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>
45 #include "workqueue_sched.h"
51 * A bound gcwq is either associated or disassociated with its CPU.
52 * While associated (!DISASSOCIATED), all workers are bound to the
53 * CPU and none has %WORKER_UNBOUND set and concurrency management
56 * While DISASSOCIATED, the cpu may be offline and all workers have
57 * %WORKER_UNBOUND set and concurrency management disabled, and may
58 * be executing on any CPU. The gcwq behaves as an unbound one.
60 * Note that DISASSOCIATED can be flipped only while holding
61 * assoc_mutex of all pools on the gcwq to avoid changing binding
62 * state while create_worker() is in progress.
64 GCWQ_DISASSOCIATED = 1 << 0, /* cpu can't serve workers */
65 GCWQ_FREEZING = 1 << 1, /* freeze in progress */
68 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
69 POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */
72 WORKER_STARTED = 1 << 0, /* started */
73 WORKER_DIE = 1 << 1, /* die die die */
74 WORKER_IDLE = 1 << 2, /* is idle */
75 WORKER_PREP = 1 << 3, /* preparing to run works */
76 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
77 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND |
82 NR_WORKER_POOLS = 2, /* # worker pools per gcwq */
84 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
85 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
86 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
88 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
89 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
92 /* call for help after 10ms
94 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
95 CREATE_COOLDOWN = HZ, /* time to breath after fail */
98 * Rescue workers are used only on emergencies and shared by
101 RESCUER_NICE_LEVEL = -20,
102 HIGHPRI_NICE_LEVEL = -20,
106 * Structure fields follow one of the following exclusion rules.
108 * I: Modifiable by initialization/destruction paths and read-only for
111 * P: Preemption protected. Disabling preemption is enough and should
112 * only be modified and accessed from the local cpu.
114 * L: gcwq->lock protected. Access with gcwq->lock held.
116 * X: During normal operation, modification requires gcwq->lock and
117 * should be done only from local cpu. Either disabling preemption
118 * on local cpu or grabbing gcwq->lock is enough for read access.
119 * If GCWQ_DISASSOCIATED is set, it's identical to L.
121 * F: wq->flush_mutex protected.
123 * W: workqueue_lock protected.
130 * The poor guys doing the actual heavy lifting. All on-duty workers
131 * are either serving the manager role, on idle list or on busy hash.
134 /* on idle list while idle, on busy hash table while busy */
136 struct list_head entry; /* L: while idle */
137 struct hlist_node hentry; /* L: while busy */
140 struct work_struct *current_work; /* L: work being processed */
141 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
142 struct list_head scheduled; /* L: scheduled works */
143 struct task_struct *task; /* I: worker task */
144 struct worker_pool *pool; /* I: the associated pool */
145 /* 64 bytes boundary on 64bit, 32 on 32bit */
146 unsigned long last_active; /* L: last active timestamp */
147 unsigned int flags; /* X: flags */
148 int id; /* I: worker id */
150 /* for rebinding worker to CPU */
151 struct work_struct rebind_work; /* L: for busy worker */
155 struct global_cwq *gcwq; /* I: the owning gcwq */
156 unsigned int flags; /* X: flags */
158 struct list_head worklist; /* L: list of pending works */
159 int nr_workers; /* L: total number of workers */
161 /* nr_idle includes the ones off idle_list for rebinding */
162 int nr_idle; /* L: currently idle ones */
164 struct list_head idle_list; /* X: list of idle workers */
165 struct timer_list idle_timer; /* L: worker idle timeout */
166 struct timer_list mayday_timer; /* L: SOS timer for workers */
168 struct mutex assoc_mutex; /* protect GCWQ_DISASSOCIATED */
169 struct ida worker_ida; /* L: for worker IDs */
173 * Global per-cpu workqueue. There's one and only one for each cpu
174 * and all works are queued and processed here regardless of their
178 spinlock_t lock; /* the gcwq lock */
179 unsigned int cpu; /* I: the associated cpu */
180 unsigned int flags; /* L: GCWQ_* flags */
182 /* workers are chained either in busy_hash or pool idle_list */
183 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
184 /* L: hash of busy workers */
186 struct worker_pool pools[NR_WORKER_POOLS];
187 /* normal and highpri pools */
188 } ____cacheline_aligned_in_smp;
191 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
192 * work_struct->data are used for flags and thus cwqs need to be
193 * aligned at two's power of the number of flag bits.
195 struct cpu_workqueue_struct {
196 struct worker_pool *pool; /* I: the associated pool */
197 struct workqueue_struct *wq; /* I: the owning workqueue */
198 int work_color; /* L: current color */
199 int flush_color; /* L: flushing color */
200 int nr_in_flight[WORK_NR_COLORS];
201 /* L: nr of in_flight works */
202 int nr_active; /* L: nr of active works */
203 int max_active; /* L: max active works */
204 struct list_head delayed_works; /* L: delayed works */
208 * Structure used to wait for workqueue flush.
211 struct list_head list; /* F: list of flushers */
212 int flush_color; /* F: flush color waiting for */
213 struct completion done; /* flush completion */
217 * All cpumasks are assumed to be always set on UP and thus can't be
218 * used to determine whether there's something to be done.
221 typedef cpumask_var_t mayday_mask_t;
222 #define mayday_test_and_set_cpu(cpu, mask) \
223 cpumask_test_and_set_cpu((cpu), (mask))
224 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
225 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
226 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
227 #define free_mayday_mask(mask) free_cpumask_var((mask))
229 typedef unsigned long mayday_mask_t;
230 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
231 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
232 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
233 #define alloc_mayday_mask(maskp, gfp) true
234 #define free_mayday_mask(mask) do { } while (0)
238 * The externally visible workqueue abstraction is an array of
239 * per-CPU workqueues:
241 struct workqueue_struct {
242 unsigned int flags; /* W: WQ_* flags */
244 struct cpu_workqueue_struct __percpu *pcpu;
245 struct cpu_workqueue_struct *single;
247 } cpu_wq; /* I: cwq's */
248 struct list_head list; /* W: list of all workqueues */
250 struct mutex flush_mutex; /* protects wq flushing */
251 int work_color; /* F: current work color */
252 int flush_color; /* F: current flush color */
253 atomic_t nr_cwqs_to_flush; /* flush in progress */
254 struct wq_flusher *first_flusher; /* F: first flusher */
255 struct list_head flusher_queue; /* F: flush waiters */
256 struct list_head flusher_overflow; /* F: flush overflow list */
258 mayday_mask_t mayday_mask; /* cpus requesting rescue */
259 struct worker *rescuer; /* I: rescue worker */
261 int nr_drainers; /* W: drain in progress */
262 int saved_max_active; /* W: saved cwq max_active */
263 #ifdef CONFIG_LOCKDEP
264 struct lockdep_map lockdep_map;
266 char name[]; /* I: workqueue name */
269 struct workqueue_struct *system_wq __read_mostly;
270 EXPORT_SYMBOL_GPL(system_wq);
271 struct workqueue_struct *system_highpri_wq __read_mostly;
272 EXPORT_SYMBOL_GPL(system_highpri_wq);
273 struct workqueue_struct *system_long_wq __read_mostly;
274 EXPORT_SYMBOL_GPL(system_long_wq);
275 struct workqueue_struct *system_unbound_wq __read_mostly;
276 EXPORT_SYMBOL_GPL(system_unbound_wq);
277 struct workqueue_struct *system_freezable_wq __read_mostly;
278 EXPORT_SYMBOL_GPL(system_freezable_wq);
280 #define CREATE_TRACE_POINTS
281 #include <trace/events/workqueue.h>
283 #define for_each_worker_pool(pool, gcwq) \
284 for ((pool) = &(gcwq)->pools[0]; \
285 (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
287 #define for_each_busy_worker(worker, i, pos, gcwq) \
288 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
289 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
291 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
294 if (cpu < nr_cpu_ids) {
296 cpu = cpumask_next(cpu, mask);
297 if (cpu < nr_cpu_ids)
301 return WORK_CPU_UNBOUND;
303 return WORK_CPU_NONE;
306 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
307 struct workqueue_struct *wq)
309 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
315 * An extra gcwq is defined for an invalid cpu number
316 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
317 * specific CPU. The following iterators are similar to
318 * for_each_*_cpu() iterators but also considers the unbound gcwq.
320 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
321 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
322 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
323 * WORK_CPU_UNBOUND for unbound workqueues
325 #define for_each_gcwq_cpu(cpu) \
326 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
327 (cpu) < WORK_CPU_NONE; \
328 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
330 #define for_each_online_gcwq_cpu(cpu) \
331 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
332 (cpu) < WORK_CPU_NONE; \
333 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
335 #define for_each_cwq_cpu(cpu, wq) \
336 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
337 (cpu) < WORK_CPU_NONE; \
338 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
340 #ifdef CONFIG_DEBUG_OBJECTS_WORK
342 static struct debug_obj_descr work_debug_descr;
344 static void *work_debug_hint(void *addr)
346 return ((struct work_struct *) addr)->func;
350 * fixup_init is called when:
351 * - an active object is initialized
353 static int work_fixup_init(void *addr, enum debug_obj_state state)
355 struct work_struct *work = addr;
358 case ODEBUG_STATE_ACTIVE:
359 cancel_work_sync(work);
360 debug_object_init(work, &work_debug_descr);
368 * fixup_activate is called when:
369 * - an active object is activated
370 * - an unknown object is activated (might be a statically initialized object)
372 static int work_fixup_activate(void *addr, enum debug_obj_state state)
374 struct work_struct *work = addr;
378 case ODEBUG_STATE_NOTAVAILABLE:
380 * This is not really a fixup. The work struct was
381 * statically initialized. We just make sure that it
382 * is tracked in the object tracker.
384 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
385 debug_object_init(work, &work_debug_descr);
386 debug_object_activate(work, &work_debug_descr);
392 case ODEBUG_STATE_ACTIVE:
401 * fixup_free is called when:
402 * - an active object is freed
404 static int work_fixup_free(void *addr, enum debug_obj_state state)
406 struct work_struct *work = addr;
409 case ODEBUG_STATE_ACTIVE:
410 cancel_work_sync(work);
411 debug_object_free(work, &work_debug_descr);
418 static struct debug_obj_descr work_debug_descr = {
419 .name = "work_struct",
420 .debug_hint = work_debug_hint,
421 .fixup_init = work_fixup_init,
422 .fixup_activate = work_fixup_activate,
423 .fixup_free = work_fixup_free,
426 static inline void debug_work_activate(struct work_struct *work)
428 debug_object_activate(work, &work_debug_descr);
431 static inline void debug_work_deactivate(struct work_struct *work)
433 debug_object_deactivate(work, &work_debug_descr);
436 void __init_work(struct work_struct *work, int onstack)
439 debug_object_init_on_stack(work, &work_debug_descr);
441 debug_object_init(work, &work_debug_descr);
443 EXPORT_SYMBOL_GPL(__init_work);
445 void destroy_work_on_stack(struct work_struct *work)
447 debug_object_free(work, &work_debug_descr);
449 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
452 static inline void debug_work_activate(struct work_struct *work) { }
453 static inline void debug_work_deactivate(struct work_struct *work) { }
456 /* Serializes the accesses to the list of workqueues. */
457 static DEFINE_SPINLOCK(workqueue_lock);
458 static LIST_HEAD(workqueues);
459 static bool workqueue_freezing; /* W: have wqs started freezing? */
462 * The almighty global cpu workqueues. nr_running is the only field
463 * which is expected to be used frequently by other cpus via
464 * try_to_wake_up(). Put it in a separate cacheline.
466 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
467 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]);
470 * Global cpu workqueue and nr_running counter for unbound gcwq. The
471 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
472 * workers have WORKER_UNBOUND set.
474 static struct global_cwq unbound_global_cwq;
475 static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = {
476 [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
479 static int worker_thread(void *__worker);
481 static int worker_pool_pri(struct worker_pool *pool)
483 return pool - pool->gcwq->pools;
486 static struct global_cwq *get_gcwq(unsigned int cpu)
488 if (cpu != WORK_CPU_UNBOUND)
489 return &per_cpu(global_cwq, cpu);
491 return &unbound_global_cwq;
494 static atomic_t *get_pool_nr_running(struct worker_pool *pool)
496 int cpu = pool->gcwq->cpu;
497 int idx = worker_pool_pri(pool);
499 if (cpu != WORK_CPU_UNBOUND)
500 return &per_cpu(pool_nr_running, cpu)[idx];
502 return &unbound_pool_nr_running[idx];
505 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
506 struct workqueue_struct *wq)
508 if (!(wq->flags & WQ_UNBOUND)) {
509 if (likely(cpu < nr_cpu_ids))
510 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
511 } else if (likely(cpu == WORK_CPU_UNBOUND))
512 return wq->cpu_wq.single;
516 static unsigned int work_color_to_flags(int color)
518 return color << WORK_STRUCT_COLOR_SHIFT;
521 static int get_work_color(struct work_struct *work)
523 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
524 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
527 static int work_next_color(int color)
529 return (color + 1) % WORK_NR_COLORS;
533 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
534 * contain the pointer to the queued cwq. Once execution starts, the flag
535 * is cleared and the high bits contain OFFQ flags and CPU number.
537 * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
538 * and clear_work_data() can be used to set the cwq, cpu or clear
539 * work->data. These functions should only be called while the work is
540 * owned - ie. while the PENDING bit is set.
542 * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
543 * a work. gcwq is available once the work has been queued anywhere after
544 * initialization until it is sync canceled. cwq is available only while
545 * the work item is queued.
547 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
548 * canceled. While being canceled, a work item may have its PENDING set
549 * but stay off timer and worklist for arbitrarily long and nobody should
550 * try to steal the PENDING bit.
552 static inline void set_work_data(struct work_struct *work, unsigned long data,
555 BUG_ON(!work_pending(work));
556 atomic_long_set(&work->data, data | flags | work_static(work));
559 static void set_work_cwq(struct work_struct *work,
560 struct cpu_workqueue_struct *cwq,
561 unsigned long extra_flags)
563 set_work_data(work, (unsigned long)cwq,
564 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
567 static void set_work_cpu_and_clear_pending(struct work_struct *work,
571 * The following wmb is paired with the implied mb in
572 * test_and_set_bit(PENDING) and ensures all updates to @work made
573 * here are visible to and precede any updates by the next PENDING
577 set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0);
580 static void clear_work_data(struct work_struct *work)
582 smp_wmb(); /* see set_work_cpu_and_clear_pending() */
583 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
586 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
588 unsigned long data = atomic_long_read(&work->data);
590 if (data & WORK_STRUCT_CWQ)
591 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
596 static struct global_cwq *get_work_gcwq(struct work_struct *work)
598 unsigned long data = atomic_long_read(&work->data);
601 if (data & WORK_STRUCT_CWQ)
602 return ((struct cpu_workqueue_struct *)
603 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq;
605 cpu = data >> WORK_OFFQ_CPU_SHIFT;
606 if (cpu == WORK_CPU_NONE)
609 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
610 return get_gcwq(cpu);
613 static void mark_work_canceling(struct work_struct *work)
615 struct global_cwq *gcwq = get_work_gcwq(work);
616 unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE;
618 set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) | WORK_OFFQ_CANCELING,
619 WORK_STRUCT_PENDING);
622 static bool work_is_canceling(struct work_struct *work)
624 unsigned long data = atomic_long_read(&work->data);
626 return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
630 * Policy functions. These define the policies on how the global worker
631 * pools are managed. Unless noted otherwise, these functions assume that
632 * they're being called with gcwq->lock held.
635 static bool __need_more_worker(struct worker_pool *pool)
637 return !atomic_read(get_pool_nr_running(pool));
641 * Need to wake up a worker? Called from anything but currently
644 * Note that, because unbound workers never contribute to nr_running, this
645 * function will always return %true for unbound gcwq as long as the
646 * worklist isn't empty.
648 static bool need_more_worker(struct worker_pool *pool)
650 return !list_empty(&pool->worklist) && __need_more_worker(pool);
653 /* Can I start working? Called from busy but !running workers. */
654 static bool may_start_working(struct worker_pool *pool)
656 return pool->nr_idle;
659 /* Do I need to keep working? Called from currently running workers. */
660 static bool keep_working(struct worker_pool *pool)
662 atomic_t *nr_running = get_pool_nr_running(pool);
664 return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
667 /* Do we need a new worker? Called from manager. */
668 static bool need_to_create_worker(struct worker_pool *pool)
670 return need_more_worker(pool) && !may_start_working(pool);
673 /* Do I need to be the manager? */
674 static bool need_to_manage_workers(struct worker_pool *pool)
676 return need_to_create_worker(pool) ||
677 (pool->flags & POOL_MANAGE_WORKERS);
680 /* Do we have too many workers and should some go away? */
681 static bool too_many_workers(struct worker_pool *pool)
683 bool managing = pool->flags & POOL_MANAGING_WORKERS;
684 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
685 int nr_busy = pool->nr_workers - nr_idle;
688 * nr_idle and idle_list may disagree if idle rebinding is in
689 * progress. Never return %true if idle_list is empty.
691 if (list_empty(&pool->idle_list))
694 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
701 /* Return the first worker. Safe with preemption disabled */
702 static struct worker *first_worker(struct worker_pool *pool)
704 if (unlikely(list_empty(&pool->idle_list)))
707 return list_first_entry(&pool->idle_list, struct worker, entry);
711 * wake_up_worker - wake up an idle worker
712 * @pool: worker pool to wake worker from
714 * Wake up the first idle worker of @pool.
717 * spin_lock_irq(gcwq->lock).
719 static void wake_up_worker(struct worker_pool *pool)
721 struct worker *worker = first_worker(pool);
724 wake_up_process(worker->task);
728 * wq_worker_waking_up - a worker is waking up
729 * @task: task waking up
730 * @cpu: CPU @task is waking up to
732 * This function is called during try_to_wake_up() when a worker is
736 * spin_lock_irq(rq->lock)
738 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
740 struct worker *worker = kthread_data(task);
742 if (!(worker->flags & WORKER_NOT_RUNNING))
743 atomic_inc(get_pool_nr_running(worker->pool));
747 * wq_worker_sleeping - a worker is going to sleep
748 * @task: task going to sleep
749 * @cpu: CPU in question, must be the current CPU number
751 * This function is called during schedule() when a busy worker is
752 * going to sleep. Worker on the same cpu can be woken up by
753 * returning pointer to its task.
756 * spin_lock_irq(rq->lock)
759 * Worker task on @cpu to wake up, %NULL if none.
761 struct task_struct *wq_worker_sleeping(struct task_struct *task,
764 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
765 struct worker_pool *pool = worker->pool;
766 atomic_t *nr_running = get_pool_nr_running(pool);
768 if (worker->flags & WORKER_NOT_RUNNING)
771 /* this can only happen on the local cpu */
772 BUG_ON(cpu != raw_smp_processor_id());
775 * The counterpart of the following dec_and_test, implied mb,
776 * worklist not empty test sequence is in insert_work().
777 * Please read comment there.
779 * NOT_RUNNING is clear. This means that we're bound to and
780 * running on the local cpu w/ rq lock held and preemption
781 * disabled, which in turn means that none else could be
782 * manipulating idle_list, so dereferencing idle_list without gcwq
785 if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
786 to_wakeup = first_worker(pool);
787 return to_wakeup ? to_wakeup->task : NULL;
791 * worker_set_flags - set worker flags and adjust nr_running accordingly
793 * @flags: flags to set
794 * @wakeup: wakeup an idle worker if necessary
796 * Set @flags in @worker->flags and adjust nr_running accordingly. If
797 * nr_running becomes zero and @wakeup is %true, an idle worker is
801 * spin_lock_irq(gcwq->lock)
803 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
806 struct worker_pool *pool = worker->pool;
808 WARN_ON_ONCE(worker->task != current);
811 * If transitioning into NOT_RUNNING, adjust nr_running and
812 * wake up an idle worker as necessary if requested by
815 if ((flags & WORKER_NOT_RUNNING) &&
816 !(worker->flags & WORKER_NOT_RUNNING)) {
817 atomic_t *nr_running = get_pool_nr_running(pool);
820 if (atomic_dec_and_test(nr_running) &&
821 !list_empty(&pool->worklist))
822 wake_up_worker(pool);
824 atomic_dec(nr_running);
827 worker->flags |= flags;
831 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
833 * @flags: flags to clear
835 * Clear @flags in @worker->flags and adjust nr_running accordingly.
838 * spin_lock_irq(gcwq->lock)
840 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
842 struct worker_pool *pool = worker->pool;
843 unsigned int oflags = worker->flags;
845 WARN_ON_ONCE(worker->task != current);
847 worker->flags &= ~flags;
850 * If transitioning out of NOT_RUNNING, increment nr_running. Note
851 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
852 * of multiple flags, not a single flag.
854 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
855 if (!(worker->flags & WORKER_NOT_RUNNING))
856 atomic_inc(get_pool_nr_running(pool));
860 * busy_worker_head - return the busy hash head for a work
861 * @gcwq: gcwq of interest
862 * @work: work to be hashed
864 * Return hash head of @gcwq for @work.
867 * spin_lock_irq(gcwq->lock).
870 * Pointer to the hash head.
872 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
873 struct work_struct *work)
875 const int base_shift = ilog2(sizeof(struct work_struct));
876 unsigned long v = (unsigned long)work;
878 /* simple shift and fold hash, do we need something better? */
880 v += v >> BUSY_WORKER_HASH_ORDER;
881 v &= BUSY_WORKER_HASH_MASK;
883 return &gcwq->busy_hash[v];
887 * __find_worker_executing_work - find worker which is executing a work
888 * @gcwq: gcwq of interest
889 * @bwh: hash head as returned by busy_worker_head()
890 * @work: work to find worker for
892 * Find a worker which is executing @work on @gcwq. @bwh should be
893 * the hash head obtained by calling busy_worker_head() with the same
897 * spin_lock_irq(gcwq->lock).
900 * Pointer to worker which is executing @work if found, NULL
903 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
904 struct hlist_head *bwh,
905 struct work_struct *work)
907 struct worker *worker;
908 struct hlist_node *tmp;
910 hlist_for_each_entry(worker, tmp, bwh, hentry)
911 if (worker->current_work == work)
917 * find_worker_executing_work - find worker which is executing a work
918 * @gcwq: gcwq of interest
919 * @work: work to find worker for
921 * Find a worker which is executing @work on @gcwq. This function is
922 * identical to __find_worker_executing_work() except that this
923 * function calculates @bwh itself.
926 * spin_lock_irq(gcwq->lock).
929 * Pointer to worker which is executing @work if found, NULL
932 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
933 struct work_struct *work)
935 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
940 * move_linked_works - move linked works to a list
941 * @work: start of series of works to be scheduled
942 * @head: target list to append @work to
943 * @nextp: out paramter for nested worklist walking
945 * Schedule linked works starting from @work to @head. Work series to
946 * be scheduled starts at @work and includes any consecutive work with
947 * WORK_STRUCT_LINKED set in its predecessor.
949 * If @nextp is not NULL, it's updated to point to the next work of
950 * the last scheduled work. This allows move_linked_works() to be
951 * nested inside outer list_for_each_entry_safe().
954 * spin_lock_irq(gcwq->lock).
956 static void move_linked_works(struct work_struct *work, struct list_head *head,
957 struct work_struct **nextp)
959 struct work_struct *n;
962 * Linked worklist will always end before the end of the list,
963 * use NULL for list head.
965 list_for_each_entry_safe_from(work, n, NULL, entry) {
966 list_move_tail(&work->entry, head);
967 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
972 * If we're already inside safe list traversal and have moved
973 * multiple works to the scheduled queue, the next position
974 * needs to be updated.
980 static void cwq_activate_delayed_work(struct work_struct *work)
982 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
984 trace_workqueue_activate_work(work);
985 move_linked_works(work, &cwq->pool->worklist, NULL);
986 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
990 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
992 struct work_struct *work = list_first_entry(&cwq->delayed_works,
993 struct work_struct, entry);
995 cwq_activate_delayed_work(work);
999 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1000 * @cwq: cwq of interest
1001 * @color: color of work which left the queue
1003 * A work either has completed or is removed from pending queue,
1004 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1007 * spin_lock_irq(gcwq->lock).
1009 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
1011 /* ignore uncolored works */
1012 if (color == WORK_NO_COLOR)
1015 cwq->nr_in_flight[color]--;
1018 if (!list_empty(&cwq->delayed_works)) {
1019 /* one down, submit a delayed one */
1020 if (cwq->nr_active < cwq->max_active)
1021 cwq_activate_first_delayed(cwq);
1024 /* is flush in progress and are we at the flushing tip? */
1025 if (likely(cwq->flush_color != color))
1028 /* are there still in-flight works? */
1029 if (cwq->nr_in_flight[color])
1032 /* this cwq is done, clear flush_color */
1033 cwq->flush_color = -1;
1036 * If this was the last cwq, wake up the first flusher. It
1037 * will handle the rest.
1039 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1040 complete(&cwq->wq->first_flusher->done);
1044 * try_to_grab_pending - steal work item from worklist and disable irq
1045 * @work: work item to steal
1046 * @is_dwork: @work is a delayed_work
1047 * @flags: place to store irq state
1049 * Try to grab PENDING bit of @work. This function can handle @work in any
1050 * stable state - idle, on timer or on worklist. Return values are
1052 * 1 if @work was pending and we successfully stole PENDING
1053 * 0 if @work was idle and we claimed PENDING
1054 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1055 * -ENOENT if someone else is canceling @work, this state may persist
1056 * for arbitrarily long
1058 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1059 * interrupted while holding PENDING and @work off queue, irq must be
1060 * disabled on entry. This, combined with delayed_work->timer being
1061 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1063 * On successful return, >= 0, irq is disabled and the caller is
1064 * responsible for releasing it using local_irq_restore(*@flags).
1066 * This function is safe to call from any context including IRQ handler.
1068 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1069 unsigned long *flags)
1071 struct global_cwq *gcwq;
1073 WARN_ON_ONCE(in_irq());
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 gcwq = get_work_gcwq(work);
1102 spin_lock(&gcwq->lock);
1103 if (!list_empty(&work->entry)) {
1105 * This work is queued, but perhaps we locked the wrong gcwq.
1106 * In that case we must see the new value after rmb(), see
1107 * insert_work()->wmb().
1110 if (gcwq == get_work_gcwq(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(&gcwq->lock);
1132 spin_unlock(&gcwq->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(gcwq->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 global_cwq *gcwq = get_gcwq(cpu);
1194 struct worker *worker;
1195 struct hlist_node *pos;
1198 spin_lock_irqsave(&gcwq->lock, flags);
1199 for_each_busy_worker(worker, i, pos, gcwq) {
1200 if (worker->task != current)
1202 spin_unlock_irqrestore(&gcwq->lock, flags);
1204 * I'm @worker, no locking necessary. See if @work
1205 * is headed to the same workqueue.
1207 return worker->current_cwq->wq == wq;
1209 spin_unlock_irqrestore(&gcwq->lock, flags);
1214 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1215 struct work_struct *work)
1217 struct global_cwq *gcwq;
1218 struct cpu_workqueue_struct *cwq;
1219 struct list_head *worklist;
1220 unsigned int work_flags;
1221 unsigned int req_cpu = cpu;
1224 * While a work item is PENDING && off queue, a task trying to
1225 * steal the PENDING will busy-loop waiting for it to either get
1226 * queued or lose PENDING. Grabbing PENDING and queueing should
1227 * happen with IRQ disabled.
1229 WARN_ON_ONCE(!irqs_disabled());
1231 debug_work_activate(work);
1233 /* if dying, only works from the same workqueue are allowed */
1234 if (unlikely(wq->flags & WQ_DRAINING) &&
1235 WARN_ON_ONCE(!is_chained_work(wq)))
1238 /* determine gcwq to use */
1239 if (!(wq->flags & WQ_UNBOUND)) {
1240 struct global_cwq *last_gcwq;
1242 if (cpu == WORK_CPU_UNBOUND)
1243 cpu = raw_smp_processor_id();
1246 * It's multi cpu. If @work was previously on a different
1247 * cpu, it might still be running there, in which case the
1248 * work needs to be queued on that cpu to guarantee
1251 gcwq = get_gcwq(cpu);
1252 last_gcwq = get_work_gcwq(work);
1254 if (last_gcwq && last_gcwq != gcwq) {
1255 struct worker *worker;
1257 spin_lock(&last_gcwq->lock);
1259 worker = find_worker_executing_work(last_gcwq, work);
1261 if (worker && worker->current_cwq->wq == wq)
1264 /* meh... not running there, queue here */
1265 spin_unlock(&last_gcwq->lock);
1266 spin_lock(&gcwq->lock);
1269 spin_lock(&gcwq->lock);
1272 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1273 spin_lock(&gcwq->lock);
1276 /* gcwq determined, get cwq and queue */
1277 cwq = get_cwq(gcwq->cpu, wq);
1278 trace_workqueue_queue_work(req_cpu, cwq, work);
1280 if (WARN_ON(!list_empty(&work->entry))) {
1281 spin_unlock(&gcwq->lock);
1285 cwq->nr_in_flight[cwq->work_color]++;
1286 work_flags = work_color_to_flags(cwq->work_color);
1288 if (likely(cwq->nr_active < cwq->max_active)) {
1289 trace_workqueue_activate_work(work);
1291 worklist = &cwq->pool->worklist;
1293 work_flags |= WORK_STRUCT_DELAYED;
1294 worklist = &cwq->delayed_works;
1297 insert_work(cwq, work, worklist, work_flags);
1299 spin_unlock(&gcwq->lock);
1303 * queue_work_on - queue work on specific cpu
1304 * @cpu: CPU number to execute work on
1305 * @wq: workqueue to use
1306 * @work: work to queue
1308 * Returns %false if @work was already on a queue, %true otherwise.
1310 * We queue the work to a specific CPU, the caller must ensure it
1313 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1314 struct work_struct *work)
1317 unsigned long flags;
1319 local_irq_save(flags);
1321 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1322 __queue_work(cpu, wq, work);
1326 local_irq_restore(flags);
1329 EXPORT_SYMBOL_GPL(queue_work_on);
1332 * queue_work - queue work on a workqueue
1333 * @wq: workqueue to use
1334 * @work: work to queue
1336 * Returns %false if @work was already on a queue, %true otherwise.
1338 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1339 * it can be processed by another CPU.
1341 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1343 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1345 EXPORT_SYMBOL_GPL(queue_work);
1347 void delayed_work_timer_fn(unsigned long __data)
1349 struct delayed_work *dwork = (struct delayed_work *)__data;
1350 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1352 /* should have been called from irqsafe timer with irq already off */
1353 __queue_work(dwork->cpu, cwq->wq, &dwork->work);
1355 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1357 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1358 struct delayed_work *dwork, unsigned long delay)
1360 struct timer_list *timer = &dwork->timer;
1361 struct work_struct *work = &dwork->work;
1364 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1365 timer->data != (unsigned long)dwork);
1366 BUG_ON(timer_pending(timer));
1367 BUG_ON(!list_empty(&work->entry));
1369 timer_stats_timer_set_start_info(&dwork->timer);
1372 * This stores cwq for the moment, for the timer_fn. Note that the
1373 * work's gcwq is preserved to allow reentrance detection for
1376 if (!(wq->flags & WQ_UNBOUND)) {
1377 struct global_cwq *gcwq = get_work_gcwq(work);
1380 * If we cannot get the last gcwq from @work directly,
1381 * select the last CPU such that it avoids unnecessarily
1382 * triggering non-reentrancy check in __queue_work().
1387 if (lcpu == WORK_CPU_UNBOUND)
1388 lcpu = raw_smp_processor_id();
1390 lcpu = WORK_CPU_UNBOUND;
1393 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1396 timer->expires = jiffies + delay;
1398 if (unlikely(cpu != WORK_CPU_UNBOUND))
1399 add_timer_on(timer, cpu);
1405 * queue_delayed_work_on - queue work on specific CPU after delay
1406 * @cpu: CPU number to execute work on
1407 * @wq: workqueue to use
1408 * @dwork: work to queue
1409 * @delay: number of jiffies to wait before queueing
1411 * Returns %false if @work was already on a queue, %true otherwise. If
1412 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1415 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1416 struct delayed_work *dwork, unsigned long delay)
1418 struct work_struct *work = &dwork->work;
1420 unsigned long flags;
1423 return queue_work_on(cpu, wq, &dwork->work);
1425 /* read the comment in __queue_work() */
1426 local_irq_save(flags);
1428 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1429 __queue_delayed_work(cpu, wq, dwork, delay);
1433 local_irq_restore(flags);
1436 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1439 * queue_delayed_work - queue work on a workqueue after delay
1440 * @wq: workqueue to use
1441 * @dwork: delayable work to queue
1442 * @delay: number of jiffies to wait before queueing
1444 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1446 bool queue_delayed_work(struct workqueue_struct *wq,
1447 struct delayed_work *dwork, unsigned long delay)
1449 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1451 EXPORT_SYMBOL_GPL(queue_delayed_work);
1454 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1455 * @cpu: CPU number to execute work on
1456 * @wq: workqueue to use
1457 * @dwork: work to queue
1458 * @delay: number of jiffies to wait before queueing
1460 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1461 * modify @dwork's timer so that it expires after @delay. If @delay is
1462 * zero, @work is guaranteed to be scheduled immediately regardless of its
1465 * Returns %false if @dwork was idle and queued, %true if @dwork was
1466 * pending and its timer was modified.
1468 * This function is safe to call from any context including IRQ handler.
1469 * See try_to_grab_pending() for details.
1471 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1472 struct delayed_work *dwork, unsigned long delay)
1474 unsigned long flags;
1478 ret = try_to_grab_pending(&dwork->work, true, &flags);
1479 } while (unlikely(ret == -EAGAIN));
1481 if (likely(ret >= 0)) {
1482 __queue_delayed_work(cpu, wq, dwork, delay);
1483 local_irq_restore(flags);
1486 /* -ENOENT from try_to_grab_pending() becomes %true */
1489 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1492 * mod_delayed_work - modify delay of or queue a delayed work
1493 * @wq: workqueue to use
1494 * @dwork: work to queue
1495 * @delay: number of jiffies to wait before queueing
1497 * mod_delayed_work_on() on local CPU.
1499 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1500 unsigned long delay)
1502 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1504 EXPORT_SYMBOL_GPL(mod_delayed_work);
1507 * worker_enter_idle - enter idle state
1508 * @worker: worker which is entering idle state
1510 * @worker is entering idle state. Update stats and idle timer if
1514 * spin_lock_irq(gcwq->lock).
1516 static void worker_enter_idle(struct worker *worker)
1518 struct worker_pool *pool = worker->pool;
1519 struct global_cwq *gcwq = pool->gcwq;
1521 BUG_ON(worker->flags & WORKER_IDLE);
1522 BUG_ON(!list_empty(&worker->entry) &&
1523 (worker->hentry.next || worker->hentry.pprev));
1525 /* can't use worker_set_flags(), also called from start_worker() */
1526 worker->flags |= WORKER_IDLE;
1528 worker->last_active = jiffies;
1530 /* idle_list is LIFO */
1531 list_add(&worker->entry, &pool->idle_list);
1533 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1534 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1537 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1538 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1539 * nr_running, the warning may trigger spuriously. Check iff
1540 * unbind is not in progress.
1542 WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
1543 pool->nr_workers == pool->nr_idle &&
1544 atomic_read(get_pool_nr_running(pool)));
1548 * worker_leave_idle - leave idle state
1549 * @worker: worker which is leaving idle state
1551 * @worker is leaving idle state. Update stats.
1554 * spin_lock_irq(gcwq->lock).
1556 static void worker_leave_idle(struct worker *worker)
1558 struct worker_pool *pool = worker->pool;
1560 BUG_ON(!(worker->flags & WORKER_IDLE));
1561 worker_clr_flags(worker, WORKER_IDLE);
1563 list_del_init(&worker->entry);
1567 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1570 * Works which are scheduled while the cpu is online must at least be
1571 * scheduled to a worker which is bound to the cpu so that if they are
1572 * flushed from cpu callbacks while cpu is going down, they are
1573 * guaranteed to execute on the cpu.
1575 * This function is to be used by rogue workers and rescuers to bind
1576 * themselves to the target cpu and may race with cpu going down or
1577 * coming online. kthread_bind() can't be used because it may put the
1578 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1579 * verbatim as it's best effort and blocking and gcwq may be
1580 * [dis]associated in the meantime.
1582 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1583 * binding against %GCWQ_DISASSOCIATED which is set during
1584 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1585 * enters idle state or fetches works without dropping lock, it can
1586 * guarantee the scheduling requirement described in the first paragraph.
1589 * Might sleep. Called without any lock but returns with gcwq->lock
1593 * %true if the associated gcwq is online (@worker is successfully
1594 * bound), %false if offline.
1596 static bool worker_maybe_bind_and_lock(struct worker *worker)
1597 __acquires(&gcwq->lock)
1599 struct global_cwq *gcwq = worker->pool->gcwq;
1600 struct task_struct *task = worker->task;
1604 * The following call may fail, succeed or succeed
1605 * without actually migrating the task to the cpu if
1606 * it races with cpu hotunplug operation. Verify
1607 * against GCWQ_DISASSOCIATED.
1609 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1610 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1612 spin_lock_irq(&gcwq->lock);
1613 if (gcwq->flags & GCWQ_DISASSOCIATED)
1615 if (task_cpu(task) == gcwq->cpu &&
1616 cpumask_equal(¤t->cpus_allowed,
1617 get_cpu_mask(gcwq->cpu)))
1619 spin_unlock_irq(&gcwq->lock);
1622 * We've raced with CPU hot[un]plug. Give it a breather
1623 * and retry migration. cond_resched() is required here;
1624 * otherwise, we might deadlock against cpu_stop trying to
1625 * bring down the CPU on non-preemptive kernel.
1633 * Rebind an idle @worker to its CPU. worker_thread() will test
1634 * list_empty(@worker->entry) before leaving idle and call this function.
1636 static void idle_worker_rebind(struct worker *worker)
1638 struct global_cwq *gcwq = worker->pool->gcwq;
1640 /* CPU may go down again inbetween, clear UNBOUND only on success */
1641 if (worker_maybe_bind_and_lock(worker))
1642 worker_clr_flags(worker, WORKER_UNBOUND);
1644 /* rebind complete, become available again */
1645 list_add(&worker->entry, &worker->pool->idle_list);
1646 spin_unlock_irq(&gcwq->lock);
1650 * Function for @worker->rebind.work used to rebind unbound busy workers to
1651 * the associated cpu which is coming back online. This is scheduled by
1652 * cpu up but can race with other cpu hotplug operations and may be
1653 * executed twice without intervening cpu down.
1655 static void busy_worker_rebind_fn(struct work_struct *work)
1657 struct worker *worker = container_of(work, struct worker, rebind_work);
1658 struct global_cwq *gcwq = worker->pool->gcwq;
1660 if (worker_maybe_bind_and_lock(worker))
1661 worker_clr_flags(worker, WORKER_UNBOUND);
1663 spin_unlock_irq(&gcwq->lock);
1667 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1668 * @gcwq: gcwq of interest
1670 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1671 * is different for idle and busy ones.
1673 * Idle ones will be removed from the idle_list and woken up. They will
1674 * add themselves back after completing rebind. This ensures that the
1675 * idle_list doesn't contain any unbound workers when re-bound busy workers
1676 * try to perform local wake-ups for concurrency management.
1678 * Busy workers can rebind after they finish their current work items.
1679 * Queueing the rebind work item at the head of the scheduled list is
1680 * enough. Note that nr_running will be properly bumped as busy workers
1683 * On return, all non-manager workers are scheduled for rebind - see
1684 * manage_workers() for the manager special case. Any idle worker
1685 * including the manager will not appear on @idle_list until rebind is
1686 * complete, making local wake-ups safe.
1688 static void rebind_workers(struct global_cwq *gcwq)
1690 struct worker_pool *pool;
1691 struct worker *worker, *n;
1692 struct hlist_node *pos;
1695 lockdep_assert_held(&gcwq->lock);
1697 for_each_worker_pool(pool, gcwq)
1698 lockdep_assert_held(&pool->assoc_mutex);
1700 /* dequeue and kick idle ones */
1701 for_each_worker_pool(pool, gcwq) {
1702 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1704 * idle workers should be off @pool->idle_list
1705 * until rebind is complete to avoid receiving
1706 * premature local wake-ups.
1708 list_del_init(&worker->entry);
1711 * worker_thread() will see the above dequeuing
1712 * and call idle_worker_rebind().
1714 wake_up_process(worker->task);
1718 /* rebind busy workers */
1719 for_each_busy_worker(worker, i, pos, gcwq) {
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 (worker_pool_pri(worker->pool))
1734 wq = system_highpri_wq;
1738 insert_work(get_cwq(gcwq->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 struct global_cwq *gcwq = pool->gcwq;
1776 const char *pri = worker_pool_pri(pool) ? "H" : "";
1777 struct worker *worker = NULL;
1780 spin_lock_irq(&gcwq->lock);
1781 while (ida_get_new(&pool->worker_ida, &id)) {
1782 spin_unlock_irq(&gcwq->lock);
1783 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1785 spin_lock_irq(&gcwq->lock);
1787 spin_unlock_irq(&gcwq->lock);
1789 worker = alloc_worker();
1793 worker->pool = pool;
1796 if (gcwq->cpu != WORK_CPU_UNBOUND)
1797 worker->task = kthread_create_on_node(worker_thread,
1798 worker, cpu_to_node(gcwq->cpu),
1799 "kworker/%u:%d%s", gcwq->cpu, id, pri);
1801 worker->task = kthread_create(worker_thread, worker,
1802 "kworker/u:%d%s", id, pri);
1803 if (IS_ERR(worker->task))
1806 if (worker_pool_pri(pool))
1807 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1810 * Determine CPU binding of the new worker depending on
1811 * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1812 * flag remains stable across this function. See the comments
1813 * above the flag definition for details.
1815 * As an unbound worker may later become a regular one if CPU comes
1816 * online, make sure every worker has %PF_THREAD_BOUND set.
1818 if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
1819 kthread_bind(worker->task, gcwq->cpu);
1821 worker->task->flags |= PF_THREAD_BOUND;
1822 worker->flags |= WORKER_UNBOUND;
1828 spin_lock_irq(&gcwq->lock);
1829 ida_remove(&pool->worker_ida, id);
1830 spin_unlock_irq(&gcwq->lock);
1837 * start_worker - start a newly created worker
1838 * @worker: worker to start
1840 * Make the gcwq aware of @worker and start it.
1843 * spin_lock_irq(gcwq->lock).
1845 static void start_worker(struct worker *worker)
1847 worker->flags |= WORKER_STARTED;
1848 worker->pool->nr_workers++;
1849 worker_enter_idle(worker);
1850 wake_up_process(worker->task);
1854 * destroy_worker - destroy a workqueue worker
1855 * @worker: worker to be destroyed
1857 * Destroy @worker and adjust @gcwq stats accordingly.
1860 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1862 static void destroy_worker(struct worker *worker)
1864 struct worker_pool *pool = worker->pool;
1865 struct global_cwq *gcwq = pool->gcwq;
1866 int id = worker->id;
1868 /* sanity check frenzy */
1869 BUG_ON(worker->current_work);
1870 BUG_ON(!list_empty(&worker->scheduled));
1872 if (worker->flags & WORKER_STARTED)
1874 if (worker->flags & WORKER_IDLE)
1877 list_del_init(&worker->entry);
1878 worker->flags |= WORKER_DIE;
1880 spin_unlock_irq(&gcwq->lock);
1882 kthread_stop(worker->task);
1885 spin_lock_irq(&gcwq->lock);
1886 ida_remove(&pool->worker_ida, id);
1889 static void idle_worker_timeout(unsigned long __pool)
1891 struct worker_pool *pool = (void *)__pool;
1892 struct global_cwq *gcwq = pool->gcwq;
1894 spin_lock_irq(&gcwq->lock);
1896 if (too_many_workers(pool)) {
1897 struct worker *worker;
1898 unsigned long expires;
1900 /* idle_list is kept in LIFO order, check the last one */
1901 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1902 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1904 if (time_before(jiffies, expires))
1905 mod_timer(&pool->idle_timer, expires);
1907 /* it's been idle for too long, wake up manager */
1908 pool->flags |= POOL_MANAGE_WORKERS;
1909 wake_up_worker(pool);
1913 spin_unlock_irq(&gcwq->lock);
1916 static bool send_mayday(struct work_struct *work)
1918 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1919 struct workqueue_struct *wq = cwq->wq;
1922 if (!(wq->flags & WQ_RESCUER))
1925 /* mayday mayday mayday */
1926 cpu = cwq->pool->gcwq->cpu;
1927 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1928 if (cpu == WORK_CPU_UNBOUND)
1930 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1931 wake_up_process(wq->rescuer->task);
1935 static void gcwq_mayday_timeout(unsigned long __pool)
1937 struct worker_pool *pool = (void *)__pool;
1938 struct global_cwq *gcwq = pool->gcwq;
1939 struct work_struct *work;
1941 spin_lock_irq(&gcwq->lock);
1943 if (need_to_create_worker(pool)) {
1945 * We've been trying to create a new worker but
1946 * haven't been successful. We might be hitting an
1947 * allocation deadlock. Send distress signals to
1950 list_for_each_entry(work, &pool->worklist, entry)
1954 spin_unlock_irq(&gcwq->lock);
1956 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1960 * maybe_create_worker - create a new worker if necessary
1961 * @pool: pool to create a new worker for
1963 * Create a new worker for @pool if necessary. @pool is guaranteed to
1964 * have at least one idle worker on return from this function. If
1965 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1966 * sent to all rescuers with works scheduled on @pool to resolve
1967 * possible allocation deadlock.
1969 * On return, need_to_create_worker() is guaranteed to be false and
1970 * may_start_working() true.
1973 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1974 * multiple times. Does GFP_KERNEL allocations. Called only from
1978 * false if no action was taken and gcwq->lock stayed locked, true
1981 static bool maybe_create_worker(struct worker_pool *pool)
1982 __releases(&gcwq->lock)
1983 __acquires(&gcwq->lock)
1985 struct global_cwq *gcwq = pool->gcwq;
1987 if (!need_to_create_worker(pool))
1990 spin_unlock_irq(&gcwq->lock);
1992 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1993 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1996 struct worker *worker;
1998 worker = create_worker(pool);
2000 del_timer_sync(&pool->mayday_timer);
2001 spin_lock_irq(&gcwq->lock);
2002 start_worker(worker);
2003 BUG_ON(need_to_create_worker(pool));
2007 if (!need_to_create_worker(pool))
2010 __set_current_state(TASK_INTERRUPTIBLE);
2011 schedule_timeout(CREATE_COOLDOWN);
2013 if (!need_to_create_worker(pool))
2017 del_timer_sync(&pool->mayday_timer);
2018 spin_lock_irq(&gcwq->lock);
2019 if (need_to_create_worker(pool))
2025 * maybe_destroy_worker - destroy workers which have been idle for a while
2026 * @pool: pool to destroy workers for
2028 * Destroy @pool workers which have been idle for longer than
2029 * IDLE_WORKER_TIMEOUT.
2032 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2033 * multiple times. Called only from manager.
2036 * false if no action was taken and gcwq->lock stayed locked, true
2039 static bool maybe_destroy_workers(struct worker_pool *pool)
2043 while (too_many_workers(pool)) {
2044 struct worker *worker;
2045 unsigned long expires;
2047 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2048 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2050 if (time_before(jiffies, expires)) {
2051 mod_timer(&pool->idle_timer, expires);
2055 destroy_worker(worker);
2063 * manage_workers - manage worker pool
2066 * Assume the manager role and manage gcwq worker pool @worker belongs
2067 * to. At any given time, there can be only zero or one manager per
2068 * gcwq. The exclusion is handled automatically by this function.
2070 * The caller can safely start processing works on false return. On
2071 * true return, it's guaranteed that need_to_create_worker() is false
2072 * and may_start_working() is true.
2075 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2076 * multiple times. Does GFP_KERNEL allocations.
2079 * false if no action was taken and gcwq->lock stayed locked, true if
2080 * some action was taken.
2082 static bool manage_workers(struct worker *worker)
2084 struct worker_pool *pool = worker->pool;
2087 if (pool->flags & POOL_MANAGING_WORKERS)
2090 pool->flags |= POOL_MANAGING_WORKERS;
2093 * To simplify both worker management and CPU hotplug, hold off
2094 * management while hotplug is in progress. CPU hotplug path can't
2095 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2096 * lead to idle worker depletion (all become busy thinking someone
2097 * else is managing) which in turn can result in deadlock under
2098 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2099 * manager against CPU hotplug.
2101 * assoc_mutex would always be free unless CPU hotplug is in
2102 * progress. trylock first without dropping @gcwq->lock.
2104 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2105 spin_unlock_irq(&pool->gcwq->lock);
2106 mutex_lock(&pool->assoc_mutex);
2108 * CPU hotplug could have happened while we were waiting
2109 * for assoc_mutex. Hotplug itself can't handle us
2110 * because manager isn't either on idle or busy list, and
2111 * @gcwq's state and ours could have deviated.
2113 * As hotplug is now excluded via assoc_mutex, we can
2114 * simply try to bind. It will succeed or fail depending
2115 * on @gcwq's current state. Try it and adjust
2116 * %WORKER_UNBOUND accordingly.
2118 if (worker_maybe_bind_and_lock(worker))
2119 worker->flags &= ~WORKER_UNBOUND;
2121 worker->flags |= WORKER_UNBOUND;
2126 pool->flags &= ~POOL_MANAGE_WORKERS;
2129 * Destroy and then create so that may_start_working() is true
2132 ret |= maybe_destroy_workers(pool);
2133 ret |= maybe_create_worker(pool);
2135 pool->flags &= ~POOL_MANAGING_WORKERS;
2136 mutex_unlock(&pool->assoc_mutex);
2141 * process_one_work - process single work
2143 * @work: work to process
2145 * Process @work. This function contains all the logics necessary to
2146 * process a single work including synchronization against and
2147 * interaction with other workers on the same cpu, queueing and
2148 * flushing. As long as context requirement is met, any worker can
2149 * call this function to process a work.
2152 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2154 static void process_one_work(struct worker *worker, struct work_struct *work)
2155 __releases(&gcwq->lock)
2156 __acquires(&gcwq->lock)
2158 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2159 struct worker_pool *pool = worker->pool;
2160 struct global_cwq *gcwq = pool->gcwq;
2161 struct hlist_head *bwh = busy_worker_head(gcwq, work);
2162 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2163 work_func_t f = work->func;
2165 struct worker *collision;
2166 #ifdef CONFIG_LOCKDEP
2168 * It is permissible to free the struct work_struct from
2169 * inside the function that is called from it, this we need to
2170 * take into account for lockdep too. To avoid bogus "held
2171 * lock freed" warnings as well as problems when looking into
2172 * work->lockdep_map, make a copy and use that here.
2174 struct lockdep_map lockdep_map;
2176 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2179 * Ensure we're on the correct CPU. DISASSOCIATED test is
2180 * necessary to avoid spurious warnings from rescuers servicing the
2181 * unbound or a disassociated gcwq.
2183 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2184 !(gcwq->flags & GCWQ_DISASSOCIATED) &&
2185 raw_smp_processor_id() != gcwq->cpu);
2188 * A single work shouldn't be executed concurrently by
2189 * multiple workers on a single cpu. Check whether anyone is
2190 * already processing the work. If so, defer the work to the
2191 * currently executing one.
2193 collision = __find_worker_executing_work(gcwq, bwh, work);
2194 if (unlikely(collision)) {
2195 move_linked_works(work, &collision->scheduled, NULL);
2199 /* claim and dequeue */
2200 debug_work_deactivate(work);
2201 hlist_add_head(&worker->hentry, bwh);
2202 worker->current_work = work;
2203 worker->current_cwq = cwq;
2204 work_color = get_work_color(work);
2206 list_del_init(&work->entry);
2209 * CPU intensive works don't participate in concurrency
2210 * management. They're the scheduler's responsibility.
2212 if (unlikely(cpu_intensive))
2213 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2216 * Unbound gcwq isn't concurrency managed and work items should be
2217 * executed ASAP. Wake up another worker if necessary.
2219 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2220 wake_up_worker(pool);
2223 * Record the last CPU and clear PENDING which should be the last
2224 * update to @work. Also, do this inside @gcwq->lock so that
2225 * PENDING and queued state changes happen together while IRQ is
2228 set_work_cpu_and_clear_pending(work, gcwq->cpu);
2230 spin_unlock_irq(&gcwq->lock);
2232 lock_map_acquire_read(&cwq->wq->lockdep_map);
2233 lock_map_acquire(&lockdep_map);
2234 trace_workqueue_execute_start(work);
2237 * While we must be careful to not use "work" after this, the trace
2238 * point will only record its address.
2240 trace_workqueue_execute_end(work);
2241 lock_map_release(&lockdep_map);
2242 lock_map_release(&cwq->wq->lockdep_map);
2244 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2245 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2246 " last function: %pf\n",
2247 current->comm, preempt_count(), task_pid_nr(current), f);
2248 debug_show_held_locks(current);
2252 spin_lock_irq(&gcwq->lock);
2254 /* clear cpu intensive status */
2255 if (unlikely(cpu_intensive))
2256 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2258 /* we're done with it, release */
2259 hlist_del_init(&worker->hentry);
2260 worker->current_work = NULL;
2261 worker->current_cwq = NULL;
2262 cwq_dec_nr_in_flight(cwq, work_color);
2266 * process_scheduled_works - process scheduled works
2269 * Process all scheduled works. Please note that the scheduled list
2270 * may change while processing a work, so this function repeatedly
2271 * fetches a work from the top and executes it.
2274 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2277 static void process_scheduled_works(struct worker *worker)
2279 while (!list_empty(&worker->scheduled)) {
2280 struct work_struct *work = list_first_entry(&worker->scheduled,
2281 struct work_struct, entry);
2282 process_one_work(worker, work);
2287 * worker_thread - the worker thread function
2290 * The gcwq worker thread function. There's a single dynamic pool of
2291 * these per each cpu. These workers process all works regardless of
2292 * their specific target workqueue. The only exception is works which
2293 * belong to workqueues with a rescuer which will be explained in
2296 static int worker_thread(void *__worker)
2298 struct worker *worker = __worker;
2299 struct worker_pool *pool = worker->pool;
2300 struct global_cwq *gcwq = pool->gcwq;
2302 /* tell the scheduler that this is a workqueue worker */
2303 worker->task->flags |= PF_WQ_WORKER;
2305 spin_lock_irq(&gcwq->lock);
2307 /* we are off idle list if destruction or rebind is requested */
2308 if (unlikely(list_empty(&worker->entry))) {
2309 spin_unlock_irq(&gcwq->lock);
2311 /* if DIE is set, destruction is requested */
2312 if (worker->flags & WORKER_DIE) {
2313 worker->task->flags &= ~PF_WQ_WORKER;
2317 /* otherwise, rebind */
2318 idle_worker_rebind(worker);
2322 worker_leave_idle(worker);
2324 /* no more worker necessary? */
2325 if (!need_more_worker(pool))
2328 /* do we need to manage? */
2329 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2333 * ->scheduled list can only be filled while a worker is
2334 * preparing to process a work or actually processing it.
2335 * Make sure nobody diddled with it while I was sleeping.
2337 BUG_ON(!list_empty(&worker->scheduled));
2340 * When control reaches this point, we're guaranteed to have
2341 * at least one idle worker or that someone else has already
2342 * assumed the manager role.
2344 worker_clr_flags(worker, WORKER_PREP);
2347 struct work_struct *work =
2348 list_first_entry(&pool->worklist,
2349 struct work_struct, entry);
2351 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2352 /* optimization path, not strictly necessary */
2353 process_one_work(worker, work);
2354 if (unlikely(!list_empty(&worker->scheduled)))
2355 process_scheduled_works(worker);
2357 move_linked_works(work, &worker->scheduled, NULL);
2358 process_scheduled_works(worker);
2360 } while (keep_working(pool));
2362 worker_set_flags(worker, WORKER_PREP, false);
2364 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2368 * gcwq->lock is held and there's no work to process and no
2369 * need to manage, sleep. Workers are woken up only while
2370 * holding gcwq->lock or from local cpu, so setting the
2371 * current state before releasing gcwq->lock is enough to
2372 * prevent losing any event.
2374 worker_enter_idle(worker);
2375 __set_current_state(TASK_INTERRUPTIBLE);
2376 spin_unlock_irq(&gcwq->lock);
2382 * rescuer_thread - the rescuer thread function
2383 * @__wq: the associated workqueue
2385 * Workqueue rescuer thread function. There's one rescuer for each
2386 * workqueue which has WQ_RESCUER set.
2388 * Regular work processing on a gcwq may block trying to create a new
2389 * worker which uses GFP_KERNEL allocation which has slight chance of
2390 * developing into deadlock if some works currently on the same queue
2391 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2392 * the problem rescuer solves.
2394 * When such condition is possible, the gcwq summons rescuers of all
2395 * workqueues which have works queued on the gcwq and let them process
2396 * those works so that forward progress can be guaranteed.
2398 * This should happen rarely.
2400 static int rescuer_thread(void *__wq)
2402 struct workqueue_struct *wq = __wq;
2403 struct worker *rescuer = wq->rescuer;
2404 struct list_head *scheduled = &rescuer->scheduled;
2405 bool is_unbound = wq->flags & WQ_UNBOUND;
2408 set_user_nice(current, RESCUER_NICE_LEVEL);
2410 set_current_state(TASK_INTERRUPTIBLE);
2412 if (kthread_should_stop())
2416 * See whether any cpu is asking for help. Unbounded
2417 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2419 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2420 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2421 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2422 struct worker_pool *pool = cwq->pool;
2423 struct global_cwq *gcwq = pool->gcwq;
2424 struct work_struct *work, *n;
2426 __set_current_state(TASK_RUNNING);
2427 mayday_clear_cpu(cpu, wq->mayday_mask);
2429 /* migrate to the target cpu if possible */
2430 rescuer->pool = pool;
2431 worker_maybe_bind_and_lock(rescuer);
2434 * Slurp in all works issued via this workqueue and
2437 BUG_ON(!list_empty(&rescuer->scheduled));
2438 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2439 if (get_work_cwq(work) == cwq)
2440 move_linked_works(work, scheduled, &n);
2442 process_scheduled_works(rescuer);
2445 * Leave this gcwq. If keep_working() is %true, notify a
2446 * regular worker; otherwise, we end up with 0 concurrency
2447 * and stalling the execution.
2449 if (keep_working(pool))
2450 wake_up_worker(pool);
2452 spin_unlock_irq(&gcwq->lock);
2460 struct work_struct work;
2461 struct completion done;
2464 static void wq_barrier_func(struct work_struct *work)
2466 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2467 complete(&barr->done);
2471 * insert_wq_barrier - insert a barrier work
2472 * @cwq: cwq to insert barrier into
2473 * @barr: wq_barrier to insert
2474 * @target: target work to attach @barr to
2475 * @worker: worker currently executing @target, NULL if @target is not executing
2477 * @barr is linked to @target such that @barr is completed only after
2478 * @target finishes execution. Please note that the ordering
2479 * guarantee is observed only with respect to @target and on the local
2482 * Currently, a queued barrier can't be canceled. This is because
2483 * try_to_grab_pending() can't determine whether the work to be
2484 * grabbed is at the head of the queue and thus can't clear LINKED
2485 * flag of the previous work while there must be a valid next work
2486 * after a work with LINKED flag set.
2488 * Note that when @worker is non-NULL, @target may be modified
2489 * underneath us, so we can't reliably determine cwq from @target.
2492 * spin_lock_irq(gcwq->lock).
2494 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2495 struct wq_barrier *barr,
2496 struct work_struct *target, struct worker *worker)
2498 struct list_head *head;
2499 unsigned int linked = 0;
2502 * debugobject calls are safe here even with gcwq->lock locked
2503 * as we know for sure that this will not trigger any of the
2504 * checks and call back into the fixup functions where we
2507 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2508 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2509 init_completion(&barr->done);
2512 * If @target is currently being executed, schedule the
2513 * barrier to the worker; otherwise, put it after @target.
2516 head = worker->scheduled.next;
2518 unsigned long *bits = work_data_bits(target);
2520 head = target->entry.next;
2521 /* there can already be other linked works, inherit and set */
2522 linked = *bits & WORK_STRUCT_LINKED;
2523 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2526 debug_work_activate(&barr->work);
2527 insert_work(cwq, &barr->work, head,
2528 work_color_to_flags(WORK_NO_COLOR) | linked);
2532 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2533 * @wq: workqueue being flushed
2534 * @flush_color: new flush color, < 0 for no-op
2535 * @work_color: new work color, < 0 for no-op
2537 * Prepare cwqs for workqueue flushing.
2539 * If @flush_color is non-negative, flush_color on all cwqs should be
2540 * -1. If no cwq has in-flight commands at the specified color, all
2541 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2542 * has in flight commands, its cwq->flush_color is set to
2543 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2544 * wakeup logic is armed and %true is returned.
2546 * The caller should have initialized @wq->first_flusher prior to
2547 * calling this function with non-negative @flush_color. If
2548 * @flush_color is negative, no flush color update is done and %false
2551 * If @work_color is non-negative, all cwqs should have the same
2552 * work_color which is previous to @work_color and all will be
2553 * advanced to @work_color.
2556 * mutex_lock(wq->flush_mutex).
2559 * %true if @flush_color >= 0 and there's something to flush. %false
2562 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2563 int flush_color, int work_color)
2568 if (flush_color >= 0) {
2569 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2570 atomic_set(&wq->nr_cwqs_to_flush, 1);
2573 for_each_cwq_cpu(cpu, wq) {
2574 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2575 struct global_cwq *gcwq = cwq->pool->gcwq;
2577 spin_lock_irq(&gcwq->lock);
2579 if (flush_color >= 0) {
2580 BUG_ON(cwq->flush_color != -1);
2582 if (cwq->nr_in_flight[flush_color]) {
2583 cwq->flush_color = flush_color;
2584 atomic_inc(&wq->nr_cwqs_to_flush);
2589 if (work_color >= 0) {
2590 BUG_ON(work_color != work_next_color(cwq->work_color));
2591 cwq->work_color = work_color;
2594 spin_unlock_irq(&gcwq->lock);
2597 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2598 complete(&wq->first_flusher->done);
2604 * flush_workqueue - ensure that any scheduled work has run to completion.
2605 * @wq: workqueue to flush
2607 * Forces execution of the workqueue and blocks until its completion.
2608 * This is typically used in driver shutdown handlers.
2610 * We sleep until all works which were queued on entry have been handled,
2611 * but we are not livelocked by new incoming ones.
2613 void flush_workqueue(struct workqueue_struct *wq)
2615 struct wq_flusher this_flusher = {
2616 .list = LIST_HEAD_INIT(this_flusher.list),
2618 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2622 lock_map_acquire(&wq->lockdep_map);
2623 lock_map_release(&wq->lockdep_map);
2625 mutex_lock(&wq->flush_mutex);
2628 * Start-to-wait phase
2630 next_color = work_next_color(wq->work_color);
2632 if (next_color != wq->flush_color) {
2634 * Color space is not full. The current work_color
2635 * becomes our flush_color and work_color is advanced
2638 BUG_ON(!list_empty(&wq->flusher_overflow));
2639 this_flusher.flush_color = wq->work_color;
2640 wq->work_color = next_color;
2642 if (!wq->first_flusher) {
2643 /* no flush in progress, become the first flusher */
2644 BUG_ON(wq->flush_color != this_flusher.flush_color);
2646 wq->first_flusher = &this_flusher;
2648 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2650 /* nothing to flush, done */
2651 wq->flush_color = next_color;
2652 wq->first_flusher = NULL;
2657 BUG_ON(wq->flush_color == this_flusher.flush_color);
2658 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2659 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2663 * Oops, color space is full, wait on overflow queue.
2664 * The next flush completion will assign us
2665 * flush_color and transfer to flusher_queue.
2667 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2670 mutex_unlock(&wq->flush_mutex);
2672 wait_for_completion(&this_flusher.done);
2675 * Wake-up-and-cascade phase
2677 * First flushers are responsible for cascading flushes and
2678 * handling overflow. Non-first flushers can simply return.
2680 if (wq->first_flusher != &this_flusher)
2683 mutex_lock(&wq->flush_mutex);
2685 /* we might have raced, check again with mutex held */
2686 if (wq->first_flusher != &this_flusher)
2689 wq->first_flusher = NULL;
2691 BUG_ON(!list_empty(&this_flusher.list));
2692 BUG_ON(wq->flush_color != this_flusher.flush_color);
2695 struct wq_flusher *next, *tmp;
2697 /* complete all the flushers sharing the current flush color */
2698 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2699 if (next->flush_color != wq->flush_color)
2701 list_del_init(&next->list);
2702 complete(&next->done);
2705 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2706 wq->flush_color != work_next_color(wq->work_color));
2708 /* this flush_color is finished, advance by one */
2709 wq->flush_color = work_next_color(wq->flush_color);
2711 /* one color has been freed, handle overflow queue */
2712 if (!list_empty(&wq->flusher_overflow)) {
2714 * Assign the same color to all overflowed
2715 * flushers, advance work_color and append to
2716 * flusher_queue. This is the start-to-wait
2717 * phase for these overflowed flushers.
2719 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2720 tmp->flush_color = wq->work_color;
2722 wq->work_color = work_next_color(wq->work_color);
2724 list_splice_tail_init(&wq->flusher_overflow,
2725 &wq->flusher_queue);
2726 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2729 if (list_empty(&wq->flusher_queue)) {
2730 BUG_ON(wq->flush_color != wq->work_color);
2735 * Need to flush more colors. Make the next flusher
2736 * the new first flusher and arm cwqs.
2738 BUG_ON(wq->flush_color == wq->work_color);
2739 BUG_ON(wq->flush_color != next->flush_color);
2741 list_del_init(&next->list);
2742 wq->first_flusher = next;
2744 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2748 * Meh... this color is already done, clear first
2749 * flusher and repeat cascading.
2751 wq->first_flusher = NULL;
2755 mutex_unlock(&wq->flush_mutex);
2757 EXPORT_SYMBOL_GPL(flush_workqueue);
2760 * drain_workqueue - drain a workqueue
2761 * @wq: workqueue to drain
2763 * Wait until the workqueue becomes empty. While draining is in progress,
2764 * only chain queueing is allowed. IOW, only currently pending or running
2765 * work items on @wq can queue further work items on it. @wq is flushed
2766 * repeatedly until it becomes empty. The number of flushing is detemined
2767 * by the depth of chaining and should be relatively short. Whine if it
2770 void drain_workqueue(struct workqueue_struct *wq)
2772 unsigned int flush_cnt = 0;
2776 * __queue_work() needs to test whether there are drainers, is much
2777 * hotter than drain_workqueue() and already looks at @wq->flags.
2778 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2780 spin_lock(&workqueue_lock);
2781 if (!wq->nr_drainers++)
2782 wq->flags |= WQ_DRAINING;
2783 spin_unlock(&workqueue_lock);
2785 flush_workqueue(wq);
2787 for_each_cwq_cpu(cpu, wq) {
2788 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2791 spin_lock_irq(&cwq->pool->gcwq->lock);
2792 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2793 spin_unlock_irq(&cwq->pool->gcwq->lock);
2798 if (++flush_cnt == 10 ||
2799 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2800 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2801 wq->name, flush_cnt);
2805 spin_lock(&workqueue_lock);
2806 if (!--wq->nr_drainers)
2807 wq->flags &= ~WQ_DRAINING;
2808 spin_unlock(&workqueue_lock);
2810 EXPORT_SYMBOL_GPL(drain_workqueue);
2812 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2814 struct worker *worker = NULL;
2815 struct global_cwq *gcwq;
2816 struct cpu_workqueue_struct *cwq;
2819 gcwq = get_work_gcwq(work);
2823 spin_lock_irq(&gcwq->lock);
2824 if (!list_empty(&work->entry)) {
2826 * See the comment near try_to_grab_pending()->smp_rmb().
2827 * If it was re-queued to a different gcwq under us, we
2828 * are not going to wait.
2831 cwq = get_work_cwq(work);
2832 if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
2835 worker = find_worker_executing_work(gcwq, work);
2838 cwq = worker->current_cwq;
2841 insert_wq_barrier(cwq, barr, work, worker);
2842 spin_unlock_irq(&gcwq->lock);
2845 * If @max_active is 1 or rescuer is in use, flushing another work
2846 * item on the same workqueue may lead to deadlock. Make sure the
2847 * flusher is not running on the same workqueue by verifying write
2850 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2851 lock_map_acquire(&cwq->wq->lockdep_map);
2853 lock_map_acquire_read(&cwq->wq->lockdep_map);
2854 lock_map_release(&cwq->wq->lockdep_map);
2858 spin_unlock_irq(&gcwq->lock);
2863 * flush_work - wait for a work to finish executing the last queueing instance
2864 * @work: the work to flush
2866 * Wait until @work has finished execution. @work is guaranteed to be idle
2867 * on return if it hasn't been requeued since flush started.
2870 * %true if flush_work() waited for the work to finish execution,
2871 * %false if it was already idle.
2873 bool flush_work(struct work_struct *work)
2875 struct wq_barrier barr;
2877 lock_map_acquire(&work->lockdep_map);
2878 lock_map_release(&work->lockdep_map);
2880 if (start_flush_work(work, &barr)) {
2881 wait_for_completion(&barr.done);
2882 destroy_work_on_stack(&barr.work);
2888 EXPORT_SYMBOL_GPL(flush_work);
2890 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2892 unsigned long flags;
2896 ret = try_to_grab_pending(work, is_dwork, &flags);
2898 * If someone else is canceling, wait for the same event it
2899 * would be waiting for before retrying.
2901 if (unlikely(ret == -ENOENT))
2903 } while (unlikely(ret < 0));
2905 /* tell other tasks trying to grab @work to back off */
2906 mark_work_canceling(work);
2907 local_irq_restore(flags);
2910 clear_work_data(work);
2915 * cancel_work_sync - cancel a work and wait for it to finish
2916 * @work: the work to cancel
2918 * Cancel @work and wait for its execution to finish. This function
2919 * can be used even if the work re-queues itself or migrates to
2920 * another workqueue. On return from this function, @work is
2921 * guaranteed to be not pending or executing on any CPU.
2923 * cancel_work_sync(&delayed_work->work) must not be used for
2924 * delayed_work's. Use cancel_delayed_work_sync() instead.
2926 * The caller must ensure that the workqueue on which @work was last
2927 * queued can't be destroyed before this function returns.
2930 * %true if @work was pending, %false otherwise.
2932 bool cancel_work_sync(struct work_struct *work)
2934 return __cancel_work_timer(work, false);
2936 EXPORT_SYMBOL_GPL(cancel_work_sync);
2939 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2940 * @dwork: the delayed work to flush
2942 * Delayed timer is cancelled and the pending work is queued for
2943 * immediate execution. Like flush_work(), this function only
2944 * considers the last queueing instance of @dwork.
2947 * %true if flush_work() waited for the work to finish execution,
2948 * %false if it was already idle.
2950 bool flush_delayed_work(struct delayed_work *dwork)
2952 local_irq_disable();
2953 if (del_timer_sync(&dwork->timer))
2954 __queue_work(dwork->cpu,
2955 get_work_cwq(&dwork->work)->wq, &dwork->work);
2957 return flush_work(&dwork->work);
2959 EXPORT_SYMBOL(flush_delayed_work);
2962 * cancel_delayed_work - cancel a delayed work
2963 * @dwork: delayed_work to cancel
2965 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2966 * and canceled; %false if wasn't pending. Note that the work callback
2967 * function may still be running on return, unless it returns %true and the
2968 * work doesn't re-arm itself. Explicitly flush or use
2969 * cancel_delayed_work_sync() to wait on it.
2971 * This function is safe to call from any context including IRQ handler.
2973 bool cancel_delayed_work(struct delayed_work *dwork)
2975 unsigned long flags;
2979 ret = try_to_grab_pending(&dwork->work, true, &flags);
2980 } while (unlikely(ret == -EAGAIN));
2982 if (unlikely(ret < 0))
2985 set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work));
2986 local_irq_restore(flags);
2989 EXPORT_SYMBOL(cancel_delayed_work);
2992 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2993 * @dwork: the delayed work cancel
2995 * This is cancel_work_sync() for delayed works.
2998 * %true if @dwork was pending, %false otherwise.
3000 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3002 return __cancel_work_timer(&dwork->work, true);
3004 EXPORT_SYMBOL(cancel_delayed_work_sync);
3007 * schedule_work_on - put work task on a specific cpu
3008 * @cpu: cpu to put the work task on
3009 * @work: job to be done
3011 * This puts a job on a specific cpu
3013 bool schedule_work_on(int cpu, struct work_struct *work)
3015 return queue_work_on(cpu, system_wq, work);
3017 EXPORT_SYMBOL(schedule_work_on);
3020 * schedule_work - put work task in global workqueue
3021 * @work: job to be done
3023 * Returns %false if @work was already on the kernel-global workqueue and
3026 * This puts a job in the kernel-global workqueue if it was not already
3027 * queued and leaves it in the same position on the kernel-global
3028 * workqueue otherwise.
3030 bool schedule_work(struct work_struct *work)
3032 return queue_work(system_wq, work);
3034 EXPORT_SYMBOL(schedule_work);
3037 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3039 * @dwork: job to be done
3040 * @delay: number of jiffies to wait
3042 * After waiting for a given time this puts a job in the kernel-global
3043 * workqueue on the specified CPU.
3045 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3046 unsigned long delay)
3048 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3050 EXPORT_SYMBOL(schedule_delayed_work_on);
3053 * schedule_delayed_work - put work task in global workqueue after delay
3054 * @dwork: job to be done
3055 * @delay: number of jiffies to wait or 0 for immediate execution
3057 * After waiting for a given time this puts a job in the kernel-global
3060 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3062 return queue_delayed_work(system_wq, dwork, delay);
3064 EXPORT_SYMBOL(schedule_delayed_work);
3067 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3068 * @func: the function to call
3070 * schedule_on_each_cpu() executes @func on each online CPU using the
3071 * system workqueue and blocks until all CPUs have completed.
3072 * schedule_on_each_cpu() is very slow.
3075 * 0 on success, -errno on failure.
3077 int schedule_on_each_cpu(work_func_t func)
3080 struct work_struct __percpu *works;
3082 works = alloc_percpu(struct work_struct);
3088 for_each_online_cpu(cpu) {
3089 struct work_struct *work = per_cpu_ptr(works, cpu);
3091 INIT_WORK(work, func);
3092 schedule_work_on(cpu, work);
3095 for_each_online_cpu(cpu)
3096 flush_work(per_cpu_ptr(works, cpu));
3104 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3106 * Forces execution of the kernel-global workqueue and blocks until its
3109 * Think twice before calling this function! It's very easy to get into
3110 * trouble if you don't take great care. Either of the following situations
3111 * will lead to deadlock:
3113 * One of the work items currently on the workqueue needs to acquire
3114 * a lock held by your code or its caller.
3116 * Your code is running in the context of a work routine.
3118 * They will be detected by lockdep when they occur, but the first might not
3119 * occur very often. It depends on what work items are on the workqueue and
3120 * what locks they need, which you have no control over.
3122 * In most situations flushing the entire workqueue is overkill; you merely
3123 * need to know that a particular work item isn't queued and isn't running.
3124 * In such cases you should use cancel_delayed_work_sync() or
3125 * cancel_work_sync() instead.
3127 void flush_scheduled_work(void)
3129 flush_workqueue(system_wq);
3131 EXPORT_SYMBOL(flush_scheduled_work);
3134 * execute_in_process_context - reliably execute the routine with user context
3135 * @fn: the function to execute
3136 * @ew: guaranteed storage for the execute work structure (must
3137 * be available when the work executes)
3139 * Executes the function immediately if process context is available,
3140 * otherwise schedules the function for delayed execution.
3142 * Returns: 0 - function was executed
3143 * 1 - function was scheduled for execution
3145 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3147 if (!in_interrupt()) {
3152 INIT_WORK(&ew->work, fn);
3153 schedule_work(&ew->work);
3157 EXPORT_SYMBOL_GPL(execute_in_process_context);
3159 int keventd_up(void)
3161 return system_wq != NULL;
3164 static int alloc_cwqs(struct workqueue_struct *wq)
3167 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3168 * Make sure that the alignment isn't lower than that of
3169 * unsigned long long.
3171 const size_t size = sizeof(struct cpu_workqueue_struct);
3172 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3173 __alignof__(unsigned long long));
3175 if (!(wq->flags & WQ_UNBOUND))
3176 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3181 * Allocate enough room to align cwq and put an extra
3182 * pointer at the end pointing back to the originally
3183 * allocated pointer which will be used for free.
3185 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3187 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3188 *(void **)(wq->cpu_wq.single + 1) = ptr;
3192 /* just in case, make sure it's actually aligned */
3193 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3194 return wq->cpu_wq.v ? 0 : -ENOMEM;
3197 static void free_cwqs(struct workqueue_struct *wq)
3199 if (!(wq->flags & WQ_UNBOUND))
3200 free_percpu(wq->cpu_wq.pcpu);
3201 else if (wq->cpu_wq.single) {
3202 /* the pointer to free is stored right after the cwq */
3203 kfree(*(void **)(wq->cpu_wq.single + 1));
3207 static int wq_clamp_max_active(int max_active, unsigned int flags,
3210 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3212 if (max_active < 1 || max_active > lim)
3213 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3214 max_active, name, 1, lim);
3216 return clamp_val(max_active, 1, lim);
3219 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3222 struct lock_class_key *key,
3223 const char *lock_name, ...)
3225 va_list args, args1;
3226 struct workqueue_struct *wq;
3230 /* determine namelen, allocate wq and format name */
3231 va_start(args, lock_name);
3232 va_copy(args1, args);
3233 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3235 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3239 vsnprintf(wq->name, namelen, fmt, args1);
3244 * Workqueues which may be used during memory reclaim should
3245 * have a rescuer to guarantee forward progress.
3247 if (flags & WQ_MEM_RECLAIM)
3248 flags |= WQ_RESCUER;
3250 max_active = max_active ?: WQ_DFL_ACTIVE;
3251 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3255 wq->saved_max_active = max_active;
3256 mutex_init(&wq->flush_mutex);
3257 atomic_set(&wq->nr_cwqs_to_flush, 0);
3258 INIT_LIST_HEAD(&wq->flusher_queue);
3259 INIT_LIST_HEAD(&wq->flusher_overflow);
3261 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3262 INIT_LIST_HEAD(&wq->list);
3264 if (alloc_cwqs(wq) < 0)
3267 for_each_cwq_cpu(cpu, wq) {
3268 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3269 struct global_cwq *gcwq = get_gcwq(cpu);
3270 int pool_idx = (bool)(flags & WQ_HIGHPRI);
3272 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3273 cwq->pool = &gcwq->pools[pool_idx];
3275 cwq->flush_color = -1;
3276 cwq->max_active = max_active;
3277 INIT_LIST_HEAD(&cwq->delayed_works);
3280 if (flags & WQ_RESCUER) {
3281 struct worker *rescuer;
3283 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3286 wq->rescuer = rescuer = alloc_worker();
3290 rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3292 if (IS_ERR(rescuer->task))
3295 rescuer->task->flags |= PF_THREAD_BOUND;
3296 wake_up_process(rescuer->task);
3300 * workqueue_lock protects global freeze state and workqueues
3301 * list. Grab it, set max_active accordingly and add the new
3302 * workqueue to workqueues list.
3304 spin_lock(&workqueue_lock);
3306 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3307 for_each_cwq_cpu(cpu, wq)
3308 get_cwq(cpu, wq)->max_active = 0;
3310 list_add(&wq->list, &workqueues);
3312 spin_unlock(&workqueue_lock);
3318 free_mayday_mask(wq->mayday_mask);
3324 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3327 * destroy_workqueue - safely terminate a workqueue
3328 * @wq: target workqueue
3330 * Safely destroy a workqueue. All work currently pending will be done first.
3332 void destroy_workqueue(struct workqueue_struct *wq)
3336 /* drain it before proceeding with destruction */
3337 drain_workqueue(wq);
3340 * wq list is used to freeze wq, remove from list after
3341 * flushing is complete in case freeze races us.
3343 spin_lock(&workqueue_lock);
3344 list_del(&wq->list);
3345 spin_unlock(&workqueue_lock);
3348 for_each_cwq_cpu(cpu, wq) {
3349 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3352 for (i = 0; i < WORK_NR_COLORS; i++)
3353 BUG_ON(cwq->nr_in_flight[i]);
3354 BUG_ON(cwq->nr_active);
3355 BUG_ON(!list_empty(&cwq->delayed_works));
3358 if (wq->flags & WQ_RESCUER) {
3359 kthread_stop(wq->rescuer->task);
3360 free_mayday_mask(wq->mayday_mask);
3367 EXPORT_SYMBOL_GPL(destroy_workqueue);
3370 * workqueue_set_max_active - adjust max_active of a workqueue
3371 * @wq: target workqueue
3372 * @max_active: new max_active value.
3374 * Set max_active of @wq to @max_active.
3377 * Don't call from IRQ context.
3379 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3383 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3385 spin_lock(&workqueue_lock);
3387 wq->saved_max_active = max_active;
3389 for_each_cwq_cpu(cpu, wq) {
3390 struct global_cwq *gcwq = get_gcwq(cpu);
3392 spin_lock_irq(&gcwq->lock);
3394 if (!(wq->flags & WQ_FREEZABLE) ||
3395 !(gcwq->flags & GCWQ_FREEZING))
3396 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3398 spin_unlock_irq(&gcwq->lock);
3401 spin_unlock(&workqueue_lock);
3403 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3406 * workqueue_congested - test whether a workqueue is congested
3407 * @cpu: CPU in question
3408 * @wq: target workqueue
3410 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3411 * no synchronization around this function and the test result is
3412 * unreliable and only useful as advisory hints or for debugging.
3415 * %true if congested, %false otherwise.
3417 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3419 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3421 return !list_empty(&cwq->delayed_works);
3423 EXPORT_SYMBOL_GPL(workqueue_congested);
3426 * work_cpu - return the last known associated cpu for @work
3427 * @work: the work of interest
3430 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3432 unsigned int work_cpu(struct work_struct *work)
3434 struct global_cwq *gcwq = get_work_gcwq(work);
3436 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3438 EXPORT_SYMBOL_GPL(work_cpu);
3441 * work_busy - test whether a work is currently pending or running
3442 * @work: the work to be tested
3444 * Test whether @work is currently pending or running. There is no
3445 * synchronization around this function and the test result is
3446 * unreliable and only useful as advisory hints or for debugging.
3447 * Especially for reentrant wqs, the pending state might hide the
3451 * OR'd bitmask of WORK_BUSY_* bits.
3453 unsigned int work_busy(struct work_struct *work)
3455 struct global_cwq *gcwq = get_work_gcwq(work);
3456 unsigned long flags;
3457 unsigned int ret = 0;
3462 spin_lock_irqsave(&gcwq->lock, flags);
3464 if (work_pending(work))
3465 ret |= WORK_BUSY_PENDING;
3466 if (find_worker_executing_work(gcwq, work))
3467 ret |= WORK_BUSY_RUNNING;
3469 spin_unlock_irqrestore(&gcwq->lock, flags);
3473 EXPORT_SYMBOL_GPL(work_busy);
3478 * There are two challenges in supporting CPU hotplug. Firstly, there
3479 * are a lot of assumptions on strong associations among work, cwq and
3480 * gcwq which make migrating pending and scheduled works very
3481 * difficult to implement without impacting hot paths. Secondly,
3482 * gcwqs serve mix of short, long and very long running works making
3483 * blocked draining impractical.
3485 * This is solved by allowing a gcwq to be disassociated from the CPU
3486 * running as an unbound one and allowing it to be reattached later if the
3487 * cpu comes back online.
3490 /* claim manager positions of all pools */
3491 static void gcwq_claim_assoc_and_lock(struct global_cwq *gcwq)
3493 struct worker_pool *pool;
3495 for_each_worker_pool(pool, gcwq)
3496 mutex_lock_nested(&pool->assoc_mutex, pool - gcwq->pools);
3497 spin_lock_irq(&gcwq->lock);
3500 /* release manager positions */
3501 static void gcwq_release_assoc_and_unlock(struct global_cwq *gcwq)
3503 struct worker_pool *pool;
3505 spin_unlock_irq(&gcwq->lock);
3506 for_each_worker_pool(pool, gcwq)
3507 mutex_unlock(&pool->assoc_mutex);
3510 static void gcwq_unbind_fn(struct work_struct *work)
3512 struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3513 struct worker_pool *pool;
3514 struct worker *worker;
3515 struct hlist_node *pos;
3518 BUG_ON(gcwq->cpu != smp_processor_id());
3520 gcwq_claim_assoc_and_lock(gcwq);
3523 * We've claimed all manager positions. Make all workers unbound
3524 * and set DISASSOCIATED. Before this, all workers except for the
3525 * ones which are still executing works from before the last CPU
3526 * down must be on the cpu. After this, they may become diasporas.
3528 for_each_worker_pool(pool, gcwq)
3529 list_for_each_entry(worker, &pool->idle_list, entry)
3530 worker->flags |= WORKER_UNBOUND;
3532 for_each_busy_worker(worker, i, pos, gcwq)
3533 worker->flags |= WORKER_UNBOUND;
3535 gcwq->flags |= GCWQ_DISASSOCIATED;
3537 gcwq_release_assoc_and_unlock(gcwq);
3540 * Call schedule() so that we cross rq->lock and thus can guarantee
3541 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3542 * as scheduler callbacks may be invoked from other cpus.
3547 * Sched callbacks are disabled now. Zap nr_running. After this,
3548 * nr_running stays zero and need_more_worker() and keep_working()
3549 * are always true as long as the worklist is not empty. @gcwq now
3550 * behaves as unbound (in terms of concurrency management) gcwq
3551 * which is served by workers tied to the CPU.
3553 * On return from this function, the current worker would trigger
3554 * unbound chain execution of pending work items if other workers
3557 for_each_worker_pool(pool, gcwq)
3558 atomic_set(get_pool_nr_running(pool), 0);
3562 * Workqueues should be brought up before normal priority CPU notifiers.
3563 * This will be registered high priority CPU notifier.
3565 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3566 unsigned long action,
3569 unsigned int cpu = (unsigned long)hcpu;
3570 struct global_cwq *gcwq = get_gcwq(cpu);
3571 struct worker_pool *pool;
3573 switch (action & ~CPU_TASKS_FROZEN) {
3574 case CPU_UP_PREPARE:
3575 for_each_worker_pool(pool, gcwq) {
3576 struct worker *worker;
3578 if (pool->nr_workers)
3581 worker = create_worker(pool);
3585 spin_lock_irq(&gcwq->lock);
3586 start_worker(worker);
3587 spin_unlock_irq(&gcwq->lock);
3591 case CPU_DOWN_FAILED:
3593 gcwq_claim_assoc_and_lock(gcwq);
3594 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3595 rebind_workers(gcwq);
3596 gcwq_release_assoc_and_unlock(gcwq);
3603 * Workqueues should be brought down after normal priority CPU notifiers.
3604 * This will be registered as low priority CPU notifier.
3606 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3607 unsigned long action,
3610 unsigned int cpu = (unsigned long)hcpu;
3611 struct work_struct unbind_work;
3613 switch (action & ~CPU_TASKS_FROZEN) {
3614 case CPU_DOWN_PREPARE:
3615 /* unbinding should happen on the local CPU */
3616 INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3617 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3618 flush_work(&unbind_work);
3626 struct work_for_cpu {
3627 struct completion completion;
3633 static int do_work_for_cpu(void *_wfc)
3635 struct work_for_cpu *wfc = _wfc;
3636 wfc->ret = wfc->fn(wfc->arg);
3637 complete(&wfc->completion);
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 task_struct *sub_thread;
3654 struct work_for_cpu wfc = {
3655 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3660 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3661 if (IS_ERR(sub_thread))
3662 return PTR_ERR(sub_thread);
3663 kthread_bind(sub_thread, cpu);
3664 wake_up_process(sub_thread);
3665 wait_for_completion(&wfc.completion);
3668 EXPORT_SYMBOL_GPL(work_on_cpu);
3669 #endif /* CONFIG_SMP */
3671 #ifdef CONFIG_FREEZER
3674 * freeze_workqueues_begin - begin freezing workqueues
3676 * Start freezing workqueues. After this function returns, all freezable
3677 * workqueues will queue new works to their frozen_works list instead of
3681 * Grabs and releases workqueue_lock and gcwq->lock's.
3683 void freeze_workqueues_begin(void)
3687 spin_lock(&workqueue_lock);
3689 BUG_ON(workqueue_freezing);
3690 workqueue_freezing = true;
3692 for_each_gcwq_cpu(cpu) {
3693 struct global_cwq *gcwq = get_gcwq(cpu);
3694 struct workqueue_struct *wq;
3696 spin_lock_irq(&gcwq->lock);
3698 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3699 gcwq->flags |= GCWQ_FREEZING;
3701 list_for_each_entry(wq, &workqueues, list) {
3702 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3704 if (cwq && wq->flags & WQ_FREEZABLE)
3705 cwq->max_active = 0;
3708 spin_unlock_irq(&gcwq->lock);
3711 spin_unlock(&workqueue_lock);
3715 * freeze_workqueues_busy - are freezable workqueues still busy?
3717 * Check whether freezing is complete. This function must be called
3718 * between freeze_workqueues_begin() and thaw_workqueues().
3721 * Grabs and releases workqueue_lock.
3724 * %true if some freezable workqueues are still busy. %false if freezing
3727 bool freeze_workqueues_busy(void)
3732 spin_lock(&workqueue_lock);
3734 BUG_ON(!workqueue_freezing);
3736 for_each_gcwq_cpu(cpu) {
3737 struct workqueue_struct *wq;
3739 * nr_active is monotonically decreasing. It's safe
3740 * to peek without lock.
3742 list_for_each_entry(wq, &workqueues, list) {
3743 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3745 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3748 BUG_ON(cwq->nr_active < 0);
3749 if (cwq->nr_active) {
3756 spin_unlock(&workqueue_lock);
3761 * thaw_workqueues - thaw workqueues
3763 * Thaw workqueues. Normal queueing is restored and all collected
3764 * frozen works are transferred to their respective gcwq worklists.
3767 * Grabs and releases workqueue_lock and gcwq->lock's.
3769 void thaw_workqueues(void)
3773 spin_lock(&workqueue_lock);
3775 if (!workqueue_freezing)
3778 for_each_gcwq_cpu(cpu) {
3779 struct global_cwq *gcwq = get_gcwq(cpu);
3780 struct worker_pool *pool;
3781 struct workqueue_struct *wq;
3783 spin_lock_irq(&gcwq->lock);
3785 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3786 gcwq->flags &= ~GCWQ_FREEZING;
3788 list_for_each_entry(wq, &workqueues, list) {
3789 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3791 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3794 /* restore max_active and repopulate worklist */
3795 cwq->max_active = wq->saved_max_active;
3797 while (!list_empty(&cwq->delayed_works) &&
3798 cwq->nr_active < cwq->max_active)
3799 cwq_activate_first_delayed(cwq);
3802 for_each_worker_pool(pool, gcwq)
3803 wake_up_worker(pool);
3805 spin_unlock_irq(&gcwq->lock);
3808 workqueue_freezing = false;
3810 spin_unlock(&workqueue_lock);
3812 #endif /* CONFIG_FREEZER */
3814 static int __init init_workqueues(void)
3819 /* make sure we have enough bits for OFFQ CPU number */
3820 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
3823 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3824 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3826 /* initialize gcwqs */
3827 for_each_gcwq_cpu(cpu) {
3828 struct global_cwq *gcwq = get_gcwq(cpu);
3829 struct worker_pool *pool;
3831 spin_lock_init(&gcwq->lock);
3833 gcwq->flags |= GCWQ_DISASSOCIATED;
3835 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3836 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3838 for_each_worker_pool(pool, gcwq) {
3840 INIT_LIST_HEAD(&pool->worklist);
3841 INIT_LIST_HEAD(&pool->idle_list);
3843 init_timer_deferrable(&pool->idle_timer);
3844 pool->idle_timer.function = idle_worker_timeout;
3845 pool->idle_timer.data = (unsigned long)pool;
3847 setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3848 (unsigned long)pool);
3850 mutex_init(&pool->assoc_mutex);
3851 ida_init(&pool->worker_ida);
3855 /* create the initial worker */
3856 for_each_online_gcwq_cpu(cpu) {
3857 struct global_cwq *gcwq = get_gcwq(cpu);
3858 struct worker_pool *pool;
3860 if (cpu != WORK_CPU_UNBOUND)
3861 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3863 for_each_worker_pool(pool, gcwq) {
3864 struct worker *worker;
3866 worker = create_worker(pool);
3868 spin_lock_irq(&gcwq->lock);
3869 start_worker(worker);
3870 spin_unlock_irq(&gcwq->lock);
3874 system_wq = alloc_workqueue("events", 0, 0);
3875 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3876 system_long_wq = alloc_workqueue("events_long", 0, 0);
3877 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3878 WQ_UNBOUND_MAX_ACTIVE);
3879 system_freezable_wq = alloc_workqueue("events_freezable",
3881 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3882 !system_unbound_wq || !system_freezable_wq);
3885 early_initcall(init_workqueues);