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
1002 * @delayed: for a delayed work
1004 * A work either has completed or is removed from pending queue,
1005 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1008 * spin_lock_irq(gcwq->lock).
1010 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1013 /* ignore uncolored works */
1014 if (color == WORK_NO_COLOR)
1017 cwq->nr_in_flight[color]--;
1021 if (!list_empty(&cwq->delayed_works)) {
1022 /* one down, submit a delayed one */
1023 if (cwq->nr_active < cwq->max_active)
1024 cwq_activate_first_delayed(cwq);
1028 /* is flush in progress and are we at the flushing tip? */
1029 if (likely(cwq->flush_color != color))
1032 /* are there still in-flight works? */
1033 if (cwq->nr_in_flight[color])
1036 /* this cwq is done, clear flush_color */
1037 cwq->flush_color = -1;
1040 * If this was the last cwq, wake up the first flusher. It
1041 * will handle the rest.
1043 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1044 complete(&cwq->wq->first_flusher->done);
1048 * try_to_grab_pending - steal work item from worklist and disable irq
1049 * @work: work item to steal
1050 * @is_dwork: @work is a delayed_work
1051 * @flags: place to store irq state
1053 * Try to grab PENDING bit of @work. This function can handle @work in any
1054 * stable state - idle, on timer or on worklist. Return values are
1056 * 1 if @work was pending and we successfully stole PENDING
1057 * 0 if @work was idle and we claimed PENDING
1058 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1059 * -ENOENT if someone else is canceling @work, this state may persist
1060 * for arbitrarily long
1062 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1063 * interrupted while holding PENDING and @work off queue, irq must be
1064 * disabled on entry. This, combined with delayed_work->timer being
1065 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1067 * On successful return, >= 0, irq is disabled and the caller is
1068 * responsible for releasing it using local_irq_restore(*@flags).
1070 * This function is safe to call from any context including IRQ handler.
1072 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1073 unsigned long *flags)
1075 struct global_cwq *gcwq;
1077 WARN_ON_ONCE(in_irq());
1079 local_irq_save(*flags);
1081 /* try to steal the timer if it exists */
1083 struct delayed_work *dwork = to_delayed_work(work);
1086 * dwork->timer is irqsafe. If del_timer() fails, it's
1087 * guaranteed that the timer is not queued anywhere and not
1088 * running on the local CPU.
1090 if (likely(del_timer(&dwork->timer)))
1094 /* try to claim PENDING the normal way */
1095 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1099 * The queueing is in progress, or it is already queued. Try to
1100 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1102 gcwq = get_work_gcwq(work);
1106 spin_lock(&gcwq->lock);
1107 if (!list_empty(&work->entry)) {
1109 * This work is queued, but perhaps we locked the wrong gcwq.
1110 * In that case we must see the new value after rmb(), see
1111 * insert_work()->wmb().
1114 if (gcwq == get_work_gcwq(work)) {
1115 debug_work_deactivate(work);
1118 * A delayed work item cannot be grabbed directly
1119 * because it might have linked NO_COLOR work items
1120 * which, if left on the delayed_list, will confuse
1121 * cwq->nr_active management later on and cause
1122 * stall. Make sure the work item is activated
1125 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1126 cwq_activate_delayed_work(work);
1128 list_del_init(&work->entry);
1129 cwq_dec_nr_in_flight(get_work_cwq(work),
1130 get_work_color(work),
1131 *work_data_bits(work) & WORK_STRUCT_DELAYED);
1133 spin_unlock(&gcwq->lock);
1137 spin_unlock(&gcwq->lock);
1139 local_irq_restore(*flags);
1140 if (work_is_canceling(work))
1147 * insert_work - insert a work into gcwq
1148 * @cwq: cwq @work belongs to
1149 * @work: work to insert
1150 * @head: insertion point
1151 * @extra_flags: extra WORK_STRUCT_* flags to set
1153 * Insert @work which belongs to @cwq into @gcwq after @head.
1154 * @extra_flags is or'd to work_struct flags.
1157 * spin_lock_irq(gcwq->lock).
1159 static void insert_work(struct cpu_workqueue_struct *cwq,
1160 struct work_struct *work, struct list_head *head,
1161 unsigned int extra_flags)
1163 struct worker_pool *pool = cwq->pool;
1165 /* we own @work, set data and link */
1166 set_work_cwq(work, cwq, extra_flags);
1169 * Ensure that we get the right work->data if we see the
1170 * result of list_add() below, see try_to_grab_pending().
1174 list_add_tail(&work->entry, head);
1177 * Ensure either worker_sched_deactivated() sees the above
1178 * list_add_tail() or we see zero nr_running to avoid workers
1179 * lying around lazily while there are works to be processed.
1183 if (__need_more_worker(pool))
1184 wake_up_worker(pool);
1188 * Test whether @work is being queued from another work executing on the
1189 * same workqueue. This is rather expensive and should only be used from
1192 static bool is_chained_work(struct workqueue_struct *wq)
1194 unsigned long flags;
1197 for_each_gcwq_cpu(cpu) {
1198 struct global_cwq *gcwq = get_gcwq(cpu);
1199 struct worker *worker;
1200 struct hlist_node *pos;
1203 spin_lock_irqsave(&gcwq->lock, flags);
1204 for_each_busy_worker(worker, i, pos, gcwq) {
1205 if (worker->task != current)
1207 spin_unlock_irqrestore(&gcwq->lock, flags);
1209 * I'm @worker, no locking necessary. See if @work
1210 * is headed to the same workqueue.
1212 return worker->current_cwq->wq == wq;
1214 spin_unlock_irqrestore(&gcwq->lock, flags);
1219 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1220 struct work_struct *work)
1222 struct global_cwq *gcwq;
1223 struct cpu_workqueue_struct *cwq;
1224 struct list_head *worklist;
1225 unsigned int work_flags;
1226 unsigned int req_cpu = cpu;
1229 * While a work item is PENDING && off queue, a task trying to
1230 * steal the PENDING will busy-loop waiting for it to either get
1231 * queued or lose PENDING. Grabbing PENDING and queueing should
1232 * happen with IRQ disabled.
1234 WARN_ON_ONCE(!irqs_disabled());
1236 debug_work_activate(work);
1238 /* if dying, only works from the same workqueue are allowed */
1239 if (unlikely(wq->flags & WQ_DRAINING) &&
1240 WARN_ON_ONCE(!is_chained_work(wq)))
1243 /* determine gcwq to use */
1244 if (!(wq->flags & WQ_UNBOUND)) {
1245 struct global_cwq *last_gcwq;
1247 if (cpu == WORK_CPU_UNBOUND)
1248 cpu = raw_smp_processor_id();
1251 * It's multi cpu. If @work was previously on a different
1252 * cpu, it might still be running there, in which case the
1253 * work needs to be queued on that cpu to guarantee
1256 gcwq = get_gcwq(cpu);
1257 last_gcwq = get_work_gcwq(work);
1259 if (last_gcwq && last_gcwq != gcwq) {
1260 struct worker *worker;
1262 spin_lock(&last_gcwq->lock);
1264 worker = find_worker_executing_work(last_gcwq, work);
1266 if (worker && worker->current_cwq->wq == wq)
1269 /* meh... not running there, queue here */
1270 spin_unlock(&last_gcwq->lock);
1271 spin_lock(&gcwq->lock);
1274 spin_lock(&gcwq->lock);
1277 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1278 spin_lock(&gcwq->lock);
1281 /* gcwq determined, get cwq and queue */
1282 cwq = get_cwq(gcwq->cpu, wq);
1283 trace_workqueue_queue_work(req_cpu, cwq, work);
1285 if (WARN_ON(!list_empty(&work->entry))) {
1286 spin_unlock(&gcwq->lock);
1290 cwq->nr_in_flight[cwq->work_color]++;
1291 work_flags = work_color_to_flags(cwq->work_color);
1293 if (likely(cwq->nr_active < cwq->max_active)) {
1294 trace_workqueue_activate_work(work);
1296 worklist = &cwq->pool->worklist;
1298 work_flags |= WORK_STRUCT_DELAYED;
1299 worklist = &cwq->delayed_works;
1302 insert_work(cwq, work, worklist, work_flags);
1304 spin_unlock(&gcwq->lock);
1308 * queue_work_on - queue work on specific cpu
1309 * @cpu: CPU number to execute work on
1310 * @wq: workqueue to use
1311 * @work: work to queue
1313 * Returns %false if @work was already on a queue, %true otherwise.
1315 * We queue the work to a specific CPU, the caller must ensure it
1318 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1319 struct work_struct *work)
1322 unsigned long flags;
1324 local_irq_save(flags);
1326 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1327 __queue_work(cpu, wq, work);
1331 local_irq_restore(flags);
1334 EXPORT_SYMBOL_GPL(queue_work_on);
1337 * queue_work - queue work on a workqueue
1338 * @wq: workqueue to use
1339 * @work: work to queue
1341 * Returns %false if @work was already on a queue, %true otherwise.
1343 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1344 * it can be processed by another CPU.
1346 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1348 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1350 EXPORT_SYMBOL_GPL(queue_work);
1352 void delayed_work_timer_fn(unsigned long __data)
1354 struct delayed_work *dwork = (struct delayed_work *)__data;
1355 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1357 /* should have been called from irqsafe timer with irq already off */
1358 __queue_work(dwork->cpu, cwq->wq, &dwork->work);
1360 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1362 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1363 struct delayed_work *dwork, unsigned long delay)
1365 struct timer_list *timer = &dwork->timer;
1366 struct work_struct *work = &dwork->work;
1369 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1370 timer->data != (unsigned long)dwork);
1371 BUG_ON(timer_pending(timer));
1372 BUG_ON(!list_empty(&work->entry));
1374 timer_stats_timer_set_start_info(&dwork->timer);
1377 * This stores cwq for the moment, for the timer_fn. Note that the
1378 * work's gcwq is preserved to allow reentrance detection for
1381 if (!(wq->flags & WQ_UNBOUND)) {
1382 struct global_cwq *gcwq = get_work_gcwq(work);
1385 * If we cannot get the last gcwq from @work directly,
1386 * select the last CPU such that it avoids unnecessarily
1387 * triggering non-reentrancy check in __queue_work().
1392 if (lcpu == WORK_CPU_UNBOUND)
1393 lcpu = raw_smp_processor_id();
1395 lcpu = WORK_CPU_UNBOUND;
1398 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1401 timer->expires = jiffies + delay;
1403 if (unlikely(cpu != WORK_CPU_UNBOUND))
1404 add_timer_on(timer, cpu);
1410 * queue_delayed_work_on - queue work on specific CPU after delay
1411 * @cpu: CPU number to execute work on
1412 * @wq: workqueue to use
1413 * @dwork: work to queue
1414 * @delay: number of jiffies to wait before queueing
1416 * Returns %false if @work was already on a queue, %true otherwise. If
1417 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1420 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1421 struct delayed_work *dwork, unsigned long delay)
1423 struct work_struct *work = &dwork->work;
1425 unsigned long flags;
1428 return queue_work_on(cpu, wq, &dwork->work);
1430 /* read the comment in __queue_work() */
1431 local_irq_save(flags);
1433 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1434 __queue_delayed_work(cpu, wq, dwork, delay);
1438 local_irq_restore(flags);
1441 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1444 * queue_delayed_work - queue work on a workqueue after delay
1445 * @wq: workqueue to use
1446 * @dwork: delayable work to queue
1447 * @delay: number of jiffies to wait before queueing
1449 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1451 bool queue_delayed_work(struct workqueue_struct *wq,
1452 struct delayed_work *dwork, unsigned long delay)
1454 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1456 EXPORT_SYMBOL_GPL(queue_delayed_work);
1459 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1460 * @cpu: CPU number to execute work on
1461 * @wq: workqueue to use
1462 * @dwork: work to queue
1463 * @delay: number of jiffies to wait before queueing
1465 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1466 * modify @dwork's timer so that it expires after @delay. If @delay is
1467 * zero, @work is guaranteed to be scheduled immediately regardless of its
1470 * Returns %false if @dwork was idle and queued, %true if @dwork was
1471 * pending and its timer was modified.
1473 * This function is safe to call from any context including IRQ handler.
1474 * See try_to_grab_pending() for details.
1476 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1477 struct delayed_work *dwork, unsigned long delay)
1479 unsigned long flags;
1483 ret = try_to_grab_pending(&dwork->work, true, &flags);
1484 } while (unlikely(ret == -EAGAIN));
1486 if (likely(ret >= 0)) {
1487 __queue_delayed_work(cpu, wq, dwork, delay);
1488 local_irq_restore(flags);
1491 /* -ENOENT from try_to_grab_pending() becomes %true */
1494 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1497 * mod_delayed_work - modify delay of or queue a delayed work
1498 * @wq: workqueue to use
1499 * @dwork: work to queue
1500 * @delay: number of jiffies to wait before queueing
1502 * mod_delayed_work_on() on local CPU.
1504 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1505 unsigned long delay)
1507 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1509 EXPORT_SYMBOL_GPL(mod_delayed_work);
1512 * worker_enter_idle - enter idle state
1513 * @worker: worker which is entering idle state
1515 * @worker is entering idle state. Update stats and idle timer if
1519 * spin_lock_irq(gcwq->lock).
1521 static void worker_enter_idle(struct worker *worker)
1523 struct worker_pool *pool = worker->pool;
1524 struct global_cwq *gcwq = pool->gcwq;
1526 BUG_ON(worker->flags & WORKER_IDLE);
1527 BUG_ON(!list_empty(&worker->entry) &&
1528 (worker->hentry.next || worker->hentry.pprev));
1530 /* can't use worker_set_flags(), also called from start_worker() */
1531 worker->flags |= WORKER_IDLE;
1533 worker->last_active = jiffies;
1535 /* idle_list is LIFO */
1536 list_add(&worker->entry, &pool->idle_list);
1538 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1539 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1542 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1543 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1544 * nr_running, the warning may trigger spuriously. Check iff
1545 * unbind is not in progress.
1547 WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
1548 pool->nr_workers == pool->nr_idle &&
1549 atomic_read(get_pool_nr_running(pool)));
1553 * worker_leave_idle - leave idle state
1554 * @worker: worker which is leaving idle state
1556 * @worker is leaving idle state. Update stats.
1559 * spin_lock_irq(gcwq->lock).
1561 static void worker_leave_idle(struct worker *worker)
1563 struct worker_pool *pool = worker->pool;
1565 BUG_ON(!(worker->flags & WORKER_IDLE));
1566 worker_clr_flags(worker, WORKER_IDLE);
1568 list_del_init(&worker->entry);
1572 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1575 * Works which are scheduled while the cpu is online must at least be
1576 * scheduled to a worker which is bound to the cpu so that if they are
1577 * flushed from cpu callbacks while cpu is going down, they are
1578 * guaranteed to execute on the cpu.
1580 * This function is to be used by rogue workers and rescuers to bind
1581 * themselves to the target cpu and may race with cpu going down or
1582 * coming online. kthread_bind() can't be used because it may put the
1583 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1584 * verbatim as it's best effort and blocking and gcwq may be
1585 * [dis]associated in the meantime.
1587 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1588 * binding against %GCWQ_DISASSOCIATED which is set during
1589 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1590 * enters idle state or fetches works without dropping lock, it can
1591 * guarantee the scheduling requirement described in the first paragraph.
1594 * Might sleep. Called without any lock but returns with gcwq->lock
1598 * %true if the associated gcwq is online (@worker is successfully
1599 * bound), %false if offline.
1601 static bool worker_maybe_bind_and_lock(struct worker *worker)
1602 __acquires(&gcwq->lock)
1604 struct global_cwq *gcwq = worker->pool->gcwq;
1605 struct task_struct *task = worker->task;
1609 * The following call may fail, succeed or succeed
1610 * without actually migrating the task to the cpu if
1611 * it races with cpu hotunplug operation. Verify
1612 * against GCWQ_DISASSOCIATED.
1614 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1615 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1617 spin_lock_irq(&gcwq->lock);
1618 if (gcwq->flags & GCWQ_DISASSOCIATED)
1620 if (task_cpu(task) == gcwq->cpu &&
1621 cpumask_equal(¤t->cpus_allowed,
1622 get_cpu_mask(gcwq->cpu)))
1624 spin_unlock_irq(&gcwq->lock);
1627 * We've raced with CPU hot[un]plug. Give it a breather
1628 * and retry migration. cond_resched() is required here;
1629 * otherwise, we might deadlock against cpu_stop trying to
1630 * bring down the CPU on non-preemptive kernel.
1638 * Rebind an idle @worker to its CPU. worker_thread() will test
1639 * list_empty(@worker->entry) before leaving idle and call this function.
1641 static void idle_worker_rebind(struct worker *worker)
1643 struct global_cwq *gcwq = worker->pool->gcwq;
1645 /* CPU may go down again inbetween, clear UNBOUND only on success */
1646 if (worker_maybe_bind_and_lock(worker))
1647 worker_clr_flags(worker, WORKER_UNBOUND);
1649 /* rebind complete, become available again */
1650 list_add(&worker->entry, &worker->pool->idle_list);
1651 spin_unlock_irq(&gcwq->lock);
1655 * Function for @worker->rebind.work used to rebind unbound busy workers to
1656 * the associated cpu which is coming back online. This is scheduled by
1657 * cpu up but can race with other cpu hotplug operations and may be
1658 * executed twice without intervening cpu down.
1660 static void busy_worker_rebind_fn(struct work_struct *work)
1662 struct worker *worker = container_of(work, struct worker, rebind_work);
1663 struct global_cwq *gcwq = worker->pool->gcwq;
1665 if (worker_maybe_bind_and_lock(worker))
1666 worker_clr_flags(worker, WORKER_UNBOUND);
1668 spin_unlock_irq(&gcwq->lock);
1672 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1673 * @gcwq: gcwq of interest
1675 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1676 * is different for idle and busy ones.
1678 * Idle ones will be removed from the idle_list and woken up. They will
1679 * add themselves back after completing rebind. This ensures that the
1680 * idle_list doesn't contain any unbound workers when re-bound busy workers
1681 * try to perform local wake-ups for concurrency management.
1683 * Busy workers can rebind after they finish their current work items.
1684 * Queueing the rebind work item at the head of the scheduled list is
1685 * enough. Note that nr_running will be properly bumped as busy workers
1688 * On return, all non-manager workers are scheduled for rebind - see
1689 * manage_workers() for the manager special case. Any idle worker
1690 * including the manager will not appear on @idle_list until rebind is
1691 * complete, making local wake-ups safe.
1693 static void rebind_workers(struct global_cwq *gcwq)
1695 struct worker_pool *pool;
1696 struct worker *worker, *n;
1697 struct hlist_node *pos;
1700 lockdep_assert_held(&gcwq->lock);
1702 for_each_worker_pool(pool, gcwq)
1703 lockdep_assert_held(&pool->assoc_mutex);
1705 /* dequeue and kick idle ones */
1706 for_each_worker_pool(pool, gcwq) {
1707 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1709 * idle workers should be off @pool->idle_list
1710 * until rebind is complete to avoid receiving
1711 * premature local wake-ups.
1713 list_del_init(&worker->entry);
1716 * worker_thread() will see the above dequeuing
1717 * and call idle_worker_rebind().
1719 wake_up_process(worker->task);
1723 /* rebind busy workers */
1724 for_each_busy_worker(worker, i, pos, gcwq) {
1725 struct work_struct *rebind_work = &worker->rebind_work;
1726 struct workqueue_struct *wq;
1728 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1729 work_data_bits(rebind_work)))
1732 debug_work_activate(rebind_work);
1735 * wq doesn't really matter but let's keep @worker->pool
1736 * and @cwq->pool consistent for sanity.
1738 if (worker_pool_pri(worker->pool))
1739 wq = system_highpri_wq;
1743 insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
1744 worker->scheduled.next,
1745 work_color_to_flags(WORK_NO_COLOR));
1749 static struct worker *alloc_worker(void)
1751 struct worker *worker;
1753 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1755 INIT_LIST_HEAD(&worker->entry);
1756 INIT_LIST_HEAD(&worker->scheduled);
1757 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1758 /* on creation a worker is in !idle && prep state */
1759 worker->flags = WORKER_PREP;
1765 * create_worker - create a new workqueue worker
1766 * @pool: pool the new worker will belong to
1768 * Create a new worker which is bound to @pool. The returned worker
1769 * can be started by calling start_worker() or destroyed using
1773 * Might sleep. Does GFP_KERNEL allocations.
1776 * Pointer to the newly created worker.
1778 static struct worker *create_worker(struct worker_pool *pool)
1780 struct global_cwq *gcwq = pool->gcwq;
1781 const char *pri = worker_pool_pri(pool) ? "H" : "";
1782 struct worker *worker = NULL;
1785 spin_lock_irq(&gcwq->lock);
1786 while (ida_get_new(&pool->worker_ida, &id)) {
1787 spin_unlock_irq(&gcwq->lock);
1788 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1790 spin_lock_irq(&gcwq->lock);
1792 spin_unlock_irq(&gcwq->lock);
1794 worker = alloc_worker();
1798 worker->pool = pool;
1801 if (gcwq->cpu != WORK_CPU_UNBOUND)
1802 worker->task = kthread_create_on_node(worker_thread,
1803 worker, cpu_to_node(gcwq->cpu),
1804 "kworker/%u:%d%s", gcwq->cpu, id, pri);
1806 worker->task = kthread_create(worker_thread, worker,
1807 "kworker/u:%d%s", id, pri);
1808 if (IS_ERR(worker->task))
1811 if (worker_pool_pri(pool))
1812 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1815 * Determine CPU binding of the new worker depending on
1816 * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1817 * flag remains stable across this function. See the comments
1818 * above the flag definition for details.
1820 * As an unbound worker may later become a regular one if CPU comes
1821 * online, make sure every worker has %PF_THREAD_BOUND set.
1823 if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
1824 kthread_bind(worker->task, gcwq->cpu);
1826 worker->task->flags |= PF_THREAD_BOUND;
1827 worker->flags |= WORKER_UNBOUND;
1833 spin_lock_irq(&gcwq->lock);
1834 ida_remove(&pool->worker_ida, id);
1835 spin_unlock_irq(&gcwq->lock);
1842 * start_worker - start a newly created worker
1843 * @worker: worker to start
1845 * Make the gcwq aware of @worker and start it.
1848 * spin_lock_irq(gcwq->lock).
1850 static void start_worker(struct worker *worker)
1852 worker->flags |= WORKER_STARTED;
1853 worker->pool->nr_workers++;
1854 worker_enter_idle(worker);
1855 wake_up_process(worker->task);
1859 * destroy_worker - destroy a workqueue worker
1860 * @worker: worker to be destroyed
1862 * Destroy @worker and adjust @gcwq stats accordingly.
1865 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1867 static void destroy_worker(struct worker *worker)
1869 struct worker_pool *pool = worker->pool;
1870 struct global_cwq *gcwq = pool->gcwq;
1871 int id = worker->id;
1873 /* sanity check frenzy */
1874 BUG_ON(worker->current_work);
1875 BUG_ON(!list_empty(&worker->scheduled));
1877 if (worker->flags & WORKER_STARTED)
1879 if (worker->flags & WORKER_IDLE)
1882 list_del_init(&worker->entry);
1883 worker->flags |= WORKER_DIE;
1885 spin_unlock_irq(&gcwq->lock);
1887 kthread_stop(worker->task);
1890 spin_lock_irq(&gcwq->lock);
1891 ida_remove(&pool->worker_ida, id);
1894 static void idle_worker_timeout(unsigned long __pool)
1896 struct worker_pool *pool = (void *)__pool;
1897 struct global_cwq *gcwq = pool->gcwq;
1899 spin_lock_irq(&gcwq->lock);
1901 if (too_many_workers(pool)) {
1902 struct worker *worker;
1903 unsigned long expires;
1905 /* idle_list is kept in LIFO order, check the last one */
1906 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1907 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1909 if (time_before(jiffies, expires))
1910 mod_timer(&pool->idle_timer, expires);
1912 /* it's been idle for too long, wake up manager */
1913 pool->flags |= POOL_MANAGE_WORKERS;
1914 wake_up_worker(pool);
1918 spin_unlock_irq(&gcwq->lock);
1921 static bool send_mayday(struct work_struct *work)
1923 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1924 struct workqueue_struct *wq = cwq->wq;
1927 if (!(wq->flags & WQ_RESCUER))
1930 /* mayday mayday mayday */
1931 cpu = cwq->pool->gcwq->cpu;
1932 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1933 if (cpu == WORK_CPU_UNBOUND)
1935 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1936 wake_up_process(wq->rescuer->task);
1940 static void gcwq_mayday_timeout(unsigned long __pool)
1942 struct worker_pool *pool = (void *)__pool;
1943 struct global_cwq *gcwq = pool->gcwq;
1944 struct work_struct *work;
1946 spin_lock_irq(&gcwq->lock);
1948 if (need_to_create_worker(pool)) {
1950 * We've been trying to create a new worker but
1951 * haven't been successful. We might be hitting an
1952 * allocation deadlock. Send distress signals to
1955 list_for_each_entry(work, &pool->worklist, entry)
1959 spin_unlock_irq(&gcwq->lock);
1961 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1965 * maybe_create_worker - create a new worker if necessary
1966 * @pool: pool to create a new worker for
1968 * Create a new worker for @pool if necessary. @pool is guaranteed to
1969 * have at least one idle worker on return from this function. If
1970 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1971 * sent to all rescuers with works scheduled on @pool to resolve
1972 * possible allocation deadlock.
1974 * On return, need_to_create_worker() is guaranteed to be false and
1975 * may_start_working() true.
1978 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1979 * multiple times. Does GFP_KERNEL allocations. Called only from
1983 * false if no action was taken and gcwq->lock stayed locked, true
1986 static bool maybe_create_worker(struct worker_pool *pool)
1987 __releases(&gcwq->lock)
1988 __acquires(&gcwq->lock)
1990 struct global_cwq *gcwq = pool->gcwq;
1992 if (!need_to_create_worker(pool))
1995 spin_unlock_irq(&gcwq->lock);
1997 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1998 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2001 struct worker *worker;
2003 worker = create_worker(pool);
2005 del_timer_sync(&pool->mayday_timer);
2006 spin_lock_irq(&gcwq->lock);
2007 start_worker(worker);
2008 BUG_ON(need_to_create_worker(pool));
2012 if (!need_to_create_worker(pool))
2015 __set_current_state(TASK_INTERRUPTIBLE);
2016 schedule_timeout(CREATE_COOLDOWN);
2018 if (!need_to_create_worker(pool))
2022 del_timer_sync(&pool->mayday_timer);
2023 spin_lock_irq(&gcwq->lock);
2024 if (need_to_create_worker(pool))
2030 * maybe_destroy_worker - destroy workers which have been idle for a while
2031 * @pool: pool to destroy workers for
2033 * Destroy @pool workers which have been idle for longer than
2034 * IDLE_WORKER_TIMEOUT.
2037 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2038 * multiple times. Called only from manager.
2041 * false if no action was taken and gcwq->lock stayed locked, true
2044 static bool maybe_destroy_workers(struct worker_pool *pool)
2048 while (too_many_workers(pool)) {
2049 struct worker *worker;
2050 unsigned long expires;
2052 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2053 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2055 if (time_before(jiffies, expires)) {
2056 mod_timer(&pool->idle_timer, expires);
2060 destroy_worker(worker);
2068 * manage_workers - manage worker pool
2071 * Assume the manager role and manage gcwq worker pool @worker belongs
2072 * to. At any given time, there can be only zero or one manager per
2073 * gcwq. The exclusion is handled automatically by this function.
2075 * The caller can safely start processing works on false return. On
2076 * true return, it's guaranteed that need_to_create_worker() is false
2077 * and may_start_working() is true.
2080 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2081 * multiple times. Does GFP_KERNEL allocations.
2084 * false if no action was taken and gcwq->lock stayed locked, true if
2085 * some action was taken.
2087 static bool manage_workers(struct worker *worker)
2089 struct worker_pool *pool = worker->pool;
2092 if (pool->flags & POOL_MANAGING_WORKERS)
2095 pool->flags |= POOL_MANAGING_WORKERS;
2098 * To simplify both worker management and CPU hotplug, hold off
2099 * management while hotplug is in progress. CPU hotplug path can't
2100 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2101 * lead to idle worker depletion (all become busy thinking someone
2102 * else is managing) which in turn can result in deadlock under
2103 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2104 * manager against CPU hotplug.
2106 * assoc_mutex would always be free unless CPU hotplug is in
2107 * progress. trylock first without dropping @gcwq->lock.
2109 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2110 spin_unlock_irq(&pool->gcwq->lock);
2111 mutex_lock(&pool->assoc_mutex);
2113 * CPU hotplug could have happened while we were waiting
2114 * for assoc_mutex. Hotplug itself can't handle us
2115 * because manager isn't either on idle or busy list, and
2116 * @gcwq's state and ours could have deviated.
2118 * As hotplug is now excluded via assoc_mutex, we can
2119 * simply try to bind. It will succeed or fail depending
2120 * on @gcwq's current state. Try it and adjust
2121 * %WORKER_UNBOUND accordingly.
2123 if (worker_maybe_bind_and_lock(worker))
2124 worker->flags &= ~WORKER_UNBOUND;
2126 worker->flags |= WORKER_UNBOUND;
2131 pool->flags &= ~POOL_MANAGE_WORKERS;
2134 * Destroy and then create so that may_start_working() is true
2137 ret |= maybe_destroy_workers(pool);
2138 ret |= maybe_create_worker(pool);
2140 pool->flags &= ~POOL_MANAGING_WORKERS;
2141 mutex_unlock(&pool->assoc_mutex);
2146 * process_one_work - process single work
2148 * @work: work to process
2150 * Process @work. This function contains all the logics necessary to
2151 * process a single work including synchronization against and
2152 * interaction with other workers on the same cpu, queueing and
2153 * flushing. As long as context requirement is met, any worker can
2154 * call this function to process a work.
2157 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2159 static void process_one_work(struct worker *worker, struct work_struct *work)
2160 __releases(&gcwq->lock)
2161 __acquires(&gcwq->lock)
2163 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2164 struct worker_pool *pool = worker->pool;
2165 struct global_cwq *gcwq = pool->gcwq;
2166 struct hlist_head *bwh = busy_worker_head(gcwq, work);
2167 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2168 work_func_t f = work->func;
2170 struct worker *collision;
2171 #ifdef CONFIG_LOCKDEP
2173 * It is permissible to free the struct work_struct from
2174 * inside the function that is called from it, this we need to
2175 * take into account for lockdep too. To avoid bogus "held
2176 * lock freed" warnings as well as problems when looking into
2177 * work->lockdep_map, make a copy and use that here.
2179 struct lockdep_map lockdep_map;
2181 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2184 * Ensure we're on the correct CPU. DISASSOCIATED test is
2185 * necessary to avoid spurious warnings from rescuers servicing the
2186 * unbound or a disassociated gcwq.
2188 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2189 !(gcwq->flags & GCWQ_DISASSOCIATED) &&
2190 raw_smp_processor_id() != gcwq->cpu);
2193 * A single work shouldn't be executed concurrently by
2194 * multiple workers on a single cpu. Check whether anyone is
2195 * already processing the work. If so, defer the work to the
2196 * currently executing one.
2198 collision = __find_worker_executing_work(gcwq, bwh, work);
2199 if (unlikely(collision)) {
2200 move_linked_works(work, &collision->scheduled, NULL);
2204 /* claim and dequeue */
2205 debug_work_deactivate(work);
2206 hlist_add_head(&worker->hentry, bwh);
2207 worker->current_work = work;
2208 worker->current_cwq = cwq;
2209 work_color = get_work_color(work);
2211 list_del_init(&work->entry);
2214 * CPU intensive works don't participate in concurrency
2215 * management. They're the scheduler's responsibility.
2217 if (unlikely(cpu_intensive))
2218 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2221 * Unbound gcwq isn't concurrency managed and work items should be
2222 * executed ASAP. Wake up another worker if necessary.
2224 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2225 wake_up_worker(pool);
2228 * Record the last CPU and clear PENDING which should be the last
2229 * update to @work. Also, do this inside @gcwq->lock so that
2230 * PENDING and queued state changes happen together while IRQ is
2233 set_work_cpu_and_clear_pending(work, gcwq->cpu);
2235 spin_unlock_irq(&gcwq->lock);
2237 lock_map_acquire_read(&cwq->wq->lockdep_map);
2238 lock_map_acquire(&lockdep_map);
2239 trace_workqueue_execute_start(work);
2242 * While we must be careful to not use "work" after this, the trace
2243 * point will only record its address.
2245 trace_workqueue_execute_end(work);
2246 lock_map_release(&lockdep_map);
2247 lock_map_release(&cwq->wq->lockdep_map);
2249 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2250 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2251 " last function: %pf\n",
2252 current->comm, preempt_count(), task_pid_nr(current), f);
2253 debug_show_held_locks(current);
2257 spin_lock_irq(&gcwq->lock);
2259 /* clear cpu intensive status */
2260 if (unlikely(cpu_intensive))
2261 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2263 /* we're done with it, release */
2264 hlist_del_init(&worker->hentry);
2265 worker->current_work = NULL;
2266 worker->current_cwq = NULL;
2267 cwq_dec_nr_in_flight(cwq, work_color, false);
2271 * process_scheduled_works - process scheduled works
2274 * Process all scheduled works. Please note that the scheduled list
2275 * may change while processing a work, so this function repeatedly
2276 * fetches a work from the top and executes it.
2279 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2282 static void process_scheduled_works(struct worker *worker)
2284 while (!list_empty(&worker->scheduled)) {
2285 struct work_struct *work = list_first_entry(&worker->scheduled,
2286 struct work_struct, entry);
2287 process_one_work(worker, work);
2292 * worker_thread - the worker thread function
2295 * The gcwq worker thread function. There's a single dynamic pool of
2296 * these per each cpu. These workers process all works regardless of
2297 * their specific target workqueue. The only exception is works which
2298 * belong to workqueues with a rescuer which will be explained in
2301 static int worker_thread(void *__worker)
2303 struct worker *worker = __worker;
2304 struct worker_pool *pool = worker->pool;
2305 struct global_cwq *gcwq = pool->gcwq;
2307 /* tell the scheduler that this is a workqueue worker */
2308 worker->task->flags |= PF_WQ_WORKER;
2310 spin_lock_irq(&gcwq->lock);
2312 /* we are off idle list if destruction or rebind is requested */
2313 if (unlikely(list_empty(&worker->entry))) {
2314 spin_unlock_irq(&gcwq->lock);
2316 /* if DIE is set, destruction is requested */
2317 if (worker->flags & WORKER_DIE) {
2318 worker->task->flags &= ~PF_WQ_WORKER;
2322 /* otherwise, rebind */
2323 idle_worker_rebind(worker);
2327 worker_leave_idle(worker);
2329 /* no more worker necessary? */
2330 if (!need_more_worker(pool))
2333 /* do we need to manage? */
2334 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2338 * ->scheduled list can only be filled while a worker is
2339 * preparing to process a work or actually processing it.
2340 * Make sure nobody diddled with it while I was sleeping.
2342 BUG_ON(!list_empty(&worker->scheduled));
2345 * When control reaches this point, we're guaranteed to have
2346 * at least one idle worker or that someone else has already
2347 * assumed the manager role.
2349 worker_clr_flags(worker, WORKER_PREP);
2352 struct work_struct *work =
2353 list_first_entry(&pool->worklist,
2354 struct work_struct, entry);
2356 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2357 /* optimization path, not strictly necessary */
2358 process_one_work(worker, work);
2359 if (unlikely(!list_empty(&worker->scheduled)))
2360 process_scheduled_works(worker);
2362 move_linked_works(work, &worker->scheduled, NULL);
2363 process_scheduled_works(worker);
2365 } while (keep_working(pool));
2367 worker_set_flags(worker, WORKER_PREP, false);
2369 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2373 * gcwq->lock is held and there's no work to process and no
2374 * need to manage, sleep. Workers are woken up only while
2375 * holding gcwq->lock or from local cpu, so setting the
2376 * current state before releasing gcwq->lock is enough to
2377 * prevent losing any event.
2379 worker_enter_idle(worker);
2380 __set_current_state(TASK_INTERRUPTIBLE);
2381 spin_unlock_irq(&gcwq->lock);
2387 * rescuer_thread - the rescuer thread function
2388 * @__wq: the associated workqueue
2390 * Workqueue rescuer thread function. There's one rescuer for each
2391 * workqueue which has WQ_RESCUER set.
2393 * Regular work processing on a gcwq may block trying to create a new
2394 * worker which uses GFP_KERNEL allocation which has slight chance of
2395 * developing into deadlock if some works currently on the same queue
2396 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2397 * the problem rescuer solves.
2399 * When such condition is possible, the gcwq summons rescuers of all
2400 * workqueues which have works queued on the gcwq and let them process
2401 * those works so that forward progress can be guaranteed.
2403 * This should happen rarely.
2405 static int rescuer_thread(void *__wq)
2407 struct workqueue_struct *wq = __wq;
2408 struct worker *rescuer = wq->rescuer;
2409 struct list_head *scheduled = &rescuer->scheduled;
2410 bool is_unbound = wq->flags & WQ_UNBOUND;
2413 set_user_nice(current, RESCUER_NICE_LEVEL);
2415 set_current_state(TASK_INTERRUPTIBLE);
2417 if (kthread_should_stop())
2421 * See whether any cpu is asking for help. Unbounded
2422 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2424 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2425 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2426 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2427 struct worker_pool *pool = cwq->pool;
2428 struct global_cwq *gcwq = pool->gcwq;
2429 struct work_struct *work, *n;
2431 __set_current_state(TASK_RUNNING);
2432 mayday_clear_cpu(cpu, wq->mayday_mask);
2434 /* migrate to the target cpu if possible */
2435 rescuer->pool = pool;
2436 worker_maybe_bind_and_lock(rescuer);
2439 * Slurp in all works issued via this workqueue and
2442 BUG_ON(!list_empty(&rescuer->scheduled));
2443 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2444 if (get_work_cwq(work) == cwq)
2445 move_linked_works(work, scheduled, &n);
2447 process_scheduled_works(rescuer);
2450 * Leave this gcwq. If keep_working() is %true, notify a
2451 * regular worker; otherwise, we end up with 0 concurrency
2452 * and stalling the execution.
2454 if (keep_working(pool))
2455 wake_up_worker(pool);
2457 spin_unlock_irq(&gcwq->lock);
2465 struct work_struct work;
2466 struct completion done;
2469 static void wq_barrier_func(struct work_struct *work)
2471 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2472 complete(&barr->done);
2476 * insert_wq_barrier - insert a barrier work
2477 * @cwq: cwq to insert barrier into
2478 * @barr: wq_barrier to insert
2479 * @target: target work to attach @barr to
2480 * @worker: worker currently executing @target, NULL if @target is not executing
2482 * @barr is linked to @target such that @barr is completed only after
2483 * @target finishes execution. Please note that the ordering
2484 * guarantee is observed only with respect to @target and on the local
2487 * Currently, a queued barrier can't be canceled. This is because
2488 * try_to_grab_pending() can't determine whether the work to be
2489 * grabbed is at the head of the queue and thus can't clear LINKED
2490 * flag of the previous work while there must be a valid next work
2491 * after a work with LINKED flag set.
2493 * Note that when @worker is non-NULL, @target may be modified
2494 * underneath us, so we can't reliably determine cwq from @target.
2497 * spin_lock_irq(gcwq->lock).
2499 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2500 struct wq_barrier *barr,
2501 struct work_struct *target, struct worker *worker)
2503 struct list_head *head;
2504 unsigned int linked = 0;
2507 * debugobject calls are safe here even with gcwq->lock locked
2508 * as we know for sure that this will not trigger any of the
2509 * checks and call back into the fixup functions where we
2512 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2513 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2514 init_completion(&barr->done);
2517 * If @target is currently being executed, schedule the
2518 * barrier to the worker; otherwise, put it after @target.
2521 head = worker->scheduled.next;
2523 unsigned long *bits = work_data_bits(target);
2525 head = target->entry.next;
2526 /* there can already be other linked works, inherit and set */
2527 linked = *bits & WORK_STRUCT_LINKED;
2528 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2531 debug_work_activate(&barr->work);
2532 insert_work(cwq, &barr->work, head,
2533 work_color_to_flags(WORK_NO_COLOR) | linked);
2537 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2538 * @wq: workqueue being flushed
2539 * @flush_color: new flush color, < 0 for no-op
2540 * @work_color: new work color, < 0 for no-op
2542 * Prepare cwqs for workqueue flushing.
2544 * If @flush_color is non-negative, flush_color on all cwqs should be
2545 * -1. If no cwq has in-flight commands at the specified color, all
2546 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2547 * has in flight commands, its cwq->flush_color is set to
2548 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2549 * wakeup logic is armed and %true is returned.
2551 * The caller should have initialized @wq->first_flusher prior to
2552 * calling this function with non-negative @flush_color. If
2553 * @flush_color is negative, no flush color update is done and %false
2556 * If @work_color is non-negative, all cwqs should have the same
2557 * work_color which is previous to @work_color and all will be
2558 * advanced to @work_color.
2561 * mutex_lock(wq->flush_mutex).
2564 * %true if @flush_color >= 0 and there's something to flush. %false
2567 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2568 int flush_color, int work_color)
2573 if (flush_color >= 0) {
2574 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2575 atomic_set(&wq->nr_cwqs_to_flush, 1);
2578 for_each_cwq_cpu(cpu, wq) {
2579 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2580 struct global_cwq *gcwq = cwq->pool->gcwq;
2582 spin_lock_irq(&gcwq->lock);
2584 if (flush_color >= 0) {
2585 BUG_ON(cwq->flush_color != -1);
2587 if (cwq->nr_in_flight[flush_color]) {
2588 cwq->flush_color = flush_color;
2589 atomic_inc(&wq->nr_cwqs_to_flush);
2594 if (work_color >= 0) {
2595 BUG_ON(work_color != work_next_color(cwq->work_color));
2596 cwq->work_color = work_color;
2599 spin_unlock_irq(&gcwq->lock);
2602 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2603 complete(&wq->first_flusher->done);
2609 * flush_workqueue - ensure that any scheduled work has run to completion.
2610 * @wq: workqueue to flush
2612 * Forces execution of the workqueue and blocks until its completion.
2613 * This is typically used in driver shutdown handlers.
2615 * We sleep until all works which were queued on entry have been handled,
2616 * but we are not livelocked by new incoming ones.
2618 void flush_workqueue(struct workqueue_struct *wq)
2620 struct wq_flusher this_flusher = {
2621 .list = LIST_HEAD_INIT(this_flusher.list),
2623 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2627 lock_map_acquire(&wq->lockdep_map);
2628 lock_map_release(&wq->lockdep_map);
2630 mutex_lock(&wq->flush_mutex);
2633 * Start-to-wait phase
2635 next_color = work_next_color(wq->work_color);
2637 if (next_color != wq->flush_color) {
2639 * Color space is not full. The current work_color
2640 * becomes our flush_color and work_color is advanced
2643 BUG_ON(!list_empty(&wq->flusher_overflow));
2644 this_flusher.flush_color = wq->work_color;
2645 wq->work_color = next_color;
2647 if (!wq->first_flusher) {
2648 /* no flush in progress, become the first flusher */
2649 BUG_ON(wq->flush_color != this_flusher.flush_color);
2651 wq->first_flusher = &this_flusher;
2653 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2655 /* nothing to flush, done */
2656 wq->flush_color = next_color;
2657 wq->first_flusher = NULL;
2662 BUG_ON(wq->flush_color == this_flusher.flush_color);
2663 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2664 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2668 * Oops, color space is full, wait on overflow queue.
2669 * The next flush completion will assign us
2670 * flush_color and transfer to flusher_queue.
2672 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2675 mutex_unlock(&wq->flush_mutex);
2677 wait_for_completion(&this_flusher.done);
2680 * Wake-up-and-cascade phase
2682 * First flushers are responsible for cascading flushes and
2683 * handling overflow. Non-first flushers can simply return.
2685 if (wq->first_flusher != &this_flusher)
2688 mutex_lock(&wq->flush_mutex);
2690 /* we might have raced, check again with mutex held */
2691 if (wq->first_flusher != &this_flusher)
2694 wq->first_flusher = NULL;
2696 BUG_ON(!list_empty(&this_flusher.list));
2697 BUG_ON(wq->flush_color != this_flusher.flush_color);
2700 struct wq_flusher *next, *tmp;
2702 /* complete all the flushers sharing the current flush color */
2703 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2704 if (next->flush_color != wq->flush_color)
2706 list_del_init(&next->list);
2707 complete(&next->done);
2710 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2711 wq->flush_color != work_next_color(wq->work_color));
2713 /* this flush_color is finished, advance by one */
2714 wq->flush_color = work_next_color(wq->flush_color);
2716 /* one color has been freed, handle overflow queue */
2717 if (!list_empty(&wq->flusher_overflow)) {
2719 * Assign the same color to all overflowed
2720 * flushers, advance work_color and append to
2721 * flusher_queue. This is the start-to-wait
2722 * phase for these overflowed flushers.
2724 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2725 tmp->flush_color = wq->work_color;
2727 wq->work_color = work_next_color(wq->work_color);
2729 list_splice_tail_init(&wq->flusher_overflow,
2730 &wq->flusher_queue);
2731 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2734 if (list_empty(&wq->flusher_queue)) {
2735 BUG_ON(wq->flush_color != wq->work_color);
2740 * Need to flush more colors. Make the next flusher
2741 * the new first flusher and arm cwqs.
2743 BUG_ON(wq->flush_color == wq->work_color);
2744 BUG_ON(wq->flush_color != next->flush_color);
2746 list_del_init(&next->list);
2747 wq->first_flusher = next;
2749 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2753 * Meh... this color is already done, clear first
2754 * flusher and repeat cascading.
2756 wq->first_flusher = NULL;
2760 mutex_unlock(&wq->flush_mutex);
2762 EXPORT_SYMBOL_GPL(flush_workqueue);
2765 * drain_workqueue - drain a workqueue
2766 * @wq: workqueue to drain
2768 * Wait until the workqueue becomes empty. While draining is in progress,
2769 * only chain queueing is allowed. IOW, only currently pending or running
2770 * work items on @wq can queue further work items on it. @wq is flushed
2771 * repeatedly until it becomes empty. The number of flushing is detemined
2772 * by the depth of chaining and should be relatively short. Whine if it
2775 void drain_workqueue(struct workqueue_struct *wq)
2777 unsigned int flush_cnt = 0;
2781 * __queue_work() needs to test whether there are drainers, is much
2782 * hotter than drain_workqueue() and already looks at @wq->flags.
2783 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2785 spin_lock(&workqueue_lock);
2786 if (!wq->nr_drainers++)
2787 wq->flags |= WQ_DRAINING;
2788 spin_unlock(&workqueue_lock);
2790 flush_workqueue(wq);
2792 for_each_cwq_cpu(cpu, wq) {
2793 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2796 spin_lock_irq(&cwq->pool->gcwq->lock);
2797 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2798 spin_unlock_irq(&cwq->pool->gcwq->lock);
2803 if (++flush_cnt == 10 ||
2804 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2805 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2806 wq->name, flush_cnt);
2810 spin_lock(&workqueue_lock);
2811 if (!--wq->nr_drainers)
2812 wq->flags &= ~WQ_DRAINING;
2813 spin_unlock(&workqueue_lock);
2815 EXPORT_SYMBOL_GPL(drain_workqueue);
2817 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2819 struct worker *worker = NULL;
2820 struct global_cwq *gcwq;
2821 struct cpu_workqueue_struct *cwq;
2824 gcwq = get_work_gcwq(work);
2828 spin_lock_irq(&gcwq->lock);
2829 if (!list_empty(&work->entry)) {
2831 * See the comment near try_to_grab_pending()->smp_rmb().
2832 * If it was re-queued to a different gcwq under us, we
2833 * are not going to wait.
2836 cwq = get_work_cwq(work);
2837 if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
2840 worker = find_worker_executing_work(gcwq, work);
2843 cwq = worker->current_cwq;
2846 insert_wq_barrier(cwq, barr, work, worker);
2847 spin_unlock_irq(&gcwq->lock);
2850 * If @max_active is 1 or rescuer is in use, flushing another work
2851 * item on the same workqueue may lead to deadlock. Make sure the
2852 * flusher is not running on the same workqueue by verifying write
2855 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2856 lock_map_acquire(&cwq->wq->lockdep_map);
2858 lock_map_acquire_read(&cwq->wq->lockdep_map);
2859 lock_map_release(&cwq->wq->lockdep_map);
2863 spin_unlock_irq(&gcwq->lock);
2868 * flush_work - wait for a work to finish executing the last queueing instance
2869 * @work: the work to flush
2871 * Wait until @work has finished execution. @work is guaranteed to be idle
2872 * on return if it hasn't been requeued since flush started.
2875 * %true if flush_work() waited for the work to finish execution,
2876 * %false if it was already idle.
2878 bool flush_work(struct work_struct *work)
2880 struct wq_barrier barr;
2882 lock_map_acquire(&work->lockdep_map);
2883 lock_map_release(&work->lockdep_map);
2885 if (start_flush_work(work, &barr)) {
2886 wait_for_completion(&barr.done);
2887 destroy_work_on_stack(&barr.work);
2893 EXPORT_SYMBOL_GPL(flush_work);
2895 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2897 unsigned long flags;
2901 ret = try_to_grab_pending(work, is_dwork, &flags);
2903 * If someone else is canceling, wait for the same event it
2904 * would be waiting for before retrying.
2906 if (unlikely(ret == -ENOENT))
2908 } while (unlikely(ret < 0));
2910 /* tell other tasks trying to grab @work to back off */
2911 mark_work_canceling(work);
2912 local_irq_restore(flags);
2915 clear_work_data(work);
2920 * cancel_work_sync - cancel a work and wait for it to finish
2921 * @work: the work to cancel
2923 * Cancel @work and wait for its execution to finish. This function
2924 * can be used even if the work re-queues itself or migrates to
2925 * another workqueue. On return from this function, @work is
2926 * guaranteed to be not pending or executing on any CPU.
2928 * cancel_work_sync(&delayed_work->work) must not be used for
2929 * delayed_work's. Use cancel_delayed_work_sync() instead.
2931 * The caller must ensure that the workqueue on which @work was last
2932 * queued can't be destroyed before this function returns.
2935 * %true if @work was pending, %false otherwise.
2937 bool cancel_work_sync(struct work_struct *work)
2939 return __cancel_work_timer(work, false);
2941 EXPORT_SYMBOL_GPL(cancel_work_sync);
2944 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2945 * @dwork: the delayed work to flush
2947 * Delayed timer is cancelled and the pending work is queued for
2948 * immediate execution. Like flush_work(), this function only
2949 * considers the last queueing instance of @dwork.
2952 * %true if flush_work() waited for the work to finish execution,
2953 * %false if it was already idle.
2955 bool flush_delayed_work(struct delayed_work *dwork)
2957 local_irq_disable();
2958 if (del_timer_sync(&dwork->timer))
2959 __queue_work(dwork->cpu,
2960 get_work_cwq(&dwork->work)->wq, &dwork->work);
2962 return flush_work(&dwork->work);
2964 EXPORT_SYMBOL(flush_delayed_work);
2967 * cancel_delayed_work - cancel a delayed work
2968 * @dwork: delayed_work to cancel
2970 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2971 * and canceled; %false if wasn't pending. Note that the work callback
2972 * function may still be running on return, unless it returns %true and the
2973 * work doesn't re-arm itself. Explicitly flush or use
2974 * cancel_delayed_work_sync() to wait on it.
2976 * This function is safe to call from any context including IRQ handler.
2978 bool cancel_delayed_work(struct delayed_work *dwork)
2980 unsigned long flags;
2984 ret = try_to_grab_pending(&dwork->work, true, &flags);
2985 } while (unlikely(ret == -EAGAIN));
2987 if (unlikely(ret < 0))
2990 set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work));
2991 local_irq_restore(flags);
2994 EXPORT_SYMBOL(cancel_delayed_work);
2997 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2998 * @dwork: the delayed work cancel
3000 * This is cancel_work_sync() for delayed works.
3003 * %true if @dwork was pending, %false otherwise.
3005 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3007 return __cancel_work_timer(&dwork->work, true);
3009 EXPORT_SYMBOL(cancel_delayed_work_sync);
3012 * schedule_work_on - put work task on a specific cpu
3013 * @cpu: cpu to put the work task on
3014 * @work: job to be done
3016 * This puts a job on a specific cpu
3018 bool schedule_work_on(int cpu, struct work_struct *work)
3020 return queue_work_on(cpu, system_wq, work);
3022 EXPORT_SYMBOL(schedule_work_on);
3025 * schedule_work - put work task in global workqueue
3026 * @work: job to be done
3028 * Returns %false if @work was already on the kernel-global workqueue and
3031 * This puts a job in the kernel-global workqueue if it was not already
3032 * queued and leaves it in the same position on the kernel-global
3033 * workqueue otherwise.
3035 bool schedule_work(struct work_struct *work)
3037 return queue_work(system_wq, work);
3039 EXPORT_SYMBOL(schedule_work);
3042 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3044 * @dwork: job to be done
3045 * @delay: number of jiffies to wait
3047 * After waiting for a given time this puts a job in the kernel-global
3048 * workqueue on the specified CPU.
3050 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3051 unsigned long delay)
3053 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3055 EXPORT_SYMBOL(schedule_delayed_work_on);
3058 * schedule_delayed_work - put work task in global workqueue after delay
3059 * @dwork: job to be done
3060 * @delay: number of jiffies to wait or 0 for immediate execution
3062 * After waiting for a given time this puts a job in the kernel-global
3065 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3067 return queue_delayed_work(system_wq, dwork, delay);
3069 EXPORT_SYMBOL(schedule_delayed_work);
3072 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3073 * @func: the function to call
3075 * schedule_on_each_cpu() executes @func on each online CPU using the
3076 * system workqueue and blocks until all CPUs have completed.
3077 * schedule_on_each_cpu() is very slow.
3080 * 0 on success, -errno on failure.
3082 int schedule_on_each_cpu(work_func_t func)
3085 struct work_struct __percpu *works;
3087 works = alloc_percpu(struct work_struct);
3093 for_each_online_cpu(cpu) {
3094 struct work_struct *work = per_cpu_ptr(works, cpu);
3096 INIT_WORK(work, func);
3097 schedule_work_on(cpu, work);
3100 for_each_online_cpu(cpu)
3101 flush_work(per_cpu_ptr(works, cpu));
3109 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3111 * Forces execution of the kernel-global workqueue and blocks until its
3114 * Think twice before calling this function! It's very easy to get into
3115 * trouble if you don't take great care. Either of the following situations
3116 * will lead to deadlock:
3118 * One of the work items currently on the workqueue needs to acquire
3119 * a lock held by your code or its caller.
3121 * Your code is running in the context of a work routine.
3123 * They will be detected by lockdep when they occur, but the first might not
3124 * occur very often. It depends on what work items are on the workqueue and
3125 * what locks they need, which you have no control over.
3127 * In most situations flushing the entire workqueue is overkill; you merely
3128 * need to know that a particular work item isn't queued and isn't running.
3129 * In such cases you should use cancel_delayed_work_sync() or
3130 * cancel_work_sync() instead.
3132 void flush_scheduled_work(void)
3134 flush_workqueue(system_wq);
3136 EXPORT_SYMBOL(flush_scheduled_work);
3139 * execute_in_process_context - reliably execute the routine with user context
3140 * @fn: the function to execute
3141 * @ew: guaranteed storage for the execute work structure (must
3142 * be available when the work executes)
3144 * Executes the function immediately if process context is available,
3145 * otherwise schedules the function for delayed execution.
3147 * Returns: 0 - function was executed
3148 * 1 - function was scheduled for execution
3150 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3152 if (!in_interrupt()) {
3157 INIT_WORK(&ew->work, fn);
3158 schedule_work(&ew->work);
3162 EXPORT_SYMBOL_GPL(execute_in_process_context);
3164 int keventd_up(void)
3166 return system_wq != NULL;
3169 static int alloc_cwqs(struct workqueue_struct *wq)
3172 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3173 * Make sure that the alignment isn't lower than that of
3174 * unsigned long long.
3176 const size_t size = sizeof(struct cpu_workqueue_struct);
3177 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3178 __alignof__(unsigned long long));
3180 if (!(wq->flags & WQ_UNBOUND))
3181 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3186 * Allocate enough room to align cwq and put an extra
3187 * pointer at the end pointing back to the originally
3188 * allocated pointer which will be used for free.
3190 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3192 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3193 *(void **)(wq->cpu_wq.single + 1) = ptr;
3197 /* just in case, make sure it's actually aligned */
3198 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3199 return wq->cpu_wq.v ? 0 : -ENOMEM;
3202 static void free_cwqs(struct workqueue_struct *wq)
3204 if (!(wq->flags & WQ_UNBOUND))
3205 free_percpu(wq->cpu_wq.pcpu);
3206 else if (wq->cpu_wq.single) {
3207 /* the pointer to free is stored right after the cwq */
3208 kfree(*(void **)(wq->cpu_wq.single + 1));
3212 static int wq_clamp_max_active(int max_active, unsigned int flags,
3215 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3217 if (max_active < 1 || max_active > lim)
3218 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3219 max_active, name, 1, lim);
3221 return clamp_val(max_active, 1, lim);
3224 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3227 struct lock_class_key *key,
3228 const char *lock_name, ...)
3230 va_list args, args1;
3231 struct workqueue_struct *wq;
3235 /* determine namelen, allocate wq and format name */
3236 va_start(args, lock_name);
3237 va_copy(args1, args);
3238 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3240 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3244 vsnprintf(wq->name, namelen, fmt, args1);
3249 * Workqueues which may be used during memory reclaim should
3250 * have a rescuer to guarantee forward progress.
3252 if (flags & WQ_MEM_RECLAIM)
3253 flags |= WQ_RESCUER;
3255 max_active = max_active ?: WQ_DFL_ACTIVE;
3256 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3260 wq->saved_max_active = max_active;
3261 mutex_init(&wq->flush_mutex);
3262 atomic_set(&wq->nr_cwqs_to_flush, 0);
3263 INIT_LIST_HEAD(&wq->flusher_queue);
3264 INIT_LIST_HEAD(&wq->flusher_overflow);
3266 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3267 INIT_LIST_HEAD(&wq->list);
3269 if (alloc_cwqs(wq) < 0)
3272 for_each_cwq_cpu(cpu, wq) {
3273 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3274 struct global_cwq *gcwq = get_gcwq(cpu);
3275 int pool_idx = (bool)(flags & WQ_HIGHPRI);
3277 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3278 cwq->pool = &gcwq->pools[pool_idx];
3280 cwq->flush_color = -1;
3281 cwq->max_active = max_active;
3282 INIT_LIST_HEAD(&cwq->delayed_works);
3285 if (flags & WQ_RESCUER) {
3286 struct worker *rescuer;
3288 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3291 wq->rescuer = rescuer = alloc_worker();
3295 rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3297 if (IS_ERR(rescuer->task))
3300 rescuer->task->flags |= PF_THREAD_BOUND;
3301 wake_up_process(rescuer->task);
3305 * workqueue_lock protects global freeze state and workqueues
3306 * list. Grab it, set max_active accordingly and add the new
3307 * workqueue to workqueues list.
3309 spin_lock(&workqueue_lock);
3311 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3312 for_each_cwq_cpu(cpu, wq)
3313 get_cwq(cpu, wq)->max_active = 0;
3315 list_add(&wq->list, &workqueues);
3317 spin_unlock(&workqueue_lock);
3323 free_mayday_mask(wq->mayday_mask);
3329 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3332 * destroy_workqueue - safely terminate a workqueue
3333 * @wq: target workqueue
3335 * Safely destroy a workqueue. All work currently pending will be done first.
3337 void destroy_workqueue(struct workqueue_struct *wq)
3341 /* drain it before proceeding with destruction */
3342 drain_workqueue(wq);
3345 * wq list is used to freeze wq, remove from list after
3346 * flushing is complete in case freeze races us.
3348 spin_lock(&workqueue_lock);
3349 list_del(&wq->list);
3350 spin_unlock(&workqueue_lock);
3353 for_each_cwq_cpu(cpu, wq) {
3354 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3357 for (i = 0; i < WORK_NR_COLORS; i++)
3358 BUG_ON(cwq->nr_in_flight[i]);
3359 BUG_ON(cwq->nr_active);
3360 BUG_ON(!list_empty(&cwq->delayed_works));
3363 if (wq->flags & WQ_RESCUER) {
3364 kthread_stop(wq->rescuer->task);
3365 free_mayday_mask(wq->mayday_mask);
3372 EXPORT_SYMBOL_GPL(destroy_workqueue);
3375 * workqueue_set_max_active - adjust max_active of a workqueue
3376 * @wq: target workqueue
3377 * @max_active: new max_active value.
3379 * Set max_active of @wq to @max_active.
3382 * Don't call from IRQ context.
3384 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3388 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3390 spin_lock(&workqueue_lock);
3392 wq->saved_max_active = max_active;
3394 for_each_cwq_cpu(cpu, wq) {
3395 struct global_cwq *gcwq = get_gcwq(cpu);
3397 spin_lock_irq(&gcwq->lock);
3399 if (!(wq->flags & WQ_FREEZABLE) ||
3400 !(gcwq->flags & GCWQ_FREEZING))
3401 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3403 spin_unlock_irq(&gcwq->lock);
3406 spin_unlock(&workqueue_lock);
3408 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3411 * workqueue_congested - test whether a workqueue is congested
3412 * @cpu: CPU in question
3413 * @wq: target workqueue
3415 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3416 * no synchronization around this function and the test result is
3417 * unreliable and only useful as advisory hints or for debugging.
3420 * %true if congested, %false otherwise.
3422 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3424 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3426 return !list_empty(&cwq->delayed_works);
3428 EXPORT_SYMBOL_GPL(workqueue_congested);
3431 * work_cpu - return the last known associated cpu for @work
3432 * @work: the work of interest
3435 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3437 unsigned int work_cpu(struct work_struct *work)
3439 struct global_cwq *gcwq = get_work_gcwq(work);
3441 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3443 EXPORT_SYMBOL_GPL(work_cpu);
3446 * work_busy - test whether a work is currently pending or running
3447 * @work: the work to be tested
3449 * Test whether @work is currently pending or running. There is no
3450 * synchronization around this function and the test result is
3451 * unreliable and only useful as advisory hints or for debugging.
3452 * Especially for reentrant wqs, the pending state might hide the
3456 * OR'd bitmask of WORK_BUSY_* bits.
3458 unsigned int work_busy(struct work_struct *work)
3460 struct global_cwq *gcwq = get_work_gcwq(work);
3461 unsigned long flags;
3462 unsigned int ret = 0;
3467 spin_lock_irqsave(&gcwq->lock, flags);
3469 if (work_pending(work))
3470 ret |= WORK_BUSY_PENDING;
3471 if (find_worker_executing_work(gcwq, work))
3472 ret |= WORK_BUSY_RUNNING;
3474 spin_unlock_irqrestore(&gcwq->lock, flags);
3478 EXPORT_SYMBOL_GPL(work_busy);
3483 * There are two challenges in supporting CPU hotplug. Firstly, there
3484 * are a lot of assumptions on strong associations among work, cwq and
3485 * gcwq which make migrating pending and scheduled works very
3486 * difficult to implement without impacting hot paths. Secondly,
3487 * gcwqs serve mix of short, long and very long running works making
3488 * blocked draining impractical.
3490 * This is solved by allowing a gcwq to be disassociated from the CPU
3491 * running as an unbound one and allowing it to be reattached later if the
3492 * cpu comes back online.
3495 /* claim manager positions of all pools */
3496 static void gcwq_claim_assoc_and_lock(struct global_cwq *gcwq)
3498 struct worker_pool *pool;
3500 for_each_worker_pool(pool, gcwq)
3501 mutex_lock_nested(&pool->assoc_mutex, pool - gcwq->pools);
3502 spin_lock_irq(&gcwq->lock);
3505 /* release manager positions */
3506 static void gcwq_release_assoc_and_unlock(struct global_cwq *gcwq)
3508 struct worker_pool *pool;
3510 spin_unlock_irq(&gcwq->lock);
3511 for_each_worker_pool(pool, gcwq)
3512 mutex_unlock(&pool->assoc_mutex);
3515 static void gcwq_unbind_fn(struct work_struct *work)
3517 struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3518 struct worker_pool *pool;
3519 struct worker *worker;
3520 struct hlist_node *pos;
3523 BUG_ON(gcwq->cpu != smp_processor_id());
3525 gcwq_claim_assoc_and_lock(gcwq);
3528 * We've claimed all manager positions. Make all workers unbound
3529 * and set DISASSOCIATED. Before this, all workers except for the
3530 * ones which are still executing works from before the last CPU
3531 * down must be on the cpu. After this, they may become diasporas.
3533 for_each_worker_pool(pool, gcwq)
3534 list_for_each_entry(worker, &pool->idle_list, entry)
3535 worker->flags |= WORKER_UNBOUND;
3537 for_each_busy_worker(worker, i, pos, gcwq)
3538 worker->flags |= WORKER_UNBOUND;
3540 gcwq->flags |= GCWQ_DISASSOCIATED;
3542 gcwq_release_assoc_and_unlock(gcwq);
3545 * Call schedule() so that we cross rq->lock and thus can guarantee
3546 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3547 * as scheduler callbacks may be invoked from other cpus.
3552 * Sched callbacks are disabled now. Zap nr_running. After this,
3553 * nr_running stays zero and need_more_worker() and keep_working()
3554 * are always true as long as the worklist is not empty. @gcwq now
3555 * behaves as unbound (in terms of concurrency management) gcwq
3556 * which is served by workers tied to the CPU.
3558 * On return from this function, the current worker would trigger
3559 * unbound chain execution of pending work items if other workers
3562 for_each_worker_pool(pool, gcwq)
3563 atomic_set(get_pool_nr_running(pool), 0);
3567 * Workqueues should be brought up before normal priority CPU notifiers.
3568 * This will be registered high priority CPU notifier.
3570 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3571 unsigned long action,
3574 unsigned int cpu = (unsigned long)hcpu;
3575 struct global_cwq *gcwq = get_gcwq(cpu);
3576 struct worker_pool *pool;
3578 switch (action & ~CPU_TASKS_FROZEN) {
3579 case CPU_UP_PREPARE:
3580 for_each_worker_pool(pool, gcwq) {
3581 struct worker *worker;
3583 if (pool->nr_workers)
3586 worker = create_worker(pool);
3590 spin_lock_irq(&gcwq->lock);
3591 start_worker(worker);
3592 spin_unlock_irq(&gcwq->lock);
3596 case CPU_DOWN_FAILED:
3598 gcwq_claim_assoc_and_lock(gcwq);
3599 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3600 rebind_workers(gcwq);
3601 gcwq_release_assoc_and_unlock(gcwq);
3608 * Workqueues should be brought down after normal priority CPU notifiers.
3609 * This will be registered as low priority CPU notifier.
3611 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3612 unsigned long action,
3615 unsigned int cpu = (unsigned long)hcpu;
3616 struct work_struct unbind_work;
3618 switch (action & ~CPU_TASKS_FROZEN) {
3619 case CPU_DOWN_PREPARE:
3620 /* unbinding should happen on the local CPU */
3621 INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3622 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3623 flush_work(&unbind_work);
3631 struct work_for_cpu {
3632 struct completion completion;
3638 static int do_work_for_cpu(void *_wfc)
3640 struct work_for_cpu *wfc = _wfc;
3641 wfc->ret = wfc->fn(wfc->arg);
3642 complete(&wfc->completion);
3647 * work_on_cpu - run a function in user context on a particular cpu
3648 * @cpu: the cpu to run on
3649 * @fn: the function to run
3650 * @arg: the function arg
3652 * This will return the value @fn returns.
3653 * It is up to the caller to ensure that the cpu doesn't go offline.
3654 * The caller must not hold any locks which would prevent @fn from completing.
3656 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3658 struct task_struct *sub_thread;
3659 struct work_for_cpu wfc = {
3660 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3665 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3666 if (IS_ERR(sub_thread))
3667 return PTR_ERR(sub_thread);
3668 kthread_bind(sub_thread, cpu);
3669 wake_up_process(sub_thread);
3670 wait_for_completion(&wfc.completion);
3673 EXPORT_SYMBOL_GPL(work_on_cpu);
3674 #endif /* CONFIG_SMP */
3676 #ifdef CONFIG_FREEZER
3679 * freeze_workqueues_begin - begin freezing workqueues
3681 * Start freezing workqueues. After this function returns, all freezable
3682 * workqueues will queue new works to their frozen_works list instead of
3686 * Grabs and releases workqueue_lock and gcwq->lock's.
3688 void freeze_workqueues_begin(void)
3692 spin_lock(&workqueue_lock);
3694 BUG_ON(workqueue_freezing);
3695 workqueue_freezing = true;
3697 for_each_gcwq_cpu(cpu) {
3698 struct global_cwq *gcwq = get_gcwq(cpu);
3699 struct workqueue_struct *wq;
3701 spin_lock_irq(&gcwq->lock);
3703 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3704 gcwq->flags |= GCWQ_FREEZING;
3706 list_for_each_entry(wq, &workqueues, list) {
3707 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3709 if (cwq && wq->flags & WQ_FREEZABLE)
3710 cwq->max_active = 0;
3713 spin_unlock_irq(&gcwq->lock);
3716 spin_unlock(&workqueue_lock);
3720 * freeze_workqueues_busy - are freezable workqueues still busy?
3722 * Check whether freezing is complete. This function must be called
3723 * between freeze_workqueues_begin() and thaw_workqueues().
3726 * Grabs and releases workqueue_lock.
3729 * %true if some freezable workqueues are still busy. %false if freezing
3732 bool freeze_workqueues_busy(void)
3737 spin_lock(&workqueue_lock);
3739 BUG_ON(!workqueue_freezing);
3741 for_each_gcwq_cpu(cpu) {
3742 struct workqueue_struct *wq;
3744 * nr_active is monotonically decreasing. It's safe
3745 * to peek without lock.
3747 list_for_each_entry(wq, &workqueues, list) {
3748 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3750 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3753 BUG_ON(cwq->nr_active < 0);
3754 if (cwq->nr_active) {
3761 spin_unlock(&workqueue_lock);
3766 * thaw_workqueues - thaw workqueues
3768 * Thaw workqueues. Normal queueing is restored and all collected
3769 * frozen works are transferred to their respective gcwq worklists.
3772 * Grabs and releases workqueue_lock and gcwq->lock's.
3774 void thaw_workqueues(void)
3778 spin_lock(&workqueue_lock);
3780 if (!workqueue_freezing)
3783 for_each_gcwq_cpu(cpu) {
3784 struct global_cwq *gcwq = get_gcwq(cpu);
3785 struct worker_pool *pool;
3786 struct workqueue_struct *wq;
3788 spin_lock_irq(&gcwq->lock);
3790 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3791 gcwq->flags &= ~GCWQ_FREEZING;
3793 list_for_each_entry(wq, &workqueues, list) {
3794 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3796 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3799 /* restore max_active and repopulate worklist */
3800 cwq->max_active = wq->saved_max_active;
3802 while (!list_empty(&cwq->delayed_works) &&
3803 cwq->nr_active < cwq->max_active)
3804 cwq_activate_first_delayed(cwq);
3807 for_each_worker_pool(pool, gcwq)
3808 wake_up_worker(pool);
3810 spin_unlock_irq(&gcwq->lock);
3813 workqueue_freezing = false;
3815 spin_unlock(&workqueue_lock);
3817 #endif /* CONFIG_FREEZER */
3819 static int __init init_workqueues(void)
3824 /* make sure we have enough bits for OFFQ CPU number */
3825 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
3828 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3829 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3831 /* initialize gcwqs */
3832 for_each_gcwq_cpu(cpu) {
3833 struct global_cwq *gcwq = get_gcwq(cpu);
3834 struct worker_pool *pool;
3836 spin_lock_init(&gcwq->lock);
3838 gcwq->flags |= GCWQ_DISASSOCIATED;
3840 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3841 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3843 for_each_worker_pool(pool, gcwq) {
3845 INIT_LIST_HEAD(&pool->worklist);
3846 INIT_LIST_HEAD(&pool->idle_list);
3848 init_timer_deferrable(&pool->idle_timer);
3849 pool->idle_timer.function = idle_worker_timeout;
3850 pool->idle_timer.data = (unsigned long)pool;
3852 setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3853 (unsigned long)pool);
3855 mutex_init(&pool->assoc_mutex);
3856 ida_init(&pool->worker_ida);
3860 /* create the initial worker */
3861 for_each_online_gcwq_cpu(cpu) {
3862 struct global_cwq *gcwq = get_gcwq(cpu);
3863 struct worker_pool *pool;
3865 if (cpu != WORK_CPU_UNBOUND)
3866 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3868 for_each_worker_pool(pool, gcwq) {
3869 struct worker *worker;
3871 worker = create_worker(pool);
3873 spin_lock_irq(&gcwq->lock);
3874 start_worker(worker);
3875 spin_unlock_irq(&gcwq->lock);
3879 system_wq = alloc_workqueue("events", 0, 0);
3880 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3881 system_long_wq = alloc_workqueue("events_long", 0, 0);
3882 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3883 WQ_UNBOUND_MAX_ACTIVE);
3884 system_freezable_wq = alloc_workqueue("events_freezable",
3886 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3887 !system_unbound_wq || !system_freezable_wq);
3890 early_initcall(init_workqueues);