linux - kernel/workqueue.c

3890 lines · 106.2 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/export.h>
27	#include <linux/kernel.h>
28	#include <linux/sched.h>
29	#include <linux/init.h>
30	#include <linux/signal.h>
31	#include <linux/completion.h>
32	#include <linux/workqueue.h>
33	#include <linux/slab.h>
34	#include <linux/cpu.h>
35	#include <linux/notifier.h>
36	#include <linux/kthread.h>
37	#include <linux/hardirq.h>
38	#include <linux/mempolicy.h>
39	#include <linux/freezer.h>
40	#include <linux/kallsyms.h>
41	#include <linux/debug_locks.h>
42	#include <linux/lockdep.h>
43	#include <linux/idr.h>
44
45	#include "workqueue_sched.h"
46
47	enum {
48	/*
49	* global_cwq flags
50	*
51	* A bound gcwq is either associated or disassociated with its CPU.
52	* While associated (!DISASSOCIATED), all workers are bound to the
53	* CPU and none has %WORKER_UNBOUND set and concurrency management
54	* is in effect.
55	*
56	* While DISASSOCIATED, the cpu may be offline and all workers have
57	* %WORKER_UNBOUND set and concurrency management disabled, and may
58	* be executing on any CPU. The gcwq behaves as an unbound one.
59	*
60	* Note that DISASSOCIATED can be flipped only while holding
61	* assoc_mutex of all pools on the gcwq to avoid changing binding
62	* state while create_worker() is in progress.
63	*/
64	GCWQ_DISASSOCIATED = 1 << 0, /* cpu can't serve workers */
65	GCWQ_FREEZING = 1 << 1, /* freeze in progress */
66
67	/* pool flags */
68	POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
69	POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */
70
71	/* worker flags */
72	WORKER_STARTED = 1 << 0, /* started */
73	WORKER_DIE = 1 << 1, /* die die die */
74	WORKER_IDLE = 1 << 2, /* is idle */
75	WORKER_PREP = 1 << 3, /* preparing to run works */
76	WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
77	WORKER_UNBOUND = 1 << 7, /* worker is unbound */
78
79	WORKER_NOT_RUNNING = WORKER_PREP \| WORKER_UNBOUND \|
80	WORKER_CPU_INTENSIVE,
81
82	NR_WORKER_POOLS = 2, /* # worker pools per gcwq */
83
84	BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
85	BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
86	BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
87
88	MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
89	IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
90
91	MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
92	/* call for help after 10ms
93	(min two ticks) */
94	MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
95	CREATE_COOLDOWN = HZ, /* time to breath after fail */
96
97	/*
98	* Rescue workers are used only on emergencies and shared by
99	* all cpus. Give -20.
100	*/
101	RESCUER_NICE_LEVEL = -20,
102	HIGHPRI_NICE_LEVEL = -20,
103	};
104
105	/*
106	* Structure fields follow one of the following exclusion rules.
107	*
108	* I: Modifiable by initialization/destruction paths and read-only for
109	* everyone else.
110	*
111	* P: Preemption protected. Disabling preemption is enough and should
112	* only be modified and accessed from the local cpu.
113	*
114	* L: gcwq->lock protected. Access with gcwq->lock held.
115	*
116	* X: During normal operation, modification requires gcwq->lock and
117	* should be done only from local cpu. Either disabling preemption
118	* on local cpu or grabbing gcwq->lock is enough for read access.
119	* If GCWQ_DISASSOCIATED is set, it's identical to L.
120	*
121	* F: wq->flush_mutex protected.
122	*
123	* W: workqueue_lock protected.
124	*/
125
126	struct global_cwq;
127	struct worker_pool;
128
129	/*
130	* The poor guys doing the actual heavy lifting. All on-duty workers
131	* are either serving the manager role, on idle list or on busy hash.
132	*/
133	struct worker {
134	/* on idle list while idle, on busy hash table while busy */
135	union {
136	struct list_head entry; /* L: while idle */
137	struct hlist_node hentry; /* L: while busy */
138	};
139
140	struct work_struct current_work; / L: work being processed */
141	struct cpu_workqueue_struct current_cwq; / L: current_work's cwq */
142	struct list_head scheduled; /* L: scheduled works */
143	struct task_struct task; / I: worker task */
144	struct worker_pool pool; / I: the associated pool */
145	/* 64 bytes boundary on 64bit, 32 on 32bit */
146	unsigned long last_active; /* L: last active timestamp */
147	unsigned int flags; /* X: flags */
148	int id; /* I: worker id */
149
150	/* for rebinding worker to CPU */
151	struct work_struct rebind_work; /* L: for busy worker */
152	};
153
154	struct worker_pool {
155	struct global_cwq gcwq; / I: the owning gcwq */
156	unsigned int flags; /* X: flags */
157
158	struct list_head worklist; /* L: list of pending works */
159	int nr_workers; /* L: total number of workers */
160
161	/* nr_idle includes the ones off idle_list for rebinding */
162	int nr_idle; /* L: currently idle ones */
163
164	struct list_head idle_list; /* X: list of idle workers */
165	struct timer_list idle_timer; /* L: worker idle timeout */
166	struct timer_list mayday_timer; /* L: SOS timer for workers */
167
168	struct mutex assoc_mutex; /* protect GCWQ_DISASSOCIATED */
169	struct ida worker_ida; /* L: for worker IDs */
170	};
171
172	/*
173	* Global per-cpu workqueue. There's one and only one for each cpu
174	* and all works are queued and processed here regardless of their
175	* target workqueues.
176	*/
177	struct global_cwq {
178	spinlock_t lock; /* the gcwq lock */
179	unsigned int cpu; /* I: the associated cpu */
180	unsigned int flags; /* L: GCWQ_* flags */
181
182	/* workers are chained either in busy_hash or pool idle_list */
183	struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
184	/* L: hash of busy workers */
185
186	struct worker_pool pools[NR_WORKER_POOLS];
187	/* normal and highpri pools */
188	} ____cacheline_aligned_in_smp;
189
190	/*
191	* The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
192	* work_struct->data are used for flags and thus cwqs need to be
193	* aligned at two's power of the number of flag bits.
194	*/
195	struct cpu_workqueue_struct {
196	struct worker_pool pool; / I: the associated pool */
197	struct workqueue_struct wq; / I: the owning workqueue */
198	int work_color; /* L: current color */
199	int flush_color; /* L: flushing color */
200	int nr_in_flight[WORK_NR_COLORS];
201	/* L: nr of in_flight works */
202	int nr_active; /* L: nr of active works */
203	int max_active; /* L: max active works */
204	struct list_head delayed_works; /* L: delayed works */
205	};
206
207	/*
208	* Structure used to wait for workqueue flush.
209	*/
210	struct wq_flusher {
211	struct list_head list; /* F: list of flushers */
212	int flush_color; /* F: flush color waiting for */
213	struct completion done; /* flush completion */
214	};
215
216	/*
217	* All cpumasks are assumed to be always set on UP and thus can't be
218	* used to determine whether there's something to be done.
219	*/
220	#ifdef CONFIG_SMP
221	typedef cpumask_var_t mayday_mask_t;
222	#define mayday_test_and_set_cpu(cpu, mask) \
223	cpumask_test_and_set_cpu((cpu), (mask))
224	#define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
225	#define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
226	#define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
227	#define free_mayday_mask(mask) free_cpumask_var((mask))
228	#else
229	typedef unsigned long mayday_mask_t;
230	#define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
231	#define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
232	#define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
233	#define alloc_mayday_mask(maskp, gfp) true
234	#define free_mayday_mask(mask) do { } while (0)
235	#endif
236
237	/*
238	* The externally visible workqueue abstraction is an array of
239	* per-CPU workqueues:
240	*/
241	struct workqueue_struct {
242	unsigned int flags; /* W: WQ_* flags */
243	union {
244	struct cpu_workqueue_struct __percpu *pcpu;
245	struct cpu_workqueue_struct *single;
246	unsigned long v;
247	} cpu_wq; /* I: cwq's */
248	struct list_head list; /* W: list of all workqueues */
249
250	struct mutex flush_mutex; /* protects wq flushing */
251	int work_color; /* F: current work color */
252	int flush_color; /* F: current flush color */
253	atomic_t nr_cwqs_to_flush; /* flush in progress */
254	struct wq_flusher first_flusher; / F: first flusher */
255	struct list_head flusher_queue; /* F: flush waiters */
256	struct list_head flusher_overflow; /* F: flush overflow list */
257
258	mayday_mask_t mayday_mask; /* cpus requesting rescue */
259	struct worker rescuer; / I: rescue worker */
260
261	int nr_drainers; /* W: drain in progress */
262	int saved_max_active; /* W: saved cwq max_active */
263	#ifdef CONFIG_LOCKDEP
264	struct lockdep_map lockdep_map;
265	#endif
266	char name[]; /* I: workqueue name */
267	};
268
269	struct workqueue_struct *system_wq __read_mostly;
270	EXPORT_SYMBOL_GPL(system_wq);
271	struct workqueue_struct *system_highpri_wq __read_mostly;
272	EXPORT_SYMBOL_GPL(system_highpri_wq);
273	struct workqueue_struct *system_long_wq __read_mostly;
274	EXPORT_SYMBOL_GPL(system_long_wq);
275	struct workqueue_struct *system_unbound_wq __read_mostly;
276	EXPORT_SYMBOL_GPL(system_unbound_wq);
277	struct workqueue_struct *system_freezable_wq __read_mostly;
278	EXPORT_SYMBOL_GPL(system_freezable_wq);
279
280	#define CREATE_TRACE_POINTS
281	#include <trace/events/workqueue.h>
282
283	#define for_each_worker_pool(pool, gcwq) \
284	for ((pool) = &(gcwq)->pools[0]; \
285	(pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
286
287	#define for_each_busy_worker(worker, i, pos, gcwq) \
288	for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
289	hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
290
291	static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
292	unsigned int sw)
293	{
294	if (cpu < nr_cpu_ids) {
295	if (sw & 1) {
296	cpu = cpumask_next(cpu, mask);
297	if (cpu < nr_cpu_ids)
298	return cpu;
299	}
300	if (sw & 2)
301	return WORK_CPU_UNBOUND;
302	}
303	return WORK_CPU_NONE;
304	}
305
306	static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
307	struct workqueue_struct *wq)
308	{
309	return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
310	}
311
312	/*
313	* CPU iterators
314	*
315	* An extra gcwq is defined for an invalid cpu number
316	* (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
317	* specific CPU. The following iterators are similar to
318	* for_each_*_cpu() iterators but also considers the unbound gcwq.
319	*
320	* for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
321	* for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
322	* for_each_cwq_cpu() : possible CPUs for bound workqueues,
323	* WORK_CPU_UNBOUND for unbound workqueues
324	*/
325	#define for_each_gcwq_cpu(cpu) \
326	for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
327	(cpu) < WORK_CPU_NONE; \
328	(cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
329
330	#define for_each_online_gcwq_cpu(cpu) \
331	for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
332	(cpu) < WORK_CPU_NONE; \
333	(cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
334
335	#define for_each_cwq_cpu(cpu, wq) \
336	for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
337	(cpu) < WORK_CPU_NONE; \
338	(cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
339
340	#ifdef CONFIG_DEBUG_OBJECTS_WORK
341
342	static struct debug_obj_descr work_debug_descr;
343
344	static void work_debug_hint(void addr)
345	{
346	return ((struct work_struct *) addr)->func;
347	}
348
349	/*
350	* fixup_init is called when:
351	* - an active object is initialized
352	*/
353	static int work_fixup_init(void *addr, enum debug_obj_state state)
354	{
355	struct work_struct *work = addr;
356
357	switch (state) {
358	case ODEBUG_STATE_ACTIVE:
359	cancel_work_sync(work);
360	debug_object_init(work, &work_debug_descr);
361	return 1;
362	default:
363	return 0;
364	}
365	}
366
367	/*
368	* fixup_activate is called when:
369	* - an active object is activated
370	* - an unknown object is activated (might be a statically initialized object)
371	*/
372	static int work_fixup_activate(void *addr, enum debug_obj_state state)
373	{
374	struct work_struct *work = addr;
375
376	switch (state) {
377
378	case ODEBUG_STATE_NOTAVAILABLE:
379	/*
380	* This is not really a fixup. The work struct was
381	* statically initialized. We just make sure that it
382	* is tracked in the object tracker.
383	*/
384	if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
385	debug_object_init(work, &work_debug_descr);
386	debug_object_activate(work, &work_debug_descr);
387	return 0;
388	}
389	WARN_ON_ONCE(1);
390	return 0;
391
392	case ODEBUG_STATE_ACTIVE:
393	WARN_ON(1);
394
395	default:
396	return 0;
397	}
398	}
399
400	/*
401	* fixup_free is called when:
402	* - an active object is freed
403	*/
404	static int work_fixup_free(void *addr, enum debug_obj_state state)
405	{
406	struct work_struct *work = addr;
407
408	switch (state) {
409	case ODEBUG_STATE_ACTIVE:
410	cancel_work_sync(work);
411	debug_object_free(work, &work_debug_descr);
412	return 1;
413	default:
414	return 0;
415	}
416	}
417
418	static struct debug_obj_descr work_debug_descr = {
419	.name = "work_struct",
420	.debug_hint = work_debug_hint,
421	.fixup_init = work_fixup_init,
422	.fixup_activate = work_fixup_activate,
423	.fixup_free = work_fixup_free,
424	};
425
426	static inline void debug_work_activate(struct work_struct *work)
427	{
428	debug_object_activate(work, &work_debug_descr);
429	}
430
431	static inline void debug_work_deactivate(struct work_struct *work)
432	{
433	debug_object_deactivate(work, &work_debug_descr);
434	}
435
436	void __init_work(struct work_struct *work, int onstack)
437	{
438	if (onstack)
439	debug_object_init_on_stack(work, &work_debug_descr);
440	else
441	debug_object_init(work, &work_debug_descr);
442	}
443	EXPORT_SYMBOL_GPL(__init_work);
444
445	void destroy_work_on_stack(struct work_struct *work)
446	{
447	debug_object_free(work, &work_debug_descr);
448	}
449	EXPORT_SYMBOL_GPL(destroy_work_on_stack);
450
451	#else
452	static inline void debug_work_activate(struct work_struct *work) { }
453	static inline void debug_work_deactivate(struct work_struct *work) { }
454	#endif
455
456	/* Serializes the accesses to the list of workqueues. */
457	static DEFINE_SPINLOCK(workqueue_lock);
458	static LIST_HEAD(workqueues);
459	static bool workqueue_freezing; /* W: have wqs started freezing? */
460
461	/*
462	* The almighty global cpu workqueues. nr_running is the only field
463	* which is expected to be used frequently by other cpus via
464	* try_to_wake_up(). Put it in a separate cacheline.
465	*/
466	static DEFINE_PER_CPU(struct global_cwq, global_cwq);
467	static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]);
468
469	/*
470	* Global cpu workqueue and nr_running counter for unbound gcwq. The
471	* gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
472	* workers have WORKER_UNBOUND set.
473	*/
474	static struct global_cwq unbound_global_cwq;
475	static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = {
476	[0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
477	};
478
479	static int worker_thread(void *__worker);
480
481	static int worker_pool_pri(struct worker_pool *pool)
482	{
483	return pool - pool->gcwq->pools;
484	}
485
486	static struct global_cwq *get_gcwq(unsigned int cpu)
487	{
488	if (cpu != WORK_CPU_UNBOUND)
489	return &per_cpu(global_cwq, cpu);
490	else
491	return &unbound_global_cwq;
492	}
493
494	static atomic_t get_pool_nr_running(struct worker_pool pool)
495	{
496	int cpu = pool->gcwq->cpu;
497	int idx = worker_pool_pri(pool);
498
499	if (cpu != WORK_CPU_UNBOUND)
500	return &per_cpu(pool_nr_running, cpu)[idx];
501	else
502	return &unbound_pool_nr_running[idx];
503	}
504
505	static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
506	struct workqueue_struct *wq)
507	{
508	if (!(wq->flags & WQ_UNBOUND)) {
509	if (likely(cpu < nr_cpu_ids))
510	return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
511	} else if (likely(cpu == WORK_CPU_UNBOUND))
512	return wq->cpu_wq.single;
513	return NULL;
514	}
515
516	static unsigned int work_color_to_flags(int color)
517	{
518	return color << WORK_STRUCT_COLOR_SHIFT;
519	}
520
521	static int get_work_color(struct work_struct *work)
522	{
523	return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
524	((1 << WORK_STRUCT_COLOR_BITS) - 1);
525	}
526
527	static int work_next_color(int color)
528	{
529	return (color + 1) % WORK_NR_COLORS;
530	}
531
532	/*
533	* While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
534	* contain the pointer to the queued cwq. Once execution starts, the flag
535	* is cleared and the high bits contain OFFQ flags and CPU number.
536	*
537	* set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
538	* and clear_work_data() can be used to set the cwq, cpu or clear
539	* work->data. These functions should only be called while the work is
540	* owned - ie. while the PENDING bit is set.
541	*
542	* get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
543	* a work. gcwq is available once the work has been queued anywhere after
544	* initialization until it is sync canceled. cwq is available only while
545	* the work item is queued.
546	*
547	* %WORK_OFFQ_CANCELING is used to mark a work item which is being
548	* canceled. While being canceled, a work item may have its PENDING set
549	* but stay off timer and worklist for arbitrarily long and nobody should
550	* try to steal the PENDING bit.
551	*/
552	static inline void set_work_data(struct work_struct *work, unsigned long data,
553	unsigned long flags)
554	{
555	BUG_ON(!work_pending(work));
556	atomic_long_set(&work->data, data \| flags \| work_static(work));
557	}
558
559	static void set_work_cwq(struct work_struct *work,
560	struct cpu_workqueue_struct *cwq,
561	unsigned long extra_flags)
562	{
563	set_work_data(work, (unsigned long)cwq,
564	WORK_STRUCT_PENDING \| WORK_STRUCT_CWQ \| extra_flags);
565	}
566
567	static void set_work_cpu_and_clear_pending(struct work_struct *work,
568	unsigned int cpu)
569	{
570	/*
571	* The following wmb is paired with the implied mb in
572	* test_and_set_bit(PENDING) and ensures all updates to @work made
573	* here are visible to and precede any updates by the next PENDING
574	* owner.
575	*/
576	smp_wmb();
577	set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0);
578	}
579
580	static void clear_work_data(struct work_struct *work)
581	{
582	smp_wmb(); /* see set_work_cpu_and_clear_pending() */
583	set_work_data(work, WORK_STRUCT_NO_CPU, 0);
584	}
585
586	static struct cpu_workqueue_struct get_work_cwq(struct work_struct work)
587	{
588	unsigned long data = atomic_long_read(&work->data);
589
590	if (data & WORK_STRUCT_CWQ)
591	return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
592	else
593	return NULL;
594	}
595
596	static struct global_cwq get_work_gcwq(struct work_struct work)
597	{
598	unsigned long data = atomic_long_read(&work->data);
599	unsigned int cpu;
600
601	if (data & WORK_STRUCT_CWQ)
602	return ((struct cpu_workqueue_struct *)
603	(data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq;
604
605	cpu = data >> WORK_OFFQ_CPU_SHIFT;
606	if (cpu == WORK_CPU_NONE)
607	return NULL;
608
609	BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
610	return get_gcwq(cpu);
611	}
612
613	static void mark_work_canceling(struct work_struct *work)
614	{
615	struct global_cwq *gcwq = get_work_gcwq(work);
616	unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE;
617
618	set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) \| WORK_OFFQ_CANCELING,
619	WORK_STRUCT_PENDING);
620	}
621
622	static bool work_is_canceling(struct work_struct *work)
623	{
624	unsigned long data = atomic_long_read(&work->data);
625
626	return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
627	}
628
629	/*
630	* Policy functions. These define the policies on how the global worker
631	* pools are managed. Unless noted otherwise, these functions assume that
632	* they're being called with gcwq->lock held.
633	*/
634
635	static bool __need_more_worker(struct worker_pool *pool)
636	{
637	return !atomic_read(get_pool_nr_running(pool));
638	}
639
640	/*
641	* Need to wake up a worker? Called from anything but currently
642	* running workers.
643	*
644	* Note that, because unbound workers never contribute to nr_running, this
645	* function will always return %true for unbound gcwq as long as the
646	* worklist isn't empty.
647	*/
648	static bool need_more_worker(struct worker_pool *pool)
649	{
650	return !list_empty(&pool->worklist) && __need_more_worker(pool);
651	}
652
653	/* Can I start working? Called from busy but !running workers. */
654	static bool may_start_working(struct worker_pool *pool)
655	{
656	return pool->nr_idle;
657	}
658
659	/* Do I need to keep working? Called from currently running workers. */
660	static bool keep_working(struct worker_pool *pool)
661	{
662	atomic_t *nr_running = get_pool_nr_running(pool);
663
664	return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
665	}
666
667	/* Do we need a new worker? Called from manager. */
668	static bool need_to_create_worker(struct worker_pool *pool)
669	{
670	return need_more_worker(pool) && !may_start_working(pool);
671	}
672
673	/* Do I need to be the manager? */
674	static bool need_to_manage_workers(struct worker_pool *pool)
675	{
676	return need_to_create_worker(pool) \|\|
677	(pool->flags & POOL_MANAGE_WORKERS);
678	}
679
680	/* Do we have too many workers and should some go away? */
681	static bool too_many_workers(struct worker_pool *pool)
682	{
683	bool managing = pool->flags & POOL_MANAGING_WORKERS;
684	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
685	int nr_busy = pool->nr_workers - nr_idle;
686
687	/*
688	* nr_idle and idle_list may disagree if idle rebinding is in
689	* progress. Never return %true if idle_list is empty.
690	*/
691	if (list_empty(&pool->idle_list))
692	return false;
693
694	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
695	}
696
697	/*
698	* Wake up functions.
699	*/
700
701	/* Return the first worker. Safe with preemption disabled */
702	static struct worker first_worker(struct worker_pool pool)
703	{
704	if (unlikely(list_empty(&pool->idle_list)))
705	return NULL;
706
707	return list_first_entry(&pool->idle_list, struct worker, entry);
708	}
709
710	/**
711	* wake_up_worker - wake up an idle worker
712	* @pool: worker pool to wake worker from
713	*
714	* Wake up the first idle worker of @pool.
715	*
716	* CONTEXT:
717	* spin_lock_irq(gcwq->lock).
718	*/
719	static void wake_up_worker(struct worker_pool *pool)
720	{
721	struct worker *worker = first_worker(pool);
722
723	if (likely(worker))
724	wake_up_process(worker->task);
725	}
726
727	/**
728	* wq_worker_waking_up - a worker is waking up
729	* @task: task waking up
730	* @cpu: CPU @task is waking up to
731	*
732	* This function is called during try_to_wake_up() when a worker is
733	* being awoken.
734	*
735	* CONTEXT:
736	* spin_lock_irq(rq->lock)
737	*/
738	void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
739	{
740	struct worker *worker = kthread_data(task);
741
742	if (!(worker->flags & WORKER_NOT_RUNNING))
743	atomic_inc(get_pool_nr_running(worker->pool));
744	}
745
746	/**
747	* wq_worker_sleeping - a worker is going to sleep
748	* @task: task going to sleep
749	* @cpu: CPU in question, must be the current CPU number
750	*
751	* This function is called during schedule() when a busy worker is
752	* going to sleep. Worker on the same cpu can be woken up by
753	* returning pointer to its task.
754	*
755	* CONTEXT:
756	* spin_lock_irq(rq->lock)
757	*
758	* RETURNS:
759	* Worker task on @cpu to wake up, %NULL if none.
760	*/
761	struct task_struct wq_worker_sleeping(struct task_struct task,
762	unsigned int cpu)
763	{
764	struct worker worker = kthread_data(task), to_wakeup = NULL;
765	struct worker_pool *pool = worker->pool;
766	atomic_t *nr_running = get_pool_nr_running(pool);
767
768	if (worker->flags & WORKER_NOT_RUNNING)
769	return NULL;
770
771	/* this can only happen on the local cpu */
772	BUG_ON(cpu != raw_smp_processor_id());
773
774	/*
775	* The counterpart of the following dec_and_test, implied mb,
776	* worklist not empty test sequence is in insert_work().
777	* Please read comment there.
778	*
779	* NOT_RUNNING is clear. This means that we're bound to and
780	* running on the local cpu w/ rq lock held and preemption
781	* disabled, which in turn means that none else could be
782	* manipulating idle_list, so dereferencing idle_list without gcwq
783	* lock is safe.
784	*/
785	if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
786	to_wakeup = first_worker(pool);
787	return to_wakeup ? to_wakeup->task : NULL;
788	}
789
790	/**
791	* worker_set_flags - set worker flags and adjust nr_running accordingly
792	* @worker: self
793	* @flags: flags to set
794	* @wakeup: wakeup an idle worker if necessary
795	*
796	* Set @flags in @worker->flags and adjust nr_running accordingly. If
797	* nr_running becomes zero and @wakeup is %true, an idle worker is
798	* woken up.
799	*
800	* CONTEXT:
801	* spin_lock_irq(gcwq->lock)
802	*/
803	static inline void worker_set_flags(struct worker *worker, unsigned int flags,
804	bool wakeup)
805	{
806	struct worker_pool *pool = worker->pool;
807
808	WARN_ON_ONCE(worker->task != current);
809
810	/*
811	* If transitioning into NOT_RUNNING, adjust nr_running and
812	* wake up an idle worker as necessary if requested by
813	* @wakeup.
814	*/
815	if ((flags & WORKER_NOT_RUNNING) &&
816	!(worker->flags & WORKER_NOT_RUNNING)) {
817	atomic_t *nr_running = get_pool_nr_running(pool);
818
819	if (wakeup) {
820	if (atomic_dec_and_test(nr_running) &&
821	!list_empty(&pool->worklist))
822	wake_up_worker(pool);
823	} else
824	atomic_dec(nr_running);
825	}
826
827	worker->flags \|= flags;
828	}
829
830	/**
831	* worker_clr_flags - clear worker flags and adjust nr_running accordingly
832	* @worker: self
833	* @flags: flags to clear
834	*
835	* Clear @flags in @worker->flags and adjust nr_running accordingly.
836	*
837	* CONTEXT:
838	* spin_lock_irq(gcwq->lock)
839	*/
840	static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
841	{
842	struct worker_pool *pool = worker->pool;
843	unsigned int oflags = worker->flags;
844
845	WARN_ON_ONCE(worker->task != current);
846
847	worker->flags &= ~flags;
848
849	/*
850	* If transitioning out of NOT_RUNNING, increment nr_running. Note
851	* that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
852	* of multiple flags, not a single flag.
853	*/
854	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
855	if (!(worker->flags & WORKER_NOT_RUNNING))
856	atomic_inc(get_pool_nr_running(pool));
857	}
858
859	/**
860	* busy_worker_head - return the busy hash head for a work
861	* @gcwq: gcwq of interest
862	* @work: work to be hashed
863	*
864	* Return hash head of @gcwq for @work.
865	*
866	* CONTEXT:
867	* spin_lock_irq(gcwq->lock).
868	*
869	* RETURNS:
870	* Pointer to the hash head.
871	*/
872	static struct hlist_head busy_worker_head(struct global_cwq gcwq,
873	struct work_struct *work)
874	{
875	const int base_shift = ilog2(sizeof(struct work_struct));
876	unsigned long v = (unsigned long)work;
877
878	/* simple shift and fold hash, do we need something better? */
879	v >>= base_shift;
880	v += v >> BUSY_WORKER_HASH_ORDER;
881	v &= BUSY_WORKER_HASH_MASK;
882
883	return &gcwq->busy_hash[v];
884	}
885
886	/**
887	* __find_worker_executing_work - find worker which is executing a work
888	* @gcwq: gcwq of interest
889	* @bwh: hash head as returned by busy_worker_head()
890	* @work: work to find worker for
891	*
892	* Find a worker which is executing @work on @gcwq. @bwh should be
893	* the hash head obtained by calling busy_worker_head() with the same
894	* work.
895	*
896	* CONTEXT:
897	* spin_lock_irq(gcwq->lock).
898	*
899	* RETURNS:
900	* Pointer to worker which is executing @work if found, NULL
901	* otherwise.
902	*/
903	static struct worker __find_worker_executing_work(struct global_cwq gcwq,
904	struct hlist_head *bwh,
905	struct work_struct *work)
906	{
907	struct worker *worker;
908	struct hlist_node *tmp;
909
910	hlist_for_each_entry(worker, tmp, bwh, hentry)
911	if (worker->current_work == work)
912	return worker;
913	return NULL;
914	}
915
916	/**
917	* find_worker_executing_work - find worker which is executing a work
918	* @gcwq: gcwq of interest
919	* @work: work to find worker for
920	*
921	* Find a worker which is executing @work on @gcwq. This function is
922	* identical to __find_worker_executing_work() except that this
923	* function calculates @bwh itself.
924	*
925	* CONTEXT:
926	* spin_lock_irq(gcwq->lock).
927	*
928	* RETURNS:
929	* Pointer to worker which is executing @work if found, NULL
930	* otherwise.
931	*/
932	static struct worker find_worker_executing_work(struct global_cwq gcwq,
933	struct work_struct *work)
934	{
935	return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
936	work);
937	}
938
939	/**
940	* move_linked_works - move linked works to a list
941	* @work: start of series of works to be scheduled
942	* @head: target list to append @work to
943	* @nextp: out paramter for nested worklist walking
944	*
945	* Schedule linked works starting from @work to @head. Work series to
946	* be scheduled starts at @work and includes any consecutive work with
947	* WORK_STRUCT_LINKED set in its predecessor.
948	*
949	* If @nextp is not NULL, it's updated to point to the next work of
950	* the last scheduled work. This allows move_linked_works() to be
951	* nested inside outer list_for_each_entry_safe().
952	*
953	* CONTEXT:
954	* spin_lock_irq(gcwq->lock).
955	*/
956	static void move_linked_works(struct work_struct work, struct list_head head,
957	struct work_struct **nextp)
958	{
959	struct work_struct *n;
960
961	/*
962	* Linked worklist will always end before the end of the list,
963	* use NULL for list head.
964	*/
965	list_for_each_entry_safe_from(work, n, NULL, entry) {
966	list_move_tail(&work->entry, head);
967	if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
968	break;
969	}
970
971	/*
972	* If we're already inside safe list traversal and have moved
973	* multiple works to the scheduled queue, the next position
974	* needs to be updated.
975	*/
976	if (nextp)
977	*nextp = n;
978	}
979
980	static void cwq_activate_delayed_work(struct work_struct *work)
981	{
982	struct cpu_workqueue_struct *cwq = get_work_cwq(work);
983
984	trace_workqueue_activate_work(work);
985	move_linked_works(work, &cwq->pool->worklist, NULL);
986	__clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
987	cwq->nr_active++;
988	}
989
990	static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
991	{
992	struct work_struct *work = list_first_entry(&cwq->delayed_works,
993	struct work_struct, entry);
994
995	cwq_activate_delayed_work(work);
996	}
997
998	/**
999	* cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1000	* @cwq: cwq of interest
1001	* @color: color of work which left the queue
1002	*
1003	* A work either has completed or is removed from pending queue,
1004	* decrement nr_in_flight of its cwq and handle workqueue flushing.
1005	*
1006	* CONTEXT:
1007	* spin_lock_irq(gcwq->lock).
1008	*/
1009	static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
1010	{
1011	/* ignore uncolored works */
1012	if (color == WORK_NO_COLOR)
1013	return;
1014
1015	cwq->nr_in_flight[color]--;
1016
1017	cwq->nr_active--;
1018	if (!list_empty(&cwq->delayed_works)) {
1019	/* one down, submit a delayed one */
1020	if (cwq->nr_active < cwq->max_active)
1021	cwq_activate_first_delayed(cwq);
1022	}
1023
1024	/* is flush in progress and are we at the flushing tip? */
1025	if (likely(cwq->flush_color != color))
1026	return;
1027
1028	/* are there still in-flight works? */
1029	if (cwq->nr_in_flight[color])
1030	return;
1031
1032	/* this cwq is done, clear flush_color */
1033	cwq->flush_color = -1;
1034
1035	/*
1036	* If this was the last cwq, wake up the first flusher. It
1037	* will handle the rest.
1038	*/
1039	if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1040	complete(&cwq->wq->first_flusher->done);
1041	}
1042
1043	/**
1044	* try_to_grab_pending - steal work item from worklist and disable irq
1045	* @work: work item to steal
1046	* @is_dwork: @work is a delayed_work
1047	* @flags: place to store irq state
1048	*
1049	* Try to grab PENDING bit of @work. This function can handle @work in any
1050	* stable state - idle, on timer or on worklist. Return values are
1051	*
1052	* 1 if @work was pending and we successfully stole PENDING
1053	* 0 if @work was idle and we claimed PENDING
1054	* -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1055	* -ENOENT if someone else is canceling @work, this state may persist
1056	* for arbitrarily long
1057	*
1058	* On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1059	* interrupted while holding PENDING and @work off queue, irq must be
1060	* disabled on entry. This, combined with delayed_work->timer being
1061	* irqsafe, ensures that we return -EAGAIN for finite short period of time.
1062	*
1063	* On successful return, >= 0, irq is disabled and the caller is
1064	* responsible for releasing it using local_irq_restore(*@flags).
1065	*
1066	* This function is safe to call from any context including IRQ handler.
1067	*/
1068	static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1069	unsigned long *flags)
1070	{
1071	struct global_cwq *gcwq;
1072
1073	local_irq_save(*flags);
1074
1075	/* try to steal the timer if it exists */
1076	if (is_dwork) {
1077	struct delayed_work *dwork = to_delayed_work(work);
1078
1079	/*
1080	* dwork->timer is irqsafe. If del_timer() fails, it's
1081	* guaranteed that the timer is not queued anywhere and not
1082	* running on the local CPU.
1083	*/
1084	if (likely(del_timer(&dwork->timer)))
1085	return 1;
1086	}
1087
1088	/* try to claim PENDING the normal way */
1089	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1090	return 0;
1091
1092	/*
1093	* The queueing is in progress, or it is already queued. Try to
1094	* steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1095	*/
1096	gcwq = get_work_gcwq(work);
1097	if (!gcwq)
1098	goto fail;
1099
1100	spin_lock(&gcwq->lock);
1101	if (!list_empty(&work->entry)) {
1102	/*
1103	* This work is queued, but perhaps we locked the wrong gcwq.
1104	* In that case we must see the new value after rmb(), see
1105	* insert_work()->wmb().
1106	*/
1107	smp_rmb();
1108	if (gcwq == get_work_gcwq(work)) {
1109	debug_work_deactivate(work);
1110
1111	/*
1112	* A delayed work item cannot be grabbed directly
1113	* because it might have linked NO_COLOR work items
1114	* which, if left on the delayed_list, will confuse
1115	* cwq->nr_active management later on and cause
1116	* stall. Make sure the work item is activated
1117	* before grabbing.
1118	*/
1119	if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1120	cwq_activate_delayed_work(work);
1121
1122	list_del_init(&work->entry);
1123	cwq_dec_nr_in_flight(get_work_cwq(work),
1124	get_work_color(work));
1125
1126	spin_unlock(&gcwq->lock);
1127	return 1;
1128	}
1129	}
1130	spin_unlock(&gcwq->lock);
1131	fail:
1132	local_irq_restore(*flags);
1133	if (work_is_canceling(work))
1134	return -ENOENT;
1135	cpu_relax();
1136	return -EAGAIN;
1137	}
1138
1139	/**
1140	* insert_work - insert a work into gcwq
1141	* @cwq: cwq @work belongs to
1142	* @work: work to insert
1143	* @head: insertion point
1144	* @extra_flags: extra WORK_STRUCT_* flags to set
1145	*
1146	* Insert @work which belongs to @cwq into @gcwq after @head.
1147	* @extra_flags is or'd to work_struct flags.
1148	*
1149	* CONTEXT:
1150	* spin_lock_irq(gcwq->lock).
1151	*/
1152	static void insert_work(struct cpu_workqueue_struct *cwq,
1153	struct work_struct work, struct list_head head,
1154	unsigned int extra_flags)
1155	{
1156	struct worker_pool *pool = cwq->pool;
1157
1158	/* we own @work, set data and link */
1159	set_work_cwq(work, cwq, extra_flags);
1160
1161	/*
1162	* Ensure that we get the right work->data if we see the
1163	* result of list_add() below, see try_to_grab_pending().
1164	*/
1165	smp_wmb();
1166
1167	list_add_tail(&work->entry, head);
1168
1169	/*
1170	* Ensure either worker_sched_deactivated() sees the above
1171	* list_add_tail() or we see zero nr_running to avoid workers
1172	* lying around lazily while there are works to be processed.
1173	*/
1174	smp_mb();
1175
1176	if (__need_more_worker(pool))
1177	wake_up_worker(pool);
1178	}
1179
1180	/*
1181	* Test whether @work is being queued from another work executing on the
1182	* same workqueue. This is rather expensive and should only be used from
1183	* cold paths.
1184	*/
1185	static bool is_chained_work(struct workqueue_struct *wq)
1186	{
1187	unsigned long flags;
1188	unsigned int cpu;
1189
1190	for_each_gcwq_cpu(cpu) {
1191	struct global_cwq *gcwq = get_gcwq(cpu);
1192	struct worker *worker;
1193	struct hlist_node *pos;
1194	int i;
1195
1196	spin_lock_irqsave(&gcwq->lock, flags);
1197	for_each_busy_worker(worker, i, pos, gcwq) {
1198	if (worker->task != current)
1199	continue;
1200	spin_unlock_irqrestore(&gcwq->lock, flags);
1201	/*
1202	* I'm @worker, no locking necessary. See if @work
1203	* is headed to the same workqueue.
1204	*/
1205	return worker->current_cwq->wq == wq;
1206	}
1207	spin_unlock_irqrestore(&gcwq->lock, flags);
1208	}
1209	return false;
1210	}
1211
1212	static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1213	struct work_struct *work)
1214	{
1215	struct global_cwq *gcwq;
1216	struct cpu_workqueue_struct *cwq;
1217	struct list_head *worklist;
1218	unsigned int work_flags;
1219	unsigned int req_cpu = cpu;
1220
1221	/*
1222	* While a work item is PENDING && off queue, a task trying to
1223	* steal the PENDING will busy-loop waiting for it to either get
1224	* queued or lose PENDING. Grabbing PENDING and queueing should
1225	* happen with IRQ disabled.
1226	*/
1227	WARN_ON_ONCE(!irqs_disabled());
1228
1229	debug_work_activate(work);
1230
1231	/* if dying, only works from the same workqueue are allowed */
1232	if (unlikely(wq->flags & WQ_DRAINING) &&
1233	WARN_ON_ONCE(!is_chained_work(wq)))
1234	return;
1235
1236	/* determine gcwq to use */
1237	if (!(wq->flags & WQ_UNBOUND)) {
1238	struct global_cwq *last_gcwq;
1239
1240	if (cpu == WORK_CPU_UNBOUND)
1241	cpu = raw_smp_processor_id();
1242
1243	/*
1244	* It's multi cpu. If @work was previously on a different
1245	* cpu, it might still be running there, in which case the
1246	* work needs to be queued on that cpu to guarantee
1247	* non-reentrancy.
1248	*/
1249	gcwq = get_gcwq(cpu);
1250	last_gcwq = get_work_gcwq(work);
1251
1252	if (last_gcwq && last_gcwq != gcwq) {
1253	struct worker *worker;
1254
1255	spin_lock(&last_gcwq->lock);
1256
1257	worker = find_worker_executing_work(last_gcwq, work);
1258
1259	if (worker && worker->current_cwq->wq == wq)
1260	gcwq = last_gcwq;
1261	else {
1262	/* meh... not running there, queue here */
1263	spin_unlock(&last_gcwq->lock);
1264	spin_lock(&gcwq->lock);
1265	}
1266	} else {
1267	spin_lock(&gcwq->lock);
1268	}
1269	} else {
1270	gcwq = get_gcwq(WORK_CPU_UNBOUND);
1271	spin_lock(&gcwq->lock);
1272	}
1273
1274	/* gcwq determined, get cwq and queue */
1275	cwq = get_cwq(gcwq->cpu, wq);
1276	trace_workqueue_queue_work(req_cpu, cwq, work);
1277
1278	if (WARN_ON(!list_empty(&work->entry))) {
1279	spin_unlock(&gcwq->lock);
1280	return;
1281	}
1282
1283	cwq->nr_in_flight[cwq->work_color]++;
1284	work_flags = work_color_to_flags(cwq->work_color);
1285
1286	if (likely(cwq->nr_active < cwq->max_active)) {
1287	trace_workqueue_activate_work(work);
1288	cwq->nr_active++;
1289	worklist = &cwq->pool->worklist;
1290	} else {
1291	work_flags \|= WORK_STRUCT_DELAYED;
1292	worklist = &cwq->delayed_works;
1293	}
1294
1295	insert_work(cwq, work, worklist, work_flags);
1296
1297	spin_unlock(&gcwq->lock);
1298	}
1299
1300	/**
1301	* queue_work_on - queue work on specific cpu
1302	* @cpu: CPU number to execute work on
1303	* @wq: workqueue to use
1304	* @work: work to queue
1305	*
1306	* Returns %false if @work was already on a queue, %true otherwise.
1307	*
1308	* We queue the work to a specific CPU, the caller must ensure it
1309	* can't go away.
1310	*/
1311	bool queue_work_on(int cpu, struct workqueue_struct *wq,
1312	struct work_struct *work)
1313	{
1314	bool ret = false;
1315	unsigned long flags;
1316
1317	local_irq_save(flags);
1318
1319	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1320	__queue_work(cpu, wq, work);
1321	ret = true;
1322	}
1323
1324	local_irq_restore(flags);
1325	return ret;
1326	}
1327	EXPORT_SYMBOL_GPL(queue_work_on);
1328
1329	/**
1330	* queue_work - queue work on a workqueue
1331	* @wq: workqueue to use
1332	* @work: work to queue
1333	*
1334	* Returns %false if @work was already on a queue, %true otherwise.
1335	*
1336	* We queue the work to the CPU on which it was submitted, but if the CPU dies
1337	* it can be processed by another CPU.
1338	*/
1339	bool queue_work(struct workqueue_struct wq, struct work_struct work)
1340	{
1341	return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1342	}
1343	EXPORT_SYMBOL_GPL(queue_work);
1344
1345	void delayed_work_timer_fn(unsigned long __data)
1346	{
1347	struct delayed_work dwork = (struct delayed_work )__data;
1348	struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1349
1350	/* should have been called from irqsafe timer with irq already off */
1351	__queue_work(dwork->cpu, cwq->wq, &dwork->work);
1352	}
1353	EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1354
1355	static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1356	struct delayed_work *dwork, unsigned long delay)
1357	{
1358	struct timer_list *timer = &dwork->timer;
1359	struct work_struct *work = &dwork->work;
1360	unsigned int lcpu;
1361
1362	WARN_ON_ONCE(timer->function != delayed_work_timer_fn \|\|
1363	timer->data != (unsigned long)dwork);
1364	BUG_ON(timer_pending(timer));
1365	BUG_ON(!list_empty(&work->entry));
1366
1367	timer_stats_timer_set_start_info(&dwork->timer);
1368
1369	/*
1370	* This stores cwq for the moment, for the timer_fn. Note that the
1371	* work's gcwq is preserved to allow reentrance detection for
1372	* delayed works.
1373	*/
1374	if (!(wq->flags & WQ_UNBOUND)) {
1375	struct global_cwq *gcwq = get_work_gcwq(work);
1376
1377	/*
1378	* If we cannot get the last gcwq from @work directly,
1379	* select the last CPU such that it avoids unnecessarily
1380	* triggering non-reentrancy check in __queue_work().
1381	*/
1382	lcpu = cpu;
1383	if (gcwq)
1384	lcpu = gcwq->cpu;
1385	if (lcpu == WORK_CPU_UNBOUND)
1386	lcpu = raw_smp_processor_id();
1387	} else {
1388	lcpu = WORK_CPU_UNBOUND;
1389	}
1390
1391	set_work_cwq(work, get_cwq(lcpu, wq), 0);
1392
1393	dwork->cpu = cpu;
1394	timer->expires = jiffies + delay;
1395
1396	if (unlikely(cpu != WORK_CPU_UNBOUND))
1397	add_timer_on(timer, cpu);
1398	else
1399	add_timer(timer);
1400	}
1401
1402	/**
1403	* queue_delayed_work_on - queue work on specific CPU after delay
1404	* @cpu: CPU number to execute work on
1405	* @wq: workqueue to use
1406	* @dwork: work to queue
1407	* @delay: number of jiffies to wait before queueing
1408	*
1409	* Returns %false if @work was already on a queue, %true otherwise. If
1410	* @delay is zero and @dwork is idle, it will be scheduled for immediate
1411	* execution.
1412	*/
1413	bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1414	struct delayed_work *dwork, unsigned long delay)
1415	{
1416	struct work_struct *work = &dwork->work;
1417	bool ret = false;
1418	unsigned long flags;
1419
1420	if (!delay)
1421	return queue_work_on(cpu, wq, &dwork->work);
1422
1423	/* read the comment in __queue_work() */
1424	local_irq_save(flags);
1425
1426	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1427	__queue_delayed_work(cpu, wq, dwork, delay);
1428	ret = true;
1429	}
1430
1431	local_irq_restore(flags);
1432	return ret;
1433	}
1434	EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1435
1436	/**
1437	* queue_delayed_work - queue work on a workqueue after delay
1438	* @wq: workqueue to use
1439	* @dwork: delayable work to queue
1440	* @delay: number of jiffies to wait before queueing
1441	*
1442	* Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1443	*/
1444	bool queue_delayed_work(struct workqueue_struct *wq,
1445	struct delayed_work *dwork, unsigned long delay)
1446	{
1447	return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1448	}
1449	EXPORT_SYMBOL_GPL(queue_delayed_work);
1450
1451	/**
1452	* mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1453	* @cpu: CPU number to execute work on
1454	* @wq: workqueue to use
1455	* @dwork: work to queue
1456	* @delay: number of jiffies to wait before queueing
1457	*
1458	* If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1459	* modify @dwork's timer so that it expires after @delay. If @delay is
1460	* zero, @work is guaranteed to be scheduled immediately regardless of its
1461	* current state.
1462	*
1463	* Returns %false if @dwork was idle and queued, %true if @dwork was
1464	* pending and its timer was modified.
1465	*
1466	* This function is safe to call from any context including IRQ handler.
1467	* See try_to_grab_pending() for details.
1468	*/
1469	bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1470	struct delayed_work *dwork, unsigned long delay)
1471	{
1472	unsigned long flags;
1473	int ret;
1474
1475	do {
1476	ret = try_to_grab_pending(&dwork->work, true, &flags);
1477	} while (unlikely(ret == -EAGAIN));
1478
1479	if (likely(ret >= 0)) {
1480	__queue_delayed_work(cpu, wq, dwork, delay);
1481	local_irq_restore(flags);
1482	}
1483
1484	/* -ENOENT from try_to_grab_pending() becomes %true */
1485	return ret;
1486	}
1487	EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1488
1489	/**
1490	* mod_delayed_work - modify delay of or queue a delayed work
1491	* @wq: workqueue to use
1492	* @dwork: work to queue
1493	* @delay: number of jiffies to wait before queueing
1494	*
1495	* mod_delayed_work_on() on local CPU.
1496	*/
1497	bool mod_delayed_work(struct workqueue_struct wq, struct delayed_work dwork,
1498	unsigned long delay)
1499	{
1500	return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1501	}
1502	EXPORT_SYMBOL_GPL(mod_delayed_work);
1503
1504	/**
1505	* worker_enter_idle - enter idle state
1506	* @worker: worker which is entering idle state
1507	*
1508	* @worker is entering idle state. Update stats and idle timer if
1509	* necessary.
1510	*
1511	* LOCKING:
1512	* spin_lock_irq(gcwq->lock).
1513	*/
1514	static void worker_enter_idle(struct worker *worker)
1515	{
1516	struct worker_pool *pool = worker->pool;
1517	struct global_cwq *gcwq = pool->gcwq;
1518
1519	BUG_ON(worker->flags & WORKER_IDLE);
1520	BUG_ON(!list_empty(&worker->entry) &&
1521	(worker->hentry.next \|\| worker->hentry.pprev));
1522
1523	/* can't use worker_set_flags(), also called from start_worker() */
1524	worker->flags \|= WORKER_IDLE;
1525	pool->nr_idle++;
1526	worker->last_active = jiffies;
1527
1528	/* idle_list is LIFO */
1529	list_add(&worker->entry, &pool->idle_list);
1530
1531	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1532	mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1533
1534	/*
1535	* Sanity check nr_running. Because gcwq_unbind_fn() releases
1536	* gcwq->lock between setting %WORKER_UNBOUND and zapping
1537	* nr_running, the warning may trigger spuriously. Check iff
1538	* unbind is not in progress.
1539	*/
1540	WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
1541	pool->nr_workers == pool->nr_idle &&
1542	atomic_read(get_pool_nr_running(pool)));
1543	}
1544
1545	/**
1546	* worker_leave_idle - leave idle state
1547	* @worker: worker which is leaving idle state
1548	*
1549	* @worker is leaving idle state. Update stats.
1550	*
1551	* LOCKING:
1552	* spin_lock_irq(gcwq->lock).
1553	*/
1554	static void worker_leave_idle(struct worker *worker)
1555	{
1556	struct worker_pool *pool = worker->pool;
1557
1558	BUG_ON(!(worker->flags & WORKER_IDLE));
1559	worker_clr_flags(worker, WORKER_IDLE);
1560	pool->nr_idle--;
1561	list_del_init(&worker->entry);
1562	}
1563
1564	/**
1565	* worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1566	* @worker: self
1567	*
1568	* Works which are scheduled while the cpu is online must at least be
1569	* scheduled to a worker which is bound to the cpu so that if they are
1570	* flushed from cpu callbacks while cpu is going down, they are
1571	* guaranteed to execute on the cpu.
1572	*
1573	* This function is to be used by rogue workers and rescuers to bind
1574	* themselves to the target cpu and may race with cpu going down or
1575	* coming online. kthread_bind() can't be used because it may put the
1576	* worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1577	* verbatim as it's best effort and blocking and gcwq may be
1578	* [dis]associated in the meantime.
1579	*
1580	* This function tries set_cpus_allowed() and locks gcwq and verifies the
1581	* binding against %GCWQ_DISASSOCIATED which is set during
1582	* %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1583	* enters idle state or fetches works without dropping lock, it can
1584	* guarantee the scheduling requirement described in the first paragraph.
1585	*
1586	* CONTEXT:
1587	* Might sleep. Called without any lock but returns with gcwq->lock
1588	* held.
1589	*
1590	* RETURNS:
1591	* %true if the associated gcwq is online (@worker is successfully
1592	* bound), %false if offline.
1593	*/
1594	static bool worker_maybe_bind_and_lock(struct worker *worker)
1595	__acquires(&gcwq->lock)
1596	{
1597	struct global_cwq *gcwq = worker->pool->gcwq;
1598	struct task_struct *task = worker->task;
1599
1600	while (true) {
1601	/*
1602	* The following call may fail, succeed or succeed
1603	* without actually migrating the task to the cpu if
1604	* it races with cpu hotunplug operation. Verify
1605	* against GCWQ_DISASSOCIATED.
1606	*/
1607	if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1608	set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1609
1610	spin_lock_irq(&gcwq->lock);
1611	if (gcwq->flags & GCWQ_DISASSOCIATED)
1612	return false;
1613	if (task_cpu(task) == gcwq->cpu &&
1614	cpumask_equal(&current->cpus_allowed,
1615	get_cpu_mask(gcwq->cpu)))
1616	return true;
1617	spin_unlock_irq(&gcwq->lock);
1618
1619	/*
1620	* We've raced with CPU hot[un]plug. Give it a breather
1621	* and retry migration. cond_resched() is required here;
1622	* otherwise, we might deadlock against cpu_stop trying to
1623	* bring down the CPU on non-preemptive kernel.
1624	*/
1625	cpu_relax();
1626	cond_resched();
1627	}
1628	}
1629
1630	/*
1631	* Rebind an idle @worker to its CPU. worker_thread() will test
1632	* list_empty(@worker->entry) before leaving idle and call this function.
1633	*/
1634	static void idle_worker_rebind(struct worker *worker)
1635	{
1636	struct global_cwq *gcwq = worker->pool->gcwq;
1637
1638	/* CPU may go down again inbetween, clear UNBOUND only on success */
1639	if (worker_maybe_bind_and_lock(worker))
1640	worker_clr_flags(worker, WORKER_UNBOUND);
1641
1642	/* rebind complete, become available again */
1643	list_add(&worker->entry, &worker->pool->idle_list);
1644	spin_unlock_irq(&gcwq->lock);
1645	}
1646
1647	/*
1648	* Function for @worker->rebind.work used to rebind unbound busy workers to
1649	* the associated cpu which is coming back online. This is scheduled by
1650	* cpu up but can race with other cpu hotplug operations and may be
1651	* executed twice without intervening cpu down.
1652	*/
1653	static void busy_worker_rebind_fn(struct work_struct *work)
1654	{
1655	struct worker *worker = container_of(work, struct worker, rebind_work);
1656	struct global_cwq *gcwq = worker->pool->gcwq;
1657
1658	if (worker_maybe_bind_and_lock(worker))
1659	worker_clr_flags(worker, WORKER_UNBOUND);
1660
1661	spin_unlock_irq(&gcwq->lock);
1662	}
1663
1664	/**
1665	* rebind_workers - rebind all workers of a gcwq to the associated CPU
1666	* @gcwq: gcwq of interest
1667	*
1668	* @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1669	* is different for idle and busy ones.
1670	*
1671	* Idle ones will be removed from the idle_list and woken up. They will
1672	* add themselves back after completing rebind. This ensures that the
1673	* idle_list doesn't contain any unbound workers when re-bound busy workers
1674	* try to perform local wake-ups for concurrency management.
1675	*
1676	* Busy workers can rebind after they finish their current work items.
1677	* Queueing the rebind work item at the head of the scheduled list is
1678	* enough. Note that nr_running will be properly bumped as busy workers
1679	* rebind.
1680	*
1681	* On return, all non-manager workers are scheduled for rebind - see
1682	* manage_workers() for the manager special case. Any idle worker
1683	* including the manager will not appear on @idle_list until rebind is
1684	* complete, making local wake-ups safe.
1685	*/
1686	static void rebind_workers(struct global_cwq *gcwq)
1687	{
1688	struct worker_pool *pool;
1689	struct worker worker, n;
1690	struct hlist_node *pos;
1691	int i;
1692
1693	lockdep_assert_held(&gcwq->lock);
1694
1695	for_each_worker_pool(pool, gcwq)
1696	lockdep_assert_held(&pool->assoc_mutex);
1697
1698	/* dequeue and kick idle ones */
1699	for_each_worker_pool(pool, gcwq) {
1700	list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1701	/*
1702	* idle workers should be off @pool->idle_list
1703	* until rebind is complete to avoid receiving
1704	* premature local wake-ups.
1705	*/
1706	list_del_init(&worker->entry);
1707
1708	/*
1709	* worker_thread() will see the above dequeuing
1710	* and call idle_worker_rebind().
1711	*/
1712	wake_up_process(worker->task);
1713	}
1714	}
1715
1716	/* rebind busy workers */
1717	for_each_busy_worker(worker, i, pos, gcwq) {
1718	struct work_struct *rebind_work = &worker->rebind_work;
1719	struct workqueue_struct *wq;
1720
1721	if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1722	work_data_bits(rebind_work)))
1723	continue;
1724
1725	debug_work_activate(rebind_work);
1726
1727	/*
1728	* wq doesn't really matter but let's keep @worker->pool
1729	* and @cwq->pool consistent for sanity.
1730	*/
1731	if (worker_pool_pri(worker->pool))
1732	wq = system_highpri_wq;
1733	else
1734	wq = system_wq;
1735
1736	insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
1737	worker->scheduled.next,
1738	work_color_to_flags(WORK_NO_COLOR));
1739	}
1740	}
1741
1742	static struct worker *alloc_worker(void)
1743	{
1744	struct worker *worker;
1745
1746	worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1747	if (worker) {
1748	INIT_LIST_HEAD(&worker->entry);
1749	INIT_LIST_HEAD(&worker->scheduled);
1750	INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1751	/* on creation a worker is in !idle && prep state */
1752	worker->flags = WORKER_PREP;
1753	}
1754	return worker;
1755	}
1756
1757	/**
1758	* create_worker - create a new workqueue worker
1759	* @pool: pool the new worker will belong to
1760	*
1761	* Create a new worker which is bound to @pool. The returned worker
1762	* can be started by calling start_worker() or destroyed using
1763	* destroy_worker().
1764	*
1765	* CONTEXT:
1766	* Might sleep. Does GFP_KERNEL allocations.
1767	*
1768	* RETURNS:
1769	* Pointer to the newly created worker.
1770	*/
1771	static struct worker create_worker(struct worker_pool pool)
1772	{
1773	struct global_cwq *gcwq = pool->gcwq;
1774	const char *pri = worker_pool_pri(pool) ? "H" : "";
1775	struct worker *worker = NULL;
1776	int id = -1;
1777
1778	spin_lock_irq(&gcwq->lock);
1779	while (ida_get_new(&pool->worker_ida, &id)) {
1780	spin_unlock_irq(&gcwq->lock);
1781	if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1782	goto fail;
1783	spin_lock_irq(&gcwq->lock);
1784	}
1785	spin_unlock_irq(&gcwq->lock);
1786
1787	worker = alloc_worker();
1788	if (!worker)
1789	goto fail;
1790
1791	worker->pool = pool;
1792	worker->id = id;
1793
1794	if (gcwq->cpu != WORK_CPU_UNBOUND)
1795	worker->task = kthread_create_on_node(worker_thread,
1796	worker, cpu_to_node(gcwq->cpu),
1797	"kworker/%u:%d%s", gcwq->cpu, id, pri);
1798	else
1799	worker->task = kthread_create(worker_thread, worker,
1800	"kworker/u:%d%s", id, pri);
1801	if (IS_ERR(worker->task))
1802	goto fail;
1803
1804	if (worker_pool_pri(pool))
1805	set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1806
1807	/*
1808	* Determine CPU binding of the new worker depending on
1809	* %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1810	* flag remains stable across this function. See the comments
1811	* above the flag definition for details.
1812	*
1813	* As an unbound worker may later become a regular one if CPU comes
1814	* online, make sure every worker has %PF_THREAD_BOUND set.
1815	*/
1816	if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
1817	kthread_bind(worker->task, gcwq->cpu);
1818	} else {
1819	worker->task->flags \|= PF_THREAD_BOUND;
1820	worker->flags \|= WORKER_UNBOUND;
1821	}
1822
1823	return worker;
1824	fail:
1825	if (id >= 0) {
1826	spin_lock_irq(&gcwq->lock);
1827	ida_remove(&pool->worker_ida, id);
1828	spin_unlock_irq(&gcwq->lock);
1829	}
1830	kfree(worker);
1831	return NULL;
1832	}
1833
1834	/**
1835	* start_worker - start a newly created worker
1836	* @worker: worker to start
1837	*
1838	* Make the gcwq aware of @worker and start it.
1839	*
1840	* CONTEXT:
1841	* spin_lock_irq(gcwq->lock).
1842	*/
1843	static void start_worker(struct worker *worker)
1844	{
1845	worker->flags \|= WORKER_STARTED;
1846	worker->pool->nr_workers++;
1847	worker_enter_idle(worker);
1848	wake_up_process(worker->task);
1849	}
1850
1851	/**
1852	* destroy_worker - destroy a workqueue worker
1853	* @worker: worker to be destroyed
1854	*
1855	* Destroy @worker and adjust @gcwq stats accordingly.
1856	*
1857	* CONTEXT:
1858	* spin_lock_irq(gcwq->lock) which is released and regrabbed.
1859	*/
1860	static void destroy_worker(struct worker *worker)
1861	{
1862	struct worker_pool *pool = worker->pool;
1863	struct global_cwq *gcwq = pool->gcwq;
1864	int id = worker->id;
1865
1866	/* sanity check frenzy */
1867	BUG_ON(worker->current_work);
1868	BUG_ON(!list_empty(&worker->scheduled));
1869
1870	if (worker->flags & WORKER_STARTED)
1871	pool->nr_workers--;
1872	if (worker->flags & WORKER_IDLE)
1873	pool->nr_idle--;
1874
1875	list_del_init(&worker->entry);
1876	worker->flags \|= WORKER_DIE;
1877
1878	spin_unlock_irq(&gcwq->lock);
1879
1880	kthread_stop(worker->task);
1881	kfree(worker);
1882
1883	spin_lock_irq(&gcwq->lock);
1884	ida_remove(&pool->worker_ida, id);
1885	}
1886
1887	static void idle_worker_timeout(unsigned long __pool)
1888	{
1889	struct worker_pool pool = (void )__pool;
1890	struct global_cwq *gcwq = pool->gcwq;
1891
1892	spin_lock_irq(&gcwq->lock);
1893
1894	if (too_many_workers(pool)) {
1895	struct worker *worker;
1896	unsigned long expires;
1897
1898	/* idle_list is kept in LIFO order, check the last one */
1899	worker = list_entry(pool->idle_list.prev, struct worker, entry);
1900	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1901
1902	if (time_before(jiffies, expires))
1903	mod_timer(&pool->idle_timer, expires);
1904	else {
1905	/* it's been idle for too long, wake up manager */
1906	pool->flags \|= POOL_MANAGE_WORKERS;
1907	wake_up_worker(pool);
1908	}
1909	}
1910
1911	spin_unlock_irq(&gcwq->lock);
1912	}
1913
1914	static bool send_mayday(struct work_struct *work)
1915	{
1916	struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1917	struct workqueue_struct *wq = cwq->wq;
1918	unsigned int cpu;
1919
1920	if (!(wq->flags & WQ_RESCUER))
1921	return false;
1922
1923	/* mayday mayday mayday */
1924	cpu = cwq->pool->gcwq->cpu;
1925	/* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1926	if (cpu == WORK_CPU_UNBOUND)
1927	cpu = 0;
1928	if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1929	wake_up_process(wq->rescuer->task);
1930	return true;
1931	}
1932
1933	static void gcwq_mayday_timeout(unsigned long __pool)
1934	{
1935	struct worker_pool pool = (void )__pool;
1936	struct global_cwq *gcwq = pool->gcwq;
1937	struct work_struct *work;
1938
1939	spin_lock_irq(&gcwq->lock);
1940
1941	if (need_to_create_worker(pool)) {
1942	/*
1943	* We've been trying to create a new worker but
1944	* haven't been successful. We might be hitting an
1945	* allocation deadlock. Send distress signals to
1946	* rescuers.
1947	*/
1948	list_for_each_entry(work, &pool->worklist, entry)
1949	send_mayday(work);
1950	}
1951
1952	spin_unlock_irq(&gcwq->lock);
1953
1954	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1955	}
1956
1957	/**
1958	* maybe_create_worker - create a new worker if necessary
1959	* @pool: pool to create a new worker for
1960	*
1961	* Create a new worker for @pool if necessary. @pool is guaranteed to
1962	* have at least one idle worker on return from this function. If
1963	* creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1964	* sent to all rescuers with works scheduled on @pool to resolve
1965	* possible allocation deadlock.
1966	*
1967	* On return, need_to_create_worker() is guaranteed to be false and
1968	* may_start_working() true.
1969	*
1970	* LOCKING:
1971	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
1972	* multiple times. Does GFP_KERNEL allocations. Called only from
1973	* manager.
1974	*
1975	* RETURNS:
1976	* false if no action was taken and gcwq->lock stayed locked, true
1977	* otherwise.
1978	*/
1979	static bool maybe_create_worker(struct worker_pool *pool)
1980	__releases(&gcwq->lock)
1981	__acquires(&gcwq->lock)
1982	{
1983	struct global_cwq *gcwq = pool->gcwq;
1984
1985	if (!need_to_create_worker(pool))
1986	return false;
1987	restart:
1988	spin_unlock_irq(&gcwq->lock);
1989
1990	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1991	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1992
1993	while (true) {
1994	struct worker *worker;
1995
1996	worker = create_worker(pool);
1997	if (worker) {
1998	del_timer_sync(&pool->mayday_timer);
1999	spin_lock_irq(&gcwq->lock);
2000	start_worker(worker);
2001	BUG_ON(need_to_create_worker(pool));
2002	return true;
2003	}
2004
2005	if (!need_to_create_worker(pool))
2006	break;
2007
2008	__set_current_state(TASK_INTERRUPTIBLE);
2009	schedule_timeout(CREATE_COOLDOWN);
2010
2011	if (!need_to_create_worker(pool))
2012	break;
2013	}
2014
2015	del_timer_sync(&pool->mayday_timer);
2016	spin_lock_irq(&gcwq->lock);
2017	if (need_to_create_worker(pool))
2018	goto restart;
2019	return true;
2020	}
2021
2022	/**
2023	* maybe_destroy_worker - destroy workers which have been idle for a while
2024	* @pool: pool to destroy workers for
2025	*
2026	* Destroy @pool workers which have been idle for longer than
2027	* IDLE_WORKER_TIMEOUT.
2028	*
2029	* LOCKING:
2030	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
2031	* multiple times. Called only from manager.
2032	*
2033	* RETURNS:
2034	* false if no action was taken and gcwq->lock stayed locked, true
2035	* otherwise.
2036	*/
2037	static bool maybe_destroy_workers(struct worker_pool *pool)
2038	{
2039	bool ret = false;
2040
2041	while (too_many_workers(pool)) {
2042	struct worker *worker;
2043	unsigned long expires;
2044
2045	worker = list_entry(pool->idle_list.prev, struct worker, entry);
2046	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2047
2048	if (time_before(jiffies, expires)) {
2049	mod_timer(&pool->idle_timer, expires);
2050	break;
2051	}
2052
2053	destroy_worker(worker);
2054	ret = true;
2055	}
2056
2057	return ret;
2058	}
2059
2060	/**
2061	* manage_workers - manage worker pool
2062	* @worker: self
2063	*
2064	* Assume the manager role and manage gcwq worker pool @worker belongs
2065	* to. At any given time, there can be only zero or one manager per
2066	* gcwq. The exclusion is handled automatically by this function.
2067	*
2068	* The caller can safely start processing works on false return. On
2069	* true return, it's guaranteed that need_to_create_worker() is false
2070	* and may_start_working() is true.
2071	*
2072	* CONTEXT:
2073	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
2074	* multiple times. Does GFP_KERNEL allocations.
2075	*
2076	* RETURNS:
2077	* false if no action was taken and gcwq->lock stayed locked, true if
2078	* some action was taken.
2079	*/
2080	static bool manage_workers(struct worker *worker)
2081	{
2082	struct worker_pool *pool = worker->pool;
2083	bool ret = false;
2084
2085	if (pool->flags & POOL_MANAGING_WORKERS)
2086	return ret;
2087
2088	pool->flags \|= POOL_MANAGING_WORKERS;
2089
2090	/*
2091	* To simplify both worker management and CPU hotplug, hold off
2092	* management while hotplug is in progress. CPU hotplug path can't
2093	* grab %POOL_MANAGING_WORKERS to achieve this because that can
2094	* lead to idle worker depletion (all become busy thinking someone
2095	* else is managing) which in turn can result in deadlock under
2096	* extreme circumstances. Use @pool->assoc_mutex to synchronize
2097	* manager against CPU hotplug.
2098	*
2099	* assoc_mutex would always be free unless CPU hotplug is in
2100	* progress. trylock first without dropping @gcwq->lock.
2101	*/
2102	if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2103	spin_unlock_irq(&pool->gcwq->lock);
2104	mutex_lock(&pool->assoc_mutex);
2105	/*
2106	* CPU hotplug could have happened while we were waiting
2107	* for assoc_mutex. Hotplug itself can't handle us
2108	* because manager isn't either on idle or busy list, and
2109	* @gcwq's state and ours could have deviated.
2110	*
2111	* As hotplug is now excluded via assoc_mutex, we can
2112	* simply try to bind. It will succeed or fail depending
2113	* on @gcwq's current state. Try it and adjust
2114	* %WORKER_UNBOUND accordingly.
2115	*/
2116	if (worker_maybe_bind_and_lock(worker))
2117	worker->flags &= ~WORKER_UNBOUND;
2118	else
2119	worker->flags \|= WORKER_UNBOUND;
2120
2121	ret = true;
2122	}
2123
2124	pool->flags &= ~POOL_MANAGE_WORKERS;
2125
2126	/*
2127	* Destroy and then create so that may_start_working() is true
2128	* on return.
2129	*/
2130	ret \|= maybe_destroy_workers(pool);
2131	ret \|= maybe_create_worker(pool);
2132
2133	pool->flags &= ~POOL_MANAGING_WORKERS;
2134	mutex_unlock(&pool->assoc_mutex);
2135	return ret;
2136	}
2137
2138	/**
2139	* process_one_work - process single work
2140	* @worker: self
2141	* @work: work to process
2142	*
2143	* Process @work. This function contains all the logics necessary to
2144	* process a single work including synchronization against and
2145	* interaction with other workers on the same cpu, queueing and
2146	* flushing. As long as context requirement is met, any worker can
2147	* call this function to process a work.
2148	*
2149	* CONTEXT:
2150	* spin_lock_irq(gcwq->lock) which is released and regrabbed.
2151	*/
2152	static void process_one_work(struct worker worker, struct work_struct work)
2153	__releases(&gcwq->lock)
2154	__acquires(&gcwq->lock)
2155	{
2156	struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2157	struct worker_pool *pool = worker->pool;
2158	struct global_cwq *gcwq = pool->gcwq;
2159	struct hlist_head *bwh = busy_worker_head(gcwq, work);
2160	bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2161	work_func_t f = work->func;
2162	int work_color;
2163	struct worker *collision;
2164	#ifdef CONFIG_LOCKDEP
2165	/*
2166	* It is permissible to free the struct work_struct from
2167	* inside the function that is called from it, this we need to
2168	* take into account for lockdep too. To avoid bogus "held
2169	* lock freed" warnings as well as problems when looking into
2170	* work->lockdep_map, make a copy and use that here.
2171	*/
2172	struct lockdep_map lockdep_map;
2173
2174	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2175	#endif
2176	/*
2177	* Ensure we're on the correct CPU. DISASSOCIATED test is
2178	* necessary to avoid spurious warnings from rescuers servicing the
2179	* unbound or a disassociated gcwq.
2180	*/
2181	WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2182	!(gcwq->flags & GCWQ_DISASSOCIATED) &&
2183	raw_smp_processor_id() != gcwq->cpu);
2184
2185	/*
2186	* A single work shouldn't be executed concurrently by
2187	* multiple workers on a single cpu. Check whether anyone is
2188	* already processing the work. If so, defer the work to the
2189	* currently executing one.
2190	*/
2191	collision = __find_worker_executing_work(gcwq, bwh, work);
2192	if (unlikely(collision)) {
2193	move_linked_works(work, &collision->scheduled, NULL);
2194	return;
2195	}
2196
2197	/* claim and dequeue */
2198	debug_work_deactivate(work);
2199	hlist_add_head(&worker->hentry, bwh);
2200	worker->current_work = work;
2201	worker->current_cwq = cwq;
2202	work_color = get_work_color(work);
2203
2204	list_del_init(&work->entry);
2205
2206	/*
2207	* CPU intensive works don't participate in concurrency
2208	* management. They're the scheduler's responsibility.
2209	*/
2210	if (unlikely(cpu_intensive))
2211	worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2212
2213	/*
2214	* Unbound gcwq isn't concurrency managed and work items should be
2215	* executed ASAP. Wake up another worker if necessary.
2216	*/
2217	if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2218	wake_up_worker(pool);
2219
2220	/*
2221	* Record the last CPU and clear PENDING which should be the last
2222	* update to @work. Also, do this inside @gcwq->lock so that
2223	* PENDING and queued state changes happen together while IRQ is
2224	* disabled.
2225	*/
2226	set_work_cpu_and_clear_pending(work, gcwq->cpu);
2227
2228	spin_unlock_irq(&gcwq->lock);
2229
2230	lock_map_acquire_read(&cwq->wq->lockdep_map);
2231	lock_map_acquire(&lockdep_map);
2232	trace_workqueue_execute_start(work);
2233	f(work);
2234	/*
2235	* While we must be careful to not use "work" after this, the trace
2236	* point will only record its address.
2237	*/
2238	trace_workqueue_execute_end(work);
2239	lock_map_release(&lockdep_map);
2240	lock_map_release(&cwq->wq->lockdep_map);
2241
2242	if (unlikely(in_atomic() \|\| lockdep_depth(current) > 0)) {
2243	pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2244	" last function: %pf\n",
2245	current->comm, preempt_count(), task_pid_nr(current), f);
2246	debug_show_held_locks(current);
2247	dump_stack();
2248	}
2249
2250	spin_lock_irq(&gcwq->lock);
2251
2252	/* clear cpu intensive status */
2253	if (unlikely(cpu_intensive))
2254	worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2255
2256	/* we're done with it, release */
2257	hlist_del_init(&worker->hentry);
2258	worker->current_work = NULL;
2259	worker->current_cwq = NULL;
2260	cwq_dec_nr_in_flight(cwq, work_color);
2261	}
2262
2263	/**
2264	* process_scheduled_works - process scheduled works
2265	* @worker: self
2266	*
2267	* Process all scheduled works. Please note that the scheduled list
2268	* may change while processing a work, so this function repeatedly
2269	* fetches a work from the top and executes it.
2270	*
2271	* CONTEXT:
2272	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
2273	* multiple times.
2274	*/
2275	static void process_scheduled_works(struct worker *worker)
2276	{
2277	while (!list_empty(&worker->scheduled)) {
2278	struct work_struct *work = list_first_entry(&worker->scheduled,
2279	struct work_struct, entry);
2280	process_one_work(worker, work);
2281	}
2282	}
2283
2284	/**
2285	* worker_thread - the worker thread function
2286	* @__worker: self
2287	*
2288	* The gcwq worker thread function. There's a single dynamic pool of
2289	* these per each cpu. These workers process all works regardless of
2290	* their specific target workqueue. The only exception is works which
2291	* belong to workqueues with a rescuer which will be explained in
2292	* rescuer_thread().
2293	*/
2294	static int worker_thread(void *__worker)
2295	{
2296	struct worker *worker = __worker;
2297	struct worker_pool *pool = worker->pool;
2298	struct global_cwq *gcwq = pool->gcwq;
2299
2300	/* tell the scheduler that this is a workqueue worker */
2301	worker->task->flags \|= PF_WQ_WORKER;
2302	woke_up:
2303	spin_lock_irq(&gcwq->lock);
2304
2305	/* we are off idle list if destruction or rebind is requested */
2306	if (unlikely(list_empty(&worker->entry))) {
2307	spin_unlock_irq(&gcwq->lock);
2308
2309	/* if DIE is set, destruction is requested */
2310	if (worker->flags & WORKER_DIE) {
2311	worker->task->flags &= ~PF_WQ_WORKER;
2312	return 0;
2313	}
2314
2315	/* otherwise, rebind */
2316	idle_worker_rebind(worker);
2317	goto woke_up;
2318	}
2319
2320	worker_leave_idle(worker);
2321	recheck:
2322	/* no more worker necessary? */
2323	if (!need_more_worker(pool))
2324	goto sleep;
2325
2326	/* do we need to manage? */
2327	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2328	goto recheck;
2329
2330	/*
2331	* ->scheduled list can only be filled while a worker is
2332	* preparing to process a work or actually processing it.
2333	* Make sure nobody diddled with it while I was sleeping.
2334	*/
2335	BUG_ON(!list_empty(&worker->scheduled));
2336
2337	/*
2338	* When control reaches this point, we're guaranteed to have
2339	* at least one idle worker or that someone else has already
2340	* assumed the manager role.
2341	*/
2342	worker_clr_flags(worker, WORKER_PREP);
2343
2344	do {
2345	struct work_struct *work =
2346	list_first_entry(&pool->worklist,
2347	struct work_struct, entry);
2348
2349	if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2350	/* optimization path, not strictly necessary */
2351	process_one_work(worker, work);
2352	if (unlikely(!list_empty(&worker->scheduled)))
2353	process_scheduled_works(worker);
2354	} else {
2355	move_linked_works(work, &worker->scheduled, NULL);
2356	process_scheduled_works(worker);
2357	}
2358	} while (keep_working(pool));
2359
2360	worker_set_flags(worker, WORKER_PREP, false);
2361	sleep:
2362	if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2363	goto recheck;
2364
2365	/*
2366	* gcwq->lock is held and there's no work to process and no
2367	* need to manage, sleep. Workers are woken up only while
2368	* holding gcwq->lock or from local cpu, so setting the
2369	* current state before releasing gcwq->lock is enough to
2370	* prevent losing any event.
2371	*/
2372	worker_enter_idle(worker);
2373	__set_current_state(TASK_INTERRUPTIBLE);
2374	spin_unlock_irq(&gcwq->lock);
2375	schedule();
2376	goto woke_up;
2377	}
2378
2379	/**
2380	* rescuer_thread - the rescuer thread function
2381	* @__wq: the associated workqueue
2382	*
2383	* Workqueue rescuer thread function. There's one rescuer for each
2384	* workqueue which has WQ_RESCUER set.
2385	*
2386	* Regular work processing on a gcwq may block trying to create a new
2387	* worker which uses GFP_KERNEL allocation which has slight chance of
2388	* developing into deadlock if some works currently on the same queue
2389	* need to be processed to satisfy the GFP_KERNEL allocation. This is
2390	* the problem rescuer solves.
2391	*
2392	* When such condition is possible, the gcwq summons rescuers of all
2393	* workqueues which have works queued on the gcwq and let them process
2394	* those works so that forward progress can be guaranteed.
2395	*
2396	* This should happen rarely.
2397	*/
2398	static int rescuer_thread(void *__wq)
2399	{
2400	struct workqueue_struct *wq = __wq;
2401	struct worker *rescuer = wq->rescuer;
2402	struct list_head *scheduled = &rescuer->scheduled;
2403	bool is_unbound = wq->flags & WQ_UNBOUND;
2404	unsigned int cpu;
2405
2406	set_user_nice(current, RESCUER_NICE_LEVEL);
2407	repeat:
2408	set_current_state(TASK_INTERRUPTIBLE);
2409
2410	if (kthread_should_stop())
2411	return 0;
2412
2413	/*
2414	* See whether any cpu is asking for help. Unbounded
2415	* workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2416	*/
2417	for_each_mayday_cpu(cpu, wq->mayday_mask) {
2418	unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2419	struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2420	struct worker_pool *pool = cwq->pool;
2421	struct global_cwq *gcwq = pool->gcwq;
2422	struct work_struct work, n;
2423
2424	__set_current_state(TASK_RUNNING);
2425	mayday_clear_cpu(cpu, wq->mayday_mask);
2426
2427	/* migrate to the target cpu if possible */
2428	rescuer->pool = pool;
2429	worker_maybe_bind_and_lock(rescuer);
2430
2431	/*
2432	* Slurp in all works issued via this workqueue and
2433	* process'em.
2434	*/
2435	BUG_ON(!list_empty(&rescuer->scheduled));
2436	list_for_each_entry_safe(work, n, &pool->worklist, entry)
2437	if (get_work_cwq(work) == cwq)
2438	move_linked_works(work, scheduled, &n);
2439
2440	process_scheduled_works(rescuer);
2441
2442	/*
2443	* Leave this gcwq. If keep_working() is %true, notify a
2444	* regular worker; otherwise, we end up with 0 concurrency
2445	* and stalling the execution.
2446	*/
2447	if (keep_working(pool))
2448	wake_up_worker(pool);
2449
2450	spin_unlock_irq(&gcwq->lock);
2451	}
2452
2453	schedule();
2454	goto repeat;
2455	}
2456
2457	struct wq_barrier {
2458	struct work_struct work;
2459	struct completion done;
2460	};
2461
2462	static void wq_barrier_func(struct work_struct *work)
2463	{
2464	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2465	complete(&barr->done);
2466	}
2467
2468	/**
2469	* insert_wq_barrier - insert a barrier work
2470	* @cwq: cwq to insert barrier into
2471	* @barr: wq_barrier to insert
2472	* @target: target work to attach @barr to
2473	* @worker: worker currently executing @target, NULL if @target is not executing
2474	*
2475	* @barr is linked to @target such that @barr is completed only after
2476	* @target finishes execution. Please note that the ordering
2477	* guarantee is observed only with respect to @target and on the local
2478	* cpu.
2479	*
2480	* Currently, a queued barrier can't be canceled. This is because
2481	* try_to_grab_pending() can't determine whether the work to be
2482	* grabbed is at the head of the queue and thus can't clear LINKED
2483	* flag of the previous work while there must be a valid next work
2484	* after a work with LINKED flag set.
2485	*
2486	* Note that when @worker is non-NULL, @target may be modified
2487	* underneath us, so we can't reliably determine cwq from @target.
2488	*
2489	* CONTEXT:
2490	* spin_lock_irq(gcwq->lock).
2491	*/
2492	static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2493	struct wq_barrier *barr,
2494	struct work_struct target, struct worker worker)
2495	{
2496	struct list_head *head;
2497	unsigned int linked = 0;
2498
2499	/*
2500	* debugobject calls are safe here even with gcwq->lock locked
2501	* as we know for sure that this will not trigger any of the
2502	* checks and call back into the fixup functions where we
2503	* might deadlock.
2504	*/
2505	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2506	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2507	init_completion(&barr->done);
2508
2509	/*
2510	* If @target is currently being executed, schedule the
2511	* barrier to the worker; otherwise, put it after @target.
2512	*/
2513	if (worker)
2514	head = worker->scheduled.next;
2515	else {
2516	unsigned long *bits = work_data_bits(target);
2517
2518	head = target->entry.next;
2519	/* there can already be other linked works, inherit and set */
2520	linked = *bits & WORK_STRUCT_LINKED;
2521	__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2522	}
2523
2524	debug_work_activate(&barr->work);
2525	insert_work(cwq, &barr->work, head,
2526	work_color_to_flags(WORK_NO_COLOR) \| linked);
2527	}
2528
2529	/**
2530	* flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2531	* @wq: workqueue being flushed
2532	* @flush_color: new flush color, < 0 for no-op
2533	* @work_color: new work color, < 0 for no-op
2534	*
2535	* Prepare cwqs for workqueue flushing.
2536	*
2537	* If @flush_color is non-negative, flush_color on all cwqs should be
2538	* -1. If no cwq has in-flight commands at the specified color, all
2539	* cwq->flush_color's stay at -1 and %false is returned. If any cwq
2540	* has in flight commands, its cwq->flush_color is set to
2541	* @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2542	* wakeup logic is armed and %true is returned.
2543	*
2544	* The caller should have initialized @wq->first_flusher prior to
2545	* calling this function with non-negative @flush_color. If
2546	* @flush_color is negative, no flush color update is done and %false
2547	* is returned.
2548	*
2549	* If @work_color is non-negative, all cwqs should have the same
2550	* work_color which is previous to @work_color and all will be
2551	* advanced to @work_color.
2552	*
2553	* CONTEXT:
2554	* mutex_lock(wq->flush_mutex).
2555	*
2556	* RETURNS:
2557	* %true if @flush_color >= 0 and there's something to flush. %false
2558	* otherwise.
2559	*/
2560	static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2561	int flush_color, int work_color)
2562	{
2563	bool wait = false;
2564	unsigned int cpu;
2565
2566	if (flush_color >= 0) {
2567	BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2568	atomic_set(&wq->nr_cwqs_to_flush, 1);
2569	}
2570
2571	for_each_cwq_cpu(cpu, wq) {
2572	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2573	struct global_cwq *gcwq = cwq->pool->gcwq;
2574
2575	spin_lock_irq(&gcwq->lock);
2576
2577	if (flush_color >= 0) {
2578	BUG_ON(cwq->flush_color != -1);
2579
2580	if (cwq->nr_in_flight[flush_color]) {
2581	cwq->flush_color = flush_color;
2582	atomic_inc(&wq->nr_cwqs_to_flush);
2583	wait = true;
2584	}
2585	}
2586
2587	if (work_color >= 0) {
2588	BUG_ON(work_color != work_next_color(cwq->work_color));
2589	cwq->work_color = work_color;
2590	}
2591
2592	spin_unlock_irq(&gcwq->lock);
2593	}
2594
2595	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2596	complete(&wq->first_flusher->done);
2597
2598	return wait;
2599	}
2600
2601	/**
2602	* flush_workqueue - ensure that any scheduled work has run to completion.
2603	* @wq: workqueue to flush
2604	*
2605	* Forces execution of the workqueue and blocks until its completion.
2606	* This is typically used in driver shutdown handlers.
2607	*
2608	* We sleep until all works which were queued on entry have been handled,
2609	* but we are not livelocked by new incoming ones.
2610	*/
2611	void flush_workqueue(struct workqueue_struct *wq)
2612	{
2613	struct wq_flusher this_flusher = {
2614	.list = LIST_HEAD_INIT(this_flusher.list),
2615	.flush_color = -1,
2616	.done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2617	};
2618	int next_color;
2619
2620	lock_map_acquire(&wq->lockdep_map);
2621	lock_map_release(&wq->lockdep_map);
2622
2623	mutex_lock(&wq->flush_mutex);
2624
2625	/*
2626	* Start-to-wait phase
2627	*/
2628	next_color = work_next_color(wq->work_color);
2629
2630	if (next_color != wq->flush_color) {
2631	/*
2632	* Color space is not full. The current work_color
2633	* becomes our flush_color and work_color is advanced
2634	* by one.
2635	*/
2636	BUG_ON(!list_empty(&wq->flusher_overflow));
2637	this_flusher.flush_color = wq->work_color;
2638	wq->work_color = next_color;
2639
2640	if (!wq->first_flusher) {
2641	/* no flush in progress, become the first flusher */
2642	BUG_ON(wq->flush_color != this_flusher.flush_color);
2643
2644	wq->first_flusher = &this_flusher;
2645
2646	if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2647	wq->work_color)) {
2648	/* nothing to flush, done */
2649	wq->flush_color = next_color;
2650	wq->first_flusher = NULL;
2651	goto out_unlock;
2652	}
2653	} else {
2654	/* wait in queue */
2655	BUG_ON(wq->flush_color == this_flusher.flush_color);
2656	list_add_tail(&this_flusher.list, &wq->flusher_queue);
2657	flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2658	}
2659	} else {
2660	/*
2661	* Oops, color space is full, wait on overflow queue.
2662	* The next flush completion will assign us
2663	* flush_color and transfer to flusher_queue.
2664	*/
2665	list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2666	}
2667
2668	mutex_unlock(&wq->flush_mutex);
2669
2670	wait_for_completion(&this_flusher.done);
2671
2672	/*
2673	* Wake-up-and-cascade phase
2674	*
2675	* First flushers are responsible for cascading flushes and
2676	* handling overflow. Non-first flushers can simply return.
2677	*/
2678	if (wq->first_flusher != &this_flusher)
2679	return;
2680
2681	mutex_lock(&wq->flush_mutex);
2682
2683	/* we might have raced, check again with mutex held */
2684	if (wq->first_flusher != &this_flusher)
2685	goto out_unlock;
2686
2687	wq->first_flusher = NULL;
2688
2689	BUG_ON(!list_empty(&this_flusher.list));
2690	BUG_ON(wq->flush_color != this_flusher.flush_color);
2691
2692	while (true) {
2693	struct wq_flusher next, tmp;
2694
2695	/* complete all the flushers sharing the current flush color */
2696	list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2697	if (next->flush_color != wq->flush_color)
2698	break;
2699	list_del_init(&next->list);
2700	complete(&next->done);
2701	}
2702
2703	BUG_ON(!list_empty(&wq->flusher_overflow) &&
2704	wq->flush_color != work_next_color(wq->work_color));
2705
2706	/* this flush_color is finished, advance by one */
2707	wq->flush_color = work_next_color(wq->flush_color);
2708
2709	/* one color has been freed, handle overflow queue */
2710	if (!list_empty(&wq->flusher_overflow)) {
2711	/*
2712	* Assign the same color to all overflowed
2713	* flushers, advance work_color and append to
2714	* flusher_queue. This is the start-to-wait
2715	* phase for these overflowed flushers.
2716	*/
2717	list_for_each_entry(tmp, &wq->flusher_overflow, list)
2718	tmp->flush_color = wq->work_color;
2719
2720	wq->work_color = work_next_color(wq->work_color);
2721
2722	list_splice_tail_init(&wq->flusher_overflow,
2723	&wq->flusher_queue);
2724	flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2725	}
2726
2727	if (list_empty(&wq->flusher_queue)) {
2728	BUG_ON(wq->flush_color != wq->work_color);
2729	break;
2730	}
2731
2732	/*
2733	* Need to flush more colors. Make the next flusher
2734	* the new first flusher and arm cwqs.
2735	*/
2736	BUG_ON(wq->flush_color == wq->work_color);
2737	BUG_ON(wq->flush_color != next->flush_color);
2738
2739	list_del_init(&next->list);
2740	wq->first_flusher = next;
2741
2742	if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2743	break;
2744
2745	/*
2746	* Meh... this color is already done, clear first
2747	* flusher and repeat cascading.
2748	*/
2749	wq->first_flusher = NULL;
2750	}
2751
2752	out_unlock:
2753	mutex_unlock(&wq->flush_mutex);
2754	}
2755	EXPORT_SYMBOL_GPL(flush_workqueue);
2756
2757	/**
2758	* drain_workqueue - drain a workqueue
2759	* @wq: workqueue to drain
2760	*
2761	* Wait until the workqueue becomes empty. While draining is in progress,
2762	* only chain queueing is allowed. IOW, only currently pending or running
2763	* work items on @wq can queue further work items on it. @wq is flushed
2764	* repeatedly until it becomes empty. The number of flushing is detemined
2765	* by the depth of chaining and should be relatively short. Whine if it
2766	* takes too long.
2767	*/
2768	void drain_workqueue(struct workqueue_struct *wq)
2769	{
2770	unsigned int flush_cnt = 0;
2771	unsigned int cpu;
2772
2773	/*
2774	* __queue_work() needs to test whether there are drainers, is much
2775	* hotter than drain_workqueue() and already looks at @wq->flags.
2776	* Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2777	*/
2778	spin_lock(&workqueue_lock);
2779	if (!wq->nr_drainers++)
2780	wq->flags \|= WQ_DRAINING;
2781	spin_unlock(&workqueue_lock);
2782	reflush:
2783	flush_workqueue(wq);
2784
2785	for_each_cwq_cpu(cpu, wq) {
2786	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2787	bool drained;
2788
2789	spin_lock_irq(&cwq->pool->gcwq->lock);
2790	drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2791	spin_unlock_irq(&cwq->pool->gcwq->lock);
2792
2793	if (drained)
2794	continue;
2795
2796	if (++flush_cnt == 10 \|\|
2797	(flush_cnt % 100 == 0 && flush_cnt <= 1000))
2798	pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2799	wq->name, flush_cnt);
2800	goto reflush;
2801	}
2802
2803	spin_lock(&workqueue_lock);
2804	if (!--wq->nr_drainers)
2805	wq->flags &= ~WQ_DRAINING;
2806	spin_unlock(&workqueue_lock);
2807	}
2808	EXPORT_SYMBOL_GPL(drain_workqueue);
2809
2810	static bool start_flush_work(struct work_struct work, struct wq_barrier barr)
2811	{
2812	struct worker *worker = NULL;
2813	struct global_cwq *gcwq;
2814	struct cpu_workqueue_struct *cwq;
2815
2816	might_sleep();
2817	gcwq = get_work_gcwq(work);
2818	if (!gcwq)
2819	return false;
2820
2821	spin_lock_irq(&gcwq->lock);
2822	if (!list_empty(&work->entry)) {
2823	/*
2824	* See the comment near try_to_grab_pending()->smp_rmb().
2825	* If it was re-queued to a different gcwq under us, we
2826	* are not going to wait.
2827	*/
2828	smp_rmb();
2829	cwq = get_work_cwq(work);
2830	if (unlikely(!cwq \|\| gcwq != cwq->pool->gcwq))
2831	goto already_gone;
2832	} else {
2833	worker = find_worker_executing_work(gcwq, work);
2834	if (!worker)
2835	goto already_gone;
2836	cwq = worker->current_cwq;
2837	}
2838
2839	insert_wq_barrier(cwq, barr, work, worker);
2840	spin_unlock_irq(&gcwq->lock);
2841
2842	/*
2843	* If @max_active is 1 or rescuer is in use, flushing another work
2844	* item on the same workqueue may lead to deadlock. Make sure the
2845	* flusher is not running on the same workqueue by verifying write
2846	* access.
2847	*/
2848	if (cwq->wq->saved_max_active == 1 \|\| cwq->wq->flags & WQ_RESCUER)
2849	lock_map_acquire(&cwq->wq->lockdep_map);
2850	else
2851	lock_map_acquire_read(&cwq->wq->lockdep_map);
2852	lock_map_release(&cwq->wq->lockdep_map);
2853
2854	return true;
2855	already_gone:
2856	spin_unlock_irq(&gcwq->lock);
2857	return false;
2858	}
2859
2860	/**
2861	* flush_work - wait for a work to finish executing the last queueing instance
2862	* @work: the work to flush
2863	*
2864	* Wait until @work has finished execution. @work is guaranteed to be idle
2865	* on return if it hasn't been requeued since flush started.
2866	*
2867	* RETURNS:
2868	* %true if flush_work() waited for the work to finish execution,
2869	* %false if it was already idle.
2870	*/
2871	bool flush_work(struct work_struct *work)
2872	{
2873	struct wq_barrier barr;
2874
2875	lock_map_acquire(&work->lockdep_map);
2876	lock_map_release(&work->lockdep_map);
2877
2878	if (start_flush_work(work, &barr)) {
2879	wait_for_completion(&barr.done);
2880	destroy_work_on_stack(&barr.work);
2881	return true;
2882	} else {
2883	return false;
2884	}
2885	}
2886	EXPORT_SYMBOL_GPL(flush_work);
2887
2888	static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2889	{
2890	unsigned long flags;
2891	int ret;
2892
2893	do {
2894	ret = try_to_grab_pending(work, is_dwork, &flags);
2895	/*
2896	* If someone else is canceling, wait for the same event it
2897	* would be waiting for before retrying.
2898	*/
2899	if (unlikely(ret == -ENOENT))
2900	flush_work(work);
2901	} while (unlikely(ret < 0));
2902
2903	/* tell other tasks trying to grab @work to back off */
2904	mark_work_canceling(work);
2905	local_irq_restore(flags);
2906
2907	flush_work(work);
2908	clear_work_data(work);
2909	return ret;
2910	}
2911
2912	/**
2913	* cancel_work_sync - cancel a work and wait for it to finish
2914	* @work: the work to cancel
2915	*
2916	* Cancel @work and wait for its execution to finish. This function
2917	* can be used even if the work re-queues itself or migrates to
2918	* another workqueue. On return from this function, @work is
2919	* guaranteed to be not pending or executing on any CPU.
2920	*
2921	* cancel_work_sync(&delayed_work->work) must not be used for
2922	* delayed_work's. Use cancel_delayed_work_sync() instead.
2923	*
2924	* The caller must ensure that the workqueue on which @work was last
2925	* queued can't be destroyed before this function returns.
2926	*
2927	* RETURNS:
2928	* %true if @work was pending, %false otherwise.
2929	*/
2930	bool cancel_work_sync(struct work_struct *work)
2931	{
2932	return __cancel_work_timer(work, false);
2933	}
2934	EXPORT_SYMBOL_GPL(cancel_work_sync);
2935
2936	/**
2937	* flush_delayed_work - wait for a dwork to finish executing the last queueing
2938	* @dwork: the delayed work to flush
2939	*
2940	* Delayed timer is cancelled and the pending work is queued for
2941	* immediate execution. Like flush_work(), this function only
2942	* considers the last queueing instance of @dwork.
2943	*
2944	* RETURNS:
2945	* %true if flush_work() waited for the work to finish execution,
2946	* %false if it was already idle.
2947	*/
2948	bool flush_delayed_work(struct delayed_work *dwork)
2949	{
2950	local_irq_disable();
2951	if (del_timer_sync(&dwork->timer))
2952	__queue_work(dwork->cpu,
2953	get_work_cwq(&dwork->work)->wq, &dwork->work);
2954	local_irq_enable();
2955	return flush_work(&dwork->work);
2956	}
2957	EXPORT_SYMBOL(flush_delayed_work);
2958
2959	/**
2960	* cancel_delayed_work - cancel a delayed work
2961	* @dwork: delayed_work to cancel
2962	*
2963	* Kill off a pending delayed_work. Returns %true if @dwork was pending
2964	* and canceled; %false if wasn't pending. Note that the work callback
2965	* function may still be running on return, unless it returns %true and the
2966	* work doesn't re-arm itself. Explicitly flush or use
2967	* cancel_delayed_work_sync() to wait on it.
2968	*
2969	* This function is safe to call from any context including IRQ handler.
2970	*/
2971	bool cancel_delayed_work(struct delayed_work *dwork)
2972	{
2973	unsigned long flags;
2974	int ret;
2975
2976	do {
2977	ret = try_to_grab_pending(&dwork->work, true, &flags);
2978	} while (unlikely(ret == -EAGAIN));
2979
2980	if (unlikely(ret < 0))
2981	return false;
2982
2983	set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work));
2984	local_irq_restore(flags);
2985	return ret;
2986	}
2987	EXPORT_SYMBOL(cancel_delayed_work);
2988
2989	/**
2990	* cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2991	* @dwork: the delayed work cancel
2992	*
2993	* This is cancel_work_sync() for delayed works.
2994	*
2995	* RETURNS:
2996	* %true if @dwork was pending, %false otherwise.
2997	*/
2998	bool cancel_delayed_work_sync(struct delayed_work *dwork)
2999	{
3000	return __cancel_work_timer(&dwork->work, true);
3001	}
3002	EXPORT_SYMBOL(cancel_delayed_work_sync);
3003
3004	/**
3005	* schedule_work_on - put work task on a specific cpu
3006	* @cpu: cpu to put the work task on
3007	* @work: job to be done
3008	*
3009	* This puts a job on a specific cpu
3010	*/
3011	bool schedule_work_on(int cpu, struct work_struct *work)
3012	{
3013	return queue_work_on(cpu, system_wq, work);
3014	}
3015	EXPORT_SYMBOL(schedule_work_on);
3016
3017	/**
3018	* schedule_work - put work task in global workqueue
3019	* @work: job to be done
3020	*
3021	* Returns %false if @work was already on the kernel-global workqueue and
3022	* %true otherwise.
3023	*
3024	* This puts a job in the kernel-global workqueue if it was not already
3025	* queued and leaves it in the same position on the kernel-global
3026	* workqueue otherwise.
3027	*/
3028	bool schedule_work(struct work_struct *work)
3029	{
3030	return queue_work(system_wq, work);
3031	}
3032	EXPORT_SYMBOL(schedule_work);
3033
3034	/**
3035	* schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3036	* @cpu: cpu to use
3037	* @dwork: job to be done
3038	* @delay: number of jiffies to wait
3039	*
3040	* After waiting for a given time this puts a job in the kernel-global
3041	* workqueue on the specified CPU.
3042	*/
3043	bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3044	unsigned long delay)
3045	{
3046	return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3047	}
3048	EXPORT_SYMBOL(schedule_delayed_work_on);
3049
3050	/**
3051	* schedule_delayed_work - put work task in global workqueue after delay
3052	* @dwork: job to be done
3053	* @delay: number of jiffies to wait or 0 for immediate execution
3054	*
3055	* After waiting for a given time this puts a job in the kernel-global
3056	* workqueue.
3057	*/
3058	bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3059	{
3060	return queue_delayed_work(system_wq, dwork, delay);
3061	}
3062	EXPORT_SYMBOL(schedule_delayed_work);
3063
3064	/**
3065	* schedule_on_each_cpu - execute a function synchronously on each online CPU
3066	* @func: the function to call
3067	*
3068	* schedule_on_each_cpu() executes @func on each online CPU using the
3069	* system workqueue and blocks until all CPUs have completed.
3070	* schedule_on_each_cpu() is very slow.
3071	*
3072	* RETURNS:
3073	* 0 on success, -errno on failure.
3074	*/
3075	int schedule_on_each_cpu(work_func_t func)
3076	{
3077	int cpu;
3078	struct work_struct __percpu *works;
3079
3080	works = alloc_percpu(struct work_struct);
3081	if (!works)
3082	return -ENOMEM;
3083
3084	get_online_cpus();
3085
3086	for_each_online_cpu(cpu) {
3087	struct work_struct *work = per_cpu_ptr(works, cpu);
3088
3089	INIT_WORK(work, func);
3090	schedule_work_on(cpu, work);
3091	}
3092
3093	for_each_online_cpu(cpu)
3094	flush_work(per_cpu_ptr(works, cpu));
3095
3096	put_online_cpus();
3097	free_percpu(works);
3098	return 0;
3099	}
3100
3101	/**
3102	* flush_scheduled_work - ensure that any scheduled work has run to completion.
3103	*
3104	* Forces execution of the kernel-global workqueue and blocks until its
3105	* completion.
3106	*
3107	* Think twice before calling this function! It's very easy to get into
3108	* trouble if you don't take great care. Either of the following situations
3109	* will lead to deadlock:
3110	*
3111	* One of the work items currently on the workqueue needs to acquire
3112	* a lock held by your code or its caller.
3113	*
3114	* Your code is running in the context of a work routine.
3115	*
3116	* They will be detected by lockdep when they occur, but the first might not
3117	* occur very often. It depends on what work items are on the workqueue and
3118	* what locks they need, which you have no control over.
3119	*
3120	* In most situations flushing the entire workqueue is overkill; you merely
3121	* need to know that a particular work item isn't queued and isn't running.
3122	* In such cases you should use cancel_delayed_work_sync() or
3123	* cancel_work_sync() instead.
3124	*/
3125	void flush_scheduled_work(void)
3126	{
3127	flush_workqueue(system_wq);
3128	}
3129	EXPORT_SYMBOL(flush_scheduled_work);
3130
3131	/**
3132	* execute_in_process_context - reliably execute the routine with user context
3133	* @fn: the function to execute
3134	* @ew: guaranteed storage for the execute work structure (must
3135	* be available when the work executes)
3136	*
3137	* Executes the function immediately if process context is available,
3138	* otherwise schedules the function for delayed execution.
3139	*
3140	* Returns: 0 - function was executed
3141	* 1 - function was scheduled for execution
3142	*/
3143	int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3144	{
3145	if (!in_interrupt()) {
3146	fn(&ew->work);
3147	return 0;
3148	}
3149
3150	INIT_WORK(&ew->work, fn);
3151	schedule_work(&ew->work);
3152
3153	return 1;
3154	}
3155	EXPORT_SYMBOL_GPL(execute_in_process_context);
3156
3157	int keventd_up(void)
3158	{
3159	return system_wq != NULL;
3160	}
3161
3162	static int alloc_cwqs(struct workqueue_struct *wq)
3163	{
3164	/*
3165	* cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3166	* Make sure that the alignment isn't lower than that of
3167	* unsigned long long.
3168	*/
3169	const size_t size = sizeof(struct cpu_workqueue_struct);
3170	const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3171	__alignof__(unsigned long long));
3172
3173	if (!(wq->flags & WQ_UNBOUND))
3174	wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3175	else {
3176	void *ptr;
3177
3178	/*
3179	* Allocate enough room to align cwq and put an extra
3180	* pointer at the end pointing back to the originally
3181	* allocated pointer which will be used for free.
3182	*/
3183	ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3184	if (ptr) {
3185	wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3186	(void *)(wq->cpu_wq.single + 1) = ptr;
3187	}
3188	}
3189
3190	/* just in case, make sure it's actually aligned */
3191	BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3192	return wq->cpu_wq.v ? 0 : -ENOMEM;
3193	}
3194
3195	static void free_cwqs(struct workqueue_struct *wq)
3196	{
3197	if (!(wq->flags & WQ_UNBOUND))
3198	free_percpu(wq->cpu_wq.pcpu);
3199	else if (wq->cpu_wq.single) {
3200	/* the pointer to free is stored right after the cwq */
3201	kfree((void *)(wq->cpu_wq.single + 1));
3202	}
3203	}
3204
3205	static int wq_clamp_max_active(int max_active, unsigned int flags,
3206	const char *name)
3207	{
3208	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3209
3210	if (max_active < 1 \|\| max_active > lim)
3211	pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3212	max_active, name, 1, lim);
3213
3214	return clamp_val(max_active, 1, lim);
3215	}
3216
3217	struct workqueue_struct __alloc_workqueue_key(const char fmt,
3218	unsigned int flags,
3219	int max_active,
3220	struct lock_class_key *key,
3221	const char *lock_name, ...)
3222	{
3223	va_list args, args1;
3224	struct workqueue_struct *wq;
3225	unsigned int cpu;
3226	size_t namelen;
3227
3228	/* determine namelen, allocate wq and format name */
3229	va_start(args, lock_name);
3230	va_copy(args1, args);
3231	namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3232
3233	wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3234	if (!wq)
3235	goto err;
3236
3237	vsnprintf(wq->name, namelen, fmt, args1);
3238	va_end(args);
3239	va_end(args1);
3240
3241	/*
3242	* Workqueues which may be used during memory reclaim should
3243	* have a rescuer to guarantee forward progress.
3244	*/
3245	if (flags & WQ_MEM_RECLAIM)
3246	flags \|= WQ_RESCUER;
3247
3248	max_active = max_active ?: WQ_DFL_ACTIVE;
3249	max_active = wq_clamp_max_active(max_active, flags, wq->name);
3250
3251	/* init wq */
3252	wq->flags = flags;
3253	wq->saved_max_active = max_active;
3254	mutex_init(&wq->flush_mutex);
3255	atomic_set(&wq->nr_cwqs_to_flush, 0);
3256	INIT_LIST_HEAD(&wq->flusher_queue);
3257	INIT_LIST_HEAD(&wq->flusher_overflow);
3258
3259	lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3260	INIT_LIST_HEAD(&wq->list);
3261
3262	if (alloc_cwqs(wq) < 0)
3263	goto err;
3264
3265	for_each_cwq_cpu(cpu, wq) {
3266	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3267	struct global_cwq *gcwq = get_gcwq(cpu);
3268	int pool_idx = (bool)(flags & WQ_HIGHPRI);
3269
3270	BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3271	cwq->pool = &gcwq->pools[pool_idx];
3272	cwq->wq = wq;
3273	cwq->flush_color = -1;
3274	cwq->max_active = max_active;
3275	INIT_LIST_HEAD(&cwq->delayed_works);
3276	}
3277
3278	if (flags & WQ_RESCUER) {
3279	struct worker *rescuer;
3280
3281	if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3282	goto err;
3283
3284	wq->rescuer = rescuer = alloc_worker();
3285	if (!rescuer)
3286	goto err;
3287
3288	rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3289	wq->name);
3290	if (IS_ERR(rescuer->task))
3291	goto err;
3292
3293	rescuer->task->flags \|= PF_THREAD_BOUND;
3294	wake_up_process(rescuer->task);
3295	}
3296
3297	/*
3298	* workqueue_lock protects global freeze state and workqueues
3299	* list. Grab it, set max_active accordingly and add the new
3300	* workqueue to workqueues list.
3301	*/
3302	spin_lock(&workqueue_lock);
3303
3304	if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3305	for_each_cwq_cpu(cpu, wq)
3306	get_cwq(cpu, wq)->max_active = 0;
3307
3308	list_add(&wq->list, &workqueues);
3309
3310	spin_unlock(&workqueue_lock);
3311
3312	return wq;
3313	err:
3314	if (wq) {
3315	free_cwqs(wq);
3316	free_mayday_mask(wq->mayday_mask);
3317	kfree(wq->rescuer);
3318	kfree(wq);
3319	}
3320	return NULL;
3321	}
3322	EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3323
3324	/**
3325	* destroy_workqueue - safely terminate a workqueue
3326	* @wq: target workqueue
3327	*
3328	* Safely destroy a workqueue. All work currently pending will be done first.
3329	*/
3330	void destroy_workqueue(struct workqueue_struct *wq)
3331	{
3332	unsigned int cpu;
3333
3334	/* drain it before proceeding with destruction */
3335	drain_workqueue(wq);
3336
3337	/*
3338	* wq list is used to freeze wq, remove from list after
3339	* flushing is complete in case freeze races us.
3340	*/
3341	spin_lock(&workqueue_lock);
3342	list_del(&wq->list);
3343	spin_unlock(&workqueue_lock);
3344
3345	/* sanity check */
3346	for_each_cwq_cpu(cpu, wq) {
3347	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3348	int i;
3349
3350	for (i = 0; i < WORK_NR_COLORS; i++)
3351	BUG_ON(cwq->nr_in_flight[i]);
3352	BUG_ON(cwq->nr_active);
3353	BUG_ON(!list_empty(&cwq->delayed_works));
3354	}
3355
3356	if (wq->flags & WQ_RESCUER) {
3357	kthread_stop(wq->rescuer->task);
3358	free_mayday_mask(wq->mayday_mask);
3359	kfree(wq->rescuer);
3360	}
3361
3362	free_cwqs(wq);
3363	kfree(wq);
3364	}
3365	EXPORT_SYMBOL_GPL(destroy_workqueue);
3366
3367	/**
3368	* cwq_set_max_active - adjust max_active of a cwq
3369	* @cwq: target cpu_workqueue_struct
3370	* @max_active: new max_active value.
3371	*
3372	* Set @cwq->max_active to @max_active and activate delayed works if
3373	* increased.
3374	*
3375	* CONTEXT:
3376	* spin_lock_irq(gcwq->lock).
3377	*/
3378	static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
3379	{
3380	cwq->max_active = max_active;
3381
3382	while (!list_empty(&cwq->delayed_works) &&
3383	cwq->nr_active < cwq->max_active)
3384	cwq_activate_first_delayed(cwq);
3385	}
3386
3387	/**
3388	* workqueue_set_max_active - adjust max_active of a workqueue
3389	* @wq: target workqueue
3390	* @max_active: new max_active value.
3391	*
3392	* Set max_active of @wq to @max_active.
3393	*
3394	* CONTEXT:
3395	* Don't call from IRQ context.
3396	*/
3397	void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3398	{
3399	unsigned int cpu;
3400
3401	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3402
3403	spin_lock(&workqueue_lock);
3404
3405	wq->saved_max_active = max_active;
3406
3407	for_each_cwq_cpu(cpu, wq) {
3408	struct global_cwq *gcwq = get_gcwq(cpu);
3409
3410	spin_lock_irq(&gcwq->lock);
3411
3412	if (!(wq->flags & WQ_FREEZABLE) \|\|
3413	!(gcwq->flags & GCWQ_FREEZING))
3414	cwq_set_max_active(get_cwq(gcwq->cpu, wq), max_active);
3415
3416	spin_unlock_irq(&gcwq->lock);
3417	}
3418
3419	spin_unlock(&workqueue_lock);
3420	}
3421	EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3422
3423	/**
3424	* workqueue_congested - test whether a workqueue is congested
3425	* @cpu: CPU in question
3426	* @wq: target workqueue
3427	*
3428	* Test whether @wq's cpu workqueue for @cpu is congested. There is
3429	* no synchronization around this function and the test result is
3430	* unreliable and only useful as advisory hints or for debugging.
3431	*
3432	* RETURNS:
3433	* %true if congested, %false otherwise.
3434	*/
3435	bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3436	{
3437	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3438
3439	return !list_empty(&cwq->delayed_works);
3440	}
3441	EXPORT_SYMBOL_GPL(workqueue_congested);
3442
3443	/**
3444	* work_cpu - return the last known associated cpu for @work
3445	* @work: the work of interest
3446	*
3447	* RETURNS:
3448	* CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3449	*/
3450	unsigned int work_cpu(struct work_struct *work)
3451	{
3452	struct global_cwq *gcwq = get_work_gcwq(work);
3453
3454	return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3455	}
3456	EXPORT_SYMBOL_GPL(work_cpu);
3457
3458	/**
3459	* work_busy - test whether a work is currently pending or running
3460	* @work: the work to be tested
3461	*
3462	* Test whether @work is currently pending or running. There is no
3463	* synchronization around this function and the test result is
3464	* unreliable and only useful as advisory hints or for debugging.
3465	* Especially for reentrant wqs, the pending state might hide the
3466	* running state.
3467	*
3468	* RETURNS:
3469	* OR'd bitmask of WORK_BUSY_* bits.
3470	*/
3471	unsigned int work_busy(struct work_struct *work)
3472	{
3473	struct global_cwq *gcwq = get_work_gcwq(work);
3474	unsigned long flags;
3475	unsigned int ret = 0;
3476
3477	if (!gcwq)
3478	return false;
3479
3480	spin_lock_irqsave(&gcwq->lock, flags);
3481
3482	if (work_pending(work))
3483	ret \|= WORK_BUSY_PENDING;
3484	if (find_worker_executing_work(gcwq, work))
3485	ret \|= WORK_BUSY_RUNNING;
3486
3487	spin_unlock_irqrestore(&gcwq->lock, flags);
3488
3489	return ret;
3490	}
3491	EXPORT_SYMBOL_GPL(work_busy);
3492
3493	/*
3494	* CPU hotplug.
3495	*
3496	* There are two challenges in supporting CPU hotplug. Firstly, there
3497	* are a lot of assumptions on strong associations among work, cwq and
3498	* gcwq which make migrating pending and scheduled works very
3499	* difficult to implement without impacting hot paths. Secondly,
3500	* gcwqs serve mix of short, long and very long running works making
3501	* blocked draining impractical.
3502	*
3503	* This is solved by allowing a gcwq to be disassociated from the CPU
3504	* running as an unbound one and allowing it to be reattached later if the
3505	* cpu comes back online.
3506	*/
3507
3508	/* claim manager positions of all pools */
3509	static void gcwq_claim_assoc_and_lock(struct global_cwq *gcwq)
3510	{
3511	struct worker_pool *pool;
3512
3513	for_each_worker_pool(pool, gcwq)
3514	mutex_lock_nested(&pool->assoc_mutex, pool - gcwq->pools);
3515	spin_lock_irq(&gcwq->lock);
3516	}
3517
3518	/* release manager positions */
3519	static void gcwq_release_assoc_and_unlock(struct global_cwq *gcwq)
3520	{
3521	struct worker_pool *pool;
3522
3523	spin_unlock_irq(&gcwq->lock);
3524	for_each_worker_pool(pool, gcwq)
3525	mutex_unlock(&pool->assoc_mutex);
3526	}
3527
3528	static void gcwq_unbind_fn(struct work_struct *work)
3529	{
3530	struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3531	struct worker_pool *pool;
3532	struct worker *worker;
3533	struct hlist_node *pos;
3534	int i;
3535
3536	BUG_ON(gcwq->cpu != smp_processor_id());
3537
3538	gcwq_claim_assoc_and_lock(gcwq);
3539
3540	/*
3541	* We've claimed all manager positions. Make all workers unbound
3542	* and set DISASSOCIATED. Before this, all workers except for the
3543	* ones which are still executing works from before the last CPU
3544	* down must be on the cpu. After this, they may become diasporas.
3545	*/
3546	for_each_worker_pool(pool, gcwq)
3547	list_for_each_entry(worker, &pool->idle_list, entry)
3548	worker->flags \|= WORKER_UNBOUND;
3549
3550	for_each_busy_worker(worker, i, pos, gcwq)
3551	worker->flags \|= WORKER_UNBOUND;
3552
3553	gcwq->flags \|= GCWQ_DISASSOCIATED;
3554
3555	gcwq_release_assoc_and_unlock(gcwq);
3556
3557	/*
3558	* Call schedule() so that we cross rq->lock and thus can guarantee
3559	* sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3560	* as scheduler callbacks may be invoked from other cpus.
3561	*/
3562	schedule();
3563
3564	/*
3565	* Sched callbacks are disabled now. Zap nr_running. After this,
3566	* nr_running stays zero and need_more_worker() and keep_working()
3567	* are always true as long as the worklist is not empty. @gcwq now
3568	* behaves as unbound (in terms of concurrency management) gcwq
3569	* which is served by workers tied to the CPU.
3570	*
3571	* On return from this function, the current worker would trigger
3572	* unbound chain execution of pending work items if other workers
3573	* didn't already.
3574	*/
3575	for_each_worker_pool(pool, gcwq)
3576	atomic_set(get_pool_nr_running(pool), 0);
3577	}
3578
3579	/*
3580	* Workqueues should be brought up before normal priority CPU notifiers.
3581	* This will be registered high priority CPU notifier.
3582	*/
3583	static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3584	unsigned long action,
3585	void *hcpu)
3586	{
3587	unsigned int cpu = (unsigned long)hcpu;
3588	struct global_cwq *gcwq = get_gcwq(cpu);
3589	struct worker_pool *pool;
3590
3591	switch (action & ~CPU_TASKS_FROZEN) {
3592	case CPU_UP_PREPARE:
3593	for_each_worker_pool(pool, gcwq) {
3594	struct worker *worker;
3595
3596	if (pool->nr_workers)
3597	continue;
3598
3599	worker = create_worker(pool);
3600	if (!worker)
3601	return NOTIFY_BAD;
3602
3603	spin_lock_irq(&gcwq->lock);
3604	start_worker(worker);
3605	spin_unlock_irq(&gcwq->lock);
3606	}
3607	break;
3608
3609	case CPU_DOWN_FAILED:
3610	case CPU_ONLINE:
3611	gcwq_claim_assoc_and_lock(gcwq);
3612	gcwq->flags &= ~GCWQ_DISASSOCIATED;
3613	rebind_workers(gcwq);
3614	gcwq_release_assoc_and_unlock(gcwq);
3615	break;
3616	}
3617	return NOTIFY_OK;
3618	}
3619
3620	/*
3621	* Workqueues should be brought down after normal priority CPU notifiers.
3622	* This will be registered as low priority CPU notifier.
3623	*/
3624	static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3625	unsigned long action,
3626	void *hcpu)
3627	{
3628	unsigned int cpu = (unsigned long)hcpu;
3629	struct work_struct unbind_work;
3630
3631	switch (action & ~CPU_TASKS_FROZEN) {
3632	case CPU_DOWN_PREPARE:
3633	/* unbinding should happen on the local CPU */
3634	INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3635	queue_work_on(cpu, system_highpri_wq, &unbind_work);
3636	flush_work(&unbind_work);
3637	break;
3638	}
3639	return NOTIFY_OK;
3640	}
3641
3642	#ifdef CONFIG_SMP
3643
3644	struct work_for_cpu {
3645	struct work_struct work;
3646	long (fn)(void );
3647	void *arg;
3648	long ret;
3649	};
3650
3651	static void work_for_cpu_fn(struct work_struct *work)
3652	{
3653	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3654
3655	wfc->ret = wfc->fn(wfc->arg);
3656	}
3657
3658	/**
3659	* work_on_cpu - run a function in user context on a particular cpu
3660	* @cpu: the cpu to run on
3661	* @fn: the function to run
3662	* @arg: the function arg
3663	*
3664	* This will return the value @fn returns.
3665	* It is up to the caller to ensure that the cpu doesn't go offline.
3666	* The caller must not hold any locks which would prevent @fn from completing.
3667	*/
3668	long work_on_cpu(unsigned int cpu, long (fn)(void ), void *arg)
3669	{
3670	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3671
3672	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3673	schedule_work_on(cpu, &wfc.work);
3674	flush_work(&wfc.work);
3675	return wfc.ret;
3676	}
3677	EXPORT_SYMBOL_GPL(work_on_cpu);
3678	#endif /* CONFIG_SMP */
3679
3680	#ifdef CONFIG_FREEZER
3681
3682	/**
3683	* freeze_workqueues_begin - begin freezing workqueues
3684	*
3685	* Start freezing workqueues. After this function returns, all freezable
3686	* workqueues will queue new works to their frozen_works list instead of
3687	* gcwq->worklist.
3688	*
3689	* CONTEXT:
3690	* Grabs and releases workqueue_lock and gcwq->lock's.
3691	*/
3692	void freeze_workqueues_begin(void)
3693	{
3694	unsigned int cpu;
3695
3696	spin_lock(&workqueue_lock);
3697
3698	BUG_ON(workqueue_freezing);
3699	workqueue_freezing = true;
3700
3701	for_each_gcwq_cpu(cpu) {
3702	struct global_cwq *gcwq = get_gcwq(cpu);
3703	struct workqueue_struct *wq;
3704
3705	spin_lock_irq(&gcwq->lock);
3706
3707	BUG_ON(gcwq->flags & GCWQ_FREEZING);
3708	gcwq->flags \|= GCWQ_FREEZING;
3709
3710	list_for_each_entry(wq, &workqueues, list) {
3711	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3712
3713	if (cwq && wq->flags & WQ_FREEZABLE)
3714	cwq->max_active = 0;
3715	}
3716
3717	spin_unlock_irq(&gcwq->lock);
3718	}
3719
3720	spin_unlock(&workqueue_lock);
3721	}
3722
3723	/**
3724	* freeze_workqueues_busy - are freezable workqueues still busy?
3725	*
3726	* Check whether freezing is complete. This function must be called
3727	* between freeze_workqueues_begin() and thaw_workqueues().
3728	*
3729	* CONTEXT:
3730	* Grabs and releases workqueue_lock.
3731	*
3732	* RETURNS:
3733	* %true if some freezable workqueues are still busy. %false if freezing
3734	* is complete.
3735	*/
3736	bool freeze_workqueues_busy(void)
3737	{
3738	unsigned int cpu;
3739	bool busy = false;
3740
3741	spin_lock(&workqueue_lock);
3742
3743	BUG_ON(!workqueue_freezing);
3744
3745	for_each_gcwq_cpu(cpu) {
3746	struct workqueue_struct *wq;
3747	/*
3748	* nr_active is monotonically decreasing. It's safe
3749	* to peek without lock.
3750	*/
3751	list_for_each_entry(wq, &workqueues, list) {
3752	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3753
3754	if (!cwq \|\| !(wq->flags & WQ_FREEZABLE))
3755	continue;
3756
3757	BUG_ON(cwq->nr_active < 0);
3758	if (cwq->nr_active) {
3759	busy = true;
3760	goto out_unlock;
3761	}
3762	}
3763	}
3764	out_unlock:
3765	spin_unlock(&workqueue_lock);
3766	return busy;
3767	}
3768
3769	/**
3770	* thaw_workqueues - thaw workqueues
3771	*
3772	* Thaw workqueues. Normal queueing is restored and all collected
3773	* frozen works are transferred to their respective gcwq worklists.
3774	*
3775	* CONTEXT:
3776	* Grabs and releases workqueue_lock and gcwq->lock's.
3777	*/
3778	void thaw_workqueues(void)
3779	{
3780	unsigned int cpu;
3781
3782	spin_lock(&workqueue_lock);
3783
3784	if (!workqueue_freezing)
3785	goto out_unlock;
3786
3787	for_each_gcwq_cpu(cpu) {
3788	struct global_cwq *gcwq = get_gcwq(cpu);
3789	struct worker_pool *pool;
3790	struct workqueue_struct *wq;
3791
3792	spin_lock_irq(&gcwq->lock);
3793
3794	BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3795	gcwq->flags &= ~GCWQ_FREEZING;
3796
3797	list_for_each_entry(wq, &workqueues, list) {
3798	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3799
3800	if (!cwq \|\| !(wq->flags & WQ_FREEZABLE))
3801	continue;
3802
3803	/* restore max_active and repopulate worklist */
3804	cwq_set_max_active(cwq, wq->saved_max_active);
3805	}
3806
3807	for_each_worker_pool(pool, gcwq)
3808	wake_up_worker(pool);
3809
3810	spin_unlock_irq(&gcwq->lock);
3811	}
3812
3813	workqueue_freezing = false;
3814	out_unlock:
3815	spin_unlock(&workqueue_lock);
3816	}
3817	#endif /* CONFIG_FREEZER */
3818
3819	static int __init init_workqueues(void)
3820	{
3821	unsigned int cpu;
3822	int i;
3823
3824	/* make sure we have enough bits for OFFQ CPU number */
3825	BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
3826	WORK_CPU_LAST);
3827
3828	cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3829	hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3830
3831	/* initialize gcwqs */
3832	for_each_gcwq_cpu(cpu) {
3833	struct global_cwq *gcwq = get_gcwq(cpu);
3834	struct worker_pool *pool;
3835
3836	spin_lock_init(&gcwq->lock);
3837	gcwq->cpu = cpu;
3838	gcwq->flags \|= GCWQ_DISASSOCIATED;
3839
3840	for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3841	INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3842
3843	for_each_worker_pool(pool, gcwq) {
3844	pool->gcwq = gcwq;
3845	INIT_LIST_HEAD(&pool->worklist);
3846	INIT_LIST_HEAD(&pool->idle_list);
3847
3848	init_timer_deferrable(&pool->idle_timer);
3849	pool->idle_timer.function = idle_worker_timeout;
3850	pool->idle_timer.data = (unsigned long)pool;
3851
3852	setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3853	(unsigned long)pool);
3854
3855	mutex_init(&pool->assoc_mutex);
3856	ida_init(&pool->worker_ida);
3857	}
3858	}
3859
3860	/* create the initial worker */
3861	for_each_online_gcwq_cpu(cpu) {
3862	struct global_cwq *gcwq = get_gcwq(cpu);
3863	struct worker_pool *pool;
3864
3865	if (cpu != WORK_CPU_UNBOUND)
3866	gcwq->flags &= ~GCWQ_DISASSOCIATED;
3867
3868	for_each_worker_pool(pool, gcwq) {
3869	struct worker *worker;
3870
3871	worker = create_worker(pool);
3872	BUG_ON(!worker);
3873	spin_lock_irq(&gcwq->lock);
3874	start_worker(worker);
3875	spin_unlock_irq(&gcwq->lock);
3876	}
3877	}
3878
3879	system_wq = alloc_workqueue("events", 0, 0);
3880	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3881	system_long_wq = alloc_workqueue("events_long", 0, 0);
3882	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3883	WQ_UNBOUND_MAX_ACTIVE);
3884	system_freezable_wq = alloc_workqueue("events_freezable",
3885	WQ_FREEZABLE, 0);
3886	BUG_ON(!system_wq \|\| !system_highpri_wq \|\| !system_long_wq \|\|
3887	!system_unbound_wq \|\| !system_freezable_wq);
3888	return 0;
3889	}
3890	early_initcall(init_workqueues);