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 BUG_ON(!list_empty(&work->entry));
1037 cwq->nr_in_flight[cwq->work_color]++;
1038 work_flags = work_color_to_flags(cwq->work_color);
1040 if (likely(cwq->nr_active < cwq->max_active)) {
1041 trace_workqueue_activate_work(work);
1043 worklist = gcwq_determine_ins_pos(gcwq, cwq);
1045 work_flags |= WORK_STRUCT_DELAYED;
1046 worklist = &cwq->delayed_works;
1049 insert_work(cwq, work, worklist, work_flags);
1051 spin_unlock_irqrestore(&gcwq->lock, flags);
1055 * queue_work - queue work on a workqueue
1056 * @wq: workqueue to use
1057 * @work: work to queue
1059 * Returns 0 if @work was already on a queue, non-zero otherwise.
1061 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1062 * it can be processed by another CPU.
1064 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1068 ret = queue_work_on(get_cpu(), wq, work);
1073 EXPORT_SYMBOL_GPL(queue_work);
1076 * queue_work_on - queue work on specific cpu
1077 * @cpu: CPU number to execute work on
1078 * @wq: workqueue to use
1079 * @work: work to queue
1081 * Returns 0 if @work was already on a queue, non-zero otherwise.
1083 * We queue the work to a specific CPU, the caller must ensure it
1087 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1091 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1092 __queue_work(cpu, wq, work);
1097 EXPORT_SYMBOL_GPL(queue_work_on);
1099 static void delayed_work_timer_fn(unsigned long __data)
1101 struct delayed_work *dwork = (struct delayed_work *)__data;
1102 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1104 __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1108 * queue_delayed_work - queue work on a workqueue after delay
1109 * @wq: workqueue to use
1110 * @dwork: delayable work to queue
1111 * @delay: number of jiffies to wait before queueing
1113 * Returns 0 if @work was already on a queue, non-zero otherwise.
1115 int queue_delayed_work(struct workqueue_struct *wq,
1116 struct delayed_work *dwork, unsigned long delay)
1119 return queue_work(wq, &dwork->work);
1121 return queue_delayed_work_on(-1, wq, dwork, delay);
1123 EXPORT_SYMBOL_GPL(queue_delayed_work);
1126 * queue_delayed_work_on - queue work on specific CPU after delay
1127 * @cpu: CPU number to execute work on
1128 * @wq: workqueue to use
1129 * @dwork: work to queue
1130 * @delay: number of jiffies to wait before queueing
1132 * Returns 0 if @work was already on a queue, non-zero otherwise.
1134 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1135 struct delayed_work *dwork, unsigned long delay)
1138 struct timer_list *timer = &dwork->timer;
1139 struct work_struct *work = &dwork->work;
1141 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1144 BUG_ON(timer_pending(timer));
1145 BUG_ON(!list_empty(&work->entry));
1147 timer_stats_timer_set_start_info(&dwork->timer);
1150 * This stores cwq for the moment, for the timer_fn.
1151 * Note that the work's gcwq is preserved to allow
1152 * reentrance detection for delayed works.
1154 if (!(wq->flags & WQ_UNBOUND)) {
1155 struct global_cwq *gcwq = get_work_gcwq(work);
1157 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1160 lcpu = raw_smp_processor_id();
1162 lcpu = WORK_CPU_UNBOUND;
1164 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1166 timer->expires = jiffies + delay;
1167 timer->data = (unsigned long)dwork;
1168 timer->function = delayed_work_timer_fn;
1170 if (unlikely(cpu >= 0))
1171 add_timer_on(timer, cpu);
1178 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1181 * worker_enter_idle - enter idle state
1182 * @worker: worker which is entering idle state
1184 * @worker is entering idle state. Update stats and idle timer if
1188 * spin_lock_irq(gcwq->lock).
1190 static void worker_enter_idle(struct worker *worker)
1192 struct global_cwq *gcwq = worker->gcwq;
1194 BUG_ON(worker->flags & WORKER_IDLE);
1195 BUG_ON(!list_empty(&worker->entry) &&
1196 (worker->hentry.next || worker->hentry.pprev));
1198 /* can't use worker_set_flags(), also called from start_worker() */
1199 worker->flags |= WORKER_IDLE;
1201 worker->last_active = jiffies;
1203 /* idle_list is LIFO */
1204 list_add(&worker->entry, &gcwq->idle_list);
1206 if (likely(!(worker->flags & WORKER_ROGUE))) {
1207 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1208 mod_timer(&gcwq->idle_timer,
1209 jiffies + IDLE_WORKER_TIMEOUT);
1211 wake_up_all(&gcwq->trustee_wait);
1214 * Sanity check nr_running. Because trustee releases gcwq->lock
1215 * between setting %WORKER_ROGUE and zapping nr_running, the
1216 * warning may trigger spuriously. Check iff trustee is idle.
1218 WARN_ON_ONCE(gcwq->trustee_state == TRUSTEE_DONE &&
1219 gcwq->nr_workers == gcwq->nr_idle &&
1220 atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1224 * worker_leave_idle - leave idle state
1225 * @worker: worker which is leaving idle state
1227 * @worker is leaving idle state. Update stats.
1230 * spin_lock_irq(gcwq->lock).
1232 static void worker_leave_idle(struct worker *worker)
1234 struct global_cwq *gcwq = worker->gcwq;
1236 BUG_ON(!(worker->flags & WORKER_IDLE));
1237 worker_clr_flags(worker, WORKER_IDLE);
1239 list_del_init(&worker->entry);
1243 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1246 * Works which are scheduled while the cpu is online must at least be
1247 * scheduled to a worker which is bound to the cpu so that if they are
1248 * flushed from cpu callbacks while cpu is going down, they are
1249 * guaranteed to execute on the cpu.
1251 * This function is to be used by rogue workers and rescuers to bind
1252 * themselves to the target cpu and may race with cpu going down or
1253 * coming online. kthread_bind() can't be used because it may put the
1254 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1255 * verbatim as it's best effort and blocking and gcwq may be
1256 * [dis]associated in the meantime.
1258 * This function tries set_cpus_allowed() and locks gcwq and verifies
1259 * the binding against GCWQ_DISASSOCIATED which is set during
1260 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1261 * idle state or fetches works without dropping lock, it can guarantee
1262 * the scheduling requirement described in the first paragraph.
1265 * Might sleep. Called without any lock but returns with gcwq->lock
1269 * %true if the associated gcwq is online (@worker is successfully
1270 * bound), %false if offline.
1272 static bool worker_maybe_bind_and_lock(struct worker *worker)
1273 __acquires(&gcwq->lock)
1275 struct global_cwq *gcwq = worker->gcwq;
1276 struct task_struct *task = worker->task;
1280 * The following call may fail, succeed or succeed
1281 * without actually migrating the task to the cpu if
1282 * it races with cpu hotunplug operation. Verify
1283 * against GCWQ_DISASSOCIATED.
1285 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1286 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1288 spin_lock_irq(&gcwq->lock);
1289 if (gcwq->flags & GCWQ_DISASSOCIATED)
1291 if (task_cpu(task) == gcwq->cpu &&
1292 cpumask_equal(¤t->cpus_allowed,
1293 get_cpu_mask(gcwq->cpu)))
1295 spin_unlock_irq(&gcwq->lock);
1298 * We've raced with CPU hot[un]plug. Give it a breather
1299 * and retry migration. cond_resched() is required here;
1300 * otherwise, we might deadlock against cpu_stop trying to
1301 * bring down the CPU on non-preemptive kernel.
1309 * Function for worker->rebind_work used to rebind rogue busy workers
1310 * to the associated cpu which is coming back online. This is
1311 * scheduled by cpu up but can race with other cpu hotplug operations
1312 * and may be executed twice without intervening cpu down.
1314 static void worker_rebind_fn(struct work_struct *work)
1316 struct worker *worker = container_of(work, struct worker, rebind_work);
1317 struct global_cwq *gcwq = worker->gcwq;
1319 if (worker_maybe_bind_and_lock(worker))
1320 worker_clr_flags(worker, WORKER_REBIND);
1322 spin_unlock_irq(&gcwq->lock);
1325 static struct worker *alloc_worker(void)
1327 struct worker *worker;
1329 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1331 INIT_LIST_HEAD(&worker->entry);
1332 INIT_LIST_HEAD(&worker->scheduled);
1333 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1334 /* on creation a worker is in !idle && prep state */
1335 worker->flags = WORKER_PREP;
1341 * create_worker - create a new workqueue worker
1342 * @gcwq: gcwq the new worker will belong to
1343 * @bind: whether to set affinity to @cpu or not
1345 * Create a new worker which is bound to @gcwq. The returned worker
1346 * can be started by calling start_worker() or destroyed using
1350 * Might sleep. Does GFP_KERNEL allocations.
1353 * Pointer to the newly created worker.
1355 static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1357 bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1358 struct worker *worker = NULL;
1361 spin_lock_irq(&gcwq->lock);
1362 while (ida_get_new(&gcwq->worker_ida, &id)) {
1363 spin_unlock_irq(&gcwq->lock);
1364 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1366 spin_lock_irq(&gcwq->lock);
1368 spin_unlock_irq(&gcwq->lock);
1370 worker = alloc_worker();
1374 worker->gcwq = gcwq;
1377 if (!on_unbound_cpu)
1378 worker->task = kthread_create_on_node(worker_thread,
1380 cpu_to_node(gcwq->cpu),
1381 "kworker/%u:%d", gcwq->cpu, id);
1383 worker->task = kthread_create(worker_thread, worker,
1384 "kworker/u:%d", id);
1385 if (IS_ERR(worker->task))
1389 * A rogue worker will become a regular one if CPU comes
1390 * online later on. Make sure every worker has
1391 * PF_THREAD_BOUND set.
1393 if (bind && !on_unbound_cpu)
1394 kthread_bind(worker->task, gcwq->cpu);
1396 worker->task->flags |= PF_THREAD_BOUND;
1398 worker->flags |= WORKER_UNBOUND;
1404 spin_lock_irq(&gcwq->lock);
1405 ida_remove(&gcwq->worker_ida, id);
1406 spin_unlock_irq(&gcwq->lock);
1413 * start_worker - start a newly created worker
1414 * @worker: worker to start
1416 * Make the gcwq aware of @worker and start it.
1419 * spin_lock_irq(gcwq->lock).
1421 static void start_worker(struct worker *worker)
1423 worker->flags |= WORKER_STARTED;
1424 worker->gcwq->nr_workers++;
1425 worker_enter_idle(worker);
1426 wake_up_process(worker->task);
1430 * destroy_worker - destroy a workqueue worker
1431 * @worker: worker to be destroyed
1433 * Destroy @worker and adjust @gcwq stats accordingly.
1436 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1438 static void destroy_worker(struct worker *worker)
1440 struct global_cwq *gcwq = worker->gcwq;
1441 int id = worker->id;
1443 /* sanity check frenzy */
1444 BUG_ON(worker->current_work);
1445 BUG_ON(!list_empty(&worker->scheduled));
1447 if (worker->flags & WORKER_STARTED)
1449 if (worker->flags & WORKER_IDLE)
1452 list_del_init(&worker->entry);
1453 worker->flags |= WORKER_DIE;
1455 spin_unlock_irq(&gcwq->lock);
1457 kthread_stop(worker->task);
1460 spin_lock_irq(&gcwq->lock);
1461 ida_remove(&gcwq->worker_ida, id);
1464 static void idle_worker_timeout(unsigned long __gcwq)
1466 struct global_cwq *gcwq = (void *)__gcwq;
1468 spin_lock_irq(&gcwq->lock);
1470 if (too_many_workers(gcwq)) {
1471 struct worker *worker;
1472 unsigned long expires;
1474 /* idle_list is kept in LIFO order, check the last one */
1475 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1476 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1478 if (time_before(jiffies, expires))
1479 mod_timer(&gcwq->idle_timer, expires);
1481 /* it's been idle for too long, wake up manager */
1482 gcwq->flags |= GCWQ_MANAGE_WORKERS;
1483 wake_up_worker(gcwq);
1487 spin_unlock_irq(&gcwq->lock);
1490 static bool send_mayday(struct work_struct *work)
1492 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1493 struct workqueue_struct *wq = cwq->wq;
1496 if (!(wq->flags & WQ_RESCUER))
1499 /* mayday mayday mayday */
1500 cpu = cwq->gcwq->cpu;
1501 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1502 if (cpu == WORK_CPU_UNBOUND)
1504 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1505 wake_up_process(wq->rescuer->task);
1509 static void gcwq_mayday_timeout(unsigned long __gcwq)
1511 struct global_cwq *gcwq = (void *)__gcwq;
1512 struct work_struct *work;
1514 spin_lock_irq(&gcwq->lock);
1516 if (need_to_create_worker(gcwq)) {
1518 * We've been trying to create a new worker but
1519 * haven't been successful. We might be hitting an
1520 * allocation deadlock. Send distress signals to
1523 list_for_each_entry(work, &gcwq->worklist, entry)
1527 spin_unlock_irq(&gcwq->lock);
1529 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1533 * maybe_create_worker - create a new worker if necessary
1534 * @gcwq: gcwq to create a new worker for
1536 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1537 * have at least one idle worker on return from this function. If
1538 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1539 * sent to all rescuers with works scheduled on @gcwq to resolve
1540 * possible allocation deadlock.
1542 * On return, need_to_create_worker() is guaranteed to be false and
1543 * may_start_working() true.
1546 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1547 * multiple times. Does GFP_KERNEL allocations. Called only from
1551 * false if no action was taken and gcwq->lock stayed locked, true
1554 static bool maybe_create_worker(struct global_cwq *gcwq)
1555 __releases(&gcwq->lock)
1556 __acquires(&gcwq->lock)
1558 if (!need_to_create_worker(gcwq))
1561 spin_unlock_irq(&gcwq->lock);
1563 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1564 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1567 struct worker *worker;
1569 worker = create_worker(gcwq, true);
1571 del_timer_sync(&gcwq->mayday_timer);
1572 spin_lock_irq(&gcwq->lock);
1573 start_worker(worker);
1574 BUG_ON(need_to_create_worker(gcwq));
1578 if (!need_to_create_worker(gcwq))
1581 __set_current_state(TASK_INTERRUPTIBLE);
1582 schedule_timeout(CREATE_COOLDOWN);
1584 if (!need_to_create_worker(gcwq))
1588 del_timer_sync(&gcwq->mayday_timer);
1589 spin_lock_irq(&gcwq->lock);
1590 if (need_to_create_worker(gcwq))
1596 * maybe_destroy_worker - destroy workers which have been idle for a while
1597 * @gcwq: gcwq to destroy workers for
1599 * Destroy @gcwq workers which have been idle for longer than
1600 * IDLE_WORKER_TIMEOUT.
1603 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1604 * multiple times. Called only from manager.
1607 * false if no action was taken and gcwq->lock stayed locked, true
1610 static bool maybe_destroy_workers(struct global_cwq *gcwq)
1614 while (too_many_workers(gcwq)) {
1615 struct worker *worker;
1616 unsigned long expires;
1618 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1619 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1621 if (time_before(jiffies, expires)) {
1622 mod_timer(&gcwq->idle_timer, expires);
1626 destroy_worker(worker);
1634 * manage_workers - manage worker pool
1637 * Assume the manager role and manage gcwq worker pool @worker belongs
1638 * to. At any given time, there can be only zero or one manager per
1639 * gcwq. The exclusion is handled automatically by this function.
1641 * The caller can safely start processing works on false return. On
1642 * true return, it's guaranteed that need_to_create_worker() is false
1643 * and may_start_working() is true.
1646 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1647 * multiple times. Does GFP_KERNEL allocations.
1650 * false if no action was taken and gcwq->lock stayed locked, true if
1651 * some action was taken.
1653 static bool manage_workers(struct worker *worker)
1655 struct global_cwq *gcwq = worker->gcwq;
1658 if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1661 gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1662 gcwq->flags |= GCWQ_MANAGING_WORKERS;
1665 * Destroy and then create so that may_start_working() is true
1668 ret |= maybe_destroy_workers(gcwq);
1669 ret |= maybe_create_worker(gcwq);
1671 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1674 * The trustee might be waiting to take over the manager
1675 * position, tell it we're done.
1677 if (unlikely(gcwq->trustee))
1678 wake_up_all(&gcwq->trustee_wait);
1684 * move_linked_works - move linked works to a list
1685 * @work: start of series of works to be scheduled
1686 * @head: target list to append @work to
1687 * @nextp: out paramter for nested worklist walking
1689 * Schedule linked works starting from @work to @head. Work series to
1690 * be scheduled starts at @work and includes any consecutive work with
1691 * WORK_STRUCT_LINKED set in its predecessor.
1693 * If @nextp is not NULL, it's updated to point to the next work of
1694 * the last scheduled work. This allows move_linked_works() to be
1695 * nested inside outer list_for_each_entry_safe().
1698 * spin_lock_irq(gcwq->lock).
1700 static void move_linked_works(struct work_struct *work, struct list_head *head,
1701 struct work_struct **nextp)
1703 struct work_struct *n;
1706 * Linked worklist will always end before the end of the list,
1707 * use NULL for list head.
1709 list_for_each_entry_safe_from(work, n, NULL, entry) {
1710 list_move_tail(&work->entry, head);
1711 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1716 * If we're already inside safe list traversal and have moved
1717 * multiple works to the scheduled queue, the next position
1718 * needs to be updated.
1724 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1726 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1727 struct work_struct, entry);
1728 struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1730 trace_workqueue_activate_work(work);
1731 move_linked_works(work, pos, NULL);
1732 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1737 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1738 * @cwq: cwq of interest
1739 * @color: color of work which left the queue
1740 * @delayed: for a delayed work
1742 * A work either has completed or is removed from pending queue,
1743 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1746 * spin_lock_irq(gcwq->lock).
1748 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1751 /* ignore uncolored works */
1752 if (color == WORK_NO_COLOR)
1755 cwq->nr_in_flight[color]--;
1759 if (!list_empty(&cwq->delayed_works)) {
1760 /* one down, submit a delayed one */
1761 if (cwq->nr_active < cwq->max_active)
1762 cwq_activate_first_delayed(cwq);
1766 /* is flush in progress and are we at the flushing tip? */
1767 if (likely(cwq->flush_color != color))
1770 /* are there still in-flight works? */
1771 if (cwq->nr_in_flight[color])
1774 /* this cwq is done, clear flush_color */
1775 cwq->flush_color = -1;
1778 * If this was the last cwq, wake up the first flusher. It
1779 * will handle the rest.
1781 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1782 complete(&cwq->wq->first_flusher->done);
1786 * process_one_work - process single work
1788 * @work: work to process
1790 * Process @work. This function contains all the logics necessary to
1791 * process a single work including synchronization against and
1792 * interaction with other workers on the same cpu, queueing and
1793 * flushing. As long as context requirement is met, any worker can
1794 * call this function to process a work.
1797 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1799 static void process_one_work(struct worker *worker, struct work_struct *work)
1800 __releases(&gcwq->lock)
1801 __acquires(&gcwq->lock)
1803 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1804 struct global_cwq *gcwq = cwq->gcwq;
1805 struct hlist_head *bwh = busy_worker_head(gcwq, work);
1806 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1807 work_func_t f = work->func;
1809 struct worker *collision;
1810 #ifdef CONFIG_LOCKDEP
1812 * It is permissible to free the struct work_struct from
1813 * inside the function that is called from it, this we need to
1814 * take into account for lockdep too. To avoid bogus "held
1815 * lock freed" warnings as well as problems when looking into
1816 * work->lockdep_map, make a copy and use that here.
1818 struct lockdep_map lockdep_map = work->lockdep_map;
1821 * A single work shouldn't be executed concurrently by
1822 * multiple workers on a single cpu. Check whether anyone is
1823 * already processing the work. If so, defer the work to the
1824 * currently executing one.
1826 collision = __find_worker_executing_work(gcwq, bwh, work);
1827 if (unlikely(collision)) {
1828 move_linked_works(work, &collision->scheduled, NULL);
1832 /* claim and process */
1833 debug_work_deactivate(work);
1834 hlist_add_head(&worker->hentry, bwh);
1835 worker->current_work = work;
1836 worker->current_cwq = cwq;
1837 work_color = get_work_color(work);
1839 /* record the current cpu number in the work data and dequeue */
1840 set_work_cpu(work, gcwq->cpu);
1841 list_del_init(&work->entry);
1844 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1845 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1847 if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1848 struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1849 struct work_struct, entry);
1851 if (!list_empty(&gcwq->worklist) &&
1852 get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1853 wake_up_worker(gcwq);
1855 gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1859 * CPU intensive works don't participate in concurrency
1860 * management. They're the scheduler's responsibility.
1862 if (unlikely(cpu_intensive))
1863 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1865 spin_unlock_irq(&gcwq->lock);
1867 work_clear_pending(work);
1868 lock_map_acquire_read(&cwq->wq->lockdep_map);
1869 lock_map_acquire(&lockdep_map);
1870 trace_workqueue_execute_start(work);
1873 * While we must be careful to not use "work" after this, the trace
1874 * point will only record its address.
1876 trace_workqueue_execute_end(work);
1877 lock_map_release(&lockdep_map);
1878 lock_map_release(&cwq->wq->lockdep_map);
1880 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1881 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1883 current->comm, preempt_count(), task_pid_nr(current));
1884 printk(KERN_ERR " last function: ");
1885 print_symbol("%s\n", (unsigned long)f);
1886 debug_show_held_locks(current);
1890 spin_lock_irq(&gcwq->lock);
1892 /* clear cpu intensive status */
1893 if (unlikely(cpu_intensive))
1894 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1896 /* we're done with it, release */
1897 hlist_del_init(&worker->hentry);
1898 worker->current_work = NULL;
1899 worker->current_cwq = NULL;
1900 cwq_dec_nr_in_flight(cwq, work_color, false);
1904 * process_scheduled_works - process scheduled works
1907 * Process all scheduled works. Please note that the scheduled list
1908 * may change while processing a work, so this function repeatedly
1909 * fetches a work from the top and executes it.
1912 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1915 static void process_scheduled_works(struct worker *worker)
1917 while (!list_empty(&worker->scheduled)) {
1918 struct work_struct *work = list_first_entry(&worker->scheduled,
1919 struct work_struct, entry);
1920 process_one_work(worker, work);
1925 * worker_thread - the worker thread function
1928 * The gcwq worker thread function. There's a single dynamic pool of
1929 * these per each cpu. These workers process all works regardless of
1930 * their specific target workqueue. The only exception is works which
1931 * belong to workqueues with a rescuer which will be explained in
1934 static int worker_thread(void *__worker)
1936 struct worker *worker = __worker;
1937 struct global_cwq *gcwq = worker->gcwq;
1939 /* tell the scheduler that this is a workqueue worker */
1940 worker->task->flags |= PF_WQ_WORKER;
1942 spin_lock_irq(&gcwq->lock);
1944 /* DIE can be set only while we're idle, checking here is enough */
1945 if (worker->flags & WORKER_DIE) {
1946 spin_unlock_irq(&gcwq->lock);
1947 worker->task->flags &= ~PF_WQ_WORKER;
1951 worker_leave_idle(worker);
1953 /* no more worker necessary? */
1954 if (!need_more_worker(gcwq))
1957 /* do we need to manage? */
1958 if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1962 * ->scheduled list can only be filled while a worker is
1963 * preparing to process a work or actually processing it.
1964 * Make sure nobody diddled with it while I was sleeping.
1966 BUG_ON(!list_empty(&worker->scheduled));
1969 * When control reaches this point, we're guaranteed to have
1970 * at least one idle worker or that someone else has already
1971 * assumed the manager role.
1973 worker_clr_flags(worker, WORKER_PREP);
1976 struct work_struct *work =
1977 list_first_entry(&gcwq->worklist,
1978 struct work_struct, entry);
1980 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1981 /* optimization path, not strictly necessary */
1982 process_one_work(worker, work);
1983 if (unlikely(!list_empty(&worker->scheduled)))
1984 process_scheduled_works(worker);
1986 move_linked_works(work, &worker->scheduled, NULL);
1987 process_scheduled_works(worker);
1989 } while (keep_working(gcwq));
1991 worker_set_flags(worker, WORKER_PREP, false);
1993 if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1997 * gcwq->lock is held and there's no work to process and no
1998 * need to manage, sleep. Workers are woken up only while
1999 * holding gcwq->lock or from local cpu, so setting the
2000 * current state before releasing gcwq->lock is enough to
2001 * prevent losing any event.
2003 worker_enter_idle(worker);
2004 __set_current_state(TASK_INTERRUPTIBLE);
2005 spin_unlock_irq(&gcwq->lock);
2011 * rescuer_thread - the rescuer thread function
2012 * @__wq: the associated workqueue
2014 * Workqueue rescuer thread function. There's one rescuer for each
2015 * workqueue which has WQ_RESCUER set.
2017 * Regular work processing on a gcwq may block trying to create a new
2018 * worker which uses GFP_KERNEL allocation which has slight chance of
2019 * developing into deadlock if some works currently on the same queue
2020 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2021 * the problem rescuer solves.
2023 * When such condition is possible, the gcwq summons rescuers of all
2024 * workqueues which have works queued on the gcwq and let them process
2025 * those works so that forward progress can be guaranteed.
2027 * This should happen rarely.
2029 static int rescuer_thread(void *__wq)
2031 struct workqueue_struct *wq = __wq;
2032 struct worker *rescuer = wq->rescuer;
2033 struct list_head *scheduled = &rescuer->scheduled;
2034 bool is_unbound = wq->flags & WQ_UNBOUND;
2037 set_user_nice(current, RESCUER_NICE_LEVEL);
2039 set_current_state(TASK_INTERRUPTIBLE);
2041 if (kthread_should_stop())
2045 * See whether any cpu is asking for help. Unbounded
2046 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2048 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2049 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2050 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2051 struct global_cwq *gcwq = cwq->gcwq;
2052 struct work_struct *work, *n;
2054 __set_current_state(TASK_RUNNING);
2055 mayday_clear_cpu(cpu, wq->mayday_mask);
2057 /* migrate to the target cpu if possible */
2058 rescuer->gcwq = gcwq;
2059 worker_maybe_bind_and_lock(rescuer);
2062 * Slurp in all works issued via this workqueue and
2065 BUG_ON(!list_empty(&rescuer->scheduled));
2066 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2067 if (get_work_cwq(work) == cwq)
2068 move_linked_works(work, scheduled, &n);
2070 process_scheduled_works(rescuer);
2073 * Leave this gcwq. If keep_working() is %true, notify a
2074 * regular worker; otherwise, we end up with 0 concurrency
2075 * and stalling the execution.
2077 if (keep_working(gcwq))
2078 wake_up_worker(gcwq);
2080 spin_unlock_irq(&gcwq->lock);
2088 struct work_struct work;
2089 struct completion done;
2092 static void wq_barrier_func(struct work_struct *work)
2094 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2095 complete(&barr->done);
2099 * insert_wq_barrier - insert a barrier work
2100 * @cwq: cwq to insert barrier into
2101 * @barr: wq_barrier to insert
2102 * @target: target work to attach @barr to
2103 * @worker: worker currently executing @target, NULL if @target is not executing
2105 * @barr is linked to @target such that @barr is completed only after
2106 * @target finishes execution. Please note that the ordering
2107 * guarantee is observed only with respect to @target and on the local
2110 * Currently, a queued barrier can't be canceled. This is because
2111 * try_to_grab_pending() can't determine whether the work to be
2112 * grabbed is at the head of the queue and thus can't clear LINKED
2113 * flag of the previous work while there must be a valid next work
2114 * after a work with LINKED flag set.
2116 * Note that when @worker is non-NULL, @target may be modified
2117 * underneath us, so we can't reliably determine cwq from @target.
2120 * spin_lock_irq(gcwq->lock).
2122 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2123 struct wq_barrier *barr,
2124 struct work_struct *target, struct worker *worker)
2126 struct list_head *head;
2127 unsigned int linked = 0;
2130 * debugobject calls are safe here even with gcwq->lock locked
2131 * as we know for sure that this will not trigger any of the
2132 * checks and call back into the fixup functions where we
2135 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2136 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2137 init_completion(&barr->done);
2140 * If @target is currently being executed, schedule the
2141 * barrier to the worker; otherwise, put it after @target.
2144 head = worker->scheduled.next;
2146 unsigned long *bits = work_data_bits(target);
2148 head = target->entry.next;
2149 /* there can already be other linked works, inherit and set */
2150 linked = *bits & WORK_STRUCT_LINKED;
2151 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2154 debug_work_activate(&barr->work);
2155 insert_work(cwq, &barr->work, head,
2156 work_color_to_flags(WORK_NO_COLOR) | linked);
2160 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2161 * @wq: workqueue being flushed
2162 * @flush_color: new flush color, < 0 for no-op
2163 * @work_color: new work color, < 0 for no-op
2165 * Prepare cwqs for workqueue flushing.
2167 * If @flush_color is non-negative, flush_color on all cwqs should be
2168 * -1. If no cwq has in-flight commands at the specified color, all
2169 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2170 * has in flight commands, its cwq->flush_color is set to
2171 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2172 * wakeup logic is armed and %true is returned.
2174 * The caller should have initialized @wq->first_flusher prior to
2175 * calling this function with non-negative @flush_color. If
2176 * @flush_color is negative, no flush color update is done and %false
2179 * If @work_color is non-negative, all cwqs should have the same
2180 * work_color which is previous to @work_color and all will be
2181 * advanced to @work_color.
2184 * mutex_lock(wq->flush_mutex).
2187 * %true if @flush_color >= 0 and there's something to flush. %false
2190 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2191 int flush_color, int work_color)
2196 if (flush_color >= 0) {
2197 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2198 atomic_set(&wq->nr_cwqs_to_flush, 1);
2201 for_each_cwq_cpu(cpu, wq) {
2202 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2203 struct global_cwq *gcwq = cwq->gcwq;
2205 spin_lock_irq(&gcwq->lock);
2207 if (flush_color >= 0) {
2208 BUG_ON(cwq->flush_color != -1);
2210 if (cwq->nr_in_flight[flush_color]) {
2211 cwq->flush_color = flush_color;
2212 atomic_inc(&wq->nr_cwqs_to_flush);
2217 if (work_color >= 0) {
2218 BUG_ON(work_color != work_next_color(cwq->work_color));
2219 cwq->work_color = work_color;
2222 spin_unlock_irq(&gcwq->lock);
2225 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2226 complete(&wq->first_flusher->done);
2232 * flush_workqueue - ensure that any scheduled work has run to completion.
2233 * @wq: workqueue to flush
2235 * Forces execution of the workqueue and blocks until its completion.
2236 * This is typically used in driver shutdown handlers.
2238 * We sleep until all works which were queued on entry have been handled,
2239 * but we are not livelocked by new incoming ones.
2241 void flush_workqueue(struct workqueue_struct *wq)
2243 struct wq_flusher this_flusher = {
2244 .list = LIST_HEAD_INIT(this_flusher.list),
2246 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2250 lock_map_acquire(&wq->lockdep_map);
2251 lock_map_release(&wq->lockdep_map);
2253 mutex_lock(&wq->flush_mutex);
2256 * Start-to-wait phase
2258 next_color = work_next_color(wq->work_color);
2260 if (next_color != wq->flush_color) {
2262 * Color space is not full. The current work_color
2263 * becomes our flush_color and work_color is advanced
2266 BUG_ON(!list_empty(&wq->flusher_overflow));
2267 this_flusher.flush_color = wq->work_color;
2268 wq->work_color = next_color;
2270 if (!wq->first_flusher) {
2271 /* no flush in progress, become the first flusher */
2272 BUG_ON(wq->flush_color != this_flusher.flush_color);
2274 wq->first_flusher = &this_flusher;
2276 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2278 /* nothing to flush, done */
2279 wq->flush_color = next_color;
2280 wq->first_flusher = NULL;
2285 BUG_ON(wq->flush_color == this_flusher.flush_color);
2286 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2287 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2291 * Oops, color space is full, wait on overflow queue.
2292 * The next flush completion will assign us
2293 * flush_color and transfer to flusher_queue.
2295 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2298 mutex_unlock(&wq->flush_mutex);
2300 wait_for_completion(&this_flusher.done);
2303 * Wake-up-and-cascade phase
2305 * First flushers are responsible for cascading flushes and
2306 * handling overflow. Non-first flushers can simply return.
2308 if (wq->first_flusher != &this_flusher)
2311 mutex_lock(&wq->flush_mutex);
2313 /* we might have raced, check again with mutex held */
2314 if (wq->first_flusher != &this_flusher)
2317 wq->first_flusher = NULL;
2319 BUG_ON(!list_empty(&this_flusher.list));
2320 BUG_ON(wq->flush_color != this_flusher.flush_color);
2323 struct wq_flusher *next, *tmp;
2325 /* complete all the flushers sharing the current flush color */
2326 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2327 if (next->flush_color != wq->flush_color)
2329 list_del_init(&next->list);
2330 complete(&next->done);
2333 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2334 wq->flush_color != work_next_color(wq->work_color));
2336 /* this flush_color is finished, advance by one */
2337 wq->flush_color = work_next_color(wq->flush_color);
2339 /* one color has been freed, handle overflow queue */
2340 if (!list_empty(&wq->flusher_overflow)) {
2342 * Assign the same color to all overflowed
2343 * flushers, advance work_color and append to
2344 * flusher_queue. This is the start-to-wait
2345 * phase for these overflowed flushers.
2347 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2348 tmp->flush_color = wq->work_color;
2350 wq->work_color = work_next_color(wq->work_color);
2352 list_splice_tail_init(&wq->flusher_overflow,
2353 &wq->flusher_queue);
2354 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2357 if (list_empty(&wq->flusher_queue)) {
2358 BUG_ON(wq->flush_color != wq->work_color);
2363 * Need to flush more colors. Make the next flusher
2364 * the new first flusher and arm cwqs.
2366 BUG_ON(wq->flush_color == wq->work_color);
2367 BUG_ON(wq->flush_color != next->flush_color);
2369 list_del_init(&next->list);
2370 wq->first_flusher = next;
2372 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2376 * Meh... this color is already done, clear first
2377 * flusher and repeat cascading.
2379 wq->first_flusher = NULL;
2383 mutex_unlock(&wq->flush_mutex);
2385 EXPORT_SYMBOL_GPL(flush_workqueue);
2388 * drain_workqueue - drain a workqueue
2389 * @wq: workqueue to drain
2391 * Wait until the workqueue becomes empty. While draining is in progress,
2392 * only chain queueing is allowed. IOW, only currently pending or running
2393 * work items on @wq can queue further work items on it. @wq is flushed
2394 * repeatedly until it becomes empty. The number of flushing is detemined
2395 * by the depth of chaining and should be relatively short. Whine if it
2398 void drain_workqueue(struct workqueue_struct *wq)
2400 unsigned int flush_cnt = 0;
2404 * __queue_work() needs to test whether there are drainers, is much
2405 * hotter than drain_workqueue() and already looks at @wq->flags.
2406 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2408 spin_lock(&workqueue_lock);
2409 if (!wq->nr_drainers++)
2410 wq->flags |= WQ_DRAINING;
2411 spin_unlock(&workqueue_lock);
2413 flush_workqueue(wq);
2415 for_each_cwq_cpu(cpu, wq) {
2416 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2419 spin_lock_irq(&cwq->gcwq->lock);
2420 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2421 spin_unlock_irq(&cwq->gcwq->lock);
2426 if (++flush_cnt == 10 ||
2427 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2428 pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n",
2429 wq->name, flush_cnt);
2433 spin_lock(&workqueue_lock);
2434 if (!--wq->nr_drainers)
2435 wq->flags &= ~WQ_DRAINING;
2436 spin_unlock(&workqueue_lock);
2438 EXPORT_SYMBOL_GPL(drain_workqueue);
2440 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2441 bool wait_executing)
2443 struct worker *worker = NULL;
2444 struct global_cwq *gcwq;
2445 struct cpu_workqueue_struct *cwq;
2448 gcwq = get_work_gcwq(work);
2452 spin_lock_irq(&gcwq->lock);
2453 if (!list_empty(&work->entry)) {
2455 * See the comment near try_to_grab_pending()->smp_rmb().
2456 * If it was re-queued to a different gcwq under us, we
2457 * are not going to wait.
2460 cwq = get_work_cwq(work);
2461 if (unlikely(!cwq || gcwq != cwq->gcwq))
2463 } else if (wait_executing) {
2464 worker = find_worker_executing_work(gcwq, work);
2467 cwq = worker->current_cwq;
2471 insert_wq_barrier(cwq, barr, work, worker);
2472 spin_unlock_irq(&gcwq->lock);
2475 * If @max_active is 1 or rescuer is in use, flushing another work
2476 * item on the same workqueue may lead to deadlock. Make sure the
2477 * flusher is not running on the same workqueue by verifying write
2480 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2481 lock_map_acquire(&cwq->wq->lockdep_map);
2483 lock_map_acquire_read(&cwq->wq->lockdep_map);
2484 lock_map_release(&cwq->wq->lockdep_map);
2488 spin_unlock_irq(&gcwq->lock);
2493 * flush_work - wait for a work to finish executing the last queueing instance
2494 * @work: the work to flush
2496 * Wait until @work has finished execution. This function considers
2497 * only the last queueing instance of @work. If @work has been
2498 * enqueued across different CPUs on a non-reentrant workqueue or on
2499 * multiple workqueues, @work might still be executing on return on
2500 * some of the CPUs from earlier queueing.
2502 * If @work was queued only on a non-reentrant, ordered or unbound
2503 * workqueue, @work is guaranteed to be idle on return if it hasn't
2504 * been requeued since flush started.
2507 * %true if flush_work() waited for the work to finish execution,
2508 * %false if it was already idle.
2510 bool flush_work(struct work_struct *work)
2512 struct wq_barrier barr;
2514 if (start_flush_work(work, &barr, true)) {
2515 wait_for_completion(&barr.done);
2516 destroy_work_on_stack(&barr.work);
2521 EXPORT_SYMBOL_GPL(flush_work);
2523 static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2525 struct wq_barrier barr;
2526 struct worker *worker;
2528 spin_lock_irq(&gcwq->lock);
2530 worker = find_worker_executing_work(gcwq, work);
2531 if (unlikely(worker))
2532 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2534 spin_unlock_irq(&gcwq->lock);
2536 if (unlikely(worker)) {
2537 wait_for_completion(&barr.done);
2538 destroy_work_on_stack(&barr.work);
2544 static bool wait_on_work(struct work_struct *work)
2551 lock_map_acquire(&work->lockdep_map);
2552 lock_map_release(&work->lockdep_map);
2554 for_each_gcwq_cpu(cpu)
2555 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2560 * flush_work_sync - wait until a work has finished execution
2561 * @work: the work to flush
2563 * Wait until @work has finished execution. On return, it's
2564 * guaranteed that all queueing instances of @work which happened
2565 * before this function is called are finished. In other words, if
2566 * @work hasn't been requeued since this function was called, @work is
2567 * guaranteed to be idle on return.
2570 * %true if flush_work_sync() waited for the work to finish execution,
2571 * %false if it was already idle.
2573 bool flush_work_sync(struct work_struct *work)
2575 struct wq_barrier barr;
2576 bool pending, waited;
2578 /* we'll wait for executions separately, queue barr only if pending */
2579 pending = start_flush_work(work, &barr, false);
2581 /* wait for executions to finish */
2582 waited = wait_on_work(work);
2584 /* wait for the pending one */
2586 wait_for_completion(&barr.done);
2587 destroy_work_on_stack(&barr.work);
2590 return pending || waited;
2592 EXPORT_SYMBOL_GPL(flush_work_sync);
2595 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2596 * so this work can't be re-armed in any way.
2598 static int try_to_grab_pending(struct work_struct *work)
2600 struct global_cwq *gcwq;
2603 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2607 * The queueing is in progress, or it is already queued. Try to
2608 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2610 gcwq = get_work_gcwq(work);
2614 spin_lock_irq(&gcwq->lock);
2615 if (!list_empty(&work->entry)) {
2617 * This work is queued, but perhaps we locked the wrong gcwq.
2618 * In that case we must see the new value after rmb(), see
2619 * insert_work()->wmb().
2622 if (gcwq == get_work_gcwq(work)) {
2623 debug_work_deactivate(work);
2624 list_del_init(&work->entry);
2625 cwq_dec_nr_in_flight(get_work_cwq(work),
2626 get_work_color(work),
2627 *work_data_bits(work) & WORK_STRUCT_DELAYED);
2631 spin_unlock_irq(&gcwq->lock);
2636 static bool __cancel_work_timer(struct work_struct *work,
2637 struct timer_list* timer)
2642 ret = (timer && likely(del_timer(timer)));
2644 ret = try_to_grab_pending(work);
2646 } while (unlikely(ret < 0));
2648 clear_work_data(work);
2653 * cancel_work_sync - cancel a work and wait for it to finish
2654 * @work: the work to cancel
2656 * Cancel @work and wait for its execution to finish. This function
2657 * can be used even if the work re-queues itself or migrates to
2658 * another workqueue. On return from this function, @work is
2659 * guaranteed to be not pending or executing on any CPU.
2661 * cancel_work_sync(&delayed_work->work) must not be used for
2662 * delayed_work's. Use cancel_delayed_work_sync() instead.
2664 * The caller must ensure that the workqueue on which @work was last
2665 * queued can't be destroyed before this function returns.
2668 * %true if @work was pending, %false otherwise.
2670 bool cancel_work_sync(struct work_struct *work)
2672 return __cancel_work_timer(work, NULL);
2674 EXPORT_SYMBOL_GPL(cancel_work_sync);
2677 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2678 * @dwork: the delayed work to flush
2680 * Delayed timer is cancelled and the pending work is queued for
2681 * immediate execution. Like flush_work(), this function only
2682 * considers the last queueing instance of @dwork.
2685 * %true if flush_work() waited for the work to finish execution,
2686 * %false if it was already idle.
2688 bool flush_delayed_work(struct delayed_work *dwork)
2690 if (del_timer_sync(&dwork->timer))
2691 __queue_work(raw_smp_processor_id(),
2692 get_work_cwq(&dwork->work)->wq, &dwork->work);
2693 return flush_work(&dwork->work);
2695 EXPORT_SYMBOL(flush_delayed_work);
2698 * flush_delayed_work_sync - wait for a dwork to finish
2699 * @dwork: the delayed work to flush
2701 * Delayed timer is cancelled and the pending work is queued for
2702 * execution immediately. Other than timer handling, its behavior
2703 * is identical to flush_work_sync().
2706 * %true if flush_work_sync() waited for the work to finish execution,
2707 * %false if it was already idle.
2709 bool flush_delayed_work_sync(struct delayed_work *dwork)
2711 if (del_timer_sync(&dwork->timer))
2712 __queue_work(raw_smp_processor_id(),
2713 get_work_cwq(&dwork->work)->wq, &dwork->work);
2714 return flush_work_sync(&dwork->work);
2716 EXPORT_SYMBOL(flush_delayed_work_sync);
2719 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2720 * @dwork: the delayed work cancel
2722 * This is cancel_work_sync() for delayed works.
2725 * %true if @dwork was pending, %false otherwise.
2727 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2729 return __cancel_work_timer(&dwork->work, &dwork->timer);
2731 EXPORT_SYMBOL(cancel_delayed_work_sync);
2734 * schedule_work - put work task in global workqueue
2735 * @work: job to be done
2737 * Returns zero if @work was already on the kernel-global workqueue and
2738 * non-zero otherwise.
2740 * This puts a job in the kernel-global workqueue if it was not already
2741 * queued and leaves it in the same position on the kernel-global
2742 * workqueue otherwise.
2744 int schedule_work(struct work_struct *work)
2746 return queue_work(system_wq, work);
2748 EXPORT_SYMBOL(schedule_work);
2751 * schedule_work_on - put work task on a specific cpu
2752 * @cpu: cpu to put the work task on
2753 * @work: job to be done
2755 * This puts a job on a specific cpu
2757 int schedule_work_on(int cpu, struct work_struct *work)
2759 return queue_work_on(cpu, system_wq, work);
2761 EXPORT_SYMBOL(schedule_work_on);
2764 * schedule_delayed_work - put work task in global workqueue after delay
2765 * @dwork: job to be done
2766 * @delay: number of jiffies to wait or 0 for immediate execution
2768 * After waiting for a given time this puts a job in the kernel-global
2771 int schedule_delayed_work(struct delayed_work *dwork,
2772 unsigned long delay)
2774 return queue_delayed_work(system_wq, dwork, delay);
2776 EXPORT_SYMBOL(schedule_delayed_work);
2779 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2781 * @dwork: job to be done
2782 * @delay: number of jiffies to wait
2784 * After waiting for a given time this puts a job in the kernel-global
2785 * workqueue on the specified CPU.
2787 int schedule_delayed_work_on(int cpu,
2788 struct delayed_work *dwork, unsigned long delay)
2790 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2792 EXPORT_SYMBOL(schedule_delayed_work_on);
2795 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2796 * @func: the function to call
2798 * schedule_on_each_cpu() executes @func on each online CPU using the
2799 * system workqueue and blocks until all CPUs have completed.
2800 * schedule_on_each_cpu() is very slow.
2803 * 0 on success, -errno on failure.
2805 int schedule_on_each_cpu(work_func_t func)
2808 struct work_struct __percpu *works;
2810 works = alloc_percpu(struct work_struct);
2816 for_each_online_cpu(cpu) {
2817 struct work_struct *work = per_cpu_ptr(works, cpu);
2819 INIT_WORK(work, func);
2820 schedule_work_on(cpu, work);
2823 for_each_online_cpu(cpu)
2824 flush_work(per_cpu_ptr(works, cpu));
2832 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2834 * Forces execution of the kernel-global workqueue and blocks until its
2837 * Think twice before calling this function! It's very easy to get into
2838 * trouble if you don't take great care. Either of the following situations
2839 * will lead to deadlock:
2841 * One of the work items currently on the workqueue needs to acquire
2842 * a lock held by your code or its caller.
2844 * Your code is running in the context of a work routine.
2846 * They will be detected by lockdep when they occur, but the first might not
2847 * occur very often. It depends on what work items are on the workqueue and
2848 * what locks they need, which you have no control over.
2850 * In most situations flushing the entire workqueue is overkill; you merely
2851 * need to know that a particular work item isn't queued and isn't running.
2852 * In such cases you should use cancel_delayed_work_sync() or
2853 * cancel_work_sync() instead.
2855 void flush_scheduled_work(void)
2857 flush_workqueue(system_wq);
2859 EXPORT_SYMBOL(flush_scheduled_work);
2862 * execute_in_process_context - reliably execute the routine with user context
2863 * @fn: the function to execute
2864 * @ew: guaranteed storage for the execute work structure (must
2865 * be available when the work executes)
2867 * Executes the function immediately if process context is available,
2868 * otherwise schedules the function for delayed execution.
2870 * Returns: 0 - function was executed
2871 * 1 - function was scheduled for execution
2873 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2875 if (!in_interrupt()) {
2880 INIT_WORK(&ew->work, fn);
2881 schedule_work(&ew->work);
2885 EXPORT_SYMBOL_GPL(execute_in_process_context);
2887 int keventd_up(void)
2889 return system_wq != NULL;
2892 static int alloc_cwqs(struct workqueue_struct *wq)
2895 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2896 * Make sure that the alignment isn't lower than that of
2897 * unsigned long long.
2899 const size_t size = sizeof(struct cpu_workqueue_struct);
2900 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2901 __alignof__(unsigned long long));
2903 if (!(wq->flags & WQ_UNBOUND))
2904 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2909 * Allocate enough room to align cwq and put an extra
2910 * pointer at the end pointing back to the originally
2911 * allocated pointer which will be used for free.
2913 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2915 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2916 *(void **)(wq->cpu_wq.single + 1) = ptr;
2920 /* just in case, make sure it's actually aligned */
2921 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2922 return wq->cpu_wq.v ? 0 : -ENOMEM;
2925 static void free_cwqs(struct workqueue_struct *wq)
2927 if (!(wq->flags & WQ_UNBOUND))
2928 free_percpu(wq->cpu_wq.pcpu);
2929 else if (wq->cpu_wq.single) {
2930 /* the pointer to free is stored right after the cwq */
2931 kfree(*(void **)(wq->cpu_wq.single + 1));
2935 static int wq_clamp_max_active(int max_active, unsigned int flags,
2938 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2940 if (max_active < 1 || max_active > lim)
2941 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2942 "is out of range, clamping between %d and %d\n",
2943 max_active, name, 1, lim);
2945 return clamp_val(max_active, 1, lim);
2948 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
2951 struct lock_class_key *key,
2952 const char *lock_name, ...)
2954 va_list args, args1;
2955 struct workqueue_struct *wq;
2959 /* determine namelen, allocate wq and format name */
2960 va_start(args, lock_name);
2961 va_copy(args1, args);
2962 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
2964 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
2968 vsnprintf(wq->name, namelen, fmt, args1);
2973 * Workqueues which may be used during memory reclaim should
2974 * have a rescuer to guarantee forward progress.
2976 if (flags & WQ_MEM_RECLAIM)
2977 flags |= WQ_RESCUER;
2980 * Unbound workqueues aren't concurrency managed and should be
2981 * dispatched to workers immediately.
2983 if (flags & WQ_UNBOUND)
2984 flags |= WQ_HIGHPRI;
2986 max_active = max_active ?: WQ_DFL_ACTIVE;
2987 max_active = wq_clamp_max_active(max_active, flags, wq->name);
2991 wq->saved_max_active = max_active;
2992 mutex_init(&wq->flush_mutex);
2993 atomic_set(&wq->nr_cwqs_to_flush, 0);
2994 INIT_LIST_HEAD(&wq->flusher_queue);
2995 INIT_LIST_HEAD(&wq->flusher_overflow);
2997 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2998 INIT_LIST_HEAD(&wq->list);
3000 if (alloc_cwqs(wq) < 0)
3003 for_each_cwq_cpu(cpu, wq) {
3004 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3005 struct global_cwq *gcwq = get_gcwq(cpu);
3007 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3010 cwq->flush_color = -1;
3011 cwq->max_active = max_active;
3012 INIT_LIST_HEAD(&cwq->delayed_works);
3015 if (flags & WQ_RESCUER) {
3016 struct worker *rescuer;
3018 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3021 wq->rescuer = rescuer = alloc_worker();
3025 rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3027 if (IS_ERR(rescuer->task))
3030 rescuer->task->flags |= PF_THREAD_BOUND;
3031 wake_up_process(rescuer->task);
3035 * workqueue_lock protects global freeze state and workqueues
3036 * list. Grab it, set max_active accordingly and add the new
3037 * workqueue to workqueues list.
3039 spin_lock(&workqueue_lock);
3041 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3042 for_each_cwq_cpu(cpu, wq)
3043 get_cwq(cpu, wq)->max_active = 0;
3045 list_add(&wq->list, &workqueues);
3047 spin_unlock(&workqueue_lock);
3053 free_mayday_mask(wq->mayday_mask);
3059 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3062 * destroy_workqueue - safely terminate a workqueue
3063 * @wq: target workqueue
3065 * Safely destroy a workqueue. All work currently pending will be done first.
3067 void destroy_workqueue(struct workqueue_struct *wq)
3071 /* drain it before proceeding with destruction */
3072 drain_workqueue(wq);
3075 * wq list is used to freeze wq, remove from list after
3076 * flushing is complete in case freeze races us.
3078 spin_lock(&workqueue_lock);
3079 list_del(&wq->list);
3080 spin_unlock(&workqueue_lock);
3083 for_each_cwq_cpu(cpu, wq) {
3084 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3087 for (i = 0; i < WORK_NR_COLORS; i++)
3088 BUG_ON(cwq->nr_in_flight[i]);
3089 BUG_ON(cwq->nr_active);
3090 BUG_ON(!list_empty(&cwq->delayed_works));
3093 if (wq->flags & WQ_RESCUER) {
3094 kthread_stop(wq->rescuer->task);
3095 free_mayday_mask(wq->mayday_mask);
3102 EXPORT_SYMBOL_GPL(destroy_workqueue);
3105 * workqueue_set_max_active - adjust max_active of a workqueue
3106 * @wq: target workqueue
3107 * @max_active: new max_active value.
3109 * Set max_active of @wq to @max_active.
3112 * Don't call from IRQ context.
3114 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3118 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3120 spin_lock(&workqueue_lock);
3122 wq->saved_max_active = max_active;
3124 for_each_cwq_cpu(cpu, wq) {
3125 struct global_cwq *gcwq = get_gcwq(cpu);
3127 spin_lock_irq(&gcwq->lock);
3129 if (!(wq->flags & WQ_FREEZABLE) ||
3130 !(gcwq->flags & GCWQ_FREEZING))
3131 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3133 spin_unlock_irq(&gcwq->lock);
3136 spin_unlock(&workqueue_lock);
3138 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3141 * workqueue_congested - test whether a workqueue is congested
3142 * @cpu: CPU in question
3143 * @wq: target workqueue
3145 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3146 * no synchronization around this function and the test result is
3147 * unreliable and only useful as advisory hints or for debugging.
3150 * %true if congested, %false otherwise.
3152 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3154 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3156 return !list_empty(&cwq->delayed_works);
3158 EXPORT_SYMBOL_GPL(workqueue_congested);
3161 * work_cpu - return the last known associated cpu for @work
3162 * @work: the work of interest
3165 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3167 unsigned int work_cpu(struct work_struct *work)
3169 struct global_cwq *gcwq = get_work_gcwq(work);
3171 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3173 EXPORT_SYMBOL_GPL(work_cpu);
3176 * work_busy - test whether a work is currently pending or running
3177 * @work: the work to be tested
3179 * Test whether @work is currently pending or running. There is no
3180 * synchronization around this function and the test result is
3181 * unreliable and only useful as advisory hints or for debugging.
3182 * Especially for reentrant wqs, the pending state might hide the
3186 * OR'd bitmask of WORK_BUSY_* bits.
3188 unsigned int work_busy(struct work_struct *work)
3190 struct global_cwq *gcwq = get_work_gcwq(work);
3191 unsigned long flags;
3192 unsigned int ret = 0;
3197 spin_lock_irqsave(&gcwq->lock, flags);
3199 if (work_pending(work))
3200 ret |= WORK_BUSY_PENDING;
3201 if (find_worker_executing_work(gcwq, work))
3202 ret |= WORK_BUSY_RUNNING;
3204 spin_unlock_irqrestore(&gcwq->lock, flags);
3208 EXPORT_SYMBOL_GPL(work_busy);
3213 * There are two challenges in supporting CPU hotplug. Firstly, there
3214 * are a lot of assumptions on strong associations among work, cwq and
3215 * gcwq which make migrating pending and scheduled works very
3216 * difficult to implement without impacting hot paths. Secondly,
3217 * gcwqs serve mix of short, long and very long running works making
3218 * blocked draining impractical.
3220 * This is solved by allowing a gcwq to be detached from CPU, running
3221 * it with unbound (rogue) workers and allowing it to be reattached
3222 * later if the cpu comes back online. A separate thread is created
3223 * to govern a gcwq in such state and is called the trustee of the
3226 * Trustee states and their descriptions.
3228 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3229 * new trustee is started with this state.
3231 * IN_CHARGE Once started, trustee will enter this state after
3232 * assuming the manager role and making all existing
3233 * workers rogue. DOWN_PREPARE waits for trustee to
3234 * enter this state. After reaching IN_CHARGE, trustee
3235 * tries to execute the pending worklist until it's empty
3236 * and the state is set to BUTCHER, or the state is set
3239 * BUTCHER Command state which is set by the cpu callback after
3240 * the cpu has went down. Once this state is set trustee
3241 * knows that there will be no new works on the worklist
3242 * and once the worklist is empty it can proceed to
3243 * killing idle workers.
3245 * RELEASE Command state which is set by the cpu callback if the
3246 * cpu down has been canceled or it has come online
3247 * again. After recognizing this state, trustee stops
3248 * trying to drain or butcher and clears ROGUE, rebinds
3249 * all remaining workers back to the cpu and releases
3252 * DONE Trustee will enter this state after BUTCHER or RELEASE
3255 * trustee CPU draining
3256 * took over down complete
3257 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3259 * | CPU is back online v return workers |
3260 * ----------------> RELEASE --------------
3264 * trustee_wait_event_timeout - timed event wait for trustee
3265 * @cond: condition to wait for
3266 * @timeout: timeout in jiffies
3268 * wait_event_timeout() for trustee to use. Handles locking and
3269 * checks for RELEASE request.
3272 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3273 * multiple times. To be used by trustee.
3276 * Positive indicating left time if @cond is satisfied, 0 if timed
3277 * out, -1 if canceled.
3279 #define trustee_wait_event_timeout(cond, timeout) ({ \
3280 long __ret = (timeout); \
3281 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3283 spin_unlock_irq(&gcwq->lock); \
3284 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3285 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3287 spin_lock_irq(&gcwq->lock); \
3289 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3293 * trustee_wait_event - event wait for trustee
3294 * @cond: condition to wait for
3296 * wait_event() for trustee to use. Automatically handles locking and
3297 * checks for CANCEL request.
3300 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3301 * multiple times. To be used by trustee.
3304 * 0 if @cond is satisfied, -1 if canceled.
3306 #define trustee_wait_event(cond) ({ \
3308 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3309 __ret1 < 0 ? -1 : 0; \
3312 static int __cpuinit trustee_thread(void *__gcwq)
3314 struct global_cwq *gcwq = __gcwq;
3315 struct worker *worker;
3316 struct work_struct *work;
3317 struct hlist_node *pos;
3321 BUG_ON(gcwq->cpu != smp_processor_id());
3323 spin_lock_irq(&gcwq->lock);
3325 * Claim the manager position and make all workers rogue.
3326 * Trustee must be bound to the target cpu and can't be
3329 BUG_ON(gcwq->cpu != smp_processor_id());
3330 rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3333 gcwq->flags |= GCWQ_MANAGING_WORKERS;
3335 list_for_each_entry(worker, &gcwq->idle_list, entry)
3336 worker->flags |= WORKER_ROGUE;
3338 for_each_busy_worker(worker, i, pos, gcwq)
3339 worker->flags |= WORKER_ROGUE;
3342 * Call schedule() so that we cross rq->lock and thus can
3343 * guarantee sched callbacks see the rogue flag. This is
3344 * necessary as scheduler callbacks may be invoked from other
3347 spin_unlock_irq(&gcwq->lock);
3349 spin_lock_irq(&gcwq->lock);
3352 * Sched callbacks are disabled now. Zap nr_running. After
3353 * this, nr_running stays zero and need_more_worker() and
3354 * keep_working() are always true as long as the worklist is
3357 atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3359 spin_unlock_irq(&gcwq->lock);
3360 del_timer_sync(&gcwq->idle_timer);
3361 spin_lock_irq(&gcwq->lock);
3364 * We're now in charge. Notify and proceed to drain. We need
3365 * to keep the gcwq running during the whole CPU down
3366 * procedure as other cpu hotunplug callbacks may need to
3367 * flush currently running tasks.
3369 gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3370 wake_up_all(&gcwq->trustee_wait);
3373 * The original cpu is in the process of dying and may go away
3374 * anytime now. When that happens, we and all workers would
3375 * be migrated to other cpus. Try draining any left work. We
3376 * want to get it over with ASAP - spam rescuers, wake up as
3377 * many idlers as necessary and create new ones till the
3378 * worklist is empty. Note that if the gcwq is frozen, there
3379 * may be frozen works in freezable cwqs. Don't declare
3380 * completion while frozen.
3382 while (gcwq->nr_workers != gcwq->nr_idle ||
3383 gcwq->flags & GCWQ_FREEZING ||
3384 gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3387 list_for_each_entry(work, &gcwq->worklist, entry) {
3392 list_for_each_entry(worker, &gcwq->idle_list, entry) {
3395 wake_up_process(worker->task);
3398 if (need_to_create_worker(gcwq)) {
3399 spin_unlock_irq(&gcwq->lock);
3400 worker = create_worker(gcwq, false);
3401 spin_lock_irq(&gcwq->lock);
3403 worker->flags |= WORKER_ROGUE;
3404 start_worker(worker);
3408 /* give a breather */
3409 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3414 * Either all works have been scheduled and cpu is down, or
3415 * cpu down has already been canceled. Wait for and butcher
3416 * all workers till we're canceled.
3419 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3420 while (!list_empty(&gcwq->idle_list))
3421 destroy_worker(list_first_entry(&gcwq->idle_list,
3422 struct worker, entry));
3423 } while (gcwq->nr_workers && rc >= 0);
3426 * At this point, either draining has completed and no worker
3427 * is left, or cpu down has been canceled or the cpu is being
3428 * brought back up. There shouldn't be any idle one left.
3429 * Tell the remaining busy ones to rebind once it finishes the
3430 * currently scheduled works by scheduling the rebind_work.
3432 WARN_ON(!list_empty(&gcwq->idle_list));
3434 for_each_busy_worker(worker, i, pos, gcwq) {
3435 struct work_struct *rebind_work = &worker->rebind_work;
3438 * Rebind_work may race with future cpu hotplug
3439 * operations. Use a separate flag to mark that
3440 * rebinding is scheduled.
3442 worker->flags |= WORKER_REBIND;
3443 worker->flags &= ~WORKER_ROGUE;
3445 /* queue rebind_work, wq doesn't matter, use the default one */
3446 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3447 work_data_bits(rebind_work)))
3450 debug_work_activate(rebind_work);
3451 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3452 worker->scheduled.next,
3453 work_color_to_flags(WORK_NO_COLOR));
3456 /* relinquish manager role */
3457 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3459 /* notify completion */
3460 gcwq->trustee = NULL;
3461 gcwq->trustee_state = TRUSTEE_DONE;
3462 wake_up_all(&gcwq->trustee_wait);
3463 spin_unlock_irq(&gcwq->lock);
3468 * wait_trustee_state - wait for trustee to enter the specified state
3469 * @gcwq: gcwq the trustee of interest belongs to
3470 * @state: target state to wait for
3472 * Wait for the trustee to reach @state. DONE is already matched.
3475 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3476 * multiple times. To be used by cpu_callback.
3478 static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3479 __releases(&gcwq->lock)
3480 __acquires(&gcwq->lock)
3482 if (!(gcwq->trustee_state == state ||
3483 gcwq->trustee_state == TRUSTEE_DONE)) {
3484 spin_unlock_irq(&gcwq->lock);
3485 __wait_event(gcwq->trustee_wait,
3486 gcwq->trustee_state == state ||
3487 gcwq->trustee_state == TRUSTEE_DONE);
3488 spin_lock_irq(&gcwq->lock);
3492 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3493 unsigned long action,
3496 unsigned int cpu = (unsigned long)hcpu;
3497 struct global_cwq *gcwq = get_gcwq(cpu);
3498 struct task_struct *new_trustee = NULL;
3499 struct worker *uninitialized_var(new_worker);
3500 unsigned long flags;
3502 action &= ~CPU_TASKS_FROZEN;
3505 case CPU_DOWN_PREPARE:
3506 new_trustee = kthread_create(trustee_thread, gcwq,
3507 "workqueue_trustee/%d\n", cpu);
3508 if (IS_ERR(new_trustee))
3509 return notifier_from_errno(PTR_ERR(new_trustee));
3510 kthread_bind(new_trustee, cpu);
3512 case CPU_UP_PREPARE:
3513 BUG_ON(gcwq->first_idle);
3514 new_worker = create_worker(gcwq, false);
3517 kthread_stop(new_trustee);
3522 /* some are called w/ irq disabled, don't disturb irq status */
3523 spin_lock_irqsave(&gcwq->lock, flags);
3526 case CPU_DOWN_PREPARE:
3527 /* initialize trustee and tell it to acquire the gcwq */
3528 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3529 gcwq->trustee = new_trustee;
3530 gcwq->trustee_state = TRUSTEE_START;
3531 wake_up_process(gcwq->trustee);
3532 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3534 case CPU_UP_PREPARE:
3535 BUG_ON(gcwq->first_idle);
3536 gcwq->first_idle = new_worker;
3541 * Before this, the trustee and all workers except for
3542 * the ones which are still executing works from
3543 * before the last CPU down must be on the cpu. After
3544 * this, they'll all be diasporas.
3546 gcwq->flags |= GCWQ_DISASSOCIATED;
3550 gcwq->trustee_state = TRUSTEE_BUTCHER;
3552 case CPU_UP_CANCELED:
3553 destroy_worker(gcwq->first_idle);
3554 gcwq->first_idle = NULL;
3557 case CPU_DOWN_FAILED:
3559 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3560 if (gcwq->trustee_state != TRUSTEE_DONE) {
3561 gcwq->trustee_state = TRUSTEE_RELEASE;
3562 wake_up_process(gcwq->trustee);
3563 wait_trustee_state(gcwq, TRUSTEE_DONE);
3567 * Trustee is done and there might be no worker left.
3568 * Put the first_idle in and request a real manager to
3571 spin_unlock_irq(&gcwq->lock);
3572 kthread_bind(gcwq->first_idle->task, cpu);
3573 spin_lock_irq(&gcwq->lock);
3574 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3575 start_worker(gcwq->first_idle);
3576 gcwq->first_idle = NULL;
3580 spin_unlock_irqrestore(&gcwq->lock, flags);
3582 return notifier_from_errno(0);
3587 struct work_for_cpu {
3588 struct completion completion;
3594 static int do_work_for_cpu(void *_wfc)
3596 struct work_for_cpu *wfc = _wfc;
3597 wfc->ret = wfc->fn(wfc->arg);
3598 complete(&wfc->completion);
3603 * work_on_cpu - run a function in user context on a particular cpu
3604 * @cpu: the cpu to run on
3605 * @fn: the function to run
3606 * @arg: the function arg
3608 * This will return the value @fn returns.
3609 * It is up to the caller to ensure that the cpu doesn't go offline.
3610 * The caller must not hold any locks which would prevent @fn from completing.
3612 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3614 struct task_struct *sub_thread;
3615 struct work_for_cpu wfc = {
3616 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3621 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3622 if (IS_ERR(sub_thread))
3623 return PTR_ERR(sub_thread);
3624 kthread_bind(sub_thread, cpu);
3625 wake_up_process(sub_thread);
3626 wait_for_completion(&wfc.completion);
3629 EXPORT_SYMBOL_GPL(work_on_cpu);
3630 #endif /* CONFIG_SMP */
3632 #ifdef CONFIG_FREEZER
3635 * freeze_workqueues_begin - begin freezing workqueues
3637 * Start freezing workqueues. After this function returns, all freezable
3638 * workqueues will queue new works to their frozen_works list instead of
3642 * Grabs and releases workqueue_lock and gcwq->lock's.
3644 void freeze_workqueues_begin(void)
3648 spin_lock(&workqueue_lock);
3650 BUG_ON(workqueue_freezing);
3651 workqueue_freezing = true;
3653 for_each_gcwq_cpu(cpu) {
3654 struct global_cwq *gcwq = get_gcwq(cpu);
3655 struct workqueue_struct *wq;
3657 spin_lock_irq(&gcwq->lock);
3659 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3660 gcwq->flags |= GCWQ_FREEZING;
3662 list_for_each_entry(wq, &workqueues, list) {
3663 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3665 if (cwq && wq->flags & WQ_FREEZABLE)
3666 cwq->max_active = 0;
3669 spin_unlock_irq(&gcwq->lock);
3672 spin_unlock(&workqueue_lock);
3676 * freeze_workqueues_busy - are freezable workqueues still busy?
3678 * Check whether freezing is complete. This function must be called
3679 * between freeze_workqueues_begin() and thaw_workqueues().
3682 * Grabs and releases workqueue_lock.
3685 * %true if some freezable workqueues are still busy. %false if freezing
3688 bool freeze_workqueues_busy(void)
3693 spin_lock(&workqueue_lock);
3695 BUG_ON(!workqueue_freezing);
3697 for_each_gcwq_cpu(cpu) {
3698 struct workqueue_struct *wq;
3700 * nr_active is monotonically decreasing. It's safe
3701 * to peek without lock.
3703 list_for_each_entry(wq, &workqueues, list) {
3704 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3706 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3709 BUG_ON(cwq->nr_active < 0);
3710 if (cwq->nr_active) {
3717 spin_unlock(&workqueue_lock);
3722 * thaw_workqueues - thaw workqueues
3724 * Thaw workqueues. Normal queueing is restored and all collected
3725 * frozen works are transferred to their respective gcwq worklists.
3728 * Grabs and releases workqueue_lock and gcwq->lock's.
3730 void thaw_workqueues(void)
3734 spin_lock(&workqueue_lock);
3736 if (!workqueue_freezing)
3739 for_each_gcwq_cpu(cpu) {
3740 struct global_cwq *gcwq = get_gcwq(cpu);
3741 struct workqueue_struct *wq;
3743 spin_lock_irq(&gcwq->lock);
3745 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3746 gcwq->flags &= ~GCWQ_FREEZING;
3748 list_for_each_entry(wq, &workqueues, list) {
3749 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3751 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3754 /* restore max_active and repopulate worklist */
3755 cwq->max_active = wq->saved_max_active;
3757 while (!list_empty(&cwq->delayed_works) &&
3758 cwq->nr_active < cwq->max_active)
3759 cwq_activate_first_delayed(cwq);
3762 wake_up_worker(gcwq);
3764 spin_unlock_irq(&gcwq->lock);
3767 workqueue_freezing = false;
3769 spin_unlock(&workqueue_lock);
3771 #endif /* CONFIG_FREEZER */
3773 static int __init init_workqueues(void)
3778 cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3780 /* initialize gcwqs */
3781 for_each_gcwq_cpu(cpu) {
3782 struct global_cwq *gcwq = get_gcwq(cpu);
3784 spin_lock_init(&gcwq->lock);
3785 INIT_LIST_HEAD(&gcwq->worklist);
3787 gcwq->flags |= GCWQ_DISASSOCIATED;
3789 INIT_LIST_HEAD(&gcwq->idle_list);
3790 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3791 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3793 init_timer_deferrable(&gcwq->idle_timer);
3794 gcwq->idle_timer.function = idle_worker_timeout;
3795 gcwq->idle_timer.data = (unsigned long)gcwq;
3797 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3798 (unsigned long)gcwq);
3800 ida_init(&gcwq->worker_ida);
3802 gcwq->trustee_state = TRUSTEE_DONE;
3803 init_waitqueue_head(&gcwq->trustee_wait);
3806 /* create the initial worker */
3807 for_each_online_gcwq_cpu(cpu) {
3808 struct global_cwq *gcwq = get_gcwq(cpu);
3809 struct worker *worker;
3811 if (cpu != WORK_CPU_UNBOUND)
3812 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3813 worker = create_worker(gcwq, true);
3815 spin_lock_irq(&gcwq->lock);
3816 start_worker(worker);
3817 spin_unlock_irq(&gcwq->lock);
3820 system_wq = alloc_workqueue("events", 0, 0);
3821 system_long_wq = alloc_workqueue("events_long", 0, 0);
3822 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3823 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3824 WQ_UNBOUND_MAX_ACTIVE);
3825 system_freezable_wq = alloc_workqueue("events_freezable",
3827 system_nrt_freezable_wq = alloc_workqueue("events_nrt_freezable",
3828 WQ_NON_REENTRANT | WQ_FREEZABLE, 0);
3829 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
3830 !system_unbound_wq || !system_freezable_wq ||
3831 !system_nrt_freezable_wq);
3834 early_initcall(init_workqueues);