done);
2708 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2709 wq->flush_color != work_next_color(wq->work_color));
2711 /* this flush_color is finished, advance by one */
2712 wq->flush_color = work_next_color(wq->flush_color);
2714 /* one color has been freed, handle overflow queue */
2715 if (!list_empty(&wq->flusher_overflow)) {
2717 * Assign the same color to all overflowed
2718 * flushers, advance work_color and append to
2719 * flusher_queue. This is the start-to-wait
2720 * phase for these overflowed flushers.
2722 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2723 tmp->flush_color = wq->work_color;
2725 wq->work_color = work_next_color(wq->work_color);
2727 list_splice_tail_init(&wq->flusher_overflow,
2728 &wq->flusher_queue);
2729 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2732 if (list_empty(&wq->flusher_queue)) {
2733 BUG_ON(wq->flush_color != wq->work_color);
2738 * Need to flush more colors. Make the next flusher
2739 * the new first flusher and arm cwqs.
2741 BUG_ON(wq->flush_color == wq->work_color);
2742 BUG_ON(wq->flush_color != next->flush_color);
2744 list_del_init(&next->list);
2745 wq->first_flusher = next;
2747 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2751 * Meh... this color is already done, clear first
2752 * flusher and repeat cascading.
2754 wq->first_flusher = NULL;
2758 mutex_unlock(&wq->flush_mutex);
2760 EXPORT_SYMBOL_GPL(flush_workqueue);
2763 * drain_workqueue - drain a workqueue
2764 * @wq: workqueue to drain
2766 * Wait until the workqueue becomes empty. While draining is in progress,
2767 * only chain queueing is allowed. IOW, only currently pending or running
2768 * work items on @wq can queue further work items on it. @wq is flushed
2769 * repeatedly until it becomes empty. The number of flushing is detemined
2770 * by the depth of chaining and should be relatively short. Whine if it
2773 void drain_workqueue(struct workqueue_struct *wq)
2775 unsigned int flush_cnt = 0;
2779 * __queue_work() needs to test whether there are drainers, is much
2780 * hotter than drain_workqueue() and already looks at @wq->flags.
2781 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2783 spin_lock(&workqueue_lock);
2784 if (!wq->nr_drainers++)
2785 wq->flags |= WQ_DRAINING;
2786 spin_unlock(&workqueue_lock);
2788 flush_workqueue(wq);
2790 for_each_cwq_cpu(cpu, wq) {
2791 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2794 spin_lock_irq(&cwq->pool->lock);
2795 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2796 spin_unlock_irq(&cwq->pool->lock);
2801 if (++flush_cnt == 10 ||
2802 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2803 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2804 wq->name, flush_cnt);
2808 spin_lock(&workqueue_lock);
2809 if (!--wq->nr_drainers)
2810 wq->flags &= ~WQ_DRAINING;
2811 spin_unlock(&workqueue_lock);
2813 EXPORT_SYMBOL_GPL(drain_workqueue);
2815 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2817 struct worker *worker = NULL;
2818 struct worker_pool *pool;
2819 struct cpu_workqueue_struct *cwq;
2822 pool = get_work_pool(work);
2826 spin_lock_irq(&pool->lock);
2827 if (!list_empty(&work->entry)) {
2829 * See the comment near try_to_grab_pending()->smp_rmb().
2830 * If it was re-queued to a different pool under us, we
2831 * are not going to wait.
2834 cwq = get_work_cwq(work);
2835 if (unlikely(!cwq || pool != cwq->pool))
2838 worker = find_worker_executing_work(pool, work);
2841 cwq = worker->current_cwq;
2844 insert_wq_barrier(cwq, barr, work, worker);
2845 spin_unlock_irq(&pool->lock);
2848 * If @max_active is 1 or rescuer is in use, flushing another work
2849 * item on the same workqueue may lead to deadlock. Make sure the
2850 * flusher is not running on the same workqueue by verifying write
2853 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2854 lock_map_acquire(&cwq->wq->lockdep_map);
2856 lock_map_acquire_read(&cwq->wq->lockdep_map);
2857 lock_map_release(&cwq->wq->lockdep_map);
2861 spin_unlock_irq(&pool->lock);
2866 * flush_work - wait for a work to finish executing the last queueing instance
2867 * @work: the work to flush
2869 * Wait until @work has finished execution. @work is guaranteed to be idle
2870 * on return if it hasn't been requeued since flush started.
2873 * %true if flush_work() waited for the work to finish execution,
2874 * %false if it was already idle.
2876 bool flush_work(struct work_struct *work)
2878 struct wq_barrier barr;
2880 lock_map_acquire(&work->lockdep_map);
2881 lock_map_release(&work->lockdep_map);
2883 if (start_flush_work(work, &barr)) {
2884 wait_for_completion(&barr.done);
2885 destroy_work_on_stack(&barr.work);
2891 EXPORT_SYMBOL_GPL(flush_work);
2893 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2895 unsigned long flags;
2899 ret = try_to_grab_pending(work, is_dwork, &flags);
2901 * If someone else is canceling, wait for the same event it
2902 * would be waiting for before retrying.
2904 if (unlikely(ret == -ENOENT))
2906 } while (unlikely(ret < 0));
2908 /* tell other tasks trying to grab @work to back off */
2909 mark_work_canceling(work);
2910 local_irq_restore(flags);
2913 clear_work_data(work);
2918 * cancel_work_sync - cancel a work and wait for it to finish
2919 * @work: the work to cancel
2921 * Cancel @work and wait for its execution to finish. This function
2922 * can be used even if the work re-queues itself or migrates to
2923 * another workqueue. On return from this function, @work is
2924 * guaranteed to be not pending or executing on any CPU.
2926 * cancel_work_sync(&delayed_work->work) must not be used for
2927 * delayed_work's. Use cancel_delayed_work_sync() instead.
2929 * The caller must ensure that the workqueue on which @work was last
2930 * queued can't be destroyed before this function returns.
2933 * %true if @work was pending, %false otherwise.
2935 bool cancel_work_sync(struct work_struct *work)
2937 return __cancel_work_timer(work, false);
2939 EXPORT_SYMBOL_GPL(cancel_work_sync);
2942 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2943 * @dwork: the delayed work to flush
2945 * Delayed timer is cancelled and the pending work is queued for
2946 * immediate execution. Like flush_work(), this function only
2947 * considers the last queueing instance of @dwork.
2950 * %true if flush_work() waited for the work to finish execution,
2951 * %false if it was already idle.
2953 bool flush_delayed_work(struct delayed_work *dwork)
2955 local_irq_disable();
2956 if (del_timer_sync(&dwork->timer))
2957 __queue_work(dwork->cpu,
2958 get_work_cwq(&dwork->work)->wq, &dwork->work);
2960 return flush_work(&dwork->work);
2962 EXPORT_SYMBOL(flush_delayed_work);
2965 * cancel_delayed_work - cancel a delayed work
2966 * @dwork: delayed_work to cancel
2968 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2969 * and canceled; %false if wasn't pending. Note that the work callback
2970 * function may still be running on return, unless it returns %true and the
2971 * work doesn't re-arm itself. Explicitly flush or use
2972 * cancel_delayed_work_sync() to wait on it.
2974 * This function is safe to call from any context including IRQ handler.
2976 bool cancel_delayed_work(struct delayed_work *dwork)
2978 unsigned long flags;
2982 ret = try_to_grab_pending(&dwork->work, true, &flags);
2983 } while (unlikely(ret == -EAGAIN));
2985 if (unlikely(ret < 0))
2988 set_work_pool_and_clear_pending(&dwork->work,
2989 get_work_pool_id(&dwork->work));
2990 local_irq_restore(flags);
2993 EXPORT_SYMBOL(cancel_delayed_work);
2996 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2997 * @dwork: the delayed work cancel
2999 * This is cancel_work_sync() for delayed works.
3002 * %true if @dwork was pending, %false otherwise.
3004 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3006 return __cancel_work_timer(&dwork->work, true);
3008 EXPORT_SYMBOL(cancel_delayed_work_sync);
3011 * schedule_work_on - put work task on a specific cpu
3012 * @cpu: cpu to put the work task on
3013 * @work: job to be done
3015 * This puts a job on a specific cpu
3017 bool schedule_work_on(int cpu, struct work_struct *work)
3019 return queue_work_on(cpu, system_wq, work);
3021 EXPORT_SYMBOL(schedule_work_on);
3024 * schedule_work - put work task in global workqueue
3025 * @work: job to be done
3027 * Returns %false if @work was already on the kernel-global workqueue and
3030 * This puts a job in the kernel-global workqueue if it was not already
3031 * queued and leaves it in the same position on the kernel-global
3032 * workqueue otherwise.
3034 bool schedule_work(struct work_struct *work)
3036 return queue_work(system_wq, work);
3038 EXPORT_SYMBOL(schedule_work);
3041 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3043 * @dwork: job to be done
3044 * @delay: number of jiffies to wait
3046 * After waiting for a given time this puts a job in the kernel-global
3047 * workqueue on the specified CPU.
3049 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3050 unsigned long delay)
3052 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3054 EXPORT_SYMBOL(schedule_delayed_work_on);
3057 * schedule_delayed_work - put work task in global workqueue after delay
3058 * @dwork: job to be done
3059 * @delay: number of jiffies to wait or 0 for immediate execution
3061 * After waiting for a given time this puts a job in the kernel-global
3064 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3066 return queue_delayed_work(system_wq, dwork, delay);
3068 EXPORT_SYMBOL(schedule_delayed_work);
3071 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3072 * @func: the function to call
3074 * schedule_on_each_cpu() executes @func on each online CPU using the
3075 * system workqueue and blocks until all CPUs have completed.
3076 * schedule_on_each_cpu() is very slow.
3079 * 0 on success, -errno on failure.
3081 int schedule_on_each_cpu(work_func_t func)
3084 struct work_struct __percpu *works;
3086 works = alloc_percpu(struct work_struct);
3092 for_each_online_cpu(cpu) {
3093 struct work_struct *work = per_cpu_ptr(works, cpu);
3095 INIT_WORK(work, func);
3096 schedule_work_on(cpu, work);
3099 for_each_online_cpu(cpu)
3100 flush_work(per_cpu_ptr(works, cpu));
3108 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3110 * Forces execution of the kernel-global workqueue and blocks until its
3113 * Think twice before calling this function! It's very easy to get into
3114 * trouble if you don't take great care. Either of the following situations
3115 * will lead to deadlock:
3117 * One of the work items currently on the workqueue needs to acquire
3118 * a lock held by your code or its caller.
3120 * Your code is running in the context of a work routine.
3122 * They will be detected by lockdep when they occur, but the first might not
3123 * occur very often. It depends on what work items are on the workqueue and
3124 * what locks they need, which you have no control over.
3126 * In most situations flushing the entire workqueue is overkill; you merely
3127 * need to know that a particular work item isn't queued and isn't running.
3128 * In such cases you should use cancel_delayed_work_sync() or
3129 * cancel_work_sync() instead.
3131 void flush_scheduled_work(void)
3133 flush_workqueue(system_wq);
3135 EXPORT_SYMBOL(flush_scheduled_work);
3138 * execute_in_process_context - reliably execute the routine with user context
3139 * @fn: the function to execute
3140 * @ew: guaranteed storage for the execute work structure (must
3141 * be available when the work executes)
3143 * Executes the function immediately if process context is available,
3144 * otherwise schedules the function for delayed execution.
3146 * Returns: 0 - function was executed
3147 * 1 - function was scheduled for execution
3149 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3151 if (!in_interrupt()) {
3156 INIT_WORK(&ew->work, fn);
3157 schedule_work(&ew->work);
3161 EXPORT_SYMBOL_GPL(execute_in_process_context);
3163 int keventd_up(void)
3165 return system_wq != NULL;
3168 static int alloc_cwqs(struct workqueue_struct *wq)
3171 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3172 * Make sure that the alignment isn't lower than that of
3173 * unsigned long long.
3175 const size_t size = sizeof(struct cpu_workqueue_struct);
3176 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3177 __alignof__(unsigned long long));
3179 if (!(wq->flags & WQ_UNBOUND))
3180 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3185 * Allocate enough room to align cwq and put an extra
3186 * pointer at the end pointing back to the originally
3187 * allocated pointer which will be used for free.
3189 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3191 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3192 *(void **)(wq->cpu_wq.single + 1) = ptr;
3196 /* just in case, make sure it's actually aligned */
3197 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3198 return wq->cpu_wq.v ? 0 : -ENOMEM;
3201 static void free_cwqs(struct workqueue_struct *wq)
3203 if (!(wq->flags & WQ_UNBOUND))
3204 free_percpu(wq->cpu_wq.pcpu);
3205 else if (wq->cpu_wq.single) {
3206 /* the pointer to free is stored right after the cwq */
3207 kfree(*(void **)(wq->cpu_wq.single + 1));
3211 static int wq_clamp_max_active(int max_active, unsigned int flags,
3214 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3216 if (max_active < 1 || max_active > lim)
3217 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3218 max_active, name, 1, lim);
3220 return clamp_val(max_active, 1, lim);
3223 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3226 struct lock_class_key *key,
3227 const char *lock_name, ...)
3229 va_list args, args1;
3230 struct workqueue_struct *wq;
3234 /* determine namelen, allocate wq and format name */
3235 va_start(args, lock_name);
3236 va_copy(args1, args);
3237 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3239 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3243 vsnprintf(wq->name, namelen, fmt, args1);
3248 * Workqueues which may be used during memory reclaim should
3249 * have a rescuer to guarantee forward progress.
3251 if (flags & WQ_MEM_RECLAIM)
3252 flags |= WQ_RESCUER;
3254 max_active = max_active ?: WQ_DFL_ACTIVE;
3255 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3259 wq->saved_max_active = max_active;
3260 mutex_init(&wq->flush_mutex);
3261 atomic_set(&wq->nr_cwqs_to_flush, 0);
3262 INIT_LIST_HEAD(&wq->flusher_queue);
3263 INIT_LIST_HEAD(&wq->flusher_overflow);
3265 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3266 INIT_LIST_HEAD(&wq->list);
3268 if (alloc_cwqs(wq) < 0)
3271 for_each_cwq_cpu(cpu, wq) {
3272 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3273 struct global_cwq *gcwq = get_gcwq(cpu);
3274 int pool_idx = (bool)(flags & WQ_HIGHPRI);
3276 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3277 cwq->pool = &gcwq->pools[pool_idx];
3279 cwq->flush_color = -1;
3280 cwq->max_active = max_active;
3281 INIT_LIST_HEAD(&cwq->delayed_works);
3284 if (flags & WQ_RESCUER) {
3285 struct worker *rescuer;
3287 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3290 wq->rescuer = rescuer = alloc_worker();
3294 rescuer->rescue_wq = wq;
3295 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3297 if (IS_ERR(rescuer->task))
3300 rescuer->task->flags |= PF_THREAD_BOUND;
3301 wake_up_process(rescuer->task);
3305 * workqueue_lock protects global freeze state and workqueues
3306 * list. Grab it, set max_active accordingly and add the new
3307 * workqueue to workqueues list.
3309 spin_lock(&workqueue_lock);
3311 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3312 for_each_cwq_cpu(cpu, wq)
3313 get_cwq(cpu, wq)->max_active = 0;
3315 list_add(&wq->list, &workqueues);
3317 spin_unlock(&workqueue_lock);
3323 free_mayday_mask(wq->mayday_mask);
3329 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3332 * destroy_workqueue - safely terminate a workqueue
3333 * @wq: target workqueue
3335 * Safely destroy a workqueue. All work currently pending will be done first.
3337 void destroy_workqueue(struct workqueue_struct *wq)
3341 /* drain it before proceeding with destruction */
3342 drain_workqueue(wq);
3345 * wq list is used to freeze wq, remove from list after
3346 * flushing is complete in case freeze races us.
3348 spin_lock(&workqueue_lock);
3349 list_del(&wq->list);
3350 spin_unlock(&workqueue_lock);
3353 for_each_cwq_cpu(cpu, wq) {
3354 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3357 for (i = 0; i < WORK_NR_COLORS; i++)
3358 BUG_ON(cwq->nr_in_flight[i]);
3359 BUG_ON(cwq->nr_active);
3360 BUG_ON(!list_empty(&cwq->delayed_works));
3363 if (wq->flags & WQ_RESCUER) {
3364 kthread_stop(wq->rescuer->task);
3365 free_mayday_mask(wq->mayday_mask);
3372 EXPORT_SYMBOL_GPL(destroy_workqueue);
3375 * cwq_set_max_active - adjust max_active of a cwq
3376 * @cwq: target cpu_workqueue_struct
3377 * @max_active: new max_active value.
3379 * Set @cwq->max_active to @max_active and activate delayed works if
3383 * spin_lock_irq(pool->lock).
3385 static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
3387 cwq->max_active = max_active;
3389 while (!list_empty(&cwq->delayed_works) &&
3390 cwq->nr_active < cwq->max_active)
3391 cwq_activate_first_delayed(cwq);
3395 * workqueue_set_max_active - adjust max_active of a workqueue
3396 * @wq: target workqueue
3397 * @max_active: new max_active value.
3399 * Set max_active of @wq to @max_active.
3402 * Don't call from IRQ context.
3404 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3408 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3410 spin_lock(&workqueue_lock);
3412 wq->saved_max_active = max_active;
3414 for_each_cwq_cpu(cpu, wq) {
3415 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3416 struct worker_pool *pool = cwq->pool;
3418 spin_lock_irq(&pool->lock);
3420 if (!(wq->flags & WQ_FREEZABLE) ||
3421 !(pool->flags & POOL_FREEZING))
3422 cwq_set_max_active(cwq, max_active);
3424 spin_unlock_irq(&pool->lock);
3427 spin_unlock(&workqueue_lock);
3429 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3432 * workqueue_congested - test whether a workqueue is congested
3433 * @cpu: CPU in question
3434 * @wq: target workqueue
3436 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3437 * no synchronization around this function and the test result is
3438 * unreliable and only useful as advisory hints or for debugging.
3441 * %true if congested, %false otherwise.
3443 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3445 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3447 return !list_empty(&cwq->delayed_works);
3449 EXPORT_SYMBOL_GPL(workqueue_congested);
3452 * work_busy - test whether a work is currently pending or running
3453 * @work: the work to be tested
3455 * Test whether @work is currently pending or running. There is no
3456 * synchronization around this function and the test result is
3457 * unreliable and only useful as advisory hints or for debugging.
3458 * Especially for reentrant wqs, the pending state might hide the
3462 * OR'd bitmask of WORK_BUSY_* bits.
3464 unsigned int work_busy(struct work_struct *work)
3466 struct worker_pool *pool = get_work_pool(work);
3467 unsigned long flags;
3468 unsigned int ret = 0;
3473 spin_lock_irqsave(&pool->lock, flags);
3475 if (work_pending(work))
3476 ret |= WORK_BUSY_PENDING;
3477 if (find_worker_executing_work(pool, work))
3478 ret |= WORK_BUSY_RUNNING;
3480 spin_unlock_irqrestore(&pool->lock, flags);
3484 EXPORT_SYMBOL_GPL(work_busy);
3489 * There are two challenges in supporting CPU hotplug. Firstly, there
3490 * are a lot of assumptions on strong associations among work, cwq and
3491 * gcwq which make migrating pending and scheduled works very
3492 * difficult to implement without impacting hot paths. Secondly,
3493 * worker pools serve mix of short, long and very long running works making
3494 * blocked draining impractical.
3496 * This is solved by allowing the pools to be disassociated from the CPU
3497 * running as an unbound one and allowing it to be reattached later if the
3498 * cpu comes back online.
3501 static void gcwq_unbind_fn(struct work_struct *work)
3503 struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3504 struct worker_pool *pool;
3505 struct worker *worker;
3506 struct hlist_node *pos;
3509 for_each_worker_pool(pool, gcwq) {
3510 BUG_ON(pool->cpu != smp_processor_id());
3512 mutex_lock(&pool->assoc_mutex);
3513 spin_lock_irq(&pool->lock);
3516 * We've claimed all manager positions. Make all workers
3517 * unbound and set DISASSOCIATED. Before this, all workers
3518 * except for the ones which are still executing works from
3519 * before the last CPU down must be on the cpu. After
3520 * this, they may become diasporas.
3522 list_for_each_entry(worker, &pool->idle_list, entry)
3523 worker->flags |= WORKER_UNBOUND;
3525 for_each_busy_worker(worker, i, pos, pool)
3526 worker->flags |= WORKER_UNBOUND;
3528 pool->flags |= POOL_DISASSOCIATED;
3530 spin_unlock_irq(&pool->lock);
3531 mutex_unlock(&pool->assoc_mutex);
3535 * Call schedule() so that we cross rq->lock and thus can guarantee
3536 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3537 * as scheduler callbacks may be invoked from other cpus.
3542 * Sched callbacks are disabled now. Zap nr_running. After this,
3543 * nr_running stays zero and need_more_worker() and keep_working()
3544 * are always true as long as the worklist is not empty. @gcwq now
3545 * behaves as unbound (in terms of concurrency management) gcwq
3546 * which is served by workers tied to the CPU.
3548 * On return from this function, the current worker would trigger
3549 * unbound chain execution of pending work items if other workers
3552 for_each_worker_pool(pool, gcwq)
3553 atomic_set(get_pool_nr_running(pool), 0);
3557 * Workqueues should be brought up before normal priority CPU notifiers.
3558 * This will be registered high priority CPU notifier.
3560 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3561 unsigned long action,
3564 unsigned int cpu = (unsigned long)hcpu;
3565 struct global_cwq *gcwq = get_gcwq(cpu);
3566 struct worker_pool *pool;
3568 switch (action & ~CPU_TASKS_FROZEN) {
3569 case CPU_UP_PREPARE:
3570 for_each_worker_pool(pool, gcwq) {
3571 struct worker *worker;
3573 if (pool->nr_workers)
3576 worker = create_worker(pool);
3580 spin_lock_irq(&pool->lock);
3581 start_worker(worker);
3582 spin_unlock_irq(&pool->lock);
3586 case CPU_DOWN_FAILED:
3588 for_each_worker_pool(pool, gcwq) {
3589 mutex_lock(&pool->assoc_mutex);
3590 spin_lock_irq(&pool->lock);
3592 pool->flags &= ~POOL_DISASSOCIATED;
3593 rebind_workers(pool);
3595 spin_unlock_irq(&pool->lock);
3596 mutex_unlock(&pool->assoc_mutex);
3604 * Workqueues should be brought down after normal priority CPU notifiers.
3605 * This will be registered as low priority CPU notifier.
3607 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3608 unsigned long action,
3611 unsigned int cpu = (unsigned long)hcpu;
3612 struct work_struct unbind_work;
3614 switch (action & ~CPU_TASKS_FROZEN) {
3615 case CPU_DOWN_PREPARE:
3616 /* unbinding should happen on the local CPU */
3617 INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3618 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3619 flush_work(&unbind_work);
3627 struct work_for_cpu {
3628 struct work_struct work;
3634 static void work_for_cpu_fn(struct work_struct *work)
3636 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3638 wfc->ret = wfc->fn(wfc->arg);
3642 * work_on_cpu - run a function in user context on a particular cpu
3643 * @cpu: the cpu to run on
3644 * @fn: the function to run
3645 * @arg: the function arg
3647 * This will return the value @fn returns.
3648 * It is up to the caller to ensure that the cpu doesn't go offline.
3649 * The caller must not hold any locks which would prevent @fn from completing.
3651 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3653 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3655 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3656 schedule_work_on(cpu, &wfc.work);
3657 flush_work(&wfc.work);
3660 EXPORT_SYMBOL_GPL(work_on_cpu);
3661 #endif /* CONFIG_SMP */
3663 #ifdef CONFIG_FREEZER
3666 * freeze_workqueues_begin - begin freezing workqueues
3668 * Start freezing workqueues. After this function returns, all freezable
3669 * workqueues will queue new works to their frozen_works list instead of
3673 * Grabs and releases workqueue_lock and pool->lock's.
3675 void freeze_workqueues_begin(void)
3679 spin_lock(&workqueue_lock);
3681 BUG_ON(workqueue_freezing);
3682 workqueue_freezing = true;
3684 for_each_gcwq_cpu(cpu) {
3685 struct global_cwq *gcwq = get_gcwq(cpu);
3686 struct worker_pool *pool;
3687 struct workqueue_struct *wq;
3689 for_each_worker_pool(pool, gcwq) {
3690 spin_lock_irq(&pool->lock);
3692 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
3693 pool->flags |= POOL_FREEZING;
3695 list_for_each_entry(wq, &workqueues, list) {
3696 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3698 if (cwq && cwq->pool == pool &&
3699 (wq->flags & WQ_FREEZABLE))
3700 cwq->max_active = 0;
3703 spin_unlock_irq(&pool->lock);
3707 spin_unlock(&workqueue_lock);
3711 * freeze_workqueues_busy - are freezable workqueues still busy?
3713 * Check whether freezing is complete. This function must be called
3714 * between freeze_workqueues_begin() and thaw_workqueues().
3717 * Grabs and releases workqueue_lock.
3720 * %true if some freezable workqueues are still busy. %false if freezing
3723 bool freeze_workqueues_busy(void)
3728 spin_lock(&workqueue_lock);
3730 BUG_ON(!workqueue_freezing);
3732 for_each_gcwq_cpu(cpu) {
3733 struct workqueue_struct *wq;
3735 * nr_active is monotonically decreasing. It's safe
3736 * to peek without lock.
3738 list_for_each_entry(wq, &workqueues, list) {
3739 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3741 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3744 BUG_ON(cwq->nr_active < 0);
3745 if (cwq->nr_active) {
3752 spin_unlock(&workqueue_lock);
3757 * thaw_workqueues - thaw workqueues
3759 * Thaw workqueues. Normal queueing is restored and all collected
3760 * frozen works are transferred to their respective gcwq worklists.
3763 * Grabs and releases workqueue_lock and pool->lock's.
3765 void thaw_workqueues(void)
3769 spin_lock(&workqueue_lock);
3771 if (!workqueue_freezing)
3774 for_each_gcwq_cpu(cpu) {
3775 struct global_cwq *gcwq = get_gcwq(cpu);
3776 struct worker_pool *pool;
3777 struct workqueue_struct *wq;
3779 for_each_worker_pool(pool, gcwq) {
3780 spin_lock_irq(&pool->lock);
3782 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
3783 pool->flags &= ~POOL_FREEZING;
3785 list_for_each_entry(wq, &workqueues, list) {
3786 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3788 if (!cwq || cwq->pool != pool ||
3789 !(wq->flags & WQ_FREEZABLE))
3792 /* restore max_active and repopulate worklist */
3793 cwq_set_max_active(cwq, wq->saved_max_active);
3796 wake_up_worker(pool);
3798 spin_unlock_irq(&pool->lock);
3802 workqueue_freezing = false;
3804 spin_unlock(&workqueue_lock);
3806 #endif /* CONFIG_FREEZER */
3808 static int __init init_workqueues(void)
3812 /* make sure we have enough bits for OFFQ pool ID */
3813 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
3814 WORK_CPU_LAST * NR_STD_WORKER_POOLS);
3816 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3817 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3819 /* initialize gcwqs */
3820 for_each_gcwq_cpu(cpu) {
3821 struct global_cwq *gcwq = get_gcwq(cpu);
3822 struct worker_pool *pool;
3824 for_each_worker_pool(pool, gcwq) {
3826 spin_lock_init(&pool->lock);
3828 pool->flags |= POOL_DISASSOCIATED;
3829 INIT_LIST_HEAD(&pool->worklist);
3830 INIT_LIST_HEAD(&pool->idle_list);
3831 hash_init(pool->busy_hash);
3833 init_timer_deferrable(&pool->idle_timer);
3834 pool->idle_timer.function = idle_worker_timeout;
3835 pool->idle_timer.data = (unsigned long)pool;
3837 setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3838 (unsigned long)pool);
3840 mutex_init(&pool->assoc_mutex);
3841 ida_init(&pool->worker_ida);
3844 BUG_ON(worker_pool_assign_id(pool));
3848 /* create the initial worker */
3849 for_each_online_gcwq_cpu(cpu) {
3850 struct global_cwq *gcwq = get_gcwq(cpu);
3851 struct worker_pool *pool;
3853 for_each_worker_pool(pool, gcwq) {
3854 struct worker *worker;
3856 if (cpu != WORK_CPU_UNBOUND)
3857 pool->flags &= ~POOL_DISASSOCIATED;
3859 worker = create_worker(pool);
3861 spin_lock_irq(&pool->lock);
3862 start_worker(worker);
3863 spin_unlock_irq(&pool->lock);
3867 system_wq = alloc_workqueue("events", 0, 0);
3868 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3869 system_long_wq = alloc_workqueue("events_long", 0, 0);
3870 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3871 WQ_UNBOUND_MAX_ACTIVE);
3872 system_freezable_wq = alloc_workqueue("events_freezable",
3874 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3875 !system_unbound_wq || !system_freezable_wq);
3878 early_initcall(init_workqueues);