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 * managership 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_REBIND = 1 << 5, /* mom is home, come back */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
80 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_REBIND | WORKER_UNBOUND |
83 NR_WORKER_POOLS = 2, /* # worker pools per gcwq */
85 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
86 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
87 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
89 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give -20.
102 RESCUER_NICE_LEVEL = -20,
103 HIGHPRI_NICE_LEVEL = -20,
107 * Structure fields follow one of the following exclusion rules.
109 * I: Modifiable by initialization/destruction paths and read-only for
112 * P: Preemption protected. Disabling preemption is enough and should
113 * only be modified and accessed from the local cpu.
115 * L: gcwq->lock protected. Access with gcwq->lock held.
117 * X: During normal operation, modification requires gcwq->lock and
118 * should be done only from local cpu. Either disabling preemption
119 * on local cpu or grabbing gcwq->lock is enough for read access.
120 * If GCWQ_DISASSOCIATED is set, it's identical to L.
122 * F: wq->flush_mutex protected.
124 * W: workqueue_lock protected.
131 * The poor guys doing the actual heavy lifting. All on-duty workers
132 * are either serving the manager role, on idle list or on busy hash.
135 /* on idle list while idle, on busy hash table while busy */
137 struct list_head entry; /* L: while idle */
138 struct hlist_node hentry; /* L: while busy */
141 struct work_struct *current_work; /* L: work being processed */
142 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
143 struct list_head scheduled; /* L: scheduled works */
144 struct task_struct *task; /* I: worker task */
145 struct worker_pool *pool; /* I: the associated pool */
146 /* 64 bytes boundary on 64bit, 32 on 32bit */
147 unsigned long last_active; /* L: last active timestamp */
148 unsigned int flags; /* X: flags */
149 int id; /* I: worker id */
151 /* for rebinding worker to CPU */
152 struct work_struct rebind_work; /* L: for busy worker */
156 struct global_cwq *gcwq; /* I: the owning gcwq */
157 unsigned int flags; /* X: flags */
159 struct list_head worklist; /* L: list of pending works */
160 int nr_workers; /* L: total number of workers */
162 /* nr_idle includes the ones off idle_list for rebinding */
163 int nr_idle; /* L: currently idle ones */
165 struct list_head idle_list; /* X: list of idle workers */
166 struct timer_list idle_timer; /* L: worker idle timeout */
167 struct timer_list mayday_timer; /* L: SOS timer for workers */
169 struct mutex manager_mutex; /* mutex manager should hold */
170 struct ida worker_ida; /* L: for worker IDs */
174 * Global per-cpu workqueue. There's one and only one for each cpu
175 * and all works are queued and processed here regardless of their
179 spinlock_t lock; /* the gcwq lock */
180 unsigned int cpu; /* I: the associated cpu */
181 unsigned int flags; /* L: GCWQ_* flags */
183 /* workers are chained either in busy_hash or pool idle_list */
184 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
185 /* L: hash of busy workers */
187 struct worker_pool pools[NR_WORKER_POOLS];
188 /* normal and highpri pools */
189 } ____cacheline_aligned_in_smp;
192 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
193 * work_struct->data are used for flags and thus cwqs need to be
194 * aligned at two's power of the number of flag bits.
196 struct cpu_workqueue_struct {
197 struct worker_pool *pool; /* I: the associated pool */
198 struct workqueue_struct *wq; /* I: the owning workqueue */
199 int work_color; /* L: current color */
200 int flush_color; /* L: flushing color */
201 int nr_in_flight[WORK_NR_COLORS];
202 /* L: nr of in_flight works */
203 int nr_active; /* L: nr of active works */
204 int max_active; /* L: max active works */
205 struct list_head delayed_works; /* L: delayed works */
209 * Structure used to wait for workqueue flush.
212 struct list_head list; /* F: list of flushers */
213 int flush_color; /* F: flush color waiting for */
214 struct completion done; /* flush completion */
218 * All cpumasks are assumed to be always set on UP and thus can't be
219 * used to determine whether there's something to be done.
222 typedef cpumask_var_t mayday_mask_t;
223 #define mayday_test_and_set_cpu(cpu, mask) \
224 cpumask_test_and_set_cpu((cpu), (mask))
225 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
226 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
227 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
228 #define free_mayday_mask(mask) free_cpumask_var((mask))
230 typedef unsigned long mayday_mask_t;
231 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
232 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
233 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
234 #define alloc_mayday_mask(maskp, gfp) true
235 #define free_mayday_mask(mask) do { } while (0)
239 * The externally visible workqueue abstraction is an array of
240 * per-CPU workqueues:
242 struct workqueue_struct {
243 unsigned int flags; /* W: WQ_* flags */
245 struct cpu_workqueue_struct __percpu *pcpu;
246 struct cpu_workqueue_struct *single;
248 } cpu_wq; /* I: cwq's */
249 struct list_head list; /* W: list of all workqueues */
251 struct mutex flush_mutex; /* protects wq flushing */
252 int work_color; /* F: current work color */
253 int flush_color; /* F: current flush color */
254 atomic_t nr_cwqs_to_flush; /* flush in progress */
255 struct wq_flusher *first_flusher; /* F: first flusher */
256 struct list_head flusher_queue; /* F: flush waiters */
257 struct list_head flusher_overflow; /* F: flush overflow list */
259 mayday_mask_t mayday_mask; /* cpus requesting rescue */
260 struct worker *rescuer; /* I: rescue worker */
262 int nr_drainers; /* W: drain in progress */
263 int saved_max_active; /* W: saved cwq max_active */
264 #ifdef CONFIG_LOCKDEP
265 struct lockdep_map lockdep_map;
267 char name[]; /* I: workqueue name */
270 struct workqueue_struct *system_wq __read_mostly;
271 EXPORT_SYMBOL_GPL(system_wq);
272 struct workqueue_struct *system_highpri_wq __read_mostly;
273 EXPORT_SYMBOL_GPL(system_highpri_wq);
274 struct workqueue_struct *system_long_wq __read_mostly;
275 EXPORT_SYMBOL_GPL(system_long_wq);
276 struct workqueue_struct *system_unbound_wq __read_mostly;
277 EXPORT_SYMBOL_GPL(system_unbound_wq);
278 struct workqueue_struct *system_freezable_wq __read_mostly;
279 EXPORT_SYMBOL_GPL(system_freezable_wq);
281 #define CREATE_TRACE_POINTS
282 #include <trace/events/workqueue.h>
284 #define for_each_worker_pool(pool, gcwq) \
285 for ((pool) = &(gcwq)->pools[0]; \
286 (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
288 #define for_each_busy_worker(worker, i, pos, gcwq) \
289 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
290 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
292 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
295 if (cpu < nr_cpu_ids) {
297 cpu = cpumask_next(cpu, mask);
298 if (cpu < nr_cpu_ids)
302 return WORK_CPU_UNBOUND;
304 return WORK_CPU_NONE;
307 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
308 struct workqueue_struct *wq)
310 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
316 * An extra gcwq is defined for an invalid cpu number
317 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
318 * specific CPU. The following iterators are similar to
319 * for_each_*_cpu() iterators but also considers the unbound gcwq.
321 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
322 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
323 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
324 * WORK_CPU_UNBOUND for unbound workqueues
326 #define for_each_gcwq_cpu(cpu) \
327 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
328 (cpu) < WORK_CPU_NONE; \
329 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
331 #define for_each_online_gcwq_cpu(cpu) \
332 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
333 (cpu) < WORK_CPU_NONE; \
334 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
336 #define for_each_cwq_cpu(cpu, wq) \
337 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
338 (cpu) < WORK_CPU_NONE; \
339 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
341 #ifdef CONFIG_DEBUG_OBJECTS_WORK
343 static struct debug_obj_descr work_debug_descr;
345 static void *work_debug_hint(void *addr)
347 return ((struct work_struct *) addr)->func;
351 * fixup_init is called when:
352 * - an active object is initialized
354 static int work_fixup_init(void *addr, enum debug_obj_state state)
356 struct work_struct *work = addr;
359 case ODEBUG_STATE_ACTIVE:
360 cancel_work_sync(work);
361 debug_object_init(work, &work_debug_descr);
369 * fixup_activate is called when:
370 * - an active object is activated
371 * - an unknown object is activated (might be a statically initialized object)
373 static int work_fixup_activate(void *addr, enum debug_obj_state state)
375 struct work_struct *work = addr;
379 case ODEBUG_STATE_NOTAVAILABLE:
381 * This is not really a fixup. The work struct was
382 * statically initialized. We just make sure that it
383 * is tracked in the object tracker.
385 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
386 debug_object_init(work, &work_debug_descr);
387 debug_object_activate(work, &work_debug_descr);
393 case ODEBUG_STATE_ACTIVE:
402 * fixup_free is called when:
403 * - an active object is freed
405 static int work_fixup_free(void *addr, enum debug_obj_state state)
407 struct work_struct *work = addr;
410 case ODEBUG_STATE_ACTIVE:
411 cancel_work_sync(work);
412 debug_object_free(work, &work_debug_descr);
419 static struct debug_obj_descr work_debug_descr = {
420 .name = "work_struct",
421 .debug_hint = work_debug_hint,
422 .fixup_init = work_fixup_init,
423 .fixup_activate = work_fixup_activate,
424 .fixup_free = work_fixup_free,
427 static inline void debug_work_activate(struct work_struct *work)
429 debug_object_activate(work, &work_debug_descr);
432 static inline void debug_work_deactivate(struct work_struct *work)
434 debug_object_deactivate(work, &work_debug_descr);
437 void __init_work(struct work_struct *work, int onstack)
440 debug_object_init_on_stack(work, &work_debug_descr);
442 debug_object_init(work, &work_debug_descr);
444 EXPORT_SYMBOL_GPL(__init_work);
446 void destroy_work_on_stack(struct work_struct *work)
448 debug_object_free(work, &work_debug_descr);
450 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
453 static inline void debug_work_activate(struct work_struct *work) { }
454 static inline void debug_work_deactivate(struct work_struct *work) { }
457 /* Serializes the accesses to the list of workqueues. */
458 static DEFINE_SPINLOCK(workqueue_lock);
459 static LIST_HEAD(workqueues);
460 static bool workqueue_freezing; /* W: have wqs started freezing? */
463 * The almighty global cpu workqueues. nr_running is the only field
464 * which is expected to be used frequently by other cpus via
465 * try_to_wake_up(). Put it in a separate cacheline.
467 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
468 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]);
471 * Global cpu workqueue and nr_running counter for unbound gcwq. The
472 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
473 * workers have WORKER_UNBOUND set.
475 static struct global_cwq unbound_global_cwq;
476 static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = {
477 [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
480 static int worker_thread(void *__worker);
482 static int worker_pool_pri(struct worker_pool *pool)
484 return pool - pool->gcwq->pools;
487 static struct global_cwq *get_gcwq(unsigned int cpu)
489 if (cpu != WORK_CPU_UNBOUND)
490 return &per_cpu(global_cwq, cpu);
492 return &unbound_global_cwq;
495 static atomic_t *get_pool_nr_running(struct worker_pool *pool)
497 int cpu = pool->gcwq->cpu;
498 int idx = worker_pool_pri(pool);
500 if (cpu != WORK_CPU_UNBOUND)
501 return &per_cpu(pool_nr_running, cpu)[idx];
503 return &unbound_pool_nr_running[idx];
506 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
507 struct workqueue_struct *wq)
509 if (!(wq->flags & WQ_UNBOUND)) {
510 if (likely(cpu < nr_cpu_ids))
511 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
512 } else if (likely(cpu == WORK_CPU_UNBOUND))
513 return wq->cpu_wq.single;
517 static unsigned int work_color_to_flags(int color)
519 return color << WORK_STRUCT_COLOR_SHIFT;
522 static int get_work_color(struct work_struct *work)
524 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
525 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
528 static int work_next_color(int color)
530 return (color + 1) % WORK_NR_COLORS;
534 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
535 * contain the pointer to the queued cwq. Once execution starts, the flag
536 * is cleared and the high bits contain OFFQ flags and CPU number.
538 * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
539 * and clear_work_data() can be used to set the cwq, cpu or clear
540 * work->data. These functions should only be called while the work is
541 * owned - ie. while the PENDING bit is set.
543 * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
544 * a work. gcwq is available once the work has been queued anywhere after
545 * initialization until it is sync canceled. cwq is available only while
546 * the work item is queued.
548 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
549 * canceled. While being canceled, a work item may have its PENDING set
550 * but stay off timer and worklist for arbitrarily long and nobody should
551 * try to steal the PENDING bit.
553 static inline void set_work_data(struct work_struct *work, unsigned long data,
556 BUG_ON(!work_pending(work));
557 atomic_long_set(&work->data, data | flags | work_static(work));
560 static void set_work_cwq(struct work_struct *work,
561 struct cpu_workqueue_struct *cwq,
562 unsigned long extra_flags)
564 set_work_data(work, (unsigned long)cwq,
565 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
568 static void set_work_cpu_and_clear_pending(struct work_struct *work,
572 * The following wmb is paired with the implied mb in
573 * test_and_set_bit(PENDING) and ensures all updates to @work made
574 * here are visible to and precede any updates by the next PENDING
578 set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0);
581 static void clear_work_data(struct work_struct *work)
583 smp_wmb(); /* see set_work_cpu_and_clear_pending() */
584 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
587 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
589 unsigned long data = atomic_long_read(&work->data);
591 if (data & WORK_STRUCT_CWQ)
592 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
597 static struct global_cwq *get_work_gcwq(struct work_struct *work)
599 unsigned long data = atomic_long_read(&work->data);
602 if (data & WORK_STRUCT_CWQ)
603 return ((struct cpu_workqueue_struct *)
604 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq;
606 cpu = data >> WORK_OFFQ_CPU_SHIFT;
607 if (cpu == WORK_CPU_NONE)
610 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
611 return get_gcwq(cpu);
614 static void mark_work_canceling(struct work_struct *work)
616 struct global_cwq *gcwq = get_work_gcwq(work);
617 unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE;
619 set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) | WORK_OFFQ_CANCELING,
620 WORK_STRUCT_PENDING);
623 static bool work_is_canceling(struct work_struct *work)
625 unsigned long data = atomic_long_read(&work->data);
627 return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
631 * Policy functions. These define the policies on how the global worker
632 * pools are managed. Unless noted otherwise, these functions assume that
633 * they're being called with gcwq->lock held.
636 static bool __need_more_worker(struct worker_pool *pool)
638 return !atomic_read(get_pool_nr_running(pool));
642 * Need to wake up a worker? Called from anything but currently
645 * Note that, because unbound workers never contribute to nr_running, this
646 * function will always return %true for unbound gcwq as long as the
647 * worklist isn't empty.
649 static bool need_more_worker(struct worker_pool *pool)
651 return !list_empty(&pool->worklist) && __need_more_worker(pool);
654 /* Can I start working? Called from busy but !running workers. */
655 static bool may_start_working(struct worker_pool *pool)
657 return pool->nr_idle;
660 /* Do I need to keep working? Called from currently running workers. */
661 static bool keep_working(struct worker_pool *pool)
663 atomic_t *nr_running = get_pool_nr_running(pool);
665 return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
668 /* Do we need a new worker? Called from manager. */
669 static bool need_to_create_worker(struct worker_pool *pool)
671 return need_more_worker(pool) && !may_start_working(pool);
674 /* Do I need to be the manager? */
675 static bool need_to_manage_workers(struct worker_pool *pool)
677 return need_to_create_worker(pool) ||
678 (pool->flags & POOL_MANAGE_WORKERS);
681 /* Do we have too many workers and should some go away? */
682 static bool too_many_workers(struct worker_pool *pool)
684 bool managing = pool->flags & POOL_MANAGING_WORKERS;
685 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
686 int nr_busy = pool->nr_workers - nr_idle;
689 * nr_idle and idle_list may disagree if idle rebinding is in
690 * progress. Never return %true if idle_list is empty.
692 if (list_empty(&pool->idle_list))
695 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
702 /* Return the first worker. Safe with preemption disabled */
703 static struct worker *first_worker(struct worker_pool *pool)
705 if (unlikely(list_empty(&pool->idle_list)))
708 return list_first_entry(&pool->idle_list, struct worker, entry);
712 * wake_up_worker - wake up an idle worker
713 * @pool: worker pool to wake worker from
715 * Wake up the first idle worker of @pool.
718 * spin_lock_irq(gcwq->lock).
720 static void wake_up_worker(struct worker_pool *pool)
722 struct worker *worker = first_worker(pool);
725 wake_up_process(worker->task);
729 * wq_worker_waking_up - a worker is waking up
730 * @task: task waking up
731 * @cpu: CPU @task is waking up to
733 * This function is called during try_to_wake_up() when a worker is
737 * spin_lock_irq(rq->lock)
739 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
741 struct worker *worker = kthread_data(task);
743 if (!(worker->flags & WORKER_NOT_RUNNING))
744 atomic_inc(get_pool_nr_running(worker->pool));
748 * wq_worker_sleeping - a worker is going to sleep
749 * @task: task going to sleep
750 * @cpu: CPU in question, must be the current CPU number
752 * This function is called during schedule() when a busy worker is
753 * going to sleep. Worker on the same cpu can be woken up by
754 * returning pointer to its task.
757 * spin_lock_irq(rq->lock)
760 * Worker task on @cpu to wake up, %NULL if none.
762 struct task_struct *wq_worker_sleeping(struct task_struct *task,
765 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
766 struct worker_pool *pool = worker->pool;
767 atomic_t *nr_running = get_pool_nr_running(pool);
769 if (worker->flags & WORKER_NOT_RUNNING)
772 /* this can only happen on the local cpu */
773 BUG_ON(cpu != raw_smp_processor_id());
776 * The counterpart of the following dec_and_test, implied mb,
777 * worklist not empty test sequence is in insert_work().
778 * Please read comment there.
780 * NOT_RUNNING is clear. This means that we're bound to and
781 * running on the local cpu w/ rq lock held and preemption
782 * disabled, which in turn means that none else could be
783 * manipulating idle_list, so dereferencing idle_list without gcwq
786 if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
787 to_wakeup = first_worker(pool);
788 return to_wakeup ? to_wakeup->task : NULL;
792 * worker_set_flags - set worker flags and adjust nr_running accordingly
794 * @flags: flags to set
795 * @wakeup: wakeup an idle worker if necessary
797 * Set @flags in @worker->flags and adjust nr_running accordingly. If
798 * nr_running becomes zero and @wakeup is %true, an idle worker is
802 * spin_lock_irq(gcwq->lock)
804 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
807 struct worker_pool *pool = worker->pool;
809 WARN_ON_ONCE(worker->task != current);
812 * If transitioning into NOT_RUNNING, adjust nr_running and
813 * wake up an idle worker as necessary if requested by
816 if ((flags & WORKER_NOT_RUNNING) &&
817 !(worker->flags & WORKER_NOT_RUNNING)) {
818 atomic_t *nr_running = get_pool_nr_running(pool);
821 if (atomic_dec_and_test(nr_running) &&
822 !list_empty(&pool->worklist))
823 wake_up_worker(pool);
825 atomic_dec(nr_running);
828 worker->flags |= flags;
832 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
834 * @flags: flags to clear
836 * Clear @flags in @worker->flags and adjust nr_running accordingly.
839 * spin_lock_irq(gcwq->lock)
841 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
843 struct worker_pool *pool = worker->pool;
844 unsigned int oflags = worker->flags;
846 WARN_ON_ONCE(worker->task != current);
848 worker->flags &= ~flags;
851 * If transitioning out of NOT_RUNNING, increment nr_running. Note
852 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
853 * of multiple flags, not a single flag.
855 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
856 if (!(worker->flags & WORKER_NOT_RUNNING))
857 atomic_inc(get_pool_nr_running(pool));
861 * busy_worker_head - return the busy hash head for a work
862 * @gcwq: gcwq of interest
863 * @work: work to be hashed
865 * Return hash head of @gcwq for @work.
868 * spin_lock_irq(gcwq->lock).
871 * Pointer to the hash head.
873 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
874 struct work_struct *work)
876 const int base_shift = ilog2(sizeof(struct work_struct));
877 unsigned long v = (unsigned long)work;
879 /* simple shift and fold hash, do we need something better? */
881 v += v >> BUSY_WORKER_HASH_ORDER;
882 v &= BUSY_WORKER_HASH_MASK;
884 return &gcwq->busy_hash[v];
888 * __find_worker_executing_work - find worker which is executing a work
889 * @gcwq: gcwq of interest
890 * @bwh: hash head as returned by busy_worker_head()
891 * @work: work to find worker for
893 * Find a worker which is executing @work on @gcwq. @bwh should be
894 * the hash head obtained by calling busy_worker_head() with the same
898 * spin_lock_irq(gcwq->lock).
901 * Pointer to worker which is executing @work if found, NULL
904 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
905 struct hlist_head *bwh,
906 struct work_struct *work)
908 struct worker *worker;
909 struct hlist_node *tmp;
911 hlist_for_each_entry(worker, tmp, bwh, hentry)
912 if (worker->current_work == work)
918 * find_worker_executing_work - find worker which is executing a work
919 * @gcwq: gcwq of interest
920 * @work: work to find worker for
922 * Find a worker which is executing @work on @gcwq. This function is
923 * identical to __find_worker_executing_work() except that this
924 * function calculates @bwh itself.
927 * spin_lock_irq(gcwq->lock).
930 * Pointer to worker which is executing @work if found, NULL
933 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
934 struct work_struct *work)
936 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
941 * move_linked_works - move linked works to a list
942 * @work: start of series of works to be scheduled
943 * @head: target list to append @work to
944 * @nextp: out paramter for nested worklist walking
946 * Schedule linked works starting from @work to @head. Work series to
947 * be scheduled starts at @work and includes any consecutive work with
948 * WORK_STRUCT_LINKED set in its predecessor.
950 * If @nextp is not NULL, it's updated to point to the next work of
951 * the last scheduled work. This allows move_linked_works() to be
952 * nested inside outer list_for_each_entry_safe().
955 * spin_lock_irq(gcwq->lock).
957 static void move_linked_works(struct work_struct *work, struct list_head *head,
958 struct work_struct **nextp)
960 struct work_struct *n;
963 * Linked worklist will always end before the end of the list,
964 * use NULL for list head.
966 list_for_each_entry_safe_from(work, n, NULL, entry) {
967 list_move_tail(&work->entry, head);
968 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
973 * If we're already inside safe list traversal and have moved
974 * multiple works to the scheduled queue, the next position
975 * needs to be updated.
981 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
983 struct work_struct *work = list_first_entry(&cwq->delayed_works,
984 struct work_struct, entry);
986 trace_workqueue_activate_work(work);
987 move_linked_works(work, &cwq->pool->worklist, NULL);
988 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
993 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
994 * @cwq: cwq of interest
995 * @color: color of work which left the queue
996 * @delayed: for a delayed work
998 * A work either has completed or is removed from pending queue,
999 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1002 * spin_lock_irq(gcwq->lock).
1004 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1007 /* ignore uncolored works */
1008 if (color == WORK_NO_COLOR)
1011 cwq->nr_in_flight[color]--;
1015 if (!list_empty(&cwq->delayed_works)) {
1016 /* one down, submit a delayed one */
1017 if (cwq->nr_active < cwq->max_active)
1018 cwq_activate_first_delayed(cwq);
1022 /* is flush in progress and are we at the flushing tip? */
1023 if (likely(cwq->flush_color != color))
1026 /* are there still in-flight works? */
1027 if (cwq->nr_in_flight[color])
1030 /* this cwq is done, clear flush_color */
1031 cwq->flush_color = -1;
1034 * If this was the last cwq, wake up the first flusher. It
1035 * will handle the rest.
1037 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1038 complete(&cwq->wq->first_flusher->done);
1042 * try_to_grab_pending - steal work item from worklist and disable irq
1043 * @work: work item to steal
1044 * @is_dwork: @work is a delayed_work
1045 * @flags: place to store irq state
1047 * Try to grab PENDING bit of @work. This function can handle @work in any
1048 * stable state - idle, on timer or on worklist. Return values are
1050 * 1 if @work was pending and we successfully stole PENDING
1051 * 0 if @work was idle and we claimed PENDING
1052 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1053 * -ENOENT if someone else is canceling @work, this state may persist
1054 * for arbitrarily long
1056 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1057 * interrupted while holding PENDING and @work off queue, irq must be
1058 * disabled on entry. This, combined with delayed_work->timer being
1059 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1061 * On successful return, >= 0, irq is disabled and the caller is
1062 * responsible for releasing it using local_irq_restore(*@flags).
1064 * This function is safe to call from any context including IRQ handler.
1066 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1067 unsigned long *flags)
1069 struct global_cwq *gcwq;
1071 WARN_ON_ONCE(in_irq());
1073 local_irq_save(*flags);
1075 /* try to steal the timer if it exists */
1077 struct delayed_work *dwork = to_delayed_work(work);
1080 * dwork->timer is irqsafe. If del_timer() fails, it's
1081 * guaranteed that the timer is not queued anywhere and not
1082 * running on the local CPU.
1084 if (likely(del_timer(&dwork->timer)))
1088 /* try to claim PENDING the normal way */
1089 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1093 * The queueing is in progress, or it is already queued. Try to
1094 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1096 gcwq = get_work_gcwq(work);
1100 spin_lock(&gcwq->lock);
1101 if (!list_empty(&work->entry)) {
1103 * This work is queued, but perhaps we locked the wrong gcwq.
1104 * In that case we must see the new value after rmb(), see
1105 * insert_work()->wmb().
1108 if (gcwq == get_work_gcwq(work)) {
1109 debug_work_deactivate(work);
1110 list_del_init(&work->entry);
1111 cwq_dec_nr_in_flight(get_work_cwq(work),
1112 get_work_color(work),
1113 *work_data_bits(work) & WORK_STRUCT_DELAYED);
1115 spin_unlock(&gcwq->lock);
1119 spin_unlock(&gcwq->lock);
1121 local_irq_restore(*flags);
1122 if (work_is_canceling(work))
1129 * insert_work - insert a work into gcwq
1130 * @cwq: cwq @work belongs to
1131 * @work: work to insert
1132 * @head: insertion point
1133 * @extra_flags: extra WORK_STRUCT_* flags to set
1135 * Insert @work which belongs to @cwq into @gcwq after @head.
1136 * @extra_flags is or'd to work_struct flags.
1139 * spin_lock_irq(gcwq->lock).
1141 static void insert_work(struct cpu_workqueue_struct *cwq,
1142 struct work_struct *work, struct list_head *head,
1143 unsigned int extra_flags)
1145 struct worker_pool *pool = cwq->pool;
1147 /* we own @work, set data and link */
1148 set_work_cwq(work, cwq, extra_flags);
1151 * Ensure that we get the right work->data if we see the
1152 * result of list_add() below, see try_to_grab_pending().
1156 list_add_tail(&work->entry, head);
1159 * Ensure either worker_sched_deactivated() sees the above
1160 * list_add_tail() or we see zero nr_running to avoid workers
1161 * lying around lazily while there are works to be processed.
1165 if (__need_more_worker(pool))
1166 wake_up_worker(pool);
1170 * Test whether @work is being queued from another work executing on the
1171 * same workqueue. This is rather expensive and should only be used from
1174 static bool is_chained_work(struct workqueue_struct *wq)
1176 unsigned long flags;
1179 for_each_gcwq_cpu(cpu) {
1180 struct global_cwq *gcwq = get_gcwq(cpu);
1181 struct worker *worker;
1182 struct hlist_node *pos;
1185 spin_lock_irqsave(&gcwq->lock, flags);
1186 for_each_busy_worker(worker, i, pos, gcwq) {
1187 if (worker->task != current)
1189 spin_unlock_irqrestore(&gcwq->lock, flags);
1191 * I'm @worker, no locking necessary. See if @work
1192 * is headed to the same workqueue.
1194 return worker->current_cwq->wq == wq;
1196 spin_unlock_irqrestore(&gcwq->lock, flags);
1201 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1202 struct work_struct *work)
1204 struct global_cwq *gcwq;
1205 struct cpu_workqueue_struct *cwq;
1206 struct list_head *worklist;
1207 unsigned int work_flags;
1208 unsigned int req_cpu = cpu;
1211 * While a work item is PENDING && off queue, a task trying to
1212 * steal the PENDING will busy-loop waiting for it to either get
1213 * queued or lose PENDING. Grabbing PENDING and queueing should
1214 * happen with IRQ disabled.
1216 WARN_ON_ONCE(!irqs_disabled());
1218 debug_work_activate(work);
1220 /* if dying, only works from the same workqueue are allowed */
1221 if (unlikely(wq->flags & WQ_DRAINING) &&
1222 WARN_ON_ONCE(!is_chained_work(wq)))
1225 /* determine gcwq to use */
1226 if (!(wq->flags & WQ_UNBOUND)) {
1227 struct global_cwq *last_gcwq;
1229 if (cpu == WORK_CPU_UNBOUND)
1230 cpu = raw_smp_processor_id();
1233 * It's multi cpu. If @work was previously on a different
1234 * cpu, it might still be running there, in which case the
1235 * work needs to be queued on that cpu to guarantee
1238 gcwq = get_gcwq(cpu);
1239 last_gcwq = get_work_gcwq(work);
1241 if (last_gcwq && last_gcwq != gcwq) {
1242 struct worker *worker;
1244 spin_lock(&last_gcwq->lock);
1246 worker = find_worker_executing_work(last_gcwq, work);
1248 if (worker && worker->current_cwq->wq == wq)
1251 /* meh... not running there, queue here */
1252 spin_unlock(&last_gcwq->lock);
1253 spin_lock(&gcwq->lock);
1256 spin_lock(&gcwq->lock);
1259 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1260 spin_lock(&gcwq->lock);
1263 /* gcwq determined, get cwq and queue */
1264 cwq = get_cwq(gcwq->cpu, wq);
1265 trace_workqueue_queue_work(req_cpu, cwq, work);
1267 if (WARN_ON(!list_empty(&work->entry))) {
1268 spin_unlock(&gcwq->lock);
1272 cwq->nr_in_flight[cwq->work_color]++;
1273 work_flags = work_color_to_flags(cwq->work_color);
1275 if (likely(cwq->nr_active < cwq->max_active)) {
1276 trace_workqueue_activate_work(work);
1278 worklist = &cwq->pool->worklist;
1280 work_flags |= WORK_STRUCT_DELAYED;
1281 worklist = &cwq->delayed_works;
1284 insert_work(cwq, work, worklist, work_flags);
1286 spin_unlock(&gcwq->lock);
1290 * queue_work_on - queue work on specific cpu
1291 * @cpu: CPU number to execute work on
1292 * @wq: workqueue to use
1293 * @work: work to queue
1295 * Returns %false if @work was already on a queue, %true otherwise.
1297 * We queue the work to a specific CPU, the caller must ensure it
1300 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1301 struct work_struct *work)
1304 unsigned long flags;
1306 local_irq_save(flags);
1308 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1309 __queue_work(cpu, wq, work);
1313 local_irq_restore(flags);
1316 EXPORT_SYMBOL_GPL(queue_work_on);
1319 * queue_work - queue work on a workqueue
1320 * @wq: workqueue to use
1321 * @work: work to queue
1323 * Returns %false if @work was already on a queue, %true otherwise.
1325 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1326 * it can be processed by another CPU.
1328 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1330 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1332 EXPORT_SYMBOL_GPL(queue_work);
1334 void delayed_work_timer_fn(unsigned long __data)
1336 struct delayed_work *dwork = (struct delayed_work *)__data;
1337 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1339 /* should have been called from irqsafe timer with irq already off */
1340 __queue_work(dwork->cpu, cwq->wq, &dwork->work);
1342 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1344 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1345 struct delayed_work *dwork, unsigned long delay)
1347 struct timer_list *timer = &dwork->timer;
1348 struct work_struct *work = &dwork->work;
1351 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1352 timer->data != (unsigned long)dwork);
1353 BUG_ON(timer_pending(timer));
1354 BUG_ON(!list_empty(&work->entry));
1356 timer_stats_timer_set_start_info(&dwork->timer);
1359 * This stores cwq for the moment, for the timer_fn. Note that the
1360 * work's gcwq is preserved to allow reentrance detection for
1363 if (!(wq->flags & WQ_UNBOUND)) {
1364 struct global_cwq *gcwq = get_work_gcwq(work);
1367 * If we cannot get the last gcwq from @work directly,
1368 * select the last CPU such that it avoids unnecessarily
1369 * triggering non-reentrancy check in __queue_work().
1374 if (lcpu == WORK_CPU_UNBOUND)
1375 lcpu = raw_smp_processor_id();
1377 lcpu = WORK_CPU_UNBOUND;
1380 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1383 timer->expires = jiffies + delay;
1385 if (unlikely(cpu != WORK_CPU_UNBOUND))
1386 add_timer_on(timer, cpu);
1392 * queue_delayed_work_on - queue work on specific CPU after delay
1393 * @cpu: CPU number to execute work on
1394 * @wq: workqueue to use
1395 * @dwork: work to queue
1396 * @delay: number of jiffies to wait before queueing
1398 * Returns %false if @work was already on a queue, %true otherwise. If
1399 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1402 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1403 struct delayed_work *dwork, unsigned long delay)
1405 struct work_struct *work = &dwork->work;
1407 unsigned long flags;
1410 return queue_work_on(cpu, wq, &dwork->work);
1412 /* read the comment in __queue_work() */
1413 local_irq_save(flags);
1415 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1416 __queue_delayed_work(cpu, wq, dwork, delay);
1420 local_irq_restore(flags);
1423 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1426 * queue_delayed_work - queue work on a workqueue after delay
1427 * @wq: workqueue to use
1428 * @dwork: delayable work to queue
1429 * @delay: number of jiffies to wait before queueing
1431 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1433 bool queue_delayed_work(struct workqueue_struct *wq,
1434 struct delayed_work *dwork, unsigned long delay)
1436 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1438 EXPORT_SYMBOL_GPL(queue_delayed_work);
1441 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1442 * @cpu: CPU number to execute work on
1443 * @wq: workqueue to use
1444 * @dwork: work to queue
1445 * @delay: number of jiffies to wait before queueing
1447 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1448 * modify @dwork's timer so that it expires after @delay. If @delay is
1449 * zero, @work is guaranteed to be scheduled immediately regardless of its
1452 * Returns %false if @dwork was idle and queued, %true if @dwork was
1453 * pending and its timer was modified.
1455 * This function is safe to call from any context including IRQ handler.
1456 * See try_to_grab_pending() for details.
1458 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1459 struct delayed_work *dwork, unsigned long delay)
1461 unsigned long flags;
1465 ret = try_to_grab_pending(&dwork->work, true, &flags);
1466 } while (unlikely(ret == -EAGAIN));
1468 if (likely(ret >= 0)) {
1469 __queue_delayed_work(cpu, wq, dwork, delay);
1470 local_irq_restore(flags);
1473 /* -ENOENT from try_to_grab_pending() becomes %true */
1476 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1479 * mod_delayed_work - modify delay of or queue a delayed work
1480 * @wq: workqueue to use
1481 * @dwork: work to queue
1482 * @delay: number of jiffies to wait before queueing
1484 * mod_delayed_work_on() on local CPU.
1486 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1487 unsigned long delay)
1489 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1491 EXPORT_SYMBOL_GPL(mod_delayed_work);
1494 * worker_enter_idle - enter idle state
1495 * @worker: worker which is entering idle state
1497 * @worker is entering idle state. Update stats and idle timer if
1501 * spin_lock_irq(gcwq->lock).
1503 static void worker_enter_idle(struct worker *worker)
1505 struct worker_pool *pool = worker->pool;
1506 struct global_cwq *gcwq = pool->gcwq;
1508 BUG_ON(worker->flags & WORKER_IDLE);
1509 BUG_ON(!list_empty(&worker->entry) &&
1510 (worker->hentry.next || worker->hentry.pprev));
1512 /* can't use worker_set_flags(), also called from start_worker() */
1513 worker->flags |= WORKER_IDLE;
1515 worker->last_active = jiffies;
1517 /* idle_list is LIFO */
1518 list_add(&worker->entry, &pool->idle_list);
1520 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1521 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1524 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1525 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1526 * nr_running, the warning may trigger spuriously. Check iff
1527 * unbind is not in progress.
1529 WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
1530 pool->nr_workers == pool->nr_idle &&
1531 atomic_read(get_pool_nr_running(pool)));
1535 * worker_leave_idle - leave idle state
1536 * @worker: worker which is leaving idle state
1538 * @worker is leaving idle state. Update stats.
1541 * spin_lock_irq(gcwq->lock).
1543 static void worker_leave_idle(struct worker *worker)
1545 struct worker_pool *pool = worker->pool;
1547 BUG_ON(!(worker->flags & WORKER_IDLE));
1548 worker_clr_flags(worker, WORKER_IDLE);
1550 list_del_init(&worker->entry);
1554 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1557 * Works which are scheduled while the cpu is online must at least be
1558 * scheduled to a worker which is bound to the cpu so that if they are
1559 * flushed from cpu callbacks while cpu is going down, they are
1560 * guaranteed to execute on the cpu.
1562 * This function is to be used by rogue workers and rescuers to bind
1563 * themselves to the target cpu and may race with cpu going down or
1564 * coming online. kthread_bind() can't be used because it may put the
1565 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1566 * verbatim as it's best effort and blocking and gcwq may be
1567 * [dis]associated in the meantime.
1569 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1570 * binding against %GCWQ_DISASSOCIATED which is set during
1571 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1572 * enters idle state or fetches works without dropping lock, it can
1573 * guarantee the scheduling requirement described in the first paragraph.
1576 * Might sleep. Called without any lock but returns with gcwq->lock
1580 * %true if the associated gcwq is online (@worker is successfully
1581 * bound), %false if offline.
1583 static bool worker_maybe_bind_and_lock(struct worker *worker)
1584 __acquires(&gcwq->lock)
1586 struct global_cwq *gcwq = worker->pool->gcwq;
1587 struct task_struct *task = worker->task;
1591 * The following call may fail, succeed or succeed
1592 * without actually migrating the task to the cpu if
1593 * it races with cpu hotunplug operation. Verify
1594 * against GCWQ_DISASSOCIATED.
1596 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1597 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1599 spin_lock_irq(&gcwq->lock);
1600 if (gcwq->flags & GCWQ_DISASSOCIATED)
1602 if (task_cpu(task) == gcwq->cpu &&
1603 cpumask_equal(¤t->cpus_allowed,
1604 get_cpu_mask(gcwq->cpu)))
1606 spin_unlock_irq(&gcwq->lock);
1609 * We've raced with CPU hot[un]plug. Give it a breather
1610 * and retry migration. cond_resched() is required here;
1611 * otherwise, we might deadlock against cpu_stop trying to
1612 * bring down the CPU on non-preemptive kernel.
1620 * Rebind an idle @worker to its CPU. worker_thread() will test
1621 * %WORKER_REBIND before leaving idle and call this function.
1623 static void idle_worker_rebind(struct worker *worker)
1625 struct global_cwq *gcwq = worker->pool->gcwq;
1628 * CPU may go down again inbetween. If rebinding fails, reinstate
1629 * UNBOUND. We're off idle_list and nobody else can do it for us.
1631 if (!worker_maybe_bind_and_lock(worker))
1632 worker->flags |= WORKER_UNBOUND;
1634 worker_clr_flags(worker, WORKER_REBIND);
1636 /* rebind complete, become available again */
1637 list_add(&worker->entry, &worker->pool->idle_list);
1638 spin_unlock_irq(&gcwq->lock);
1642 * Function for @worker->rebind.work used to rebind unbound busy workers to
1643 * the associated cpu which is coming back online. This is scheduled by
1644 * cpu up but can race with other cpu hotplug operations and may be
1645 * executed twice without intervening cpu down.
1647 static void busy_worker_rebind_fn(struct work_struct *work)
1649 struct worker *worker = container_of(work, struct worker, rebind_work);
1650 struct global_cwq *gcwq = worker->pool->gcwq;
1652 if (worker_maybe_bind_and_lock(worker))
1653 worker_clr_flags(worker, WORKER_UNBOUND);
1655 spin_unlock_irq(&gcwq->lock);
1659 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1660 * @gcwq: gcwq of interest
1662 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1663 * is different for idle and busy ones.
1665 * Idle ones will be removed from the idle_list and woken up. They will
1666 * add themselves back after completing rebind. This ensures that the
1667 * idle_list doesn't contain any unbound workers when re-bound busy workers
1668 * try to perform local wake-ups for concurrency management.
1670 * Busy workers can rebind after they finish their current work items.
1671 * Queueing the rebind work item at the head of the scheduled list is
1672 * enough. Note that nr_running will be properly bumped as busy workers
1675 * On return, all non-manager workers are scheduled for rebind - see
1676 * manage_workers() for the manager special case. Any idle worker
1677 * including the manager will not appear on @idle_list until rebind is
1678 * complete, making local wake-ups safe.
1680 static void rebind_workers(struct global_cwq *gcwq)
1682 struct worker_pool *pool;
1683 struct worker *worker, *n;
1684 struct hlist_node *pos;
1687 lockdep_assert_held(&gcwq->lock);
1689 for_each_worker_pool(pool, gcwq)
1690 lockdep_assert_held(&pool->manager_mutex);
1692 /* set REBIND and kick idle ones */
1693 for_each_worker_pool(pool, gcwq) {
1694 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1695 unsigned long worker_flags = worker->flags;
1697 /* morph UNBOUND to REBIND atomically */
1698 worker_flags &= ~WORKER_UNBOUND;
1699 worker_flags |= WORKER_REBIND;
1700 ACCESS_ONCE(worker->flags) = worker_flags;
1703 * idle workers should be off @pool->idle_list
1704 * until rebind is complete to avoid receiving
1705 * premature local wake-ups.
1707 list_del_init(&worker->entry);
1709 /* worker_thread() will call idle_worker_rebind() */
1710 wake_up_process(worker->task);
1714 /* rebind busy workers */
1715 for_each_busy_worker(worker, i, pos, gcwq) {
1716 struct work_struct *rebind_work = &worker->rebind_work;
1717 struct workqueue_struct *wq;
1719 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1720 work_data_bits(rebind_work)))
1723 debug_work_activate(rebind_work);
1726 * wq doesn't really matter but let's keep @worker->pool
1727 * and @cwq->pool consistent for sanity.
1729 if (worker_pool_pri(worker->pool))
1730 wq = system_highpri_wq;
1734 insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
1735 worker->scheduled.next,
1736 work_color_to_flags(WORK_NO_COLOR));
1740 static struct worker *alloc_worker(void)
1742 struct worker *worker;
1744 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1746 INIT_LIST_HEAD(&worker->entry);
1747 INIT_LIST_HEAD(&worker->scheduled);
1748 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1749 /* on creation a worker is in !idle && prep state */
1750 worker->flags = WORKER_PREP;
1756 * create_worker - create a new workqueue worker
1757 * @pool: pool the new worker will belong to
1759 * Create a new worker which is bound to @pool. The returned worker
1760 * can be started by calling start_worker() or destroyed using
1764 * Might sleep. Does GFP_KERNEL allocations.
1767 * Pointer to the newly created worker.
1769 static struct worker *create_worker(struct worker_pool *pool)
1771 struct global_cwq *gcwq = pool->gcwq;
1772 const char *pri = worker_pool_pri(pool) ? "H" : "";
1773 struct worker *worker = NULL;
1776 spin_lock_irq(&gcwq->lock);
1777 while (ida_get_new(&pool->worker_ida, &id)) {
1778 spin_unlock_irq(&gcwq->lock);
1779 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1781 spin_lock_irq(&gcwq->lock);
1783 spin_unlock_irq(&gcwq->lock);
1785 worker = alloc_worker();
1789 worker->pool = pool;
1792 if (gcwq->cpu != WORK_CPU_UNBOUND)
1793 worker->task = kthread_create_on_node(worker_thread,
1794 worker, cpu_to_node(gcwq->cpu),
1795 "kworker/%u:%d%s", gcwq->cpu, id, pri);
1797 worker->task = kthread_create(worker_thread, worker,
1798 "kworker/u:%d%s", id, pri);
1799 if (IS_ERR(worker->task))
1802 if (worker_pool_pri(pool))
1803 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1806 * Determine CPU binding of the new worker depending on
1807 * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1808 * flag remains stable across this function. See the comments
1809 * above the flag definition for details.
1811 * As an unbound worker may later become a regular one if CPU comes
1812 * online, make sure every worker has %PF_THREAD_BOUND set.
1814 if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
1815 kthread_bind(worker->task, gcwq->cpu);
1817 worker->task->flags |= PF_THREAD_BOUND;
1818 worker->flags |= WORKER_UNBOUND;
1824 spin_lock_irq(&gcwq->lock);
1825 ida_remove(&pool->worker_ida, id);
1826 spin_unlock_irq(&gcwq->lock);
1833 * start_worker - start a newly created worker
1834 * @worker: worker to start
1836 * Make the gcwq aware of @worker and start it.
1839 * spin_lock_irq(gcwq->lock).
1841 static void start_worker(struct worker *worker)
1843 worker->flags |= WORKER_STARTED;
1844 worker->pool->nr_workers++;
1845 worker_enter_idle(worker);
1846 wake_up_process(worker->task);
1850 * destroy_worker - destroy a workqueue worker
1851 * @worker: worker to be destroyed
1853 * Destroy @worker and adjust @gcwq stats accordingly.
1856 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1858 static void destroy_worker(struct worker *worker)
1860 struct worker_pool *pool = worker->pool;
1861 struct global_cwq *gcwq = pool->gcwq;
1862 int id = worker->id;
1864 /* sanity check frenzy */
1865 BUG_ON(worker->current_work);
1866 BUG_ON(!list_empty(&worker->scheduled));
1868 if (worker->flags & WORKER_STARTED)
1870 if (worker->flags & WORKER_IDLE)
1873 list_del_init(&worker->entry);
1874 worker->flags |= WORKER_DIE;
1876 spin_unlock_irq(&gcwq->lock);
1878 kthread_stop(worker->task);
1881 spin_lock_irq(&gcwq->lock);
1882 ida_remove(&pool->worker_ida, id);
1885 static void idle_worker_timeout(unsigned long __pool)
1887 struct worker_pool *pool = (void *)__pool;
1888 struct global_cwq *gcwq = pool->gcwq;
1890 spin_lock_irq(&gcwq->lock);
1892 if (too_many_workers(pool)) {
1893 struct worker *worker;
1894 unsigned long expires;
1896 /* idle_list is kept in LIFO order, check the last one */
1897 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1898 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1900 if (time_before(jiffies, expires))
1901 mod_timer(&pool->idle_timer, expires);
1903 /* it's been idle for too long, wake up manager */
1904 pool->flags |= POOL_MANAGE_WORKERS;
1905 wake_up_worker(pool);
1909 spin_unlock_irq(&gcwq->lock);
1912 static bool send_mayday(struct work_struct *work)
1914 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1915 struct workqueue_struct *wq = cwq->wq;
1918 if (!(wq->flags & WQ_RESCUER))
1921 /* mayday mayday mayday */
1922 cpu = cwq->pool->gcwq->cpu;
1923 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1924 if (cpu == WORK_CPU_UNBOUND)
1926 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1927 wake_up_process(wq->rescuer->task);
1931 static void gcwq_mayday_timeout(unsigned long __pool)
1933 struct worker_pool *pool = (void *)__pool;
1934 struct global_cwq *gcwq = pool->gcwq;
1935 struct work_struct *work;
1937 spin_lock_irq(&gcwq->lock);
1939 if (need_to_create_worker(pool)) {
1941 * We've been trying to create a new worker but
1942 * haven't been successful. We might be hitting an
1943 * allocation deadlock. Send distress signals to
1946 list_for_each_entry(work, &pool->worklist, entry)
1950 spin_unlock_irq(&gcwq->lock);
1952 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1956 * maybe_create_worker - create a new worker if necessary
1957 * @pool: pool to create a new worker for
1959 * Create a new worker for @pool if necessary. @pool is guaranteed to
1960 * have at least one idle worker on return from this function. If
1961 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1962 * sent to all rescuers with works scheduled on @pool to resolve
1963 * possible allocation deadlock.
1965 * On return, need_to_create_worker() is guaranteed to be false and
1966 * may_start_working() true.
1969 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1970 * multiple times. Does GFP_KERNEL allocations. Called only from
1974 * false if no action was taken and gcwq->lock stayed locked, true
1977 static bool maybe_create_worker(struct worker_pool *pool)
1978 __releases(&gcwq->lock)
1979 __acquires(&gcwq->lock)
1981 struct global_cwq *gcwq = pool->gcwq;
1983 if (!need_to_create_worker(pool))
1986 spin_unlock_irq(&gcwq->lock);
1988 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1989 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1992 struct worker *worker;
1994 worker = create_worker(pool);
1996 del_timer_sync(&pool->mayday_timer);
1997 spin_lock_irq(&gcwq->lock);
1998 start_worker(worker);
1999 BUG_ON(need_to_create_worker(pool));
2003 if (!need_to_create_worker(pool))
2006 __set_current_state(TASK_INTERRUPTIBLE);
2007 schedule_timeout(CREATE_COOLDOWN);
2009 if (!need_to_create_worker(pool))
2013 del_timer_sync(&pool->mayday_timer);
2014 spin_lock_irq(&gcwq->lock);
2015 if (need_to_create_worker(pool))
2021 * maybe_destroy_worker - destroy workers which have been idle for a while
2022 * @pool: pool to destroy workers for
2024 * Destroy @pool workers which have been idle for longer than
2025 * IDLE_WORKER_TIMEOUT.
2028 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2029 * multiple times. Called only from manager.
2032 * false if no action was taken and gcwq->lock stayed locked, true
2035 static bool maybe_destroy_workers(struct worker_pool *pool)
2039 while (too_many_workers(pool)) {
2040 struct worker *worker;
2041 unsigned long expires;
2043 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2044 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2046 if (time_before(jiffies, expires)) {
2047 mod_timer(&pool->idle_timer, expires);
2051 destroy_worker(worker);
2059 * manage_workers - manage worker pool
2062 * Assume the manager role and manage gcwq worker pool @worker belongs
2063 * to. At any given time, there can be only zero or one manager per
2064 * gcwq. The exclusion is handled automatically by this function.
2066 * The caller can safely start processing works on false return. On
2067 * true return, it's guaranteed that need_to_create_worker() is false
2068 * and may_start_working() is true.
2071 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2072 * multiple times. Does GFP_KERNEL allocations.
2075 * false if no action was taken and gcwq->lock stayed locked, true if
2076 * some action was taken.
2078 static bool manage_workers(struct worker *worker)
2080 struct worker_pool *pool = worker->pool;
2083 if (pool->flags & POOL_MANAGING_WORKERS)
2086 pool->flags |= POOL_MANAGING_WORKERS;
2089 * To simplify both worker management and CPU hotplug, hold off
2090 * management while hotplug is in progress. CPU hotplug path can't
2091 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2092 * lead to idle worker depletion (all become busy thinking someone
2093 * else is managing) which in turn can result in deadlock under
2094 * extreme circumstances. Use @pool->manager_mutex to synchronize
2095 * manager against CPU hotplug.
2097 * manager_mutex would always be free unless CPU hotplug is in
2098 * progress. trylock first without dropping @gcwq->lock.
2100 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2101 spin_unlock_irq(&pool->gcwq->lock);
2102 mutex_lock(&pool->manager_mutex);
2104 * CPU hotplug could have happened while we were waiting
2105 * for manager_mutex. Hotplug itself can't handle us
2106 * because manager isn't either on idle or busy list, and
2107 * @gcwq's state and ours could have deviated.
2109 * As hotplug is now excluded via manager_mutex, we can
2110 * simply try to bind. It will succeed or fail depending
2111 * on @gcwq's current state. Try it and adjust
2112 * %WORKER_UNBOUND accordingly.
2114 if (worker_maybe_bind_and_lock(worker))
2115 worker->flags &= ~WORKER_UNBOUND;
2117 worker->flags |= WORKER_UNBOUND;
2122 pool->flags &= ~POOL_MANAGE_WORKERS;
2125 * Destroy and then create so that may_start_working() is true
2128 ret |= maybe_destroy_workers(pool);
2129 ret |= maybe_create_worker(pool);
2131 pool->flags &= ~POOL_MANAGING_WORKERS;
2132 mutex_unlock(&pool->manager_mutex);
2137 * process_one_work - process single work
2139 * @work: work to process
2141 * Process @work. This function contains all the logics necessary to
2142 * process a single work including synchronization against and
2143 * interaction with other workers on the same cpu, queueing and
2144 * flushing. As long as context requirement is met, any worker can
2145 * call this function to process a work.
2148 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2150 static void process_one_work(struct worker *worker, struct work_struct *work)
2151 __releases(&gcwq->lock)
2152 __acquires(&gcwq->lock)
2154 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2155 struct worker_pool *pool = worker->pool;
2156 struct global_cwq *gcwq = pool->gcwq;
2157 struct hlist_head *bwh = busy_worker_head(gcwq, work);
2158 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2159 work_func_t f = work->func;
2161 struct worker *collision;
2162 #ifdef CONFIG_LOCKDEP
2164 * It is permissible to free the struct work_struct from
2165 * inside the function that is called from it, this we need to
2166 * take into account for lockdep too. To avoid bogus "held
2167 * lock freed" warnings as well as problems when looking into
2168 * work->lockdep_map, make a copy and use that here.
2170 struct lockdep_map lockdep_map;
2172 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2175 * Ensure we're on the correct CPU. DISASSOCIATED test is
2176 * necessary to avoid spurious warnings from rescuers servicing the
2177 * unbound or a disassociated gcwq.
2179 WARN_ON_ONCE(!(worker->flags & (WORKER_UNBOUND | WORKER_REBIND)) &&
2180 !(gcwq->flags & GCWQ_DISASSOCIATED) &&
2181 raw_smp_processor_id() != gcwq->cpu);
2184 * A single work shouldn't be executed concurrently by
2185 * multiple workers on a single cpu. Check whether anyone is
2186 * already processing the work. If so, defer the work to the
2187 * currently executing one.
2189 collision = __find_worker_executing_work(gcwq, bwh, work);
2190 if (unlikely(collision)) {
2191 move_linked_works(work, &collision->scheduled, NULL);
2195 /* claim and dequeue */
2196 debug_work_deactivate(work);
2197 hlist_add_head(&worker->hentry, bwh);
2198 worker->current_work = work;
2199 worker->current_cwq = cwq;
2200 work_color = get_work_color(work);
2202 list_del_init(&work->entry);
2205 * CPU intensive works don't participate in concurrency
2206 * management. They're the scheduler's responsibility.
2208 if (unlikely(cpu_intensive))
2209 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2212 * Unbound gcwq isn't concurrency managed and work items should be
2213 * executed ASAP. Wake up another worker if necessary.
2215 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2216 wake_up_worker(pool);
2219 * Record the last CPU and clear PENDING which should be the last
2220 * update to @work. Also, do this inside @gcwq->lock so that
2221 * PENDING and queued state changes happen together while IRQ is
2224 set_work_cpu_and_clear_pending(work, gcwq->cpu);
2226 spin_unlock_irq(&gcwq->lock);
2228 lock_map_acquire_read(&cwq->wq->lockdep_map);
2229 lock_map_acquire(&lockdep_map);
2230 trace_workqueue_execute_start(work);
2233 * While we must be careful to not use "work" after this, the trace
2234 * point will only record its address.
2236 trace_workqueue_execute_end(work);
2237 lock_map_release(&lockdep_map);
2238 lock_map_release(&cwq->wq->lockdep_map);
2240 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2241 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2242 " last function: %pf\n",
2243 current->comm, preempt_count(), task_pid_nr(current), f);
2244 debug_show_held_locks(current);
2248 spin_lock_irq(&gcwq->lock);
2250 /* clear cpu intensive status */
2251 if (unlikely(cpu_intensive))
2252 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2254 /* we're done with it, release */
2255 hlist_del_init(&worker->hentry);
2256 worker->current_work = NULL;
2257 worker->current_cwq = NULL;
2258 cwq_dec_nr_in_flight(cwq, work_color, false);
2262 * process_scheduled_works - process scheduled works
2265 * Process all scheduled works. Please note that the scheduled list
2266 * may change while processing a work, so this function repeatedly
2267 * fetches a work from the top and executes it.
2270 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2273 static void process_scheduled_works(struct worker *worker)
2275 while (!list_empty(&worker->scheduled)) {
2276 struct work_struct *work = list_first_entry(&worker->scheduled,
2277 struct work_struct, entry);
2278 process_one_work(worker, work);
2283 * worker_thread - the worker thread function
2286 * The gcwq worker thread function. There's a single dynamic pool of
2287 * these per each cpu. These workers process all works regardless of
2288 * their specific target workqueue. The only exception is works which
2289 * belong to workqueues with a rescuer which will be explained in
2292 static int worker_thread(void *__worker)
2294 struct worker *worker = __worker;
2295 struct worker_pool *pool = worker->pool;
2296 struct global_cwq *gcwq = pool->gcwq;
2298 /* tell the scheduler that this is a workqueue worker */
2299 worker->task->flags |= PF_WQ_WORKER;
2301 spin_lock_irq(&gcwq->lock);
2304 * DIE can be set only while idle and REBIND set while busy has
2305 * @worker->rebind_work scheduled. Checking here is enough.
2307 if (unlikely(worker->flags & (WORKER_REBIND | WORKER_DIE))) {
2308 spin_unlock_irq(&gcwq->lock);
2310 if (worker->flags & WORKER_DIE) {
2311 worker->task->flags &= ~PF_WQ_WORKER;
2315 idle_worker_rebind(worker);
2319 worker_leave_idle(worker);
2321 /* no more worker necessary? */
2322 if (!need_more_worker(pool))
2325 /* do we need to manage? */
2326 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2330 * ->scheduled list can only be filled while a worker is
2331 * preparing to process a work or actually processing it.
2332 * Make sure nobody diddled with it while I was sleeping.
2334 BUG_ON(!list_empty(&worker->scheduled));
2337 * When control reaches this point, we're guaranteed to have
2338 * at least one idle worker or that someone else has already
2339 * assumed the manager role.
2341 worker_clr_flags(worker, WORKER_PREP);
2344 struct work_struct *work =
2345 list_first_entry(&pool->worklist,
2346 struct work_struct, entry);
2348 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2349 /* optimization path, not strictly necessary */
2350 process_one_work(worker, work);
2351 if (unlikely(!list_empty(&worker->scheduled)))
2352 process_scheduled_works(worker);
2354 move_linked_works(work, &worker->scheduled, NULL);
2355 process_scheduled_works(worker);
2357 } while (keep_working(pool));
2359 worker_set_flags(worker, WORKER_PREP, false);
2361 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2365 * gcwq->lock is held and there's no work to process and no
2366 * need to manage, sleep. Workers are woken up only while
2367 * holding gcwq->lock or from local cpu, so setting the
2368 * current state before releasing gcwq->lock is enough to
2369 * prevent losing any event.
2371 worker_enter_idle(worker);
2372 __set_current_state(TASK_INTERRUPTIBLE);
2373 spin_unlock_irq(&gcwq->lock);
2379 * rescuer_thread - the rescuer thread function
2380 * @__wq: the associated workqueue
2382 * Workqueue rescuer thread function. There's one rescuer for each
2383 * workqueue which has WQ_RESCUER set.
2385 * Regular work processing on a gcwq may block trying to create a new
2386 * worker which uses GFP_KERNEL allocation which has slight chance of
2387 * developing into deadlock if some works currently on the same queue
2388 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2389 * the problem rescuer solves.
2391 * When such condition is possible, the gcwq summons rescuers of all
2392 * workqueues which have works queued on the gcwq and let them process
2393 * those works so that forward progress can be guaranteed.
2395 * This should happen rarely.
2397 static int rescuer_thread(void *__wq)
2399 struct workqueue_struct *wq = __wq;
2400 struct worker *rescuer = wq->rescuer;
2401 struct list_head *scheduled = &rescuer->scheduled;
2402 bool is_unbound = wq->flags & WQ_UNBOUND;
2405 set_user_nice(current, RESCUER_NICE_LEVEL);
2407 set_current_state(TASK_INTERRUPTIBLE);
2409 if (kthread_should_stop())
2413 * See whether any cpu is asking for help. Unbounded
2414 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2416 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2417 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2418 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2419 struct worker_pool *pool = cwq->pool;
2420 struct global_cwq *gcwq = pool->gcwq;
2421 struct work_struct *work, *n;
2423 __set_current_state(TASK_RUNNING);
2424 mayday_clear_cpu(cpu, wq->mayday_mask);
2426 /* migrate to the target cpu if possible */
2427 rescuer->pool = pool;
2428 worker_maybe_bind_and_lock(rescuer);
2431 * Slurp in all works issued via this workqueue and
2434 BUG_ON(!list_empty(&rescuer->scheduled));
2435 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2436 if (get_work_cwq(work) == cwq)
2437 move_linked_works(work, scheduled, &n);
2439 process_scheduled_works(rescuer);
2442 * Leave this gcwq. If keep_working() is %true, notify a
2443 * regular worker; otherwise, we end up with 0 concurrency
2444 * and stalling the execution.
2446 if (keep_working(pool))
2447 wake_up_worker(pool);
2449 spin_unlock_irq(&gcwq->lock);
2457 struct work_struct work;
2458 struct completion done;
2461 static void wq_barrier_func(struct work_struct *work)
2463 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2464 complete(&barr->done);
2468 * insert_wq_barrier - insert a barrier work
2469 * @cwq: cwq to insert barrier into
2470 * @barr: wq_barrier to insert
2471 * @target: target work to attach @barr to
2472 * @worker: worker currently executing @target, NULL if @target is not executing
2474 * @barr is linked to @target such that @barr is completed only after
2475 * @target finishes execution. Please note that the ordering
2476 * guarantee is observed only with respect to @target and on the local
2479 * Currently, a queued barrier can't be canceled. This is because
2480 * try_to_grab_pending() can't determine whether the work to be
2481 * grabbed is at the head of the queue and thus can't clear LINKED
2482 * flag of the previous work while there must be a valid next work
2483 * after a work with LINKED flag set.
2485 * Note that when @worker is non-NULL, @target may be modified
2486 * underneath us, so we can't reliably determine cwq from @target.
2489 * spin_lock_irq(gcwq->lock).
2491 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2492 struct wq_barrier *barr,
2493 struct work_struct *target, struct worker *worker)
2495 struct list_head *head;
2496 unsigned int linked = 0;
2499 * debugobject calls are safe here even with gcwq->lock locked
2500 * as we know for sure that this will not trigger any of the
2501 * checks and call back into the fixup functions where we
2504 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2505 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2506 init_completion(&barr->done);
2509 * If @target is currently being executed, schedule the
2510 * barrier to the worker; otherwise, put it after @target.
2513 head = worker->scheduled.next;
2515 unsigned long *bits = work_data_bits(target);
2517 head = target->entry.next;
2518 /* there can already be other linked works, inherit and set */
2519 linked = *bits & WORK_STRUCT_LINKED;
2520 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2523 debug_work_activate(&barr->work);
2524 insert_work(cwq, &barr->work, head,
2525 work_color_to_flags(WORK_NO_COLOR) | linked);
2529 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2530 * @wq: workqueue being flushed
2531 * @flush_color: new flush color, < 0 for no-op
2532 * @work_color: new work color, < 0 for no-op
2534 * Prepare cwqs for workqueue flushing.
2536 * If @flush_color is non-negative, flush_color on all cwqs should be
2537 * -1. If no cwq has in-flight commands at the specified color, all
2538 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2539 * has in flight commands, its cwq->flush_color is set to
2540 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2541 * wakeup logic is armed and %true is returned.
2543 * The caller should have initialized @wq->first_flusher prior to
2544 * calling this function with non-negative @flush_color. If
2545 * @flush_color is negative, no flush color update is done and %false
2548 * If @work_color is non-negative, all cwqs should have the same
2549 * work_color which is previous to @work_color and all will be
2550 * advanced to @work_color.
2553 * mutex_lock(wq->flush_mutex).
2556 * %true if @flush_color >= 0 and there's something to flush. %false
2559 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2560 int flush_color, int work_color)
2565 if (flush_color >= 0) {
2566 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2567 atomic_set(&wq->nr_cwqs_to_flush, 1);
2570 for_each_cwq_cpu(cpu, wq) {
2571 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2572 struct global_cwq *gcwq = cwq->pool->gcwq;
2574 spin_lock_irq(&gcwq->lock);
2576 if (flush_color >= 0) {
2577 BUG_ON(cwq->flush_color != -1);
2579 if (cwq->nr_in_flight[flush_color]) {
2580 cwq->flush_color = flush_color;
2581 atomic_inc(&wq->nr_cwqs_to_flush);
2586 if (work_color >= 0) {
2587 BUG_ON(work_color != work_next_color(cwq->work_color));
2588 cwq->work_color = work_color;
2591 spin_unlock_irq(&gcwq->lock);
2594 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2595 complete(&wq->first_flusher->done);
2601 * flush_workqueue - ensure that any scheduled work has run to completion.
2602 * @wq: workqueue to flush
2604 * Forces execution of the workqueue and blocks until its completion.
2605 * This is typically used in driver shutdown handlers.
2607 * We sleep until all works which were queued on entry have been handled,
2608 * but we are not livelocked by new incoming ones.
2610 void flush_workqueue(struct workqueue_struct *wq)
2612 struct wq_flusher this_flusher = {
2613 .list = LIST_HEAD_INIT(this_flusher.list),
2615 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2619 lock_map_acquire(&wq->lockdep_map);
2620 lock_map_release(&wq->lockdep_map);
2622 mutex_lock(&wq->flush_mutex);
2625 * Start-to-wait phase
2627 next_color = work_next_color(wq->work_color);
2629 if (next_color != wq->flush_color) {
2631 * Color space is not full. The current work_color
2632 * becomes our flush_color and work_color is advanced
2635 BUG_ON(!list_empty(&wq->flusher_overflow));
2636 this_flusher.flush_color = wq->work_color;
2637 wq->work_color = next_color;
2639 if (!wq->first_flusher) {
2640 /* no flush in progress, become the first flusher */
2641 BUG_ON(wq->flush_color != this_flusher.flush_color);
2643 wq->first_flusher = &this_flusher;
2645 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2647 /* nothing to flush, done */
2648 wq->flush_color = next_color;
2649 wq->first_flusher = NULL;
2654 BUG_ON(wq->flush_color == this_flusher.flush_color);
2655 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2656 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2660 * Oops, color space is full, wait on overflow queue.
2661 * The next flush completion will assign us
2662 * flush_color and transfer to flusher_queue.
2664 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2667 mutex_unlock(&wq->flush_mutex);
2669 wait_for_completion(&this_flusher.done);
2672 * Wake-up-and-cascade phase
2674 * First flushers are responsible for cascading flushes and
2675 * handling overflow. Non-first flushers can simply return.
2677 if (wq->first_flusher != &this_flusher)
2680 mutex_lock(&wq->flush_mutex);
2682 /* we might have raced, check again with mutex held */
2683 if (wq->first_flusher != &this_flusher)
2686 wq->first_flusher = NULL;
2688 BUG_ON(!list_empty(&this_flusher.list));
2689 BUG_ON(wq->flush_color != this_flusher.flush_color);
2692 struct wq_flusher *next, *tmp;
2694 /* complete all the flushers sharing the current flush color */
2695 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2696 if (next->flush_color != wq->flush_color)
2698 list_del_init(&next->list);
2699 complete(&next->done);
2702 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2703 wq->flush_color != work_next_color(wq->work_color));
2705 /* this flush_color is finished, advance by one */
2706 wq->flush_color = work_next_color(wq->flush_color);
2708 /* one color has been freed, handle overflow queue */
2709 if (!list_empty(&wq->flusher_overflow)) {
2711 * Assign the same color to all overflowed
2712 * flushers, advance work_color and append to
2713 * flusher_queue. This is the start-to-wait
2714 * phase for these overflowed flushers.
2716 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2717 tmp->flush_color = wq->work_color;
2719 wq->work_color = work_next_color(wq->work_color);
2721 list_splice_tail_init(&wq->flusher_overflow,
2722 &wq->flusher_queue);
2723 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2726 if (list_empty(&wq->flusher_queue)) {
2727 BUG_ON(wq->flush_color != wq->work_color);
2732 * Need to flush more colors. Make the next flusher
2733 * the new first flusher and arm cwqs.
2735 BUG_ON(wq->flush_color == wq->work_color);
2736 BUG_ON(wq->flush_color != next->flush_color);
2738 list_del_init(&next->list);
2739 wq->first_flusher = next;
2741 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2745 * Meh... this color is already done, clear first
2746 * flusher and repeat cascading.
2748 wq->first_flusher = NULL;
2752 mutex_unlock(&wq->flush_mutex);
2754 EXPORT_SYMBOL_GPL(flush_workqueue);
2757 * drain_workqueue - drain a workqueue
2758 * @wq: workqueue to drain
2760 * Wait until the workqueue becomes empty. While draining is in progress,
2761 * only chain queueing is allowed. IOW, only currently pending or running
2762 * work items on @wq can queue further work items on it. @wq is flushed
2763 * repeatedly until it becomes empty. The number of flushing is detemined
2764 * by the depth of chaining and should be relatively short. Whine if it
2767 void drain_workqueue(struct workqueue_struct *wq)
2769 unsigned int flush_cnt = 0;
2773 * __queue_work() needs to test whether there are drainers, is much
2774 * hotter than drain_workqueue() and already looks at @wq->flags.
2775 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2777 spin_lock(&workqueue_lock);
2778 if (!wq->nr_drainers++)
2779 wq->flags |= WQ_DRAINING;
2780 spin_unlock(&workqueue_lock);
2782 flush_workqueue(wq);
2784 for_each_cwq_cpu(cpu, wq) {
2785 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2788 spin_lock_irq(&cwq->pool->gcwq->lock);
2789 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2790 spin_unlock_irq(&cwq->pool->gcwq->lock);
2795 if (++flush_cnt == 10 ||
2796 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2797 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2798 wq->name, flush_cnt);
2802 spin_lock(&workqueue_lock);
2803 if (!--wq->nr_drainers)
2804 wq->flags &= ~WQ_DRAINING;
2805 spin_unlock(&workqueue_lock);
2807 EXPORT_SYMBOL_GPL(drain_workqueue);
2809 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2811 struct worker *worker = NULL;
2812 struct global_cwq *gcwq;
2813 struct cpu_workqueue_struct *cwq;
2816 gcwq = get_work_gcwq(work);
2820 spin_lock_irq(&gcwq->lock);
2821 if (!list_empty(&work->entry)) {
2823 * See the comment near try_to_grab_pending()->smp_rmb().
2824 * If it was re-queued to a different gcwq under us, we
2825 * are not going to wait.
2828 cwq = get_work_cwq(work);
2829 if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
2832 worker = find_worker_executing_work(gcwq, work);
2835 cwq = worker->current_cwq;
2838 insert_wq_barrier(cwq, barr, work, worker);
2839 spin_unlock_irq(&gcwq->lock);
2842 * If @max_active is 1 or rescuer is in use, flushing another work
2843 * item on the same workqueue may lead to deadlock. Make sure the
2844 * flusher is not running on the same workqueue by verifying write
2847 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2848 lock_map_acquire(&cwq->wq->lockdep_map);
2850 lock_map_acquire_read(&cwq->wq->lockdep_map);
2851 lock_map_release(&cwq->wq->lockdep_map);
2855 spin_unlock_irq(&gcwq->lock);
2860 * flush_work - wait for a work to finish executing the last queueing instance
2861 * @work: the work to flush
2863 * Wait until @work has finished execution. @work is guaranteed to be idle
2864 * on return if it hasn't been requeued since flush started.
2867 * %true if flush_work() waited for the work to finish execution,
2868 * %false if it was already idle.
2870 bool flush_work(struct work_struct *work)
2872 struct wq_barrier barr;
2874 lock_map_acquire(&work->lockdep_map);
2875 lock_map_release(&work->lockdep_map);
2877 if (start_flush_work(work, &barr)) {
2878 wait_for_completion(&barr.done);
2879 destroy_work_on_stack(&barr.work);
2885 EXPORT_SYMBOL_GPL(flush_work);
2887 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2889 unsigned long flags;
2893 ret = try_to_grab_pending(work, is_dwork, &flags);
2895 * If someone else is canceling, wait for the same event it
2896 * would be waiting for before retrying.
2898 if (unlikely(ret == -ENOENT))
2900 } while (unlikely(ret < 0));
2902 /* tell other tasks trying to grab @work to back off */
2903 mark_work_canceling(work);
2904 local_irq_restore(flags);
2907 clear_work_data(work);
2912 * cancel_work_sync - cancel a work and wait for it to finish
2913 * @work: the work to cancel
2915 * Cancel @work and wait for its execution to finish. This function
2916 * can be used even if the work re-queues itself or migrates to
2917 * another workqueue. On return from this function, @work is
2918 * guaranteed to be not pending or executing on any CPU.
2920 * cancel_work_sync(&delayed_work->work) must not be used for
2921 * delayed_work's. Use cancel_delayed_work_sync() instead.
2923 * The caller must ensure that the workqueue on which @work was last
2924 * queued can't be destroyed before this function returns.
2927 * %true if @work was pending, %false otherwise.
2929 bool cancel_work_sync(struct work_struct *work)
2931 return __cancel_work_timer(work, false);
2933 EXPORT_SYMBOL_GPL(cancel_work_sync);
2936 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2937 * @dwork: the delayed work to flush
2939 * Delayed timer is cancelled and the pending work is queued for
2940 * immediate execution. Like flush_work(), this function only
2941 * considers the last queueing instance of @dwork.
2944 * %true if flush_work() waited for the work to finish execution,
2945 * %false if it was already idle.
2947 bool flush_delayed_work(struct delayed_work *dwork)
2949 local_irq_disable();
2950 if (del_timer_sync(&dwork->timer))
2951 __queue_work(dwork->cpu,
2952 get_work_cwq(&dwork->work)->wq, &dwork->work);
2954 return flush_work(&dwork->work);
2956 EXPORT_SYMBOL(flush_delayed_work);
2959 * cancel_delayed_work - cancel a delayed work
2960 * @dwork: delayed_work to cancel
2962 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2963 * and canceled; %false if wasn't pending. Note that the work callback
2964 * function may still be running on return, unless it returns %true and the
2965 * work doesn't re-arm itself. Explicitly flush or use
2966 * cancel_delayed_work_sync() to wait on it.
2968 * This function is safe to call from any context including IRQ handler.
2970 bool cancel_delayed_work(struct delayed_work *dwork)
2972 unsigned long flags;
2976 ret = try_to_grab_pending(&dwork->work, true, &flags);
2977 } while (unlikely(ret == -EAGAIN));
2979 if (unlikely(ret < 0))
2982 set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work));
2983 local_irq_restore(flags);
2986 EXPORT_SYMBOL(cancel_delayed_work);
2989 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2990 * @dwork: the delayed work cancel
2992 * This is cancel_work_sync() for delayed works.
2995 * %true if @dwork was pending, %false otherwise.
2997 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2999 return __cancel_work_timer(&dwork->work, true);
3001 EXPORT_SYMBOL(cancel_delayed_work_sync);
3004 * schedule_work_on - put work task on a specific cpu
3005 * @cpu: cpu to put the work task on
3006 * @work: job to be done
3008 * This puts a job on a specific cpu
3010 bool schedule_work_on(int cpu, struct work_struct *work)
3012 return queue_work_on(cpu, system_wq, work);
3014 EXPORT_SYMBOL(schedule_work_on);
3017 * schedule_work - put work task in global workqueue
3018 * @work: job to be done
3020 * Returns %false if @work was already on the kernel-global workqueue and
3023 * This puts a job in the kernel-global workqueue if it was not already
3024 * queued and leaves it in the same position on the kernel-global
3025 * workqueue otherwise.
3027 bool schedule_work(struct work_struct *work)
3029 return queue_work(system_wq, work);
3031 EXPORT_SYMBOL(schedule_work);
3034 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3036 * @dwork: job to be done
3037 * @delay: number of jiffies to wait
3039 * After waiting for a given time this puts a job in the kernel-global
3040 * workqueue on the specified CPU.
3042 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3043 unsigned long delay)
3045 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3047 EXPORT_SYMBOL(schedule_delayed_work_on);
3050 * schedule_delayed_work - put work task in global workqueue after delay
3051 * @dwork: job to be done
3052 * @delay: number of jiffies to wait or 0 for immediate execution
3054 * After waiting for a given time this puts a job in the kernel-global
3057 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3059 return queue_delayed_work(system_wq, dwork, delay);
3061 EXPORT_SYMBOL(schedule_delayed_work);
3064 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3065 * @func: the function to call
3067 * schedule_on_each_cpu() executes @func on each online CPU using the
3068 * system workqueue and blocks until all CPUs have completed.
3069 * schedule_on_each_cpu() is very slow.
3072 * 0 on success, -errno on failure.
3074 int schedule_on_each_cpu(work_func_t func)
3077 struct work_struct __percpu *works;
3079 works = alloc_percpu(struct work_struct);
3085 for_each_online_cpu(cpu) {
3086 struct work_struct *work = per_cpu_ptr(works, cpu);
3088 INIT_WORK(work, func);
3089 schedule_work_on(cpu, work);
3092 for_each_online_cpu(cpu)
3093 flush_work(per_cpu_ptr(works, cpu));
3101 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3103 * Forces execution of the kernel-global workqueue and blocks until its
3106 * Think twice before calling this function! It's very easy to get into
3107 * trouble if you don't take great care. Either of the following situations
3108 * will lead to deadlock:
3110 * One of the work items currently on the workqueue needs to acquire
3111 * a lock held by your code or its caller.
3113 * Your code is running in the context of a work routine.
3115 * They will be detected by lockdep when they occur, but the first might not
3116 * occur very often. It depends on what work items are on the workqueue and
3117 * what locks they need, which you have no control over.
3119 * In most situations flushing the entire workqueue is overkill; you merely
3120 * need to know that a particular work item isn't queued and isn't running.
3121 * In such cases you should use cancel_delayed_work_sync() or
3122 * cancel_work_sync() instead.
3124 void flush_scheduled_work(void)
3126 flush_workqueue(system_wq);
3128 EXPORT_SYMBOL(flush_scheduled_work);
3131 * execute_in_process_context - reliably execute the routine with user context
3132 * @fn: the function to execute
3133 * @ew: guaranteed storage for the execute work structure (must
3134 * be available when the work executes)
3136 * Executes the function immediately if process context is available,
3137 * otherwise schedules the function for delayed execution.
3139 * Returns: 0 - function was executed
3140 * 1 - function was scheduled for execution
3142 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3144 if (!in_interrupt()) {
3149 INIT_WORK(&ew->work, fn);
3150 schedule_work(&ew->work);
3154 EXPORT_SYMBOL_GPL(execute_in_process_context);
3156 int keventd_up(void)
3158 return system_wq != NULL;
3161 static int alloc_cwqs(struct workqueue_struct *wq)
3164 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3165 * Make sure that the alignment isn't lower than that of
3166 * unsigned long long.
3168 const size_t size = sizeof(struct cpu_workqueue_struct);
3169 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3170 __alignof__(unsigned long long));
3172 if (!(wq->flags & WQ_UNBOUND))
3173 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3178 * Allocate enough room to align cwq and put an extra
3179 * pointer at the end pointing back to the originally
3180 * allocated pointer which will be used for free.
3182 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3184 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3185 *(void **)(wq->cpu_wq.single + 1) = ptr;
3189 /* just in case, make sure it's actually aligned */
3190 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3191 return wq->cpu_wq.v ? 0 : -ENOMEM;
3194 static void free_cwqs(struct workqueue_struct *wq)
3196 if (!(wq->flags & WQ_UNBOUND))
3197 free_percpu(wq->cpu_wq.pcpu);
3198 else if (wq->cpu_wq.single) {
3199 /* the pointer to free is stored right after the cwq */
3200 kfree(*(void **)(wq->cpu_wq.single + 1));
3204 static int wq_clamp_max_active(int max_active, unsigned int flags,
3207 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3209 if (max_active < 1 || max_active > lim)
3210 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3211 max_active, name, 1, lim);
3213 return clamp_val(max_active, 1, lim);
3216 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3219 struct lock_class_key *key,
3220 const char *lock_name, ...)
3222 va_list args, args1;
3223 struct workqueue_struct *wq;
3227 /* determine namelen, allocate wq and format name */
3228 va_start(args, lock_name);
3229 va_copy(args1, args);
3230 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3232 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3236 vsnprintf(wq->name, namelen, fmt, args1);
3241 * Workqueues which may be used during memory reclaim should
3242 * have a rescuer to guarantee forward progress.
3244 if (flags & WQ_MEM_RECLAIM)
3245 flags |= WQ_RESCUER;
3247 max_active = max_active ?: WQ_DFL_ACTIVE;
3248 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3252 wq->saved_max_active = max_active;
3253 mutex_init(&wq->flush_mutex);
3254 atomic_set(&wq->nr_cwqs_to_flush, 0);
3255 INIT_LIST_HEAD(&wq->flusher_queue);
3256 INIT_LIST_HEAD(&wq->flusher_overflow);
3258 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3259 INIT_LIST_HEAD(&wq->list);
3261 if (alloc_cwqs(wq) < 0)
3264 for_each_cwq_cpu(cpu, wq) {
3265 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3266 struct global_cwq *gcwq = get_gcwq(cpu);
3267 int pool_idx = (bool)(flags & WQ_HIGHPRI);
3269 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3270 cwq->pool = &gcwq->pools[pool_idx];
3272 cwq->flush_color = -1;
3273 cwq->max_active = max_active;
3274 INIT_LIST_HEAD(&cwq->delayed_works);
3277 if (flags & WQ_RESCUER) {
3278 struct worker *rescuer;
3280 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3283 wq->rescuer = rescuer = alloc_worker();
3287 rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3289 if (IS_ERR(rescuer->task))
3292 rescuer->task->flags |= PF_THREAD_BOUND;
3293 wake_up_process(rescuer->task);
3297 * workqueue_lock protects global freeze state and workqueues
3298 * list. Grab it, set max_active accordingly and add the new
3299 * workqueue to workqueues list.
3301 spin_lock(&workqueue_lock);
3303 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3304 for_each_cwq_cpu(cpu, wq)
3305 get_cwq(cpu, wq)->max_active = 0;
3307 list_add(&wq->list, &workqueues);
3309 spin_unlock(&workqueue_lock);
3315 free_mayday_mask(wq->mayday_mask);
3321 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3324 * destroy_workqueue - safely terminate a workqueue
3325 * @wq: target workqueue
3327 * Safely destroy a workqueue. All work currently pending will be done first.
3329 void destroy_workqueue(struct workqueue_struct *wq)
3333 /* drain it before proceeding with destruction */
3334 drain_workqueue(wq);
3337 * wq list is used to freeze wq, remove from list after
3338 * flushing is complete in case freeze races us.
3340 spin_lock(&workqueue_lock);
3341 list_del(&wq->list);
3342 spin_unlock(&workqueue_lock);
3345 for_each_cwq_cpu(cpu, wq) {
3346 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3349 for (i = 0; i < WORK_NR_COLORS; i++)
3350 BUG_ON(cwq->nr_in_flight[i]);
3351 BUG_ON(cwq->nr_active);
3352 BUG_ON(!list_empty(&cwq->delayed_works));
3355 if (wq->flags & WQ_RESCUER) {
3356 kthread_stop(wq->rescuer->task);
3357 free_mayday_mask(wq->mayday_mask);
3364 EXPORT_SYMBOL_GPL(destroy_workqueue);
3367 * workqueue_set_max_active - adjust max_active of a workqueue
3368 * @wq: target workqueue
3369 * @max_active: new max_active value.
3371 * Set max_active of @wq to @max_active.
3374 * Don't call from IRQ context.
3376 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3380 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3382 spin_lock(&workqueue_lock);
3384 wq->saved_max_active = max_active;
3386 for_each_cwq_cpu(cpu, wq) {
3387 struct global_cwq *gcwq = get_gcwq(cpu);
3389 spin_lock_irq(&gcwq->lock);
3391 if (!(wq->flags & WQ_FREEZABLE) ||
3392 !(gcwq->flags & GCWQ_FREEZING))
3393 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3395 spin_unlock_irq(&gcwq->lock);
3398 spin_unlock(&workqueue_lock);
3400 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3403 * workqueue_congested - test whether a workqueue is congested
3404 * @cpu: CPU in question
3405 * @wq: target workqueue
3407 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3408 * no synchronization around this function and the test result is
3409 * unreliable and only useful as advisory hints or for debugging.
3412 * %true if congested, %false otherwise.
3414 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3416 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3418 return !list_empty(&cwq->delayed_works);
3420 EXPORT_SYMBOL_GPL(workqueue_congested);
3423 * work_cpu - return the last known associated cpu for @work
3424 * @work: the work of interest
3427 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3429 unsigned int work_cpu(struct work_struct *work)
3431 struct global_cwq *gcwq = get_work_gcwq(work);
3433 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3435 EXPORT_SYMBOL_GPL(work_cpu);
3438 * work_busy - test whether a work is currently pending or running
3439 * @work: the work to be tested
3441 * Test whether @work is currently pending or running. There is no
3442 * synchronization around this function and the test result is
3443 * unreliable and only useful as advisory hints or for debugging.
3444 * Especially for reentrant wqs, the pending state might hide the
3448 * OR'd bitmask of WORK_BUSY_* bits.
3450 unsigned int work_busy(struct work_struct *work)
3452 struct global_cwq *gcwq = get_work_gcwq(work);
3453 unsigned long flags;
3454 unsigned int ret = 0;
3459 spin_lock_irqsave(&gcwq->lock, flags);
3461 if (work_pending(work))
3462 ret |= WORK_BUSY_PENDING;
3463 if (find_worker_executing_work(gcwq, work))
3464 ret |= WORK_BUSY_RUNNING;
3466 spin_unlock_irqrestore(&gcwq->lock, flags);
3470 EXPORT_SYMBOL_GPL(work_busy);
3475 * There are two challenges in supporting CPU hotplug. Firstly, there
3476 * are a lot of assumptions on strong associations among work, cwq and
3477 * gcwq which make migrating pending and scheduled works very
3478 * difficult to implement without impacting hot paths. Secondly,
3479 * gcwqs serve mix of short, long and very long running works making
3480 * blocked draining impractical.
3482 * This is solved by allowing a gcwq to be disassociated from the CPU
3483 * running as an unbound one and allowing it to be reattached later if the
3484 * cpu comes back online.
3487 /* claim manager positions of all pools */
3488 static void gcwq_claim_management_and_lock(struct global_cwq *gcwq)
3490 struct worker_pool *pool;
3492 for_each_worker_pool(pool, gcwq)
3493 mutex_lock_nested(&pool->manager_mutex, pool - gcwq->pools);
3494 spin_lock_irq(&gcwq->lock);
3497 /* release manager positions */
3498 static void gcwq_release_management_and_unlock(struct global_cwq *gcwq)
3500 struct worker_pool *pool;
3502 spin_unlock_irq(&gcwq->lock);
3503 for_each_worker_pool(pool, gcwq)
3504 mutex_unlock(&pool->manager_mutex);
3507 static void gcwq_unbind_fn(struct work_struct *work)
3509 struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3510 struct worker_pool *pool;
3511 struct worker *worker;
3512 struct hlist_node *pos;
3515 BUG_ON(gcwq->cpu != smp_processor_id());
3517 gcwq_claim_management_and_lock(gcwq);
3520 * We've claimed all manager positions. Make all workers unbound
3521 * and set DISASSOCIATED. Before this, all workers except for the
3522 * ones which are still executing works from before the last CPU
3523 * down must be on the cpu. After this, they may become diasporas.
3525 for_each_worker_pool(pool, gcwq)
3526 list_for_each_entry(worker, &pool->idle_list, entry)
3527 worker->flags |= WORKER_UNBOUND;
3529 for_each_busy_worker(worker, i, pos, gcwq)
3530 worker->flags |= WORKER_UNBOUND;
3532 gcwq->flags |= GCWQ_DISASSOCIATED;
3534 gcwq_release_management_and_unlock(gcwq);
3537 * Call schedule() so that we cross rq->lock and thus can guarantee
3538 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3539 * as scheduler callbacks may be invoked from other cpus.
3544 * Sched callbacks are disabled now. Zap nr_running. After this,
3545 * nr_running stays zero and need_more_worker() and keep_working()
3546 * are always true as long as the worklist is not empty. @gcwq now
3547 * behaves as unbound (in terms of concurrency management) gcwq
3548 * which is served by workers tied to the CPU.
3550 * On return from this function, the current worker would trigger
3551 * unbound chain execution of pending work items if other workers
3554 for_each_worker_pool(pool, gcwq)
3555 atomic_set(get_pool_nr_running(pool), 0);
3559 * Workqueues should be brought up before normal priority CPU notifiers.
3560 * This will be registered high priority CPU notifier.
3562 static int __devinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3563 unsigned long action,
3566 unsigned int cpu = (unsigned long)hcpu;
3567 struct global_cwq *gcwq = get_gcwq(cpu);
3568 struct worker_pool *pool;
3570 switch (action & ~CPU_TASKS_FROZEN) {
3571 case CPU_UP_PREPARE:
3572 for_each_worker_pool(pool, gcwq) {
3573 struct worker *worker;
3575 if (pool->nr_workers)
3578 worker = create_worker(pool);
3582 spin_lock_irq(&gcwq->lock);
3583 start_worker(worker);
3584 spin_unlock_irq(&gcwq->lock);
3588 case CPU_DOWN_FAILED:
3590 gcwq_claim_management_and_lock(gcwq);
3591 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3592 rebind_workers(gcwq);
3593 gcwq_release_management_and_unlock(gcwq);
3600 * Workqueues should be brought down after normal priority CPU notifiers.
3601 * This will be registered as low priority CPU notifier.
3603 static int __devinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3604 unsigned long action,
3607 unsigned int cpu = (unsigned long)hcpu;
3608 struct work_struct unbind_work;
3610 switch (action & ~CPU_TASKS_FROZEN) {
3611 case CPU_DOWN_PREPARE:
3612 /* unbinding should happen on the local CPU */
3613 INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3614 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3615 flush_work(&unbind_work);
3623 struct work_for_cpu {
3624 struct completion completion;
3630 static int do_work_for_cpu(void *_wfc)
3632 struct work_for_cpu *wfc = _wfc;
3633 wfc->ret = wfc->fn(wfc->arg);
3634 complete(&wfc->completion);
3639 * work_on_cpu - run a function in user context on a particular cpu
3640 * @cpu: the cpu to run on
3641 * @fn: the function to run
3642 * @arg: the function arg
3644 * This will return the value @fn returns.
3645 * It is up to the caller to ensure that the cpu doesn't go offline.
3646 * The caller must not hold any locks which would prevent @fn from completing.
3648 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3650 struct task_struct *sub_thread;
3651 struct work_for_cpu wfc = {
3652 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3657 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3658 if (IS_ERR(sub_thread))
3659 return PTR_ERR(sub_thread);
3660 kthread_bind(sub_thread, cpu);
3661 wake_up_process(sub_thread);
3662 wait_for_completion(&wfc.completion);
3665 EXPORT_SYMBOL_GPL(work_on_cpu);
3666 #endif /* CONFIG_SMP */
3668 #ifdef CONFIG_FREEZER
3671 * freeze_workqueues_begin - begin freezing workqueues
3673 * Start freezing workqueues. After this function returns, all freezable
3674 * workqueues will queue new works to their frozen_works list instead of
3678 * Grabs and releases workqueue_lock and gcwq->lock's.
3680 void freeze_workqueues_begin(void)
3684 spin_lock(&workqueue_lock);
3686 BUG_ON(workqueue_freezing);
3687 workqueue_freezing = true;
3689 for_each_gcwq_cpu(cpu) {
3690 struct global_cwq *gcwq = get_gcwq(cpu);
3691 struct workqueue_struct *wq;
3693 spin_lock_irq(&gcwq->lock);
3695 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3696 gcwq->flags |= GCWQ_FREEZING;
3698 list_for_each_entry(wq, &workqueues, list) {
3699 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3701 if (cwq && wq->flags & WQ_FREEZABLE)
3702 cwq->max_active = 0;
3705 spin_unlock_irq(&gcwq->lock);
3708 spin_unlock(&workqueue_lock);
3712 * freeze_workqueues_busy - are freezable workqueues still busy?
3714 * Check whether freezing is complete. This function must be called
3715 * between freeze_workqueues_begin() and thaw_workqueues().
3718 * Grabs and releases workqueue_lock.
3721 * %true if some freezable workqueues are still busy. %false if freezing
3724 bool freeze_workqueues_busy(void)
3729 spin_lock(&workqueue_lock);
3731 BUG_ON(!workqueue_freezing);
3733 for_each_gcwq_cpu(cpu) {
3734 struct workqueue_struct *wq;
3736 * nr_active is monotonically decreasing. It's safe
3737 * to peek without lock.
3739 list_for_each_entry(wq, &workqueues, list) {
3740 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3742 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3745 BUG_ON(cwq->nr_active < 0);
3746 if (cwq->nr_active) {
3753 spin_unlock(&workqueue_lock);
3758 * thaw_workqueues - thaw workqueues
3760 * Thaw workqueues. Normal queueing is restored and all collected
3761 * frozen works are transferred to their respective gcwq worklists.
3764 * Grabs and releases workqueue_lock and gcwq->lock's.
3766 void thaw_workqueues(void)
3770 spin_lock(&workqueue_lock);
3772 if (!workqueue_freezing)
3775 for_each_gcwq_cpu(cpu) {
3776 struct global_cwq *gcwq = get_gcwq(cpu);
3777 struct worker_pool *pool;
3778 struct workqueue_struct *wq;
3780 spin_lock_irq(&gcwq->lock);
3782 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3783 gcwq->flags &= ~GCWQ_FREEZING;
3785 list_for_each_entry(wq, &workqueues, list) {
3786 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3788 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3791 /* restore max_active and repopulate worklist */
3792 cwq->max_active = wq->saved_max_active;
3794 while (!list_empty(&cwq->delayed_works) &&
3795 cwq->nr_active < cwq->max_active)
3796 cwq_activate_first_delayed(cwq);
3799 for_each_worker_pool(pool, gcwq)
3800 wake_up_worker(pool);
3802 spin_unlock_irq(&gcwq->lock);
3805 workqueue_freezing = false;
3807 spin_unlock(&workqueue_lock);
3809 #endif /* CONFIG_FREEZER */
3811 static int __init init_workqueues(void)
3816 /* make sure we have enough bits for OFFQ CPU number */
3817 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
3820 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3821 cpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3823 /* initialize gcwqs */
3824 for_each_gcwq_cpu(cpu) {
3825 struct global_cwq *gcwq = get_gcwq(cpu);
3826 struct worker_pool *pool;
3828 spin_lock_init(&gcwq->lock);
3830 gcwq->flags |= GCWQ_DISASSOCIATED;
3832 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3833 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3835 for_each_worker_pool(pool, gcwq) {
3837 INIT_LIST_HEAD(&pool->worklist);
3838 INIT_LIST_HEAD(&pool->idle_list);
3840 init_timer_deferrable(&pool->idle_timer);
3841 pool->idle_timer.function = idle_worker_timeout;
3842 pool->idle_timer.data = (unsigned long)pool;
3844 setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3845 (unsigned long)pool);
3847 mutex_init(&pool->manager_mutex);
3848 ida_init(&pool->worker_ida);
3852 /* create the initial worker */
3853 for_each_online_gcwq_cpu(cpu) {
3854 struct global_cwq *gcwq = get_gcwq(cpu);
3855 struct worker_pool *pool;
3857 if (cpu != WORK_CPU_UNBOUND)
3858 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3860 for_each_worker_pool(pool, gcwq) {
3861 struct worker *worker;
3863 worker = create_worker(pool);
3865 spin_lock_irq(&gcwq->lock);
3866 start_worker(worker);
3867 spin_unlock_irq(&gcwq->lock);
3871 system_wq = alloc_workqueue("events", 0, 0);
3872 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3873 system_long_wq = alloc_workqueue("events_long", 0, 0);
3874 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3875 WQ_UNBOUND_MAX_ACTIVE);
3876 system_freezable_wq = alloc_workqueue("events_freezable",
3878 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3879 !system_unbound_wq || !system_freezable_wq);
3882 early_initcall(init_workqueues);