linux - kernel/workqueue.c

3800 lines · 103.0 KiB

1	/*
2	* kernel/workqueue.c - generic async execution with shared worker pool
3	*
4	* Copyright (C) 2002 Ingo Molnar
5	*
6	* Derived from the taskqueue/keventd code by:
7	* David Woodhouse <[email protected]>
8	* Andrew Morton
9	* Kai Petzke <[email protected]>
10	* Theodore Ts'o <[email protected]>
11	*
12	* Made to use alloc_percpu by Christoph Lameter.
13	*
14	* Copyright (C) 2010 SUSE Linux Products GmbH
15	* Copyright (C) 2010 Tejun Heo <[email protected]>
16	*
17	* This is the generic async execution mechanism. Work items as are
18	* executed in process context. The worker pool is shared and
19	* automatically managed. There is one worker pool for each CPU and
20	* one extra for works which are better served by workers which are
21	* not bound to any specific CPU.
22	*
23	* Please read Documentation/workqueue.txt for details.
24	*/
25
26	#include <linux/module.h>
27	#include <linux/kernel.h>
28	#include <linux/sched.h>
29	#include <linux/init.h>
30	#include <linux/signal.h>
31	#include <linux/completion.h>
32	#include <linux/workqueue.h>
33	#include <linux/slab.h>
34	#include <linux/cpu.h>
35	#include <linux/notifier.h>
36	#include <linux/kthread.h>
37	#include <linux/hardirq.h>
38	#include <linux/mempolicy.h>
39	#include <linux/freezer.h>
40	#include <linux/kallsyms.h>
41	#include <linux/debug_locks.h>
42	#include <linux/lockdep.h>
43	#include <linux/idr.h>
44
45	#include "workqueue_sched.h"
46
47	enum {
48	/* global_cwq flags */
49	GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
50	GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */
51	GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
52	GCWQ_FREEZING = 1 << 3, /* freeze in progress */
53	GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */
54
55	/* worker flags */
56	WORKER_STARTED = 1 << 0, /* started */
57	WORKER_DIE = 1 << 1, /* die die die */
58	WORKER_IDLE = 1 << 2, /* is idle */
59	WORKER_PREP = 1 << 3, /* preparing to run works */
60	WORKER_ROGUE = 1 << 4, /* not bound to any cpu */
61	WORKER_REBIND = 1 << 5, /* mom is home, come back */
62	WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
63	WORKER_UNBOUND = 1 << 7, /* worker is unbound */
64
65	WORKER_NOT_RUNNING = WORKER_PREP \| WORKER_ROGUE \| WORKER_REBIND \|
66	WORKER_CPU_INTENSIVE \| WORKER_UNBOUND,
67
68	/* gcwq->trustee_state */
69	TRUSTEE_START = 0, /* start */
70	TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */
71	TRUSTEE_BUTCHER = 2, /* butcher workers */
72	TRUSTEE_RELEASE = 3, /* release workers */
73	TRUSTEE_DONE = 4, /* trustee is done */
74
75	BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
76	BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
77	BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
78
79	MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
80	IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
81
82	MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
83	/* call for help after 10ms
84	(min two ticks) */
85	MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
86	CREATE_COOLDOWN = HZ, /* time to breath after fail */
87	TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
88
89	/*
90	* Rescue workers are used only on emergencies and shared by
91	* all cpus. Give -20.
92	*/
93	RESCUER_NICE_LEVEL = -20,
94	};
95
96	/*
97	* Structure fields follow one of the following exclusion rules.
98	*
99	* I: Modifiable by initialization/destruction paths and read-only for
100	* everyone else.
101	*
102	* P: Preemption protected. Disabling preemption is enough and should
103	* only be modified and accessed from the local cpu.
104	*
105	* L: gcwq->lock protected. Access with gcwq->lock held.
106	*
107	* X: During normal operation, modification requires gcwq->lock and
108	* should be done only from local cpu. Either disabling preemption
109	* on local cpu or grabbing gcwq->lock is enough for read access.
110	* If GCWQ_DISASSOCIATED is set, it's identical to L.
111	*
112	* F: wq->flush_mutex protected.
113	*
114	* W: workqueue_lock protected.
115	*/
116
117	struct global_cwq;
118
119	/*
120	* The poor guys doing the actual heavy lifting. All on-duty workers
121	* are either serving the manager role, on idle list or on busy hash.
122	*/
123	struct worker {
124	/* on idle list while idle, on busy hash table while busy */
125	union {
126	struct list_head entry; /* L: while idle */
127	struct hlist_node hentry; /* L: while busy */
128	};
129
130	struct work_struct current_work; / L: work being processed */
131	struct cpu_workqueue_struct current_cwq; / L: current_work's cwq */
132	struct list_head scheduled; /* L: scheduled works */
133	struct task_struct task; / I: worker task */
134	struct global_cwq gcwq; / I: the associated gcwq */
135	/* 64 bytes boundary on 64bit, 32 on 32bit */
136	unsigned long last_active; /* L: last active timestamp */
137	unsigned int flags; /* X: flags */
138	int id; /* I: worker id */
139	struct work_struct rebind_work; /* L: rebind worker to cpu */
140	};
141
142	/*
143	* Global per-cpu workqueue. There's one and only one for each cpu
144	* and all works are queued and processed here regardless of their
145	* target workqueues.
146	*/
147	struct global_cwq {
148	spinlock_t lock; /* the gcwq lock */
149	struct list_head worklist; /* L: list of pending works */
150	unsigned int cpu; /* I: the associated cpu */
151	unsigned int flags; /* L: GCWQ_* flags */
152
153	int nr_workers; /* L: total number of workers */
154	int nr_idle; /* L: currently idle ones */
155
156	/* workers are chained either in the idle_list or busy_hash */
157	struct list_head idle_list; /* X: list of idle workers */
158	struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
159	/* L: hash of busy workers */
160
161	struct timer_list idle_timer; /* L: worker idle timeout */
162	struct timer_list mayday_timer; /* L: SOS timer for dworkers */
163
164	struct ida worker_ida; /* L: for worker IDs */
165
166	struct task_struct trustee; / L: for gcwq shutdown */
167	unsigned int trustee_state; /* L: trustee state */
168	wait_queue_head_t trustee_wait; /* trustee wait */
169	struct worker first_idle; / L: first idle worker */
170	} ____cacheline_aligned_in_smp;
171
172	/*
173	* The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
174	* work_struct->data are used for flags and thus cwqs need to be
175	* aligned at two's power of the number of flag bits.
176	*/
177	struct cpu_workqueue_struct {
178	struct global_cwq gcwq; / I: the associated gcwq */
179	struct workqueue_struct wq; / I: the owning workqueue */
180	int work_color; /* L: current color */
181	int flush_color; /* L: flushing color */
182	int nr_in_flight[WORK_NR_COLORS];
183	/* L: nr of in_flight works */
184	int nr_active; /* L: nr of active works */
185	int max_active; /* L: max active works */
186	struct list_head delayed_works; /* L: delayed works */
187	};
188
189	/*
190	* Structure used to wait for workqueue flush.
191	*/
192	struct wq_flusher {
193	struct list_head list; /* F: list of flushers */
194	int flush_color; /* F: flush color waiting for */
195	struct completion done; /* flush completion */
196	};
197
198	/*
199	* All cpumasks are assumed to be always set on UP and thus can't be
200	* used to determine whether there's something to be done.
201	*/
202	#ifdef CONFIG_SMP
203	typedef cpumask_var_t mayday_mask_t;
204	#define mayday_test_and_set_cpu(cpu, mask) \
205	cpumask_test_and_set_cpu((cpu), (mask))
206	#define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
207	#define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
208	#define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
209	#define free_mayday_mask(mask) free_cpumask_var((mask))
210	#else
211	typedef unsigned long mayday_mask_t;
212	#define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
213	#define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
214	#define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
215	#define alloc_mayday_mask(maskp, gfp) true
216	#define free_mayday_mask(mask) do { } while (0)
217	#endif
218
219	/*
220	* The externally visible workqueue abstraction is an array of
221	* per-CPU workqueues:
222	*/
223	struct workqueue_struct {
224	unsigned int flags; /* I: WQ_* flags */
225	union {
226	struct cpu_workqueue_struct __percpu *pcpu;
227	struct cpu_workqueue_struct *single;
228	unsigned long v;
229	} cpu_wq; /* I: cwq's */
230	struct list_head list; /* W: list of all workqueues */
231
232	struct mutex flush_mutex; /* protects wq flushing */
233	int work_color; /* F: current work color */
234	int flush_color; /* F: current flush color */
235	atomic_t nr_cwqs_to_flush; /* flush in progress */
236	struct wq_flusher first_flusher; / F: first flusher */
237	struct list_head flusher_queue; /* F: flush waiters */
238	struct list_head flusher_overflow; /* F: flush overflow list */
239
240	mayday_mask_t mayday_mask; /* cpus requesting rescue */
241	struct worker rescuer; / I: rescue worker */
242
243	int saved_max_active; /* W: saved cwq max_active */
244	const char name; / I: workqueue name */
245	#ifdef CONFIG_LOCKDEP
246	struct lockdep_map lockdep_map;
247	#endif
248	};
249
250	struct workqueue_struct *system_wq __read_mostly;
251	struct workqueue_struct *system_long_wq __read_mostly;
252	struct workqueue_struct *system_nrt_wq __read_mostly;
253	struct workqueue_struct *system_unbound_wq __read_mostly;
254	struct workqueue_struct *system_freezable_wq __read_mostly;
255	EXPORT_SYMBOL_GPL(system_wq);
256	EXPORT_SYMBOL_GPL(system_long_wq);
257	EXPORT_SYMBOL_GPL(system_nrt_wq);
258	EXPORT_SYMBOL_GPL(system_unbound_wq);
259	EXPORT_SYMBOL_GPL(system_freezable_wq);
260
261	#define CREATE_TRACE_POINTS
262	#include <trace/events/workqueue.h>
263
264	#define for_each_busy_worker(worker, i, pos, gcwq) \
265	for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
266	hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
267
268	static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
269	unsigned int sw)
270	{
271	if (cpu < nr_cpu_ids) {
272	if (sw & 1) {
273	cpu = cpumask_next(cpu, mask);
274	if (cpu < nr_cpu_ids)
275	return cpu;
276	}
277	if (sw & 2)
278	return WORK_CPU_UNBOUND;
279	}
280	return WORK_CPU_NONE;
281	}
282
283	static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
284	struct workqueue_struct *wq)
285	{
286	return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
287	}
288
289	/*
290	* CPU iterators
291	*
292	* An extra gcwq is defined for an invalid cpu number
293	* (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
294	* specific CPU. The following iterators are similar to
295	* for_each_*_cpu() iterators but also considers the unbound gcwq.
296	*
297	* for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
298	* for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
299	* for_each_cwq_cpu() : possible CPUs for bound workqueues,
300	* WORK_CPU_UNBOUND for unbound workqueues
301	*/
302	#define for_each_gcwq_cpu(cpu) \
303	for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
304	(cpu) < WORK_CPU_NONE; \
305	(cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
306
307	#define for_each_online_gcwq_cpu(cpu) \
308	for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
309	(cpu) < WORK_CPU_NONE; \
310	(cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
311
312	#define for_each_cwq_cpu(cpu, wq) \
313	for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
314	(cpu) < WORK_CPU_NONE; \
315	(cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
316
317	#ifdef CONFIG_DEBUG_OBJECTS_WORK
318
319	static struct debug_obj_descr work_debug_descr;
320
321	static void work_debug_hint(void addr)
322	{
323	return ((struct work_struct *) addr)->func;
324	}
325
326	/*
327	* fixup_init is called when:
328	* - an active object is initialized
329	*/
330	static int work_fixup_init(void *addr, enum debug_obj_state state)
331	{
332	struct work_struct *work = addr;
333
334	switch (state) {
335	case ODEBUG_STATE_ACTIVE:
336	cancel_work_sync(work);
337	debug_object_init(work, &work_debug_descr);
338	return 1;
339	default:
340	return 0;
341	}
342	}
343
344	/*
345	* fixup_activate is called when:
346	* - an active object is activated
347	* - an unknown object is activated (might be a statically initialized object)
348	*/
349	static int work_fixup_activate(void *addr, enum debug_obj_state state)
350	{
351	struct work_struct *work = addr;
352
353	switch (state) {
354
355	case ODEBUG_STATE_NOTAVAILABLE:
356	/*
357	* This is not really a fixup. The work struct was
358	* statically initialized. We just make sure that it
359	* is tracked in the object tracker.
360	*/
361	if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
362	debug_object_init(work, &work_debug_descr);
363	debug_object_activate(work, &work_debug_descr);
364	return 0;
365	}
366	WARN_ON_ONCE(1);
367	return 0;
368
369	case ODEBUG_STATE_ACTIVE:
370	WARN_ON(1);
371
372	default:
373	return 0;
374	}
375	}
376
377	/*
378	* fixup_free is called when:
379	* - an active object is freed
380	*/
381	static int work_fixup_free(void *addr, enum debug_obj_state state)
382	{
383	struct work_struct *work = addr;
384
385	switch (state) {
386	case ODEBUG_STATE_ACTIVE:
387	cancel_work_sync(work);
388	debug_object_free(work, &work_debug_descr);
389	return 1;
390	default:
391	return 0;
392	}
393	}
394
395	static struct debug_obj_descr work_debug_descr = {
396	.name = "work_struct",
397	.debug_hint = work_debug_hint,
398	.fixup_init = work_fixup_init,
399	.fixup_activate = work_fixup_activate,
400	.fixup_free = work_fixup_free,
401	};
402
403	static inline void debug_work_activate(struct work_struct *work)
404	{
405	debug_object_activate(work, &work_debug_descr);
406	}
407
408	static inline void debug_work_deactivate(struct work_struct *work)
409	{
410	debug_object_deactivate(work, &work_debug_descr);
411	}
412
413	void __init_work(struct work_struct *work, int onstack)
414	{
415	if (onstack)
416	debug_object_init_on_stack(work, &work_debug_descr);
417	else
418	debug_object_init(work, &work_debug_descr);
419	}
420	EXPORT_SYMBOL_GPL(__init_work);
421
422	void destroy_work_on_stack(struct work_struct *work)
423	{
424	debug_object_free(work, &work_debug_descr);
425	}
426	EXPORT_SYMBOL_GPL(destroy_work_on_stack);
427
428	#else
429	static inline void debug_work_activate(struct work_struct *work) { }
430	static inline void debug_work_deactivate(struct work_struct *work) { }
431	#endif
432
433	/* Serializes the accesses to the list of workqueues. */
434	static DEFINE_SPINLOCK(workqueue_lock);
435	static LIST_HEAD(workqueues);
436	static bool workqueue_freezing; /* W: have wqs started freezing? */
437
438	/*
439	* The almighty global cpu workqueues. nr_running is the only field
440	* which is expected to be used frequently by other cpus via
441	* try_to_wake_up(). Put it in a separate cacheline.
442	*/
443	static DEFINE_PER_CPU(struct global_cwq, global_cwq);
444	static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
445
446	/*
447	* Global cpu workqueue and nr_running counter for unbound gcwq. The
448	* gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
449	* workers have WORKER_UNBOUND set.
450	*/
451	static struct global_cwq unbound_global_cwq;
452	static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
453
454	static int worker_thread(void *__worker);
455
456	static struct global_cwq *get_gcwq(unsigned int cpu)
457	{
458	if (cpu != WORK_CPU_UNBOUND)
459	return &per_cpu(global_cwq, cpu);
460	else
461	return &unbound_global_cwq;
462	}
463
464	static atomic_t *get_gcwq_nr_running(unsigned int cpu)
465	{
466	if (cpu != WORK_CPU_UNBOUND)
467	return &per_cpu(gcwq_nr_running, cpu);
468	else
469	return &unbound_gcwq_nr_running;
470	}
471
472	static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
473	struct workqueue_struct *wq)
474	{
475	if (!(wq->flags & WQ_UNBOUND)) {
476	if (likely(cpu < nr_cpu_ids)) {
477	#ifdef CONFIG_SMP
478	return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
479	#else
480	return wq->cpu_wq.single;
481	#endif
482	}
483	} else if (likely(cpu == WORK_CPU_UNBOUND))
484	return wq->cpu_wq.single;
485	return NULL;
486	}
487
488	static unsigned int work_color_to_flags(int color)
489	{
490	return color << WORK_STRUCT_COLOR_SHIFT;
491	}
492
493	static int get_work_color(struct work_struct *work)
494	{
495	return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
496	((1 << WORK_STRUCT_COLOR_BITS) - 1);
497	}
498
499	static int work_next_color(int color)
500	{
501	return (color + 1) % WORK_NR_COLORS;
502	}
503
504	/*
505	* A work's data points to the cwq with WORK_STRUCT_CWQ set while the
506	* work is on queue. Once execution starts, WORK_STRUCT_CWQ is
507	* cleared and the work data contains the cpu number it was last on.
508	*
509	* set_work_{cwq\|cpu}() and clear_work_data() can be used to set the
510	* cwq, cpu or clear work->data. These functions should only be
511	* called while the work is owned - ie. while the PENDING bit is set.
512	*
513	* get_work_[g]cwq() can be used to obtain the gcwq or cwq
514	* corresponding to a work. gcwq is available once the work has been
515	* queued anywhere after initialization. cwq is available only from
516	* queueing until execution starts.
517	*/
518	static inline void set_work_data(struct work_struct *work, unsigned long data,
519	unsigned long flags)
520	{
521	BUG_ON(!work_pending(work));
522	atomic_long_set(&work->data, data \| flags \| work_static(work));
523	}
524
525	static void set_work_cwq(struct work_struct *work,
526	struct cpu_workqueue_struct *cwq,
527	unsigned long extra_flags)
528	{
529	set_work_data(work, (unsigned long)cwq,
530	WORK_STRUCT_PENDING \| WORK_STRUCT_CWQ \| extra_flags);
531	}
532
533	static void set_work_cpu(struct work_struct *work, unsigned int cpu)
534	{
535	set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
536	}
537
538	static void clear_work_data(struct work_struct *work)
539	{
540	set_work_data(work, WORK_STRUCT_NO_CPU, 0);
541	}
542
543	static struct cpu_workqueue_struct get_work_cwq(struct work_struct work)
544	{
545	unsigned long data = atomic_long_read(&work->data);
546
547	if (data & WORK_STRUCT_CWQ)
548	return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
549	else
550	return NULL;
551	}
552
553	static struct global_cwq get_work_gcwq(struct work_struct work)
554	{
555	unsigned long data = atomic_long_read(&work->data);
556	unsigned int cpu;
557
558	if (data & WORK_STRUCT_CWQ)
559	return ((struct cpu_workqueue_struct *)
560	(data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
561
562	cpu = data >> WORK_STRUCT_FLAG_BITS;
563	if (cpu == WORK_CPU_NONE)
564	return NULL;
565
566	BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
567	return get_gcwq(cpu);
568	}
569
570	/*
571	* Policy functions. These define the policies on how the global
572	* worker pool is managed. Unless noted otherwise, these functions
573	* assume that they're being called with gcwq->lock held.
574	*/
575
576	static bool __need_more_worker(struct global_cwq *gcwq)
577	{
578	return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) \|\|
579	gcwq->flags & GCWQ_HIGHPRI_PENDING;
580	}
581
582	/*
583	* Need to wake up a worker? Called from anything but currently
584	* running workers.
585	*/
586	static bool need_more_worker(struct global_cwq *gcwq)
587	{
588	return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
589	}
590
591	/* Can I start working? Called from busy but !running workers. */
592	static bool may_start_working(struct global_cwq *gcwq)
593	{
594	return gcwq->nr_idle;
595	}
596
597	/* Do I need to keep working? Called from currently running workers. */
598	static bool keep_working(struct global_cwq *gcwq)
599	{
600	atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
601
602	return !list_empty(&gcwq->worklist) &&
603	(atomic_read(nr_running) <= 1 \|\|
604	gcwq->flags & GCWQ_HIGHPRI_PENDING);
605	}
606
607	/* Do we need a new worker? Called from manager. */
608	static bool need_to_create_worker(struct global_cwq *gcwq)
609	{
610	return need_more_worker(gcwq) && !may_start_working(gcwq);
611	}
612
613	/* Do I need to be the manager? */
614	static bool need_to_manage_workers(struct global_cwq *gcwq)
615	{
616	return need_to_create_worker(gcwq) \|\| gcwq->flags & GCWQ_MANAGE_WORKERS;
617	}
618
619	/* Do we have too many workers and should some go away? */
620	static bool too_many_workers(struct global_cwq *gcwq)
621	{
622	bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
623	int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
624	int nr_busy = gcwq->nr_workers - nr_idle;
625
626	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
627	}
628
629	/*
630	* Wake up functions.
631	*/
632
633	/* Return the first worker. Safe with preemption disabled */
634	static struct worker first_worker(struct global_cwq gcwq)
635	{
636	if (unlikely(list_empty(&gcwq->idle_list)))
637	return NULL;
638
639	return list_first_entry(&gcwq->idle_list, struct worker, entry);
640	}
641
642	/**
643	* wake_up_worker - wake up an idle worker
644	* @gcwq: gcwq to wake worker for
645	*
646	* Wake up the first idle worker of @gcwq.
647	*
648	* CONTEXT:
649	* spin_lock_irq(gcwq->lock).
650	*/
651	static void wake_up_worker(struct global_cwq *gcwq)
652	{
653	struct worker *worker = first_worker(gcwq);
654
655	if (likely(worker))
656	wake_up_process(worker->task);
657	}
658
659	/**
660	* wq_worker_waking_up - a worker is waking up
661	* @task: task waking up
662	* @cpu: CPU @task is waking up to
663	*
664	* This function is called during try_to_wake_up() when a worker is
665	* being awoken.
666	*
667	* CONTEXT:
668	* spin_lock_irq(rq->lock)
669	*/
670	void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
671	{
672	struct worker *worker = kthread_data(task);
673
674	if (!(worker->flags & WORKER_NOT_RUNNING))
675	atomic_inc(get_gcwq_nr_running(cpu));
676	}
677
678	/**
679	* wq_worker_sleeping - a worker is going to sleep
680	* @task: task going to sleep
681	* @cpu: CPU in question, must be the current CPU number
682	*
683	* This function is called during schedule() when a busy worker is
684	* going to sleep. Worker on the same cpu can be woken up by
685	* returning pointer to its task.
686	*
687	* CONTEXT:
688	* spin_lock_irq(rq->lock)
689	*
690	* RETURNS:
691	* Worker task on @cpu to wake up, %NULL if none.
692	*/
693	struct task_struct wq_worker_sleeping(struct task_struct task,
694	unsigned int cpu)
695	{
696	struct worker worker = kthread_data(task), to_wakeup = NULL;
697	struct global_cwq *gcwq = get_gcwq(cpu);
698	atomic_t *nr_running = get_gcwq_nr_running(cpu);
699
700	if (worker->flags & WORKER_NOT_RUNNING)
701	return NULL;
702
703	/* this can only happen on the local cpu */
704	BUG_ON(cpu != raw_smp_processor_id());
705
706	/*
707	* The counterpart of the following dec_and_test, implied mb,
708	* worklist not empty test sequence is in insert_work().
709	* Please read comment there.
710	*
711	* NOT_RUNNING is clear. This means that trustee is not in
712	* charge and we're running on the local cpu w/ rq lock held
713	* and preemption disabled, which in turn means that none else
714	* could be manipulating idle_list, so dereferencing idle_list
715	* without gcwq lock is safe.
716	*/
717	if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
718	to_wakeup = first_worker(gcwq);
719	return to_wakeup ? to_wakeup->task : NULL;
720	}
721
722	/**
723	* worker_set_flags - set worker flags and adjust nr_running accordingly
724	* @worker: self
725	* @flags: flags to set
726	* @wakeup: wakeup an idle worker if necessary
727	*
728	* Set @flags in @worker->flags and adjust nr_running accordingly. If
729	* nr_running becomes zero and @wakeup is %true, an idle worker is
730	* woken up.
731	*
732	* CONTEXT:
733	* spin_lock_irq(gcwq->lock)
734	*/
735	static inline void worker_set_flags(struct worker *worker, unsigned int flags,
736	bool wakeup)
737	{
738	struct global_cwq *gcwq = worker->gcwq;
739
740	WARN_ON_ONCE(worker->task != current);
741
742	/*
743	* If transitioning into NOT_RUNNING, adjust nr_running and
744	* wake up an idle worker as necessary if requested by
745	* @wakeup.
746	*/
747	if ((flags & WORKER_NOT_RUNNING) &&
748	!(worker->flags & WORKER_NOT_RUNNING)) {
749	atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
750
751	if (wakeup) {
752	if (atomic_dec_and_test(nr_running) &&
753	!list_empty(&gcwq->worklist))
754	wake_up_worker(gcwq);
755	} else
756	atomic_dec(nr_running);
757	}
758
759	worker->flags \|= flags;
760	}
761
762	/**
763	* worker_clr_flags - clear worker flags and adjust nr_running accordingly
764	* @worker: self
765	* @flags: flags to clear
766	*
767	* Clear @flags in @worker->flags and adjust nr_running accordingly.
768	*
769	* CONTEXT:
770	* spin_lock_irq(gcwq->lock)
771	*/
772	static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
773	{
774	struct global_cwq *gcwq = worker->gcwq;
775	unsigned int oflags = worker->flags;
776
777	WARN_ON_ONCE(worker->task != current);
778
779	worker->flags &= ~flags;
780
781	/*
782	* If transitioning out of NOT_RUNNING, increment nr_running. Note
783	* that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
784	* of multiple flags, not a single flag.
785	*/
786	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
787	if (!(worker->flags & WORKER_NOT_RUNNING))
788	atomic_inc(get_gcwq_nr_running(gcwq->cpu));
789	}
790
791	/**
792	* busy_worker_head - return the busy hash head for a work
793	* @gcwq: gcwq of interest
794	* @work: work to be hashed
795	*
796	* Return hash head of @gcwq for @work.
797	*
798	* CONTEXT:
799	* spin_lock_irq(gcwq->lock).
800	*
801	* RETURNS:
802	* Pointer to the hash head.
803	*/
804	static struct hlist_head busy_worker_head(struct global_cwq gcwq,
805	struct work_struct *work)
806	{
807	const int base_shift = ilog2(sizeof(struct work_struct));
808	unsigned long v = (unsigned long)work;
809
810	/* simple shift and fold hash, do we need something better? */
811	v >>= base_shift;
812	v += v >> BUSY_WORKER_HASH_ORDER;
813	v &= BUSY_WORKER_HASH_MASK;
814
815	return &gcwq->busy_hash[v];
816	}
817
818	/**
819	* __find_worker_executing_work - find worker which is executing a work
820	* @gcwq: gcwq of interest
821	* @bwh: hash head as returned by busy_worker_head()
822	* @work: work to find worker for
823	*
824	* Find a worker which is executing @work on @gcwq. @bwh should be
825	* the hash head obtained by calling busy_worker_head() with the same
826	* work.
827	*
828	* CONTEXT:
829	* spin_lock_irq(gcwq->lock).
830	*
831	* RETURNS:
832	* Pointer to worker which is executing @work if found, NULL
833	* otherwise.
834	*/
835	static struct worker __find_worker_executing_work(struct global_cwq gcwq,
836	struct hlist_head *bwh,
837	struct work_struct *work)
838	{
839	struct worker *worker;
840	struct hlist_node *tmp;
841
842	hlist_for_each_entry(worker, tmp, bwh, hentry)
843	if (worker->current_work == work)
844	return worker;
845	return NULL;
846	}
847
848	/**
849	* find_worker_executing_work - find worker which is executing a work
850	* @gcwq: gcwq of interest
851	* @work: work to find worker for
852	*
853	* Find a worker which is executing @work on @gcwq. This function is
854	* identical to __find_worker_executing_work() except that this
855	* function calculates @bwh itself.
856	*
857	* CONTEXT:
858	* spin_lock_irq(gcwq->lock).
859	*
860	* RETURNS:
861	* Pointer to worker which is executing @work if found, NULL
862	* otherwise.
863	*/
864	static struct worker find_worker_executing_work(struct global_cwq gcwq,
865	struct work_struct *work)
866	{
867	return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
868	work);
869	}
870
871	/**
872	* gcwq_determine_ins_pos - find insertion position
873	* @gcwq: gcwq of interest
874	* @cwq: cwq a work is being queued for
875	*
876	* A work for @cwq is about to be queued on @gcwq, determine insertion
877	* position for the work. If @cwq is for HIGHPRI wq, the work is
878	* queued at the head of the queue but in FIFO order with respect to
879	* other HIGHPRI works; otherwise, at the end of the queue. This
880	* function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
881	* there are HIGHPRI works pending.
882	*
883	* CONTEXT:
884	* spin_lock_irq(gcwq->lock).
885	*
886	* RETURNS:
887	* Pointer to inserstion position.
888	*/
889	static inline struct list_head gcwq_determine_ins_pos(struct global_cwq gcwq,
890	struct cpu_workqueue_struct *cwq)
891	{
892	struct work_struct *twork;
893
894	if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
895	return &gcwq->worklist;
896
897	list_for_each_entry(twork, &gcwq->worklist, entry) {
898	struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
899
900	if (!(tcwq->wq->flags & WQ_HIGHPRI))
901	break;
902	}
903
904	gcwq->flags \|= GCWQ_HIGHPRI_PENDING;
905	return &twork->entry;
906	}
907
908	/**
909	* insert_work - insert a work into gcwq
910	* @cwq: cwq @work belongs to
911	* @work: work to insert
912	* @head: insertion point
913	* @extra_flags: extra WORK_STRUCT_* flags to set
914	*
915	* Insert @work which belongs to @cwq into @gcwq after @head.
916	* @extra_flags is or'd to work_struct flags.
917	*
918	* CONTEXT:
919	* spin_lock_irq(gcwq->lock).
920	*/
921	static void insert_work(struct cpu_workqueue_struct *cwq,
922	struct work_struct work, struct list_head head,
923	unsigned int extra_flags)
924	{
925	struct global_cwq *gcwq = cwq->gcwq;
926
927	/* we own @work, set data and link */
928	set_work_cwq(work, cwq, extra_flags);
929
930	/*
931	* Ensure that we get the right work->data if we see the
932	* result of list_add() below, see try_to_grab_pending().
933	*/
934	smp_wmb();
935
936	list_add_tail(&work->entry, head);
937
938	/*
939	* Ensure either worker_sched_deactivated() sees the above
940	* list_add_tail() or we see zero nr_running to avoid workers
941	* lying around lazily while there are works to be processed.
942	*/
943	smp_mb();
944
945	if (__need_more_worker(gcwq))
946	wake_up_worker(gcwq);
947	}
948
949	/*
950	* Test whether @work is being queued from another work executing on the
951	* same workqueue. This is rather expensive and should only be used from
952	* cold paths.
953	*/
954	static bool is_chained_work(struct workqueue_struct *wq)
955	{
956	unsigned long flags;
957	unsigned int cpu;
958
959	for_each_gcwq_cpu(cpu) {
960	struct global_cwq *gcwq = get_gcwq(cpu);
961	struct worker *worker;
962	struct hlist_node *pos;
963	int i;
964
965	spin_lock_irqsave(&gcwq->lock, flags);
966	for_each_busy_worker(worker, i, pos, gcwq) {
967	if (worker->task != current)
968	continue;
969	spin_unlock_irqrestore(&gcwq->lock, flags);
970	/*
971	* I'm @worker, no locking necessary. See if @work
972	* is headed to the same workqueue.
973	*/
974	return worker->current_cwq->wq == wq;
975	}
976	spin_unlock_irqrestore(&gcwq->lock, flags);
977	}
978	return false;
979	}
980
981	static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
982	struct work_struct *work)
983	{
984	struct global_cwq *gcwq;
985	struct cpu_workqueue_struct *cwq;
986	struct list_head *worklist;
987	unsigned int work_flags;
988	unsigned long flags;
989
990	debug_work_activate(work);
991
992	/* if dying, only works from the same workqueue are allowed */
993	if (unlikely(wq->flags & WQ_DYING) &&
994	WARN_ON_ONCE(!is_chained_work(wq)))
995	return;
996
997	/* determine gcwq to use */
998	if (!(wq->flags & WQ_UNBOUND)) {
999	struct global_cwq *last_gcwq;
1000
1001	if (unlikely(cpu == WORK_CPU_UNBOUND))
1002	cpu = raw_smp_processor_id();
1003
1004	/*
1005	* It's multi cpu. If @wq is non-reentrant and @work
1006	* was previously on a different cpu, it might still
1007	* be running there, in which case the work needs to
1008	* be queued on that cpu to guarantee non-reentrance.
1009	*/
1010	gcwq = get_gcwq(cpu);
1011	if (wq->flags & WQ_NON_REENTRANT &&
1012	(last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1013	struct worker *worker;
1014
1015	spin_lock_irqsave(&last_gcwq->lock, flags);
1016
1017	worker = find_worker_executing_work(last_gcwq, work);
1018
1019	if (worker && worker->current_cwq->wq == wq)
1020	gcwq = last_gcwq;
1021	else {
1022	/* meh... not running there, queue here */
1023	spin_unlock_irqrestore(&last_gcwq->lock, flags);
1024	spin_lock_irqsave(&gcwq->lock, flags);
1025	}
1026	} else
1027	spin_lock_irqsave(&gcwq->lock, flags);
1028	} else {
1029	gcwq = get_gcwq(WORK_CPU_UNBOUND);
1030	spin_lock_irqsave(&gcwq->lock, flags);
1031	}
1032
1033	/* gcwq determined, get cwq and queue */
1034	cwq = get_cwq(gcwq->cpu, wq);
1035	trace_workqueue_queue_work(cpu, cwq, work);
1036
1037	BUG_ON(!list_empty(&work->entry));
1038
1039	cwq->nr_in_flight[cwq->work_color]++;
1040	work_flags = work_color_to_flags(cwq->work_color);
1041
1042	if (likely(cwq->nr_active < cwq->max_active)) {
1043	trace_workqueue_activate_work(work);
1044	cwq->nr_active++;
1045	worklist = gcwq_determine_ins_pos(gcwq, cwq);
1046	} else {
1047	work_flags \|= WORK_STRUCT_DELAYED;
1048	worklist = &cwq->delayed_works;
1049	}
1050
1051	insert_work(cwq, work, worklist, work_flags);
1052
1053	spin_unlock_irqrestore(&gcwq->lock, flags);
1054	}
1055
1056	/**
1057	* queue_work - queue work on a workqueue
1058	* @wq: workqueue to use
1059	* @work: work to queue
1060	*
1061	* Returns 0 if @work was already on a queue, non-zero otherwise.
1062	*
1063	* We queue the work to the CPU on which it was submitted, but if the CPU dies
1064	* it can be processed by another CPU.
1065	*/
1066	int queue_work(struct workqueue_struct wq, struct work_struct work)
1067	{
1068	int ret;
1069
1070	ret = queue_work_on(get_cpu(), wq, work);
1071	put_cpu();
1072
1073	return ret;
1074	}
1075	EXPORT_SYMBOL_GPL(queue_work);
1076
1077	/**
1078	* queue_work_on - queue work on specific cpu
1079	* @cpu: CPU number to execute work on
1080	* @wq: workqueue to use
1081	* @work: work to queue
1082	*
1083	* Returns 0 if @work was already on a queue, non-zero otherwise.
1084	*
1085	* We queue the work to a specific CPU, the caller must ensure it
1086	* can't go away.
1087	*/
1088	int
1089	queue_work_on(int cpu, struct workqueue_struct wq, struct work_struct work)
1090	{
1091	int ret = 0;
1092
1093	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1094	__queue_work(cpu, wq, work);
1095	ret = 1;
1096	}
1097	return ret;
1098	}
1099	EXPORT_SYMBOL_GPL(queue_work_on);
1100
1101	static void delayed_work_timer_fn(unsigned long __data)
1102	{
1103	struct delayed_work dwork = (struct delayed_work )__data;
1104	struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1105
1106	__queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1107	}
1108
1109	/**
1110	* queue_delayed_work - queue work on a workqueue after delay
1111	* @wq: workqueue to use
1112	* @dwork: delayable work to queue
1113	* @delay: number of jiffies to wait before queueing
1114	*
1115	* Returns 0 if @work was already on a queue, non-zero otherwise.
1116	*/
1117	int queue_delayed_work(struct workqueue_struct *wq,
1118	struct delayed_work *dwork, unsigned long delay)
1119	{
1120	if (delay == 0)
1121	return queue_work(wq, &dwork->work);
1122
1123	return queue_delayed_work_on(-1, wq, dwork, delay);
1124	}
1125	EXPORT_SYMBOL_GPL(queue_delayed_work);
1126
1127	/**
1128	* queue_delayed_work_on - queue work on specific CPU after delay
1129	* @cpu: CPU number to execute work on
1130	* @wq: workqueue to use
1131	* @dwork: work to queue
1132	* @delay: number of jiffies to wait before queueing
1133	*
1134	* Returns 0 if @work was already on a queue, non-zero otherwise.
1135	*/
1136	int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1137	struct delayed_work *dwork, unsigned long delay)
1138	{
1139	int ret = 0;
1140	struct timer_list *timer = &dwork->timer;
1141	struct work_struct *work = &dwork->work;
1142
1143	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1144	unsigned int lcpu;
1145
1146	BUG_ON(timer_pending(timer));
1147	BUG_ON(!list_empty(&work->entry));
1148
1149	timer_stats_timer_set_start_info(&dwork->timer);
1150
1151	/*
1152	* This stores cwq for the moment, for the timer_fn.
1153	* Note that the work's gcwq is preserved to allow
1154	* reentrance detection for delayed works.
1155	*/
1156	if (!(wq->flags & WQ_UNBOUND)) {
1157	struct global_cwq *gcwq = get_work_gcwq(work);
1158
1159	if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1160	lcpu = gcwq->cpu;
1161	else
1162	lcpu = raw_smp_processor_id();
1163	} else
1164	lcpu = WORK_CPU_UNBOUND;
1165
1166	set_work_cwq(work, get_cwq(lcpu, wq), 0);
1167
1168	timer->expires = jiffies + delay;
1169	timer->data = (unsigned long)dwork;
1170	timer->function = delayed_work_timer_fn;
1171
1172	if (unlikely(cpu >= 0))
1173	add_timer_on(timer, cpu);
1174	else
1175	add_timer(timer);
1176	ret = 1;
1177	}
1178	return ret;
1179	}
1180	EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1181
1182	/**
1183	* worker_enter_idle - enter idle state
1184	* @worker: worker which is entering idle state
1185	*
1186	* @worker is entering idle state. Update stats and idle timer if
1187	* necessary.
1188	*
1189	* LOCKING:
1190	* spin_lock_irq(gcwq->lock).
1191	*/
1192	static void worker_enter_idle(struct worker *worker)
1193	{
1194	struct global_cwq *gcwq = worker->gcwq;
1195
1196	BUG_ON(worker->flags & WORKER_IDLE);
1197	BUG_ON(!list_empty(&worker->entry) &&
1198	(worker->hentry.next \|\| worker->hentry.pprev));
1199
1200	/* can't use worker_set_flags(), also called from start_worker() */
1201	worker->flags \|= WORKER_IDLE;
1202	gcwq->nr_idle++;
1203	worker->last_active = jiffies;
1204
1205	/* idle_list is LIFO */
1206	list_add(&worker->entry, &gcwq->idle_list);
1207
1208	if (likely(!(worker->flags & WORKER_ROGUE))) {
1209	if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1210	mod_timer(&gcwq->idle_timer,
1211	jiffies + IDLE_WORKER_TIMEOUT);
1212	} else
1213	wake_up_all(&gcwq->trustee_wait);
1214
1215	/* sanity check nr_running */
1216	WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
1217	atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1218	}
1219
1220	/**
1221	* worker_leave_idle - leave idle state
1222	* @worker: worker which is leaving idle state
1223	*
1224	* @worker is leaving idle state. Update stats.
1225	*
1226	* LOCKING:
1227	* spin_lock_irq(gcwq->lock).
1228	*/
1229	static void worker_leave_idle(struct worker *worker)
1230	{
1231	struct global_cwq *gcwq = worker->gcwq;
1232
1233	BUG_ON(!(worker->flags & WORKER_IDLE));
1234	worker_clr_flags(worker, WORKER_IDLE);
1235	gcwq->nr_idle--;
1236	list_del_init(&worker->entry);
1237	}
1238
1239	/**
1240	* worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1241	* @worker: self
1242	*
1243	* Works which are scheduled while the cpu is online must at least be
1244	* scheduled to a worker which is bound to the cpu so that if they are
1245	* flushed from cpu callbacks while cpu is going down, they are
1246	* guaranteed to execute on the cpu.
1247	*
1248	* This function is to be used by rogue workers and rescuers to bind
1249	* themselves to the target cpu and may race with cpu going down or
1250	* coming online. kthread_bind() can't be used because it may put the
1251	* worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1252	* verbatim as it's best effort and blocking and gcwq may be
1253	* [dis]associated in the meantime.
1254	*
1255	* This function tries set_cpus_allowed() and locks gcwq and verifies
1256	* the binding against GCWQ_DISASSOCIATED which is set during
1257	* CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1258	* idle state or fetches works without dropping lock, it can guarantee
1259	* the scheduling requirement described in the first paragraph.
1260	*
1261	* CONTEXT:
1262	* Might sleep. Called without any lock but returns with gcwq->lock
1263	* held.
1264	*
1265	* RETURNS:
1266	* %true if the associated gcwq is online (@worker is successfully
1267	* bound), %false if offline.
1268	*/
1269	static bool worker_maybe_bind_and_lock(struct worker *worker)
1270	__acquires(&gcwq->lock)
1271	{
1272	struct global_cwq *gcwq = worker->gcwq;
1273	struct task_struct *task = worker->task;
1274
1275	while (true) {
1276	/*
1277	* The following call may fail, succeed or succeed
1278	* without actually migrating the task to the cpu if
1279	* it races with cpu hotunplug operation. Verify
1280	* against GCWQ_DISASSOCIATED.
1281	*/
1282	if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1283	set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1284
1285	spin_lock_irq(&gcwq->lock);
1286	if (gcwq->flags & GCWQ_DISASSOCIATED)
1287	return false;
1288	if (task_cpu(task) == gcwq->cpu &&
1289	cpumask_equal(&current->cpus_allowed,
1290	get_cpu_mask(gcwq->cpu)))
1291	return true;
1292	spin_unlock_irq(&gcwq->lock);
1293
1294	/*
1295	* We've raced with CPU hot[un]plug. Give it a breather
1296	* and retry migration. cond_resched() is required here;
1297	* otherwise, we might deadlock against cpu_stop trying to
1298	* bring down the CPU on non-preemptive kernel.
1299	*/
1300	cpu_relax();
1301	cond_resched();
1302	}
1303	}
1304
1305	/*
1306	* Function for worker->rebind_work used to rebind rogue busy workers
1307	* to the associated cpu which is coming back online. This is
1308	* scheduled by cpu up but can race with other cpu hotplug operations
1309	* and may be executed twice without intervening cpu down.
1310	*/
1311	static void worker_rebind_fn(struct work_struct *work)
1312	{
1313	struct worker *worker = container_of(work, struct worker, rebind_work);
1314	struct global_cwq *gcwq = worker->gcwq;
1315
1316	if (worker_maybe_bind_and_lock(worker))
1317	worker_clr_flags(worker, WORKER_REBIND);
1318
1319	spin_unlock_irq(&gcwq->lock);
1320	}
1321
1322	static struct worker *alloc_worker(void)
1323	{
1324	struct worker *worker;
1325
1326	worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1327	if (worker) {
1328	INIT_LIST_HEAD(&worker->entry);
1329	INIT_LIST_HEAD(&worker->scheduled);
1330	INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1331	/* on creation a worker is in !idle && prep state */
1332	worker->flags = WORKER_PREP;
1333	}
1334	return worker;
1335	}
1336
1337	/**
1338	* create_worker - create a new workqueue worker
1339	* @gcwq: gcwq the new worker will belong to
1340	* @bind: whether to set affinity to @cpu or not
1341	*
1342	* Create a new worker which is bound to @gcwq. The returned worker
1343	* can be started by calling start_worker() or destroyed using
1344	* destroy_worker().
1345	*
1346	* CONTEXT:
1347	* Might sleep. Does GFP_KERNEL allocations.
1348	*
1349	* RETURNS:
1350	* Pointer to the newly created worker.
1351	*/
1352	static struct worker create_worker(struct global_cwq gcwq, bool bind)
1353	{
1354	bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1355	struct worker *worker = NULL;
1356	int id = -1;
1357
1358	spin_lock_irq(&gcwq->lock);
1359	while (ida_get_new(&gcwq->worker_ida, &id)) {
1360	spin_unlock_irq(&gcwq->lock);
1361	if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1362	goto fail;
1363	spin_lock_irq(&gcwq->lock);
1364	}
1365	spin_unlock_irq(&gcwq->lock);
1366
1367	worker = alloc_worker();
1368	if (!worker)
1369	goto fail;
1370
1371	worker->gcwq = gcwq;
1372	worker->id = id;
1373
1374	if (!on_unbound_cpu)
1375	worker->task = kthread_create_on_node(worker_thread,
1376	worker,
1377	cpu_to_node(gcwq->cpu),
1378	"kworker/%u:%d", gcwq->cpu, id);
1379	else
1380	worker->task = kthread_create(worker_thread, worker,
1381	"kworker/u:%d", id);
1382	if (IS_ERR(worker->task))
1383	goto fail;
1384
1385	/*
1386	* A rogue worker will become a regular one if CPU comes
1387	* online later on. Make sure every worker has
1388	* PF_THREAD_BOUND set.
1389	*/
1390	if (bind && !on_unbound_cpu)
1391	kthread_bind(worker->task, gcwq->cpu);
1392	else {
1393	worker->task->flags \|= PF_THREAD_BOUND;
1394	if (on_unbound_cpu)
1395	worker->flags \|= WORKER_UNBOUND;
1396	}
1397
1398	return worker;
1399	fail:
1400	if (id >= 0) {
1401	spin_lock_irq(&gcwq->lock);
1402	ida_remove(&gcwq->worker_ida, id);
1403	spin_unlock_irq(&gcwq->lock);
1404	}
1405	kfree(worker);
1406	return NULL;
1407	}
1408
1409	/**
1410	* start_worker - start a newly created worker
1411	* @worker: worker to start
1412	*
1413	* Make the gcwq aware of @worker and start it.
1414	*
1415	* CONTEXT:
1416	* spin_lock_irq(gcwq->lock).
1417	*/
1418	static void start_worker(struct worker *worker)
1419	{
1420	worker->flags \|= WORKER_STARTED;
1421	worker->gcwq->nr_workers++;
1422	worker_enter_idle(worker);
1423	wake_up_process(worker->task);
1424	}
1425
1426	/**
1427	* destroy_worker - destroy a workqueue worker
1428	* @worker: worker to be destroyed
1429	*
1430	* Destroy @worker and adjust @gcwq stats accordingly.
1431	*
1432	* CONTEXT:
1433	* spin_lock_irq(gcwq->lock) which is released and regrabbed.
1434	*/
1435	static void destroy_worker(struct worker *worker)
1436	{
1437	struct global_cwq *gcwq = worker->gcwq;
1438	int id = worker->id;
1439
1440	/* sanity check frenzy */
1441	BUG_ON(worker->current_work);
1442	BUG_ON(!list_empty(&worker->scheduled));
1443
1444	if (worker->flags & WORKER_STARTED)
1445	gcwq->nr_workers--;
1446	if (worker->flags & WORKER_IDLE)
1447	gcwq->nr_idle--;
1448
1449	list_del_init(&worker->entry);
1450	worker->flags \|= WORKER_DIE;
1451
1452	spin_unlock_irq(&gcwq->lock);
1453
1454	kthread_stop(worker->task);
1455	kfree(worker);
1456
1457	spin_lock_irq(&gcwq->lock);
1458	ida_remove(&gcwq->worker_ida, id);
1459	}
1460
1461	static void idle_worker_timeout(unsigned long __gcwq)
1462	{
1463	struct global_cwq gcwq = (void )__gcwq;
1464
1465	spin_lock_irq(&gcwq->lock);
1466
1467	if (too_many_workers(gcwq)) {
1468	struct worker *worker;
1469	unsigned long expires;
1470
1471	/* idle_list is kept in LIFO order, check the last one */
1472	worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1473	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1474
1475	if (time_before(jiffies, expires))
1476	mod_timer(&gcwq->idle_timer, expires);
1477	else {
1478	/* it's been idle for too long, wake up manager */
1479	gcwq->flags \|= GCWQ_MANAGE_WORKERS;
1480	wake_up_worker(gcwq);
1481	}
1482	}
1483
1484	spin_unlock_irq(&gcwq->lock);
1485	}
1486
1487	static bool send_mayday(struct work_struct *work)
1488	{
1489	struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1490	struct workqueue_struct *wq = cwq->wq;
1491	unsigned int cpu;
1492
1493	if (!(wq->flags & WQ_RESCUER))
1494	return false;
1495
1496	/* mayday mayday mayday */
1497	cpu = cwq->gcwq->cpu;
1498	/* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1499	if (cpu == WORK_CPU_UNBOUND)
1500	cpu = 0;
1501	if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1502	wake_up_process(wq->rescuer->task);
1503	return true;
1504	}
1505
1506	static void gcwq_mayday_timeout(unsigned long __gcwq)
1507	{
1508	struct global_cwq gcwq = (void )__gcwq;
1509	struct work_struct *work;
1510
1511	spin_lock_irq(&gcwq->lock);
1512
1513	if (need_to_create_worker(gcwq)) {
1514	/*
1515	* We've been trying to create a new worker but
1516	* haven't been successful. We might be hitting an
1517	* allocation deadlock. Send distress signals to
1518	* rescuers.
1519	*/
1520	list_for_each_entry(work, &gcwq->worklist, entry)
1521	send_mayday(work);
1522	}
1523
1524	spin_unlock_irq(&gcwq->lock);
1525
1526	mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1527	}
1528
1529	/**
1530	* maybe_create_worker - create a new worker if necessary
1531	* @gcwq: gcwq to create a new worker for
1532	*
1533	* Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1534	* have at least one idle worker on return from this function. If
1535	* creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1536	* sent to all rescuers with works scheduled on @gcwq to resolve
1537	* possible allocation deadlock.
1538	*
1539	* On return, need_to_create_worker() is guaranteed to be false and
1540	* may_start_working() true.
1541	*
1542	* LOCKING:
1543	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
1544	* multiple times. Does GFP_KERNEL allocations. Called only from
1545	* manager.
1546	*
1547	* RETURNS:
1548	* false if no action was taken and gcwq->lock stayed locked, true
1549	* otherwise.
1550	*/
1551	static bool maybe_create_worker(struct global_cwq *gcwq)
1552	__releases(&gcwq->lock)
1553	__acquires(&gcwq->lock)
1554	{
1555	if (!need_to_create_worker(gcwq))
1556	return false;
1557	restart:
1558	spin_unlock_irq(&gcwq->lock);
1559
1560	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1561	mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1562
1563	while (true) {
1564	struct worker *worker;
1565
1566	worker = create_worker(gcwq, true);
1567	if (worker) {
1568	del_timer_sync(&gcwq->mayday_timer);
1569	spin_lock_irq(&gcwq->lock);
1570	start_worker(worker);
1571	BUG_ON(need_to_create_worker(gcwq));
1572	return true;
1573	}
1574
1575	if (!need_to_create_worker(gcwq))
1576	break;
1577
1578	__set_current_state(TASK_INTERRUPTIBLE);
1579	schedule_timeout(CREATE_COOLDOWN);
1580
1581	if (!need_to_create_worker(gcwq))
1582	break;
1583	}
1584
1585	del_timer_sync(&gcwq->mayday_timer);
1586	spin_lock_irq(&gcwq->lock);
1587	if (need_to_create_worker(gcwq))
1588	goto restart;
1589	return true;
1590	}
1591
1592	/**
1593	* maybe_destroy_worker - destroy workers which have been idle for a while
1594	* @gcwq: gcwq to destroy workers for
1595	*
1596	* Destroy @gcwq workers which have been idle for longer than
1597	* IDLE_WORKER_TIMEOUT.
1598	*
1599	* LOCKING:
1600	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
1601	* multiple times. Called only from manager.
1602	*
1603	* RETURNS:
1604	* false if no action was taken and gcwq->lock stayed locked, true
1605	* otherwise.
1606	*/
1607	static bool maybe_destroy_workers(struct global_cwq *gcwq)
1608	{
1609	bool ret = false;
1610
1611	while (too_many_workers(gcwq)) {
1612	struct worker *worker;
1613	unsigned long expires;
1614
1615	worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1616	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1617
1618	if (time_before(jiffies, expires)) {
1619	mod_timer(&gcwq->idle_timer, expires);
1620	break;
1621	}
1622
1623	destroy_worker(worker);
1624	ret = true;
1625	}
1626
1627	return ret;
1628	}
1629
1630	/**
1631	* manage_workers - manage worker pool
1632	* @worker: self
1633	*
1634	* Assume the manager role and manage gcwq worker pool @worker belongs
1635	* to. At any given time, there can be only zero or one manager per
1636	* gcwq. The exclusion is handled automatically by this function.
1637	*
1638	* The caller can safely start processing works on false return. On
1639	* true return, it's guaranteed that need_to_create_worker() is false
1640	* and may_start_working() is true.
1641	*
1642	* CONTEXT:
1643	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
1644	* multiple times. Does GFP_KERNEL allocations.
1645	*
1646	* RETURNS:
1647	* false if no action was taken and gcwq->lock stayed locked, true if
1648	* some action was taken.
1649	*/
1650	static bool manage_workers(struct worker *worker)
1651	{
1652	struct global_cwq *gcwq = worker->gcwq;
1653	bool ret = false;
1654
1655	if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1656	return ret;
1657
1658	gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1659	gcwq->flags \|= GCWQ_MANAGING_WORKERS;
1660
1661	/*
1662	* Destroy and then create so that may_start_working() is true
1663	* on return.
1664	*/
1665	ret \|= maybe_destroy_workers(gcwq);
1666	ret \|= maybe_create_worker(gcwq);
1667
1668	gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1669
1670	/*
1671	* The trustee might be waiting to take over the manager
1672	* position, tell it we're done.
1673	*/
1674	if (unlikely(gcwq->trustee))
1675	wake_up_all(&gcwq->trustee_wait);
1676
1677	return ret;
1678	}
1679
1680	/**
1681	* move_linked_works - move linked works to a list
1682	* @work: start of series of works to be scheduled
1683	* @head: target list to append @work to
1684	* @nextp: out paramter for nested worklist walking
1685	*
1686	* Schedule linked works starting from @work to @head. Work series to
1687	* be scheduled starts at @work and includes any consecutive work with
1688	* WORK_STRUCT_LINKED set in its predecessor.
1689	*
1690	* If @nextp is not NULL, it's updated to point to the next work of
1691	* the last scheduled work. This allows move_linked_works() to be
1692	* nested inside outer list_for_each_entry_safe().
1693	*
1694	* CONTEXT:
1695	* spin_lock_irq(gcwq->lock).
1696	*/
1697	static void move_linked_works(struct work_struct work, struct list_head head,
1698	struct work_struct **nextp)
1699	{
1700	struct work_struct *n;
1701
1702	/*
1703	* Linked worklist will always end before the end of the list,
1704	* use NULL for list head.
1705	*/
1706	list_for_each_entry_safe_from(work, n, NULL, entry) {
1707	list_move_tail(&work->entry, head);
1708	if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1709	break;
1710	}
1711
1712	/*
1713	* If we're already inside safe list traversal and have moved
1714	* multiple works to the scheduled queue, the next position
1715	* needs to be updated.
1716	*/
1717	if (nextp)
1718	*nextp = n;
1719	}
1720
1721	static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1722	{
1723	struct work_struct *work = list_first_entry(&cwq->delayed_works,
1724	struct work_struct, entry);
1725	struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1726
1727	trace_workqueue_activate_work(work);
1728	move_linked_works(work, pos, NULL);
1729	__clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1730	cwq->nr_active++;
1731	}
1732
1733	/**
1734	* cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1735	* @cwq: cwq of interest
1736	* @color: color of work which left the queue
1737	* @delayed: for a delayed work
1738	*
1739	* A work either has completed or is removed from pending queue,
1740	* decrement nr_in_flight of its cwq and handle workqueue flushing.
1741	*
1742	* CONTEXT:
1743	* spin_lock_irq(gcwq->lock).
1744	*/
1745	static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1746	bool delayed)
1747	{
1748	/* ignore uncolored works */
1749	if (color == WORK_NO_COLOR)
1750	return;
1751
1752	cwq->nr_in_flight[color]--;
1753
1754	if (!delayed) {
1755	cwq->nr_active--;
1756	if (!list_empty(&cwq->delayed_works)) {
1757	/* one down, submit a delayed one */
1758	if (cwq->nr_active < cwq->max_active)
1759	cwq_activate_first_delayed(cwq);
1760	}
1761	}
1762
1763	/* is flush in progress and are we at the flushing tip? */
1764	if (likely(cwq->flush_color != color))
1765	return;
1766
1767	/* are there still in-flight works? */
1768	if (cwq->nr_in_flight[color])
1769	return;
1770
1771	/* this cwq is done, clear flush_color */
1772	cwq->flush_color = -1;
1773
1774	/*
1775	* If this was the last cwq, wake up the first flusher. It
1776	* will handle the rest.
1777	*/
1778	if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1779	complete(&cwq->wq->first_flusher->done);
1780	}
1781
1782	/**
1783	* process_one_work - process single work
1784	* @worker: self
1785	* @work: work to process
1786	*
1787	* Process @work. This function contains all the logics necessary to
1788	* process a single work including synchronization against and
1789	* interaction with other workers on the same cpu, queueing and
1790	* flushing. As long as context requirement is met, any worker can
1791	* call this function to process a work.
1792	*
1793	* CONTEXT:
1794	* spin_lock_irq(gcwq->lock) which is released and regrabbed.
1795	*/
1796	static void process_one_work(struct worker worker, struct work_struct work)
1797	__releases(&gcwq->lock)
1798	__acquires(&gcwq->lock)
1799	{
1800	struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1801	struct global_cwq *gcwq = cwq->gcwq;
1802	struct hlist_head *bwh = busy_worker_head(gcwq, work);
1803	bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1804	work_func_t f = work->func;
1805	int work_color;
1806	struct worker *collision;
1807	#ifdef CONFIG_LOCKDEP
1808	/*
1809	* It is permissible to free the struct work_struct from
1810	* inside the function that is called from it, this we need to
1811	* take into account for lockdep too. To avoid bogus "held
1812	* lock freed" warnings as well as problems when looking into
1813	* work->lockdep_map, make a copy and use that here.
1814	*/
1815	struct lockdep_map lockdep_map = work->lockdep_map;
1816	#endif
1817	/*
1818	* A single work shouldn't be executed concurrently by
1819	* multiple workers on a single cpu. Check whether anyone is
1820	* already processing the work. If so, defer the work to the
1821	* currently executing one.
1822	*/
1823	collision = __find_worker_executing_work(gcwq, bwh, work);
1824	if (unlikely(collision)) {
1825	move_linked_works(work, &collision->scheduled, NULL);
1826	return;
1827	}
1828
1829	/* claim and process */
1830	debug_work_deactivate(work);
1831	hlist_add_head(&worker->hentry, bwh);
1832	worker->current_work = work;
1833	worker->current_cwq = cwq;
1834	work_color = get_work_color(work);
1835
1836	/* record the current cpu number in the work data and dequeue */
1837	set_work_cpu(work, gcwq->cpu);
1838	list_del_init(&work->entry);
1839
1840	/*
1841	* If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1842	* wake up another worker; otherwise, clear HIGHPRI_PENDING.
1843	*/
1844	if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1845	struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1846	struct work_struct, entry);
1847
1848	if (!list_empty(&gcwq->worklist) &&
1849	get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1850	wake_up_worker(gcwq);
1851	else
1852	gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1853	}
1854
1855	/*
1856	* CPU intensive works don't participate in concurrency
1857	* management. They're the scheduler's responsibility.
1858	*/
1859	if (unlikely(cpu_intensive))
1860	worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1861
1862	spin_unlock_irq(&gcwq->lock);
1863
1864	work_clear_pending(work);
1865	lock_map_acquire_read(&cwq->wq->lockdep_map);
1866	lock_map_acquire(&lockdep_map);
1867	trace_workqueue_execute_start(work);
1868	f(work);
1869	/*
1870	* While we must be careful to not use "work" after this, the trace
1871	* point will only record its address.
1872	*/
1873	trace_workqueue_execute_end(work);
1874	lock_map_release(&lockdep_map);
1875	lock_map_release(&cwq->wq->lockdep_map);
1876
1877	if (unlikely(in_atomic() \|\| lockdep_depth(current) > 0)) {
1878	printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1879	"%s/0x%08x/%d\n",
1880	current->comm, preempt_count(), task_pid_nr(current));
1881	printk(KERN_ERR " last function: ");
1882	print_symbol("%s\n", (unsigned long)f);
1883	debug_show_held_locks(current);
1884	dump_stack();
1885	}
1886
1887	spin_lock_irq(&gcwq->lock);
1888
1889	/* clear cpu intensive status */
1890	if (unlikely(cpu_intensive))
1891	worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1892
1893	/* we're done with it, release */
1894	hlist_del_init(&worker->hentry);
1895	worker->current_work = NULL;
1896	worker->current_cwq = NULL;
1897	cwq_dec_nr_in_flight(cwq, work_color, false);
1898	}
1899
1900	/**
1901	* process_scheduled_works - process scheduled works
1902	* @worker: self
1903	*
1904	* Process all scheduled works. Please note that the scheduled list
1905	* may change while processing a work, so this function repeatedly
1906	* fetches a work from the top and executes it.
1907	*
1908	* CONTEXT:
1909	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
1910	* multiple times.
1911	*/
1912	static void process_scheduled_works(struct worker *worker)
1913	{
1914	while (!list_empty(&worker->scheduled)) {
1915	struct work_struct *work = list_first_entry(&worker->scheduled,
1916	struct work_struct, entry);
1917	process_one_work(worker, work);
1918	}
1919	}
1920
1921	/**
1922	* worker_thread - the worker thread function
1923	* @__worker: self
1924	*
1925	* The gcwq worker thread function. There's a single dynamic pool of
1926	* these per each cpu. These workers process all works regardless of
1927	* their specific target workqueue. The only exception is works which
1928	* belong to workqueues with a rescuer which will be explained in
1929	* rescuer_thread().
1930	*/
1931	static int worker_thread(void *__worker)
1932	{
1933	struct worker *worker = __worker;
1934	struct global_cwq *gcwq = worker->gcwq;
1935
1936	/* tell the scheduler that this is a workqueue worker */
1937	worker->task->flags \|= PF_WQ_WORKER;
1938	woke_up:
1939	spin_lock_irq(&gcwq->lock);
1940
1941	/* DIE can be set only while we're idle, checking here is enough */
1942	if (worker->flags & WORKER_DIE) {
1943	spin_unlock_irq(&gcwq->lock);
1944	worker->task->flags &= ~PF_WQ_WORKER;
1945	return 0;
1946	}
1947
1948	worker_leave_idle(worker);
1949	recheck:
1950	/* no more worker necessary? */
1951	if (!need_more_worker(gcwq))
1952	goto sleep;
1953
1954	/* do we need to manage? */
1955	if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1956	goto recheck;
1957
1958	/*
1959	* ->scheduled list can only be filled while a worker is
1960	* preparing to process a work or actually processing it.
1961	* Make sure nobody diddled with it while I was sleeping.
1962	*/
1963	BUG_ON(!list_empty(&worker->scheduled));
1964
1965	/*
1966	* When control reaches this point, we're guaranteed to have
1967	* at least one idle worker or that someone else has already
1968	* assumed the manager role.
1969	*/
1970	worker_clr_flags(worker, WORKER_PREP);
1971
1972	do {
1973	struct work_struct *work =
1974	list_first_entry(&gcwq->worklist,
1975	struct work_struct, entry);
1976
1977	if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1978	/* optimization path, not strictly necessary */
1979	process_one_work(worker, work);
1980	if (unlikely(!list_empty(&worker->scheduled)))
1981	process_scheduled_works(worker);
1982	} else {
1983	move_linked_works(work, &worker->scheduled, NULL);
1984	process_scheduled_works(worker);
1985	}
1986	} while (keep_working(gcwq));
1987
1988	worker_set_flags(worker, WORKER_PREP, false);
1989	sleep:
1990	if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1991	goto recheck;
1992
1993	/*
1994	* gcwq->lock is held and there's no work to process and no
1995	* need to manage, sleep. Workers are woken up only while
1996	* holding gcwq->lock or from local cpu, so setting the
1997	* current state before releasing gcwq->lock is enough to
1998	* prevent losing any event.
1999	*/
2000	worker_enter_idle(worker);
2001	__set_current_state(TASK_INTERRUPTIBLE);
2002	spin_unlock_irq(&gcwq->lock);
2003	schedule();
2004	goto woke_up;
2005	}
2006
2007	/**
2008	* rescuer_thread - the rescuer thread function
2009	* @__wq: the associated workqueue
2010	*
2011	* Workqueue rescuer thread function. There's one rescuer for each
2012	* workqueue which has WQ_RESCUER set.
2013	*
2014	* Regular work processing on a gcwq may block trying to create a new
2015	* worker which uses GFP_KERNEL allocation which has slight chance of
2016	* developing into deadlock if some works currently on the same queue
2017	* need to be processed to satisfy the GFP_KERNEL allocation. This is
2018	* the problem rescuer solves.
2019	*
2020	* When such condition is possible, the gcwq summons rescuers of all
2021	* workqueues which have works queued on the gcwq and let them process
2022	* those works so that forward progress can be guaranteed.
2023	*
2024	* This should happen rarely.
2025	*/
2026	static int rescuer_thread(void *__wq)
2027	{
2028	struct workqueue_struct *wq = __wq;
2029	struct worker *rescuer = wq->rescuer;
2030	struct list_head *scheduled = &rescuer->scheduled;
2031	bool is_unbound = wq->flags & WQ_UNBOUND;
2032	unsigned int cpu;
2033
2034	set_user_nice(current, RESCUER_NICE_LEVEL);
2035	repeat:
2036	set_current_state(TASK_INTERRUPTIBLE);
2037
2038	if (kthread_should_stop())
2039	return 0;
2040
2041	/*
2042	* See whether any cpu is asking for help. Unbounded
2043	* workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2044	*/
2045	for_each_mayday_cpu(cpu, wq->mayday_mask) {
2046	unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2047	struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2048	struct global_cwq *gcwq = cwq->gcwq;
2049	struct work_struct work, n;
2050
2051	__set_current_state(TASK_RUNNING);
2052	mayday_clear_cpu(cpu, wq->mayday_mask);
2053
2054	/* migrate to the target cpu if possible */
2055	rescuer->gcwq = gcwq;
2056	worker_maybe_bind_and_lock(rescuer);
2057
2058	/*
2059	* Slurp in all works issued via this workqueue and
2060	* process'em.
2061	*/
2062	BUG_ON(!list_empty(&rescuer->scheduled));
2063	list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2064	if (get_work_cwq(work) == cwq)
2065	move_linked_works(work, scheduled, &n);
2066
2067	process_scheduled_works(rescuer);
2068
2069	/*
2070	* Leave this gcwq. If keep_working() is %true, notify a
2071	* regular worker; otherwise, we end up with 0 concurrency
2072	* and stalling the execution.
2073	*/
2074	if (keep_working(gcwq))
2075	wake_up_worker(gcwq);
2076
2077	spin_unlock_irq(&gcwq->lock);
2078	}
2079
2080	schedule();
2081	goto repeat;
2082	}
2083
2084	struct wq_barrier {
2085	struct work_struct work;
2086	struct completion done;
2087	};
2088
2089	static void wq_barrier_func(struct work_struct *work)
2090	{
2091	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2092	complete(&barr->done);
2093	}
2094
2095	/**
2096	* insert_wq_barrier - insert a barrier work
2097	* @cwq: cwq to insert barrier into
2098	* @barr: wq_barrier to insert
2099	* @target: target work to attach @barr to
2100	* @worker: worker currently executing @target, NULL if @target is not executing
2101	*
2102	* @barr is linked to @target such that @barr is completed only after
2103	* @target finishes execution. Please note that the ordering
2104	* guarantee is observed only with respect to @target and on the local
2105	* cpu.
2106	*
2107	* Currently, a queued barrier can't be canceled. This is because
2108	* try_to_grab_pending() can't determine whether the work to be
2109	* grabbed is at the head of the queue and thus can't clear LINKED
2110	* flag of the previous work while there must be a valid next work
2111	* after a work with LINKED flag set.
2112	*
2113	* Note that when @worker is non-NULL, @target may be modified
2114	* underneath us, so we can't reliably determine cwq from @target.
2115	*
2116	* CONTEXT:
2117	* spin_lock_irq(gcwq->lock).
2118	*/
2119	static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2120	struct wq_barrier *barr,
2121	struct work_struct target, struct worker worker)
2122	{
2123	struct list_head *head;
2124	unsigned int linked = 0;
2125
2126	/*
2127	* debugobject calls are safe here even with gcwq->lock locked
2128	* as we know for sure that this will not trigger any of the
2129	* checks and call back into the fixup functions where we
2130	* might deadlock.
2131	*/
2132	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2133	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2134	init_completion(&barr->done);
2135
2136	/*
2137	* If @target is currently being executed, schedule the
2138	* barrier to the worker; otherwise, put it after @target.
2139	*/
2140	if (worker)
2141	head = worker->scheduled.next;
2142	else {
2143	unsigned long *bits = work_data_bits(target);
2144
2145	head = target->entry.next;
2146	/* there can already be other linked works, inherit and set */
2147	linked = *bits & WORK_STRUCT_LINKED;
2148	__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2149	}
2150
2151	debug_work_activate(&barr->work);
2152	insert_work(cwq, &barr->work, head,
2153	work_color_to_flags(WORK_NO_COLOR) \| linked);
2154	}
2155
2156	/**
2157	* flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2158	* @wq: workqueue being flushed
2159	* @flush_color: new flush color, < 0 for no-op
2160	* @work_color: new work color, < 0 for no-op
2161	*
2162	* Prepare cwqs for workqueue flushing.
2163	*
2164	* If @flush_color is non-negative, flush_color on all cwqs should be
2165	* -1. If no cwq has in-flight commands at the specified color, all
2166	* cwq->flush_color's stay at -1 and %false is returned. If any cwq
2167	* has in flight commands, its cwq->flush_color is set to
2168	* @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2169	* wakeup logic is armed and %true is returned.
2170	*
2171	* The caller should have initialized @wq->first_flusher prior to
2172	* calling this function with non-negative @flush_color. If
2173	* @flush_color is negative, no flush color update is done and %false
2174	* is returned.
2175	*
2176	* If @work_color is non-negative, all cwqs should have the same
2177	* work_color which is previous to @work_color and all will be
2178	* advanced to @work_color.
2179	*
2180	* CONTEXT:
2181	* mutex_lock(wq->flush_mutex).
2182	*
2183	* RETURNS:
2184	* %true if @flush_color >= 0 and there's something to flush. %false
2185	* otherwise.
2186	*/
2187	static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2188	int flush_color, int work_color)
2189	{
2190	bool wait = false;
2191	unsigned int cpu;
2192
2193	if (flush_color >= 0) {
2194	BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2195	atomic_set(&wq->nr_cwqs_to_flush, 1);
2196	}
2197
2198	for_each_cwq_cpu(cpu, wq) {
2199	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2200	struct global_cwq *gcwq = cwq->gcwq;
2201
2202	spin_lock_irq(&gcwq->lock);
2203
2204	if (flush_color >= 0) {
2205	BUG_ON(cwq->flush_color != -1);
2206
2207	if (cwq->nr_in_flight[flush_color]) {
2208	cwq->flush_color = flush_color;
2209	atomic_inc(&wq->nr_cwqs_to_flush);
2210	wait = true;
2211	}
2212	}
2213
2214	if (work_color >= 0) {
2215	BUG_ON(work_color != work_next_color(cwq->work_color));
2216	cwq->work_color = work_color;
2217	}
2218
2219	spin_unlock_irq(&gcwq->lock);
2220	}
2221
2222	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2223	complete(&wq->first_flusher->done);
2224
2225	return wait;
2226	}
2227
2228	/**
2229	* flush_workqueue - ensure that any scheduled work has run to completion.
2230	* @wq: workqueue to flush
2231	*
2232	* Forces execution of the workqueue and blocks until its completion.
2233	* This is typically used in driver shutdown handlers.
2234	*
2235	* We sleep until all works which were queued on entry have been handled,
2236	* but we are not livelocked by new incoming ones.
2237	*/
2238	void flush_workqueue(struct workqueue_struct *wq)
2239	{
2240	struct wq_flusher this_flusher = {
2241	.list = LIST_HEAD_INIT(this_flusher.list),
2242	.flush_color = -1,
2243	.done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2244	};
2245	int next_color;
2246
2247	lock_map_acquire(&wq->lockdep_map);
2248	lock_map_release(&wq->lockdep_map);
2249
2250	mutex_lock(&wq->flush_mutex);
2251
2252	/*
2253	* Start-to-wait phase
2254	*/
2255	next_color = work_next_color(wq->work_color);
2256
2257	if (next_color != wq->flush_color) {
2258	/*
2259	* Color space is not full. The current work_color
2260	* becomes our flush_color and work_color is advanced
2261	* by one.
2262	*/
2263	BUG_ON(!list_empty(&wq->flusher_overflow));
2264	this_flusher.flush_color = wq->work_color;
2265	wq->work_color = next_color;
2266
2267	if (!wq->first_flusher) {
2268	/* no flush in progress, become the first flusher */
2269	BUG_ON(wq->flush_color != this_flusher.flush_color);
2270
2271	wq->first_flusher = &this_flusher;
2272
2273	if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2274	wq->work_color)) {
2275	/* nothing to flush, done */
2276	wq->flush_color = next_color;
2277	wq->first_flusher = NULL;
2278	goto out_unlock;
2279	}
2280	} else {
2281	/* wait in queue */
2282	BUG_ON(wq->flush_color == this_flusher.flush_color);
2283	list_add_tail(&this_flusher.list, &wq->flusher_queue);
2284	flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2285	}
2286	} else {
2287	/*
2288	* Oops, color space is full, wait on overflow queue.
2289	* The next flush completion will assign us
2290	* flush_color and transfer to flusher_queue.
2291	*/
2292	list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2293	}
2294
2295	mutex_unlock(&wq->flush_mutex);
2296
2297	wait_for_completion(&this_flusher.done);
2298
2299	/*
2300	* Wake-up-and-cascade phase
2301	*
2302	* First flushers are responsible for cascading flushes and
2303	* handling overflow. Non-first flushers can simply return.
2304	*/
2305	if (wq->first_flusher != &this_flusher)
2306	return;
2307
2308	mutex_lock(&wq->flush_mutex);
2309
2310	/* we might have raced, check again with mutex held */
2311	if (wq->first_flusher != &this_flusher)
2312	goto out_unlock;
2313
2314	wq->first_flusher = NULL;
2315
2316	BUG_ON(!list_empty(&this_flusher.list));
2317	BUG_ON(wq->flush_color != this_flusher.flush_color);
2318
2319	while (true) {
2320	struct wq_flusher next, tmp;
2321
2322	/* complete all the flushers sharing the current flush color */
2323	list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2324	if (next->flush_color != wq->flush_color)
2325	break;
2326	list_del_init(&next->list);
2327	complete(&next->done);
2328	}
2329
2330	BUG_ON(!list_empty(&wq->flusher_overflow) &&
2331	wq->flush_color != work_next_color(wq->work_color));
2332
2333	/* this flush_color is finished, advance by one */
2334	wq->flush_color = work_next_color(wq->flush_color);
2335
2336	/* one color has been freed, handle overflow queue */
2337	if (!list_empty(&wq->flusher_overflow)) {
2338	/*
2339	* Assign the same color to all overflowed
2340	* flushers, advance work_color and append to
2341	* flusher_queue. This is the start-to-wait
2342	* phase for these overflowed flushers.
2343	*/
2344	list_for_each_entry(tmp, &wq->flusher_overflow, list)
2345	tmp->flush_color = wq->work_color;
2346
2347	wq->work_color = work_next_color(wq->work_color);
2348
2349	list_splice_tail_init(&wq->flusher_overflow,
2350	&wq->flusher_queue);
2351	flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2352	}
2353
2354	if (list_empty(&wq->flusher_queue)) {
2355	BUG_ON(wq->flush_color != wq->work_color);
2356	break;
2357	}
2358
2359	/*
2360	* Need to flush more colors. Make the next flusher
2361	* the new first flusher and arm cwqs.
2362	*/
2363	BUG_ON(wq->flush_color == wq->work_color);
2364	BUG_ON(wq->flush_color != next->flush_color);
2365
2366	list_del_init(&next->list);
2367	wq->first_flusher = next;
2368
2369	if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2370	break;
2371
2372	/*
2373	* Meh... this color is already done, clear first
2374	* flusher and repeat cascading.
2375	*/
2376	wq->first_flusher = NULL;
2377	}
2378
2379	out_unlock:
2380	mutex_unlock(&wq->flush_mutex);
2381	}
2382	EXPORT_SYMBOL_GPL(flush_workqueue);
2383
2384	static bool start_flush_work(struct work_struct work, struct wq_barrier barr,
2385	bool wait_executing)
2386	{
2387	struct worker *worker = NULL;
2388	struct global_cwq *gcwq;
2389	struct cpu_workqueue_struct *cwq;
2390
2391	might_sleep();
2392	gcwq = get_work_gcwq(work);
2393	if (!gcwq)
2394	return false;
2395
2396	spin_lock_irq(&gcwq->lock);
2397	if (!list_empty(&work->entry)) {
2398	/*
2399	* See the comment near try_to_grab_pending()->smp_rmb().
2400	* If it was re-queued to a different gcwq under us, we
2401	* are not going to wait.
2402	*/
2403	smp_rmb();
2404	cwq = get_work_cwq(work);
2405	if (unlikely(!cwq \|\| gcwq != cwq->gcwq))
2406	goto already_gone;
2407	} else if (wait_executing) {
2408	worker = find_worker_executing_work(gcwq, work);
2409	if (!worker)
2410	goto already_gone;
2411	cwq = worker->current_cwq;
2412	} else
2413	goto already_gone;
2414
2415	insert_wq_barrier(cwq, barr, work, worker);
2416	spin_unlock_irq(&gcwq->lock);
2417
2418	/*
2419	* If @max_active is 1 or rescuer is in use, flushing another work
2420	* item on the same workqueue may lead to deadlock. Make sure the
2421	* flusher is not running on the same workqueue by verifying write
2422	* access.
2423	*/
2424	if (cwq->wq->saved_max_active == 1 \|\| cwq->wq->flags & WQ_RESCUER)
2425	lock_map_acquire(&cwq->wq->lockdep_map);
2426	else
2427	lock_map_acquire_read(&cwq->wq->lockdep_map);
2428	lock_map_release(&cwq->wq->lockdep_map);
2429
2430	return true;
2431	already_gone:
2432	spin_unlock_irq(&gcwq->lock);
2433	return false;
2434	}
2435
2436	/**
2437	* flush_work - wait for a work to finish executing the last queueing instance
2438	* @work: the work to flush
2439	*
2440	* Wait until @work has finished execution. This function considers
2441	* only the last queueing instance of @work. If @work has been
2442	* enqueued across different CPUs on a non-reentrant workqueue or on
2443	* multiple workqueues, @work might still be executing on return on
2444	* some of the CPUs from earlier queueing.
2445	*
2446	* If @work was queued only on a non-reentrant, ordered or unbound
2447	* workqueue, @work is guaranteed to be idle on return if it hasn't
2448	* been requeued since flush started.
2449	*
2450	* RETURNS:
2451	* %true if flush_work() waited for the work to finish execution,
2452	* %false if it was already idle.
2453	*/
2454	bool flush_work(struct work_struct *work)
2455	{
2456	struct wq_barrier barr;
2457
2458	if (start_flush_work(work, &barr, true)) {
2459	wait_for_completion(&barr.done);
2460	destroy_work_on_stack(&barr.work);
2461	return true;
2462	} else
2463	return false;
2464	}
2465	EXPORT_SYMBOL_GPL(flush_work);
2466
2467	static bool wait_on_cpu_work(struct global_cwq gcwq, struct work_struct work)
2468	{
2469	struct wq_barrier barr;
2470	struct worker *worker;
2471
2472	spin_lock_irq(&gcwq->lock);
2473
2474	worker = find_worker_executing_work(gcwq, work);
2475	if (unlikely(worker))
2476	insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2477
2478	spin_unlock_irq(&gcwq->lock);
2479
2480	if (unlikely(worker)) {
2481	wait_for_completion(&barr.done);
2482	destroy_work_on_stack(&barr.work);
2483	return true;
2484	} else
2485	return false;
2486	}
2487
2488	static bool wait_on_work(struct work_struct *work)
2489	{
2490	bool ret = false;
2491	int cpu;
2492
2493	might_sleep();
2494
2495	lock_map_acquire(&work->lockdep_map);
2496	lock_map_release(&work->lockdep_map);
2497
2498	for_each_gcwq_cpu(cpu)
2499	ret \|= wait_on_cpu_work(get_gcwq(cpu), work);
2500	return ret;
2501	}
2502
2503	/**
2504	* flush_work_sync - wait until a work has finished execution
2505	* @work: the work to flush
2506	*
2507	* Wait until @work has finished execution. On return, it's
2508	* guaranteed that all queueing instances of @work which happened
2509	* before this function is called are finished. In other words, if
2510	* @work hasn't been requeued since this function was called, @work is
2511	* guaranteed to be idle on return.
2512	*
2513	* RETURNS:
2514	* %true if flush_work_sync() waited for the work to finish execution,
2515	* %false if it was already idle.
2516	*/
2517	bool flush_work_sync(struct work_struct *work)
2518	{
2519	struct wq_barrier barr;
2520	bool pending, waited;
2521
2522	/* we'll wait for executions separately, queue barr only if pending */
2523	pending = start_flush_work(work, &barr, false);
2524
2525	/* wait for executions to finish */
2526	waited = wait_on_work(work);
2527
2528	/* wait for the pending one */
2529	if (pending) {
2530	wait_for_completion(&barr.done);
2531	destroy_work_on_stack(&barr.work);
2532	}
2533
2534	return pending \|\| waited;
2535	}
2536	EXPORT_SYMBOL_GPL(flush_work_sync);
2537
2538	/*
2539	* Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2540	* so this work can't be re-armed in any way.
2541	*/
2542	static int try_to_grab_pending(struct work_struct *work)
2543	{
2544	struct global_cwq *gcwq;
2545	int ret = -1;
2546
2547	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2548	return 0;
2549
2550	/*
2551	* The queueing is in progress, or it is already queued. Try to
2552	* steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2553	*/
2554	gcwq = get_work_gcwq(work);
2555	if (!gcwq)
2556	return ret;
2557
2558	spin_lock_irq(&gcwq->lock);
2559	if (!list_empty(&work->entry)) {
2560	/*
2561	* This work is queued, but perhaps we locked the wrong gcwq.
2562	* In that case we must see the new value after rmb(), see
2563	* insert_work()->wmb().
2564	*/
2565	smp_rmb();
2566	if (gcwq == get_work_gcwq(work)) {
2567	debug_work_deactivate(work);
2568	list_del_init(&work->entry);
2569	cwq_dec_nr_in_flight(get_work_cwq(work),
2570	get_work_color(work),
2571	*work_data_bits(work) & WORK_STRUCT_DELAYED);
2572	ret = 1;
2573	}
2574	}
2575	spin_unlock_irq(&gcwq->lock);
2576
2577	return ret;
2578	}
2579
2580	static bool __cancel_work_timer(struct work_struct *work,
2581	struct timer_list* timer)
2582	{
2583	int ret;
2584
2585	do {
2586	ret = (timer && likely(del_timer(timer)));
2587	if (!ret)
2588	ret = try_to_grab_pending(work);
2589	wait_on_work(work);
2590	} while (unlikely(ret < 0));
2591
2592	clear_work_data(work);
2593	return ret;
2594	}
2595
2596	/**
2597	* cancel_work_sync - cancel a work and wait for it to finish
2598	* @work: the work to cancel
2599	*
2600	* Cancel @work and wait for its execution to finish. This function
2601	* can be used even if the work re-queues itself or migrates to
2602	* another workqueue. On return from this function, @work is
2603	* guaranteed to be not pending or executing on any CPU.
2604	*
2605	* cancel_work_sync(&delayed_work->work) must not be used for
2606	* delayed_work's. Use cancel_delayed_work_sync() instead.
2607	*
2608	* The caller must ensure that the workqueue on which @work was last
2609	* queued can't be destroyed before this function returns.
2610	*
2611	* RETURNS:
2612	* %true if @work was pending, %false otherwise.
2613	*/
2614	bool cancel_work_sync(struct work_struct *work)
2615	{
2616	return __cancel_work_timer(work, NULL);
2617	}
2618	EXPORT_SYMBOL_GPL(cancel_work_sync);
2619
2620	/**
2621	* flush_delayed_work - wait for a dwork to finish executing the last queueing
2622	* @dwork: the delayed work to flush
2623	*
2624	* Delayed timer is cancelled and the pending work is queued for
2625	* immediate execution. Like flush_work(), this function only
2626	* considers the last queueing instance of @dwork.
2627	*
2628	* RETURNS:
2629	* %true if flush_work() waited for the work to finish execution,
2630	* %false if it was already idle.
2631	*/
2632	bool flush_delayed_work(struct delayed_work *dwork)
2633	{
2634	if (del_timer_sync(&dwork->timer))
2635	__queue_work(raw_smp_processor_id(),
2636	get_work_cwq(&dwork->work)->wq, &dwork->work);
2637	return flush_work(&dwork->work);
2638	}
2639	EXPORT_SYMBOL(flush_delayed_work);
2640
2641	/**
2642	* flush_delayed_work_sync - wait for a dwork to finish
2643	* @dwork: the delayed work to flush
2644	*
2645	* Delayed timer is cancelled and the pending work is queued for
2646	* execution immediately. Other than timer handling, its behavior
2647	* is identical to flush_work_sync().
2648	*
2649	* RETURNS:
2650	* %true if flush_work_sync() waited for the work to finish execution,
2651	* %false if it was already idle.
2652	*/
2653	bool flush_delayed_work_sync(struct delayed_work *dwork)
2654	{
2655	if (del_timer_sync(&dwork->timer))
2656	__queue_work(raw_smp_processor_id(),
2657	get_work_cwq(&dwork->work)->wq, &dwork->work);
2658	return flush_work_sync(&dwork->work);
2659	}
2660	EXPORT_SYMBOL(flush_delayed_work_sync);
2661
2662	/**
2663	* cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2664	* @dwork: the delayed work cancel
2665	*
2666	* This is cancel_work_sync() for delayed works.
2667	*
2668	* RETURNS:
2669	* %true if @dwork was pending, %false otherwise.
2670	*/
2671	bool cancel_delayed_work_sync(struct delayed_work *dwork)
2672	{
2673	return __cancel_work_timer(&dwork->work, &dwork->timer);
2674	}
2675	EXPORT_SYMBOL(cancel_delayed_work_sync);
2676
2677	/**
2678	* schedule_work - put work task in global workqueue
2679	* @work: job to be done
2680	*
2681	* Returns zero if @work was already on the kernel-global workqueue and
2682	* non-zero otherwise.
2683	*
2684	* This puts a job in the kernel-global workqueue if it was not already
2685	* queued and leaves it in the same position on the kernel-global
2686	* workqueue otherwise.
2687	*/
2688	int schedule_work(struct work_struct *work)
2689	{
2690	return queue_work(system_wq, work);
2691	}
2692	EXPORT_SYMBOL(schedule_work);
2693
2694	/*
2695	* schedule_work_on - put work task on a specific cpu
2696	* @cpu: cpu to put the work task on
2697	* @work: job to be done
2698	*
2699	* This puts a job on a specific cpu
2700	*/
2701	int schedule_work_on(int cpu, struct work_struct *work)
2702	{
2703	return queue_work_on(cpu, system_wq, work);
2704	}
2705	EXPORT_SYMBOL(schedule_work_on);
2706
2707	/**
2708	* schedule_delayed_work - put work task in global workqueue after delay
2709	* @dwork: job to be done
2710	* @delay: number of jiffies to wait or 0 for immediate execution
2711	*
2712	* After waiting for a given time this puts a job in the kernel-global
2713	* workqueue.
2714	*/
2715	int schedule_delayed_work(struct delayed_work *dwork,
2716	unsigned long delay)
2717	{
2718	return queue_delayed_work(system_wq, dwork, delay);
2719	}
2720	EXPORT_SYMBOL(schedule_delayed_work);
2721
2722	/**
2723	* schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2724	* @cpu: cpu to use
2725	* @dwork: job to be done
2726	* @delay: number of jiffies to wait
2727	*
2728	* After waiting for a given time this puts a job in the kernel-global
2729	* workqueue on the specified CPU.
2730	*/
2731	int schedule_delayed_work_on(int cpu,
2732	struct delayed_work *dwork, unsigned long delay)
2733	{
2734	return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2735	}
2736	EXPORT_SYMBOL(schedule_delayed_work_on);
2737
2738	/**
2739	* schedule_on_each_cpu - execute a function synchronously on each online CPU
2740	* @func: the function to call
2741	*
2742	* schedule_on_each_cpu() executes @func on each online CPU using the
2743	* system workqueue and blocks until all CPUs have completed.
2744	* schedule_on_each_cpu() is very slow.
2745	*
2746	* RETURNS:
2747	* 0 on success, -errno on failure.
2748	*/
2749	int schedule_on_each_cpu(work_func_t func)
2750	{
2751	int cpu;
2752	struct work_struct __percpu *works;
2753
2754	works = alloc_percpu(struct work_struct);
2755	if (!works)
2756	return -ENOMEM;
2757
2758	get_online_cpus();
2759
2760	for_each_online_cpu(cpu) {
2761	struct work_struct *work = per_cpu_ptr(works, cpu);
2762
2763	INIT_WORK(work, func);
2764	schedule_work_on(cpu, work);
2765	}
2766
2767	for_each_online_cpu(cpu)
2768	flush_work(per_cpu_ptr(works, cpu));
2769
2770	put_online_cpus();
2771	free_percpu(works);
2772	return 0;
2773	}
2774
2775	/**
2776	* flush_scheduled_work - ensure that any scheduled work has run to completion.
2777	*
2778	* Forces execution of the kernel-global workqueue and blocks until its
2779	* completion.
2780	*
2781	* Think twice before calling this function! It's very easy to get into
2782	* trouble if you don't take great care. Either of the following situations
2783	* will lead to deadlock:
2784	*
2785	* One of the work items currently on the workqueue needs to acquire
2786	* a lock held by your code or its caller.
2787	*
2788	* Your code is running in the context of a work routine.
2789	*
2790	* They will be detected by lockdep when they occur, but the first might not
2791	* occur very often. It depends on what work items are on the workqueue and
2792	* what locks they need, which you have no control over.
2793	*
2794	* In most situations flushing the entire workqueue is overkill; you merely
2795	* need to know that a particular work item isn't queued and isn't running.
2796	* In such cases you should use cancel_delayed_work_sync() or
2797	* cancel_work_sync() instead.
2798	*/
2799	void flush_scheduled_work(void)
2800	{
2801	flush_workqueue(system_wq);
2802	}
2803	EXPORT_SYMBOL(flush_scheduled_work);
2804
2805	/**
2806	* execute_in_process_context - reliably execute the routine with user context
2807	* @fn: the function to execute
2808	* @ew: guaranteed storage for the execute work structure (must
2809	* be available when the work executes)
2810	*
2811	* Executes the function immediately if process context is available,
2812	* otherwise schedules the function for delayed execution.
2813	*
2814	* Returns: 0 - function was executed
2815	* 1 - function was scheduled for execution
2816	*/
2817	int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2818	{
2819	if (!in_interrupt()) {
2820	fn(&ew->work);
2821	return 0;
2822	}
2823
2824	INIT_WORK(&ew->work, fn);
2825	schedule_work(&ew->work);
2826
2827	return 1;
2828	}
2829	EXPORT_SYMBOL_GPL(execute_in_process_context);
2830
2831	int keventd_up(void)
2832	{
2833	return system_wq != NULL;
2834	}
2835
2836	static int alloc_cwqs(struct workqueue_struct *wq)
2837	{
2838	/*
2839	* cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2840	* Make sure that the alignment isn't lower than that of
2841	* unsigned long long.
2842	*/
2843	const size_t size = sizeof(struct cpu_workqueue_struct);
2844	const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2845	__alignof__(unsigned long long));
2846	#ifdef CONFIG_SMP
2847	bool percpu = !(wq->flags & WQ_UNBOUND);
2848	#else
2849	bool percpu = false;
2850	#endif
2851
2852	if (percpu)
2853	wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2854	else {
2855	void *ptr;
2856
2857	/*
2858	* Allocate enough room to align cwq and put an extra
2859	* pointer at the end pointing back to the originally
2860	* allocated pointer which will be used for free.
2861	*/
2862	ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2863	if (ptr) {
2864	wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2865	(void *)(wq->cpu_wq.single + 1) = ptr;
2866	}
2867	}
2868
2869	/* just in case, make sure it's actually aligned
2870	* - this is affected by PERCPU() alignment in vmlinux.lds.S
2871	*/
2872	BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2873	return wq->cpu_wq.v ? 0 : -ENOMEM;
2874	}
2875
2876	static void free_cwqs(struct workqueue_struct *wq)
2877	{
2878	#ifdef CONFIG_SMP
2879	bool percpu = !(wq->flags & WQ_UNBOUND);
2880	#else
2881	bool percpu = false;
2882	#endif
2883
2884	if (percpu)
2885	free_percpu(wq->cpu_wq.pcpu);
2886	else if (wq->cpu_wq.single) {
2887	/* the pointer to free is stored right after the cwq */
2888	kfree((void *)(wq->cpu_wq.single + 1));
2889	}
2890	}
2891
2892	static int wq_clamp_max_active(int max_active, unsigned int flags,
2893	const char *name)
2894	{
2895	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2896
2897	if (max_active < 1 \|\| max_active > lim)
2898	printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2899	"is out of range, clamping between %d and %d\n",
2900	max_active, name, 1, lim);
2901
2902	return clamp_val(max_active, 1, lim);
2903	}
2904
2905	struct workqueue_struct __alloc_workqueue_key(const char name,
2906	unsigned int flags,
2907	int max_active,
2908	struct lock_class_key *key,
2909	const char *lock_name)
2910	{
2911	struct workqueue_struct *wq;
2912	unsigned int cpu;
2913
2914	/*
2915	* Workqueues which may be used during memory reclaim should
2916	* have a rescuer to guarantee forward progress.
2917	*/
2918	if (flags & WQ_MEM_RECLAIM)
2919	flags \|= WQ_RESCUER;
2920
2921	/*
2922	* Unbound workqueues aren't concurrency managed and should be
2923	* dispatched to workers immediately.
2924	*/
2925	if (flags & WQ_UNBOUND)
2926	flags \|= WQ_HIGHPRI;
2927
2928	max_active = max_active ?: WQ_DFL_ACTIVE;
2929	max_active = wq_clamp_max_active(max_active, flags, name);
2930
2931	wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2932	if (!wq)
2933	goto err;
2934
2935	wq->flags = flags;
2936	wq->saved_max_active = max_active;
2937	mutex_init(&wq->flush_mutex);
2938	atomic_set(&wq->nr_cwqs_to_flush, 0);
2939	INIT_LIST_HEAD(&wq->flusher_queue);
2940	INIT_LIST_HEAD(&wq->flusher_overflow);
2941
2942	wq->name = name;
2943	lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2944	INIT_LIST_HEAD(&wq->list);
2945
2946	if (alloc_cwqs(wq) < 0)
2947	goto err;
2948
2949	for_each_cwq_cpu(cpu, wq) {
2950	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2951	struct global_cwq *gcwq = get_gcwq(cpu);
2952
2953	BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
2954	cwq->gcwq = gcwq;
2955	cwq->wq = wq;
2956	cwq->flush_color = -1;
2957	cwq->max_active = max_active;
2958	INIT_LIST_HEAD(&cwq->delayed_works);
2959	}
2960
2961	if (flags & WQ_RESCUER) {
2962	struct worker *rescuer;
2963
2964	if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
2965	goto err;
2966
2967	wq->rescuer = rescuer = alloc_worker();
2968	if (!rescuer)
2969	goto err;
2970
2971	rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
2972	if (IS_ERR(rescuer->task))
2973	goto err;
2974
2975	rescuer->task->flags \|= PF_THREAD_BOUND;
2976	wake_up_process(rescuer->task);
2977	}
2978
2979	/*
2980	* workqueue_lock protects global freeze state and workqueues
2981	* list. Grab it, set max_active accordingly and add the new
2982	* workqueue to workqueues list.
2983	*/
2984	spin_lock(&workqueue_lock);
2985
2986	if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
2987	for_each_cwq_cpu(cpu, wq)
2988	get_cwq(cpu, wq)->max_active = 0;
2989
2990	list_add(&wq->list, &workqueues);
2991
2992	spin_unlock(&workqueue_lock);
2993
2994	return wq;
2995	err:
2996	if (wq) {
2997	free_cwqs(wq);
2998	free_mayday_mask(wq->mayday_mask);
2999	kfree(wq->rescuer);
3000	kfree(wq);
3001	}
3002	return NULL;
3003	}
3004	EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3005
3006	/**
3007	* destroy_workqueue - safely terminate a workqueue
3008	* @wq: target workqueue
3009	*
3010	* Safely destroy a workqueue. All work currently pending will be done first.
3011	*/
3012	void destroy_workqueue(struct workqueue_struct *wq)
3013	{
3014	unsigned int flush_cnt = 0;
3015	unsigned int cpu;
3016
3017	/*
3018	* Mark @wq dying and drain all pending works. Once WQ_DYING is
3019	* set, only chain queueing is allowed. IOW, only currently
3020	* pending or running work items on @wq can queue further work
3021	* items on it. @wq is flushed repeatedly until it becomes empty.
3022	* The number of flushing is detemined by the depth of chaining and
3023	* should be relatively short. Whine if it takes too long.
3024	*/
3025	wq->flags \|= WQ_DYING;
3026	reflush:
3027	flush_workqueue(wq);
3028
3029	for_each_cwq_cpu(cpu, wq) {
3030	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3031
3032	if (!cwq->nr_active && list_empty(&cwq->delayed_works))
3033	continue;
3034
3035	if (++flush_cnt == 10 \|\|
3036	(flush_cnt % 100 == 0 && flush_cnt <= 1000))
3037	printk(KERN_WARNING "workqueue %s: flush on "
3038	"destruction isn't complete after %u tries\n",
3039	wq->name, flush_cnt);
3040	goto reflush;
3041	}
3042
3043	/*
3044	* wq list is used to freeze wq, remove from list after
3045	* flushing is complete in case freeze races us.
3046	*/
3047	spin_lock(&workqueue_lock);
3048	list_del(&wq->list);
3049	spin_unlock(&workqueue_lock);
3050
3051	/* sanity check */
3052	for_each_cwq_cpu(cpu, wq) {
3053	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3054	int i;
3055
3056	for (i = 0; i < WORK_NR_COLORS; i++)
3057	BUG_ON(cwq->nr_in_flight[i]);
3058	BUG_ON(cwq->nr_active);
3059	BUG_ON(!list_empty(&cwq->delayed_works));
3060	}
3061
3062	if (wq->flags & WQ_RESCUER) {
3063	kthread_stop(wq->rescuer->task);
3064	free_mayday_mask(wq->mayday_mask);
3065	kfree(wq->rescuer);
3066	}
3067
3068	free_cwqs(wq);
3069	kfree(wq);
3070	}
3071	EXPORT_SYMBOL_GPL(destroy_workqueue);
3072
3073	/**
3074	* workqueue_set_max_active - adjust max_active of a workqueue
3075	* @wq: target workqueue
3076	* @max_active: new max_active value.
3077	*
3078	* Set max_active of @wq to @max_active.
3079	*
3080	* CONTEXT:
3081	* Don't call from IRQ context.
3082	*/
3083	void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3084	{
3085	unsigned int cpu;
3086
3087	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3088
3089	spin_lock(&workqueue_lock);
3090
3091	wq->saved_max_active = max_active;
3092
3093	for_each_cwq_cpu(cpu, wq) {
3094	struct global_cwq *gcwq = get_gcwq(cpu);
3095
3096	spin_lock_irq(&gcwq->lock);
3097
3098	if (!(wq->flags & WQ_FREEZABLE) \|\|
3099	!(gcwq->flags & GCWQ_FREEZING))
3100	get_cwq(gcwq->cpu, wq)->max_active = max_active;
3101
3102	spin_unlock_irq(&gcwq->lock);
3103	}
3104
3105	spin_unlock(&workqueue_lock);
3106	}
3107	EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3108
3109	/**
3110	* workqueue_congested - test whether a workqueue is congested
3111	* @cpu: CPU in question
3112	* @wq: target workqueue
3113	*
3114	* Test whether @wq's cpu workqueue for @cpu is congested. There is
3115	* no synchronization around this function and the test result is
3116	* unreliable and only useful as advisory hints or for debugging.
3117	*
3118	* RETURNS:
3119	* %true if congested, %false otherwise.
3120	*/
3121	bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3122	{
3123	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3124
3125	return !list_empty(&cwq->delayed_works);
3126	}
3127	EXPORT_SYMBOL_GPL(workqueue_congested);
3128
3129	/**
3130	* work_cpu - return the last known associated cpu for @work
3131	* @work: the work of interest
3132	*
3133	* RETURNS:
3134	* CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3135	*/
3136	unsigned int work_cpu(struct work_struct *work)
3137	{
3138	struct global_cwq *gcwq = get_work_gcwq(work);
3139
3140	return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3141	}
3142	EXPORT_SYMBOL_GPL(work_cpu);
3143
3144	/**
3145	* work_busy - test whether a work is currently pending or running
3146	* @work: the work to be tested
3147	*
3148	* Test whether @work is currently pending or running. There is no
3149	* synchronization around this function and the test result is
3150	* unreliable and only useful as advisory hints or for debugging.
3151	* Especially for reentrant wqs, the pending state might hide the
3152	* running state.
3153	*
3154	* RETURNS:
3155	* OR'd bitmask of WORK_BUSY_* bits.
3156	*/
3157	unsigned int work_busy(struct work_struct *work)
3158	{
3159	struct global_cwq *gcwq = get_work_gcwq(work);
3160	unsigned long flags;
3161	unsigned int ret = 0;
3162
3163	if (!gcwq)
3164	return false;
3165
3166	spin_lock_irqsave(&gcwq->lock, flags);
3167
3168	if (work_pending(work))
3169	ret \|= WORK_BUSY_PENDING;
3170	if (find_worker_executing_work(gcwq, work))
3171	ret \|= WORK_BUSY_RUNNING;
3172
3173	spin_unlock_irqrestore(&gcwq->lock, flags);
3174
3175	return ret;
3176	}
3177	EXPORT_SYMBOL_GPL(work_busy);
3178
3179	/*
3180	* CPU hotplug.
3181	*
3182	* There are two challenges in supporting CPU hotplug. Firstly, there
3183	* are a lot of assumptions on strong associations among work, cwq and
3184	* gcwq which make migrating pending and scheduled works very
3185	* difficult to implement without impacting hot paths. Secondly,
3186	* gcwqs serve mix of short, long and very long running works making
3187	* blocked draining impractical.
3188	*
3189	* This is solved by allowing a gcwq to be detached from CPU, running
3190	* it with unbound (rogue) workers and allowing it to be reattached
3191	* later if the cpu comes back online. A separate thread is created
3192	* to govern a gcwq in such state and is called the trustee of the
3193	* gcwq.
3194	*
3195	* Trustee states and their descriptions.
3196	*
3197	* START Command state used on startup. On CPU_DOWN_PREPARE, a
3198	* new trustee is started with this state.
3199	*
3200	* IN_CHARGE Once started, trustee will enter this state after
3201	* assuming the manager role and making all existing
3202	* workers rogue. DOWN_PREPARE waits for trustee to
3203	* enter this state. After reaching IN_CHARGE, trustee
3204	* tries to execute the pending worklist until it's empty
3205	* and the state is set to BUTCHER, or the state is set
3206	* to RELEASE.
3207	*
3208	* BUTCHER Command state which is set by the cpu callback after
3209	* the cpu has went down. Once this state is set trustee
3210	* knows that there will be no new works on the worklist
3211	* and once the worklist is empty it can proceed to
3212	* killing idle workers.
3213	*
3214	* RELEASE Command state which is set by the cpu callback if the
3215	* cpu down has been canceled or it has come online
3216	* again. After recognizing this state, trustee stops
3217	* trying to drain or butcher and clears ROGUE, rebinds
3218	* all remaining workers back to the cpu and releases
3219	* manager role.
3220	*
3221	* DONE Trustee will enter this state after BUTCHER or RELEASE
3222	* is complete.
3223	*
3224	* trustee CPU draining
3225	* took over down complete
3226	* START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3227	* \| \| ^
3228	* \| CPU is back online v return workers \|
3229	* ----------------> RELEASE --------------
3230	*/
3231
3232	/**
3233	* trustee_wait_event_timeout - timed event wait for trustee
3234	* @cond: condition to wait for
3235	* @timeout: timeout in jiffies
3236	*
3237	* wait_event_timeout() for trustee to use. Handles locking and
3238	* checks for RELEASE request.
3239	*
3240	* CONTEXT:
3241	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
3242	* multiple times. To be used by trustee.
3243	*
3244	* RETURNS:
3245	* Positive indicating left time if @cond is satisfied, 0 if timed
3246	* out, -1 if canceled.
3247	*/
3248	#define trustee_wait_event_timeout(cond, timeout) ({ \
3249	long __ret = (timeout); \
3250	while (!((cond) \|\| (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3251	__ret) { \
3252	spin_unlock_irq(&gcwq->lock); \
3253	__wait_event_timeout(gcwq->trustee_wait, (cond) \|\| \
3254	(gcwq->trustee_state == TRUSTEE_RELEASE), \
3255	__ret); \
3256	spin_lock_irq(&gcwq->lock); \
3257	} \
3258	gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3259	})
3260
3261	/**
3262	* trustee_wait_event - event wait for trustee
3263	* @cond: condition to wait for
3264	*
3265	* wait_event() for trustee to use. Automatically handles locking and
3266	* checks for CANCEL request.
3267	*
3268	* CONTEXT:
3269	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
3270	* multiple times. To be used by trustee.
3271	*
3272	* RETURNS:
3273	* 0 if @cond is satisfied, -1 if canceled.
3274	*/
3275	#define trustee_wait_event(cond) ({ \
3276	long __ret1; \
3277	__ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3278	__ret1 < 0 ? -1 : 0; \
3279	})
3280
3281	static int __cpuinit trustee_thread(void *__gcwq)
3282	{
3283	struct global_cwq *gcwq = __gcwq;
3284	struct worker *worker;
3285	struct work_struct *work;
3286	struct hlist_node *pos;
3287	long rc;
3288	int i;
3289
3290	BUG_ON(gcwq->cpu != smp_processor_id());
3291
3292	spin_lock_irq(&gcwq->lock);
3293	/*
3294	* Claim the manager position and make all workers rogue.
3295	* Trustee must be bound to the target cpu and can't be
3296	* cancelled.
3297	*/
3298	BUG_ON(gcwq->cpu != smp_processor_id());
3299	rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3300	BUG_ON(rc < 0);
3301
3302	gcwq->flags \|= GCWQ_MANAGING_WORKERS;
3303
3304	list_for_each_entry(worker, &gcwq->idle_list, entry)
3305	worker->flags \|= WORKER_ROGUE;
3306
3307	for_each_busy_worker(worker, i, pos, gcwq)
3308	worker->flags \|= WORKER_ROGUE;
3309
3310	/*
3311	* Call schedule() so that we cross rq->lock and thus can
3312	* guarantee sched callbacks see the rogue flag. This is
3313	* necessary as scheduler callbacks may be invoked from other
3314	* cpus.
3315	*/
3316	spin_unlock_irq(&gcwq->lock);
3317	schedule();
3318	spin_lock_irq(&gcwq->lock);
3319
3320	/*
3321	* Sched callbacks are disabled now. Zap nr_running. After
3322	* this, nr_running stays zero and need_more_worker() and
3323	* keep_working() are always true as long as the worklist is
3324	* not empty.
3325	*/
3326	atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3327
3328	spin_unlock_irq(&gcwq->lock);
3329	del_timer_sync(&gcwq->idle_timer);
3330	spin_lock_irq(&gcwq->lock);
3331
3332	/*
3333	* We're now in charge. Notify and proceed to drain. We need
3334	* to keep the gcwq running during the whole CPU down
3335	* procedure as other cpu hotunplug callbacks may need to
3336	* flush currently running tasks.
3337	*/
3338	gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3339	wake_up_all(&gcwq->trustee_wait);
3340
3341	/*
3342	* The original cpu is in the process of dying and may go away
3343	* anytime now. When that happens, we and all workers would
3344	* be migrated to other cpus. Try draining any left work. We
3345	* want to get it over with ASAP - spam rescuers, wake up as
3346	* many idlers as necessary and create new ones till the
3347	* worklist is empty. Note that if the gcwq is frozen, there
3348	* may be frozen works in freezable cwqs. Don't declare
3349	* completion while frozen.
3350	*/
3351	while (gcwq->nr_workers != gcwq->nr_idle \|\|
3352	gcwq->flags & GCWQ_FREEZING \|\|
3353	gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3354	int nr_works = 0;
3355
3356	list_for_each_entry(work, &gcwq->worklist, entry) {
3357	send_mayday(work);
3358	nr_works++;
3359	}
3360
3361	list_for_each_entry(worker, &gcwq->idle_list, entry) {
3362	if (!nr_works--)
3363	break;
3364	wake_up_process(worker->task);
3365	}
3366
3367	if (need_to_create_worker(gcwq)) {
3368	spin_unlock_irq(&gcwq->lock);
3369	worker = create_worker(gcwq, false);
3370	spin_lock_irq(&gcwq->lock);
3371	if (worker) {
3372	worker->flags \|= WORKER_ROGUE;
3373	start_worker(worker);
3374	}
3375	}
3376
3377	/* give a breather */
3378	if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3379	break;
3380	}
3381
3382	/*
3383	* Either all works have been scheduled and cpu is down, or
3384	* cpu down has already been canceled. Wait for and butcher
3385	* all workers till we're canceled.
3386	*/
3387	do {
3388	rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3389	while (!list_empty(&gcwq->idle_list))
3390	destroy_worker(list_first_entry(&gcwq->idle_list,
3391	struct worker, entry));
3392	} while (gcwq->nr_workers && rc >= 0);
3393
3394	/*
3395	* At this point, either draining has completed and no worker
3396	* is left, or cpu down has been canceled or the cpu is being
3397	* brought back up. There shouldn't be any idle one left.
3398	* Tell the remaining busy ones to rebind once it finishes the
3399	* currently scheduled works by scheduling the rebind_work.
3400	*/
3401	WARN_ON(!list_empty(&gcwq->idle_list));
3402
3403	for_each_busy_worker(worker, i, pos, gcwq) {
3404	struct work_struct *rebind_work = &worker->rebind_work;
3405
3406	/*
3407	* Rebind_work may race with future cpu hotplug
3408	* operations. Use a separate flag to mark that
3409	* rebinding is scheduled.
3410	*/
3411	worker->flags \|= WORKER_REBIND;
3412	worker->flags &= ~WORKER_ROGUE;
3413
3414	/* queue rebind_work, wq doesn't matter, use the default one */
3415	if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3416	work_data_bits(rebind_work)))
3417	continue;
3418
3419	debug_work_activate(rebind_work);
3420	insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3421	worker->scheduled.next,
3422	work_color_to_flags(WORK_NO_COLOR));
3423	}
3424
3425	/* relinquish manager role */
3426	gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3427
3428	/* notify completion */
3429	gcwq->trustee = NULL;
3430	gcwq->trustee_state = TRUSTEE_DONE;
3431	wake_up_all(&gcwq->trustee_wait);
3432	spin_unlock_irq(&gcwq->lock);
3433	return 0;
3434	}
3435
3436	/**
3437	* wait_trustee_state - wait for trustee to enter the specified state
3438	* @gcwq: gcwq the trustee of interest belongs to
3439	* @state: target state to wait for
3440	*
3441	* Wait for the trustee to reach @state. DONE is already matched.
3442	*
3443	* CONTEXT:
3444	* spin_lock_irq(gcwq->lock) which may be released and regrabbed
3445	* multiple times. To be used by cpu_callback.
3446	*/
3447	static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3448	__releases(&gcwq->lock)
3449	__acquires(&gcwq->lock)
3450	{
3451	if (!(gcwq->trustee_state == state \|\|
3452	gcwq->trustee_state == TRUSTEE_DONE)) {
3453	spin_unlock_irq(&gcwq->lock);
3454	__wait_event(gcwq->trustee_wait,
3455	gcwq->trustee_state == state \|\|
3456	gcwq->trustee_state == TRUSTEE_DONE);
3457	spin_lock_irq(&gcwq->lock);
3458	}
3459	}
3460
3461	static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3462	unsigned long action,
3463	void *hcpu)
3464	{
3465	unsigned int cpu = (unsigned long)hcpu;
3466	struct global_cwq *gcwq = get_gcwq(cpu);
3467	struct task_struct *new_trustee = NULL;
3468	struct worker *uninitialized_var(new_worker);
3469	unsigned long flags;
3470
3471	action &= ~CPU_TASKS_FROZEN;
3472
3473	switch (action) {
3474	case CPU_DOWN_PREPARE:
3475	new_trustee = kthread_create(trustee_thread, gcwq,
3476	"workqueue_trustee/%d\n", cpu);
3477	if (IS_ERR(new_trustee))
3478	return notifier_from_errno(PTR_ERR(new_trustee));
3479	kthread_bind(new_trustee, cpu);
3480	/* fall through */
3481	case CPU_UP_PREPARE:
3482	BUG_ON(gcwq->first_idle);
3483	new_worker = create_worker(gcwq, false);
3484	if (!new_worker) {
3485	if (new_trustee)
3486	kthread_stop(new_trustee);
3487	return NOTIFY_BAD;
3488	}
3489	}
3490
3491	/* some are called w/ irq disabled, don't disturb irq status */
3492	spin_lock_irqsave(&gcwq->lock, flags);
3493
3494	switch (action) {
3495	case CPU_DOWN_PREPARE:
3496	/* initialize trustee and tell it to acquire the gcwq */
3497	BUG_ON(gcwq->trustee \|\| gcwq->trustee_state != TRUSTEE_DONE);
3498	gcwq->trustee = new_trustee;
3499	gcwq->trustee_state = TRUSTEE_START;
3500	wake_up_process(gcwq->trustee);
3501	wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3502	/* fall through */
3503	case CPU_UP_PREPARE:
3504	BUG_ON(gcwq->first_idle);
3505	gcwq->first_idle = new_worker;
3506	break;
3507
3508	case CPU_DYING:
3509	/*
3510	* Before this, the trustee and all workers except for
3511	* the ones which are still executing works from
3512	* before the last CPU down must be on the cpu. After
3513	* this, they'll all be diasporas.
3514	*/
3515	gcwq->flags \|= GCWQ_DISASSOCIATED;
3516	break;
3517
3518	case CPU_POST_DEAD:
3519	gcwq->trustee_state = TRUSTEE_BUTCHER;
3520	/* fall through */
3521	case CPU_UP_CANCELED:
3522	destroy_worker(gcwq->first_idle);
3523	gcwq->first_idle = NULL;
3524	break;
3525
3526	case CPU_DOWN_FAILED:
3527	case CPU_ONLINE:
3528	gcwq->flags &= ~GCWQ_DISASSOCIATED;
3529	if (gcwq->trustee_state != TRUSTEE_DONE) {
3530	gcwq->trustee_state = TRUSTEE_RELEASE;
3531	wake_up_process(gcwq->trustee);
3532	wait_trustee_state(gcwq, TRUSTEE_DONE);
3533	}
3534
3535	/*
3536	* Trustee is done and there might be no worker left.
3537	* Put the first_idle in and request a real manager to
3538	* take a look.
3539	*/
3540	spin_unlock_irq(&gcwq->lock);
3541	kthread_bind(gcwq->first_idle->task, cpu);
3542	spin_lock_irq(&gcwq->lock);
3543	gcwq->flags \|= GCWQ_MANAGE_WORKERS;
3544	start_worker(gcwq->first_idle);
3545	gcwq->first_idle = NULL;
3546	break;
3547	}
3548
3549	spin_unlock_irqrestore(&gcwq->lock, flags);
3550
3551	return notifier_from_errno(0);
3552	}
3553
3554	#ifdef CONFIG_SMP
3555
3556	struct work_for_cpu {
3557	struct completion completion;
3558	long (fn)(void );
3559	void *arg;
3560	long ret;
3561	};
3562
3563	static int do_work_for_cpu(void *_wfc)
3564	{
3565	struct work_for_cpu *wfc = _wfc;
3566	wfc->ret = wfc->fn(wfc->arg);
3567	complete(&wfc->completion);
3568	return 0;
3569	}
3570
3571	/**
3572	* work_on_cpu - run a function in user context on a particular cpu
3573	* @cpu: the cpu to run on
3574	* @fn: the function to run
3575	* @arg: the function arg
3576	*
3577	* This will return the value @fn returns.
3578	* It is up to the caller to ensure that the cpu doesn't go offline.
3579	* The caller must not hold any locks which would prevent @fn from completing.
3580	*/
3581	long work_on_cpu(unsigned int cpu, long (fn)(void ), void *arg)
3582	{
3583	struct task_struct *sub_thread;
3584	struct work_for_cpu wfc = {
3585	.completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3586	.fn = fn,
3587	.arg = arg,
3588	};
3589
3590	sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3591	if (IS_ERR(sub_thread))
3592	return PTR_ERR(sub_thread);
3593	kthread_bind(sub_thread, cpu);
3594	wake_up_process(sub_thread);
3595	wait_for_completion(&wfc.completion);
3596	return wfc.ret;
3597	}
3598	EXPORT_SYMBOL_GPL(work_on_cpu);
3599	#endif /* CONFIG_SMP */
3600
3601	#ifdef CONFIG_FREEZER
3602
3603	/**
3604	* freeze_workqueues_begin - begin freezing workqueues
3605	*
3606	* Start freezing workqueues. After this function returns, all freezable
3607	* workqueues will queue new works to their frozen_works list instead of
3608	* gcwq->worklist.
3609	*
3610	* CONTEXT:
3611	* Grabs and releases workqueue_lock and gcwq->lock's.
3612	*/
3613	void freeze_workqueues_begin(void)
3614	{
3615	unsigned int cpu;
3616
3617	spin_lock(&workqueue_lock);
3618
3619	BUG_ON(workqueue_freezing);
3620	workqueue_freezing = true;
3621
3622	for_each_gcwq_cpu(cpu) {
3623	struct global_cwq *gcwq = get_gcwq(cpu);
3624	struct workqueue_struct *wq;
3625
3626	spin_lock_irq(&gcwq->lock);
3627
3628	BUG_ON(gcwq->flags & GCWQ_FREEZING);
3629	gcwq->flags \|= GCWQ_FREEZING;
3630
3631	list_for_each_entry(wq, &workqueues, list) {
3632	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3633
3634	if (cwq && wq->flags & WQ_FREEZABLE)
3635	cwq->max_active = 0;
3636	}
3637
3638	spin_unlock_irq(&gcwq->lock);
3639	}
3640
3641	spin_unlock(&workqueue_lock);
3642	}
3643
3644	/**
3645	* freeze_workqueues_busy - are freezable workqueues still busy?
3646	*
3647	* Check whether freezing is complete. This function must be called
3648	* between freeze_workqueues_begin() and thaw_workqueues().
3649	*
3650	* CONTEXT:
3651	* Grabs and releases workqueue_lock.
3652	*
3653	* RETURNS:
3654	* %true if some freezable workqueues are still busy. %false if freezing
3655	* is complete.
3656	*/
3657	bool freeze_workqueues_busy(void)
3658	{
3659	unsigned int cpu;
3660	bool busy = false;
3661
3662	spin_lock(&workqueue_lock);
3663
3664	BUG_ON(!workqueue_freezing);
3665
3666	for_each_gcwq_cpu(cpu) {
3667	struct workqueue_struct *wq;
3668	/*
3669	* nr_active is monotonically decreasing. It's safe
3670	* to peek without lock.
3671	*/
3672	list_for_each_entry(wq, &workqueues, list) {
3673	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3674
3675	if (!cwq \|\| !(wq->flags & WQ_FREEZABLE))
3676	continue;
3677
3678	BUG_ON(cwq->nr_active < 0);
3679	if (cwq->nr_active) {
3680	busy = true;
3681	goto out_unlock;
3682	}
3683	}
3684	}
3685	out_unlock:
3686	spin_unlock(&workqueue_lock);
3687	return busy;
3688	}
3689
3690	/**
3691	* thaw_workqueues - thaw workqueues
3692	*
3693	* Thaw workqueues. Normal queueing is restored and all collected
3694	* frozen works are transferred to their respective gcwq worklists.
3695	*
3696	* CONTEXT:
3697	* Grabs and releases workqueue_lock and gcwq->lock's.
3698	*/
3699	void thaw_workqueues(void)
3700	{
3701	unsigned int cpu;
3702
3703	spin_lock(&workqueue_lock);
3704
3705	if (!workqueue_freezing)
3706	goto out_unlock;
3707
3708	for_each_gcwq_cpu(cpu) {
3709	struct global_cwq *gcwq = get_gcwq(cpu);
3710	struct workqueue_struct *wq;
3711
3712	spin_lock_irq(&gcwq->lock);
3713
3714	BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3715	gcwq->flags &= ~GCWQ_FREEZING;
3716
3717	list_for_each_entry(wq, &workqueues, list) {
3718	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3719
3720	if (!cwq \|\| !(wq->flags & WQ_FREEZABLE))
3721	continue;
3722
3723	/* restore max_active and repopulate worklist */
3724	cwq->max_active = wq->saved_max_active;
3725
3726	while (!list_empty(&cwq->delayed_works) &&
3727	cwq->nr_active < cwq->max_active)
3728	cwq_activate_first_delayed(cwq);
3729	}
3730
3731	wake_up_worker(gcwq);
3732
3733	spin_unlock_irq(&gcwq->lock);
3734	}
3735
3736	workqueue_freezing = false;
3737	out_unlock:
3738	spin_unlock(&workqueue_lock);
3739	}
3740	#endif /* CONFIG_FREEZER */
3741
3742	static int __init init_workqueues(void)
3743	{
3744	unsigned int cpu;
3745	int i;
3746
3747	cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3748
3749	/* initialize gcwqs */
3750	for_each_gcwq_cpu(cpu) {
3751	struct global_cwq *gcwq = get_gcwq(cpu);
3752
3753	spin_lock_init(&gcwq->lock);
3754	INIT_LIST_HEAD(&gcwq->worklist);
3755	gcwq->cpu = cpu;
3756	gcwq->flags \|= GCWQ_DISASSOCIATED;
3757
3758	INIT_LIST_HEAD(&gcwq->idle_list);
3759	for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3760	INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3761
3762	init_timer_deferrable(&gcwq->idle_timer);
3763	gcwq->idle_timer.function = idle_worker_timeout;
3764	gcwq->idle_timer.data = (unsigned long)gcwq;
3765
3766	setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3767	(unsigned long)gcwq);
3768
3769	ida_init(&gcwq->worker_ida);
3770
3771	gcwq->trustee_state = TRUSTEE_DONE;
3772	init_waitqueue_head(&gcwq->trustee_wait);
3773	}
3774
3775	/* create the initial worker */
3776	for_each_online_gcwq_cpu(cpu) {
3777	struct global_cwq *gcwq = get_gcwq(cpu);
3778	struct worker *worker;
3779
3780	if (cpu != WORK_CPU_UNBOUND)
3781	gcwq->flags &= ~GCWQ_DISASSOCIATED;
3782	worker = create_worker(gcwq, true);
3783	BUG_ON(!worker);
3784	spin_lock_irq(&gcwq->lock);
3785	start_worker(worker);
3786	spin_unlock_irq(&gcwq->lock);
3787	}
3788
3789	system_wq = alloc_workqueue("events", 0, 0);
3790	system_long_wq = alloc_workqueue("events_long", 0, 0);
3791	system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3792	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3793	WQ_UNBOUND_MAX_ACTIVE);
3794	system_freezable_wq = alloc_workqueue("events_freezable",
3795	WQ_FREEZABLE, 0);
3796	BUG_ON(!system_wq \|\| !system_long_wq \|\| !system_nrt_wq \|\|
3797	!system_unbound_wq \|\| !system_freezable_wq);
3798	return 0;
3799	}
3800	early_initcall(init_workqueues);