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"
48 /* global_cwq flags */
49 GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
50 GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */
51 GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
52 GCWQ_FREEZING = 1 << 3, /* freeze in progress */
53 GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */
56 WORKER_STARTED = 1 << 0, /* started */
57 WORKER_DIE = 1 << 1, /* die die die */
58 WORKER_IDLE = 1 << 2, /* is idle */
59 WORKER_PREP = 1 << 3, /* preparing to run works */
60 WORKER_ROGUE = 1 << 4, /* not bound to any cpu */
61 WORKER_REBIND = 1 << 5, /* mom is home, come back */
62 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
63 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
65 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND |
66 WORKER_CPU_INTENSIVE | WORKER_UNBOUND,
68 /* gcwq->trustee_state */
69 TRUSTEE_START = 0, /* start */
70 TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */
71 TRUSTEE_BUTCHER = 2, /* butcher workers */
72 TRUSTEE_RELEASE = 3, /* release workers */
73 TRUSTEE_DONE = 4, /* trustee is done */
75 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
76 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
77 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
79 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
80 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
82 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
83 /* call for help after 10ms
85 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
86 CREATE_COOLDOWN = HZ, /* time to breath after fail */
87 TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
90 * Rescue workers are used only on emergencies and shared by
93 RESCUER_NICE_LEVEL = -20,
97 * Structure fields follow one of the following exclusion rules.
99 * I: Modifiable by initialization/destruction paths and read-only for
102 * P: Preemption protected. Disabling preemption is enough and should
103 * only be modified and accessed from the local cpu.
105 * L: gcwq->lock protected. Access with gcwq->lock held.
107 * X: During normal operation, modification requires gcwq->lock and
108 * should be done only from local cpu. Either disabling preemption
109 * on local cpu or grabbing gcwq->lock is enough for read access.
110 * If GCWQ_DISASSOCIATED is set, it's identical to L.
112 * F: wq->flush_mutex protected.
114 * W: workqueue_lock protected.
120 * The poor guys doing the actual heavy lifting. All on-duty workers
121 * are either serving the manager role, on idle list or on busy hash.
124 /* on idle list while idle, on busy hash table while busy */
126 struct list_head entry; /* L: while idle */
127 struct hlist_node hentry; /* L: while busy */
130 struct work_struct *current_work; /* L: work being processed */
131 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
132 struct list_head scheduled; /* L: scheduled works */
133 struct task_struct *task; /* I: worker task */
134 struct global_cwq *gcwq; /* I: the associated gcwq */
135 /* 64 bytes boundary on 64bit, 32 on 32bit */
136 unsigned long last_active; /* L: last active timestamp */
137 unsigned int flags; /* X: flags */
138 int id; /* I: worker id */
139 struct work_struct rebind_work; /* L: rebind worker to cpu */
143 * Global per-cpu workqueue. There's one and only one for each cpu
144 * and all works are queued and processed here regardless of their
148 spinlock_t lock; /* the gcwq lock */
149 struct list_head worklist; /* L: list of pending works */
150 unsigned int cpu; /* I: the associated cpu */
151 unsigned int flags; /* L: GCWQ_* flags */
153 int nr_workers; /* L: total number of workers */
154 int nr_idle; /* L: currently idle ones */
156 /* workers are chained either in the idle_list or busy_hash */
157 struct list_head idle_list; /* X: list of idle workers */
158 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
159 /* L: hash of busy workers */
161 struct timer_list idle_timer; /* L: worker idle timeout */
162 struct timer_list mayday_timer; /* L: SOS timer for dworkers */
164 struct ida worker_ida; /* L: for worker IDs */
166 struct task_struct *trustee; /* L: for gcwq shutdown */
167 unsigned int trustee_state; /* L: trustee state */
168 wait_queue_head_t trustee_wait; /* trustee wait */
169 struct worker *first_idle; /* L: first idle worker */
170 } ____cacheline_aligned_in_smp;
173 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
174 * work_struct->data are used for flags and thus cwqs need to be
175 * aligned at two's power of the number of flag bits.
177 struct cpu_workqueue_struct {
178 struct global_cwq *gcwq; /* I: the associated gcwq */
179 struct workqueue_struct *wq; /* I: the owning workqueue */
180 int work_color; /* L: current color */
181 int flush_color; /* L: flushing color */
182 int nr_in_flight[WORK_NR_COLORS];
183 /* L: nr of in_flight works */
184 int nr_active; /* L: nr of active works */
185 int max_active; /* L: max active works */
186 struct list_head delayed_works; /* L: delayed works */
190 * Structure used to wait for workqueue flush.
193 struct list_head list; /* F: list of flushers */
194 int flush_color; /* F: flush color waiting for */
195 struct completion done; /* flush completion */
199 * All cpumasks are assumed to be always set on UP and thus can't be
200 * used to determine whether there's something to be done.
203 typedef cpumask_var_t mayday_mask_t;
204 #define mayday_test_and_set_cpu(cpu, mask) \
205 cpumask_test_and_set_cpu((cpu), (mask))
206 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
207 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
208 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
209 #define free_mayday_mask(mask) free_cpumask_var((mask))
211 typedef unsigned long mayday_mask_t;
212 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
213 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
214 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
215 #define alloc_mayday_mask(maskp, gfp) true
216 #define free_mayday_mask(mask) do { } while (0)
220 * The externally visible workqueue abstraction is an array of
221 * per-CPU workqueues:
223 struct workqueue_struct {
224 unsigned int flags; /* W: WQ_* flags */
226 struct cpu_workqueue_struct __percpu *pcpu;
227 struct cpu_workqueue_struct *single;
229 } cpu_wq; /* I: cwq's */
230 struct list_head list; /* W: list of all workqueues */
232 struct mutex flush_mutex; /* protects wq flushing */
233 int work_color; /* F: current work color */
234 int flush_color; /* F: current flush color */
235 atomic_t nr_cwqs_to_flush; /* flush in progress */
236 struct wq_flusher *first_flusher; /* F: first flusher */
237 struct list_head flusher_queue; /* F: flush waiters */
238 struct list_head flusher_overflow; /* F: flush overflow list */
240 mayday_mask_t mayday_mask; /* cpus requesting rescue */
241 struct worker *rescuer; /* I: rescue worker */
243 int nr_drainers; /* W: drain in progress */
244 int saved_max_active; /* W: saved cwq max_active */
245 #ifdef CONFIG_LOCKDEP
246 struct lockdep_map lockdep_map;
248 char name[]; /* I: workqueue name */
251 struct workqueue_struct *system_wq __read_mostly;
252 struct workqueue_struct *system_long_wq __read_mostly;
253 struct workqueue_struct *system_nrt_wq __read_mostly;
254 struct workqueue_struct *system_unbound_wq __read_mostly;
255 struct workqueue_struct *system_freezable_wq __read_mostly;
256 struct workqueue_struct *system_nrt_freezable_wq __read_mostly;
257 EXPORT_SYMBOL_GPL(system_wq);
258 EXPORT_SYMBOL_GPL(system_long_wq);
259 EXPORT_SYMBOL_GPL(system_nrt_wq);
260 EXPORT_SYMBOL_GPL(system_unbound_wq);
261 EXPORT_SYMBOL_GPL(system_freezable_wq);
262 EXPORT_SYMBOL_GPL(system_nrt_freezable_wq);
264 #define CREATE_TRACE_POINTS
265 #include <trace/events/workqueue.h>
267 #define for_each_busy_worker(worker, i, pos, gcwq) \
268 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
269 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
271 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
274 if (cpu < nr_cpu_ids) {
276 cpu = cpumask_next(cpu, mask);
277 if (cpu < nr_cpu_ids)
281 return WORK_CPU_UNBOUND;
283 return WORK_CPU_NONE;
286 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
287 struct workqueue_struct *wq)
289 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
295 * An extra gcwq is defined for an invalid cpu number
296 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
297 * specific CPU. The following iterators are similar to
298 * for_each_*_cpu() iterators but also considers the unbound gcwq.
300 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
301 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
302 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
303 * WORK_CPU_UNBOUND for unbound workqueues
305 #define for_each_gcwq_cpu(cpu) \
306 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
307 (cpu) < WORK_CPU_NONE; \
308 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
310 #define for_each_online_gcwq_cpu(cpu) \
311 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
312 (cpu) < WORK_CPU_NONE; \
313 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
315 #define for_each_cwq_cpu(cpu, wq) \
316 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
317 (cpu) < WORK_CPU_NONE; \
318 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
320 #ifdef CONFIG_DEBUG_OBJECTS_WORK
322 static struct debug_obj_descr work_debug_descr;
324 static void *work_debug_hint(void *addr)
326 return ((struct work_struct *) addr)->func;
330 * fixup_init is called when:
331 * - an active object is initialized
333 static int work_fixup_init(void *addr, enum debug_obj_state state)
335 struct work_struct *work = addr;
338 case ODEBUG_STATE_ACTIVE:
339 cancel_work_sync(work);
340 debug_object_init(work, &work_debug_descr);
348 * fixup_activate is called when:
349 * - an active object is activated
350 * - an unknown object is activated (might be a statically initialized object)
352 static int work_fixup_activate(void *addr, enum debug_obj_state state)
354 struct work_struct *work = addr;
358 case ODEBUG_STATE_NOTAVAILABLE:
360 * This is not really a fixup. The work struct was
361 * statically initialized. We just make sure that it
362 * is tracked in the object tracker.
364 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
365 debug_object_init(work, &work_debug_descr);
366 debug_object_activate(work, &work_debug_descr);
372 case ODEBUG_STATE_ACTIVE:
381 * fixup_free is called when:
382 * - an active object is freed
384 static int work_fixup_free(void *addr, enum debug_obj_state state)
386 struct work_struct *work = addr;
389 case ODEBUG_STATE_ACTIVE:
390 cancel_work_sync(work);
391 debug_object_free(work, &work_debug_descr);
398 static struct debug_obj_descr work_debug_descr = {
399 .name = "work_struct",
400 .debug_hint = work_debug_hint,
401 .fixup_init = work_fixup_init,
402 .fixup_activate = work_fixup_activate,
403 .fixup_free = work_fixup_free,
406 static inline void debug_work_activate(struct work_struct *work)
408 debug_object_activate(work, &work_debug_descr);
411 static inline void debug_work_deactivate(struct work_struct *work)
413 debug_object_deactivate(work, &work_debug_descr);
416 void __init_work(struct work_struct *work, int onstack)
419 debug_object_init_on_stack(work, &work_debug_descr);
421 debug_object_init(work, &work_debug_descr);
423 EXPORT_SYMBOL_GPL(__init_work);
425 void destroy_work_on_stack(struct work_struct *work)
427 debug_object_free(work, &work_debug_descr);
429 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
432 static inline void debug_work_activate(struct work_struct *work) { }
433 static inline void debug_work_deactivate(struct work_struct *work) { }
436 /* Serializes the accesses to the list of workqueues. */
437 static DEFINE_SPINLOCK(workqueue_lock);
438 static LIST_HEAD(workqueues);
439 static bool workqueue_freezing; /* W: have wqs started freezing? */
442 * The almighty global cpu workqueues. nr_running is the only field
443 * which is expected to be used frequently by other cpus via
444 * try_to_wake_up(). Put it in a separate cacheline.
446 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
447 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
450 * Global cpu workqueue and nr_running counter for unbound gcwq. The
451 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
452 * workers have WORKER_UNBOUND set.
454 static struct global_cwq unbound_global_cwq;
455 static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
457 static int worker_thread(void *__worker);
459 static struct global_cwq *get_gcwq(unsigned int cpu)
461 if (cpu != WORK_CPU_UNBOUND)
462 return &per_cpu(global_cwq, cpu);
464 return &unbound_global_cwq;
467 static atomic_t *get_gcwq_nr_running(unsigned int cpu)
469 if (cpu != WORK_CPU_UNBOUND)
470 return &per_cpu(gcwq_nr_running, cpu);
472 return &unbound_gcwq_nr_running;
475 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
476 struct workqueue_struct *wq)
478 if (!(wq->flags & WQ_UNBOUND)) {
479 if (likely(cpu < nr_cpu_ids))
480 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
481 } else if (likely(cpu == WORK_CPU_UNBOUND))
482 return wq->cpu_wq.single;
486 static unsigned int work_color_to_flags(int color)
488 return color << WORK_STRUCT_COLOR_SHIFT;
491 static int get_work_color(struct work_struct *work)
493 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
494 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
497 static int work_next_color(int color)
499 return (color + 1) % WORK_NR_COLORS;
503 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
504 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
505 * cleared and the work data contains the cpu number it was last on.
507 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
508 * cwq, cpu or clear work->data. These functions should only be
509 * called while the work is owned - ie. while the PENDING bit is set.
511 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
512 * corresponding to a work. gcwq is available once the work has been
513 * queued anywhere after initialization. cwq is available only from
514 * queueing until execution starts.
516 static inline void set_work_data(struct work_struct *work, unsigned long data,
519 BUG_ON(!work_pending(work));
520 atomic_long_set(&work->data, data | flags | work_static(work));
523 static void set_work_cwq(struct work_struct *work,
524 struct cpu_workqueue_struct *cwq,
525 unsigned long extra_flags)
527 set_work_data(work, (unsigned long)cwq,
528 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
531 static void set_work_cpu(struct work_struct *work, unsigned int cpu)
533 set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
536 static void clear_work_data(struct work_struct *work)
538 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
541 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
543 unsigned long data = atomic_long_read(&work->data);
545 if (data & WORK_STRUCT_CWQ)
546 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
551 static struct global_cwq *get_work_gcwq(struct work_struct *work)
553 unsigned long data = atomic_long_read(&work->data);
556 if (data & WORK_STRUCT_CWQ)
557 return ((struct cpu_workqueue_struct *)
558 (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
560 cpu = data >> WORK_STRUCT_FLAG_BITS;
561 if (cpu == WORK_CPU_NONE)
564 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
565 return get_gcwq(cpu);
569 * Policy functions. These define the policies on how the global
570 * worker pool is managed. Unless noted otherwise, these functions
571 * assume that they're being called with gcwq->lock held.
574 static bool __need_more_worker(struct global_cwq *gcwq)
576 return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
577 gcwq->flags & GCWQ_HIGHPRI_PENDING;
581 * Need to wake up a worker? Called from anything but currently
584 static bool need_more_worker(struct global_cwq *gcwq)
586 return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
589 /* Can I start working? Called from busy but !running workers. */
590 static bool may_start_working(struct global_cwq *gcwq)
592 return gcwq->nr_idle;
595 /* Do I need to keep working? Called from currently running workers. */
596 static bool keep_working(struct global_cwq *gcwq)
598 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
600 return !list_empty(&gcwq->worklist) &&
601 (atomic_read(nr_running) <= 1 ||
602 gcwq->flags & GCWQ_HIGHPRI_PENDING);
605 /* Do we need a new worker? Called from manager. */
606 static bool need_to_create_worker(struct global_cwq *gcwq)
608 return need_more_worker(gcwq) && !may_start_working(gcwq);
611 /* Do I need to be the manager? */
612 static bool need_to_manage_workers(struct global_cwq *gcwq)
614 return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
617 /* Do we have too many workers and should some go away? */
618 static bool too_many_workers(struct global_cwq *gcwq)
620 bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
621 int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
622 int nr_busy = gcwq->nr_workers - nr_idle;
624 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
631 /* Return the first worker. Safe with preemption disabled */
632 static struct worker *first_worker(struct global_cwq *gcwq)
634 if (unlikely(list_empty(&gcwq->idle_list)))
637 return list_first_entry(&gcwq->idle_list, struct worker, entry);
641 * wake_up_worker - wake up an idle worker
642 * @gcwq: gcwq to wake worker for
644 * Wake up the first idle worker of @gcwq.
647 * spin_lock_irq(gcwq->lock).
649 static void wake_up_worker(struct global_cwq *gcwq)
651 struct worker *worker = first_worker(gcwq);
654 wake_up_process(worker->task);
658 * wq_worker_waking_up - a worker is waking up
659 * @task: task waking up
660 * @cpu: CPU @task is waking up to
662 * This function is called during try_to_wake_up() when a worker is
666 * spin_lock_irq(rq->lock)
668 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
670 struct worker *worker = kthread_data(task);
672 if (!(worker->flags & WORKER_NOT_RUNNING))
673 atomic_inc(get_gcwq_nr_running(cpu));
677 * wq_worker_sleeping - a worker is going to sleep
678 * @task: task going to sleep
679 * @cpu: CPU in question, must be the current CPU number
681 * This function is called during schedule() when a busy worker is
682 * going to sleep. Worker on the same cpu can be woken up by
683 * returning pointer to its task.
686 * spin_lock_irq(rq->lock)
689 * Worker task on @cpu to wake up, %NULL if none.
691 struct task_struct *wq_worker_sleeping(struct task_struct *task,
694 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
695 struct global_cwq *gcwq = get_gcwq(cpu);
696 atomic_t *nr_running = get_gcwq_nr_running(cpu);
698 if (worker->flags & WORKER_NOT_RUNNING)
701 /* this can only happen on the local cpu */
702 BUG_ON(cpu != raw_smp_processor_id());
705 * The counterpart of the following dec_and_test, implied mb,
706 * worklist not empty test sequence is in insert_work().
707 * Please read comment there.
709 * NOT_RUNNING is clear. This means that trustee is not in
710 * charge and we're running on the local cpu w/ rq lock held
711 * and preemption disabled, which in turn means that none else
712 * could be manipulating idle_list, so dereferencing idle_list
713 * without gcwq lock is safe.
715 if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
716 to_wakeup = first_worker(gcwq);
717 return to_wakeup ? to_wakeup->task : NULL;
721 * worker_set_flags - set worker flags and adjust nr_running accordingly
723 * @flags: flags to set
724 * @wakeup: wakeup an idle worker if necessary
726 * Set @flags in @worker->flags and adjust nr_running accordingly. If
727 * nr_running becomes zero and @wakeup is %true, an idle worker is
731 * spin_lock_irq(gcwq->lock)
733 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
736 struct global_cwq *gcwq = worker->gcwq;
738 WARN_ON_ONCE(worker->task != current);
741 * If transitioning into NOT_RUNNING, adjust nr_running and
742 * wake up an idle worker as necessary if requested by
745 if ((flags & WORKER_NOT_RUNNING) &&
746 !(worker->flags & WORKER_NOT_RUNNING)) {
747 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
750 if (atomic_dec_and_test(nr_running) &&
751 !list_empty(&gcwq->worklist))
752 wake_up_worker(gcwq);
754 atomic_dec(nr_running);
757 worker->flags |= flags;
761 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
763 * @flags: flags to clear
765 * Clear @flags in @worker->flags and adjust nr_running accordingly.
768 * spin_lock_irq(gcwq->lock)
770 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
772 struct global_cwq *gcwq = worker->gcwq;
773 unsigned int oflags = worker->flags;
775 WARN_ON_ONCE(worker->task != current);
777 worker->flags &= ~flags;
780 * If transitioning out of NOT_RUNNING, increment nr_running. Note
781 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
782 * of multiple flags, not a single flag.
784 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
785 if (!(worker->flags & WORKER_NOT_RUNNING))
786 atomic_inc(get_gcwq_nr_running(gcwq->cpu));
790 * busy_worker_head - return the busy hash head for a work
791 * @gcwq: gcwq of interest
792 * @work: work to be hashed
794 * Return hash head of @gcwq for @work.
797 * spin_lock_irq(gcwq->lock).
800 * Pointer to the hash head.
802 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
803 struct work_struct *work)
805 const int base_shift = ilog2(sizeof(struct work_struct));
806 unsigned long v = (unsigned long)work;
808 /* simple shift and fold hash, do we need something better? */
810 v += v >> BUSY_WORKER_HASH_ORDER;
811 v &= BUSY_WORKER_HASH_MASK;
813 return &gcwq->busy_hash[v];
817 * __find_worker_executing_work - find worker which is executing a work
818 * @gcwq: gcwq of interest
819 * @bwh: hash head as returned by busy_worker_head()
820 * @work: work to find worker for
822 * Find a worker which is executing @work on @gcwq. @bwh should be
823 * the hash head obtained by calling busy_worker_head() with the same
827 * spin_lock_irq(gcwq->lock).
830 * Pointer to worker which is executing @work if found, NULL
833 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
834 struct hlist_head *bwh,
835 struct work_struct *work)
837 struct worker *worker;
838 struct hlist_node *tmp;
840 hlist_for_each_entry(worker, tmp, bwh, hentry)
841 if (worker->current_work == work)
847 * find_worker_executing_work - find worker which is executing a work
848 * @gcwq: gcwq of interest
849 * @work: work to find worker for
851 * Find a worker which is executing @work on @gcwq. This function is
852 * identical to __find_worker_executing_work() except that this
853 * function calculates @bwh itself.
856 * spin_lock_irq(gcwq->lock).
859 * Pointer to worker which is executing @work if found, NULL
862 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
863 struct work_struct *work)
865 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
870 * gcwq_determine_ins_pos - find insertion position
871 * @gcwq: gcwq of interest
872 * @cwq: cwq a work is being queued for
874 * A work for @cwq is about to be queued on @gcwq, determine insertion
875 * position for the work. If @cwq is for HIGHPRI wq, the work is
876 * queued at the head of the queue but in FIFO order with respect to
877 * other HIGHPRI works; otherwise, at the end of the queue. This
878 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
879 * there are HIGHPRI works pending.
882 * spin_lock_irq(gcwq->lock).
885 * Pointer to inserstion position.
887 static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
888 struct cpu_workqueue_struct *cwq)
890 struct work_struct *twork;
892 if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
893 return &gcwq->worklist;
895 list_for_each_entry(twork, &gcwq->worklist, entry) {
896 struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
898 if (!(tcwq->wq->flags & WQ_HIGHPRI))
902 gcwq->flags |= GCWQ_HIGHPRI_PENDING;
903 return &twork->entry;
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 global_cwq *gcwq = cwq->gcwq;
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(gcwq))
944 wake_up_worker(gcwq);
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 debug_work_activate(work);
990 /* if dying, only works from the same workqueue are allowed */
991 if (unlikely(wq->flags & WQ_DRAINING) &&
992 WARN_ON_ONCE(!is_chained_work(wq)))
995 /* determine gcwq to use */
996 if (!(wq->flags & WQ_UNBOUND)) {
997 struct global_cwq *last_gcwq;
999 if (unlikely(cpu == WORK_CPU_UNBOUND))
1000 cpu = raw_smp_processor_id();
1003 * It's multi cpu. If @wq is non-reentrant and @work
1004 * was previously on a different cpu, it might still
1005 * be running there, in which case the work needs to
1006 * be queued on that cpu to guarantee non-reentrance.
1008 gcwq = get_gcwq(cpu);
1009 if (wq->flags & WQ_NON_REENTRANT &&
1010 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1011 struct worker *worker;
1013 spin_lock_irqsave(&last_gcwq->lock, flags);
1015 worker = find_worker_executing_work(last_gcwq, work);
1017 if (worker && worker->current_cwq->wq == wq)
1020 /* meh... not running there, queue here */
1021 spin_unlock_irqrestore(&last_gcwq->lock, flags);
1022 spin_lock_irqsave(&gcwq->lock, flags);
1025 spin_lock_irqsave(&gcwq->lock, flags);
1027 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1028 spin_lock_irqsave(&gcwq->lock, flags);
1031 /* gcwq determined, get cwq and queue */
1032 cwq = get_cwq(gcwq->cpu, wq);
1033 trace_workqueue_queue_work(cpu, cwq, work);
1035 if (WARN_ON(!list_empty(&work->entry))) {
1036 spin_unlock_irqrestore(&gcwq->lock, flags);
1040 cwq->nr_in_flight[cwq->work_color]++;
1041 work_flags = work_color_to_flags(cwq->work_color);
1043 if (likely(cwq->nr_active < cwq->max_active)) {
1044 trace_workqueue_activate_work(work);
1046 worklist = gcwq_determine_ins_pos(gcwq, cwq);
1048 work_flags |= WORK_STRUCT_DELAYED;
1049 worklist = &cwq->delayed_works;
1052 insert_work(cwq, work, worklist, work_flags);
1054 spin_unlock_irqrestore(&gcwq->lock, flags);
1058 * queue_work - queue work on a workqueue
1059 * @wq: workqueue to use
1060 * @work: work to queue
1062 * Returns 0 if @work was already on a queue, non-zero otherwise.
1064 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1065 * it can be processed by another CPU.
1067 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1071 ret = queue_work_on(get_cpu(), wq, work);
1076 EXPORT_SYMBOL_GPL(queue_work);
1079 * queue_work_on - queue work on specific cpu
1080 * @cpu: CPU number to execute work on
1081 * @wq: workqueue to use
1082 * @work: work to queue
1084 * Returns 0 if @work was already on a queue, non-zero otherwise.
1086 * We queue the work to a specific CPU, the caller must ensure it
1090 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1094 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1095 __queue_work(cpu, wq, work);
1100 EXPORT_SYMBOL_GPL(queue_work_on);
1102 static void delayed_work_timer_fn(unsigned long __data)
1104 struct delayed_work *dwork = (struct delayed_work *)__data;
1105 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1107 __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1111 * queue_delayed_work - queue work on a workqueue after delay
1112 * @wq: workqueue to use
1113 * @dwork: delayable work to queue
1114 * @delay: number of jiffies to wait before queueing
1116 * Returns 0 if @work was already on a queue, non-zero otherwise.
1118 int queue_delayed_work(struct workqueue_struct *wq,
1119 struct delayed_work *dwork, unsigned long delay)
1122 return queue_work(wq, &dwork->work);
1124 return queue_delayed_work_on(-1, wq, dwork, delay);
1126 EXPORT_SYMBOL_GPL(queue_delayed_work);
1129 * queue_delayed_work_on - queue work on specific CPU after delay
1130 * @cpu: CPU number to execute work on
1131 * @wq: workqueue to use
1132 * @dwork: work to queue
1133 * @delay: number of jiffies to wait before queueing
1135 * Returns 0 if @work was already on a queue, non-zero otherwise.
1137 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1138 struct delayed_work *dwork, unsigned long delay)
1141 struct timer_list *timer = &dwork->timer;
1142 struct work_struct *work = &dwork->work;
1144 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1147 BUG_ON(timer_pending(timer));
1148 BUG_ON(!list_empty(&work->entry));
1150 timer_stats_timer_set_start_info(&dwork->timer);
1153 * This stores cwq for the moment, for the timer_fn.
1154 * Note that the work's gcwq is preserved to allow
1155 * reentrance detection for delayed works.
1157 if (!(wq->flags & WQ_UNBOUND)) {
1158 struct global_cwq *gcwq = get_work_gcwq(work);
1160 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1163 lcpu = raw_smp_processor_id();
1165 lcpu = WORK_CPU_UNBOUND;
1167 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1169 timer->expires = jiffies + delay;
1170 timer->data = (unsigned long)dwork;
1171 timer->function = delayed_work_timer_fn;
1173 if (unlikely(cpu >= 0))
1174 add_timer_on(timer, cpu);
1181 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1184 * worker_enter_idle - enter idle state
1185 * @worker: worker which is entering idle state
1187 * @worker is entering idle state. Update stats and idle timer if
1191 * spin_lock_irq(gcwq->lock).
1193 static void worker_enter_idle(struct worker *worker)
1195 struct global_cwq *gcwq = worker->gcwq;
1197 BUG_ON(worker->flags & WORKER_IDLE);
1198 BUG_ON(!list_empty(&worker->entry) &&
1199 (worker->hentry.next || worker->hentry.pprev));
1201 /* can't use worker_set_flags(), also called from start_worker() */
1202 worker->flags |= WORKER_IDLE;
1204 worker->last_active = jiffies;
1206 /* idle_list is LIFO */
1207 list_add(&worker->entry, &gcwq->idle_list);
1209 if (likely(!(worker->flags & WORKER_ROGUE))) {
1210 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1211 mod_timer(&gcwq->idle_timer,
1212 jiffies + IDLE_WORKER_TIMEOUT);
1214 wake_up_all(&gcwq->trustee_wait);
1216 /* sanity check nr_running */
1217 WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
1218 atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1222 * worker_leave_idle - leave idle state
1223 * @worker: worker which is leaving idle state
1225 * @worker is leaving idle state. Update stats.
1228 * spin_lock_irq(gcwq->lock).
1230 static void worker_leave_idle(struct worker *worker)
1232 struct global_cwq *gcwq = worker->gcwq;
1234 BUG_ON(!(worker->flags & WORKER_IDLE));
1235 worker_clr_flags(worker, WORKER_IDLE);
1237 list_del_init(&worker->entry);
1241 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1244 * Works which are scheduled while the cpu is online must at least be
1245 * scheduled to a worker which is bound to the cpu so that if they are
1246 * flushed from cpu callbacks while cpu is going down, they are
1247 * guaranteed to execute on the cpu.
1249 * This function is to be used by rogue workers and rescuers to bind
1250 * themselves to the target cpu and may race with cpu going down or
1251 * coming online. kthread_bind() can't be used because it may put the
1252 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1253 * verbatim as it's best effort and blocking and gcwq may be
1254 * [dis]associated in the meantime.
1256 * This function tries set_cpus_allowed() and locks gcwq and verifies
1257 * the binding against GCWQ_DISASSOCIATED which is set during
1258 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1259 * idle state or fetches works without dropping lock, it can guarantee
1260 * the scheduling requirement described in the first paragraph.
1263 * Might sleep. Called without any lock but returns with gcwq->lock
1267 * %true if the associated gcwq is online (@worker is successfully
1268 * bound), %false if offline.
1270 static bool worker_maybe_bind_and_lock(struct worker *worker)
1271 __acquires(&gcwq->lock)
1273 struct global_cwq *gcwq = worker->gcwq;
1274 struct task_struct *task = worker->task;
1278 * The following call may fail, succeed or succeed
1279 * without actually migrating the task to the cpu if
1280 * it races with cpu hotunplug operation. Verify
1281 * against GCWQ_DISASSOCIATED.
1283 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1284 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1286 spin_lock_irq(&gcwq->lock);
1287 if (gcwq->flags & GCWQ_DISASSOCIATED)
1289 if (task_cpu(task) == gcwq->cpu &&
1290 cpumask_equal(¤t->cpus_allowed,
1291 get_cpu_mask(gcwq->cpu)))
1293 spin_unlock_irq(&gcwq->lock);
1296 * We've raced with CPU hot[un]plug. Give it a breather
1297 * and retry migration. cond_resched() is required here;
1298 * otherwise, we might deadlock against cpu_stop trying to
1299 * bring down the CPU on non-preemptive kernel.
1307 * Function for worker->rebind_work used to rebind rogue busy workers
1308 * to the associated cpu which is coming back online. This is
1309 * scheduled by cpu up but can race with other cpu hotplug operations
1310 * and may be executed twice without intervening cpu down.
1312 static void worker_rebind_fn(struct work_struct *work)
1314 struct worker *worker = container_of(work, struct worker, rebind_work);
1315 struct global_cwq *gcwq = worker->gcwq;
1317 if (worker_maybe_bind_and_lock(worker))
1318 worker_clr_flags(worker, WORKER_REBIND);
1320 spin_unlock_irq(&gcwq->lock);
1323 static struct worker *alloc_worker(void)
1325 struct worker *worker;
1327 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1329 INIT_LIST_HEAD(&worker->entry);
1330 INIT_LIST_HEAD(&worker->scheduled);
1331 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1332 /* on creation a worker is in !idle && prep state */
1333 worker->flags = WORKER_PREP;
1339 * create_worker - create a new workqueue worker
1340 * @gcwq: gcwq the new worker will belong to
1341 * @bind: whether to set affinity to @cpu or not
1343 * Create a new worker which is bound to @gcwq. The returned worker
1344 * can be started by calling start_worker() or destroyed using
1348 * Might sleep. Does GFP_KERNEL allocations.
1351 * Pointer to the newly created worker.
1353 static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1355 bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1356 struct worker *worker = NULL;
1359 spin_lock_irq(&gcwq->lock);
1360 while (ida_get_new(&gcwq->worker_ida, &id)) {
1361 spin_unlock_irq(&gcwq->lock);
1362 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1364 spin_lock_irq(&gcwq->lock);
1366 spin_unlock_irq(&gcwq->lock);
1368 worker = alloc_worker();
1372 worker->gcwq = gcwq;
1375 if (!on_unbound_cpu)
1376 worker->task = kthread_create_on_node(worker_thread,
1378 cpu_to_node(gcwq->cpu),
1379 "kworker/%u:%d", gcwq->cpu, id);
1381 worker->task = kthread_create(worker_thread, worker,
1382 "kworker/u:%d", id);
1383 if (IS_ERR(worker->task))
1387 * A rogue worker will become a regular one if CPU comes
1388 * online later on. Make sure every worker has
1389 * PF_THREAD_BOUND set.
1391 if (bind && !on_unbound_cpu)
1392 kthread_bind(worker->task, gcwq->cpu);
1394 worker->task->flags |= PF_THREAD_BOUND;
1396 worker->flags |= WORKER_UNBOUND;
1402 spin_lock_irq(&gcwq->lock);
1403 ida_remove(&gcwq->worker_ida, id);
1404 spin_unlock_irq(&gcwq->lock);
1411 * start_worker - start a newly created worker
1412 * @worker: worker to start
1414 * Make the gcwq aware of @worker and start it.
1417 * spin_lock_irq(gcwq->lock).
1419 static void start_worker(struct worker *worker)
1421 worker->flags |= WORKER_STARTED;
1422 worker->gcwq->nr_workers++;
1423 worker_enter_idle(worker);
1424 wake_up_process(worker->task);
1428 * destroy_worker - destroy a workqueue worker
1429 * @worker: worker to be destroyed
1431 * Destroy @worker and adjust @gcwq stats accordingly.
1434 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1436 static void destroy_worker(struct worker *worker)
1438 struct global_cwq *gcwq = worker->gcwq;
1439 int id = worker->id;
1441 /* sanity check frenzy */
1442 BUG_ON(worker->current_work);
1443 BUG_ON(!list_empty(&worker->scheduled));
1445 if (worker->flags & WORKER_STARTED)
1447 if (worker->flags & WORKER_IDLE)
1450 list_del_init(&worker->entry);
1451 worker->flags |= WORKER_DIE;
1453 spin_unlock_irq(&gcwq->lock);
1455 kthread_stop(worker->task);
1458 spin_lock_irq(&gcwq->lock);
1459 ida_remove(&gcwq->worker_ida, id);
1462 static void idle_worker_timeout(unsigned long __gcwq)
1464 struct global_cwq *gcwq = (void *)__gcwq;
1466 spin_lock_irq(&gcwq->lock);
1468 if (too_many_workers(gcwq)) {
1469 struct worker *worker;
1470 unsigned long expires;
1472 /* idle_list is kept in LIFO order, check the last one */
1473 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1474 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1476 if (time_before(jiffies, expires))
1477 mod_timer(&gcwq->idle_timer, expires);
1479 /* it's been idle for too long, wake up manager */
1480 gcwq->flags |= GCWQ_MANAGE_WORKERS;
1481 wake_up_worker(gcwq);
1485 spin_unlock_irq(&gcwq->lock);
1488 static bool send_mayday(struct work_struct *work)
1490 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1491 struct workqueue_struct *wq = cwq->wq;
1494 if (!(wq->flags & WQ_RESCUER))
1497 /* mayday mayday mayday */
1498 cpu = cwq->gcwq->cpu;
1499 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1500 if (cpu == WORK_CPU_UNBOUND)
1502 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1503 wake_up_process(wq->rescuer->task);
1507 static void gcwq_mayday_timeout(unsigned long __gcwq)
1509 struct global_cwq *gcwq = (void *)__gcwq;
1510 struct work_struct *work;
1512 spin_lock_irq(&gcwq->lock);
1514 if (need_to_create_worker(gcwq)) {
1516 * We've been trying to create a new worker but
1517 * haven't been successful. We might be hitting an
1518 * allocation deadlock. Send distress signals to
1521 list_for_each_entry(work, &gcwq->worklist, entry)
1525 spin_unlock_irq(&gcwq->lock);
1527 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1531 * maybe_create_worker - create a new worker if necessary
1532 * @gcwq: gcwq to create a new worker for
1534 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1535 * have at least one idle worker on return from this function. If
1536 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1537 * sent to all rescuers with works scheduled on @gcwq to resolve
1538 * possible allocation deadlock.
1540 * On return, need_to_create_worker() is guaranteed to be false and
1541 * may_start_working() true.
1544 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1545 * multiple times. Does GFP_KERNEL allocations. Called only from
1549 * false if no action was taken and gcwq->lock stayed locked, true
1552 static bool maybe_create_worker(struct global_cwq *gcwq)
1553 __releases(&gcwq->lock)
1554 __acquires(&gcwq->lock)
1556 if (!need_to_create_worker(gcwq))
1559 spin_unlock_irq(&gcwq->lock);
1561 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1562 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1565 struct worker *worker;
1567 worker = create_worker(gcwq, true);
1569 del_timer_sync(&gcwq->mayday_timer);
1570 spin_lock_irq(&gcwq->lock);
1571 start_worker(worker);
1572 BUG_ON(need_to_create_worker(gcwq));
1576 if (!need_to_create_worker(gcwq))
1579 __set_current_state(TASK_INTERRUPTIBLE);
1580 schedule_timeout(CREATE_COOLDOWN);
1582 if (!need_to_create_worker(gcwq))
1586 del_timer_sync(&gcwq->mayday_timer);
1587 spin_lock_irq(&gcwq->lock);
1588 if (need_to_create_worker(gcwq))
1594 * maybe_destroy_worker - destroy workers which have been idle for a while
1595 * @gcwq: gcwq to destroy workers for
1597 * Destroy @gcwq workers which have been idle for longer than
1598 * IDLE_WORKER_TIMEOUT.
1601 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1602 * multiple times. Called only from manager.
1605 * false if no action was taken and gcwq->lock stayed locked, true
1608 static bool maybe_destroy_workers(struct global_cwq *gcwq)
1612 while (too_many_workers(gcwq)) {
1613 struct worker *worker;
1614 unsigned long expires;
1616 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1617 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1619 if (time_before(jiffies, expires)) {
1620 mod_timer(&gcwq->idle_timer, expires);
1624 destroy_worker(worker);
1632 * manage_workers - manage worker pool
1635 * Assume the manager role and manage gcwq worker pool @worker belongs
1636 * to. At any given time, there can be only zero or one manager per
1637 * gcwq. The exclusion is handled automatically by this function.
1639 * The caller can safely start processing works on false return. On
1640 * true return, it's guaranteed that need_to_create_worker() is false
1641 * and may_start_working() is true.
1644 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1645 * multiple times. Does GFP_KERNEL allocations.
1648 * false if no action was taken and gcwq->lock stayed locked, true if
1649 * some action was taken.
1651 static bool manage_workers(struct worker *worker)
1653 struct global_cwq *gcwq = worker->gcwq;
1656 if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1659 gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1660 gcwq->flags |= GCWQ_MANAGING_WORKERS;
1663 * Destroy and then create so that may_start_working() is true
1666 ret |= maybe_destroy_workers(gcwq);
1667 ret |= maybe_create_worker(gcwq);
1669 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1672 * The trustee might be waiting to take over the manager
1673 * position, tell it we're done.
1675 if (unlikely(gcwq->trustee))
1676 wake_up_all(&gcwq->trustee_wait);
1682 * move_linked_works - move linked works to a list
1683 * @work: start of series of works to be scheduled
1684 * @head: target list to append @work to
1685 * @nextp: out paramter for nested worklist walking
1687 * Schedule linked works starting from @work to @head. Work series to
1688 * be scheduled starts at @work and includes any consecutive work with
1689 * WORK_STRUCT_LINKED set in its predecessor.
1691 * If @nextp is not NULL, it's updated to point to the next work of
1692 * the last scheduled work. This allows move_linked_works() to be
1693 * nested inside outer list_for_each_entry_safe().
1696 * spin_lock_irq(gcwq->lock).
1698 static void move_linked_works(struct work_struct *work, struct list_head *head,
1699 struct work_struct **nextp)
1701 struct work_struct *n;
1704 * Linked worklist will always end before the end of the list,
1705 * use NULL for list head.
1707 list_for_each_entry_safe_from(work, n, NULL, entry) {
1708 list_move_tail(&work->entry, head);
1709 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1714 * If we're already inside safe list traversal and have moved
1715 * multiple works to the scheduled queue, the next position
1716 * needs to be updated.
1722 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1724 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1725 struct work_struct, entry);
1726 struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1728 trace_workqueue_activate_work(work);
1729 move_linked_works(work, pos, NULL);
1730 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1735 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1736 * @cwq: cwq of interest
1737 * @color: color of work which left the queue
1738 * @delayed: for a delayed work
1740 * A work either has completed or is removed from pending queue,
1741 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1744 * spin_lock_irq(gcwq->lock).
1746 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1749 /* ignore uncolored works */
1750 if (color == WORK_NO_COLOR)
1753 cwq->nr_in_flight[color]--;
1757 if (!list_empty(&cwq->delayed_works)) {
1758 /* one down, submit a delayed one */
1759 if (cwq->nr_active < cwq->max_active)
1760 cwq_activate_first_delayed(cwq);
1764 /* is flush in progress and are we at the flushing tip? */
1765 if (likely(cwq->flush_color != color))
1768 /* are there still in-flight works? */
1769 if (cwq->nr_in_flight[color])
1772 /* this cwq is done, clear flush_color */
1773 cwq->flush_color = -1;
1776 * If this was the last cwq, wake up the first flusher. It
1777 * will handle the rest.
1779 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1780 complete(&cwq->wq->first_flusher->done);
1784 * process_one_work - process single work
1786 * @work: work to process
1788 * Process @work. This function contains all the logics necessary to
1789 * process a single work including synchronization against and
1790 * interaction with other workers on the same cpu, queueing and
1791 * flushing. As long as context requirement is met, any worker can
1792 * call this function to process a work.
1795 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1797 static void process_one_work(struct worker *worker, struct work_struct *work)
1798 __releases(&gcwq->lock)
1799 __acquires(&gcwq->lock)
1801 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1802 struct global_cwq *gcwq = cwq->gcwq;
1803 struct hlist_head *bwh = busy_worker_head(gcwq, work);
1804 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1805 work_func_t f = work->func;
1807 struct worker *collision;
1808 #ifdef CONFIG_LOCKDEP
1810 * It is permissible to free the struct work_struct from
1811 * inside the function that is called from it, this we need to
1812 * take into account for lockdep too. To avoid bogus "held
1813 * lock freed" warnings as well as problems when looking into
1814 * work->lockdep_map, make a copy and use that here.
1816 struct lockdep_map lockdep_map = work->lockdep_map;
1819 * A single work shouldn't be executed concurrently by
1820 * multiple workers on a single cpu. Check whether anyone is
1821 * already processing the work. If so, defer the work to the
1822 * currently executing one.
1824 collision = __find_worker_executing_work(gcwq, bwh, work);
1825 if (unlikely(collision)) {
1826 move_linked_works(work, &collision->scheduled, NULL);
1830 /* claim and process */
1831 debug_work_deactivate(work);
1832 hlist_add_head(&worker->hentry, bwh);
1833 worker->current_work = work;
1834 worker->current_cwq = cwq;
1835 work_color = get_work_color(work);
1837 /* record the current cpu number in the work data and dequeue */
1838 set_work_cpu(work, gcwq->cpu);
1839 list_del_init(&work->entry);
1842 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1843 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1845 if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1846 struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1847 struct work_struct, entry);
1849 if (!list_empty(&gcwq->worklist) &&
1850 get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1851 wake_up_worker(gcwq);
1853 gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1857 * CPU intensive works don't participate in concurrency
1858 * management. They're the scheduler's responsibility.
1860 if (unlikely(cpu_intensive))
1861 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1863 spin_unlock_irq(&gcwq->lock);
1865 work_clear_pending(work);
1866 lock_map_acquire_read(&cwq->wq->lockdep_map);
1867 lock_map_acquire(&lockdep_map);
1868 trace_workqueue_execute_start(work);
1871 * While we must be careful to not use "work" after this, the trace
1872 * point will only record its address.
1874 trace_workqueue_execute_end(work);
1875 lock_map_release(&lockdep_map);
1876 lock_map_release(&cwq->wq->lockdep_map);
1878 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1879 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1881 current->comm, preempt_count(), task_pid_nr(current));
1882 printk(KERN_ERR " last function: ");
1883 print_symbol("%s\n", (unsigned long)f);
1884 debug_show_held_locks(current);
1888 spin_lock_irq(&gcwq->lock);
1890 /* clear cpu intensive status */
1891 if (unlikely(cpu_intensive))
1892 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1894 /* we're done with it, release */
1895 hlist_del_init(&worker->hentry);
1896 worker->current_work = NULL;
1897 worker->current_cwq = NULL;
1898 cwq_dec_nr_in_flight(cwq, work_color, false);
1902 * process_scheduled_works - process scheduled works
1905 * Process all scheduled works. Please note that the scheduled list
1906 * may change while processing a work, so this function repeatedly
1907 * fetches a work from the top and executes it.
1910 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1913 static void process_scheduled_works(struct worker *worker)
1915 while (!list_empty(&worker->scheduled)) {
1916 struct work_struct *work = list_first_entry(&worker->scheduled,
1917 struct work_struct, entry);
1918 process_one_work(worker, work);
1923 * worker_thread - the worker thread function
1926 * The gcwq worker thread function. There's a single dynamic pool of
1927 * these per each cpu. These workers process all works regardless of
1928 * their specific target workqueue. The only exception is works which
1929 * belong to workqueues with a rescuer which will be explained in
1932 static int worker_thread(void *__worker)
1934 struct worker *worker = __worker;
1935 struct global_cwq *gcwq = worker->gcwq;
1937 /* tell the scheduler that this is a workqueue worker */
1938 worker->task->flags |= PF_WQ_WORKER;
1940 spin_lock_irq(&gcwq->lock);
1942 /* DIE can be set only while we're idle, checking here is enough */
1943 if (worker->flags & WORKER_DIE) {
1944 spin_unlock_irq(&gcwq->lock);
1945 worker->task->flags &= ~PF_WQ_WORKER;
1949 worker_leave_idle(worker);
1951 /* no more worker necessary? */
1952 if (!need_more_worker(gcwq))
1955 /* do we need to manage? */
1956 if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1960 * ->scheduled list can only be filled while a worker is
1961 * preparing to process a work or actually processing it.
1962 * Make sure nobody diddled with it while I was sleeping.
1964 BUG_ON(!list_empty(&worker->scheduled));
1967 * When control reaches this point, we're guaranteed to have
1968 * at least one idle worker or that someone else has already
1969 * assumed the manager role.
1971 worker_clr_flags(worker, WORKER_PREP);
1974 struct work_struct *work =
1975 list_first_entry(&gcwq->worklist,
1976 struct work_struct, entry);
1978 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1979 /* optimization path, not strictly necessary */
1980 process_one_work(worker, work);
1981 if (unlikely(!list_empty(&worker->scheduled)))
1982 process_scheduled_works(worker);
1984 move_linked_works(work, &worker->scheduled, NULL);
1985 process_scheduled_works(worker);
1987 } while (keep_working(gcwq));
1989 worker_set_flags(worker, WORKER_PREP, false);
1991 if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1995 * gcwq->lock is held and there's no work to process and no
1996 * need to manage, sleep. Workers are woken up only while
1997 * holding gcwq->lock or from local cpu, so setting the
1998 * current state before releasing gcwq->lock is enough to
1999 * prevent losing any event.
2001 worker_enter_idle(worker);
2002 __set_current_state(TASK_INTERRUPTIBLE);
2003 spin_unlock_irq(&gcwq->lock);
2009 * rescuer_thread - the rescuer thread function
2010 * @__wq: the associated workqueue
2012 * Workqueue rescuer thread function. There's one rescuer for each
2013 * workqueue which has WQ_RESCUER set.
2015 * Regular work processing on a gcwq may block trying to create a new
2016 * worker which uses GFP_KERNEL allocation which has slight chance of
2017 * developing into deadlock if some works currently on the same queue
2018 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2019 * the problem rescuer solves.
2021 * When such condition is possible, the gcwq summons rescuers of all
2022 * workqueues which have works queued on the gcwq and let them process
2023 * those works so that forward progress can be guaranteed.
2025 * This should happen rarely.
2027 static int rescuer_thread(void *__wq)
2029 struct workqueue_struct *wq = __wq;
2030 struct worker *rescuer = wq->rescuer;
2031 struct list_head *scheduled = &rescuer->scheduled;
2032 bool is_unbound = wq->flags & WQ_UNBOUND;
2035 set_user_nice(current, RESCUER_NICE_LEVEL);
2037 set_current_state(TASK_INTERRUPTIBLE);
2039 if (kthread_should_stop())
2043 * See whether any cpu is asking for help. Unbounded
2044 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2046 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2047 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2048 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2049 struct global_cwq *gcwq = cwq->gcwq;
2050 struct work_struct *work, *n;
2052 __set_current_state(TASK_RUNNING);
2053 mayday_clear_cpu(cpu, wq->mayday_mask);
2055 /* migrate to the target cpu if possible */
2056 rescuer->gcwq = gcwq;
2057 worker_maybe_bind_and_lock(rescuer);
2060 * Slurp in all works issued via this workqueue and
2063 BUG_ON(!list_empty(&rescuer->scheduled));
2064 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2065 if (get_work_cwq(work) == cwq)
2066 move_linked_works(work, scheduled, &n);
2068 process_scheduled_works(rescuer);
2071 * Leave this gcwq. If keep_working() is %true, notify a
2072 * regular worker; otherwise, we end up with 0 concurrency
2073 * and stalling the execution.
2075 if (keep_working(gcwq))
2076 wake_up_worker(gcwq);
2078 spin_unlock_irq(&gcwq->lock);
2086 struct work_struct work;
2087 struct completion done;
2090 static void wq_barrier_func(struct work_struct *work)
2092 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2093 complete(&barr->done);
2097 * insert_wq_barrier - insert a barrier work
2098 * @cwq: cwq to insert barrier into
2099 * @barr: wq_barrier to insert
2100 * @target: target work to attach @barr to
2101 * @worker: worker currently executing @target, NULL if @target is not executing
2103 * @barr is linked to @target such that @barr is completed only after
2104 * @target finishes execution. Please note that the ordering
2105 * guarantee is observed only with respect to @target and on the local
2108 * Currently, a queued barrier can't be canceled. This is because
2109 * try_to_grab_pending() can't determine whether the work to be
2110 * grabbed is at the head of the queue and thus can't clear LINKED
2111 * flag of the previous work while there must be a valid next work
2112 * after a work with LINKED flag set.
2114 * Note that when @worker is non-NULL, @target may be modified
2115 * underneath us, so we can't reliably determine cwq from @target.
2118 * spin_lock_irq(gcwq->lock).
2120 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2121 struct wq_barrier *barr,
2122 struct work_struct *target, struct worker *worker)
2124 struct list_head *head;
2125 unsigned int linked = 0;
2128 * debugobject calls are safe here even with gcwq->lock locked
2129 * as we know for sure that this will not trigger any of the
2130 * checks and call back into the fixup functions where we
2133 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2134 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2135 init_completion(&barr->done);
2138 * If @target is currently being executed, schedule the
2139 * barrier to the worker; otherwise, put it after @target.
2142 head = worker->scheduled.next;
2144 unsigned long *bits = work_data_bits(target);
2146 head = target->entry.next;
2147 /* there can already be other linked works, inherit and set */
2148 linked = *bits & WORK_STRUCT_LINKED;
2149 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2152 debug_work_activate(&barr->work);
2153 insert_work(cwq, &barr->work, head,
2154 work_color_to_flags(WORK_NO_COLOR) | linked);
2158 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2159 * @wq: workqueue being flushed
2160 * @flush_color: new flush color, < 0 for no-op
2161 * @work_color: new work color, < 0 for no-op
2163 * Prepare cwqs for workqueue flushing.
2165 * If @flush_color is non-negative, flush_color on all cwqs should be
2166 * -1. If no cwq has in-flight commands at the specified color, all
2167 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2168 * has in flight commands, its cwq->flush_color is set to
2169 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2170 * wakeup logic is armed and %true is returned.
2172 * The caller should have initialized @wq->first_flusher prior to
2173 * calling this function with non-negative @flush_color. If
2174 * @flush_color is negative, no flush color update is done and %false
2177 * If @work_color is non-negative, all cwqs should have the same
2178 * work_color which is previous to @work_color and all will be
2179 * advanced to @work_color.
2182 * mutex_lock(wq->flush_mutex).
2185 * %true if @flush_color >= 0 and there's something to flush. %false
2188 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2189 int flush_color, int work_color)
2194 if (flush_color >= 0) {
2195 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2196 atomic_set(&wq->nr_cwqs_to_flush, 1);
2199 for_each_cwq_cpu(cpu, wq) {
2200 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2201 struct global_cwq *gcwq = cwq->gcwq;
2203 spin_lock_irq(&gcwq->lock);
2205 if (flush_color >= 0) {
2206 BUG_ON(cwq->flush_color != -1);
2208 if (cwq->nr_in_flight[flush_color]) {
2209 cwq->flush_color = flush_color;
2210 atomic_inc(&wq->nr_cwqs_to_flush);
2215 if (work_color >= 0) {
2216 BUG_ON(work_color != work_next_color(cwq->work_color));
2217 cwq->work_color = work_color;
2220 spin_unlock_irq(&gcwq->lock);
2223 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2224 complete(&wq->first_flusher->done);
2230 * flush_workqueue - ensure that any scheduled work has run to completion.
2231 * @wq: workqueue to flush
2233 * Forces execution of the workqueue and blocks until its completion.
2234 * This is typically used in driver shutdown handlers.
2236 * We sleep until all works which were queued on entry have been handled,
2237 * but we are not livelocked by new incoming ones.
2239 void flush_workqueue(struct workqueue_struct *wq)
2241 struct wq_flusher this_flusher = {
2242 .list = LIST_HEAD_INIT(this_flusher.list),
2244 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2248 lock_map_acquire(&wq->lockdep_map);
2249 lock_map_release(&wq->lockdep_map);
2251 mutex_lock(&wq->flush_mutex);
2254 * Start-to-wait phase
2256 next_color = work_next_color(wq->work_color);
2258 if (next_color != wq->flush_color) {
2260 * Color space is not full. The current work_color
2261 * becomes our flush_color and work_color is advanced
2264 BUG_ON(!list_empty(&wq->flusher_overflow));
2265 this_flusher.flush_color = wq->work_color;
2266 wq->work_color = next_color;
2268 if (!wq->first_flusher) {
2269 /* no flush in progress, become the first flusher */
2270 BUG_ON(wq->flush_color != this_flusher.flush_color);
2272 wq->first_flusher = &this_flusher;
2274 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2276 /* nothing to flush, done */
2277 wq->flush_color = next_color;
2278 wq->first_flusher = NULL;
2283 BUG_ON(wq->flush_color == this_flusher.flush_color);
2284 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2285 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2289 * Oops, color space is full, wait on overflow queue.
2290 * The next flush completion will assign us
2291 * flush_color and transfer to flusher_queue.
2293 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2296 mutex_unlock(&wq->flush_mutex);
2298 wait_for_completion(&this_flusher.done);
2301 * Wake-up-and-cascade phase
2303 * First flushers are responsible for cascading flushes and
2304 * handling overflow. Non-first flushers can simply return.
2306 if (wq->first_flusher != &this_flusher)
2309 mutex_lock(&wq->flush_mutex);
2311 /* we might have raced, check again with mutex held */
2312 if (wq->first_flusher != &this_flusher)
2315 wq->first_flusher = NULL;
2317 BUG_ON(!list_empty(&this_flusher.list));
2318 BUG_ON(wq->flush_color != this_flusher.flush_color);
2321 struct wq_flusher *next, *tmp;
2323 /* complete all the flushers sharing the current flush color */
2324 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2325 if (next->flush_color != wq->flush_color)
2327 list_del_init(&next->list);
2328 complete(&next->done);
2331 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2332 wq->flush_color != work_next_color(wq->work_color));
2334 /* this flush_color is finished, advance by one */
2335 wq->flush_color = work_next_color(wq->flush_color);
2337 /* one color has been freed, handle overflow queue */
2338 if (!list_empty(&wq->flusher_overflow)) {
2340 * Assign the same color to all overflowed
2341 * flushers, advance work_color and append to
2342 * flusher_queue. This is the start-to-wait
2343 * phase for these overflowed flushers.
2345 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2346 tmp->flush_color = wq->work_color;
2348 wq->work_color = work_next_color(wq->work_color);
2350 list_splice_tail_init(&wq->flusher_overflow,
2351 &wq->flusher_queue);
2352 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2355 if (list_empty(&wq->flusher_queue)) {
2356 BUG_ON(wq->flush_color != wq->work_color);
2361 * Need to flush more colors. Make the next flusher
2362 * the new first flusher and arm cwqs.
2364 BUG_ON(wq->flush_color == wq->work_color);
2365 BUG_ON(wq->flush_color != next->flush_color);
2367 list_del_init(&next->list);
2368 wq->first_flusher = next;
2370 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2374 * Meh... this color is already done, clear first
2375 * flusher and repeat cascading.
2377 wq->first_flusher = NULL;
2381 mutex_unlock(&wq->flush_mutex);
2383 EXPORT_SYMBOL_GPL(flush_workqueue);
2386 * drain_workqueue - drain a workqueue
2387 * @wq: workqueue to drain
2389 * Wait until the workqueue becomes empty. While draining is in progress,
2390 * only chain queueing is allowed. IOW, only currently pending or running
2391 * work items on @wq can queue further work items on it. @wq is flushed
2392 * repeatedly until it becomes empty. The number of flushing is detemined
2393 * by the depth of chaining and should be relatively short. Whine if it
2396 void drain_workqueue(struct workqueue_struct *wq)
2398 unsigned int flush_cnt = 0;
2402 * __queue_work() needs to test whether there are drainers, is much
2403 * hotter than drain_workqueue() and already looks at @wq->flags.
2404 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2406 spin_lock(&workqueue_lock);
2407 if (!wq->nr_drainers++)
2408 wq->flags |= WQ_DRAINING;
2409 spin_unlock(&workqueue_lock);
2411 flush_workqueue(wq);
2413 for_each_cwq_cpu(cpu, wq) {
2414 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2417 spin_lock_irq(&cwq->gcwq->lock);
2418 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2419 spin_unlock_irq(&cwq->gcwq->lock);
2424 if (++flush_cnt == 10 ||
2425 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2426 pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n",
2427 wq->name, flush_cnt);
2431 spin_lock(&workqueue_lock);
2432 if (!--wq->nr_drainers)
2433 wq->flags &= ~WQ_DRAINING;
2434 spin_unlock(&workqueue_lock);
2436 EXPORT_SYMBOL_GPL(drain_workqueue);
2438 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2439 bool wait_executing)
2441 struct worker *worker = NULL;
2442 struct global_cwq *gcwq;
2443 struct cpu_workqueue_struct *cwq;
2446 gcwq = get_work_gcwq(work);
2450 spin_lock_irq(&gcwq->lock);
2451 if (!list_empty(&work->entry)) {
2453 * See the comment near try_to_grab_pending()->smp_rmb().
2454 * If it was re-queued to a different gcwq under us, we
2455 * are not going to wait.
2458 cwq = get_work_cwq(work);
2459 if (unlikely(!cwq || gcwq != cwq->gcwq))
2461 } else if (wait_executing) {
2462 worker = find_worker_executing_work(gcwq, work);
2465 cwq = worker->current_cwq;
2469 insert_wq_barrier(cwq, barr, work, worker);
2470 spin_unlock_irq(&gcwq->lock);
2473 * If @max_active is 1 or rescuer is in use, flushing another work
2474 * item on the same workqueue may lead to deadlock. Make sure the
2475 * flusher is not running on the same workqueue by verifying write
2478 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2479 lock_map_acquire(&cwq->wq->lockdep_map);
2481 lock_map_acquire_read(&cwq->wq->lockdep_map);
2482 lock_map_release(&cwq->wq->lockdep_map);
2486 spin_unlock_irq(&gcwq->lock);
2491 * flush_work - wait for a work to finish executing the last queueing instance
2492 * @work: the work to flush
2494 * Wait until @work has finished execution. This function considers
2495 * only the last queueing instance of @work. If @work has been
2496 * enqueued across different CPUs on a non-reentrant workqueue or on
2497 * multiple workqueues, @work might still be executing on return on
2498 * some of the CPUs from earlier queueing.
2500 * If @work was queued only on a non-reentrant, ordered or unbound
2501 * workqueue, @work is guaranteed to be idle on return if it hasn't
2502 * been requeued since flush started.
2505 * %true if flush_work() waited for the work to finish execution,
2506 * %false if it was already idle.
2508 bool flush_work(struct work_struct *work)
2510 struct wq_barrier barr;
2512 lock_map_acquire(&work->lockdep_map);
2513 lock_map_release(&work->lockdep_map);
2515 if (start_flush_work(work, &barr, true)) {
2516 wait_for_completion(&barr.done);
2517 destroy_work_on_stack(&barr.work);
2522 EXPORT_SYMBOL_GPL(flush_work);
2524 static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2526 struct wq_barrier barr;
2527 struct worker *worker;
2529 spin_lock_irq(&gcwq->lock);
2531 worker = find_worker_executing_work(gcwq, work);
2532 if (unlikely(worker))
2533 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2535 spin_unlock_irq(&gcwq->lock);
2537 if (unlikely(worker)) {
2538 wait_for_completion(&barr.done);
2539 destroy_work_on_stack(&barr.work);
2545 static bool wait_on_work(struct work_struct *work)
2552 lock_map_acquire(&work->lockdep_map);
2553 lock_map_release(&work->lockdep_map);
2555 for_each_gcwq_cpu(cpu)
2556 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2561 * flush_work_sync - wait until a work has finished execution
2562 * @work: the work to flush
2564 * Wait until @work has finished execution. On return, it's
2565 * guaranteed that all queueing instances of @work which happened
2566 * before this function is called are finished. In other words, if
2567 * @work hasn't been requeued since this function was called, @work is
2568 * guaranteed to be idle on return.
2571 * %true if flush_work_sync() waited for the work to finish execution,
2572 * %false if it was already idle.
2574 bool flush_work_sync(struct work_struct *work)
2576 struct wq_barrier barr;
2577 bool pending, waited;
2579 /* we'll wait for executions separately, queue barr only if pending */
2580 pending = start_flush_work(work, &barr, false);
2582 /* wait for executions to finish */
2583 waited = wait_on_work(work);
2585 /* wait for the pending one */
2587 wait_for_completion(&barr.done);
2588 destroy_work_on_stack(&barr.work);
2591 return pending || waited;
2593 EXPORT_SYMBOL_GPL(flush_work_sync);
2596 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2597 * so this work can't be re-armed in any way.
2599 static int try_to_grab_pending(struct work_struct *work)
2601 struct global_cwq *gcwq;
2604 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2608 * The queueing is in progress, or it is already queued. Try to
2609 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2611 gcwq = get_work_gcwq(work);
2615 spin_lock_irq(&gcwq->lock);
2616 if (!list_empty(&work->entry)) {
2618 * This work is queued, but perhaps we locked the wrong gcwq.
2619 * In that case we must see the new value after rmb(), see
2620 * insert_work()->wmb().
2623 if (gcwq == get_work_gcwq(work)) {
2624 debug_work_deactivate(work);
2625 list_del_init(&work->entry);
2626 cwq_dec_nr_in_flight(get_work_cwq(work),
2627 get_work_color(work),
2628 *work_data_bits(work) & WORK_STRUCT_DELAYED);
2632 spin_unlock_irq(&gcwq->lock);
2637 static bool __cancel_work_timer(struct work_struct *work,
2638 struct timer_list* timer)
2643 ret = (timer && likely(del_timer(timer)));
2645 ret = try_to_grab_pending(work);
2647 } while (unlikely(ret < 0));
2649 clear_work_data(work);
2654 * cancel_work_sync - cancel a work and wait for it to finish
2655 * @work: the work to cancel
2657 * Cancel @work and wait for its execution to finish. This function
2658 * can be used even if the work re-queues itself or migrates to
2659 * another workqueue. On return from this function, @work is
2660 * guaranteed to be not pending or executing on any CPU.
2662 * cancel_work_sync(&delayed_work->work) must not be used for
2663 * delayed_work's. Use cancel_delayed_work_sync() instead.
2665 * The caller must ensure that the workqueue on which @work was last
2666 * queued can't be destroyed before this function returns.
2669 * %true if @work was pending, %false otherwise.
2671 bool cancel_work_sync(struct work_struct *work)
2673 return __cancel_work_timer(work, NULL);
2675 EXPORT_SYMBOL_GPL(cancel_work_sync);
2678 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2679 * @dwork: the delayed work to flush
2681 * Delayed timer is cancelled and the pending work is queued for
2682 * immediate execution. Like flush_work(), this function only
2683 * considers the last queueing instance of @dwork.
2686 * %true if flush_work() waited for the work to finish execution,
2687 * %false if it was already idle.
2689 bool flush_delayed_work(struct delayed_work *dwork)
2691 if (del_timer_sync(&dwork->timer))
2692 __queue_work(raw_smp_processor_id(),
2693 get_work_cwq(&dwork->work)->wq, &dwork->work);
2694 return flush_work(&dwork->work);
2696 EXPORT_SYMBOL(flush_delayed_work);
2699 * flush_delayed_work_sync - wait for a dwork to finish
2700 * @dwork: the delayed work to flush
2702 * Delayed timer is cancelled and the pending work is queued for
2703 * execution immediately. Other than timer handling, its behavior
2704 * is identical to flush_work_sync().
2707 * %true if flush_work_sync() waited for the work to finish execution,
2708 * %false if it was already idle.
2710 bool flush_delayed_work_sync(struct delayed_work *dwork)
2712 if (del_timer_sync(&dwork->timer))
2713 __queue_work(raw_smp_processor_id(),
2714 get_work_cwq(&dwork->work)->wq, &dwork->work);
2715 return flush_work_sync(&dwork->work);
2717 EXPORT_SYMBOL(flush_delayed_work_sync);
2720 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2721 * @dwork: the delayed work cancel
2723 * This is cancel_work_sync() for delayed works.
2726 * %true if @dwork was pending, %false otherwise.
2728 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2730 return __cancel_work_timer(&dwork->work, &dwork->timer);
2732 EXPORT_SYMBOL(cancel_delayed_work_sync);
2735 * schedule_work - put work task in global workqueue
2736 * @work: job to be done
2738 * Returns zero if @work was already on the kernel-global workqueue and
2739 * non-zero otherwise.
2741 * This puts a job in the kernel-global workqueue if it was not already
2742 * queued and leaves it in the same position on the kernel-global
2743 * workqueue otherwise.
2745 int schedule_work(struct work_struct *work)
2747 return queue_work(system_wq, work);
2749 EXPORT_SYMBOL(schedule_work);
2752 * schedule_work_on - put work task on a specific cpu
2753 * @cpu: cpu to put the work task on
2754 * @work: job to be done
2756 * This puts a job on a specific cpu
2758 int schedule_work_on(int cpu, struct work_struct *work)
2760 return queue_work_on(cpu, system_wq, work);
2762 EXPORT_SYMBOL(schedule_work_on);
2765 * schedule_delayed_work - put work task in global workqueue after delay
2766 * @dwork: job to be done
2767 * @delay: number of jiffies to wait or 0 for immediate execution
2769 * After waiting for a given time this puts a job in the kernel-global
2772 int schedule_delayed_work(struct delayed_work *dwork,
2773 unsigned long delay)
2775 return queue_delayed_work(system_wq, dwork, delay);
2777 EXPORT_SYMBOL(schedule_delayed_work);
2780 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2782 * @dwork: job to be done
2783 * @delay: number of jiffies to wait
2785 * After waiting for a given time this puts a job in the kernel-global
2786 * workqueue on the specified CPU.
2788 int schedule_delayed_work_on(int cpu,
2789 struct delayed_work *dwork, unsigned long delay)
2791 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2793 EXPORT_SYMBOL(schedule_delayed_work_on);
2796 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2797 * @func: the function to call
2799 * schedule_on_each_cpu() executes @func on each online CPU using the
2800 * system workqueue and blocks until all CPUs have completed.
2801 * schedule_on_each_cpu() is very slow.
2804 * 0 on success, -errno on failure.
2806 int schedule_on_each_cpu(work_func_t func)
2809 struct work_struct __percpu *works;
2811 works = alloc_percpu(struct work_struct);
2817 for_each_online_cpu(cpu) {
2818 struct work_struct *work = per_cpu_ptr(works, cpu);
2820 INIT_WORK(work, func);
2821 schedule_work_on(cpu, work);
2824 for_each_online_cpu(cpu)
2825 flush_work(per_cpu_ptr(works, cpu));
2833 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2835 * Forces execution of the kernel-global workqueue and blocks until its
2838 * Think twice before calling this function! It's very easy to get into
2839 * trouble if you don't take great care. Either of the following situations
2840 * will lead to deadlock:
2842 * One of the work items currently on the workqueue needs to acquire
2843 * a lock held by your code or its caller.
2845 * Your code is running in the context of a work routine.
2847 * They will be detected by lockdep when they occur, but the first might not
2848 * occur very often. It depends on what work items are on the workqueue and
2849 * what locks they need, which you have no control over.
2851 * In most situations flushing the entire workqueue is overkill; you merely
2852 * need to know that a particular work item isn't queued and isn't running.
2853 * In such cases you should use cancel_delayed_work_sync() or
2854 * cancel_work_sync() instead.
2856 void flush_scheduled_work(void)
2858 flush_workqueue(system_wq);
2860 EXPORT_SYMBOL(flush_scheduled_work);
2863 * execute_in_process_context - reliably execute the routine with user context
2864 * @fn: the function to execute
2865 * @ew: guaranteed storage for the execute work structure (must
2866 * be available when the work executes)
2868 * Executes the function immediately if process context is available,
2869 * otherwise schedules the function for delayed execution.
2871 * Returns: 0 - function was executed
2872 * 1 - function was scheduled for execution
2874 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2876 if (!in_interrupt()) {
2881 INIT_WORK(&ew->work, fn);
2882 schedule_work(&ew->work);
2886 EXPORT_SYMBOL_GPL(execute_in_process_context);
2888 int keventd_up(void)
2890 return system_wq != NULL;
2893 static int alloc_cwqs(struct workqueue_struct *wq)
2896 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2897 * Make sure that the alignment isn't lower than that of
2898 * unsigned long long.
2900 const size_t size = sizeof(struct cpu_workqueue_struct);
2901 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2902 __alignof__(unsigned long long));
2904 if (!(wq->flags & WQ_UNBOUND))
2905 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2910 * Allocate enough room to align cwq and put an extra
2911 * pointer at the end pointing back to the originally
2912 * allocated pointer which will be used for free.
2914 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2916 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2917 *(void **)(wq->cpu_wq.single + 1) = ptr;
2921 /* just in case, make sure it's actually aligned */
2922 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2923 return wq->cpu_wq.v ? 0 : -ENOMEM;
2926 static void free_cwqs(struct workqueue_struct *wq)
2928 if (!(wq->flags & WQ_UNBOUND))
2929 free_percpu(wq->cpu_wq.pcpu);
2930 else if (wq->cpu_wq.single) {
2931 /* the pointer to free is stored right after the cwq */
2932 kfree(*(void **)(wq->cpu_wq.single + 1));
2936 static int wq_clamp_max_active(int max_active, unsigned int flags,
2939 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2941 if (max_active < 1 || max_active > lim)
2942 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2943 "is out of range, clamping between %d and %d\n",
2944 max_active, name, 1, lim);
2946 return clamp_val(max_active, 1, lim);
2949 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
2952 struct lock_class_key *key,
2953 const char *lock_name, ...)
2955 va_list args, args1;
2956 struct workqueue_struct *wq;
2960 /* determine namelen, allocate wq and format name */
2961 va_start(args, lock_name);
2962 va_copy(args1, args);
2963 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
2965 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
2969 vsnprintf(wq->name, namelen, fmt, args1);
2974 * Workqueues which may be used during memory reclaim should
2975 * have a rescuer to guarantee forward progress.
2977 if (flags & WQ_MEM_RECLAIM)
2978 flags |= WQ_RESCUER;
2981 * Unbound workqueues aren't concurrency managed and should be
2982 * dispatched to workers immediately.
2984 if (flags & WQ_UNBOUND)
2985 flags |= WQ_HIGHPRI;
2987 max_active = max_active ?: WQ_DFL_ACTIVE;
2988 max_active = wq_clamp_max_active(max_active, flags, wq->name);
2992 wq->saved_max_active = max_active;
2993 mutex_init(&wq->flush_mutex);
2994 atomic_set(&wq->nr_cwqs_to_flush, 0);
2995 INIT_LIST_HEAD(&wq->flusher_queue);
2996 INIT_LIST_HEAD(&wq->flusher_overflow);
2998 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2999 INIT_LIST_HEAD(&wq->list);
3001 if (alloc_cwqs(wq) < 0)
3004 for_each_cwq_cpu(cpu, wq) {
3005 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3006 struct global_cwq *gcwq = get_gcwq(cpu);
3008 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3011 cwq->flush_color = -1;
3012 cwq->max_active = max_active;
3013 INIT_LIST_HEAD(&cwq->delayed_works);
3016 if (flags & WQ_RESCUER) {
3017 struct worker *rescuer;
3019 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3022 wq->rescuer = rescuer = alloc_worker();
3026 rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3028 if (IS_ERR(rescuer->task))
3031 rescuer->task->flags |= PF_THREAD_BOUND;
3032 wake_up_process(rescuer->task);
3036 * workqueue_lock protects global freeze state and workqueues
3037 * list. Grab it, set max_active accordingly and add the new
3038 * workqueue to workqueues list.
3040 spin_lock(&workqueue_lock);
3042 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3043 for_each_cwq_cpu(cpu, wq)
3044 get_cwq(cpu, wq)->max_active = 0;
3046 list_add(&wq->list, &workqueues);
3048 spin_unlock(&workqueue_lock);
3054 free_mayday_mask(wq->mayday_mask);
3060 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3063 * destroy_workqueue - safely terminate a workqueue
3064 * @wq: target workqueue
3066 * Safely destroy a workqueue. All work currently pending will be done first.
3068 void destroy_workqueue(struct workqueue_struct *wq)
3072 /* drain it before proceeding with destruction */
3073 drain_workqueue(wq);
3076 * wq list is used to freeze wq, remove from list after
3077 * flushing is complete in case freeze races us.
3079 spin_lock(&workqueue_lock);
3080 list_del(&wq->list);
3081 spin_unlock(&workqueue_lock);
3084 for_each_cwq_cpu(cpu, wq) {
3085 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3088 for (i = 0; i < WORK_NR_COLORS; i++)
3089 BUG_ON(cwq->nr_in_flight[i]);
3090 BUG_ON(cwq->nr_active);
3091 BUG_ON(!list_empty(&cwq->delayed_works));
3094 if (wq->flags & WQ_RESCUER) {
3095 kthread_stop(wq->rescuer->task);
3096 free_mayday_mask(wq->mayday_mask);
3103 EXPORT_SYMBOL_GPL(destroy_workqueue);
3106 * workqueue_set_max_active - adjust max_active of a workqueue
3107 * @wq: target workqueue
3108 * @max_active: new max_active value.
3110 * Set max_active of @wq to @max_active.
3113 * Don't call from IRQ context.
3115 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3119 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3121 spin_lock(&workqueue_lock);
3123 wq->saved_max_active = max_active;
3125 for_each_cwq_cpu(cpu, wq) {
3126 struct global_cwq *gcwq = get_gcwq(cpu);
3128 spin_lock_irq(&gcwq->lock);
3130 if (!(wq->flags & WQ_FREEZABLE) ||
3131 !(gcwq->flags & GCWQ_FREEZING))
3132 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3134 spin_unlock_irq(&gcwq->lock);
3137 spin_unlock(&workqueue_lock);
3139 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3142 * workqueue_congested - test whether a workqueue is congested
3143 * @cpu: CPU in question
3144 * @wq: target workqueue
3146 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3147 * no synchronization around this function and the test result is
3148 * unreliable and only useful as advisory hints or for debugging.
3151 * %true if congested, %false otherwise.
3153 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3155 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3157 return !list_empty(&cwq->delayed_works);
3159 EXPORT_SYMBOL_GPL(workqueue_congested);
3162 * work_cpu - return the last known associated cpu for @work
3163 * @work: the work of interest
3166 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3168 unsigned int work_cpu(struct work_struct *work)
3170 struct global_cwq *gcwq = get_work_gcwq(work);
3172 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3174 EXPORT_SYMBOL_GPL(work_cpu);
3177 * work_busy - test whether a work is currently pending or running
3178 * @work: the work to be tested
3180 * Test whether @work is currently pending or running. There is no
3181 * synchronization around this function and the test result is
3182 * unreliable and only useful as advisory hints or for debugging.
3183 * Especially for reentrant wqs, the pending state might hide the
3187 * OR'd bitmask of WORK_BUSY_* bits.
3189 unsigned int work_busy(struct work_struct *work)
3191 struct global_cwq *gcwq = get_work_gcwq(work);
3192 unsigned long flags;
3193 unsigned int ret = 0;
3198 spin_lock_irqsave(&gcwq->lock, flags);
3200 if (work_pending(work))
3201 ret |= WORK_BUSY_PENDING;
3202 if (find_worker_executing_work(gcwq, work))
3203 ret |= WORK_BUSY_RUNNING;
3205 spin_unlock_irqrestore(&gcwq->lock, flags);
3209 EXPORT_SYMBOL_GPL(work_busy);
3214 * There are two challenges in supporting CPU hotplug. Firstly, there
3215 * are a lot of assumptions on strong associations among work, cwq and
3216 * gcwq which make migrating pending and scheduled works very
3217 * difficult to implement without impacting hot paths. Secondly,
3218 * gcwqs serve mix of short, long and very long running works making
3219 * blocked draining impractical.
3221 * This is solved by allowing a gcwq to be detached from CPU, running
3222 * it with unbound (rogue) workers and allowing it to be reattached
3223 * later if the cpu comes back online. A separate thread is created
3224 * to govern a gcwq in such state and is called the trustee of the
3227 * Trustee states and their descriptions.
3229 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3230 * new trustee is started with this state.
3232 * IN_CHARGE Once started, trustee will enter this state after
3233 * assuming the manager role and making all existing
3234 * workers rogue. DOWN_PREPARE waits for trustee to
3235 * enter this state. After reaching IN_CHARGE, trustee
3236 * tries to execute the pending worklist until it's empty
3237 * and the state is set to BUTCHER, or the state is set
3240 * BUTCHER Command state which is set by the cpu callback after
3241 * the cpu has went down. Once this state is set trustee
3242 * knows that there will be no new works on the worklist
3243 * and once the worklist is empty it can proceed to
3244 * killing idle workers.
3246 * RELEASE Command state which is set by the cpu callback if the
3247 * cpu down has been canceled or it has come online
3248 * again. After recognizing this state, trustee stops
3249 * trying to drain or butcher and clears ROGUE, rebinds
3250 * all remaining workers back to the cpu and releases
3253 * DONE Trustee will enter this state after BUTCHER or RELEASE
3256 * trustee CPU draining
3257 * took over down complete
3258 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3260 * | CPU is back online v return workers |
3261 * ----------------> RELEASE --------------
3265 * trustee_wait_event_timeout - timed event wait for trustee
3266 * @cond: condition to wait for
3267 * @timeout: timeout in jiffies
3269 * wait_event_timeout() for trustee to use. Handles locking and
3270 * checks for RELEASE request.
3273 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3274 * multiple times. To be used by trustee.
3277 * Positive indicating left time if @cond is satisfied, 0 if timed
3278 * out, -1 if canceled.
3280 #define trustee_wait_event_timeout(cond, timeout) ({ \
3281 long __ret = (timeout); \
3282 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3284 spin_unlock_irq(&gcwq->lock); \
3285 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3286 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3288 spin_lock_irq(&gcwq->lock); \
3290 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3294 * trustee_wait_event - event wait for trustee
3295 * @cond: condition to wait for
3297 * wait_event() for trustee to use. Automatically handles locking and
3298 * checks for CANCEL request.
3301 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3302 * multiple times. To be used by trustee.
3305 * 0 if @cond is satisfied, -1 if canceled.
3307 #define trustee_wait_event(cond) ({ \
3309 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3310 __ret1 < 0 ? -1 : 0; \
3313 static int __cpuinit trustee_thread(void *__gcwq)
3315 struct global_cwq *gcwq = __gcwq;
3316 struct worker *worker;
3317 struct work_struct *work;
3318 struct hlist_node *pos;
3322 BUG_ON(gcwq->cpu != smp_processor_id());
3324 spin_lock_irq(&gcwq->lock);
3326 * Claim the manager position and make all workers rogue.
3327 * Trustee must be bound to the target cpu and can't be
3330 BUG_ON(gcwq->cpu != smp_processor_id());
3331 rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3334 gcwq->flags |= GCWQ_MANAGING_WORKERS;
3336 list_for_each_entry(worker, &gcwq->idle_list, entry)
3337 worker->flags |= WORKER_ROGUE;
3339 for_each_busy_worker(worker, i, pos, gcwq)
3340 worker->flags |= WORKER_ROGUE;
3343 * Call schedule() so that we cross rq->lock and thus can
3344 * guarantee sched callbacks see the rogue flag. This is
3345 * necessary as scheduler callbacks may be invoked from other
3348 spin_unlock_irq(&gcwq->lock);
3350 spin_lock_irq(&gcwq->lock);
3353 * Sched callbacks are disabled now. Zap nr_running. After
3354 * this, nr_running stays zero and need_more_worker() and
3355 * keep_working() are always true as long as the worklist is
3358 atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3360 spin_unlock_irq(&gcwq->lock);
3361 del_timer_sync(&gcwq->idle_timer);
3362 spin_lock_irq(&gcwq->lock);
3365 * We're now in charge. Notify and proceed to drain. We need
3366 * to keep the gcwq running during the whole CPU down
3367 * procedure as other cpu hotunplug callbacks may need to
3368 * flush currently running tasks.
3370 gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3371 wake_up_all(&gcwq->trustee_wait);
3374 * The original cpu is in the process of dying and may go away
3375 * anytime now. When that happens, we and all workers would
3376 * be migrated to other cpus. Try draining any left work. We
3377 * want to get it over with ASAP - spam rescuers, wake up as
3378 * many idlers as necessary and create new ones till the
3379 * worklist is empty. Note that if the gcwq is frozen, there
3380 * may be frozen works in freezable cwqs. Don't declare
3381 * completion while frozen.
3383 while (gcwq->nr_workers != gcwq->nr_idle ||
3384 gcwq->flags & GCWQ_FREEZING ||
3385 gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3388 list_for_each_entry(work, &gcwq->worklist, entry) {
3393 list_for_each_entry(worker, &gcwq->idle_list, entry) {
3396 wake_up_process(worker->task);
3399 if (need_to_create_worker(gcwq)) {
3400 spin_unlock_irq(&gcwq->lock);
3401 worker = create_worker(gcwq, false);
3402 spin_lock_irq(&gcwq->lock);
3404 worker->flags |= WORKER_ROGUE;
3405 start_worker(worker);
3409 /* give a breather */
3410 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3415 * Either all works have been scheduled and cpu is down, or
3416 * cpu down has already been canceled. Wait for and butcher
3417 * all workers till we're canceled.
3420 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3421 while (!list_empty(&gcwq->idle_list))
3422 destroy_worker(list_first_entry(&gcwq->idle_list,
3423 struct worker, entry));
3424 } while (gcwq->nr_workers && rc >= 0);
3427 * At this point, either draining has completed and no worker
3428 * is left, or cpu down has been canceled or the cpu is being
3429 * brought back up. There shouldn't be any idle one left.
3430 * Tell the remaining busy ones to rebind once it finishes the
3431 * currently scheduled works by scheduling the rebind_work.
3433 WARN_ON(!list_empty(&gcwq->idle_list));
3435 for_each_busy_worker(worker, i, pos, gcwq) {
3436 struct work_struct *rebind_work = &worker->rebind_work;
3439 * Rebind_work may race with future cpu hotplug
3440 * operations. Use a separate flag to mark that
3441 * rebinding is scheduled.
3443 worker->flags |= WORKER_REBIND;
3444 worker->flags &= ~WORKER_ROGUE;
3446 /* queue rebind_work, wq doesn't matter, use the default one */
3447 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3448 work_data_bits(rebind_work)))
3451 debug_work_activate(rebind_work);
3452 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3453 worker->scheduled.next,
3454 work_color_to_flags(WORK_NO_COLOR));
3457 /* relinquish manager role */
3458 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3460 /* notify completion */
3461 gcwq->trustee = NULL;
3462 gcwq->trustee_state = TRUSTEE_DONE;
3463 wake_up_all(&gcwq->trustee_wait);
3464 spin_unlock_irq(&gcwq->lock);
3469 * wait_trustee_state - wait for trustee to enter the specified state
3470 * @gcwq: gcwq the trustee of interest belongs to
3471 * @state: target state to wait for
3473 * Wait for the trustee to reach @state. DONE is already matched.
3476 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3477 * multiple times. To be used by cpu_callback.
3479 static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3480 __releases(&gcwq->lock)
3481 __acquires(&gcwq->lock)
3483 if (!(gcwq->trustee_state == state ||
3484 gcwq->trustee_state == TRUSTEE_DONE)) {
3485 spin_unlock_irq(&gcwq->lock);
3486 __wait_event(gcwq->trustee_wait,
3487 gcwq->trustee_state == state ||
3488 gcwq->trustee_state == TRUSTEE_DONE);
3489 spin_lock_irq(&gcwq->lock);
3493 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3494 unsigned long action,
3497 unsigned int cpu = (unsigned long)hcpu;
3498 struct global_cwq *gcwq = get_gcwq(cpu);
3499 struct task_struct *new_trustee = NULL;
3500 struct worker *uninitialized_var(new_worker);
3501 unsigned long flags;
3503 action &= ~CPU_TASKS_FROZEN;
3506 case CPU_DOWN_PREPARE:
3507 new_trustee = kthread_create(trustee_thread, gcwq,
3508 "workqueue_trustee/%d\n", cpu);
3509 if (IS_ERR(new_trustee))
3510 return notifier_from_errno(PTR_ERR(new_trustee));
3511 kthread_bind(new_trustee, cpu);
3513 case CPU_UP_PREPARE:
3514 BUG_ON(gcwq->first_idle);
3515 new_worker = create_worker(gcwq, false);
3518 kthread_stop(new_trustee);
3523 /* some are called w/ irq disabled, don't disturb irq status */
3524 spin_lock_irqsave(&gcwq->lock, flags);
3527 case CPU_DOWN_PREPARE:
3528 /* initialize trustee and tell it to acquire the gcwq */
3529 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3530 gcwq->trustee = new_trustee;
3531 gcwq->trustee_state = TRUSTEE_START;
3532 wake_up_process(gcwq->trustee);
3533 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3535 case CPU_UP_PREPARE:
3536 BUG_ON(gcwq->first_idle);
3537 gcwq->first_idle = new_worker;
3542 * Before this, the trustee and all workers except for
3543 * the ones which are still executing works from
3544 * before the last CPU down must be on the cpu. After
3545 * this, they'll all be diasporas.
3547 gcwq->flags |= GCWQ_DISASSOCIATED;
3551 gcwq->trustee_state = TRUSTEE_BUTCHER;
3553 case CPU_UP_CANCELED:
3554 destroy_worker(gcwq->first_idle);
3555 gcwq->first_idle = NULL;
3558 case CPU_DOWN_FAILED:
3560 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3561 if (gcwq->trustee_state != TRUSTEE_DONE) {
3562 gcwq->trustee_state = TRUSTEE_RELEASE;
3563 wake_up_process(gcwq->trustee);
3564 wait_trustee_state(gcwq, TRUSTEE_DONE);
3568 * Trustee is done and there might be no worker left.
3569 * Put the first_idle in and request a real manager to
3572 spin_unlock_irq(&gcwq->lock);
3573 kthread_bind(gcwq->first_idle->task, cpu);
3574 spin_lock_irq(&gcwq->lock);
3575 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3576 start_worker(gcwq->first_idle);
3577 gcwq->first_idle = NULL;
3581 spin_unlock_irqrestore(&gcwq->lock, flags);
3583 return notifier_from_errno(0);
3588 struct work_for_cpu {
3589 struct completion completion;
3595 static int do_work_for_cpu(void *_wfc)
3597 struct work_for_cpu *wfc = _wfc;
3598 wfc->ret = wfc->fn(wfc->arg);
3599 complete(&wfc->completion);
3604 * work_on_cpu - run a function in user context on a particular cpu
3605 * @cpu: the cpu to run on
3606 * @fn: the function to run
3607 * @arg: the function arg
3609 * This will return the value @fn returns.
3610 * It is up to the caller to ensure that the cpu doesn't go offline.
3611 * The caller must not hold any locks which would prevent @fn from completing.
3613 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3615 struct task_struct *sub_thread;
3616 struct work_for_cpu wfc = {
3617 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3622 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3623 if (IS_ERR(sub_thread))
3624 return PTR_ERR(sub_thread);
3625 kthread_bind(sub_thread, cpu);
3626 wake_up_process(sub_thread);
3627 wait_for_completion(&wfc.completion);
3630 EXPORT_SYMBOL_GPL(work_on_cpu);
3631 #endif /* CONFIG_SMP */
3633 #ifdef CONFIG_FREEZER
3636 * freeze_workqueues_begin - begin freezing workqueues
3638 * Start freezing workqueues. After this function returns, all freezable
3639 * workqueues will queue new works to their frozen_works list instead of
3643 * Grabs and releases workqueue_lock and gcwq->lock's.
3645 void freeze_workqueues_begin(void)
3649 spin_lock(&workqueue_lock);
3651 BUG_ON(workqueue_freezing);
3652 workqueue_freezing = true;
3654 for_each_gcwq_cpu(cpu) {
3655 struct global_cwq *gcwq = get_gcwq(cpu);
3656 struct workqueue_struct *wq;
3658 spin_lock_irq(&gcwq->lock);
3660 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3661 gcwq->flags |= GCWQ_FREEZING;
3663 list_for_each_entry(wq, &workqueues, list) {
3664 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3666 if (cwq && wq->flags & WQ_FREEZABLE)
3667 cwq->max_active = 0;
3670 spin_unlock_irq(&gcwq->lock);
3673 spin_unlock(&workqueue_lock);
3677 * freeze_workqueues_busy - are freezable workqueues still busy?
3679 * Check whether freezing is complete. This function must be called
3680 * between freeze_workqueues_begin() and thaw_workqueues().
3683 * Grabs and releases workqueue_lock.
3686 * %true if some freezable workqueues are still busy. %false if freezing
3689 bool freeze_workqueues_busy(void)
3694 spin_lock(&workqueue_lock);
3696 BUG_ON(!workqueue_freezing);
3698 for_each_gcwq_cpu(cpu) {
3699 struct workqueue_struct *wq;
3701 * nr_active is monotonically decreasing. It's safe
3702 * to peek without lock.
3704 list_for_each_entry(wq, &workqueues, list) {
3705 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3707 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3710 BUG_ON(cwq->nr_active < 0);
3711 if (cwq->nr_active) {
3718 spin_unlock(&workqueue_lock);
3723 * thaw_workqueues - thaw workqueues
3725 * Thaw workqueues. Normal queueing is restored and all collected
3726 * frozen works are transferred to their respective gcwq worklists.
3729 * Grabs and releases workqueue_lock and gcwq->lock's.
3731 void thaw_workqueues(void)
3735 spin_lock(&workqueue_lock);
3737 if (!workqueue_freezing)
3740 for_each_gcwq_cpu(cpu) {
3741 struct global_cwq *gcwq = get_gcwq(cpu);
3742 struct workqueue_struct *wq;
3744 spin_lock_irq(&gcwq->lock);
3746 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3747 gcwq->flags &= ~GCWQ_FREEZING;
3749 list_for_each_entry(wq, &workqueues, list) {
3750 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3752 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3755 /* restore max_active and repopulate worklist */
3756 cwq->max_active = wq->saved_max_active;
3758 while (!list_empty(&cwq->delayed_works) &&
3759 cwq->nr_active < cwq->max_active)
3760 cwq_activate_first_delayed(cwq);
3763 wake_up_worker(gcwq);
3765 spin_unlock_irq(&gcwq->lock);
3768 workqueue_freezing = false;
3770 spin_unlock(&workqueue_lock);
3772 #endif /* CONFIG_FREEZER */
3774 static int __init init_workqueues(void)
3779 cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3781 /* initialize gcwqs */
3782 for_each_gcwq_cpu(cpu) {
3783 struct global_cwq *gcwq = get_gcwq(cpu);
3785 spin_lock_init(&gcwq->lock);
3786 INIT_LIST_HEAD(&gcwq->worklist);
3788 gcwq->flags |= GCWQ_DISASSOCIATED;
3790 INIT_LIST_HEAD(&gcwq->idle_list);
3791 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3792 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3794 init_timer_deferrable(&gcwq->idle_timer);
3795 gcwq->idle_timer.function = idle_worker_timeout;
3796 gcwq->idle_timer.data = (unsigned long)gcwq;
3798 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3799 (unsigned long)gcwq);
3801 ida_init(&gcwq->worker_ida);
3803 gcwq->trustee_state = TRUSTEE_DONE;
3804 init_waitqueue_head(&gcwq->trustee_wait);
3807 /* create the initial worker */
3808 for_each_online_gcwq_cpu(cpu) {
3809 struct global_cwq *gcwq = get_gcwq(cpu);
3810 struct worker *worker;
3812 if (cpu != WORK_CPU_UNBOUND)
3813 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3814 worker = create_worker(gcwq, true);
3816 spin_lock_irq(&gcwq->lock);
3817 start_worker(worker);
3818 spin_unlock_irq(&gcwq->lock);
3821 system_wq = alloc_workqueue("events", 0, 0);
3822 system_long_wq = alloc_workqueue("events_long", 0, 0);
3823 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3824 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3825 WQ_UNBOUND_MAX_ACTIVE);
3826 system_freezable_wq = alloc_workqueue("events_freezable",
3828 system_nrt_freezable_wq = alloc_workqueue("events_nrt_freezable",
3829 WQ_NON_REENTRANT | WQ_FREEZABLE, 0);
3830 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
3831 !system_unbound_wq || !system_freezable_wq ||
3832 !system_nrt_freezable_wq);
3835 early_initcall(init_workqueues);