linux - kernel/workqueue.c

3698 lines · 100.0 KiB

1	/*
2	* kernel/workqueue.c - generic async execution with shared worker pool
3	*
4	* Copyright (C) 2002 Ingo Molnar
5	*
6	* Derived from the taskqueue/keventd code by:
7	* David Woodhouse <[email protected]>
8	* Andrew Morton
9	* Kai Petzke <[email protected]>
10	* Theodore Ts'o <[email protected]>
11	*
12	* Made to use alloc_percpu by Christoph Lameter.
13	*
14	* Copyright (C) 2010 SUSE Linux Products GmbH
15	* Copyright (C) 2010 Tejun Heo <[email protected]>
16	*
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.
22	*
23	* Please read Documentation/workqueue.txt for details.
24	*/
25
26	#include <linux/module.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>
44
45	#include "workqueue_sched.h"
46
47	enum {
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 */
54
55	/* worker flags */
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 */
64
65	WORKER_NOT_RUNNING = WORKER_PREP \| WORKER_ROGUE \| WORKER_REBIND \|
66	WORKER_CPU_INTENSIVE \| WORKER_UNBOUND,
67
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 */
74
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,
78
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 */
81
82	MAYDAY_INITIAL_TIMEOUT = HZ / 100, /* call for help after 10ms */
83	MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
84	CREATE_COOLDOWN = HZ, /* time to breath after fail */
85	TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
86
87	/*
88	* Rescue workers are used only on emergencies and shared by
89	* all cpus. Give -20.
90	*/
91	RESCUER_NICE_LEVEL = -20,
92	};
93
94	/*
95	* Structure fields follow one of the following exclusion rules.
96	*
97	* I: Modifiable by initialization/destruction paths and read-only for
98	* everyone else.
99	*
100	* P: Preemption protected. Disabling preemption is enough and should
101	* only be modified and accessed from the local cpu.
102	*
103	* L: gcwq->lock protected. Access with gcwq->lock held.
104	*
105	* X: During normal operation, modification requires gcwq->lock and
106	* should be done only from local cpu. Either disabling preemption
107	* on local cpu or grabbing gcwq->lock is enough for read access.
108	* If GCWQ_DISASSOCIATED is set, it's identical to L.
109	*
110	* F: wq->flush_mutex protected.
111	*
112	* W: workqueue_lock protected.
113	*/
114
115	struct global_cwq;
116
117	/*
118	* The poor guys doing the actual heavy lifting. All on-duty workers
119	* are either serving the manager role, on idle list or on busy hash.
120	*/
121	struct worker {
122	/* on idle list while idle, on busy hash table while busy */
123	union {
124	struct list_head entry; /* L: while idle */
125	struct hlist_node hentry; /* L: while busy */
126	};
127
128	struct work_struct current_work; / L: work being processed */
129	struct cpu_workqueue_struct current_cwq; / L: current_work's cwq */
130	struct list_head scheduled; /* L: scheduled works */
131	struct task_struct task; / I: worker task */
132	struct global_cwq gcwq; / I: the associated gcwq */
133	/* 64 bytes boundary on 64bit, 32 on 32bit */
134	unsigned long last_active; /* L: last active timestamp */
135	unsigned int flags; /* X: flags */
136	int id; /* I: worker id */
137	struct work_struct rebind_work; /* L: rebind worker to cpu */
138	};
139
140	/*
141	* Global per-cpu workqueue. There's one and only one for each cpu
142	* and all works are queued and processed here regardless of their
143	* target workqueues.
144	*/
145	struct global_cwq {
146	spinlock_t lock; /* the gcwq lock */
147	struct list_head worklist; /* L: list of pending works */
148	unsigned int cpu; /* I: the associated cpu */
149	unsigned int flags; /* L: GCWQ_* flags */
150
151	int nr_workers; /* L: total number of workers */
152	int nr_idle; /* L: currently idle ones */
153
154	/* workers are chained either in the idle_list or busy_hash */
155	struct list_head idle_list; /* X: list of idle workers */
156	struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
157	/* L: hash of busy workers */
158
159	struct timer_list idle_timer; /* L: worker idle timeout */
160	struct timer_list mayday_timer; /* L: SOS timer for dworkers */
161
162	struct ida worker_ida; /* L: for worker IDs */
163
164	struct task_struct trustee; / L: for gcwq shutdown */
165	unsigned int trustee_state; /* L: trustee state */
166	wait_queue_head_t trustee_wait; /* trustee wait */
167	struct worker first_idle; / L: first idle worker */
168	} ____cacheline_aligned_in_smp;
169
170	/*
171	* The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
172	* work_struct->data are used for flags and thus cwqs need to be
173	* aligned at two's power of the number of flag bits.
174	*/
175	struct cpu_workqueue_struct {
176	struct global_cwq gcwq; / I: the associated gcwq */
177	struct workqueue_struct wq; / I: the owning workqueue */
178	int work_color; /* L: current color */
179	int flush_color; /* L: flushing color */
180	int nr_in_flight[WORK_NR_COLORS];
181	/* L: nr of in_flight works */
182	int nr_active; /* L: nr of active works */
183	int max_active; /* L: max active works */
184	struct list_head delayed_works; /* L: delayed works */
185	};
186
187	/*
188	* Structure used to wait for workqueue flush.
189	*/
190	struct wq_flusher {
191	struct list_head list; /* F: list of flushers */
192	int flush_color; /* F: flush color waiting for */
193	struct completion done; /* flush completion */
194	};
195
196	/*
197	* All cpumasks are assumed to be always set on UP and thus can't be
198	* used to determine whether there's something to be done.
199	*/
200	#ifdef CONFIG_SMP
201	typedef cpumask_var_t mayday_mask_t;
202	#define mayday_test_and_set_cpu(cpu, mask) \
203	cpumask_test_and_set_cpu((cpu), (mask))
204	#define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
205	#define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
206	#define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
207	#define free_mayday_mask(mask) free_cpumask_var((mask))
208	#else
209	typedef unsigned long mayday_mask_t;
210	#define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
211	#define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
212	#define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
213	#define alloc_mayday_mask(maskp, gfp) true
214	#define free_mayday_mask(mask) do { } while (0)
215	#endif
216
217	/*
218	* The externally visible workqueue abstraction is an array of
219	* per-CPU workqueues:
220	*/
221	struct workqueue_struct {
222	unsigned int flags; /* I: WQ_* flags */
223	union {
224	struct cpu_workqueue_struct __percpu *pcpu;
225	struct cpu_workqueue_struct *single;
226	unsigned long v;
227	} cpu_wq; /* I: cwq's */
228	struct list_head list; /* W: list of all workqueues */
229
230	struct mutex flush_mutex; /* protects wq flushing */
231	int work_color; /* F: current work color */
232	int flush_color; /* F: current flush color */
233	atomic_t nr_cwqs_to_flush; /* flush in progress */
234	struct wq_flusher first_flusher; / F: first flusher */
235	struct list_head flusher_queue; /* F: flush waiters */
236	struct list_head flusher_overflow; /* F: flush overflow list */
237
238	mayday_mask_t mayday_mask; /* cpus requesting rescue */
239	struct worker rescuer; / I: rescue worker */
240
241	int saved_max_active; /* W: saved cwq max_active */
242	const char name; / I: workqueue name */
243	#ifdef CONFIG_LOCKDEP
244	struct lockdep_map lockdep_map;
245	#endif
246	};
247
248	struct workqueue_struct *system_wq __read_mostly;
249	struct workqueue_struct *system_long_wq __read_mostly;
250	struct workqueue_struct *system_nrt_wq __read_mostly;
251	struct workqueue_struct *system_unbound_wq __read_mostly;
252	EXPORT_SYMBOL_GPL(system_wq);
253	EXPORT_SYMBOL_GPL(system_long_wq);
254	EXPORT_SYMBOL_GPL(system_nrt_wq);
255	EXPORT_SYMBOL_GPL(system_unbound_wq);
256
257	#define CREATE_TRACE_POINTS
258	#include <trace/events/workqueue.h>
259
260	#define for_each_busy_worker(worker, i, pos, gcwq) \
261	for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
262	hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
263
264	static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
265	unsigned int sw)
266	{
267	if (cpu < nr_cpu_ids) {
268	if (sw & 1) {
269	cpu = cpumask_next(cpu, mask);
270	if (cpu < nr_cpu_ids)
271	return cpu;
272	}
273	if (sw & 2)
274	return WORK_CPU_UNBOUND;
275	}
276	return WORK_CPU_NONE;
277	}
278
279	static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
280	struct workqueue_struct *wq)
281	{
282	return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
283	}
284
285	/*
286	* CPU iterators
287	*
288	* An extra gcwq is defined for an invalid cpu number
289	* (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
290	* specific CPU. The following iterators are similar to
291	* for_each_*_cpu() iterators but also considers the unbound gcwq.
292	*
293	* for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
294	* for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
295	* for_each_cwq_cpu() : possible CPUs for bound workqueues,
296	* WORK_CPU_UNBOUND for unbound workqueues
297	*/
298	#define for_each_gcwq_cpu(cpu) \
299	for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
300	(cpu) < WORK_CPU_NONE; \
301	(cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
302
303	#define for_each_online_gcwq_cpu(cpu) \
304	for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
305	(cpu) < WORK_CPU_NONE; \
306	(cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
307
308	#define for_each_cwq_cpu(cpu, wq) \
309	for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
310	(cpu) < WORK_CPU_NONE; \
311	(cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
312
313	#ifdef CONFIG_DEBUG_OBJECTS_WORK
314
315	static struct debug_obj_descr work_debug_descr;
316
317	/*
318	* fixup_init is called when:
319	* - an active object is initialized
320	*/
321	static int work_fixup_init(void *addr, enum debug_obj_state state)
322	{
323	struct work_struct *work = addr;
324
325	switch (state) {
326	case ODEBUG_STATE_ACTIVE:
327	cancel_work_sync(work);
328	debug_object_init(work, &work_debug_descr);
329	return 1;
330	default:
331	return 0;
332	}
333	}
334
335	/*
336	* fixup_activate is called when:
337	* - an active object is activated
338	* - an unknown object is activated (might be a statically initialized object)
339	*/
340	static int work_fixup_activate(void *addr, enum debug_obj_state state)
341	{
342	struct work_struct *work = addr;
343
344	switch (state) {
345
346	case ODEBUG_STATE_NOTAVAILABLE:
347	/*
348	* This is not really a fixup. The work struct was
349	* statically initialized. We just make sure that it
350	* is tracked in the object tracker.
351	*/
352	if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
353	debug_object_init(work, &work_debug_descr);
354	debug_object_activate(work, &work_debug_descr);
355	return 0;
356	}
357	WARN_ON_ONCE(1);
358	return 0;
359
360	case ODEBUG_STATE_ACTIVE:
361	WARN_ON(1);
362
363	default:
364	return 0;
365	}
366	}
367
368	/*
369	* fixup_free is called when:
370	* - an active object is freed
371	*/
372	static int work_fixup_free(void *addr, enum debug_obj_state state)
373	{
374	struct work_struct *work = addr;
375
376	switch (state) {
377	case ODEBUG_STATE_ACTIVE:
378	cancel_work_sync(work);
379	debug_object_free(work, &work_debug_descr);
380	return 1;
381	default:
382	return 0;
383	}
384	}
385
386	static struct debug_obj_descr work_debug_descr = {
387	.name = "work_struct",
388	.fixup_init = work_fixup_init,
389	.fixup_activate = work_fixup_activate,
390	.fixup_free = work_fixup_free,
391	};
392
393	static inline void debug_work_activate(struct work_struct *work)
394	{
395	debug_object_activate(work, &work_debug_descr);
396	}
397
398	static inline void debug_work_deactivate(struct work_struct *work)
399	{
400	debug_object_deactivate(work, &work_debug_descr);
401	}
402
403	void __init_work(struct work_struct *work, int onstack)
404	{
405	if (onstack)
406	debug_object_init_on_stack(work, &work_debug_descr);
407	else
408	debug_object_init(work, &work_debug_descr);
409	}
410	EXPORT_SYMBOL_GPL(__init_work);
411
412	void destroy_work_on_stack(struct work_struct *work)
413	{
414	debug_object_free(work, &work_debug_descr);
415	}
416	EXPORT_SYMBOL_GPL(destroy_work_on_stack);
417
418	#else
419	static inline void debug_work_activate(struct work_struct *work) { }
420	static inline void debug_work_deactivate(struct work_struct *work) { }
421	#endif
422
423	/* Serializes the accesses to the list of workqueues. */
424	static DEFINE_SPINLOCK(workqueue_lock);
425	static LIST_HEAD(workqueues);
426	static bool workqueue_freezing; /* W: have wqs started freezing? */
427
428	/*
429	* The almighty global cpu workqueues. nr_running is the only field
430	* which is expected to be used frequently by other cpus via
431	* try_to_wake_up(). Put it in a separate cacheline.
432	*/
433	static DEFINE_PER_CPU(struct global_cwq, global_cwq);
434	static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
435
436	/*
437	* Global cpu workqueue and nr_running counter for unbound gcwq. The
438	* gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
439	* workers have WORKER_UNBOUND set.
440	*/
441	static struct global_cwq unbound_global_cwq;
442	static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
443
444	static int worker_thread(void *__worker);
445
446	static struct global_cwq *get_gcwq(unsigned int cpu)
447	{
448	if (cpu != WORK_CPU_UNBOUND)
449	return &per_cpu(global_cwq, cpu);
450	else
451	return &unbound_global_cwq;
452	}
453
454	static atomic_t *get_gcwq_nr_running(unsigned int cpu)
455	{
456	if (cpu != WORK_CPU_UNBOUND)
457	return &per_cpu(gcwq_nr_running, cpu);
458	else
459	return &unbound_gcwq_nr_running;
460	}
461
462	static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
463	struct workqueue_struct *wq)
464	{
465	if (!(wq->flags & WQ_UNBOUND)) {
466	if (likely(cpu < nr_cpu_ids)) {
467	#ifdef CONFIG_SMP
468	return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
469	#else
470	return wq->cpu_wq.single;
471	#endif
472	}
473	} else if (likely(cpu == WORK_CPU_UNBOUND))
474	return wq->cpu_wq.single;
475	return NULL;
476	}
477
478	static unsigned int work_color_to_flags(int color)
479	{
480	return color << WORK_STRUCT_COLOR_SHIFT;
481	}
482
483	static int get_work_color(struct work_struct *work)
484	{
485	return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
486	((1 << WORK_STRUCT_COLOR_BITS) - 1);
487	}
488
489	static int work_next_color(int color)
490	{
491	return (color + 1) % WORK_NR_COLORS;
492	}
493
494	/*
495	* A work's data points to the cwq with WORK_STRUCT_CWQ set while the
496	* work is on queue. Once execution starts, WORK_STRUCT_CWQ is
497	* cleared and the work data contains the cpu number it was last on.
498	*
499	* set_work_{cwq\|cpu}() and clear_work_data() can be used to set the
500	* cwq, cpu or clear work->data. These functions should only be
501	* called while the work is owned - ie. while the PENDING bit is set.
502	*
503	* get_work_[g]cwq() can be used to obtain the gcwq or cwq
504	* corresponding to a work. gcwq is available once the work has been
505	* queued anywhere after initialization. cwq is available only from
506	* queueing until execution starts.
507	*/
508	static inline void set_work_data(struct work_struct *work, unsigned long data,
509	unsigned long flags)
510	{
511	BUG_ON(!work_pending(work));
512	atomic_long_set(&work->data, data \| flags \| work_static(work));
513	}
514
515	static void set_work_cwq(struct work_struct *work,
516	struct cpu_workqueue_struct *cwq,
517	unsigned long extra_flags)
518	{
519	set_work_data(work, (unsigned long)cwq,
520	WORK_STRUCT_PENDING \| WORK_STRUCT_CWQ \| extra_flags);
521	}
522
523	static void set_work_cpu(struct work_struct *work, unsigned int cpu)
524	{
525	set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
526	}
527
528	static void clear_work_data(struct work_struct *work)
529	{
530	set_work_data(work, WORK_STRUCT_NO_CPU, 0);
531	}
532
533	static struct cpu_workqueue_struct get_work_cwq(struct work_struct work)
534	{
535	unsigned long data = atomic_long_read(&work->data);
536
537	if (data & WORK_STRUCT_CWQ)
538	return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
539	else
540	return NULL;
541	}
542
543	static struct global_cwq get_work_gcwq(struct work_struct work)
544	{
545	unsigned long data = atomic_long_read(&work->data);
546	unsigned int cpu;
547
548	if (data & WORK_STRUCT_CWQ)
549	return ((struct cpu_workqueue_struct *)
550	(data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
551
552	cpu = data >> WORK_STRUCT_FLAG_BITS;
553	if (cpu == WORK_CPU_NONE)
554	return NULL;
555
556	BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
557	return get_gcwq(cpu);
558	}
559
560	/*
561	* Policy functions. These define the policies on how the global
562	* worker pool is managed. Unless noted otherwise, these functions
563	* assume that they're being called with gcwq->lock held.
564	*/
565
566	static bool __need_more_worker(struct global_cwq *gcwq)
567	{
568	return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) \|\|
569	gcwq->flags & GCWQ_HIGHPRI_PENDING;
570	}
571
572	/*
573	* Need to wake up a worker? Called from anything but currently
574	* running workers.
575	*/
576	static bool need_more_worker(struct global_cwq *gcwq)
577	{
578	return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
579	}
580
581	/* Can I start working? Called from busy but !running workers. */
582	static bool may_start_working(struct global_cwq *gcwq)
583	{
584	return gcwq->nr_idle;
585	}
586
587	/* Do I need to keep working? Called from currently running workers. */
588	static bool keep_working(struct global_cwq *gcwq)
589	{
590	atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
591
592	return !list_empty(&gcwq->worklist) &&
593	(atomic_read(nr_running) <= 1 \|\|
594	gcwq->flags & GCWQ_HIGHPRI_PENDING);
595	}
596
597	/* Do we need a new worker? Called from manager. */
598	static bool need_to_create_worker(struct global_cwq *gcwq)
599	{
600	return need_more_worker(gcwq) && !may_start_working(gcwq);
601	}
602
603	/* Do I need to be the manager? */
604	static bool need_to_manage_workers(struct global_cwq *gcwq)
605	{
606	return need_to_create_worker(gcwq) \|\| gcwq->flags & GCWQ_MANAGE_WORKERS;
607	}
608
609	/* Do we have too many workers and should some go away? */
610	static bool too_many_workers(struct global_cwq *gcwq)
611	{
612	bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
613	int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
614	int nr_busy = gcwq->nr_workers - nr_idle;
615
616	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
617	}
618
619	/*
620	* Wake up functions.
621	*/
622
623	/* Return the first worker. Safe with preemption disabled */
624	static struct worker first_worker(struct global_cwq gcwq)
625	{
626	if (unlikely(list_empty(&gcwq->idle_list)))
627	return NULL;
628
629	return list_first_entry(&gcwq->idle_list, struct worker, entry);
630	}
631
632	/**
633	* wake_up_worker - wake up an idle worker
634	* @gcwq: gcwq to wake worker for
635	*
636	* Wake up the first idle worker of @gcwq.
637	*
638	* CONTEXT:
639	* spin_lock_irq(gcwq->lock).
640	*/
641	static void wake_up_worker(struct global_cwq *gcwq)
642	{
643	struct worker *worker = first_worker(gcwq);
644
645	if (likely(worker))
646	wake_up_process(worker->task);
647	}
648
649	/**
650	* wq_worker_waking_up - a worker is waking up
651	* @task: task waking up
652	* @cpu: CPU @task is waking up to
653	*
654	* This function is called during try_to_wake_up() when a worker is
655	* being awoken.
656	*
657	* CONTEXT:
658	* spin_lock_irq(rq->lock)
659	*/
660	void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
661	{
662	struct worker *worker = kthread_data(task);
663
664	if (likely(!(worker->flags & WORKER_NOT_RUNNING)))
665	atomic_inc(get_gcwq_nr_running(cpu));
666	}
667
668	/**
669	* wq_worker_sleeping - a worker is going to sleep
670	* @task: task going to sleep
671	* @cpu: CPU in question, must be the current CPU number
672	*
673	* This function is called during schedule() when a busy worker is
674	* going to sleep. Worker on the same cpu can be woken up by
675	* returning pointer to its task.
676	*
677	* CONTEXT:
678	* spin_lock_irq(rq->lock)
679	*
680	* RETURNS:
681	* Worker task on @cpu to wake up, %NULL if none.
682	*/
683	struct task_struct wq_worker_sleeping(struct task_struct task,
684	unsigned int cpu)
685	{
686	struct worker worker = kthread_data(task), to_wakeup = NULL;
687	struct global_cwq *gcwq = get_gcwq(cpu);
688	atomic_t *nr_running = get_gcwq_nr_running(cpu);
689
690	if (unlikely(worker->flags & WORKER_NOT_RUNNING))
691	return NULL;
692
693	/* this can only happen on the local cpu */
694	BUG_ON(cpu != raw_smp_processor_id());
695
696	/*
697	* The counterpart of the following dec_and_test, implied mb,
698	* worklist not empty test sequence is in insert_work().
699	* Please read comment there.
700	*
701	* NOT_RUNNING is clear. This means that trustee is not in
702	* charge and we're running on the local cpu w/ rq lock held
703	* and preemption disabled, which in turn means that none else
704	* could be manipulating idle_list, so dereferencing idle_list
705	* without gcwq lock is safe.
706	*/
707	if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
708	to_wakeup = first_worker(gcwq);
709	return to_wakeup ? to_wakeup->task : NULL;
710	}
711
712	/**
713	* worker_set_flags - set worker flags and adjust nr_running accordingly
714	* @worker: self
715	* @flags: flags to set
716	* @wakeup: wakeup an idle worker if necessary
717	*
718	* Set @flags in @worker->flags and adjust nr_running accordingly. If
719	* nr_running becomes zero and @wakeup is %true, an idle worker is
720	* woken up.
721	*
722	* CONTEXT:
723	* spin_lock_irq(gcwq->lock)
724	*/
725	static inline void worker_set_flags(struct worker *worker, unsigned int flags,
726	bool wakeup)
727	{
728	struct global_cwq *gcwq = worker->gcwq;
729
730	WARN_ON_ONCE(worker->task != current);
731
732	/*
733	* If transitioning into NOT_RUNNING, adjust nr_running and
734	* wake up an idle worker as necessary if requested by
735	* @wakeup.
736	*/
737	if ((flags & WORKER_NOT_RUNNING) &&
738	!(worker->flags & WORKER_NOT_RUNNING)) {
739	atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
740
741	if (wakeup) {
742	if (atomic_dec_and_test(nr_running) &&
743	!list_empty(&gcwq->worklist))
744	wake_up_worker(gcwq);
745	} else
746	atomic_dec(nr_running);
747	}
748
749	worker->flags \|= flags;
750	}
751
752	/**
753	* worker_clr_flags - clear worker flags and adjust nr_running accordingly
754	* @worker: self
755	* @flags: flags to clear
756	*
757	* Clear @flags in @worker->flags and adjust nr_running accordingly.
758	*
759	* CONTEXT:
760	* spin_lock_irq(gcwq->lock)
761	*/
762	static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
763	{
764	struct global_cwq *gcwq = worker->gcwq;
765	unsigned int oflags = worker->flags;
766
767	WARN_ON_ONCE(worker->task != current);
768
769	worker->flags &= ~flags;
770
771	/* if transitioning out of NOT_RUNNING, increment nr_running */
772	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
773	if (!(worker->flags & WORKER_NOT_RUNNING))
774	atomic_inc(get_gcwq_nr_running(gcwq->cpu));
775	}
776
777	/**
778	* busy_worker_head - return the busy hash head for a work
779	* @gcwq: gcwq of interest
780	* @work: work to be hashed
781	*
782	* Return hash head of @gcwq for @work.
783	*
784	* CONTEXT:
785	* spin_lock_irq(gcwq->lock).
786	*
787	* RETURNS:
788	* Pointer to the hash head.
789	*/
790	static struct hlist_head busy_worker_head(struct global_cwq gcwq,
791	struct work_struct *work)
792	{
793	const int base_shift = ilog2(sizeof(struct work_struct));
794	unsigned long v = (unsigned long)work;
795
796	/* simple shift and fold hash, do we need something better? */
797	v >>= base_shift;
798	v += v >> BUSY_WORKER_HASH_ORDER;
799	v &= BUSY_WORKER_HASH_MASK;
800
801	return &gcwq->busy_hash[v];
802	}
803
804	/**
805	* __find_worker_executing_work - find worker which is executing a work
806	* @gcwq: gcwq of interest
807	* @bwh: hash head as returned by busy_worker_head()
808	* @work: work to find worker for
809	*
810	* Find a worker which is executing @work on @gcwq. @bwh should be
811	* the hash head obtained by calling busy_worker_head() with the same
812	* work.
813	*
814	* CONTEXT:
815	* spin_lock_irq(gcwq->lock).
816	*
817	* RETURNS:
818	* Pointer to worker which is executing @work if found, NULL
819	* otherwise.
820	*/
821	static struct worker __find_worker_executing_work(struct global_cwq gcwq,
822	struct hlist_head *bwh,
823	struct work_struct *work)
824	{
825	struct worker *worker;
826	struct hlist_node *tmp;
827
828	hlist_for_each_entry(worker, tmp, bwh, hentry)
829	if (worker->current_work == work)
830	return worker;
831	return NULL;
832	}
833
834	/**
835	* find_worker_executing_work - find worker which is executing a work
836	* @gcwq: gcwq of interest
837	* @work: work to find worker for
838	*
839	* Find a worker which is executing @work on @gcwq. This function is
840	* identical to __find_worker_executing_work() except that this
841	* function calculates @bwh itself.
842	*
843	* CONTEXT:
844	* spin_lock_irq(gcwq->lock).
845	*
846	* RETURNS:
847	* Pointer to worker which is executing @work if found, NULL
848	* otherwise.
849	*/
850	static struct worker find_worker_executing_work(struct global_cwq gcwq,
851	struct work_struct *work)
852	{
853	return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
854	work);
855	}
856
857	/**
858	* gcwq_determine_ins_pos - find insertion position
859	* @gcwq: gcwq of interest
860	* @cwq: cwq a work is being queued for
861	*
862	* A work for @cwq is about to be queued on @gcwq, determine insertion
863	* position for the work. If @cwq is for HIGHPRI wq, the work is
864	* queued at the head of the queue but in FIFO order with respect to
865	* other HIGHPRI works; otherwise, at the end of the queue. This
866	* function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
867	* there are HIGHPRI works pending.
868	*
869	* CONTEXT:
870	* spin_lock_irq(gcwq->lock).
871	*
872	* RETURNS:
873	* Pointer to inserstion position.
874	*/
875	static inline struct list_head gcwq_determine_ins_pos(struct global_cwq gcwq,
876	struct cpu_workqueue_struct *cwq)
877	{
878	struct work_struct *twork;
879
880	if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
881	return &gcwq->worklist;
882
883	list_for_each_entry(twork, &gcwq->worklist, entry) {
884	struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
885
886	if (!(tcwq->wq->flags & WQ_HIGHPRI))
887	break;
888	}
889
890	gcwq->flags \|= GCWQ_HIGHPRI_PENDING;
891	return &twork->entry;
892	}
893
894	/**
895	* insert_work - insert a work into gcwq
896	* @cwq: cwq @work belongs to
897	* @work: work to insert
898	* @head: insertion point
899	* @extra_flags: extra WORK_STRUCT_* flags to set
900	*
901	* Insert @work which belongs to @cwq into @gcwq after @head.
902	* @extra_flags is or'd to work_struct flags.
903	*
904	* CONTEXT:
905	* spin_lock_irq(gcwq->lock).
906	*/
907	static void insert_work(struct cpu_workqueue_struct *cwq,
908	struct work_struct work, struct list_head head,
909	unsigned int extra_flags)
910	{
911	struct global_cwq *gcwq = cwq->gcwq;
912
913	/* we own @work, set data and link */
914	set_work_cwq(work, cwq, extra_flags);
915
916	/*
917	* Ensure that we get the right work->data if we see the
918	* result of list_add() below, see try_to_grab_pending().
919	*/
920	smp_wmb();
921
922	list_add_tail(&work->entry, head);
923
924	/*
925	* Ensure either worker_sched_deactivated() sees the above
926	* list_add_tail() or we see zero nr_running to avoid workers
927	* lying around lazily while there are works to be processed.
928	*/
929	smp_mb();
930
931	if (__need_more_worker(gcwq))
932	wake_up_worker(gcwq);
933	}
934
935	static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
936	struct work_struct *work)
937	{
938	struct global_cwq *gcwq;
939	struct cpu_workqueue_struct *cwq;
940	struct list_head *worklist;
941	unsigned int work_flags;
942	unsigned long flags;
943
944	debug_work_activate(work);
945
946	if (WARN_ON_ONCE(wq->flags & WQ_DYING))
947	return;
948
949	/* determine gcwq to use */
950	if (!(wq->flags & WQ_UNBOUND)) {
951	struct global_cwq *last_gcwq;
952
953	if (unlikely(cpu == WORK_CPU_UNBOUND))
954	cpu = raw_smp_processor_id();
955
956	/*
957	* It's multi cpu. If @wq is non-reentrant and @work
958	* was previously on a different cpu, it might still
959	* be running there, in which case the work needs to
960	* be queued on that cpu to guarantee non-reentrance.
961	*/
962	gcwq = get_gcwq(cpu);
963	if (wq->flags & WQ_NON_REENTRANT &&
964	(last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
965	struct worker *worker;
966
967	spin_lock_irqsave(&last_gcwq->lock, flags);
968
969	worker = find_worker_executing_work(last_gcwq, work);
970
971	if (worker && worker->current_cwq->wq == wq)
972	gcwq = last_gcwq;
973	else {
974	/* meh... not running there, queue here */
975	spin_unlock_irqrestore(&last_gcwq->lock, flags);
976	spin_lock_irqsave(&gcwq->lock, flags);
977	}
978	} else
979	spin_lock_irqsave(&gcwq->lock, flags);
980	} else {
981	gcwq = get_gcwq(WORK_CPU_UNBOUND);
982	spin_lock_irqsave(&gcwq->lock, flags);
983	}
984
985	/* gcwq determined, get cwq and queue */
986	cwq = get_cwq(gcwq->cpu, wq);
987	trace_workqueue_queue_work(cpu, cwq, work);
988
989	BUG_ON(!list_empty(&work->entry));
990
991	cwq->nr_in_flight[cwq->work_color]++;
992	work_flags = work_color_to_flags(cwq->work_color);
993
994	if (likely(cwq->nr_active < cwq->max_active)) {
995	trace_workqueue_activate_work(work);
996	cwq->nr_active++;
997	worklist = gcwq_determine_ins_pos(gcwq, cwq);
998	} else {
999	work_flags \|= WORK_STRUCT_DELAYED;
1000	worklist = &cwq->delayed_works;
1001	}
1002
1003	insert_work(cwq, work, worklist, work_flags);
1004
1005	spin_unlock_irqrestore(&gcwq->lock, flags);
1006	}
1007
1008	/**
1009	* queue_work - queue work on a workqueue
1010	* @wq: workqueue to use
1011	* @work: work to queue
1012	*
1013	* Returns 0 if @work was already on a queue, non-zero otherwise.
1014	*
1015	* We queue the work to the CPU on which it was submitted, but if the CPU dies
1016	* it can be processed by another CPU.
1017	*/
1018	int queue_work(struct workqueue_struct wq, struct work_struct work)
1019	{
1020	int ret;
1021
1022	ret = queue_work_on(get_cpu(), wq, work);
1023	put_cpu();
1024
1025	return ret;
1026	}
1027	EXPORT_SYMBOL_GPL(queue_work);
1028
1029	/**
1030	* queue_work_on - queue work on specific cpu
1031	* @cpu: CPU number to execute work on
1032	* @wq: workqueue to use
1033	* @work: work to queue
1034	*
1035	* Returns 0 if @work was already on a queue, non-zero otherwise.
1036	*
1037	* We queue the work to a specific CPU, the caller must ensure it
1038	* can't go away.
1039	*/
1040	int
1041	queue_work_on(int cpu, struct workqueue_struct wq, struct work_struct work)
1042	{
1043	int ret = 0;
1044
1045	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1046	__queue_work(cpu, wq, work);
1047	ret = 1;
1048	}
1049	return ret;
1050	}
1051	EXPORT_SYMBOL_GPL(queue_work_on);
1052
1053	static void delayed_work_timer_fn(unsigned long __data)
1054	{
1055	struct delayed_work dwork = (struct delayed_work )__data;
1056	struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1057
1058	__queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1059	}
1060
1061	/**
1062	* queue_delayed_work - queue work on a workqueue after delay
1063	* @wq: workqueue to use
1064	* @dwork: delayable work to queue
1065	* @delay: number of jiffies to wait before queueing
1066	*
1067	* Returns 0 if @work was already on a queue, non-zero otherwise.
1068	*/
1069	int queue_delayed_work(struct workqueue_struct *wq,
1070	struct delayed_work *dwork, unsigned long delay)
1071	{
1072	if (delay == 0)
1073	return queue_work(wq, &dwork->work);
1074
1075	return queue_delayed_work_on(-1, wq, dwork, delay);
1076	}
1077	EXPORT_SYMBOL_GPL(queue_delayed_work);
1078
1079	/**
1080	* queue_delayed_work_on - queue work on specific CPU after delay
1081	* @cpu: CPU number to execute work on
1082	* @wq: workqueue to use
1083	* @dwork: work to queue
1084	* @delay: number of jiffies to wait before queueing
1085	*
1086	* Returns 0 if @work was already on a queue, non-zero otherwise.
1087	*/
1088	int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1089	struct delayed_work *dwork, unsigned long delay)
1090	{
1091	int ret = 0;
1092	struct timer_list *timer = &dwork->timer;
1093	struct work_struct *work = &dwork->work;
1094
1095	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1096	unsigned int lcpu;
1097
1098	BUG_ON(timer_pending(timer));
1099	BUG_ON(!list_empty(&work->entry));
1100
1101	timer_stats_timer_set_start_info(&dwork->timer);
1102
1103	/*
1104	* This stores cwq for the moment, for the timer_fn.
1105	* Note that the work's gcwq is preserved to allow
1106	* reentrance detection for delayed works.
1107	*/
1108	if (!(wq->flags & WQ_UNBOUND)) {
1109	struct global_cwq *gcwq = get_work_gcwq(work);
1110
1111	if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1112	lcpu = gcwq->cpu;
1113	else
1114	lcpu = raw_smp_processor_id();
1115	} else
1116	lcpu = WORK_CPU_UNBOUND;
1117
1118	set_work_cwq(work, get_cwq(lcpu, wq), 0);
1119
1120	timer->expires = jiffies + delay;
1121	timer->data = (unsigned long)dwork;
1122	timer->function = delayed_work_timer_fn;
1123
1124	if (unlikely(cpu >= 0))
1125	add_timer_on(timer, cpu);
1126	else
1127	add_timer(timer);
1128	ret = 1;
1129	}
1130	return ret;
1131	}
1132	EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1133
1134	/**
1135	* worker_enter_idle - enter idle state
1136	* @worker: worker which is entering idle state
1137	*
1138	* @worker is entering idle state. Update stats and idle timer if
1139	* necessary.
1140	*
1141	* LOCKING:
1142	* spin_lock_irq(gcwq->lock).
1143	*/
1144	static void worker_enter_idle(struct worker *worker)
1145	{
1146	struct global_cwq *gcwq = worker->gcwq;
1147
1148	BUG_ON(worker->flags & WORKER_IDLE);
1149	BUG_ON(!list_empty(&worker->entry) &&
1150	(worker->hentry.next \|\| worker->hentry.pprev));
1151
1152	/* can't use worker_set_flags(), also called from start_worker() */
1153	worker->flags \|= WORKER_IDLE;
1154	gcwq->nr_idle++;
1155	worker->last_active = jiffies;
1156
1157	/* idle_list is LIFO */
1158	list_add(&worker->entry, &gcwq->idle_list);
1159
1160	if (likely(!(worker->flags & WORKER_ROGUE))) {
1161	if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1162	mod_timer(&gcwq->idle_timer,
1163	jiffies + IDLE_WORKER_TIMEOUT);
1164	} else
1165	wake_up_all(&gcwq->trustee_wait);
1166
1167	/* sanity check nr_running */
1168	WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
1169	atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1170	}
1171
1172	/**
1173	* worker_leave_idle - leave idle state
1174	* @worker: worker which is leaving idle state
1175	*
1176	* @worker is leaving idle state. Update stats.
1177	*
1178	* LOCKING:
1179	* spin_lock_irq(gcwq->lock).
1180	*/
1181	static void worker_leave_idle(struct worker *worker)
1182	{
1183	struct global_cwq *gcwq = worker->gcwq;
1184
1185	BUG_ON(!(worker->flags & WORKER_IDLE));
1186	worker_clr_flags(worker, WORKER_IDLE);
1187	gcwq->nr_idle--;
1188	list_del_init(&worker->entry);
1189	}
1190
1191	/**
1192	* worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1193	* @worker: self
1194	*
1195	* Works which are scheduled while the cpu is online must at least be
1196	* scheduled to a worker which is bound to the cpu so that if they are
1197	* flushed from cpu callbacks while cpu is going down, they are
1198	* guaranteed to execute on the cpu.
1199	*
1200	* This function is to be used by rogue workers and rescuers to bind
1201	* themselves to the target cpu and may race with cpu going down or
1202	* coming online. kthread_bind() can't be used because it may put the
1203	* worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1204	* verbatim as it's best effort and blocking and gcwq may be
1205	* [dis]associated in the meantime.
1206	*
1207	* This function tries set_cpus_allowed() and locks gcwq and verifies
1208	* the binding against GCWQ_DISASSOCIATED which is set during
1209	* CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1210	* idle state or fetches works without dropping lock, it can guarantee
1211	* the scheduling requirement described in the first paragraph.
1212	*
1213	* CONTEXT:
1214	* Might sleep. Called without any lock but returns with gcwq->lock
1215	* held.
1216	*
1217	* RETURNS:
1218	* %true if the associated gcwq is online (@worker is successfully
1219	* bound), %false if offline.
1220	*/
1221	static bool worker_maybe_bind_and_lock(struct worker *worker)
1222	__acquires(&gcwq->lock)
1223	{
1224	struct global_cwq *gcwq = worker->gcwq;
1225	struct task_struct *task = worker->task;
1226
1227	while (true) {
1228	/*
1229	* The following call may fail, succeed or succeed
1230	* without actually migrating the task to the cpu if
1231	* it races with cpu hotunplug operation. Verify
1232	* against GCWQ_DISASSOCIATED.
1233	*/
1234	if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1235	set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1236
1237	spin_lock_irq(&gcwq->lock);
1238	if (gcwq->flags & GCWQ_DISASSOCIATED)
1239	return false;
1240	if (task_cpu(task) == gcwq->cpu &&
1241	cpumask_equal(&current->cpus_allowed,
1242	get_cpu_mask(gcwq->cpu)))
1243	return true;
1244	spin_unlock_irq(&gcwq->lock);
1245
1246	/* CPU has come up inbetween, retry migration */
1247	cpu_relax();
1248	}
1249	}
1250
1251	/*
1252	* Function for worker->rebind_work used to rebind rogue busy workers
1253	* to the associated cpu which is coming back online. This is
1254	* scheduled by cpu up but can race with other cpu hotplug operations
1255	* and may be executed twice without intervening cpu down.
1256	*/
1257	static void worker_rebind_fn(struct work_struct *work)
1258	{
1259	struct worker *worker = container_of(work, struct worker, rebind_work);
1260	struct global_cwq *gcwq = worker->gcwq;
1261
1262	if (worker_maybe_bind_and_lock(worker))
1263	worker_clr_flags(worker, WORKER_REBIND);
1264
1265	spin_unlock_irq(&gcwq->lock);
1266	}
1267
1268	static struct worker *alloc_worker(void)
1269	{
1270	struct worker *worker;
1271
1272	worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1273	if (worker) {
1274	INIT_LIST_HEAD(&worker->entry);
1275	INIT_LIST_HEAD(&worker->scheduled);
1276	INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1277	/* on creation a worker is in !idle && prep state */
1278	worker->flags = WORKER_PREP;
1279	}
1280	return worker;
1281	}
1282
1283	/**
1284	* create_worker - create a new workqueue worker
1285	* @gcwq: gcwq the new worker will belong to
1286	* @bind: whether to set affinity to @cpu or not
1287	*
1288	* Create a new worker which is bound to @gcwq. The returned worker
1289	* can be started by calling start_worker() or destroyed using
1290	* destroy_worker().
1291	*
1292	* CONTEXT:
1293	* Might sleep. Does GFP_KERNEL allocations.
1294	*
1295	* RETURNS:
1296	* Pointer to the newly created worker.
1297	*/
1298	static struct worker create_worker(struct global_cwq gcwq, bool bind)
1299	{
1300	bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1301	struct worker *worker = NULL;
1302	int id = -1;
1303
1304	spin_lock_irq(&gcwq->lock);
1305	while (ida_get_new(&gcwq->worker_ida, &id)) {
1306	spin_unlock_irq(&gcwq->lock);
1307	if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1308	goto fail;
1309	spin_lock_irq(&gcwq->lock);
1310	}
1311	spin_unlock_irq(&gcwq->lock);
1312
1313	worker = alloc_worker();
1314	if (!worker)
1315	goto fail;
1316
1317	worker->gcwq = gcwq;
1318	worker->id = id;
1319
1320	if (!on_unbound_cpu)
1321	worker->task = kthread_create(worker_thread, worker,
1322	"kworker/%u:%d", gcwq->cpu, id);
1323	else
1324	worker->task = kthread_create(worker_thread, worker,
1325	"kworker/u:%d", id);
1326	if (IS_ERR(worker->task))
1327	goto fail;
1328
1329	/*
1330	* A rogue worker will become a regular one if CPU comes
1331	* online later on. Make sure every worker has
1332	* PF_THREAD_BOUND set.
1333	*/
1334	if (bind && !on_unbound_cpu)
1335	kthread_bind(worker->task, gcwq->cpu);
1336	else {
1337	worker->task->flags \|= PF_THREAD_BOUND;
1338	if (on_unbound_cpu)
1339	worker->flags \|= WORKER_UNBOUND;
1340	}
1341
1342	return worker;
1343	fail:
1344	if (id >= 0) {
1345	spin_lock_irq(&gcwq->lock);
1346	ida_remove(&gcwq->worker_ida, id);
1347	spin_unlock_irq(&gcwq->lock);
1348	}
1349	kfree(worker);
1350	return NULL;
1351	}
1352
1353	/**
1354	* start_worker - start a newly created worker
1355	* @worker: worker to start
1356	*
1357	* Make the gcwq aware of @worker and start it.
1358	*
1359	* CONTEXT:
1360	* spin_lock_irq(gcwq->lock).
1361	*/
1362	static void start_worker(struct worker *worker)
1363	{
1364	worker->flags \|= WORKER_STARTED;
1365	worker->gcwq->nr_workers++;
1366	worker_enter_idle(worker);
1367	wake_up_process(worker->task);
1368	}
1369
1370	/**
1371	* destroy_worker - destroy a workqueue worker
1372	* @worker: worker to be destroyed
1373	*
1374	* Destroy @worker and adjust @gcwq stats accordingly.
1375	*
1376	* CONTEXT:
1377	* spin_lock_irq(gcwq->lock) which is released and regrabbed.
1378	*/
1379	static void destroy_worker(struct worker *worker)
1380	{
1381	struct global_cwq *gcwq = worker->gcwq;
1382	int id = worker->id;
1383
1384	/* sanity check frenzy */
1385	BUG_ON(worker->current_work);
1386	BUG_ON(!list_empty(&worker->scheduled));
1387
1388	if (worker->flags & WORKER_STARTED)
1389	gcwq->nr_workers--;
1390	if (worker->flags & WORKER_IDLE)
1391	gcwq->nr_idle--;
1392
1393	list_del_init(&worker->entry);
1394	worker->flags \|= WORKER_DIE;
1395
1396	spin_unlock_irq(&gcwq->lock);
1397
1398	kthread_stop(worker->task);
1399	kfree(worker);
1400
1401	spin_lock_irq(&gcwq->lock);
1402	ida_remove(&gcwq->worker_ida, id);
1403	}
1404
1405	static void idle_worker_timeout(unsigned long __gcwq)
1406	{
1407	struct global_cwq gcwq = (void )__gcwq;
1408
1409	spin_lock_irq(&gcwq->lock);
1410
1411	if (too_many_workers(gcwq)) {
1412	struct worker *worker;
1413	unsigned long expires;
1414
1415	/* idle_list is kept in LIFO order, check the last one */
1416	worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1417	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1418
1419	if (time_before(jiffies, expires))
1420	mod_timer(&gcwq->idle_timer, expires);
1421	else {
1422	/* it's been idle for too long, wake up manager */
1423	gcwq->flags \|= GCWQ_MANAGE_WORKERS;
1424	wake_up_worker(gcwq);
1425	}
1426	}
1427
1428	spin_unlock_irq(&gcwq->lock);
1429	}
1430
1431	static bool send_mayday(struct work_struct *work)
1432	{
1433	struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1434	struct workqueue_struct *wq = cwq->wq;
1435	unsigned int cpu;
1436
1437	if (!(wq->flags & WQ_RESCUER))
1438	return false;
1439
1440	/* mayday mayday mayday */
1441	cpu = cwq->gcwq->cpu;
1442	/* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1443	if (cpu == WORK_CPU_UNBOUND)
1444	cpu = 0;
1445	if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1446	wake_up_process(wq->rescuer->task);
1447	return true;
1448	}
1449
1450	static void gcwq_mayday_timeout(unsigned long __gcwq)
1451	{
1452	struct global_cwq gcwq = (void )__gcwq;
1453	struct work_struct *work;
1454
1455	spin_lock_irq(&gcwq->lock);
1456
1457	if (need_to_create_worker(gcwq)) {
1458	/*
1459	* We've been trying to create a new worker but
1460	* haven't been successful. We might be hitting an
1461	* allocation deadlock. Send distress signals to
1462	* rescuers.
1463	*/
1464	list_for_each_entry(work, &gcwq->worklist, entry)
1465	send_mayday(work);
1466	}
1467
1468	spin_unlock_irq(&gcwq->lock);
1469
1470	mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1471	}
1472
1473	/**
1474	* maybe_create_worker - create a new worker if necessary
1475	* @gcwq: gcwq to create a new worker for
1476	*
1477	* Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1478	* have at least one idle worker on return from this function. If
1479	* creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1480	* sent to all rescuers with works scheduled on @gcwq to resolve
1481	* possible allocation deadlock.
1482	*
1483	* On return, need_to_create_worker() is guaranteed to be false and
1484	* may_start_working() true.
1485	*
1486	* LOCKING:
1487	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
1488	* multiple times. Does GFP_KERNEL allocations. Called only from
1489	* manager.
1490	*
1491	* RETURNS:
1492	* false if no action was taken and gcwq->lock stayed locked, true
1493	* otherwise.
1494	*/
1495	static bool maybe_create_worker(struct global_cwq *gcwq)
1496	__releases(&gcwq->lock)
1497	__acquires(&gcwq->lock)
1498	{
1499	if (!need_to_create_worker(gcwq))
1500	return false;
1501	restart:
1502	spin_unlock_irq(&gcwq->lock);
1503
1504	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1505	mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1506
1507	while (true) {
1508	struct worker *worker;
1509
1510	worker = create_worker(gcwq, true);
1511	if (worker) {
1512	del_timer_sync(&gcwq->mayday_timer);
1513	spin_lock_irq(&gcwq->lock);
1514	start_worker(worker);
1515	BUG_ON(need_to_create_worker(gcwq));
1516	return true;
1517	}
1518
1519	if (!need_to_create_worker(gcwq))
1520	break;
1521
1522	__set_current_state(TASK_INTERRUPTIBLE);
1523	schedule_timeout(CREATE_COOLDOWN);
1524
1525	if (!need_to_create_worker(gcwq))
1526	break;
1527	}
1528
1529	del_timer_sync(&gcwq->mayday_timer);
1530	spin_lock_irq(&gcwq->lock);
1531	if (need_to_create_worker(gcwq))
1532	goto restart;
1533	return true;
1534	}
1535
1536	/**
1537	* maybe_destroy_worker - destroy workers which have been idle for a while
1538	* @gcwq: gcwq to destroy workers for
1539	*
1540	* Destroy @gcwq workers which have been idle for longer than
1541	* IDLE_WORKER_TIMEOUT.
1542	*
1543	* LOCKING:
1544	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
1545	* multiple times. Called only from manager.
1546	*
1547	* RETURNS:
1548	* false if no action was taken and gcwq->lock stayed locked, true
1549	* otherwise.
1550	*/
1551	static bool maybe_destroy_workers(struct global_cwq *gcwq)
1552	{
1553	bool ret = false;
1554
1555	while (too_many_workers(gcwq)) {
1556	struct worker *worker;
1557	unsigned long expires;
1558
1559	worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1560	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1561
1562	if (time_before(jiffies, expires)) {
1563	mod_timer(&gcwq->idle_timer, expires);
1564	break;
1565	}
1566
1567	destroy_worker(worker);
1568	ret = true;
1569	}
1570
1571	return ret;
1572	}
1573
1574	/**
1575	* manage_workers - manage worker pool
1576	* @worker: self
1577	*
1578	* Assume the manager role and manage gcwq worker pool @worker belongs
1579	* to. At any given time, there can be only zero or one manager per
1580	* gcwq. The exclusion is handled automatically by this function.
1581	*
1582	* The caller can safely start processing works on false return. On
1583	* true return, it's guaranteed that need_to_create_worker() is false
1584	* and may_start_working() is true.
1585	*
1586	* CONTEXT:
1587	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
1588	* multiple times. Does GFP_KERNEL allocations.
1589	*
1590	* RETURNS:
1591	* false if no action was taken and gcwq->lock stayed locked, true if
1592	* some action was taken.
1593	*/
1594	static bool manage_workers(struct worker *worker)
1595	{
1596	struct global_cwq *gcwq = worker->gcwq;
1597	bool ret = false;
1598
1599	if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1600	return ret;
1601
1602	gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1603	gcwq->flags \|= GCWQ_MANAGING_WORKERS;
1604
1605	/*
1606	* Destroy and then create so that may_start_working() is true
1607	* on return.
1608	*/
1609	ret \|= maybe_destroy_workers(gcwq);
1610	ret \|= maybe_create_worker(gcwq);
1611
1612	gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1613
1614	/*
1615	* The trustee might be waiting to take over the manager
1616	* position, tell it we're done.
1617	*/
1618	if (unlikely(gcwq->trustee))
1619	wake_up_all(&gcwq->trustee_wait);
1620
1621	return ret;
1622	}
1623
1624	/**
1625	* move_linked_works - move linked works to a list
1626	* @work: start of series of works to be scheduled
1627	* @head: target list to append @work to
1628	* @nextp: out paramter for nested worklist walking
1629	*
1630	* Schedule linked works starting from @work to @head. Work series to
1631	* be scheduled starts at @work and includes any consecutive work with
1632	* WORK_STRUCT_LINKED set in its predecessor.
1633	*
1634	* If @nextp is not NULL, it's updated to point to the next work of
1635	* the last scheduled work. This allows move_linked_works() to be
1636	* nested inside outer list_for_each_entry_safe().
1637	*
1638	* CONTEXT:
1639	* spin_lock_irq(gcwq->lock).
1640	*/
1641	static void move_linked_works(struct work_struct work, struct list_head head,
1642	struct work_struct **nextp)
1643	{
1644	struct work_struct *n;
1645
1646	/*
1647	* Linked worklist will always end before the end of the list,
1648	* use NULL for list head.
1649	*/
1650	list_for_each_entry_safe_from(work, n, NULL, entry) {
1651	list_move_tail(&work->entry, head);
1652	if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1653	break;
1654	}
1655
1656	/*
1657	* If we're already inside safe list traversal and have moved
1658	* multiple works to the scheduled queue, the next position
1659	* needs to be updated.
1660	*/
1661	if (nextp)
1662	*nextp = n;
1663	}
1664
1665	static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1666	{
1667	struct work_struct *work = list_first_entry(&cwq->delayed_works,
1668	struct work_struct, entry);
1669	struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1670
1671	trace_workqueue_activate_work(work);
1672	move_linked_works(work, pos, NULL);
1673	__clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1674	cwq->nr_active++;
1675	}
1676
1677	/**
1678	* cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1679	* @cwq: cwq of interest
1680	* @color: color of work which left the queue
1681	* @delayed: for a delayed work
1682	*
1683	* A work either has completed or is removed from pending queue,
1684	* decrement nr_in_flight of its cwq and handle workqueue flushing.
1685	*
1686	* CONTEXT:
1687	* spin_lock_irq(gcwq->lock).
1688	*/
1689	static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1690	bool delayed)
1691	{
1692	/* ignore uncolored works */
1693	if (color == WORK_NO_COLOR)
1694	return;
1695
1696	cwq->nr_in_flight[color]--;
1697
1698	if (!delayed) {
1699	cwq->nr_active--;
1700	if (!list_empty(&cwq->delayed_works)) {
1701	/* one down, submit a delayed one */
1702	if (cwq->nr_active < cwq->max_active)
1703	cwq_activate_first_delayed(cwq);
1704	}
1705	}
1706
1707	/* is flush in progress and are we at the flushing tip? */
1708	if (likely(cwq->flush_color != color))
1709	return;
1710
1711	/* are there still in-flight works? */
1712	if (cwq->nr_in_flight[color])
1713	return;
1714
1715	/* this cwq is done, clear flush_color */
1716	cwq->flush_color = -1;
1717
1718	/*
1719	* If this was the last cwq, wake up the first flusher. It
1720	* will handle the rest.
1721	*/
1722	if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1723	complete(&cwq->wq->first_flusher->done);
1724	}
1725
1726	/**
1727	* process_one_work - process single work
1728	* @worker: self
1729	* @work: work to process
1730	*
1731	* Process @work. This function contains all the logics necessary to
1732	* process a single work including synchronization against and
1733	* interaction with other workers on the same cpu, queueing and
1734	* flushing. As long as context requirement is met, any worker can
1735	* call this function to process a work.
1736	*
1737	* CONTEXT:
1738	* spin_lock_irq(gcwq->lock) which is released and regrabbed.
1739	*/
1740	static void process_one_work(struct worker worker, struct work_struct work)
1741	__releases(&gcwq->lock)
1742	__acquires(&gcwq->lock)
1743	{
1744	struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1745	struct global_cwq *gcwq = cwq->gcwq;
1746	struct hlist_head *bwh = busy_worker_head(gcwq, work);
1747	bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1748	work_func_t f = work->func;
1749	int work_color;
1750	struct worker *collision;
1751	#ifdef CONFIG_LOCKDEP
1752	/*
1753	* It is permissible to free the struct work_struct from
1754	* inside the function that is called from it, this we need to
1755	* take into account for lockdep too. To avoid bogus "held
1756	* lock freed" warnings as well as problems when looking into
1757	* work->lockdep_map, make a copy and use that here.
1758	*/
1759	struct lockdep_map lockdep_map = work->lockdep_map;
1760	#endif
1761	/*
1762	* A single work shouldn't be executed concurrently by
1763	* multiple workers on a single cpu. Check whether anyone is
1764	* already processing the work. If so, defer the work to the
1765	* currently executing one.
1766	*/
1767	collision = __find_worker_executing_work(gcwq, bwh, work);
1768	if (unlikely(collision)) {
1769	move_linked_works(work, &collision->scheduled, NULL);
1770	return;
1771	}
1772
1773	/* claim and process */
1774	debug_work_deactivate(work);
1775	hlist_add_head(&worker->hentry, bwh);
1776	worker->current_work = work;
1777	worker->current_cwq = cwq;
1778	work_color = get_work_color(work);
1779
1780	/* record the current cpu number in the work data and dequeue */
1781	set_work_cpu(work, gcwq->cpu);
1782	list_del_init(&work->entry);
1783
1784	/*
1785	* If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1786	* wake up another worker; otherwise, clear HIGHPRI_PENDING.
1787	*/
1788	if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1789	struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1790	struct work_struct, entry);
1791
1792	if (!list_empty(&gcwq->worklist) &&
1793	get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1794	wake_up_worker(gcwq);
1795	else
1796	gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1797	}
1798
1799	/*
1800	* CPU intensive works don't participate in concurrency
1801	* management. They're the scheduler's responsibility.
1802	*/
1803	if (unlikely(cpu_intensive))
1804	worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1805
1806	spin_unlock_irq(&gcwq->lock);
1807
1808	work_clear_pending(work);
1809	lock_map_acquire(&cwq->wq->lockdep_map);
1810	lock_map_acquire(&lockdep_map);
1811	trace_workqueue_execute_start(work);
1812	f(work);
1813	/*
1814	* While we must be careful to not use "work" after this, the trace
1815	* point will only record its address.
1816	*/
1817	trace_workqueue_execute_end(work);
1818	lock_map_release(&lockdep_map);
1819	lock_map_release(&cwq->wq->lockdep_map);
1820
1821	if (unlikely(in_atomic() \|\| lockdep_depth(current) > 0)) {
1822	printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1823	"%s/0x%08x/%d\n",
1824	current->comm, preempt_count(), task_pid_nr(current));
1825	printk(KERN_ERR " last function: ");
1826	print_symbol("%s\n", (unsigned long)f);
1827	debug_show_held_locks(current);
1828	dump_stack();
1829	}
1830
1831	spin_lock_irq(&gcwq->lock);
1832
1833	/* clear cpu intensive status */
1834	if (unlikely(cpu_intensive))
1835	worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1836
1837	/* we're done with it, release */
1838	hlist_del_init(&worker->hentry);
1839	worker->current_work = NULL;
1840	worker->current_cwq = NULL;
1841	cwq_dec_nr_in_flight(cwq, work_color, false);
1842	}
1843
1844	/**
1845	* process_scheduled_works - process scheduled works
1846	* @worker: self
1847	*
1848	* Process all scheduled works. Please note that the scheduled list
1849	* may change while processing a work, so this function repeatedly
1850	* fetches a work from the top and executes it.
1851	*
1852	* CONTEXT:
1853	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
1854	* multiple times.
1855	*/
1856	static void process_scheduled_works(struct worker *worker)
1857	{
1858	while (!list_empty(&worker->scheduled)) {
1859	struct work_struct *work = list_first_entry(&worker->scheduled,
1860	struct work_struct, entry);
1861	process_one_work(worker, work);
1862	}
1863	}
1864
1865	/**
1866	* worker_thread - the worker thread function
1867	* @__worker: self
1868	*
1869	* The gcwq worker thread function. There's a single dynamic pool of
1870	* these per each cpu. These workers process all works regardless of
1871	* their specific target workqueue. The only exception is works which
1872	* belong to workqueues with a rescuer which will be explained in
1873	* rescuer_thread().
1874	*/
1875	static int worker_thread(void *__worker)
1876	{
1877	struct worker *worker = __worker;
1878	struct global_cwq *gcwq = worker->gcwq;
1879
1880	/* tell the scheduler that this is a workqueue worker */
1881	worker->task->flags \|= PF_WQ_WORKER;
1882	woke_up:
1883	spin_lock_irq(&gcwq->lock);
1884
1885	/* DIE can be set only while we're idle, checking here is enough */
1886	if (worker->flags & WORKER_DIE) {
1887	spin_unlock_irq(&gcwq->lock);
1888	worker->task->flags &= ~PF_WQ_WORKER;
1889	return 0;
1890	}
1891
1892	worker_leave_idle(worker);
1893	recheck:
1894	/* no more worker necessary? */
1895	if (!need_more_worker(gcwq))
1896	goto sleep;
1897
1898	/* do we need to manage? */
1899	if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1900	goto recheck;
1901
1902	/*
1903	* ->scheduled list can only be filled while a worker is
1904	* preparing to process a work or actually processing it.
1905	* Make sure nobody diddled with it while I was sleeping.
1906	*/
1907	BUG_ON(!list_empty(&worker->scheduled));
1908
1909	/*
1910	* When control reaches this point, we're guaranteed to have
1911	* at least one idle worker or that someone else has already
1912	* assumed the manager role.
1913	*/
1914	worker_clr_flags(worker, WORKER_PREP);
1915
1916	do {
1917	struct work_struct *work =
1918	list_first_entry(&gcwq->worklist,
1919	struct work_struct, entry);
1920
1921	if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1922	/* optimization path, not strictly necessary */
1923	process_one_work(worker, work);
1924	if (unlikely(!list_empty(&worker->scheduled)))
1925	process_scheduled_works(worker);
1926	} else {
1927	move_linked_works(work, &worker->scheduled, NULL);
1928	process_scheduled_works(worker);
1929	}
1930	} while (keep_working(gcwq));
1931
1932	worker_set_flags(worker, WORKER_PREP, false);
1933	sleep:
1934	if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1935	goto recheck;
1936
1937	/*
1938	* gcwq->lock is held and there's no work to process and no
1939	* need to manage, sleep. Workers are woken up only while
1940	* holding gcwq->lock or from local cpu, so setting the
1941	* current state before releasing gcwq->lock is enough to
1942	* prevent losing any event.
1943	*/
1944	worker_enter_idle(worker);
1945	__set_current_state(TASK_INTERRUPTIBLE);
1946	spin_unlock_irq(&gcwq->lock);
1947	schedule();
1948	goto woke_up;
1949	}
1950
1951	/**
1952	* rescuer_thread - the rescuer thread function
1953	* @__wq: the associated workqueue
1954	*
1955	* Workqueue rescuer thread function. There's one rescuer for each
1956	* workqueue which has WQ_RESCUER set.
1957	*
1958	* Regular work processing on a gcwq may block trying to create a new
1959	* worker which uses GFP_KERNEL allocation which has slight chance of
1960	* developing into deadlock if some works currently on the same queue
1961	* need to be processed to satisfy the GFP_KERNEL allocation. This is
1962	* the problem rescuer solves.
1963	*
1964	* When such condition is possible, the gcwq summons rescuers of all
1965	* workqueues which have works queued on the gcwq and let them process
1966	* those works so that forward progress can be guaranteed.
1967	*
1968	* This should happen rarely.
1969	*/
1970	static int rescuer_thread(void *__wq)
1971	{
1972	struct workqueue_struct *wq = __wq;
1973	struct worker *rescuer = wq->rescuer;
1974	struct list_head *scheduled = &rescuer->scheduled;
1975	bool is_unbound = wq->flags & WQ_UNBOUND;
1976	unsigned int cpu;
1977
1978	set_user_nice(current, RESCUER_NICE_LEVEL);
1979	repeat:
1980	set_current_state(TASK_INTERRUPTIBLE);
1981
1982	if (kthread_should_stop())
1983	return 0;
1984
1985	/*
1986	* See whether any cpu is asking for help. Unbounded
1987	* workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
1988	*/
1989	for_each_mayday_cpu(cpu, wq->mayday_mask) {
1990	unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
1991	struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
1992	struct global_cwq *gcwq = cwq->gcwq;
1993	struct work_struct work, n;
1994
1995	__set_current_state(TASK_RUNNING);
1996	mayday_clear_cpu(cpu, wq->mayday_mask);
1997
1998	/* migrate to the target cpu if possible */
1999	rescuer->gcwq = gcwq;
2000	worker_maybe_bind_and_lock(rescuer);
2001
2002	/*
2003	* Slurp in all works issued via this workqueue and
2004	* process'em.
2005	*/
2006	BUG_ON(!list_empty(&rescuer->scheduled));
2007	list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2008	if (get_work_cwq(work) == cwq)
2009	move_linked_works(work, scheduled, &n);
2010
2011	process_scheduled_works(rescuer);
2012	spin_unlock_irq(&gcwq->lock);
2013	}
2014
2015	schedule();
2016	goto repeat;
2017	}
2018
2019	struct wq_barrier {
2020	struct work_struct work;
2021	struct completion done;
2022	};
2023
2024	static void wq_barrier_func(struct work_struct *work)
2025	{
2026	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2027	complete(&barr->done);
2028	}
2029
2030	/**
2031	* insert_wq_barrier - insert a barrier work
2032	* @cwq: cwq to insert barrier into
2033	* @barr: wq_barrier to insert
2034	* @target: target work to attach @barr to
2035	* @worker: worker currently executing @target, NULL if @target is not executing
2036	*
2037	* @barr is linked to @target such that @barr is completed only after
2038	* @target finishes execution. Please note that the ordering
2039	* guarantee is observed only with respect to @target and on the local
2040	* cpu.
2041	*
2042	* Currently, a queued barrier can't be canceled. This is because
2043	* try_to_grab_pending() can't determine whether the work to be
2044	* grabbed is at the head of the queue and thus can't clear LINKED
2045	* flag of the previous work while there must be a valid next work
2046	* after a work with LINKED flag set.
2047	*
2048	* Note that when @worker is non-NULL, @target may be modified
2049	* underneath us, so we can't reliably determine cwq from @target.
2050	*
2051	* CONTEXT:
2052	* spin_lock_irq(gcwq->lock).
2053	*/
2054	static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2055	struct wq_barrier *barr,
2056	struct work_struct target, struct worker worker)
2057	{
2058	struct list_head *head;
2059	unsigned int linked = 0;
2060
2061	/*
2062	* debugobject calls are safe here even with gcwq->lock locked
2063	* as we know for sure that this will not trigger any of the
2064	* checks and call back into the fixup functions where we
2065	* might deadlock.
2066	*/
2067	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2068	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2069	init_completion(&barr->done);
2070
2071	/*
2072	* If @target is currently being executed, schedule the
2073	* barrier to the worker; otherwise, put it after @target.
2074	*/
2075	if (worker)
2076	head = worker->scheduled.next;
2077	else {
2078	unsigned long *bits = work_data_bits(target);
2079
2080	head = target->entry.next;
2081	/* there can already be other linked works, inherit and set */
2082	linked = *bits & WORK_STRUCT_LINKED;
2083	__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2084	}
2085
2086	debug_work_activate(&barr->work);
2087	insert_work(cwq, &barr->work, head,
2088	work_color_to_flags(WORK_NO_COLOR) \| linked);
2089	}
2090
2091	/**
2092	* flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2093	* @wq: workqueue being flushed
2094	* @flush_color: new flush color, < 0 for no-op
2095	* @work_color: new work color, < 0 for no-op
2096	*
2097	* Prepare cwqs for workqueue flushing.
2098	*
2099	* If @flush_color is non-negative, flush_color on all cwqs should be
2100	* -1. If no cwq has in-flight commands at the specified color, all
2101	* cwq->flush_color's stay at -1 and %false is returned. If any cwq
2102	* has in flight commands, its cwq->flush_color is set to
2103	* @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2104	* wakeup logic is armed and %true is returned.
2105	*
2106	* The caller should have initialized @wq->first_flusher prior to
2107	* calling this function with non-negative @flush_color. If
2108	* @flush_color is negative, no flush color update is done and %false
2109	* is returned.
2110	*
2111	* If @work_color is non-negative, all cwqs should have the same
2112	* work_color which is previous to @work_color and all will be
2113	* advanced to @work_color.
2114	*
2115	* CONTEXT:
2116	* mutex_lock(wq->flush_mutex).
2117	*
2118	* RETURNS:
2119	* %true if @flush_color >= 0 and there's something to flush. %false
2120	* otherwise.
2121	*/
2122	static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2123	int flush_color, int work_color)
2124	{
2125	bool wait = false;
2126	unsigned int cpu;
2127
2128	if (flush_color >= 0) {
2129	BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2130	atomic_set(&wq->nr_cwqs_to_flush, 1);
2131	}
2132
2133	for_each_cwq_cpu(cpu, wq) {
2134	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2135	struct global_cwq *gcwq = cwq->gcwq;
2136
2137	spin_lock_irq(&gcwq->lock);
2138
2139	if (flush_color >= 0) {
2140	BUG_ON(cwq->flush_color != -1);
2141
2142	if (cwq->nr_in_flight[flush_color]) {
2143	cwq->flush_color = flush_color;
2144	atomic_inc(&wq->nr_cwqs_to_flush);
2145	wait = true;
2146	}
2147	}
2148
2149	if (work_color >= 0) {
2150	BUG_ON(work_color != work_next_color(cwq->work_color));
2151	cwq->work_color = work_color;
2152	}
2153
2154	spin_unlock_irq(&gcwq->lock);
2155	}
2156
2157	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2158	complete(&wq->first_flusher->done);
2159
2160	return wait;
2161	}
2162
2163	/**
2164	* flush_workqueue - ensure that any scheduled work has run to completion.
2165	* @wq: workqueue to flush
2166	*
2167	* Forces execution of the workqueue and blocks until its completion.
2168	* This is typically used in driver shutdown handlers.
2169	*
2170	* We sleep until all works which were queued on entry have been handled,
2171	* but we are not livelocked by new incoming ones.
2172	*/
2173	void flush_workqueue(struct workqueue_struct *wq)
2174	{
2175	struct wq_flusher this_flusher = {
2176	.list = LIST_HEAD_INIT(this_flusher.list),
2177	.flush_color = -1,
2178	.done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2179	};
2180	int next_color;
2181
2182	lock_map_acquire(&wq->lockdep_map);
2183	lock_map_release(&wq->lockdep_map);
2184
2185	mutex_lock(&wq->flush_mutex);
2186
2187	/*
2188	* Start-to-wait phase
2189	*/
2190	next_color = work_next_color(wq->work_color);
2191
2192	if (next_color != wq->flush_color) {
2193	/*
2194	* Color space is not full. The current work_color
2195	* becomes our flush_color and work_color is advanced
2196	* by one.
2197	*/
2198	BUG_ON(!list_empty(&wq->flusher_overflow));
2199	this_flusher.flush_color = wq->work_color;
2200	wq->work_color = next_color;
2201
2202	if (!wq->first_flusher) {
2203	/* no flush in progress, become the first flusher */
2204	BUG_ON(wq->flush_color != this_flusher.flush_color);
2205
2206	wq->first_flusher = &this_flusher;
2207
2208	if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2209	wq->work_color)) {
2210	/* nothing to flush, done */
2211	wq->flush_color = next_color;
2212	wq->first_flusher = NULL;
2213	goto out_unlock;
2214	}
2215	} else {
2216	/* wait in queue */
2217	BUG_ON(wq->flush_color == this_flusher.flush_color);
2218	list_add_tail(&this_flusher.list, &wq->flusher_queue);
2219	flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2220	}
2221	} else {
2222	/*
2223	* Oops, color space is full, wait on overflow queue.
2224	* The next flush completion will assign us
2225	* flush_color and transfer to flusher_queue.
2226	*/
2227	list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2228	}
2229
2230	mutex_unlock(&wq->flush_mutex);
2231
2232	wait_for_completion(&this_flusher.done);
2233
2234	/*
2235	* Wake-up-and-cascade phase
2236	*
2237	* First flushers are responsible for cascading flushes and
2238	* handling overflow. Non-first flushers can simply return.
2239	*/
2240	if (wq->first_flusher != &this_flusher)
2241	return;
2242
2243	mutex_lock(&wq->flush_mutex);
2244
2245	/* we might have raced, check again with mutex held */
2246	if (wq->first_flusher != &this_flusher)
2247	goto out_unlock;
2248
2249	wq->first_flusher = NULL;
2250
2251	BUG_ON(!list_empty(&this_flusher.list));
2252	BUG_ON(wq->flush_color != this_flusher.flush_color);
2253
2254	while (true) {
2255	struct wq_flusher next, tmp;
2256
2257	/* complete all the flushers sharing the current flush color */
2258	list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2259	if (next->flush_color != wq->flush_color)
2260	break;
2261	list_del_init(&next->list);
2262	complete(&next->done);
2263	}
2264
2265	BUG_ON(!list_empty(&wq->flusher_overflow) &&
2266	wq->flush_color != work_next_color(wq->work_color));
2267
2268	/* this flush_color is finished, advance by one */
2269	wq->flush_color = work_next_color(wq->flush_color);
2270
2271	/* one color has been freed, handle overflow queue */
2272	if (!list_empty(&wq->flusher_overflow)) {
2273	/*
2274	* Assign the same color to all overflowed
2275	* flushers, advance work_color and append to
2276	* flusher_queue. This is the start-to-wait
2277	* phase for these overflowed flushers.
2278	*/
2279	list_for_each_entry(tmp, &wq->flusher_overflow, list)
2280	tmp->flush_color = wq->work_color;
2281
2282	wq->work_color = work_next_color(wq->work_color);
2283
2284	list_splice_tail_init(&wq->flusher_overflow,
2285	&wq->flusher_queue);
2286	flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2287	}
2288
2289	if (list_empty(&wq->flusher_queue)) {
2290	BUG_ON(wq->flush_color != wq->work_color);
2291	break;
2292	}
2293
2294	/*
2295	* Need to flush more colors. Make the next flusher
2296	* the new first flusher and arm cwqs.
2297	*/
2298	BUG_ON(wq->flush_color == wq->work_color);
2299	BUG_ON(wq->flush_color != next->flush_color);
2300
2301	list_del_init(&next->list);
2302	wq->first_flusher = next;
2303
2304	if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2305	break;
2306
2307	/*
2308	* Meh... this color is already done, clear first
2309	* flusher and repeat cascading.
2310	*/
2311	wq->first_flusher = NULL;
2312	}
2313
2314	out_unlock:
2315	mutex_unlock(&wq->flush_mutex);
2316	}
2317	EXPORT_SYMBOL_GPL(flush_workqueue);
2318
2319	static bool start_flush_work(struct work_struct work, struct wq_barrier barr,
2320	bool wait_executing)
2321	{
2322	struct worker *worker = NULL;
2323	struct global_cwq *gcwq;
2324	struct cpu_workqueue_struct *cwq;
2325
2326	might_sleep();
2327	gcwq = get_work_gcwq(work);
2328	if (!gcwq)
2329	return false;
2330
2331	spin_lock_irq(&gcwq->lock);
2332	if (!list_empty(&work->entry)) {
2333	/*
2334	* See the comment near try_to_grab_pending()->smp_rmb().
2335	* If it was re-queued to a different gcwq under us, we
2336	* are not going to wait.
2337	*/
2338	smp_rmb();
2339	cwq = get_work_cwq(work);
2340	if (unlikely(!cwq \|\| gcwq != cwq->gcwq))
2341	goto already_gone;
2342	} else if (wait_executing) {
2343	worker = find_worker_executing_work(gcwq, work);
2344	if (!worker)
2345	goto already_gone;
2346	cwq = worker->current_cwq;
2347	} else
2348	goto already_gone;
2349
2350	insert_wq_barrier(cwq, barr, work, worker);
2351	spin_unlock_irq(&gcwq->lock);
2352
2353	lock_map_acquire(&cwq->wq->lockdep_map);
2354	lock_map_release(&cwq->wq->lockdep_map);
2355	return true;
2356	already_gone:
2357	spin_unlock_irq(&gcwq->lock);
2358	return false;
2359	}
2360
2361	/**
2362	* flush_work - wait for a work to finish executing the last queueing instance
2363	* @work: the work to flush
2364	*
2365	* Wait until @work has finished execution. This function considers
2366	* only the last queueing instance of @work. If @work has been
2367	* enqueued across different CPUs on a non-reentrant workqueue or on
2368	* multiple workqueues, @work might still be executing on return on
2369	* some of the CPUs from earlier queueing.
2370	*
2371	* If @work was queued only on a non-reentrant, ordered or unbound
2372	* workqueue, @work is guaranteed to be idle on return if it hasn't
2373	* been requeued since flush started.
2374	*
2375	* RETURNS:
2376	* %true if flush_work() waited for the work to finish execution,
2377	* %false if it was already idle.
2378	*/
2379	bool flush_work(struct work_struct *work)
2380	{
2381	struct wq_barrier barr;
2382
2383	if (start_flush_work(work, &barr, true)) {
2384	wait_for_completion(&barr.done);
2385	destroy_work_on_stack(&barr.work);
2386	return true;
2387	} else
2388	return false;
2389	}
2390	EXPORT_SYMBOL_GPL(flush_work);
2391
2392	static bool wait_on_cpu_work(struct global_cwq gcwq, struct work_struct work)
2393	{
2394	struct wq_barrier barr;
2395	struct worker *worker;
2396
2397	spin_lock_irq(&gcwq->lock);
2398
2399	worker = find_worker_executing_work(gcwq, work);
2400	if (unlikely(worker))
2401	insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2402
2403	spin_unlock_irq(&gcwq->lock);
2404
2405	if (unlikely(worker)) {
2406	wait_for_completion(&barr.done);
2407	destroy_work_on_stack(&barr.work);
2408	return true;
2409	} else
2410	return false;
2411	}
2412
2413	static bool wait_on_work(struct work_struct *work)
2414	{
2415	bool ret = false;
2416	int cpu;
2417
2418	might_sleep();
2419
2420	lock_map_acquire(&work->lockdep_map);
2421	lock_map_release(&work->lockdep_map);
2422
2423	for_each_gcwq_cpu(cpu)
2424	ret \|= wait_on_cpu_work(get_gcwq(cpu), work);
2425	return ret;
2426	}
2427
2428	/**
2429	* flush_work_sync - wait until a work has finished execution
2430	* @work: the work to flush
2431	*
2432	* Wait until @work has finished execution. On return, it's
2433	* guaranteed that all queueing instances of @work which happened
2434	* before this function is called are finished. In other words, if
2435	* @work hasn't been requeued since this function was called, @work is
2436	* guaranteed to be idle on return.
2437	*
2438	* RETURNS:
2439	* %true if flush_work_sync() waited for the work to finish execution,
2440	* %false if it was already idle.
2441	*/
2442	bool flush_work_sync(struct work_struct *work)
2443	{
2444	struct wq_barrier barr;
2445	bool pending, waited;
2446
2447	/* we'll wait for executions separately, queue barr only if pending */
2448	pending = start_flush_work(work, &barr, false);
2449
2450	/* wait for executions to finish */
2451	waited = wait_on_work(work);
2452
2453	/* wait for the pending one */
2454	if (pending) {
2455	wait_for_completion(&barr.done);
2456	destroy_work_on_stack(&barr.work);
2457	}
2458
2459	return pending \|\| waited;
2460	}
2461	EXPORT_SYMBOL_GPL(flush_work_sync);
2462
2463	/*
2464	* Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2465	* so this work can't be re-armed in any way.
2466	*/
2467	static int try_to_grab_pending(struct work_struct *work)
2468	{
2469	struct global_cwq *gcwq;
2470	int ret = -1;
2471
2472	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2473	return 0;
2474
2475	/*
2476	* The queueing is in progress, or it is already queued. Try to
2477	* steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2478	*/
2479	gcwq = get_work_gcwq(work);
2480	if (!gcwq)
2481	return ret;
2482
2483	spin_lock_irq(&gcwq->lock);
2484	if (!list_empty(&work->entry)) {
2485	/*
2486	* This work is queued, but perhaps we locked the wrong gcwq.
2487	* In that case we must see the new value after rmb(), see
2488	* insert_work()->wmb().
2489	*/
2490	smp_rmb();
2491	if (gcwq == get_work_gcwq(work)) {
2492	debug_work_deactivate(work);
2493	list_del_init(&work->entry);
2494	cwq_dec_nr_in_flight(get_work_cwq(work),
2495	get_work_color(work),
2496	*work_data_bits(work) & WORK_STRUCT_DELAYED);
2497	ret = 1;
2498	}
2499	}
2500	spin_unlock_irq(&gcwq->lock);
2501
2502	return ret;
2503	}
2504
2505	static bool __cancel_work_timer(struct work_struct *work,
2506	struct timer_list* timer)
2507	{
2508	int ret;
2509
2510	do {
2511	ret = (timer && likely(del_timer(timer)));
2512	if (!ret)
2513	ret = try_to_grab_pending(work);
2514	wait_on_work(work);
2515	} while (unlikely(ret < 0));
2516
2517	clear_work_data(work);
2518	return ret;
2519	}
2520
2521	/**
2522	* cancel_work_sync - cancel a work and wait for it to finish
2523	* @work: the work to cancel
2524	*
2525	* Cancel @work and wait for its execution to finish. This function
2526	* can be used even if the work re-queues itself or migrates to
2527	* another workqueue. On return from this function, @work is
2528	* guaranteed to be not pending or executing on any CPU.
2529	*
2530	* cancel_work_sync(&delayed_work->work) must not be used for
2531	* delayed_work's. Use cancel_delayed_work_sync() instead.
2532	*
2533	* The caller must ensure that the workqueue on which @work was last
2534	* queued can't be destroyed before this function returns.
2535	*
2536	* RETURNS:
2537	* %true if @work was pending, %false otherwise.
2538	*/
2539	bool cancel_work_sync(struct work_struct *work)
2540	{
2541	return __cancel_work_timer(work, NULL);
2542	}
2543	EXPORT_SYMBOL_GPL(cancel_work_sync);
2544
2545	/**
2546	* flush_delayed_work - wait for a dwork to finish executing the last queueing
2547	* @dwork: the delayed work to flush
2548	*
2549	* Delayed timer is cancelled and the pending work is queued for
2550	* immediate execution. Like flush_work(), this function only
2551	* considers the last queueing instance of @dwork.
2552	*
2553	* RETURNS:
2554	* %true if flush_work() waited for the work to finish execution,
2555	* %false if it was already idle.
2556	*/
2557	bool flush_delayed_work(struct delayed_work *dwork)
2558	{
2559	if (del_timer_sync(&dwork->timer))
2560	__queue_work(raw_smp_processor_id(),
2561	get_work_cwq(&dwork->work)->wq, &dwork->work);
2562	return flush_work(&dwork->work);
2563	}
2564	EXPORT_SYMBOL(flush_delayed_work);
2565
2566	/**
2567	* flush_delayed_work_sync - wait for a dwork to finish
2568	* @dwork: the delayed work to flush
2569	*
2570	* Delayed timer is cancelled and the pending work is queued for
2571	* execution immediately. Other than timer handling, its behavior
2572	* is identical to flush_work_sync().
2573	*
2574	* RETURNS:
2575	* %true if flush_work_sync() waited for the work to finish execution,
2576	* %false if it was already idle.
2577	*/
2578	bool flush_delayed_work_sync(struct delayed_work *dwork)
2579	{
2580	if (del_timer_sync(&dwork->timer))
2581	__queue_work(raw_smp_processor_id(),
2582	get_work_cwq(&dwork->work)->wq, &dwork->work);
2583	return flush_work_sync(&dwork->work);
2584	}
2585	EXPORT_SYMBOL(flush_delayed_work_sync);
2586
2587	/**
2588	* cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2589	* @dwork: the delayed work cancel
2590	*
2591	* This is cancel_work_sync() for delayed works.
2592	*
2593	* RETURNS:
2594	* %true if @dwork was pending, %false otherwise.
2595	*/
2596	bool cancel_delayed_work_sync(struct delayed_work *dwork)
2597	{
2598	return __cancel_work_timer(&dwork->work, &dwork->timer);
2599	}
2600	EXPORT_SYMBOL(cancel_delayed_work_sync);
2601
2602	/**
2603	* schedule_work - put work task in global workqueue
2604	* @work: job to be done
2605	*
2606	* Returns zero if @work was already on the kernel-global workqueue and
2607	* non-zero otherwise.
2608	*
2609	* This puts a job in the kernel-global workqueue if it was not already
2610	* queued and leaves it in the same position on the kernel-global
2611	* workqueue otherwise.
2612	*/
2613	int schedule_work(struct work_struct *work)
2614	{
2615	return queue_work(system_wq, work);
2616	}
2617	EXPORT_SYMBOL(schedule_work);
2618
2619	/*
2620	* schedule_work_on - put work task on a specific cpu
2621	* @cpu: cpu to put the work task on
2622	* @work: job to be done
2623	*
2624	* This puts a job on a specific cpu
2625	*/
2626	int schedule_work_on(int cpu, struct work_struct *work)
2627	{
2628	return queue_work_on(cpu, system_wq, work);
2629	}
2630	EXPORT_SYMBOL(schedule_work_on);
2631
2632	/**
2633	* schedule_delayed_work - put work task in global workqueue after delay
2634	* @dwork: job to be done
2635	* @delay: number of jiffies to wait or 0 for immediate execution
2636	*
2637	* After waiting for a given time this puts a job in the kernel-global
2638	* workqueue.
2639	*/
2640	int schedule_delayed_work(struct delayed_work *dwork,
2641	unsigned long delay)
2642	{
2643	return queue_delayed_work(system_wq, dwork, delay);
2644	}
2645	EXPORT_SYMBOL(schedule_delayed_work);
2646
2647	/**
2648	* schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2649	* @cpu: cpu to use
2650	* @dwork: job to be done
2651	* @delay: number of jiffies to wait
2652	*
2653	* After waiting for a given time this puts a job in the kernel-global
2654	* workqueue on the specified CPU.
2655	*/
2656	int schedule_delayed_work_on(int cpu,
2657	struct delayed_work *dwork, unsigned long delay)
2658	{
2659	return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2660	}
2661	EXPORT_SYMBOL(schedule_delayed_work_on);
2662
2663	/**
2664	* schedule_on_each_cpu - execute a function synchronously on each online CPU
2665	* @func: the function to call
2666	*
2667	* schedule_on_each_cpu() executes @func on each online CPU using the
2668	* system workqueue and blocks until all CPUs have completed.
2669	* schedule_on_each_cpu() is very slow.
2670	*
2671	* RETURNS:
2672	* 0 on success, -errno on failure.
2673	*/
2674	int schedule_on_each_cpu(work_func_t func)
2675	{
2676	int cpu;
2677	struct work_struct __percpu *works;
2678
2679	works = alloc_percpu(struct work_struct);
2680	if (!works)
2681	return -ENOMEM;
2682
2683	get_online_cpus();
2684
2685	for_each_online_cpu(cpu) {
2686	struct work_struct *work = per_cpu_ptr(works, cpu);
2687
2688	INIT_WORK(work, func);
2689	schedule_work_on(cpu, work);
2690	}
2691
2692	for_each_online_cpu(cpu)
2693	flush_work(per_cpu_ptr(works, cpu));
2694
2695	put_online_cpus();
2696	free_percpu(works);
2697	return 0;
2698	}
2699
2700	/**
2701	* flush_scheduled_work - ensure that any scheduled work has run to completion.
2702	*
2703	* Forces execution of the kernel-global workqueue and blocks until its
2704	* completion.
2705	*
2706	* Think twice before calling this function! It's very easy to get into
2707	* trouble if you don't take great care. Either of the following situations
2708	* will lead to deadlock:
2709	*
2710	* One of the work items currently on the workqueue needs to acquire
2711	* a lock held by your code or its caller.
2712	*
2713	* Your code is running in the context of a work routine.
2714	*
2715	* They will be detected by lockdep when they occur, but the first might not
2716	* occur very often. It depends on what work items are on the workqueue and
2717	* what locks they need, which you have no control over.
2718	*
2719	* In most situations flushing the entire workqueue is overkill; you merely
2720	* need to know that a particular work item isn't queued and isn't running.
2721	* In such cases you should use cancel_delayed_work_sync() or
2722	* cancel_work_sync() instead.
2723	*/
2724	void flush_scheduled_work(void)
2725	{
2726	flush_workqueue(system_wq);
2727	}
2728	EXPORT_SYMBOL(flush_scheduled_work);
2729
2730	/**
2731	* execute_in_process_context - reliably execute the routine with user context
2732	* @fn: the function to execute
2733	* @ew: guaranteed storage for the execute work structure (must
2734	* be available when the work executes)
2735	*
2736	* Executes the function immediately if process context is available,
2737	* otherwise schedules the function for delayed execution.
2738	*
2739	* Returns: 0 - function was executed
2740	* 1 - function was scheduled for execution
2741	*/
2742	int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2743	{
2744	if (!in_interrupt()) {
2745	fn(&ew->work);
2746	return 0;
2747	}
2748
2749	INIT_WORK(&ew->work, fn);
2750	schedule_work(&ew->work);
2751
2752	return 1;
2753	}
2754	EXPORT_SYMBOL_GPL(execute_in_process_context);
2755
2756	int keventd_up(void)
2757	{
2758	return system_wq != NULL;
2759	}
2760
2761	static int alloc_cwqs(struct workqueue_struct *wq)
2762	{
2763	/*
2764	* cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2765	* Make sure that the alignment isn't lower than that of
2766	* unsigned long long.
2767	*/
2768	const size_t size = sizeof(struct cpu_workqueue_struct);
2769	const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2770	__alignof__(unsigned long long));
2771	#ifdef CONFIG_SMP
2772	bool percpu = !(wq->flags & WQ_UNBOUND);
2773	#else
2774	bool percpu = false;
2775	#endif
2776
2777	if (percpu)
2778	wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2779	else {
2780	void *ptr;
2781
2782	/*
2783	* Allocate enough room to align cwq and put an extra
2784	* pointer at the end pointing back to the originally
2785	* allocated pointer which will be used for free.
2786	*/
2787	ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2788	if (ptr) {
2789	wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2790	(void *)(wq->cpu_wq.single + 1) = ptr;
2791	}
2792	}
2793
2794	/* just in case, make sure it's actually aligned
2795	* - this is affected by PERCPU() alignment in vmlinux.lds.S
2796	*/
2797	BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2798	return wq->cpu_wq.v ? 0 : -ENOMEM;
2799	}
2800
2801	static void free_cwqs(struct workqueue_struct *wq)
2802	{
2803	#ifdef CONFIG_SMP
2804	bool percpu = !(wq->flags & WQ_UNBOUND);
2805	#else
2806	bool percpu = false;
2807	#endif
2808
2809	if (percpu)
2810	free_percpu(wq->cpu_wq.pcpu);
2811	else if (wq->cpu_wq.single) {
2812	/* the pointer to free is stored right after the cwq */
2813	kfree((void *)(wq->cpu_wq.single + 1));
2814	}
2815	}
2816
2817	static int wq_clamp_max_active(int max_active, unsigned int flags,
2818	const char *name)
2819	{
2820	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2821
2822	if (max_active < 1 \|\| max_active > lim)
2823	printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2824	"is out of range, clamping between %d and %d\n",
2825	max_active, name, 1, lim);
2826
2827	return clamp_val(max_active, 1, lim);
2828	}
2829
2830	struct workqueue_struct __alloc_workqueue_key(const char name,
2831	unsigned int flags,
2832	int max_active,
2833	struct lock_class_key *key,
2834	const char *lock_name)
2835	{
2836	struct workqueue_struct *wq;
2837	unsigned int cpu;
2838
2839	/*
2840	* Workqueues which may be used during memory reclaim should
2841	* have a rescuer to guarantee forward progress.
2842	*/
2843	if (flags & WQ_MEM_RECLAIM)
2844	flags \|= WQ_RESCUER;
2845
2846	/*
2847	* Unbound workqueues aren't concurrency managed and should be
2848	* dispatched to workers immediately.
2849	*/
2850	if (flags & WQ_UNBOUND)
2851	flags \|= WQ_HIGHPRI;
2852
2853	max_active = max_active ?: WQ_DFL_ACTIVE;
2854	max_active = wq_clamp_max_active(max_active, flags, name);
2855
2856	wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2857	if (!wq)
2858	goto err;
2859
2860	wq->flags = flags;
2861	wq->saved_max_active = max_active;
2862	mutex_init(&wq->flush_mutex);
2863	atomic_set(&wq->nr_cwqs_to_flush, 0);
2864	INIT_LIST_HEAD(&wq->flusher_queue);
2865	INIT_LIST_HEAD(&wq->flusher_overflow);
2866
2867	wq->name = name;
2868	lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2869	INIT_LIST_HEAD(&wq->list);
2870
2871	if (alloc_cwqs(wq) < 0)
2872	goto err;
2873
2874	for_each_cwq_cpu(cpu, wq) {
2875	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2876	struct global_cwq *gcwq = get_gcwq(cpu);
2877
2878	BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
2879	cwq->gcwq = gcwq;
2880	cwq->wq = wq;
2881	cwq->flush_color = -1;
2882	cwq->max_active = max_active;
2883	INIT_LIST_HEAD(&cwq->delayed_works);
2884	}
2885
2886	if (flags & WQ_RESCUER) {
2887	struct worker *rescuer;
2888
2889	if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
2890	goto err;
2891
2892	wq->rescuer = rescuer = alloc_worker();
2893	if (!rescuer)
2894	goto err;
2895
2896	rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
2897	if (IS_ERR(rescuer->task))
2898	goto err;
2899
2900	rescuer->task->flags \|= PF_THREAD_BOUND;
2901	wake_up_process(rescuer->task);
2902	}
2903
2904	/*
2905	* workqueue_lock protects global freeze state and workqueues
2906	* list. Grab it, set max_active accordingly and add the new
2907	* workqueue to workqueues list.
2908	*/
2909	spin_lock(&workqueue_lock);
2910
2911	if (workqueue_freezing && wq->flags & WQ_FREEZEABLE)
2912	for_each_cwq_cpu(cpu, wq)
2913	get_cwq(cpu, wq)->max_active = 0;
2914
2915	list_add(&wq->list, &workqueues);
2916
2917	spin_unlock(&workqueue_lock);
2918
2919	return wq;
2920	err:
2921	if (wq) {
2922	free_cwqs(wq);
2923	free_mayday_mask(wq->mayday_mask);
2924	kfree(wq->rescuer);
2925	kfree(wq);
2926	}
2927	return NULL;
2928	}
2929	EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
2930
2931	/**
2932	* destroy_workqueue - safely terminate a workqueue
2933	* @wq: target workqueue
2934	*
2935	* Safely destroy a workqueue. All work currently pending will be done first.
2936	*/
2937	void destroy_workqueue(struct workqueue_struct *wq)
2938	{
2939	unsigned int cpu;
2940
2941	wq->flags \|= WQ_DYING;
2942	flush_workqueue(wq);
2943
2944	/*
2945	* wq list is used to freeze wq, remove from list after
2946	* flushing is complete in case freeze races us.
2947	*/
2948	spin_lock(&workqueue_lock);
2949	list_del(&wq->list);
2950	spin_unlock(&workqueue_lock);
2951
2952	/* sanity check */
2953	for_each_cwq_cpu(cpu, wq) {
2954	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2955	int i;
2956
2957	for (i = 0; i < WORK_NR_COLORS; i++)
2958	BUG_ON(cwq->nr_in_flight[i]);
2959	BUG_ON(cwq->nr_active);
2960	BUG_ON(!list_empty(&cwq->delayed_works));
2961	}
2962
2963	if (wq->flags & WQ_RESCUER) {
2964	kthread_stop(wq->rescuer->task);
2965	free_mayday_mask(wq->mayday_mask);
2966	kfree(wq->rescuer);
2967	}
2968
2969	free_cwqs(wq);
2970	kfree(wq);
2971	}
2972	EXPORT_SYMBOL_GPL(destroy_workqueue);
2973
2974	/**
2975	* workqueue_set_max_active - adjust max_active of a workqueue
2976	* @wq: target workqueue
2977	* @max_active: new max_active value.
2978	*
2979	* Set max_active of @wq to @max_active.
2980	*
2981	* CONTEXT:
2982	* Don't call from IRQ context.
2983	*/
2984	void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
2985	{
2986	unsigned int cpu;
2987
2988	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
2989
2990	spin_lock(&workqueue_lock);
2991
2992	wq->saved_max_active = max_active;
2993
2994	for_each_cwq_cpu(cpu, wq) {
2995	struct global_cwq *gcwq = get_gcwq(cpu);
2996
2997	spin_lock_irq(&gcwq->lock);
2998
2999	if (!(wq->flags & WQ_FREEZEABLE) \|\|
3000	!(gcwq->flags & GCWQ_FREEZING))
3001	get_cwq(gcwq->cpu, wq)->max_active = max_active;
3002
3003	spin_unlock_irq(&gcwq->lock);
3004	}
3005
3006	spin_unlock(&workqueue_lock);
3007	}
3008	EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3009
3010	/**
3011	* workqueue_congested - test whether a workqueue is congested
3012	* @cpu: CPU in question
3013	* @wq: target workqueue
3014	*
3015	* Test whether @wq's cpu workqueue for @cpu is congested. There is
3016	* no synchronization around this function and the test result is
3017	* unreliable and only useful as advisory hints or for debugging.
3018	*
3019	* RETURNS:
3020	* %true if congested, %false otherwise.
3021	*/
3022	bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3023	{
3024	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3025
3026	return !list_empty(&cwq->delayed_works);
3027	}
3028	EXPORT_SYMBOL_GPL(workqueue_congested);
3029
3030	/**
3031	* work_cpu - return the last known associated cpu for @work
3032	* @work: the work of interest
3033	*
3034	* RETURNS:
3035	* CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3036	*/
3037	unsigned int work_cpu(struct work_struct *work)
3038	{
3039	struct global_cwq *gcwq = get_work_gcwq(work);
3040
3041	return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3042	}
3043	EXPORT_SYMBOL_GPL(work_cpu);
3044
3045	/**
3046	* work_busy - test whether a work is currently pending or running
3047	* @work: the work to be tested
3048	*
3049	* Test whether @work is currently pending or running. There is no
3050	* synchronization around this function and the test result is
3051	* unreliable and only useful as advisory hints or for debugging.
3052	* Especially for reentrant wqs, the pending state might hide the
3053	* running state.
3054	*
3055	* RETURNS:
3056	* OR'd bitmask of WORK_BUSY_* bits.
3057	*/
3058	unsigned int work_busy(struct work_struct *work)
3059	{
3060	struct global_cwq *gcwq = get_work_gcwq(work);
3061	unsigned long flags;
3062	unsigned int ret = 0;
3063
3064	if (!gcwq)
3065	return false;
3066
3067	spin_lock_irqsave(&gcwq->lock, flags);
3068
3069	if (work_pending(work))
3070	ret \|= WORK_BUSY_PENDING;
3071	if (find_worker_executing_work(gcwq, work))
3072	ret \|= WORK_BUSY_RUNNING;
3073
3074	spin_unlock_irqrestore(&gcwq->lock, flags);
3075
3076	return ret;
3077	}
3078	EXPORT_SYMBOL_GPL(work_busy);
3079
3080	/*
3081	* CPU hotplug.
3082	*
3083	* There are two challenges in supporting CPU hotplug. Firstly, there
3084	* are a lot of assumptions on strong associations among work, cwq and
3085	* gcwq which make migrating pending and scheduled works very
3086	* difficult to implement without impacting hot paths. Secondly,
3087	* gcwqs serve mix of short, long and very long running works making
3088	* blocked draining impractical.
3089	*
3090	* This is solved by allowing a gcwq to be detached from CPU, running
3091	* it with unbound (rogue) workers and allowing it to be reattached
3092	* later if the cpu comes back online. A separate thread is created
3093	* to govern a gcwq in such state and is called the trustee of the
3094	* gcwq.
3095	*
3096	* Trustee states and their descriptions.
3097	*
3098	* START Command state used on startup. On CPU_DOWN_PREPARE, a
3099	* new trustee is started with this state.
3100	*
3101	* IN_CHARGE Once started, trustee will enter this state after
3102	* assuming the manager role and making all existing
3103	* workers rogue. DOWN_PREPARE waits for trustee to
3104	* enter this state. After reaching IN_CHARGE, trustee
3105	* tries to execute the pending worklist until it's empty
3106	* and the state is set to BUTCHER, or the state is set
3107	* to RELEASE.
3108	*
3109	* BUTCHER Command state which is set by the cpu callback after
3110	* the cpu has went down. Once this state is set trustee
3111	* knows that there will be no new works on the worklist
3112	* and once the worklist is empty it can proceed to
3113	* killing idle workers.
3114	*
3115	* RELEASE Command state which is set by the cpu callback if the
3116	* cpu down has been canceled or it has come online
3117	* again. After recognizing this state, trustee stops
3118	* trying to drain or butcher and clears ROGUE, rebinds
3119	* all remaining workers back to the cpu and releases
3120	* manager role.
3121	*
3122	* DONE Trustee will enter this state after BUTCHER or RELEASE
3123	* is complete.
3124	*
3125	* trustee CPU draining
3126	* took over down complete
3127	* START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3128	* \| \| ^
3129	* \| CPU is back online v return workers \|
3130	* ----------------> RELEASE --------------
3131	*/
3132
3133	/**
3134	* trustee_wait_event_timeout - timed event wait for trustee
3135	* @cond: condition to wait for
3136	* @timeout: timeout in jiffies
3137	*
3138	* wait_event_timeout() for trustee to use. Handles locking and
3139	* checks for RELEASE request.
3140	*
3141	* CONTEXT:
3142	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
3143	* multiple times. To be used by trustee.
3144	*
3145	* RETURNS:
3146	* Positive indicating left time if @cond is satisfied, 0 if timed
3147	* out, -1 if canceled.
3148	*/
3149	#define trustee_wait_event_timeout(cond, timeout) ({ \
3150	long __ret = (timeout); \
3151	while (!((cond) \|\| (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3152	__ret) { \
3153	spin_unlock_irq(&gcwq->lock); \
3154	__wait_event_timeout(gcwq->trustee_wait, (cond) \|\| \
3155	(gcwq->trustee_state == TRUSTEE_RELEASE), \
3156	__ret); \
3157	spin_lock_irq(&gcwq->lock); \
3158	} \
3159	gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3160	})
3161
3162	/**
3163	* trustee_wait_event - event wait for trustee
3164	* @cond: condition to wait for
3165	*
3166	* wait_event() for trustee to use. Automatically handles locking and
3167	* checks for CANCEL request.
3168	*
3169	* CONTEXT:
3170	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
3171	* multiple times. To be used by trustee.
3172	*
3173	* RETURNS:
3174	* 0 if @cond is satisfied, -1 if canceled.
3175	*/
3176	#define trustee_wait_event(cond) ({ \
3177	long __ret1; \
3178	__ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3179	__ret1 < 0 ? -1 : 0; \
3180	})
3181
3182	static int __cpuinit trustee_thread(void *__gcwq)
3183	{
3184	struct global_cwq *gcwq = __gcwq;
3185	struct worker *worker;
3186	struct work_struct *work;
3187	struct hlist_node *pos;
3188	long rc;
3189	int i;
3190
3191	BUG_ON(gcwq->cpu != smp_processor_id());
3192
3193	spin_lock_irq(&gcwq->lock);
3194	/*
3195	* Claim the manager position and make all workers rogue.
3196	* Trustee must be bound to the target cpu and can't be
3197	* cancelled.
3198	*/
3199	BUG_ON(gcwq->cpu != smp_processor_id());
3200	rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3201	BUG_ON(rc < 0);
3202
3203	gcwq->flags \|= GCWQ_MANAGING_WORKERS;
3204
3205	list_for_each_entry(worker, &gcwq->idle_list, entry)
3206	worker->flags \|= WORKER_ROGUE;
3207
3208	for_each_busy_worker(worker, i, pos, gcwq)
3209	worker->flags \|= WORKER_ROGUE;
3210
3211	/*
3212	* Call schedule() so that we cross rq->lock and thus can
3213	* guarantee sched callbacks see the rogue flag. This is
3214	* necessary as scheduler callbacks may be invoked from other
3215	* cpus.
3216	*/
3217	spin_unlock_irq(&gcwq->lock);
3218	schedule();
3219	spin_lock_irq(&gcwq->lock);
3220
3221	/*
3222	* Sched callbacks are disabled now. Zap nr_running. After
3223	* this, nr_running stays zero and need_more_worker() and
3224	* keep_working() are always true as long as the worklist is
3225	* not empty.
3226	*/
3227	atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3228
3229	spin_unlock_irq(&gcwq->lock);
3230	del_timer_sync(&gcwq->idle_timer);
3231	spin_lock_irq(&gcwq->lock);
3232
3233	/*
3234	* We're now in charge. Notify and proceed to drain. We need
3235	* to keep the gcwq running during the whole CPU down
3236	* procedure as other cpu hotunplug callbacks may need to
3237	* flush currently running tasks.
3238	*/
3239	gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3240	wake_up_all(&gcwq->trustee_wait);
3241
3242	/*
3243	* The original cpu is in the process of dying and may go away
3244	* anytime now. When that happens, we and all workers would
3245	* be migrated to other cpus. Try draining any left work. We
3246	* want to get it over with ASAP - spam rescuers, wake up as
3247	* many idlers as necessary and create new ones till the
3248	* worklist is empty. Note that if the gcwq is frozen, there
3249	* may be frozen works in freezeable cwqs. Don't declare
3250	* completion while frozen.
3251	*/
3252	while (gcwq->nr_workers != gcwq->nr_idle \|\|
3253	gcwq->flags & GCWQ_FREEZING \|\|
3254	gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3255	int nr_works = 0;
3256
3257	list_for_each_entry(work, &gcwq->worklist, entry) {
3258	send_mayday(work);
3259	nr_works++;
3260	}
3261
3262	list_for_each_entry(worker, &gcwq->idle_list, entry) {
3263	if (!nr_works--)
3264	break;
3265	wake_up_process(worker->task);
3266	}
3267
3268	if (need_to_create_worker(gcwq)) {
3269	spin_unlock_irq(&gcwq->lock);
3270	worker = create_worker(gcwq, false);
3271	spin_lock_irq(&gcwq->lock);
3272	if (worker) {
3273	worker->flags \|= WORKER_ROGUE;
3274	start_worker(worker);
3275	}
3276	}
3277
3278	/* give a breather */
3279	if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3280	break;
3281	}
3282
3283	/*
3284	* Either all works have been scheduled and cpu is down, or
3285	* cpu down has already been canceled. Wait for and butcher
3286	* all workers till we're canceled.
3287	*/
3288	do {
3289	rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3290	while (!list_empty(&gcwq->idle_list))
3291	destroy_worker(list_first_entry(&gcwq->idle_list,
3292	struct worker, entry));
3293	} while (gcwq->nr_workers && rc >= 0);
3294
3295	/*
3296	* At this point, either draining has completed and no worker
3297	* is left, or cpu down has been canceled or the cpu is being
3298	* brought back up. There shouldn't be any idle one left.
3299	* Tell the remaining busy ones to rebind once it finishes the
3300	* currently scheduled works by scheduling the rebind_work.
3301	*/
3302	WARN_ON(!list_empty(&gcwq->idle_list));
3303
3304	for_each_busy_worker(worker, i, pos, gcwq) {
3305	struct work_struct *rebind_work = &worker->rebind_work;
3306
3307	/*
3308	* Rebind_work may race with future cpu hotplug
3309	* operations. Use a separate flag to mark that
3310	* rebinding is scheduled.
3311	*/
3312	worker->flags \|= WORKER_REBIND;
3313	worker->flags &= ~WORKER_ROGUE;
3314
3315	/* queue rebind_work, wq doesn't matter, use the default one */
3316	if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3317	work_data_bits(rebind_work)))
3318	continue;
3319
3320	debug_work_activate(rebind_work);
3321	insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3322	worker->scheduled.next,
3323	work_color_to_flags(WORK_NO_COLOR));
3324	}
3325
3326	/* relinquish manager role */
3327	gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3328
3329	/* notify completion */
3330	gcwq->trustee = NULL;
3331	gcwq->trustee_state = TRUSTEE_DONE;
3332	wake_up_all(&gcwq->trustee_wait);
3333	spin_unlock_irq(&gcwq->lock);
3334	return 0;
3335	}
3336
3337	/**
3338	* wait_trustee_state - wait for trustee to enter the specified state
3339	* @gcwq: gcwq the trustee of interest belongs to
3340	* @state: target state to wait for
3341	*
3342	* Wait for the trustee to reach @state. DONE is already matched.
3343	*
3344	* CONTEXT:
3345	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
3346	* multiple times. To be used by cpu_callback.
3347	*/
3348	static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3349	__releases(&gcwq->lock)
3350	__acquires(&gcwq->lock)
3351	{
3352	if (!(gcwq->trustee_state == state \|\|
3353	gcwq->trustee_state == TRUSTEE_DONE)) {
3354	spin_unlock_irq(&gcwq->lock);
3355	__wait_event(gcwq->trustee_wait,
3356	gcwq->trustee_state == state \|\|
3357	gcwq->trustee_state == TRUSTEE_DONE);
3358	spin_lock_irq(&gcwq->lock);
3359	}
3360	}
3361
3362	static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3363	unsigned long action,
3364	void *hcpu)
3365	{
3366	unsigned int cpu = (unsigned long)hcpu;
3367	struct global_cwq *gcwq = get_gcwq(cpu);
3368	struct task_struct *new_trustee = NULL;
3369	struct worker *uninitialized_var(new_worker);
3370	unsigned long flags;
3371
3372	action &= ~CPU_TASKS_FROZEN;
3373
3374	switch (action) {
3375	case CPU_DOWN_PREPARE:
3376	new_trustee = kthread_create(trustee_thread, gcwq,
3377	"workqueue_trustee/%d\n", cpu);
3378	if (IS_ERR(new_trustee))
3379	return notifier_from_errno(PTR_ERR(new_trustee));
3380	kthread_bind(new_trustee, cpu);
3381	/* fall through */
3382	case CPU_UP_PREPARE:
3383	BUG_ON(gcwq->first_idle);
3384	new_worker = create_worker(gcwq, false);
3385	if (!new_worker) {
3386	if (new_trustee)
3387	kthread_stop(new_trustee);
3388	return NOTIFY_BAD;
3389	}
3390	}
3391
3392	/* some are called w/ irq disabled, don't disturb irq status */
3393	spin_lock_irqsave(&gcwq->lock, flags);
3394
3395	switch (action) {
3396	case CPU_DOWN_PREPARE:
3397	/* initialize trustee and tell it to acquire the gcwq */
3398	BUG_ON(gcwq->trustee \|\| gcwq->trustee_state != TRUSTEE_DONE);
3399	gcwq->trustee = new_trustee;
3400	gcwq->trustee_state = TRUSTEE_START;
3401	wake_up_process(gcwq->trustee);
3402	wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3403	/* fall through */
3404	case CPU_UP_PREPARE:
3405	BUG_ON(gcwq->first_idle);
3406	gcwq->first_idle = new_worker;
3407	break;
3408
3409	case CPU_DYING:
3410	/*
3411	* Before this, the trustee and all workers except for
3412	* the ones which are still executing works from
3413	* before the last CPU down must be on the cpu. After
3414	* this, they'll all be diasporas.
3415	*/
3416	gcwq->flags \|= GCWQ_DISASSOCIATED;
3417	break;
3418
3419	case CPU_POST_DEAD:
3420	gcwq->trustee_state = TRUSTEE_BUTCHER;
3421	/* fall through */
3422	case CPU_UP_CANCELED:
3423	destroy_worker(gcwq->first_idle);
3424	gcwq->first_idle = NULL;
3425	break;
3426
3427	case CPU_DOWN_FAILED:
3428	case CPU_ONLINE:
3429	gcwq->flags &= ~GCWQ_DISASSOCIATED;
3430	if (gcwq->trustee_state != TRUSTEE_DONE) {
3431	gcwq->trustee_state = TRUSTEE_RELEASE;
3432	wake_up_process(gcwq->trustee);
3433	wait_trustee_state(gcwq, TRUSTEE_DONE);
3434	}
3435
3436	/*
3437	* Trustee is done and there might be no worker left.
3438	* Put the first_idle in and request a real manager to
3439	* take a look.
3440	*/
3441	spin_unlock_irq(&gcwq->lock);
3442	kthread_bind(gcwq->first_idle->task, cpu);
3443	spin_lock_irq(&gcwq->lock);
3444	gcwq->flags \|= GCWQ_MANAGE_WORKERS;
3445	start_worker(gcwq->first_idle);
3446	gcwq->first_idle = NULL;
3447	break;
3448	}
3449
3450	spin_unlock_irqrestore(&gcwq->lock, flags);
3451
3452	return notifier_from_errno(0);
3453	}
3454
3455	#ifdef CONFIG_SMP
3456
3457	struct work_for_cpu {
3458	struct completion completion;
3459	long (fn)(void );
3460	void *arg;
3461	long ret;
3462	};
3463
3464	static int do_work_for_cpu(void *_wfc)
3465	{
3466	struct work_for_cpu *wfc = _wfc;
3467	wfc->ret = wfc->fn(wfc->arg);
3468	complete(&wfc->completion);
3469	return 0;
3470	}
3471
3472	/**
3473	* work_on_cpu - run a function in user context on a particular cpu
3474	* @cpu: the cpu to run on
3475	* @fn: the function to run
3476	* @arg: the function arg
3477	*
3478	* This will return the value @fn returns.
3479	* It is up to the caller to ensure that the cpu doesn't go offline.
3480	* The caller must not hold any locks which would prevent @fn from completing.
3481	*/
3482	long work_on_cpu(unsigned int cpu, long (fn)(void ), void *arg)
3483	{
3484	struct task_struct *sub_thread;
3485	struct work_for_cpu wfc = {
3486	.completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3487	.fn = fn,
3488	.arg = arg,
3489	};
3490
3491	sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3492	if (IS_ERR(sub_thread))
3493	return PTR_ERR(sub_thread);
3494	kthread_bind(sub_thread, cpu);
3495	wake_up_process(sub_thread);
3496	wait_for_completion(&wfc.completion);
3497	return wfc.ret;
3498	}
3499	EXPORT_SYMBOL_GPL(work_on_cpu);
3500	#endif /* CONFIG_SMP */
3501
3502	#ifdef CONFIG_FREEZER
3503
3504	/**
3505	* freeze_workqueues_begin - begin freezing workqueues
3506	*
3507	* Start freezing workqueues. After this function returns, all
3508	* freezeable workqueues will queue new works to their frozen_works
3509	* list instead of gcwq->worklist.
3510	*
3511	* CONTEXT:
3512	* Grabs and releases workqueue_lock and gcwq->lock's.
3513	*/
3514	void freeze_workqueues_begin(void)
3515	{
3516	unsigned int cpu;
3517
3518	spin_lock(&workqueue_lock);
3519
3520	BUG_ON(workqueue_freezing);
3521	workqueue_freezing = true;
3522
3523	for_each_gcwq_cpu(cpu) {
3524	struct global_cwq *gcwq = get_gcwq(cpu);
3525	struct workqueue_struct *wq;
3526
3527	spin_lock_irq(&gcwq->lock);
3528
3529	BUG_ON(gcwq->flags & GCWQ_FREEZING);
3530	gcwq->flags \|= GCWQ_FREEZING;
3531
3532	list_for_each_entry(wq, &workqueues, list) {
3533	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3534
3535	if (cwq && wq->flags & WQ_FREEZEABLE)
3536	cwq->max_active = 0;
3537	}
3538
3539	spin_unlock_irq(&gcwq->lock);
3540	}
3541
3542	spin_unlock(&workqueue_lock);
3543	}
3544
3545	/**
3546	* freeze_workqueues_busy - are freezeable workqueues still busy?
3547	*
3548	* Check whether freezing is complete. This function must be called
3549	* between freeze_workqueues_begin() and thaw_workqueues().
3550	*
3551	* CONTEXT:
3552	* Grabs and releases workqueue_lock.
3553	*
3554	* RETURNS:
3555	* %true if some freezeable workqueues are still busy. %false if
3556	* freezing is complete.
3557	*/
3558	bool freeze_workqueues_busy(void)
3559	{
3560	unsigned int cpu;
3561	bool busy = false;
3562
3563	spin_lock(&workqueue_lock);
3564
3565	BUG_ON(!workqueue_freezing);
3566
3567	for_each_gcwq_cpu(cpu) {
3568	struct workqueue_struct *wq;
3569	/*
3570	* nr_active is monotonically decreasing. It's safe
3571	* to peek without lock.
3572	*/
3573	list_for_each_entry(wq, &workqueues, list) {
3574	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3575
3576	if (!cwq \|\| !(wq->flags & WQ_FREEZEABLE))
3577	continue;
3578
3579	BUG_ON(cwq->nr_active < 0);
3580	if (cwq->nr_active) {
3581	busy = true;
3582	goto out_unlock;
3583	}
3584	}
3585	}
3586	out_unlock:
3587	spin_unlock(&workqueue_lock);
3588	return busy;
3589	}
3590
3591	/**
3592	* thaw_workqueues - thaw workqueues
3593	*
3594	* Thaw workqueues. Normal queueing is restored and all collected
3595	* frozen works are transferred to their respective gcwq worklists.
3596	*
3597	* CONTEXT:
3598	* Grabs and releases workqueue_lock and gcwq->lock's.
3599	*/
3600	void thaw_workqueues(void)
3601	{
3602	unsigned int cpu;
3603
3604	spin_lock(&workqueue_lock);
3605
3606	if (!workqueue_freezing)
3607	goto out_unlock;
3608
3609	for_each_gcwq_cpu(cpu) {
3610	struct global_cwq *gcwq = get_gcwq(cpu);
3611	struct workqueue_struct *wq;
3612
3613	spin_lock_irq(&gcwq->lock);
3614
3615	BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3616	gcwq->flags &= ~GCWQ_FREEZING;
3617
3618	list_for_each_entry(wq, &workqueues, list) {
3619	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3620
3621	if (!cwq \|\| !(wq->flags & WQ_FREEZEABLE))
3622	continue;
3623
3624	/* restore max_active and repopulate worklist */
3625	cwq->max_active = wq->saved_max_active;
3626
3627	while (!list_empty(&cwq->delayed_works) &&
3628	cwq->nr_active < cwq->max_active)
3629	cwq_activate_first_delayed(cwq);
3630	}
3631
3632	wake_up_worker(gcwq);
3633
3634	spin_unlock_irq(&gcwq->lock);
3635	}
3636
3637	workqueue_freezing = false;
3638	out_unlock:
3639	spin_unlock(&workqueue_lock);
3640	}
3641	#endif /* CONFIG_FREEZER */
3642
3643	static int __init init_workqueues(void)
3644	{
3645	unsigned int cpu;
3646	int i;
3647
3648	cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3649
3650	/* initialize gcwqs */
3651	for_each_gcwq_cpu(cpu) {
3652	struct global_cwq *gcwq = get_gcwq(cpu);
3653
3654	spin_lock_init(&gcwq->lock);
3655	INIT_LIST_HEAD(&gcwq->worklist);
3656	gcwq->cpu = cpu;
3657	gcwq->flags \|= GCWQ_DISASSOCIATED;
3658
3659	INIT_LIST_HEAD(&gcwq->idle_list);
3660	for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3661	INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3662
3663	init_timer_deferrable(&gcwq->idle_timer);
3664	gcwq->idle_timer.function = idle_worker_timeout;
3665	gcwq->idle_timer.data = (unsigned long)gcwq;
3666
3667	setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3668	(unsigned long)gcwq);
3669
3670	ida_init(&gcwq->worker_ida);
3671
3672	gcwq->trustee_state = TRUSTEE_DONE;
3673	init_waitqueue_head(&gcwq->trustee_wait);
3674	}
3675
3676	/* create the initial worker */
3677	for_each_online_gcwq_cpu(cpu) {
3678	struct global_cwq *gcwq = get_gcwq(cpu);
3679	struct worker *worker;
3680
3681	if (cpu != WORK_CPU_UNBOUND)
3682	gcwq->flags &= ~GCWQ_DISASSOCIATED;
3683	worker = create_worker(gcwq, true);
3684	BUG_ON(!worker);
3685	spin_lock_irq(&gcwq->lock);
3686	start_worker(worker);
3687	spin_unlock_irq(&gcwq->lock);
3688	}
3689
3690	system_wq = alloc_workqueue("events", 0, 0);
3691	system_long_wq = alloc_workqueue("events_long", 0, 0);
3692	system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3693	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3694	WQ_UNBOUND_MAX_ACTIVE);
3695	BUG_ON(!system_wq \|\| !system_long_wq \|\| !system_nrt_wq);
3696	return 0;
3697	}
3698	early_initcall(init_workqueues);