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 */
71 WORKER_STARTED = 1 << 0, /* started */
72 WORKER_DIE = 1 << 1, /* die die die */
73 WORKER_IDLE = 1 << 2, /* is idle */
74 WORKER_PREP = 1 << 3, /* preparing to run works */
75 WORKER_REBIND = 1 << 5, /* mom is home, come back */
76 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
77 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_REBIND | WORKER_UNBOUND |
82 NR_WORKER_POOLS = 2, /* # worker pools per gcwq */
84 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
85 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
86 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
88 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
89 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
92 /* call for help after 10ms
94 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
95 CREATE_COOLDOWN = HZ, /* time to breath after fail */
98 * Rescue workers are used only on emergencies and shared by
101 RESCUER_NICE_LEVEL = -20,
102 HIGHPRI_NICE_LEVEL = -20,
106 * Structure fields follow one of the following exclusion rules.
108 * I: Modifiable by initialization/destruction paths and read-only for
111 * P: Preemption protected. Disabling preemption is enough and should
112 * only be modified and accessed from the local cpu.
114 * L: gcwq->lock protected. Access with gcwq->lock held.
116 * X: During normal operation, modification requires gcwq->lock and
117 * should be done only from local cpu. Either disabling preemption
118 * on local cpu or grabbing gcwq->lock is enough for read access.
119 * If GCWQ_DISASSOCIATED is set, it's identical to L.
121 * F: wq->flush_mutex protected.
123 * W: workqueue_lock protected.
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 idle_rebind *idle_rebind; /* L: for idle worker */
153 struct work_struct rebind_work; /* L: for busy worker */
157 struct global_cwq *gcwq; /* I: the owning gcwq */
158 unsigned int flags; /* X: flags */
160 struct list_head worklist; /* L: list of pending works */
161 int nr_workers; /* L: total number of workers */
162 int nr_idle; /* L: currently idle ones */
164 struct list_head idle_list; /* X: list of idle workers */
165 struct timer_list idle_timer; /* L: worker idle timeout */
166 struct timer_list mayday_timer; /* L: SOS timer for workers */
168 struct mutex manager_mutex; /* mutex manager should hold */
169 struct ida worker_ida; /* L: for worker IDs */
173 * Global per-cpu workqueue. There's one and only one for each cpu
174 * and all works are queued and processed here regardless of their
178 spinlock_t lock; /* the gcwq lock */
179 unsigned int cpu; /* I: the associated cpu */
180 unsigned int flags; /* L: GCWQ_* flags */
182 /* workers are chained either in busy_hash or pool idle_list */
183 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
184 /* L: hash of busy workers */
186 struct worker_pool pools[2]; /* normal and highpri pools */
188 wait_queue_head_t rebind_hold; /* rebind hold wait */
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 struct workqueue_struct *system_long_wq __read_mostly;
272 struct workqueue_struct *system_nrt_wq __read_mostly;
273 struct workqueue_struct *system_unbound_wq __read_mostly;
274 struct workqueue_struct *system_freezable_wq __read_mostly;
275 struct workqueue_struct *system_nrt_freezable_wq __read_mostly;
276 EXPORT_SYMBOL_GPL(system_wq);
277 EXPORT_SYMBOL_GPL(system_long_wq);
278 EXPORT_SYMBOL_GPL(system_nrt_wq);
279 EXPORT_SYMBOL_GPL(system_unbound_wq);
280 EXPORT_SYMBOL_GPL(system_freezable_wq);
281 EXPORT_SYMBOL_GPL(system_nrt_freezable_wq);
283 #define CREATE_TRACE_POINTS
284 #include <trace/events/workqueue.h>
286 #define for_each_worker_pool(pool, gcwq) \
287 for ((pool) = &(gcwq)->pools[0]; \
288 (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
290 #define for_each_busy_worker(worker, i, pos, gcwq) \
291 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
292 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
294 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
297 if (cpu < nr_cpu_ids) {
299 cpu = cpumask_next(cpu, mask);
300 if (cpu < nr_cpu_ids)
304 return WORK_CPU_UNBOUND;
306 return WORK_CPU_NONE;
309 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
310 struct workqueue_struct *wq)
312 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
318 * An extra gcwq is defined for an invalid cpu number
319 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
320 * specific CPU. The following iterators are similar to
321 * for_each_*_cpu() iterators but also considers the unbound gcwq.
323 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
324 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
325 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
326 * WORK_CPU_UNBOUND for unbound workqueues
328 #define for_each_gcwq_cpu(cpu) \
329 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
330 (cpu) < WORK_CPU_NONE; \
331 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
333 #define for_each_online_gcwq_cpu(cpu) \
334 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
335 (cpu) < WORK_CPU_NONE; \
336 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
338 #define for_each_cwq_cpu(cpu, wq) \
339 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
340 (cpu) < WORK_CPU_NONE; \
341 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
343 #ifdef CONFIG_DEBUG_OBJECTS_WORK
345 static struct debug_obj_descr work_debug_descr;
347 static void *work_debug_hint(void *addr)
349 return ((struct work_struct *) addr)->func;
353 * fixup_init is called when:
354 * - an active object is initialized
356 static int work_fixup_init(void *addr, enum debug_obj_state state)
358 struct work_struct *work = addr;
361 case ODEBUG_STATE_ACTIVE:
362 cancel_work_sync(work);
363 debug_object_init(work, &work_debug_descr);
371 * fixup_activate is called when:
372 * - an active object is activated
373 * - an unknown object is activated (might be a statically initialized object)
375 static int work_fixup_activate(void *addr, enum debug_obj_state state)
377 struct work_struct *work = addr;
381 case ODEBUG_STATE_NOTAVAILABLE:
383 * This is not really a fixup. The work struct was
384 * statically initialized. We just make sure that it
385 * is tracked in the object tracker.
387 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
388 debug_object_init(work, &work_debug_descr);
389 debug_object_activate(work, &work_debug_descr);
395 case ODEBUG_STATE_ACTIVE:
404 * fixup_free is called when:
405 * - an active object is freed
407 static int work_fixup_free(void *addr, enum debug_obj_state state)
409 struct work_struct *work = addr;
412 case ODEBUG_STATE_ACTIVE:
413 cancel_work_sync(work);
414 debug_object_free(work, &work_debug_descr);
421 static struct debug_obj_descr work_debug_descr = {
422 .name = "work_struct",
423 .debug_hint = work_debug_hint,
424 .fixup_init = work_fixup_init,
425 .fixup_activate = work_fixup_activate,
426 .fixup_free = work_fixup_free,
429 static inline void debug_work_activate(struct work_struct *work)
431 debug_object_activate(work, &work_debug_descr);
434 static inline void debug_work_deactivate(struct work_struct *work)
436 debug_object_deactivate(work, &work_debug_descr);
439 void __init_work(struct work_struct *work, int onstack)
442 debug_object_init_on_stack(work, &work_debug_descr);
444 debug_object_init(work, &work_debug_descr);
446 EXPORT_SYMBOL_GPL(__init_work);
448 void destroy_work_on_stack(struct work_struct *work)
450 debug_object_free(work, &work_debug_descr);
452 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
455 static inline void debug_work_activate(struct work_struct *work) { }
456 static inline void debug_work_deactivate(struct work_struct *work) { }
459 /* Serializes the accesses to the list of workqueues. */
460 static DEFINE_SPINLOCK(workqueue_lock);
461 static LIST_HEAD(workqueues);
462 static bool workqueue_freezing; /* W: have wqs started freezing? */
465 * The almighty global cpu workqueues. nr_running is the only field
466 * which is expected to be used frequently by other cpus via
467 * try_to_wake_up(). Put it in a separate cacheline.
469 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
470 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]);
473 * Global cpu workqueue and nr_running counter for unbound gcwq. The
474 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
475 * workers have WORKER_UNBOUND set.
477 static struct global_cwq unbound_global_cwq;
478 static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = {
479 [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
482 static int worker_thread(void *__worker);
484 static int worker_pool_pri(struct worker_pool *pool)
486 return pool - pool->gcwq->pools;
489 static struct global_cwq *get_gcwq(unsigned int cpu)
491 if (cpu != WORK_CPU_UNBOUND)
492 return &per_cpu(global_cwq, cpu);
494 return &unbound_global_cwq;
497 static atomic_t *get_pool_nr_running(struct worker_pool *pool)
499 int cpu = pool->gcwq->cpu;
500 int idx = worker_pool_pri(pool);
502 if (cpu != WORK_CPU_UNBOUND)
503 return &per_cpu(pool_nr_running, cpu)[idx];
505 return &unbound_pool_nr_running[idx];
508 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
509 struct workqueue_struct *wq)
511 if (!(wq->flags & WQ_UNBOUND)) {
512 if (likely(cpu < nr_cpu_ids))
513 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
514 } else if (likely(cpu == WORK_CPU_UNBOUND))
515 return wq->cpu_wq.single;
519 static unsigned int work_color_to_flags(int color)
521 return color << WORK_STRUCT_COLOR_SHIFT;
524 static int get_work_color(struct work_struct *work)
526 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
527 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
530 static int work_next_color(int color)
532 return (color + 1) % WORK_NR_COLORS;
536 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
537 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
538 * cleared and the work data contains the cpu number it was last on.
540 * set_work_cwq(), set_work_cpu_and_clear_pending() and clear_work_data()
541 * can be used to set the cwq, cpu or clear work->data. These functions
542 * should only be called while the work is owned - ie. while the PENDING
545 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
546 * corresponding to a work. gcwq is available once the work has been
547 * queued anywhere after initialization. cwq is available only from
548 * queueing until execution starts.
550 static inline void set_work_data(struct work_struct *work, unsigned long data,
553 BUG_ON(!work_pending(work));
554 atomic_long_set(&work->data, data | flags | work_static(work));
557 static void set_work_cwq(struct work_struct *work,
558 struct cpu_workqueue_struct *cwq,
559 unsigned long extra_flags)
561 set_work_data(work, (unsigned long)cwq,
562 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
565 static void set_work_cpu_and_clear_pending(struct work_struct *work,
568 set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, 0);
571 static void clear_work_data(struct work_struct *work)
573 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
576 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
578 unsigned long data = atomic_long_read(&work->data);
580 if (data & WORK_STRUCT_CWQ)
581 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
586 static struct global_cwq *get_work_gcwq(struct work_struct *work)
588 unsigned long data = atomic_long_read(&work->data);
591 if (data & WORK_STRUCT_CWQ)
592 return ((struct cpu_workqueue_struct *)
593 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq;
595 cpu = data >> WORK_STRUCT_FLAG_BITS;
596 if (cpu == WORK_CPU_NONE)
599 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
600 return get_gcwq(cpu);
604 * Policy functions. These define the policies on how the global worker
605 * pools are managed. Unless noted otherwise, these functions assume that
606 * they're being called with gcwq->lock held.
609 static bool __need_more_worker(struct worker_pool *pool)
611 return !atomic_read(get_pool_nr_running(pool));
615 * Need to wake up a worker? Called from anything but currently
618 * Note that, because unbound workers never contribute to nr_running, this
619 * function will always return %true for unbound gcwq as long as the
620 * worklist isn't empty.
622 static bool need_more_worker(struct worker_pool *pool)
624 return !list_empty(&pool->worklist) && __need_more_worker(pool);
627 /* Can I start working? Called from busy but !running workers. */
628 static bool may_start_working(struct worker_pool *pool)
630 return pool->nr_idle;
633 /* Do I need to keep working? Called from currently running workers. */
634 static bool keep_working(struct worker_pool *pool)
636 atomic_t *nr_running = get_pool_nr_running(pool);
638 return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
641 /* Do we need a new worker? Called from manager. */
642 static bool need_to_create_worker(struct worker_pool *pool)
644 return need_more_worker(pool) && !may_start_working(pool);
647 /* Do I need to be the manager? */
648 static bool need_to_manage_workers(struct worker_pool *pool)
650 return need_to_create_worker(pool) ||
651 (pool->flags & POOL_MANAGE_WORKERS);
654 /* Do we have too many workers and should some go away? */
655 static bool too_many_workers(struct worker_pool *pool)
657 bool managing = mutex_is_locked(&pool->manager_mutex);
658 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
659 int nr_busy = pool->nr_workers - nr_idle;
661 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
668 /* Return the first worker. Safe with preemption disabled */
669 static struct worker *first_worker(struct worker_pool *pool)
671 if (unlikely(list_empty(&pool->idle_list)))
674 return list_first_entry(&pool->idle_list, struct worker, entry);
678 * wake_up_worker - wake up an idle worker
679 * @pool: worker pool to wake worker from
681 * Wake up the first idle worker of @pool.
684 * spin_lock_irq(gcwq->lock).
686 static void wake_up_worker(struct worker_pool *pool)
688 struct worker *worker = first_worker(pool);
691 wake_up_process(worker->task);
695 * wq_worker_waking_up - a worker is waking up
696 * @task: task waking up
697 * @cpu: CPU @task is waking up to
699 * This function is called during try_to_wake_up() when a worker is
703 * spin_lock_irq(rq->lock)
705 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
707 struct worker *worker = kthread_data(task);
709 if (!(worker->flags & WORKER_NOT_RUNNING))
710 atomic_inc(get_pool_nr_running(worker->pool));
714 * wq_worker_sleeping - a worker is going to sleep
715 * @task: task going to sleep
716 * @cpu: CPU in question, must be the current CPU number
718 * This function is called during schedule() when a busy worker is
719 * going to sleep. Worker on the same cpu can be woken up by
720 * returning pointer to its task.
723 * spin_lock_irq(rq->lock)
726 * Worker task on @cpu to wake up, %NULL if none.
728 struct task_struct *wq_worker_sleeping(struct task_struct *task,
731 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
732 struct worker_pool *pool = worker->pool;
733 atomic_t *nr_running = get_pool_nr_running(pool);
735 if (worker->flags & WORKER_NOT_RUNNING)
738 /* this can only happen on the local cpu */
739 BUG_ON(cpu != raw_smp_processor_id());
742 * The counterpart of the following dec_and_test, implied mb,
743 * worklist not empty test sequence is in insert_work().
744 * Please read comment there.
746 * NOT_RUNNING is clear. This means that we're bound to and
747 * running on the local cpu w/ rq lock held and preemption
748 * disabled, which in turn means that none else could be
749 * manipulating idle_list, so dereferencing idle_list without gcwq
752 if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
753 to_wakeup = first_worker(pool);
754 return to_wakeup ? to_wakeup->task : NULL;
758 * worker_set_flags - set worker flags and adjust nr_running accordingly
760 * @flags: flags to set
761 * @wakeup: wakeup an idle worker if necessary
763 * Set @flags in @worker->flags and adjust nr_running accordingly. If
764 * nr_running becomes zero and @wakeup is %true, an idle worker is
768 * spin_lock_irq(gcwq->lock)
770 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
773 struct worker_pool *pool = worker->pool;
775 WARN_ON_ONCE(worker->task != current);
778 * If transitioning into NOT_RUNNING, adjust nr_running and
779 * wake up an idle worker as necessary if requested by
782 if ((flags & WORKER_NOT_RUNNING) &&
783 !(worker->flags & WORKER_NOT_RUNNING)) {
784 atomic_t *nr_running = get_pool_nr_running(pool);
787 if (atomic_dec_and_test(nr_running) &&
788 !list_empty(&pool->worklist))
789 wake_up_worker(pool);
791 atomic_dec(nr_running);
794 worker->flags |= flags;
798 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
800 * @flags: flags to clear
802 * Clear @flags in @worker->flags and adjust nr_running accordingly.
805 * spin_lock_irq(gcwq->lock)
807 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
809 struct worker_pool *pool = worker->pool;
810 unsigned int oflags = worker->flags;
812 WARN_ON_ONCE(worker->task != current);
814 worker->flags &= ~flags;
817 * If transitioning out of NOT_RUNNING, increment nr_running. Note
818 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
819 * of multiple flags, not a single flag.
821 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
822 if (!(worker->flags & WORKER_NOT_RUNNING))
823 atomic_inc(get_pool_nr_running(pool));
827 * busy_worker_head - return the busy hash head for a work
828 * @gcwq: gcwq of interest
829 * @work: work to be hashed
831 * Return hash head of @gcwq for @work.
834 * spin_lock_irq(gcwq->lock).
837 * Pointer to the hash head.
839 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
840 struct work_struct *work)
842 const int base_shift = ilog2(sizeof(struct work_struct));
843 unsigned long v = (unsigned long)work;
845 /* simple shift and fold hash, do we need something better? */
847 v += v >> BUSY_WORKER_HASH_ORDER;
848 v &= BUSY_WORKER_HASH_MASK;
850 return &gcwq->busy_hash[v];
854 * __find_worker_executing_work - find worker which is executing a work
855 * @gcwq: gcwq of interest
856 * @bwh: hash head as returned by busy_worker_head()
857 * @work: work to find worker for
859 * Find a worker which is executing @work on @gcwq. @bwh should be
860 * the hash head obtained by calling busy_worker_head() with the same
864 * spin_lock_irq(gcwq->lock).
867 * Pointer to worker which is executing @work if found, NULL
870 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
871 struct hlist_head *bwh,
872 struct work_struct *work)
874 struct worker *worker;
875 struct hlist_node *tmp;
877 hlist_for_each_entry(worker, tmp, bwh, hentry)
878 if (worker->current_work == work)
884 * find_worker_executing_work - find worker which is executing a work
885 * @gcwq: gcwq of interest
886 * @work: work to find worker for
888 * Find a worker which is executing @work on @gcwq. This function is
889 * identical to __find_worker_executing_work() except that this
890 * function calculates @bwh itself.
893 * spin_lock_irq(gcwq->lock).
896 * Pointer to worker which is executing @work if found, NULL
899 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
900 struct work_struct *work)
902 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
907 * insert_work - insert a work into gcwq
908 * @cwq: cwq @work belongs to
909 * @work: work to insert
910 * @head: insertion point
911 * @extra_flags: extra WORK_STRUCT_* flags to set
913 * Insert @work which belongs to @cwq into @gcwq after @head.
914 * @extra_flags is or'd to work_struct flags.
917 * spin_lock_irq(gcwq->lock).
919 static void insert_work(struct cpu_workqueue_struct *cwq,
920 struct work_struct *work, struct list_head *head,
921 unsigned int extra_flags)
923 struct worker_pool *pool = cwq->pool;
925 /* we own @work, set data and link */
926 set_work_cwq(work, cwq, extra_flags);
929 * Ensure that we get the right work->data if we see the
930 * result of list_add() below, see try_to_grab_pending().
934 list_add_tail(&work->entry, head);
937 * Ensure either worker_sched_deactivated() sees the above
938 * list_add_tail() or we see zero nr_running to avoid workers
939 * lying around lazily while there are works to be processed.
943 if (__need_more_worker(pool))
944 wake_up_worker(pool);
948 * Test whether @work is being queued from another work executing on the
949 * same workqueue. This is rather expensive and should only be used from
952 static bool is_chained_work(struct workqueue_struct *wq)
957 for_each_gcwq_cpu(cpu) {
958 struct global_cwq *gcwq = get_gcwq(cpu);
959 struct worker *worker;
960 struct hlist_node *pos;
963 spin_lock_irqsave(&gcwq->lock, flags);
964 for_each_busy_worker(worker, i, pos, gcwq) {
965 if (worker->task != current)
967 spin_unlock_irqrestore(&gcwq->lock, flags);
969 * I'm @worker, no locking necessary. See if @work
970 * is headed to the same workqueue.
972 return worker->current_cwq->wq == wq;
974 spin_unlock_irqrestore(&gcwq->lock, flags);
979 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
980 struct work_struct *work)
982 struct global_cwq *gcwq;
983 struct cpu_workqueue_struct *cwq;
984 struct list_head *worklist;
985 unsigned int work_flags;
988 * While a work item is PENDING && off queue, a task trying to
989 * steal the PENDING will busy-loop waiting for it to either get
990 * queued or lose PENDING. Grabbing PENDING and queueing should
991 * happen with IRQ disabled.
993 WARN_ON_ONCE(!irqs_disabled());
995 debug_work_activate(work);
997 /* if dying, only works from the same workqueue are allowed */
998 if (unlikely(wq->flags & WQ_DRAINING) &&
999 WARN_ON_ONCE(!is_chained_work(wq)))
1002 /* determine gcwq to use */
1003 if (!(wq->flags & WQ_UNBOUND)) {
1004 struct global_cwq *last_gcwq;
1006 if (cpu == WORK_CPU_UNBOUND)
1007 cpu = raw_smp_processor_id();
1010 * It's multi cpu. If @wq is non-reentrant and @work
1011 * was previously on a different cpu, it might still
1012 * be running there, in which case the work needs to
1013 * be queued on that cpu to guarantee non-reentrance.
1015 gcwq = get_gcwq(cpu);
1016 if (wq->flags & WQ_NON_REENTRANT &&
1017 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1018 struct worker *worker;
1020 spin_lock(&last_gcwq->lock);
1022 worker = find_worker_executing_work(last_gcwq, work);
1024 if (worker && worker->current_cwq->wq == wq)
1027 /* meh... not running there, queue here */
1028 spin_unlock(&last_gcwq->lock);
1029 spin_lock(&gcwq->lock);
1032 spin_lock(&gcwq->lock);
1035 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1036 spin_lock(&gcwq->lock);
1039 /* gcwq determined, get cwq and queue */
1040 cwq = get_cwq(gcwq->cpu, wq);
1041 trace_workqueue_queue_work(cpu, cwq, work);
1043 if (WARN_ON(!list_empty(&work->entry))) {
1044 spin_unlock(&gcwq->lock);
1048 cwq->nr_in_flight[cwq->work_color]++;
1049 work_flags = work_color_to_flags(cwq->work_color);
1051 if (likely(cwq->nr_active < cwq->max_active)) {
1052 trace_workqueue_activate_work(work);
1054 worklist = &cwq->pool->worklist;
1056 work_flags |= WORK_STRUCT_DELAYED;
1057 worklist = &cwq->delayed_works;
1060 insert_work(cwq, work, worklist, work_flags);
1062 spin_unlock(&gcwq->lock);
1066 * queue_work_on - queue work on specific cpu
1067 * @cpu: CPU number to execute work on
1068 * @wq: workqueue to use
1069 * @work: work to queue
1071 * Returns %false if @work was already on a queue, %true otherwise.
1073 * We queue the work to a specific CPU, the caller must ensure it
1076 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1077 struct work_struct *work)
1080 unsigned long flags;
1082 local_irq_save(flags);
1084 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1085 __queue_work(cpu, wq, work);
1089 local_irq_restore(flags);
1092 EXPORT_SYMBOL_GPL(queue_work_on);
1095 * queue_work - queue work on a workqueue
1096 * @wq: workqueue to use
1097 * @work: work to queue
1099 * Returns %false if @work was already on a queue, %true otherwise.
1101 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1102 * it can be processed by another CPU.
1104 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1106 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1108 EXPORT_SYMBOL_GPL(queue_work);
1110 void delayed_work_timer_fn(unsigned long __data)
1112 struct delayed_work *dwork = (struct delayed_work *)__data;
1113 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1115 local_irq_disable();
1116 __queue_work(WORK_CPU_UNBOUND, cwq->wq, &dwork->work);
1119 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1122 * queue_delayed_work_on - queue work on specific CPU after delay
1123 * @cpu: CPU number to execute work on
1124 * @wq: workqueue to use
1125 * @dwork: work to queue
1126 * @delay: number of jiffies to wait before queueing
1128 * Returns %false if @work was already on a queue, %true otherwise. If
1129 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1132 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1133 struct delayed_work *dwork, unsigned long delay)
1135 struct timer_list *timer = &dwork->timer;
1136 struct work_struct *work = &dwork->work;
1138 unsigned long flags;
1141 return queue_work_on(cpu, wq, &dwork->work);
1143 /* read the comment in __queue_work() */
1144 local_irq_save(flags);
1146 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1149 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1150 timer->data != (unsigned long)dwork);
1151 BUG_ON(timer_pending(timer));
1152 BUG_ON(!list_empty(&work->entry));
1154 timer_stats_timer_set_start_info(&dwork->timer);
1157 * This stores cwq for the moment, for the timer_fn.
1158 * Note that the work's gcwq is preserved to allow
1159 * reentrance detection for delayed works.
1161 if (!(wq->flags & WQ_UNBOUND)) {
1162 struct global_cwq *gcwq = get_work_gcwq(work);
1164 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1167 lcpu = raw_smp_processor_id();
1169 lcpu = WORK_CPU_UNBOUND;
1171 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1173 timer->expires = jiffies + delay;
1175 if (unlikely(cpu != WORK_CPU_UNBOUND))
1176 add_timer_on(timer, cpu);
1182 local_irq_restore(flags);
1185 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1188 * queue_delayed_work - queue work on a workqueue after delay
1189 * @wq: workqueue to use
1190 * @dwork: delayable work to queue
1191 * @delay: number of jiffies to wait before queueing
1193 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1195 bool queue_delayed_work(struct workqueue_struct *wq,
1196 struct delayed_work *dwork, unsigned long delay)
1198 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1200 EXPORT_SYMBOL_GPL(queue_delayed_work);
1203 * worker_enter_idle - enter idle state
1204 * @worker: worker which is entering idle state
1206 * @worker is entering idle state. Update stats and idle timer if
1210 * spin_lock_irq(gcwq->lock).
1212 static void worker_enter_idle(struct worker *worker)
1214 struct worker_pool *pool = worker->pool;
1215 struct global_cwq *gcwq = pool->gcwq;
1217 BUG_ON(worker->flags & WORKER_IDLE);
1218 BUG_ON(!list_empty(&worker->entry) &&
1219 (worker->hentry.next || worker->hentry.pprev));
1221 /* can't use worker_set_flags(), also called from start_worker() */
1222 worker->flags |= WORKER_IDLE;
1224 worker->last_active = jiffies;
1226 /* idle_list is LIFO */
1227 list_add(&worker->entry, &pool->idle_list);
1229 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1230 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1233 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1234 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1235 * nr_running, the warning may trigger spuriously. Check iff
1236 * unbind is not in progress.
1238 WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
1239 pool->nr_workers == pool->nr_idle &&
1240 atomic_read(get_pool_nr_running(pool)));
1244 * worker_leave_idle - leave idle state
1245 * @worker: worker which is leaving idle state
1247 * @worker is leaving idle state. Update stats.
1250 * spin_lock_irq(gcwq->lock).
1252 static void worker_leave_idle(struct worker *worker)
1254 struct worker_pool *pool = worker->pool;
1256 BUG_ON(!(worker->flags & WORKER_IDLE));
1257 worker_clr_flags(worker, WORKER_IDLE);
1259 list_del_init(&worker->entry);
1263 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1266 * Works which are scheduled while the cpu is online must at least be
1267 * scheduled to a worker which is bound to the cpu so that if they are
1268 * flushed from cpu callbacks while cpu is going down, they are
1269 * guaranteed to execute on the cpu.
1271 * This function is to be used by rogue workers and rescuers to bind
1272 * themselves to the target cpu and may race with cpu going down or
1273 * coming online. kthread_bind() can't be used because it may put the
1274 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1275 * verbatim as it's best effort and blocking and gcwq may be
1276 * [dis]associated in the meantime.
1278 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1279 * binding against %GCWQ_DISASSOCIATED which is set during
1280 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1281 * enters idle state or fetches works without dropping lock, it can
1282 * guarantee the scheduling requirement described in the first paragraph.
1285 * Might sleep. Called without any lock but returns with gcwq->lock
1289 * %true if the associated gcwq is online (@worker is successfully
1290 * bound), %false if offline.
1292 static bool worker_maybe_bind_and_lock(struct worker *worker)
1293 __acquires(&gcwq->lock)
1295 struct global_cwq *gcwq = worker->pool->gcwq;
1296 struct task_struct *task = worker->task;
1300 * The following call may fail, succeed or succeed
1301 * without actually migrating the task to the cpu if
1302 * it races with cpu hotunplug operation. Verify
1303 * against GCWQ_DISASSOCIATED.
1305 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1306 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1308 spin_lock_irq(&gcwq->lock);
1309 if (gcwq->flags & GCWQ_DISASSOCIATED)
1311 if (task_cpu(task) == gcwq->cpu &&
1312 cpumask_equal(¤t->cpus_allowed,
1313 get_cpu_mask(gcwq->cpu)))
1315 spin_unlock_irq(&gcwq->lock);
1318 * We've raced with CPU hot[un]plug. Give it a breather
1319 * and retry migration. cond_resched() is required here;
1320 * otherwise, we might deadlock against cpu_stop trying to
1321 * bring down the CPU on non-preemptive kernel.
1328 struct idle_rebind {
1329 int cnt; /* # workers to be rebound */
1330 struct completion done; /* all workers rebound */
1334 * Rebind an idle @worker to its CPU. During CPU onlining, this has to
1335 * happen synchronously for idle workers. worker_thread() will test
1336 * %WORKER_REBIND before leaving idle and call this function.
1338 static void idle_worker_rebind(struct worker *worker)
1340 struct global_cwq *gcwq = worker->pool->gcwq;
1342 /* CPU must be online at this point */
1343 WARN_ON(!worker_maybe_bind_and_lock(worker));
1344 if (!--worker->idle_rebind->cnt)
1345 complete(&worker->idle_rebind->done);
1346 spin_unlock_irq(&worker->pool->gcwq->lock);
1348 /* we did our part, wait for rebind_workers() to finish up */
1349 wait_event(gcwq->rebind_hold, !(worker->flags & WORKER_REBIND));
1353 * Function for @worker->rebind.work used to rebind unbound busy workers to
1354 * the associated cpu which is coming back online. This is scheduled by
1355 * cpu up but can race with other cpu hotplug operations and may be
1356 * executed twice without intervening cpu down.
1358 static void busy_worker_rebind_fn(struct work_struct *work)
1360 struct worker *worker = container_of(work, struct worker, rebind_work);
1361 struct global_cwq *gcwq = worker->pool->gcwq;
1363 if (worker_maybe_bind_and_lock(worker))
1364 worker_clr_flags(worker, WORKER_REBIND);
1366 spin_unlock_irq(&gcwq->lock);
1370 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1371 * @gcwq: gcwq of interest
1373 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1374 * is different for idle and busy ones.
1376 * The idle ones should be rebound synchronously and idle rebinding should
1377 * be complete before any worker starts executing work items with
1378 * concurrency management enabled; otherwise, scheduler may oops trying to
1379 * wake up non-local idle worker from wq_worker_sleeping().
1381 * This is achieved by repeatedly requesting rebinding until all idle
1382 * workers are known to have been rebound under @gcwq->lock and holding all
1383 * idle workers from becoming busy until idle rebinding is complete.
1385 * Once idle workers are rebound, busy workers can be rebound as they
1386 * finish executing their current work items. Queueing the rebind work at
1387 * the head of their scheduled lists is enough. Note that nr_running will
1388 * be properbly bumped as busy workers rebind.
1390 * On return, all workers are guaranteed to either be bound or have rebind
1391 * work item scheduled.
1393 static void rebind_workers(struct global_cwq *gcwq)
1394 __releases(&gcwq->lock) __acquires(&gcwq->lock)
1396 struct idle_rebind idle_rebind;
1397 struct worker_pool *pool;
1398 struct worker *worker;
1399 struct hlist_node *pos;
1402 lockdep_assert_held(&gcwq->lock);
1404 for_each_worker_pool(pool, gcwq)
1405 lockdep_assert_held(&pool->manager_mutex);
1408 * Rebind idle workers. Interlocked both ways. We wait for
1409 * workers to rebind via @idle_rebind.done. Workers will wait for
1410 * us to finish up by watching %WORKER_REBIND.
1412 init_completion(&idle_rebind.done);
1414 idle_rebind.cnt = 1;
1415 INIT_COMPLETION(idle_rebind.done);
1417 /* set REBIND and kick idle ones, we'll wait for these later */
1418 for_each_worker_pool(pool, gcwq) {
1419 list_for_each_entry(worker, &pool->idle_list, entry) {
1420 if (worker->flags & WORKER_REBIND)
1423 /* morph UNBOUND to REBIND */
1424 worker->flags &= ~WORKER_UNBOUND;
1425 worker->flags |= WORKER_REBIND;
1428 worker->idle_rebind = &idle_rebind;
1430 /* worker_thread() will call idle_worker_rebind() */
1431 wake_up_process(worker->task);
1435 if (--idle_rebind.cnt) {
1436 spin_unlock_irq(&gcwq->lock);
1437 wait_for_completion(&idle_rebind.done);
1438 spin_lock_irq(&gcwq->lock);
1439 /* busy ones might have become idle while waiting, retry */
1444 * All idle workers are rebound and waiting for %WORKER_REBIND to
1445 * be cleared inside idle_worker_rebind(). Clear and release.
1446 * Clearing %WORKER_REBIND from this foreign context is safe
1447 * because these workers are still guaranteed to be idle.
1449 for_each_worker_pool(pool, gcwq)
1450 list_for_each_entry(worker, &pool->idle_list, entry)
1451 worker->flags &= ~WORKER_REBIND;
1453 wake_up_all(&gcwq->rebind_hold);
1455 /* rebind busy workers */
1456 for_each_busy_worker(worker, i, pos, gcwq) {
1457 struct work_struct *rebind_work = &worker->rebind_work;
1459 /* morph UNBOUND to REBIND */
1460 worker->flags &= ~WORKER_UNBOUND;
1461 worker->flags |= WORKER_REBIND;
1463 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1464 work_data_bits(rebind_work)))
1467 /* wq doesn't matter, use the default one */
1468 debug_work_activate(rebind_work);
1469 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
1470 worker->scheduled.next,
1471 work_color_to_flags(WORK_NO_COLOR));
1475 static struct worker *alloc_worker(void)
1477 struct worker *worker;
1479 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1481 INIT_LIST_HEAD(&worker->entry);
1482 INIT_LIST_HEAD(&worker->scheduled);
1483 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1484 /* on creation a worker is in !idle && prep state */
1485 worker->flags = WORKER_PREP;
1491 * create_worker - create a new workqueue worker
1492 * @pool: pool the new worker will belong to
1494 * Create a new worker which is bound to @pool. The returned worker
1495 * can be started by calling start_worker() or destroyed using
1499 * Might sleep. Does GFP_KERNEL allocations.
1502 * Pointer to the newly created worker.
1504 static struct worker *create_worker(struct worker_pool *pool)
1506 struct global_cwq *gcwq = pool->gcwq;
1507 const char *pri = worker_pool_pri(pool) ? "H" : "";
1508 struct worker *worker = NULL;
1511 spin_lock_irq(&gcwq->lock);
1512 while (ida_get_new(&pool->worker_ida, &id)) {
1513 spin_unlock_irq(&gcwq->lock);
1514 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1516 spin_lock_irq(&gcwq->lock);
1518 spin_unlock_irq(&gcwq->lock);
1520 worker = alloc_worker();
1524 worker->pool = pool;
1527 if (gcwq->cpu != WORK_CPU_UNBOUND)
1528 worker->task = kthread_create_on_node(worker_thread,
1529 worker, cpu_to_node(gcwq->cpu),
1530 "kworker/%u:%d%s", gcwq->cpu, id, pri);
1532 worker->task = kthread_create(worker_thread, worker,
1533 "kworker/u:%d%s", id, pri);
1534 if (IS_ERR(worker->task))
1537 if (worker_pool_pri(pool))
1538 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1541 * Determine CPU binding of the new worker depending on
1542 * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1543 * flag remains stable across this function. See the comments
1544 * above the flag definition for details.
1546 * As an unbound worker may later become a regular one if CPU comes
1547 * online, make sure every worker has %PF_THREAD_BOUND set.
1549 if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
1550 kthread_bind(worker->task, gcwq->cpu);
1552 worker->task->flags |= PF_THREAD_BOUND;
1553 worker->flags |= WORKER_UNBOUND;
1559 spin_lock_irq(&gcwq->lock);
1560 ida_remove(&pool->worker_ida, id);
1561 spin_unlock_irq(&gcwq->lock);
1568 * start_worker - start a newly created worker
1569 * @worker: worker to start
1571 * Make the gcwq aware of @worker and start it.
1574 * spin_lock_irq(gcwq->lock).
1576 static void start_worker(struct worker *worker)
1578 worker->flags |= WORKER_STARTED;
1579 worker->pool->nr_workers++;
1580 worker_enter_idle(worker);
1581 wake_up_process(worker->task);
1585 * destroy_worker - destroy a workqueue worker
1586 * @worker: worker to be destroyed
1588 * Destroy @worker and adjust @gcwq stats accordingly.
1591 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1593 static void destroy_worker(struct worker *worker)
1595 struct worker_pool *pool = worker->pool;
1596 struct global_cwq *gcwq = pool->gcwq;
1597 int id = worker->id;
1599 /* sanity check frenzy */
1600 BUG_ON(worker->current_work);
1601 BUG_ON(!list_empty(&worker->scheduled));
1603 if (worker->flags & WORKER_STARTED)
1605 if (worker->flags & WORKER_IDLE)
1608 list_del_init(&worker->entry);
1609 worker->flags |= WORKER_DIE;
1611 spin_unlock_irq(&gcwq->lock);
1613 kthread_stop(worker->task);
1616 spin_lock_irq(&gcwq->lock);
1617 ida_remove(&pool->worker_ida, id);
1620 static void idle_worker_timeout(unsigned long __pool)
1622 struct worker_pool *pool = (void *)__pool;
1623 struct global_cwq *gcwq = pool->gcwq;
1625 spin_lock_irq(&gcwq->lock);
1627 if (too_many_workers(pool)) {
1628 struct worker *worker;
1629 unsigned long expires;
1631 /* idle_list is kept in LIFO order, check the last one */
1632 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1633 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1635 if (time_before(jiffies, expires))
1636 mod_timer(&pool->idle_timer, expires);
1638 /* it's been idle for too long, wake up manager */
1639 pool->flags |= POOL_MANAGE_WORKERS;
1640 wake_up_worker(pool);
1644 spin_unlock_irq(&gcwq->lock);
1647 static bool send_mayday(struct work_struct *work)
1649 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1650 struct workqueue_struct *wq = cwq->wq;
1653 if (!(wq->flags & WQ_RESCUER))
1656 /* mayday mayday mayday */
1657 cpu = cwq->pool->gcwq->cpu;
1658 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1659 if (cpu == WORK_CPU_UNBOUND)
1661 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1662 wake_up_process(wq->rescuer->task);
1666 static void gcwq_mayday_timeout(unsigned long __pool)
1668 struct worker_pool *pool = (void *)__pool;
1669 struct global_cwq *gcwq = pool->gcwq;
1670 struct work_struct *work;
1672 spin_lock_irq(&gcwq->lock);
1674 if (need_to_create_worker(pool)) {
1676 * We've been trying to create a new worker but
1677 * haven't been successful. We might be hitting an
1678 * allocation deadlock. Send distress signals to
1681 list_for_each_entry(work, &pool->worklist, entry)
1685 spin_unlock_irq(&gcwq->lock);
1687 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1691 * maybe_create_worker - create a new worker if necessary
1692 * @pool: pool to create a new worker for
1694 * Create a new worker for @pool if necessary. @pool is guaranteed to
1695 * have at least one idle worker on return from this function. If
1696 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1697 * sent to all rescuers with works scheduled on @pool to resolve
1698 * possible allocation deadlock.
1700 * On return, need_to_create_worker() is guaranteed to be false and
1701 * may_start_working() true.
1704 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1705 * multiple times. Does GFP_KERNEL allocations. Called only from
1709 * false if no action was taken and gcwq->lock stayed locked, true
1712 static bool maybe_create_worker(struct worker_pool *pool)
1713 __releases(&gcwq->lock)
1714 __acquires(&gcwq->lock)
1716 struct global_cwq *gcwq = pool->gcwq;
1718 if (!need_to_create_worker(pool))
1721 spin_unlock_irq(&gcwq->lock);
1723 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1724 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1727 struct worker *worker;
1729 worker = create_worker(pool);
1731 del_timer_sync(&pool->mayday_timer);
1732 spin_lock_irq(&gcwq->lock);
1733 start_worker(worker);
1734 BUG_ON(need_to_create_worker(pool));
1738 if (!need_to_create_worker(pool))
1741 __set_current_state(TASK_INTERRUPTIBLE);
1742 schedule_timeout(CREATE_COOLDOWN);
1744 if (!need_to_create_worker(pool))
1748 del_timer_sync(&pool->mayday_timer);
1749 spin_lock_irq(&gcwq->lock);
1750 if (need_to_create_worker(pool))
1756 * maybe_destroy_worker - destroy workers which have been idle for a while
1757 * @pool: pool to destroy workers for
1759 * Destroy @pool workers which have been idle for longer than
1760 * IDLE_WORKER_TIMEOUT.
1763 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1764 * multiple times. Called only from manager.
1767 * false if no action was taken and gcwq->lock stayed locked, true
1770 static bool maybe_destroy_workers(struct worker_pool *pool)
1774 while (too_many_workers(pool)) {
1775 struct worker *worker;
1776 unsigned long expires;
1778 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1779 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1781 if (time_before(jiffies, expires)) {
1782 mod_timer(&pool->idle_timer, expires);
1786 destroy_worker(worker);
1794 * manage_workers - manage worker pool
1797 * Assume the manager role and manage gcwq worker pool @worker belongs
1798 * to. At any given time, there can be only zero or one manager per
1799 * gcwq. The exclusion is handled automatically by this function.
1801 * The caller can safely start processing works on false return. On
1802 * true return, it's guaranteed that need_to_create_worker() is false
1803 * and may_start_working() is true.
1806 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1807 * multiple times. Does GFP_KERNEL allocations.
1810 * false if no action was taken and gcwq->lock stayed locked, true if
1811 * some action was taken.
1813 static bool manage_workers(struct worker *worker)
1815 struct worker_pool *pool = worker->pool;
1818 if (!mutex_trylock(&pool->manager_mutex))
1821 pool->flags &= ~POOL_MANAGE_WORKERS;
1824 * Destroy and then create so that may_start_working() is true
1827 ret |= maybe_destroy_workers(pool);
1828 ret |= maybe_create_worker(pool);
1830 mutex_unlock(&pool->manager_mutex);
1835 * move_linked_works - move linked works to a list
1836 * @work: start of series of works to be scheduled
1837 * @head: target list to append @work to
1838 * @nextp: out paramter for nested worklist walking
1840 * Schedule linked works starting from @work to @head. Work series to
1841 * be scheduled starts at @work and includes any consecutive work with
1842 * WORK_STRUCT_LINKED set in its predecessor.
1844 * If @nextp is not NULL, it's updated to point to the next work of
1845 * the last scheduled work. This allows move_linked_works() to be
1846 * nested inside outer list_for_each_entry_safe().
1849 * spin_lock_irq(gcwq->lock).
1851 static void move_linked_works(struct work_struct *work, struct list_head *head,
1852 struct work_struct **nextp)
1854 struct work_struct *n;
1857 * Linked worklist will always end before the end of the list,
1858 * use NULL for list head.
1860 list_for_each_entry_safe_from(work, n, NULL, entry) {
1861 list_move_tail(&work->entry, head);
1862 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1867 * If we're already inside safe list traversal and have moved
1868 * multiple works to the scheduled queue, the next position
1869 * needs to be updated.
1875 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1877 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1878 struct work_struct, entry);
1880 trace_workqueue_activate_work(work);
1881 move_linked_works(work, &cwq->pool->worklist, NULL);
1882 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1887 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1888 * @cwq: cwq of interest
1889 * @color: color of work which left the queue
1890 * @delayed: for a delayed work
1892 * A work either has completed or is removed from pending queue,
1893 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1896 * spin_lock_irq(gcwq->lock).
1898 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1901 /* ignore uncolored works */
1902 if (color == WORK_NO_COLOR)
1905 cwq->nr_in_flight[color]--;
1909 if (!list_empty(&cwq->delayed_works)) {
1910 /* one down, submit a delayed one */
1911 if (cwq->nr_active < cwq->max_active)
1912 cwq_activate_first_delayed(cwq);
1916 /* is flush in progress and are we at the flushing tip? */
1917 if (likely(cwq->flush_color != color))
1920 /* are there still in-flight works? */
1921 if (cwq->nr_in_flight[color])
1924 /* this cwq is done, clear flush_color */
1925 cwq->flush_color = -1;
1928 * If this was the last cwq, wake up the first flusher. It
1929 * will handle the rest.
1931 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1932 complete(&cwq->wq->first_flusher->done);
1936 * process_one_work - process single work
1938 * @work: work to process
1940 * Process @work. This function contains all the logics necessary to
1941 * process a single work including synchronization against and
1942 * interaction with other workers on the same cpu, queueing and
1943 * flushing. As long as context requirement is met, any worker can
1944 * call this function to process a work.
1947 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1949 static void process_one_work(struct worker *worker, struct work_struct *work)
1950 __releases(&gcwq->lock)
1951 __acquires(&gcwq->lock)
1953 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1954 struct worker_pool *pool = worker->pool;
1955 struct global_cwq *gcwq = pool->gcwq;
1956 struct hlist_head *bwh = busy_worker_head(gcwq, work);
1957 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1958 work_func_t f = work->func;
1960 struct worker *collision;
1961 #ifdef CONFIG_LOCKDEP
1963 * It is permissible to free the struct work_struct from
1964 * inside the function that is called from it, this we need to
1965 * take into account for lockdep too. To avoid bogus "held
1966 * lock freed" warnings as well as problems when looking into
1967 * work->lockdep_map, make a copy and use that here.
1969 struct lockdep_map lockdep_map;
1971 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
1974 * Ensure we're on the correct CPU. DISASSOCIATED test is
1975 * necessary to avoid spurious warnings from rescuers servicing the
1976 * unbound or a disassociated gcwq.
1978 WARN_ON_ONCE(!(worker->flags & (WORKER_UNBOUND | WORKER_REBIND)) &&
1979 !(gcwq->flags & GCWQ_DISASSOCIATED) &&
1980 raw_smp_processor_id() != gcwq->cpu);
1983 * A single work shouldn't be executed concurrently by
1984 * multiple workers on a single cpu. Check whether anyone is
1985 * already processing the work. If so, defer the work to the
1986 * currently executing one.
1988 collision = __find_worker_executing_work(gcwq, bwh, work);
1989 if (unlikely(collision)) {
1990 move_linked_works(work, &collision->scheduled, NULL);
1994 /* claim and dequeue */
1995 debug_work_deactivate(work);
1996 hlist_add_head(&worker->hentry, bwh);
1997 worker->current_work = work;
1998 worker->current_cwq = cwq;
1999 work_color = get_work_color(work);
2001 list_del_init(&work->entry);
2004 * CPU intensive works don't participate in concurrency
2005 * management. They're the scheduler's responsibility.
2007 if (unlikely(cpu_intensive))
2008 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2011 * Unbound gcwq isn't concurrency managed and work items should be
2012 * executed ASAP. Wake up another worker if necessary.
2014 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2015 wake_up_worker(pool);
2018 * Record the last CPU and clear PENDING. The following wmb is
2019 * paired with the implied mb in test_and_set_bit(PENDING) and
2020 * ensures all updates to @work made here are visible to and
2021 * precede any updates by the next PENDING owner. Also, clear
2022 * PENDING inside @gcwq->lock so that PENDING and queued state
2023 * changes happen together while IRQ is disabled.
2026 set_work_cpu_and_clear_pending(work, gcwq->cpu);
2028 spin_unlock_irq(&gcwq->lock);
2030 lock_map_acquire_read(&cwq->wq->lockdep_map);
2031 lock_map_acquire(&lockdep_map);
2032 trace_workqueue_execute_start(work);
2035 * While we must be careful to not use "work" after this, the trace
2036 * point will only record its address.
2038 trace_workqueue_execute_end(work);
2039 lock_map_release(&lockdep_map);
2040 lock_map_release(&cwq->wq->lockdep_map);
2042 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2043 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
2045 current->comm, preempt_count(), task_pid_nr(current));
2046 printk(KERN_ERR " last function: ");
2047 print_symbol("%s\n", (unsigned long)f);
2048 debug_show_held_locks(current);
2052 spin_lock_irq(&gcwq->lock);
2054 /* clear cpu intensive status */
2055 if (unlikely(cpu_intensive))
2056 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2058 /* we're done with it, release */
2059 hlist_del_init(&worker->hentry);
2060 worker->current_work = NULL;
2061 worker->current_cwq = NULL;
2062 cwq_dec_nr_in_flight(cwq, work_color, false);
2066 * process_scheduled_works - process scheduled works
2069 * Process all scheduled works. Please note that the scheduled list
2070 * may change while processing a work, so this function repeatedly
2071 * fetches a work from the top and executes it.
2074 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2077 static void process_scheduled_works(struct worker *worker)
2079 while (!list_empty(&worker->scheduled)) {
2080 struct work_struct *work = list_first_entry(&worker->scheduled,
2081 struct work_struct, entry);
2082 process_one_work(worker, work);
2087 * worker_thread - the worker thread function
2090 * The gcwq worker thread function. There's a single dynamic pool of
2091 * these per each cpu. These workers process all works regardless of
2092 * their specific target workqueue. The only exception is works which
2093 * belong to workqueues with a rescuer which will be explained in
2096 static int worker_thread(void *__worker)
2098 struct worker *worker = __worker;
2099 struct worker_pool *pool = worker->pool;
2100 struct global_cwq *gcwq = pool->gcwq;
2102 /* tell the scheduler that this is a workqueue worker */
2103 worker->task->flags |= PF_WQ_WORKER;
2105 spin_lock_irq(&gcwq->lock);
2108 * DIE can be set only while idle and REBIND set while busy has
2109 * @worker->rebind_work scheduled. Checking here is enough.
2111 if (unlikely(worker->flags & (WORKER_REBIND | WORKER_DIE))) {
2112 spin_unlock_irq(&gcwq->lock);
2114 if (worker->flags & WORKER_DIE) {
2115 worker->task->flags &= ~PF_WQ_WORKER;
2119 idle_worker_rebind(worker);
2123 worker_leave_idle(worker);
2125 /* no more worker necessary? */
2126 if (!need_more_worker(pool))
2129 /* do we need to manage? */
2130 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2134 * ->scheduled list can only be filled while a worker is
2135 * preparing to process a work or actually processing it.
2136 * Make sure nobody diddled with it while I was sleeping.
2138 BUG_ON(!list_empty(&worker->scheduled));
2141 * When control reaches this point, we're guaranteed to have
2142 * at least one idle worker or that someone else has already
2143 * assumed the manager role.
2145 worker_clr_flags(worker, WORKER_PREP);
2148 struct work_struct *work =
2149 list_first_entry(&pool->worklist,
2150 struct work_struct, entry);
2152 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2153 /* optimization path, not strictly necessary */
2154 process_one_work(worker, work);
2155 if (unlikely(!list_empty(&worker->scheduled)))
2156 process_scheduled_works(worker);
2158 move_linked_works(work, &worker->scheduled, NULL);
2159 process_scheduled_works(worker);
2161 } while (keep_working(pool));
2163 worker_set_flags(worker, WORKER_PREP, false);
2165 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2169 * gcwq->lock is held and there's no work to process and no
2170 * need to manage, sleep. Workers are woken up only while
2171 * holding gcwq->lock or from local cpu, so setting the
2172 * current state before releasing gcwq->lock is enough to
2173 * prevent losing any event.
2175 worker_enter_idle(worker);
2176 __set_current_state(TASK_INTERRUPTIBLE);
2177 spin_unlock_irq(&gcwq->lock);
2183 * rescuer_thread - the rescuer thread function
2184 * @__wq: the associated workqueue
2186 * Workqueue rescuer thread function. There's one rescuer for each
2187 * workqueue which has WQ_RESCUER set.
2189 * Regular work processing on a gcwq may block trying to create a new
2190 * worker which uses GFP_KERNEL allocation which has slight chance of
2191 * developing into deadlock if some works currently on the same queue
2192 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2193 * the problem rescuer solves.
2195 * When such condition is possible, the gcwq summons rescuers of all
2196 * workqueues which have works queued on the gcwq and let them process
2197 * those works so that forward progress can be guaranteed.
2199 * This should happen rarely.
2201 static int rescuer_thread(void *__wq)
2203 struct workqueue_struct *wq = __wq;
2204 struct worker *rescuer = wq->rescuer;
2205 struct list_head *scheduled = &rescuer->scheduled;
2206 bool is_unbound = wq->flags & WQ_UNBOUND;
2209 set_user_nice(current, RESCUER_NICE_LEVEL);
2211 set_current_state(TASK_INTERRUPTIBLE);
2213 if (kthread_should_stop())
2217 * See whether any cpu is asking for help. Unbounded
2218 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2220 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2221 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2222 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2223 struct worker_pool *pool = cwq->pool;
2224 struct global_cwq *gcwq = pool->gcwq;
2225 struct work_struct *work, *n;
2227 __set_current_state(TASK_RUNNING);
2228 mayday_clear_cpu(cpu, wq->mayday_mask);
2230 /* migrate to the target cpu if possible */
2231 rescuer->pool = pool;
2232 worker_maybe_bind_and_lock(rescuer);
2235 * Slurp in all works issued via this workqueue and
2238 BUG_ON(!list_empty(&rescuer->scheduled));
2239 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2240 if (get_work_cwq(work) == cwq)
2241 move_linked_works(work, scheduled, &n);
2243 process_scheduled_works(rescuer);
2246 * Leave this gcwq. If keep_working() is %true, notify a
2247 * regular worker; otherwise, we end up with 0 concurrency
2248 * and stalling the execution.
2250 if (keep_working(pool))
2251 wake_up_worker(pool);
2253 spin_unlock_irq(&gcwq->lock);
2261 struct work_struct work;
2262 struct completion done;
2265 static void wq_barrier_func(struct work_struct *work)
2267 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2268 complete(&barr->done);
2272 * insert_wq_barrier - insert a barrier work
2273 * @cwq: cwq to insert barrier into
2274 * @barr: wq_barrier to insert
2275 * @target: target work to attach @barr to
2276 * @worker: worker currently executing @target, NULL if @target is not executing
2278 * @barr is linked to @target such that @barr is completed only after
2279 * @target finishes execution. Please note that the ordering
2280 * guarantee is observed only with respect to @target and on the local
2283 * Currently, a queued barrier can't be canceled. This is because
2284 * try_to_grab_pending() can't determine whether the work to be
2285 * grabbed is at the head of the queue and thus can't clear LINKED
2286 * flag of the previous work while there must be a valid next work
2287 * after a work with LINKED flag set.
2289 * Note that when @worker is non-NULL, @target may be modified
2290 * underneath us, so we can't reliably determine cwq from @target.
2293 * spin_lock_irq(gcwq->lock).
2295 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2296 struct wq_barrier *barr,
2297 struct work_struct *target, struct worker *worker)
2299 struct list_head *head;
2300 unsigned int linked = 0;
2303 * debugobject calls are safe here even with gcwq->lock locked
2304 * as we know for sure that this will not trigger any of the
2305 * checks and call back into the fixup functions where we
2308 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2309 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2310 init_completion(&barr->done);
2313 * If @target is currently being executed, schedule the
2314 * barrier to the worker; otherwise, put it after @target.
2317 head = worker->scheduled.next;
2319 unsigned long *bits = work_data_bits(target);
2321 head = target->entry.next;
2322 /* there can already be other linked works, inherit and set */
2323 linked = *bits & WORK_STRUCT_LINKED;
2324 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2327 debug_work_activate(&barr->work);
2328 insert_work(cwq, &barr->work, head,
2329 work_color_to_flags(WORK_NO_COLOR) | linked);
2333 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2334 * @wq: workqueue being flushed
2335 * @flush_color: new flush color, < 0 for no-op
2336 * @work_color: new work color, < 0 for no-op
2338 * Prepare cwqs for workqueue flushing.
2340 * If @flush_color is non-negative, flush_color on all cwqs should be
2341 * -1. If no cwq has in-flight commands at the specified color, all
2342 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2343 * has in flight commands, its cwq->flush_color is set to
2344 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2345 * wakeup logic is armed and %true is returned.
2347 * The caller should have initialized @wq->first_flusher prior to
2348 * calling this function with non-negative @flush_color. If
2349 * @flush_color is negative, no flush color update is done and %false
2352 * If @work_color is non-negative, all cwqs should have the same
2353 * work_color which is previous to @work_color and all will be
2354 * advanced to @work_color.
2357 * mutex_lock(wq->flush_mutex).
2360 * %true if @flush_color >= 0 and there's something to flush. %false
2363 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2364 int flush_color, int work_color)
2369 if (flush_color >= 0) {
2370 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2371 atomic_set(&wq->nr_cwqs_to_flush, 1);
2374 for_each_cwq_cpu(cpu, wq) {
2375 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2376 struct global_cwq *gcwq = cwq->pool->gcwq;
2378 spin_lock_irq(&gcwq->lock);
2380 if (flush_color >= 0) {
2381 BUG_ON(cwq->flush_color != -1);
2383 if (cwq->nr_in_flight[flush_color]) {
2384 cwq->flush_color = flush_color;
2385 atomic_inc(&wq->nr_cwqs_to_flush);
2390 if (work_color >= 0) {
2391 BUG_ON(work_color != work_next_color(cwq->work_color));
2392 cwq->work_color = work_color;
2395 spin_unlock_irq(&gcwq->lock);
2398 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2399 complete(&wq->first_flusher->done);
2405 * flush_workqueue - ensure that any scheduled work has run to completion.
2406 * @wq: workqueue to flush
2408 * Forces execution of the workqueue and blocks until its completion.
2409 * This is typically used in driver shutdown handlers.
2411 * We sleep until all works which were queued on entry have been handled,
2412 * but we are not livelocked by new incoming ones.
2414 void flush_workqueue(struct workqueue_struct *wq)
2416 struct wq_flusher this_flusher = {
2417 .list = LIST_HEAD_INIT(this_flusher.list),
2419 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2423 lock_map_acquire(&wq->lockdep_map);
2424 lock_map_release(&wq->lockdep_map);
2426 mutex_lock(&wq->flush_mutex);
2429 * Start-to-wait phase
2431 next_color = work_next_color(wq->work_color);
2433 if (next_color != wq->flush_color) {
2435 * Color space is not full. The current work_color
2436 * becomes our flush_color and work_color is advanced
2439 BUG_ON(!list_empty(&wq->flusher_overflow));
2440 this_flusher.flush_color = wq->work_color;
2441 wq->work_color = next_color;
2443 if (!wq->first_flusher) {
2444 /* no flush in progress, become the first flusher */
2445 BUG_ON(wq->flush_color != this_flusher.flush_color);
2447 wq->first_flusher = &this_flusher;
2449 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2451 /* nothing to flush, done */
2452 wq->flush_color = next_color;
2453 wq->first_flusher = NULL;
2458 BUG_ON(wq->flush_color == this_flusher.flush_color);
2459 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2460 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2464 * Oops, color space is full, wait on overflow queue.
2465 * The next flush completion will assign us
2466 * flush_color and transfer to flusher_queue.
2468 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2471 mutex_unlock(&wq->flush_mutex);
2473 wait_for_completion(&this_flusher.done);
2476 * Wake-up-and-cascade phase
2478 * First flushers are responsible for cascading flushes and
2479 * handling overflow. Non-first flushers can simply return.
2481 if (wq->first_flusher != &this_flusher)
2484 mutex_lock(&wq->flush_mutex);
2486 /* we might have raced, check again with mutex held */
2487 if (wq->first_flusher != &this_flusher)
2490 wq->first_flusher = NULL;
2492 BUG_ON(!list_empty(&this_flusher.list));
2493 BUG_ON(wq->flush_color != this_flusher.flush_color);
2496 struct wq_flusher *next, *tmp;
2498 /* complete all the flushers sharing the current flush color */
2499 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2500 if (next->flush_color != wq->flush_color)
2502 list_del_init(&next->list);
2503 complete(&next->done);
2506 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2507 wq->flush_color != work_next_color(wq->work_color));
2509 /* this flush_color is finished, advance by one */
2510 wq->flush_color = work_next_color(wq->flush_color);
2512 /* one color has been freed, handle overflow queue */
2513 if (!list_empty(&wq->flusher_overflow)) {
2515 * Assign the same color to all overflowed
2516 * flushers, advance work_color and append to
2517 * flusher_queue. This is the start-to-wait
2518 * phase for these overflowed flushers.
2520 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2521 tmp->flush_color = wq->work_color;
2523 wq->work_color = work_next_color(wq->work_color);
2525 list_splice_tail_init(&wq->flusher_overflow,
2526 &wq->flusher_queue);
2527 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2530 if (list_empty(&wq->flusher_queue)) {
2531 BUG_ON(wq->flush_color != wq->work_color);
2536 * Need to flush more colors. Make the next flusher
2537 * the new first flusher and arm cwqs.
2539 BUG_ON(wq->flush_color == wq->work_color);
2540 BUG_ON(wq->flush_color != next->flush_color);
2542 list_del_init(&next->list);
2543 wq->first_flusher = next;
2545 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2549 * Meh... this color is already done, clear first
2550 * flusher and repeat cascading.
2552 wq->first_flusher = NULL;
2556 mutex_unlock(&wq->flush_mutex);
2558 EXPORT_SYMBOL_GPL(flush_workqueue);
2561 * drain_workqueue - drain a workqueue
2562 * @wq: workqueue to drain
2564 * Wait until the workqueue becomes empty. While draining is in progress,
2565 * only chain queueing is allowed. IOW, only currently pending or running
2566 * work items on @wq can queue further work items on it. @wq is flushed
2567 * repeatedly until it becomes empty. The number of flushing is detemined
2568 * by the depth of chaining and should be relatively short. Whine if it
2571 void drain_workqueue(struct workqueue_struct *wq)
2573 unsigned int flush_cnt = 0;
2577 * __queue_work() needs to test whether there are drainers, is much
2578 * hotter than drain_workqueue() and already looks at @wq->flags.
2579 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2581 spin_lock(&workqueue_lock);
2582 if (!wq->nr_drainers++)
2583 wq->flags |= WQ_DRAINING;
2584 spin_unlock(&workqueue_lock);
2586 flush_workqueue(wq);
2588 for_each_cwq_cpu(cpu, wq) {
2589 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2592 spin_lock_irq(&cwq->pool->gcwq->lock);
2593 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2594 spin_unlock_irq(&cwq->pool->gcwq->lock);
2599 if (++flush_cnt == 10 ||
2600 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2601 pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n",
2602 wq->name, flush_cnt);
2606 spin_lock(&workqueue_lock);
2607 if (!--wq->nr_drainers)
2608 wq->flags &= ~WQ_DRAINING;
2609 spin_unlock(&workqueue_lock);
2611 EXPORT_SYMBOL_GPL(drain_workqueue);
2613 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2614 bool wait_executing)
2616 struct worker *worker = NULL;
2617 struct global_cwq *gcwq;
2618 struct cpu_workqueue_struct *cwq;
2621 gcwq = get_work_gcwq(work);
2625 spin_lock_irq(&gcwq->lock);
2626 if (!list_empty(&work->entry)) {
2628 * See the comment near try_to_grab_pending()->smp_rmb().
2629 * If it was re-queued to a different gcwq under us, we
2630 * are not going to wait.
2633 cwq = get_work_cwq(work);
2634 if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
2636 } else if (wait_executing) {
2637 worker = find_worker_executing_work(gcwq, work);
2640 cwq = worker->current_cwq;
2644 insert_wq_barrier(cwq, barr, work, worker);
2645 spin_unlock_irq(&gcwq->lock);
2648 * If @max_active is 1 or rescuer is in use, flushing another work
2649 * item on the same workqueue may lead to deadlock. Make sure the
2650 * flusher is not running on the same workqueue by verifying write
2653 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2654 lock_map_acquire(&cwq->wq->lockdep_map);
2656 lock_map_acquire_read(&cwq->wq->lockdep_map);
2657 lock_map_release(&cwq->wq->lockdep_map);
2661 spin_unlock_irq(&gcwq->lock);
2666 * flush_work - wait for a work to finish executing the last queueing instance
2667 * @work: the work to flush
2669 * Wait until @work has finished execution. This function considers
2670 * only the last queueing instance of @work. If @work has been
2671 * enqueued across different CPUs on a non-reentrant workqueue or on
2672 * multiple workqueues, @work might still be executing on return on
2673 * some of the CPUs from earlier queueing.
2675 * If @work was queued only on a non-reentrant, ordered or unbound
2676 * workqueue, @work is guaranteed to be idle on return if it hasn't
2677 * been requeued since flush started.
2680 * %true if flush_work() waited for the work to finish execution,
2681 * %false if it was already idle.
2683 bool flush_work(struct work_struct *work)
2685 struct wq_barrier barr;
2687 lock_map_acquire(&work->lockdep_map);
2688 lock_map_release(&work->lockdep_map);
2690 if (start_flush_work(work, &barr, true)) {
2691 wait_for_completion(&barr.done);
2692 destroy_work_on_stack(&barr.work);
2697 EXPORT_SYMBOL_GPL(flush_work);
2699 static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2701 struct wq_barrier barr;
2702 struct worker *worker;
2704 spin_lock_irq(&gcwq->lock);
2706 worker = find_worker_executing_work(gcwq, work);
2707 if (unlikely(worker))
2708 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2710 spin_unlock_irq(&gcwq->lock);
2712 if (unlikely(worker)) {
2713 wait_for_completion(&barr.done);
2714 destroy_work_on_stack(&barr.work);
2720 static bool wait_on_work(struct work_struct *work)
2727 lock_map_acquire(&work->lockdep_map);
2728 lock_map_release(&work->lockdep_map);
2730 for_each_gcwq_cpu(cpu)
2731 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2736 * flush_work_sync - wait until a work has finished execution
2737 * @work: the work to flush
2739 * Wait until @work has finished execution. On return, it's
2740 * guaranteed that all queueing instances of @work which happened
2741 * before this function is called are finished. In other words, if
2742 * @work hasn't been requeued since this function was called, @work is
2743 * guaranteed to be idle on return.
2746 * %true if flush_work_sync() waited for the work to finish execution,
2747 * %false if it was already idle.
2749 bool flush_work_sync(struct work_struct *work)
2751 struct wq_barrier barr;
2752 bool pending, waited;
2754 /* we'll wait for executions separately, queue barr only if pending */
2755 pending = start_flush_work(work, &barr, false);
2757 /* wait for executions to finish */
2758 waited = wait_on_work(work);
2760 /* wait for the pending one */
2762 wait_for_completion(&barr.done);
2763 destroy_work_on_stack(&barr.work);
2766 return pending || waited;
2768 EXPORT_SYMBOL_GPL(flush_work_sync);
2771 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2772 * so this work can't be re-armed in any way.
2774 static int try_to_grab_pending(struct work_struct *work)
2776 struct global_cwq *gcwq;
2779 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2783 * The queueing is in progress, or it is already queued. Try to
2784 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2786 gcwq = get_work_gcwq(work);
2790 spin_lock_irq(&gcwq->lock);
2791 if (!list_empty(&work->entry)) {
2793 * This work is queued, but perhaps we locked the wrong gcwq.
2794 * In that case we must see the new value after rmb(), see
2795 * insert_work()->wmb().
2798 if (gcwq == get_work_gcwq(work)) {
2799 debug_work_deactivate(work);
2800 list_del_init(&work->entry);
2801 cwq_dec_nr_in_flight(get_work_cwq(work),
2802 get_work_color(work),
2803 *work_data_bits(work) & WORK_STRUCT_DELAYED);
2807 spin_unlock_irq(&gcwq->lock);
2812 static bool __cancel_work_timer(struct work_struct *work,
2813 struct timer_list* timer)
2818 ret = (timer && likely(del_timer(timer)));
2820 ret = try_to_grab_pending(work);
2822 } while (unlikely(ret < 0));
2824 clear_work_data(work);
2829 * cancel_work_sync - cancel a work and wait for it to finish
2830 * @work: the work to cancel
2832 * Cancel @work and wait for its execution to finish. This function
2833 * can be used even if the work re-queues itself or migrates to
2834 * another workqueue. On return from this function, @work is
2835 * guaranteed to be not pending or executing on any CPU.
2837 * cancel_work_sync(&delayed_work->work) must not be used for
2838 * delayed_work's. Use cancel_delayed_work_sync() instead.
2840 * The caller must ensure that the workqueue on which @work was last
2841 * queued can't be destroyed before this function returns.
2844 * %true if @work was pending, %false otherwise.
2846 bool cancel_work_sync(struct work_struct *work)
2848 return __cancel_work_timer(work, NULL);
2850 EXPORT_SYMBOL_GPL(cancel_work_sync);
2853 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2854 * @dwork: the delayed work to flush
2856 * Delayed timer is cancelled and the pending work is queued for
2857 * immediate execution. Like flush_work(), this function only
2858 * considers the last queueing instance of @dwork.
2861 * %true if flush_work() waited for the work to finish execution,
2862 * %false if it was already idle.
2864 bool flush_delayed_work(struct delayed_work *dwork)
2866 local_irq_disable();
2867 if (del_timer_sync(&dwork->timer))
2868 __queue_work(WORK_CPU_UNBOUND,
2869 get_work_cwq(&dwork->work)->wq, &dwork->work);
2871 return flush_work(&dwork->work);
2873 EXPORT_SYMBOL(flush_delayed_work);
2876 * flush_delayed_work_sync - wait for a dwork to finish
2877 * @dwork: the delayed work to flush
2879 * Delayed timer is cancelled and the pending work is queued for
2880 * execution immediately. Other than timer handling, its behavior
2881 * is identical to flush_work_sync().
2884 * %true if flush_work_sync() waited for the work to finish execution,
2885 * %false if it was already idle.
2887 bool flush_delayed_work_sync(struct delayed_work *dwork)
2889 local_irq_disable();
2890 if (del_timer_sync(&dwork->timer))
2891 __queue_work(WORK_CPU_UNBOUND,
2892 get_work_cwq(&dwork->work)->wq, &dwork->work);
2894 return flush_work_sync(&dwork->work);
2896 EXPORT_SYMBOL(flush_delayed_work_sync);
2899 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2900 * @dwork: the delayed work cancel
2902 * This is cancel_work_sync() for delayed works.
2905 * %true if @dwork was pending, %false otherwise.
2907 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2909 return __cancel_work_timer(&dwork->work, &dwork->timer);
2911 EXPORT_SYMBOL(cancel_delayed_work_sync);
2914 * schedule_work_on - put work task on a specific cpu
2915 * @cpu: cpu to put the work task on
2916 * @work: job to be done
2918 * This puts a job on a specific cpu
2920 bool schedule_work_on(int cpu, struct work_struct *work)
2922 return queue_work_on(cpu, system_wq, work);
2924 EXPORT_SYMBOL(schedule_work_on);
2927 * schedule_work - put work task in global workqueue
2928 * @work: job to be done
2930 * Returns %false if @work was already on the kernel-global workqueue and
2933 * This puts a job in the kernel-global workqueue if it was not already
2934 * queued and leaves it in the same position on the kernel-global
2935 * workqueue otherwise.
2937 bool schedule_work(struct work_struct *work)
2939 return queue_work(system_wq, work);
2941 EXPORT_SYMBOL(schedule_work);
2944 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2946 * @dwork: job to be done
2947 * @delay: number of jiffies to wait
2949 * After waiting for a given time this puts a job in the kernel-global
2950 * workqueue on the specified CPU.
2952 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
2953 unsigned long delay)
2955 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2957 EXPORT_SYMBOL(schedule_delayed_work_on);
2960 * schedule_delayed_work - put work task in global workqueue after delay
2961 * @dwork: job to be done
2962 * @delay: number of jiffies to wait or 0 for immediate execution
2964 * After waiting for a given time this puts a job in the kernel-global
2967 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
2969 return queue_delayed_work(system_wq, dwork, delay);
2971 EXPORT_SYMBOL(schedule_delayed_work);
2974 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2975 * @func: the function to call
2977 * schedule_on_each_cpu() executes @func on each online CPU using the
2978 * system workqueue and blocks until all CPUs have completed.
2979 * schedule_on_each_cpu() is very slow.
2982 * 0 on success, -errno on failure.
2984 int schedule_on_each_cpu(work_func_t func)
2987 struct work_struct __percpu *works;
2989 works = alloc_percpu(struct work_struct);
2995 for_each_online_cpu(cpu) {
2996 struct work_struct *work = per_cpu_ptr(works, cpu);
2998 INIT_WORK(work, func);
2999 schedule_work_on(cpu, work);
3002 for_each_online_cpu(cpu)
3003 flush_work(per_cpu_ptr(works, cpu));
3011 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3013 * Forces execution of the kernel-global workqueue and blocks until its
3016 * Think twice before calling this function! It's very easy to get into
3017 * trouble if you don't take great care. Either of the following situations
3018 * will lead to deadlock:
3020 * One of the work items currently on the workqueue needs to acquire
3021 * a lock held by your code or its caller.
3023 * Your code is running in the context of a work routine.
3025 * They will be detected by lockdep when they occur, but the first might not
3026 * occur very often. It depends on what work items are on the workqueue and
3027 * what locks they need, which you have no control over.
3029 * In most situations flushing the entire workqueue is overkill; you merely
3030 * need to know that a particular work item isn't queued and isn't running.
3031 * In such cases you should use cancel_delayed_work_sync() or
3032 * cancel_work_sync() instead.
3034 void flush_scheduled_work(void)
3036 flush_workqueue(system_wq);
3038 EXPORT_SYMBOL(flush_scheduled_work);
3041 * execute_in_process_context - reliably execute the routine with user context
3042 * @fn: the function to execute
3043 * @ew: guaranteed storage for the execute work structure (must
3044 * be available when the work executes)
3046 * Executes the function immediately if process context is available,
3047 * otherwise schedules the function for delayed execution.
3049 * Returns: 0 - function was executed
3050 * 1 - function was scheduled for execution
3052 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3054 if (!in_interrupt()) {
3059 INIT_WORK(&ew->work, fn);
3060 schedule_work(&ew->work);
3064 EXPORT_SYMBOL_GPL(execute_in_process_context);
3066 int keventd_up(void)
3068 return system_wq != NULL;
3071 static int alloc_cwqs(struct workqueue_struct *wq)
3074 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3075 * Make sure that the alignment isn't lower than that of
3076 * unsigned long long.
3078 const size_t size = sizeof(struct cpu_workqueue_struct);
3079 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3080 __alignof__(unsigned long long));
3082 if (!(wq->flags & WQ_UNBOUND))
3083 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3088 * Allocate enough room to align cwq and put an extra
3089 * pointer at the end pointing back to the originally
3090 * allocated pointer which will be used for free.
3092 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3094 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3095 *(void **)(wq->cpu_wq.single + 1) = ptr;
3099 /* just in case, make sure it's actually aligned */
3100 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3101 return wq->cpu_wq.v ? 0 : -ENOMEM;
3104 static void free_cwqs(struct workqueue_struct *wq)
3106 if (!(wq->flags & WQ_UNBOUND))
3107 free_percpu(wq->cpu_wq.pcpu);
3108 else if (wq->cpu_wq.single) {
3109 /* the pointer to free is stored right after the cwq */
3110 kfree(*(void **)(wq->cpu_wq.single + 1));
3114 static int wq_clamp_max_active(int max_active, unsigned int flags,
3117 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3119 if (max_active < 1 || max_active > lim)
3120 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
3121 "is out of range, clamping between %d and %d\n",
3122 max_active, name, 1, lim);
3124 return clamp_val(max_active, 1, lim);
3127 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3130 struct lock_class_key *key,
3131 const char *lock_name, ...)
3133 va_list args, args1;
3134 struct workqueue_struct *wq;
3138 /* determine namelen, allocate wq and format name */
3139 va_start(args, lock_name);
3140 va_copy(args1, args);
3141 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3143 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3147 vsnprintf(wq->name, namelen, fmt, args1);
3152 * Workqueues which may be used during memory reclaim should
3153 * have a rescuer to guarantee forward progress.
3155 if (flags & WQ_MEM_RECLAIM)
3156 flags |= WQ_RESCUER;
3158 max_active = max_active ?: WQ_DFL_ACTIVE;
3159 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3163 wq->saved_max_active = max_active;
3164 mutex_init(&wq->flush_mutex);
3165 atomic_set(&wq->nr_cwqs_to_flush, 0);
3166 INIT_LIST_HEAD(&wq->flusher_queue);
3167 INIT_LIST_HEAD(&wq->flusher_overflow);
3169 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3170 INIT_LIST_HEAD(&wq->list);
3172 if (alloc_cwqs(wq) < 0)
3175 for_each_cwq_cpu(cpu, wq) {
3176 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3177 struct global_cwq *gcwq = get_gcwq(cpu);
3178 int pool_idx = (bool)(flags & WQ_HIGHPRI);
3180 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3181 cwq->pool = &gcwq->pools[pool_idx];
3183 cwq->flush_color = -1;
3184 cwq->max_active = max_active;
3185 INIT_LIST_HEAD(&cwq->delayed_works);
3188 if (flags & WQ_RESCUER) {
3189 struct worker *rescuer;
3191 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3194 wq->rescuer = rescuer = alloc_worker();
3198 rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3200 if (IS_ERR(rescuer->task))
3203 rescuer->task->flags |= PF_THREAD_BOUND;
3204 wake_up_process(rescuer->task);
3208 * workqueue_lock protects global freeze state and workqueues
3209 * list. Grab it, set max_active accordingly and add the new
3210 * workqueue to workqueues list.
3212 spin_lock(&workqueue_lock);
3214 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3215 for_each_cwq_cpu(cpu, wq)
3216 get_cwq(cpu, wq)->max_active = 0;
3218 list_add(&wq->list, &workqueues);
3220 spin_unlock(&workqueue_lock);
3226 free_mayday_mask(wq->mayday_mask);
3232 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3235 * destroy_workqueue - safely terminate a workqueue
3236 * @wq: target workqueue
3238 * Safely destroy a workqueue. All work currently pending will be done first.
3240 void destroy_workqueue(struct workqueue_struct *wq)
3244 /* drain it before proceeding with destruction */
3245 drain_workqueue(wq);
3248 * wq list is used to freeze wq, remove from list after
3249 * flushing is complete in case freeze races us.
3251 spin_lock(&workqueue_lock);
3252 list_del(&wq->list);
3253 spin_unlock(&workqueue_lock);
3256 for_each_cwq_cpu(cpu, wq) {
3257 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3260 for (i = 0; i < WORK_NR_COLORS; i++)
3261 BUG_ON(cwq->nr_in_flight[i]);
3262 BUG_ON(cwq->nr_active);
3263 BUG_ON(!list_empty(&cwq->delayed_works));
3266 if (wq->flags & WQ_RESCUER) {
3267 kthread_stop(wq->rescuer->task);
3268 free_mayday_mask(wq->mayday_mask);
3275 EXPORT_SYMBOL_GPL(destroy_workqueue);
3278 * workqueue_set_max_active - adjust max_active of a workqueue
3279 * @wq: target workqueue
3280 * @max_active: new max_active value.
3282 * Set max_active of @wq to @max_active.
3285 * Don't call from IRQ context.
3287 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3291 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3293 spin_lock(&workqueue_lock);
3295 wq->saved_max_active = max_active;
3297 for_each_cwq_cpu(cpu, wq) {
3298 struct global_cwq *gcwq = get_gcwq(cpu);
3300 spin_lock_irq(&gcwq->lock);
3302 if (!(wq->flags & WQ_FREEZABLE) ||
3303 !(gcwq->flags & GCWQ_FREEZING))
3304 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3306 spin_unlock_irq(&gcwq->lock);
3309 spin_unlock(&workqueue_lock);
3311 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3314 * workqueue_congested - test whether a workqueue is congested
3315 * @cpu: CPU in question
3316 * @wq: target workqueue
3318 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3319 * no synchronization around this function and the test result is
3320 * unreliable and only useful as advisory hints or for debugging.
3323 * %true if congested, %false otherwise.
3325 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3327 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3329 return !list_empty(&cwq->delayed_works);
3331 EXPORT_SYMBOL_GPL(workqueue_congested);
3334 * work_cpu - return the last known associated cpu for @work
3335 * @work: the work of interest
3338 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3340 unsigned int work_cpu(struct work_struct *work)
3342 struct global_cwq *gcwq = get_work_gcwq(work);
3344 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3346 EXPORT_SYMBOL_GPL(work_cpu);
3349 * work_busy - test whether a work is currently pending or running
3350 * @work: the work to be tested
3352 * Test whether @work is currently pending or running. There is no
3353 * synchronization around this function and the test result is
3354 * unreliable and only useful as advisory hints or for debugging.
3355 * Especially for reentrant wqs, the pending state might hide the
3359 * OR'd bitmask of WORK_BUSY_* bits.
3361 unsigned int work_busy(struct work_struct *work)
3363 struct global_cwq *gcwq = get_work_gcwq(work);
3364 unsigned long flags;
3365 unsigned int ret = 0;
3370 spin_lock_irqsave(&gcwq->lock, flags);
3372 if (work_pending(work))
3373 ret |= WORK_BUSY_PENDING;
3374 if (find_worker_executing_work(gcwq, work))
3375 ret |= WORK_BUSY_RUNNING;
3377 spin_unlock_irqrestore(&gcwq->lock, flags);
3381 EXPORT_SYMBOL_GPL(work_busy);
3386 * There are two challenges in supporting CPU hotplug. Firstly, there
3387 * are a lot of assumptions on strong associations among work, cwq and
3388 * gcwq which make migrating pending and scheduled works very
3389 * difficult to implement without impacting hot paths. Secondly,
3390 * gcwqs serve mix of short, long and very long running works making
3391 * blocked draining impractical.
3393 * This is solved by allowing a gcwq to be disassociated from the CPU
3394 * running as an unbound one and allowing it to be reattached later if the
3395 * cpu comes back online.
3398 /* claim manager positions of all pools */
3399 static void gcwq_claim_management_and_lock(struct global_cwq *gcwq)
3401 struct worker_pool *pool;
3403 for_each_worker_pool(pool, gcwq)
3404 mutex_lock_nested(&pool->manager_mutex, pool - gcwq->pools);
3405 spin_lock_irq(&gcwq->lock);
3408 /* release manager positions */
3409 static void gcwq_release_management_and_unlock(struct global_cwq *gcwq)
3411 struct worker_pool *pool;
3413 spin_unlock_irq(&gcwq->lock);
3414 for_each_worker_pool(pool, gcwq)
3415 mutex_unlock(&pool->manager_mutex);
3418 static void gcwq_unbind_fn(struct work_struct *work)
3420 struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3421 struct worker_pool *pool;
3422 struct worker *worker;
3423 struct hlist_node *pos;
3426 BUG_ON(gcwq->cpu != smp_processor_id());
3428 gcwq_claim_management_and_lock(gcwq);
3431 * We've claimed all manager positions. Make all workers unbound
3432 * and set DISASSOCIATED. Before this, all workers except for the
3433 * ones which are still executing works from before the last CPU
3434 * down must be on the cpu. After this, they may become diasporas.
3436 for_each_worker_pool(pool, gcwq)
3437 list_for_each_entry(worker, &pool->idle_list, entry)
3438 worker->flags |= WORKER_UNBOUND;
3440 for_each_busy_worker(worker, i, pos, gcwq)
3441 worker->flags |= WORKER_UNBOUND;
3443 gcwq->flags |= GCWQ_DISASSOCIATED;
3445 gcwq_release_management_and_unlock(gcwq);
3448 * Call schedule() so that we cross rq->lock and thus can guarantee
3449 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3450 * as scheduler callbacks may be invoked from other cpus.
3455 * Sched callbacks are disabled now. Zap nr_running. After this,
3456 * nr_running stays zero and need_more_worker() and keep_working()
3457 * are always true as long as the worklist is not empty. @gcwq now
3458 * behaves as unbound (in terms of concurrency management) gcwq
3459 * which is served by workers tied to the CPU.
3461 * On return from this function, the current worker would trigger
3462 * unbound chain execution of pending work items if other workers
3465 for_each_worker_pool(pool, gcwq)
3466 atomic_set(get_pool_nr_running(pool), 0);
3470 * Workqueues should be brought up before normal priority CPU notifiers.
3471 * This will be registered high priority CPU notifier.
3473 static int __devinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3474 unsigned long action,
3477 unsigned int cpu = (unsigned long)hcpu;
3478 struct global_cwq *gcwq = get_gcwq(cpu);
3479 struct worker_pool *pool;
3481 switch (action & ~CPU_TASKS_FROZEN) {
3482 case CPU_UP_PREPARE:
3483 for_each_worker_pool(pool, gcwq) {
3484 struct worker *worker;
3486 if (pool->nr_workers)
3489 worker = create_worker(pool);
3493 spin_lock_irq(&gcwq->lock);
3494 start_worker(worker);
3495 spin_unlock_irq(&gcwq->lock);
3499 case CPU_DOWN_FAILED:
3501 gcwq_claim_management_and_lock(gcwq);
3502 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3503 rebind_workers(gcwq);
3504 gcwq_release_management_and_unlock(gcwq);
3511 * Workqueues should be brought down after normal priority CPU notifiers.
3512 * This will be registered as low priority CPU notifier.
3514 static int __devinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3515 unsigned long action,
3518 unsigned int cpu = (unsigned long)hcpu;
3519 struct work_struct unbind_work;
3521 switch (action & ~CPU_TASKS_FROZEN) {
3522 case CPU_DOWN_PREPARE:
3523 /* unbinding should happen on the local CPU */
3524 INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3525 schedule_work_on(cpu, &unbind_work);
3526 flush_work(&unbind_work);
3534 struct work_for_cpu {
3535 struct completion completion;
3541 static int do_work_for_cpu(void *_wfc)
3543 struct work_for_cpu *wfc = _wfc;
3544 wfc->ret = wfc->fn(wfc->arg);
3545 complete(&wfc->completion);
3550 * work_on_cpu - run a function in user context on a particular cpu
3551 * @cpu: the cpu to run on
3552 * @fn: the function to run
3553 * @arg: the function arg
3555 * This will return the value @fn returns.
3556 * It is up to the caller to ensure that the cpu doesn't go offline.
3557 * The caller must not hold any locks which would prevent @fn from completing.
3559 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3561 struct task_struct *sub_thread;
3562 struct work_for_cpu wfc = {
3563 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3568 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3569 if (IS_ERR(sub_thread))
3570 return PTR_ERR(sub_thread);
3571 kthread_bind(sub_thread, cpu);
3572 wake_up_process(sub_thread);
3573 wait_for_completion(&wfc.completion);
3576 EXPORT_SYMBOL_GPL(work_on_cpu);
3577 #endif /* CONFIG_SMP */
3579 #ifdef CONFIG_FREEZER
3582 * freeze_workqueues_begin - begin freezing workqueues
3584 * Start freezing workqueues. After this function returns, all freezable
3585 * workqueues will queue new works to their frozen_works list instead of
3589 * Grabs and releases workqueue_lock and gcwq->lock's.
3591 void freeze_workqueues_begin(void)
3595 spin_lock(&workqueue_lock);
3597 BUG_ON(workqueue_freezing);
3598 workqueue_freezing = true;
3600 for_each_gcwq_cpu(cpu) {
3601 struct global_cwq *gcwq = get_gcwq(cpu);
3602 struct workqueue_struct *wq;
3604 spin_lock_irq(&gcwq->lock);
3606 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3607 gcwq->flags |= GCWQ_FREEZING;
3609 list_for_each_entry(wq, &workqueues, list) {
3610 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3612 if (cwq && wq->flags & WQ_FREEZABLE)
3613 cwq->max_active = 0;
3616 spin_unlock_irq(&gcwq->lock);
3619 spin_unlock(&workqueue_lock);
3623 * freeze_workqueues_busy - are freezable workqueues still busy?
3625 * Check whether freezing is complete. This function must be called
3626 * between freeze_workqueues_begin() and thaw_workqueues().
3629 * Grabs and releases workqueue_lock.
3632 * %true if some freezable workqueues are still busy. %false if freezing
3635 bool freeze_workqueues_busy(void)
3640 spin_lock(&workqueue_lock);
3642 BUG_ON(!workqueue_freezing);
3644 for_each_gcwq_cpu(cpu) {
3645 struct workqueue_struct *wq;
3647 * nr_active is monotonically decreasing. It's safe
3648 * to peek without lock.
3650 list_for_each_entry(wq, &workqueues, list) {
3651 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3653 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3656 BUG_ON(cwq->nr_active < 0);
3657 if (cwq->nr_active) {
3664 spin_unlock(&workqueue_lock);
3669 * thaw_workqueues - thaw workqueues
3671 * Thaw workqueues. Normal queueing is restored and all collected
3672 * frozen works are transferred to their respective gcwq worklists.
3675 * Grabs and releases workqueue_lock and gcwq->lock's.
3677 void thaw_workqueues(void)
3681 spin_lock(&workqueue_lock);
3683 if (!workqueue_freezing)
3686 for_each_gcwq_cpu(cpu) {
3687 struct global_cwq *gcwq = get_gcwq(cpu);
3688 struct worker_pool *pool;
3689 struct workqueue_struct *wq;
3691 spin_lock_irq(&gcwq->lock);
3693 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3694 gcwq->flags &= ~GCWQ_FREEZING;
3696 list_for_each_entry(wq, &workqueues, list) {
3697 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3699 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3702 /* restore max_active and repopulate worklist */
3703 cwq->max_active = wq->saved_max_active;
3705 while (!list_empty(&cwq->delayed_works) &&
3706 cwq->nr_active < cwq->max_active)
3707 cwq_activate_first_delayed(cwq);
3710 for_each_worker_pool(pool, gcwq)
3711 wake_up_worker(pool);
3713 spin_unlock_irq(&gcwq->lock);
3716 workqueue_freezing = false;
3718 spin_unlock(&workqueue_lock);
3720 #endif /* CONFIG_FREEZER */
3722 static int __init init_workqueues(void)
3727 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3728 cpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3730 /* initialize gcwqs */
3731 for_each_gcwq_cpu(cpu) {
3732 struct global_cwq *gcwq = get_gcwq(cpu);
3733 struct worker_pool *pool;
3735 spin_lock_init(&gcwq->lock);
3737 gcwq->flags |= GCWQ_DISASSOCIATED;
3739 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3740 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3742 for_each_worker_pool(pool, gcwq) {
3744 INIT_LIST_HEAD(&pool->worklist);
3745 INIT_LIST_HEAD(&pool->idle_list);
3747 init_timer_deferrable(&pool->idle_timer);
3748 pool->idle_timer.function = idle_worker_timeout;
3749 pool->idle_timer.data = (unsigned long)pool;
3751 setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3752 (unsigned long)pool);
3754 mutex_init(&pool->manager_mutex);
3755 ida_init(&pool->worker_ida);
3758 init_waitqueue_head(&gcwq->rebind_hold);
3761 /* create the initial worker */
3762 for_each_online_gcwq_cpu(cpu) {
3763 struct global_cwq *gcwq = get_gcwq(cpu);
3764 struct worker_pool *pool;
3766 if (cpu != WORK_CPU_UNBOUND)
3767 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3769 for_each_worker_pool(pool, gcwq) {
3770 struct worker *worker;
3772 worker = create_worker(pool);
3774 spin_lock_irq(&gcwq->lock);
3775 start_worker(worker);
3776 spin_unlock_irq(&gcwq->lock);
3780 system_wq = alloc_workqueue("events", 0, 0);
3781 system_long_wq = alloc_workqueue("events_long", 0, 0);
3782 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3783 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3784 WQ_UNBOUND_MAX_ACTIVE);
3785 system_freezable_wq = alloc_workqueue("events_freezable",
3787 system_nrt_freezable_wq = alloc_workqueue("events_nrt_freezable",
3788 WQ_NON_REENTRANT | WQ_FREEZABLE, 0);
3789 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
3790 !system_unbound_wq || !system_freezable_wq ||
3791 !system_nrt_freezable_wq);
3794 early_initcall(init_workqueues);