linux - kernel/workqueue.c

5541 lines · 152.2 KiB

1	/*
2	* kernel/workqueue.c - generic async execution with shared worker pool
3	*
4	* Copyright (C) 2002 Ingo Molnar
5	*
6	* Derived from the taskqueue/keventd code by:
7	* David Woodhouse <[email protected]>
8	* Andrew Morton
9	* Kai Petzke <[email protected]>
10	* Theodore Ts'o <[email protected]>
11	*
12	* Made to use alloc_percpu by Christoph Lameter.
13	*
14	* Copyright (C) 2010 SUSE Linux Products GmbH
15	* Copyright (C) 2010 Tejun Heo <[email protected]>
16	*
17	* This is the generic async execution mechanism. Work items as are
18	* executed in process context. The worker pool is shared and
19	* automatically managed. There are two worker pools for each CPU (one for
20	* normal work items and the other for high priority ones) and some extra
21	* pools for workqueues which are not bound to any specific CPU - the
22	* number of these backing pools is dynamic.
23	*
24	* Please read Documentation/workqueue.txt for details.
25	*/
26
27	#include <linux/export.h>
28	#include <linux/kernel.h>
29	#include <linux/sched.h>
30	#include <linux/init.h>
31	#include <linux/signal.h>
32	#include <linux/completion.h>
33	#include <linux/workqueue.h>
34	#include <linux/slab.h>
35	#include <linux/cpu.h>
36	#include <linux/notifier.h>
37	#include <linux/kthread.h>
38	#include <linux/hardirq.h>
39	#include <linux/mempolicy.h>
40	#include <linux/freezer.h>
41	#include <linux/kallsyms.h>
42	#include <linux/debug_locks.h>
43	#include <linux/lockdep.h>
44	#include <linux/idr.h>
45	#include <linux/jhash.h>
46	#include <linux/hashtable.h>
47	#include <linux/rculist.h>
48	#include <linux/nodemask.h>
49	#include <linux/moduleparam.h>
50	#include <linux/uaccess.h>
51
52	#include "workqueue_internal.h"
53
54	enum {
55	/*
56	* worker_pool flags
57	*
58	* A bound pool is either associated or disassociated with its CPU.
59	* While associated (!DISASSOCIATED), all workers are bound to the
60	* CPU and none has %WORKER_UNBOUND set and concurrency management
61	* is in effect.
62	*
63	* While DISASSOCIATED, the cpu may be offline and all workers have
64	* %WORKER_UNBOUND set and concurrency management disabled, and may
65	* be executing on any CPU. The pool behaves as an unbound one.
66	*
67	* Note that DISASSOCIATED should be flipped only while holding
68	* attach_mutex to avoid changing binding state while
69	* worker_attach_to_pool() is in progress.
70	*/
71	POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
72
73	/* worker flags */
74	WORKER_DIE = 1 << 1, /* die die die */
75	WORKER_IDLE = 1 << 2, /* is idle */
76	WORKER_PREP = 1 << 3, /* preparing to run works */
77	WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78	WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79	WORKER_REBOUND = 1 << 8, /* worker was rebound */
80
81	WORKER_NOT_RUNNING = WORKER_PREP \| WORKER_CPU_INTENSIVE \|
82	WORKER_UNBOUND \| WORKER_REBOUND,
83
84	NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
85
86	UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87	BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
88
89	MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90	IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91
92	MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93	/* call for help after 10ms
94	(min two ticks) */
95	MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96	CREATE_COOLDOWN = HZ, /* time to breath after fail */
97
98	/*
99	* Rescue workers are used only on emergencies and shared by
100	* all cpus. Give MIN_NICE.
101	*/
102	RESCUER_NICE_LEVEL = MIN_NICE,
103	HIGHPRI_NICE_LEVEL = MIN_NICE,
104
105	WQ_NAME_LEN = 24,
106	};
107
108	/*
109	* Structure fields follow one of the following exclusion rules.
110	*
111	* I: Modifiable by initialization/destruction paths and read-only for
112	* everyone else.
113	*
114	* P: Preemption protected. Disabling preemption is enough and should
115	* only be modified and accessed from the local cpu.
116	*
117	* L: pool->lock protected. Access with pool->lock held.
118	*
119	* X: During normal operation, modification requires pool->lock and should
120	* be done only from local cpu. Either disabling preemption on local
121	* cpu or grabbing pool->lock is enough for read access. If
122	* POOL_DISASSOCIATED is set, it's identical to L.
123	*
124	* A: pool->attach_mutex protected.
125	*
126	* PL: wq_pool_mutex protected.
127	*
128	* PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
129	*
130	* PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
131	*
132	* PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133	* sched-RCU for reads.
134	*
135	* WQ: wq->mutex protected.
136	*
137	* WR: wq->mutex protected for writes. Sched-RCU protected for reads.
138	*
139	* MD: wq_mayday_lock protected.
140	*/
141
142	/* struct worker is defined in workqueue_internal.h */
143
144	struct worker_pool {
145	spinlock_t lock; /* the pool lock */
146	int cpu; /* I: the associated cpu */
147	int node; /* I: the associated node ID */
148	int id; /* I: pool ID */
149	unsigned int flags; /* X: flags */
150
151	unsigned long watchdog_ts; /* L: watchdog timestamp */
152
153	struct list_head worklist; /* L: list of pending works */
154	int nr_workers; /* L: total number of workers */
155
156	/* nr_idle includes the ones off idle_list for rebinding */
157	int nr_idle; /* L: currently idle ones */
158
159	struct list_head idle_list; /* X: list of idle workers */
160	struct timer_list idle_timer; /* L: worker idle timeout */
161	struct timer_list mayday_timer; /* L: SOS timer for workers */
162
163	/* a workers is either on busy_hash or idle_list, or the manager */
164	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
165	/* L: hash of busy workers */
166
167	/* see manage_workers() for details on the two manager mutexes */
168	struct mutex manager_arb; /* manager arbitration */
169	struct worker manager; / L: purely informational */
170	struct mutex attach_mutex; /* attach/detach exclusion */
171	struct list_head workers; /* A: attached workers */
172	struct completion detach_completion; / all workers detached */
173
174	struct ida worker_ida; /* worker IDs for task name */
175
176	struct workqueue_attrs attrs; / I: worker attributes */
177	struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178	int refcnt; /* PL: refcnt for unbound pools */
179
180	/*
181	* The current concurrency level. As it's likely to be accessed
182	* from other CPUs during try_to_wake_up(), put it in a separate
183	* cacheline.
184	*/
185	atomic_t nr_running ____cacheline_aligned_in_smp;
186
187	/*
188	* Destruction of pool is sched-RCU protected to allow dereferences
189	* from get_work_pool().
190	*/
191	struct rcu_head rcu;
192	} ____cacheline_aligned_in_smp;
193
194	/*
195	* The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196	* of work_struct->data are used for flags and the remaining high bits
197	* point to the pwq; thus, pwqs need to be aligned at two's power of the
198	* number of flag bits.
199	*/
200	struct pool_workqueue {
201	struct worker_pool pool; / I: the associated pool */
202	struct workqueue_struct wq; / I: the owning workqueue */
203	int work_color; /* L: current color */
204	int flush_color; /* L: flushing color */
205	int refcnt; /* L: reference count */
206	int nr_in_flight[WORK_NR_COLORS];
207	/* L: nr of in_flight works */
208	int nr_active; /* L: nr of active works */
209	int max_active; /* L: max active works */
210	struct list_head delayed_works; /* L: delayed works */
211	struct list_head pwqs_node; /* WR: node on wq->pwqs */
212	struct list_head mayday_node; /* MD: node on wq->maydays */
213
214	/*
215	* Release of unbound pwq is punted to system_wq. See put_pwq()
216	* and pwq_unbound_release_workfn() for details. pool_workqueue
217	* itself is also sched-RCU protected so that the first pwq can be
218	* determined without grabbing wq->mutex.
219	*/
220	struct work_struct unbound_release_work;
221	struct rcu_head rcu;
222	} __aligned(1 << WORK_STRUCT_FLAG_BITS);
223
224	/*
225	* Structure used to wait for workqueue flush.
226	*/
227	struct wq_flusher {
228	struct list_head list; /* WQ: list of flushers */
229	int flush_color; /* WQ: flush color waiting for */
230	struct completion done; /* flush completion */
231	};
232
233	struct wq_device;
234
235	/*
236	* The externally visible workqueue. It relays the issued work items to
237	* the appropriate worker_pool through its pool_workqueues.
238	*/
239	struct workqueue_struct {
240	struct list_head pwqs; /* WR: all pwqs of this wq */
241	struct list_head list; /* PR: list of all workqueues */
242
243	struct mutex mutex; /* protects this wq */
244	int work_color; /* WQ: current work color */
245	int flush_color; /* WQ: current flush color */
246	atomic_t nr_pwqs_to_flush; /* flush in progress */
247	struct wq_flusher first_flusher; / WQ: first flusher */
248	struct list_head flusher_queue; /* WQ: flush waiters */
249	struct list_head flusher_overflow; /* WQ: flush overflow list */
250
251	struct list_head maydays; /* MD: pwqs requesting rescue */
252	struct worker rescuer; / I: rescue worker */
253
254	int nr_drainers; /* WQ: drain in progress */
255	int saved_max_active; /* WQ: saved pwq max_active */
256
257	struct workqueue_attrs unbound_attrs; / PW: only for unbound wqs */
258	struct pool_workqueue dfl_pwq; / PW: only for unbound wqs */
259
260	#ifdef CONFIG_SYSFS
261	struct wq_device wq_dev; / I: for sysfs interface */
262	#endif
263	#ifdef CONFIG_LOCKDEP
264	struct lockdep_map lockdep_map;
265	#endif
266	char name[WQ_NAME_LEN]; /* I: workqueue name */
267
268	/*
269	* Destruction of workqueue_struct is sched-RCU protected to allow
270	* walking the workqueues list without grabbing wq_pool_mutex.
271	* This is used to dump all workqueues from sysrq.
272	*/
273	struct rcu_head rcu;
274
275	/* hot fields used during command issue, aligned to cacheline */
276	unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277	struct pool_workqueue __percpu cpu_pwqs; / I: per-cpu pwqs */
278	struct pool_workqueue __rcu numa_pwq_tbl[]; / PWR: unbound pwqs indexed by node */
279	};
280
281	static struct kmem_cache *pwq_cache;
282
283	static cpumask_var_t *wq_numa_possible_cpumask;
284	/* possible CPUs of each node */
285
286	static bool wq_disable_numa;
287	module_param_named(disable_numa, wq_disable_numa, bool, 0444);
288
289	/* see the comment above the definition of WQ_POWER_EFFICIENT */
290	static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291	module_param_named(power_efficient, wq_power_efficient, bool, 0444);
292
293	static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
294
295	/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
296	static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
297
298	static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
299	static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
300
301	static LIST_HEAD(workqueues); /* PR: list of all workqueues */
302	static bool workqueue_freezing; /* PL: have wqs started freezing? */
303
304	/* PL: allowable cpus for unbound wqs and work items */
305	static cpumask_var_t wq_unbound_cpumask;
306
307	/* CPU where unbound work was last round robin scheduled from this CPU */
308	static DEFINE_PER_CPU(int, wq_rr_cpu_last);
309
310	/*
311	* Local execution of unbound work items is no longer guaranteed. The
312	* following always forces round-robin CPU selection on unbound work items
313	* to uncover usages which depend on it.
314	*/
315	#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
316	static bool wq_debug_force_rr_cpu = true;
317	#else
318	static bool wq_debug_force_rr_cpu = false;
319	#endif
320	module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
321
322	/* the per-cpu worker pools */
323	static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
324
325	static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
326
327	/* PL: hash of all unbound pools keyed by pool->attrs */
328	static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
329
330	/* I: attributes used when instantiating standard unbound pools on demand */
331	static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
332
333	/* I: attributes used when instantiating ordered pools on demand */
334	static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
335
336	struct workqueue_struct *system_wq __read_mostly;
337	EXPORT_SYMBOL(system_wq);
338	struct workqueue_struct *system_highpri_wq __read_mostly;
339	EXPORT_SYMBOL_GPL(system_highpri_wq);
340	struct workqueue_struct *system_long_wq __read_mostly;
341	EXPORT_SYMBOL_GPL(system_long_wq);
342	struct workqueue_struct *system_unbound_wq __read_mostly;
343	EXPORT_SYMBOL_GPL(system_unbound_wq);
344	struct workqueue_struct *system_freezable_wq __read_mostly;
345	EXPORT_SYMBOL_GPL(system_freezable_wq);
346	struct workqueue_struct *system_power_efficient_wq __read_mostly;
347	EXPORT_SYMBOL_GPL(system_power_efficient_wq);
348	struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
349	EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
350
351	static int worker_thread(void *__worker);
352	static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
353
354	#define CREATE_TRACE_POINTS
355	#include <trace/events/workqueue.h>
356
357	#define assert_rcu_or_pool_mutex() \
358	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
359	!lockdep_is_held(&wq_pool_mutex), \
360	"sched RCU or wq_pool_mutex should be held")
361
362	#define assert_rcu_or_wq_mutex(wq) \
363	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
364	!lockdep_is_held(&wq->mutex), \
365	"sched RCU or wq->mutex should be held")
366
367	#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
368	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
369	!lockdep_is_held(&wq->mutex) && \
370	!lockdep_is_held(&wq_pool_mutex), \
371	"sched RCU, wq->mutex or wq_pool_mutex should be held")
372
373	#define for_each_cpu_worker_pool(pool, cpu) \
374	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
375	(pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
376	(pool)++)
377
378	/**
379	* for_each_pool - iterate through all worker_pools in the system
380	* @pool: iteration cursor
381	* @pi: integer used for iteration
382	*
383	* This must be called either with wq_pool_mutex held or sched RCU read
384	* locked. If the pool needs to be used beyond the locking in effect, the
385	* caller is responsible for guaranteeing that the pool stays online.
386	*
387	* The if/else clause exists only for the lockdep assertion and can be
388	* ignored.
389	*/
390	#define for_each_pool(pool, pi) \
391	idr_for_each_entry(&worker_pool_idr, pool, pi) \
392	if (({ assert_rcu_or_pool_mutex(); false; })) { } \
393	else
394
395	/**
396	* for_each_pool_worker - iterate through all workers of a worker_pool
397	* @worker: iteration cursor
398	* @pool: worker_pool to iterate workers of
399	*
400	* This must be called with @pool->attach_mutex.
401	*
402	* The if/else clause exists only for the lockdep assertion and can be
403	* ignored.
404	*/
405	#define for_each_pool_worker(worker, pool) \
406	list_for_each_entry((worker), &(pool)->workers, node) \
407	if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
408	else
409
410	/**
411	* for_each_pwq - iterate through all pool_workqueues of the specified workqueue
412	* @pwq: iteration cursor
413	* @wq: the target workqueue
414	*
415	* This must be called either with wq->mutex held or sched RCU read locked.
416	* If the pwq needs to be used beyond the locking in effect, the caller is
417	* responsible for guaranteeing that the pwq stays online.
418	*
419	* The if/else clause exists only for the lockdep assertion and can be
420	* ignored.
421	*/
422	#define for_each_pwq(pwq, wq) \
423	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
424	if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
425	else
426
427	#ifdef CONFIG_DEBUG_OBJECTS_WORK
428
429	static struct debug_obj_descr work_debug_descr;
430
431	static void work_debug_hint(void addr)
432	{
433	return ((struct work_struct *) addr)->func;
434	}
435
436	static bool work_is_static_object(void *addr)
437	{
438	struct work_struct *work = addr;
439
440	return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
441	}
442
443	/*
444	* fixup_init is called when:
445	* - an active object is initialized
446	*/
447	static bool work_fixup_init(void *addr, enum debug_obj_state state)
448	{
449	struct work_struct *work = addr;
450
451	switch (state) {
452	case ODEBUG_STATE_ACTIVE:
453	cancel_work_sync(work);
454	debug_object_init(work, &work_debug_descr);
455	return true;
456	default:
457	return false;
458	}
459	}
460
461	/*
462	* fixup_free is called when:
463	* - an active object is freed
464	*/
465	static bool work_fixup_free(void *addr, enum debug_obj_state state)
466	{
467	struct work_struct *work = addr;
468
469	switch (state) {
470	case ODEBUG_STATE_ACTIVE:
471	cancel_work_sync(work);
472	debug_object_free(work, &work_debug_descr);
473	return true;
474	default:
475	return false;
476	}
477	}
478
479	static struct debug_obj_descr work_debug_descr = {
480	.name = "work_struct",
481	.debug_hint = work_debug_hint,
482	.is_static_object = work_is_static_object,
483	.fixup_init = work_fixup_init,
484	.fixup_free = work_fixup_free,
485	};
486
487	static inline void debug_work_activate(struct work_struct *work)
488	{
489	debug_object_activate(work, &work_debug_descr);
490	}
491
492	static inline void debug_work_deactivate(struct work_struct *work)
493	{
494	debug_object_deactivate(work, &work_debug_descr);
495	}
496
497	void __init_work(struct work_struct *work, int onstack)
498	{
499	if (onstack)
500	debug_object_init_on_stack(work, &work_debug_descr);
501	else
502	debug_object_init(work, &work_debug_descr);
503	}
504	EXPORT_SYMBOL_GPL(__init_work);
505
506	void destroy_work_on_stack(struct work_struct *work)
507	{
508	debug_object_free(work, &work_debug_descr);
509	}
510	EXPORT_SYMBOL_GPL(destroy_work_on_stack);
511
512	void destroy_delayed_work_on_stack(struct delayed_work *work)
513	{
514	destroy_timer_on_stack(&work->timer);
515	debug_object_free(&work->work, &work_debug_descr);
516	}
517	EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
518
519	#else
520	static inline void debug_work_activate(struct work_struct *work) { }
521	static inline void debug_work_deactivate(struct work_struct *work) { }
522	#endif
523
524	/**
525	* worker_pool_assign_id - allocate ID and assing it to @pool
526	* @pool: the pool pointer of interest
527	*
528	* Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
529	* successfully, -errno on failure.
530	*/
531	static int worker_pool_assign_id(struct worker_pool *pool)
532	{
533	int ret;
534
535	lockdep_assert_held(&wq_pool_mutex);
536
537	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
538	GFP_KERNEL);
539	if (ret >= 0) {
540	pool->id = ret;
541	return 0;
542	}
543	return ret;
544	}
545
546	/**
547	* unbound_pwq_by_node - return the unbound pool_workqueue for the given node
548	* @wq: the target workqueue
549	* @node: the node ID
550	*
551	* This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
552	* read locked.
553	* If the pwq needs to be used beyond the locking in effect, the caller is
554	* responsible for guaranteeing that the pwq stays online.
555	*
556	* Return: The unbound pool_workqueue for @node.
557	*/
558	static struct pool_workqueue unbound_pwq_by_node(struct workqueue_struct wq,
559	int node)
560	{
561	assert_rcu_or_wq_mutex_or_pool_mutex(wq);
562
563	/*
564	* XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
565	* delayed item is pending. The plan is to keep CPU -> NODE
566	* mapping valid and stable across CPU on/offlines. Once that
567	* happens, this workaround can be removed.
568	*/
569	if (unlikely(node == NUMA_NO_NODE))
570	return wq->dfl_pwq;
571
572	return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
573	}
574
575	static unsigned int work_color_to_flags(int color)
576	{
577	return color << WORK_STRUCT_COLOR_SHIFT;
578	}
579
580	static int get_work_color(struct work_struct *work)
581	{
582	return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
583	((1 << WORK_STRUCT_COLOR_BITS) - 1);
584	}
585
586	static int work_next_color(int color)
587	{
588	return (color + 1) % WORK_NR_COLORS;
589	}
590
591	/*
592	* While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
593	* contain the pointer to the queued pwq. Once execution starts, the flag
594	* is cleared and the high bits contain OFFQ flags and pool ID.
595	*
596	* set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
597	* and clear_work_data() can be used to set the pwq, pool or clear
598	* work->data. These functions should only be called while the work is
599	* owned - ie. while the PENDING bit is set.
600	*
601	* get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
602	* corresponding to a work. Pool is available once the work has been
603	* queued anywhere after initialization until it is sync canceled. pwq is
604	* available only while the work item is queued.
605	*
606	* %WORK_OFFQ_CANCELING is used to mark a work item which is being
607	* canceled. While being canceled, a work item may have its PENDING set
608	* but stay off timer and worklist for arbitrarily long and nobody should
609	* try to steal the PENDING bit.
610	*/
611	static inline void set_work_data(struct work_struct *work, unsigned long data,
612	unsigned long flags)
613	{
614	WARN_ON_ONCE(!work_pending(work));
615	atomic_long_set(&work->data, data \| flags \| work_static(work));
616	}
617
618	static void set_work_pwq(struct work_struct work, struct pool_workqueue pwq,
619	unsigned long extra_flags)
620	{
621	set_work_data(work, (unsigned long)pwq,
622	WORK_STRUCT_PENDING \| WORK_STRUCT_PWQ \| extra_flags);
623	}
624
625	static void set_work_pool_and_keep_pending(struct work_struct *work,
626	int pool_id)
627	{
628	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
629	WORK_STRUCT_PENDING);
630	}
631
632	static void set_work_pool_and_clear_pending(struct work_struct *work,
633	int pool_id)
634	{
635	/*
636	* The following wmb is paired with the implied mb in
637	* test_and_set_bit(PENDING) and ensures all updates to @work made
638	* here are visible to and precede any updates by the next PENDING
639	* owner.
640	*/
641	smp_wmb();
642	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
643	/*
644	* The following mb guarantees that previous clear of a PENDING bit
645	* will not be reordered with any speculative LOADS or STORES from
646	* work->current_func, which is executed afterwards. This possible
647	* reordering can lead to a missed execution on attempt to qeueue
648	* the same @work. E.g. consider this case:
649	*
650	* CPU#0 CPU#1
651	* ---------------------------- --------------------------------
652	*
653	* 1 STORE event_indicated
654	* 2 queue_work_on() {
655	* 3 test_and_set_bit(PENDING)
656	* 4 } set_..._and_clear_pending() {
657	* 5 set_work_data() # clear bit
658	* 6 smp_mb()
659	* 7 work->current_func() {
660	* 8 LOAD event_indicated
661	* }
662	*
663	* Without an explicit full barrier speculative LOAD on line 8 can
664	* be executed before CPU#0 does STORE on line 1. If that happens,
665	* CPU#0 observes the PENDING bit is still set and new execution of
666	* a @work is not queued in a hope, that CPU#1 will eventually
667	* finish the queued @work. Meanwhile CPU#1 does not see
668	* event_indicated is set, because speculative LOAD was executed
669	* before actual STORE.
670	*/
671	smp_mb();
672	}
673
674	static void clear_work_data(struct work_struct *work)
675	{
676	smp_wmb(); /* see set_work_pool_and_clear_pending() */
677	set_work_data(work, WORK_STRUCT_NO_POOL, 0);
678	}
679
680	static struct pool_workqueue get_work_pwq(struct work_struct work)
681	{
682	unsigned long data = atomic_long_read(&work->data);
683
684	if (data & WORK_STRUCT_PWQ)
685	return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
686	else
687	return NULL;
688	}
689
690	/**
691	* get_work_pool - return the worker_pool a given work was associated with
692	* @work: the work item of interest
693	*
694	* Pools are created and destroyed under wq_pool_mutex, and allows read
695	* access under sched-RCU read lock. As such, this function should be
696	* called under wq_pool_mutex or with preemption disabled.
697	*
698	* All fields of the returned pool are accessible as long as the above
699	* mentioned locking is in effect. If the returned pool needs to be used
700	* beyond the critical section, the caller is responsible for ensuring the
701	* returned pool is and stays online.
702	*
703	* Return: The worker_pool @work was last associated with. %NULL if none.
704	*/
705	static struct worker_pool get_work_pool(struct work_struct work)
706	{
707	unsigned long data = atomic_long_read(&work->data);
708	int pool_id;
709
710	assert_rcu_or_pool_mutex();
711
712	if (data & WORK_STRUCT_PWQ)
713	return ((struct pool_workqueue *)
714	(data & WORK_STRUCT_WQ_DATA_MASK))->pool;
715
716	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
717	if (pool_id == WORK_OFFQ_POOL_NONE)
718	return NULL;
719
720	return idr_find(&worker_pool_idr, pool_id);
721	}
722
723	/**
724	* get_work_pool_id - return the worker pool ID a given work is associated with
725	* @work: the work item of interest
726	*
727	* Return: The worker_pool ID @work was last associated with.
728	* %WORK_OFFQ_POOL_NONE if none.
729	*/
730	static int get_work_pool_id(struct work_struct *work)
731	{
732	unsigned long data = atomic_long_read(&work->data);
733
734	if (data & WORK_STRUCT_PWQ)
735	return ((struct pool_workqueue *)
736	(data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
737
738	return data >> WORK_OFFQ_POOL_SHIFT;
739	}
740
741	static void mark_work_canceling(struct work_struct *work)
742	{
743	unsigned long pool_id = get_work_pool_id(work);
744
745	pool_id <<= WORK_OFFQ_POOL_SHIFT;
746	set_work_data(work, pool_id \| WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
747	}
748
749	static bool work_is_canceling(struct work_struct *work)
750	{
751	unsigned long data = atomic_long_read(&work->data);
752
753	return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
754	}
755
756	/*
757	* Policy functions. These define the policies on how the global worker
758	* pools are managed. Unless noted otherwise, these functions assume that
759	* they're being called with pool->lock held.
760	*/
761
762	static bool __need_more_worker(struct worker_pool *pool)
763	{
764	return !atomic_read(&pool->nr_running);
765	}
766
767	/*
768	* Need to wake up a worker? Called from anything but currently
769	* running workers.
770	*
771	* Note that, because unbound workers never contribute to nr_running, this
772	* function will always return %true for unbound pools as long as the
773	* worklist isn't empty.
774	*/
775	static bool need_more_worker(struct worker_pool *pool)
776	{
777	return !list_empty(&pool->worklist) && __need_more_worker(pool);
778	}
779
780	/* Can I start working? Called from busy but !running workers. */
781	static bool may_start_working(struct worker_pool *pool)
782	{
783	return pool->nr_idle;
784	}
785
786	/* Do I need to keep working? Called from currently running workers. */
787	static bool keep_working(struct worker_pool *pool)
788	{
789	return !list_empty(&pool->worklist) &&
790	atomic_read(&pool->nr_running) <= 1;
791	}
792
793	/* Do we need a new worker? Called from manager. */
794	static bool need_to_create_worker(struct worker_pool *pool)
795	{
796	return need_more_worker(pool) && !may_start_working(pool);
797	}
798
799	/* Do we have too many workers and should some go away? */
800	static bool too_many_workers(struct worker_pool *pool)
801	{
802	bool managing = mutex_is_locked(&pool->manager_arb);
803	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
804	int nr_busy = pool->nr_workers - nr_idle;
805
806	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
807	}
808
809	/*
810	* Wake up functions.
811	*/
812
813	/* Return the first idle worker. Safe with preemption disabled */
814	static struct worker first_idle_worker(struct worker_pool pool)
815	{
816	if (unlikely(list_empty(&pool->idle_list)))
817	return NULL;
818
819	return list_first_entry(&pool->idle_list, struct worker, entry);
820	}
821
822	/**
823	* wake_up_worker - wake up an idle worker
824	* @pool: worker pool to wake worker from
825	*
826	* Wake up the first idle worker of @pool.
827	*
828	* CONTEXT:
829	* spin_lock_irq(pool->lock).
830	*/
831	static void wake_up_worker(struct worker_pool *pool)
832	{
833	struct worker *worker = first_idle_worker(pool);
834
835	if (likely(worker))
836	wake_up_process(worker->task);
837	}
838
839	/**
840	* wq_worker_waking_up - a worker is waking up
841	* @task: task waking up
842	* @cpu: CPU @task is waking up to
843	*
844	* This function is called during try_to_wake_up() when a worker is
845	* being awoken.
846	*
847	* CONTEXT:
848	* spin_lock_irq(rq->lock)
849	*/
850	void wq_worker_waking_up(struct task_struct *task, int cpu)
851	{
852	struct worker *worker = kthread_data(task);
853
854	if (!(worker->flags & WORKER_NOT_RUNNING)) {
855	WARN_ON_ONCE(worker->pool->cpu != cpu);
856	atomic_inc(&worker->pool->nr_running);
857	}
858	}
859
860	/**
861	* wq_worker_sleeping - a worker is going to sleep
862	* @task: task going to sleep
863	*
864	* This function is called during schedule() when a busy worker is
865	* going to sleep. Worker on the same cpu can be woken up by
866	* returning pointer to its task.
867	*
868	* CONTEXT:
869	* spin_lock_irq(rq->lock)
870	*
871	* Return:
872	* Worker task on @cpu to wake up, %NULL if none.
873	*/
874	struct task_struct wq_worker_sleeping(struct task_struct task)
875	{
876	struct worker worker = kthread_data(task), to_wakeup = NULL;
877	struct worker_pool *pool;
878
879	/*
880	* Rescuers, which may not have all the fields set up like normal
881	* workers, also reach here, let's not access anything before
882	* checking NOT_RUNNING.
883	*/
884	if (worker->flags & WORKER_NOT_RUNNING)
885	return NULL;
886
887	pool = worker->pool;
888
889	/* this can only happen on the local cpu */
890	if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
891	return NULL;
892
893	/*
894	* The counterpart of the following dec_and_test, implied mb,
895	* worklist not empty test sequence is in insert_work().
896	* Please read comment there.
897	*
898	* NOT_RUNNING is clear. This means that we're bound to and
899	* running on the local cpu w/ rq lock held and preemption
900	* disabled, which in turn means that none else could be
901	* manipulating idle_list, so dereferencing idle_list without pool
902	* lock is safe.
903	*/
904	if (atomic_dec_and_test(&pool->nr_running) &&
905	!list_empty(&pool->worklist))
906	to_wakeup = first_idle_worker(pool);
907	return to_wakeup ? to_wakeup->task : NULL;
908	}
909
910	/**
911	* worker_set_flags - set worker flags and adjust nr_running accordingly
912	* @worker: self
913	* @flags: flags to set
914	*
915	* Set @flags in @worker->flags and adjust nr_running accordingly.
916	*
917	* CONTEXT:
918	* spin_lock_irq(pool->lock)
919	*/
920	static inline void worker_set_flags(struct worker *worker, unsigned int flags)
921	{
922	struct worker_pool *pool = worker->pool;
923
924	WARN_ON_ONCE(worker->task != current);
925
926	/* If transitioning into NOT_RUNNING, adjust nr_running. */
927	if ((flags & WORKER_NOT_RUNNING) &&
928	!(worker->flags & WORKER_NOT_RUNNING)) {
929	atomic_dec(&pool->nr_running);
930	}
931
932	worker->flags \|= flags;
933	}
934
935	/**
936	* worker_clr_flags - clear worker flags and adjust nr_running accordingly
937	* @worker: self
938	* @flags: flags to clear
939	*
940	* Clear @flags in @worker->flags and adjust nr_running accordingly.
941	*
942	* CONTEXT:
943	* spin_lock_irq(pool->lock)
944	*/
945	static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
946	{
947	struct worker_pool *pool = worker->pool;
948	unsigned int oflags = worker->flags;
949
950	WARN_ON_ONCE(worker->task != current);
951
952	worker->flags &= ~flags;
953
954	/*
955	* If transitioning out of NOT_RUNNING, increment nr_running. Note
956	* that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
957	* of multiple flags, not a single flag.
958	*/
959	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
960	if (!(worker->flags & WORKER_NOT_RUNNING))
961	atomic_inc(&pool->nr_running);
962	}
963
964	/**
965	* find_worker_executing_work - find worker which is executing a work
966	* @pool: pool of interest
967	* @work: work to find worker for
968	*
969	* Find a worker which is executing @work on @pool by searching
970	* @pool->busy_hash which is keyed by the address of @work. For a worker
971	* to match, its current execution should match the address of @work and
972	* its work function. This is to avoid unwanted dependency between
973	* unrelated work executions through a work item being recycled while still
974	* being executed.
975	*
976	* This is a bit tricky. A work item may be freed once its execution
977	* starts and nothing prevents the freed area from being recycled for
978	* another work item. If the same work item address ends up being reused
979	* before the original execution finishes, workqueue will identify the
980	* recycled work item as currently executing and make it wait until the
981	* current execution finishes, introducing an unwanted dependency.
982	*
983	* This function checks the work item address and work function to avoid
984	* false positives. Note that this isn't complete as one may construct a
985	* work function which can introduce dependency onto itself through a
986	* recycled work item. Well, if somebody wants to shoot oneself in the
987	* foot that badly, there's only so much we can do, and if such deadlock
988	* actually occurs, it should be easy to locate the culprit work function.
989	*
990	* CONTEXT:
991	* spin_lock_irq(pool->lock).
992	*
993	* Return:
994	* Pointer to worker which is executing @work if found, %NULL
995	* otherwise.
996	*/
997	static struct worker find_worker_executing_work(struct worker_pool pool,
998	struct work_struct *work)
999	{
1000	struct worker *worker;
1001
1002	hash_for_each_possible(pool->busy_hash, worker, hentry,
1003	(unsigned long)work)
1004	if (worker->current_work == work &&
1005	worker->current_func == work->func)
1006	return worker;
1007
1008	return NULL;
1009	}
1010
1011	/**
1012	* move_linked_works - move linked works to a list
1013	* @work: start of series of works to be scheduled
1014	* @head: target list to append @work to
1015	* @nextp: out parameter for nested worklist walking
1016	*
1017	* Schedule linked works starting from @work to @head. Work series to
1018	* be scheduled starts at @work and includes any consecutive work with
1019	* WORK_STRUCT_LINKED set in its predecessor.
1020	*
1021	* If @nextp is not NULL, it's updated to point to the next work of
1022	* the last scheduled work. This allows move_linked_works() to be
1023	* nested inside outer list_for_each_entry_safe().
1024	*
1025	* CONTEXT:
1026	* spin_lock_irq(pool->lock).
1027	*/
1028	static void move_linked_works(struct work_struct work, struct list_head head,
1029	struct work_struct **nextp)
1030	{
1031	struct work_struct *n;
1032
1033	/*
1034	* Linked worklist will always end before the end of the list,
1035	* use NULL for list head.
1036	*/
1037	list_for_each_entry_safe_from(work, n, NULL, entry) {
1038	list_move_tail(&work->entry, head);
1039	if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1040	break;
1041	}
1042
1043	/*
1044	* If we're already inside safe list traversal and have moved
1045	* multiple works to the scheduled queue, the next position
1046	* needs to be updated.
1047	*/
1048	if (nextp)
1049	*nextp = n;
1050	}
1051
1052	/**
1053	* get_pwq - get an extra reference on the specified pool_workqueue
1054	* @pwq: pool_workqueue to get
1055	*
1056	* Obtain an extra reference on @pwq. The caller should guarantee that
1057	* @pwq has positive refcnt and be holding the matching pool->lock.
1058	*/
1059	static void get_pwq(struct pool_workqueue *pwq)
1060	{
1061	lockdep_assert_held(&pwq->pool->lock);
1062	WARN_ON_ONCE(pwq->refcnt <= 0);
1063	pwq->refcnt++;
1064	}
1065
1066	/**
1067	* put_pwq - put a pool_workqueue reference
1068	* @pwq: pool_workqueue to put
1069	*
1070	* Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1071	* destruction. The caller should be holding the matching pool->lock.
1072	*/
1073	static void put_pwq(struct pool_workqueue *pwq)
1074	{
1075	lockdep_assert_held(&pwq->pool->lock);
1076	if (likely(--pwq->refcnt))
1077	return;
1078	if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1079	return;
1080	/*
1081	* @pwq can't be released under pool->lock, bounce to
1082	* pwq_unbound_release_workfn(). This never recurses on the same
1083	* pool->lock as this path is taken only for unbound workqueues and
1084	* the release work item is scheduled on a per-cpu workqueue. To
1085	* avoid lockdep warning, unbound pool->locks are given lockdep
1086	* subclass of 1 in get_unbound_pool().
1087	*/
1088	schedule_work(&pwq->unbound_release_work);
1089	}
1090
1091	/**
1092	* put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1093	* @pwq: pool_workqueue to put (can be %NULL)
1094	*
1095	* put_pwq() with locking. This function also allows %NULL @pwq.
1096	*/
1097	static void put_pwq_unlocked(struct pool_workqueue *pwq)
1098	{
1099	if (pwq) {
1100	/*
1101	* As both pwqs and pools are sched-RCU protected, the
1102	* following lock operations are safe.
1103	*/
1104	spin_lock_irq(&pwq->pool->lock);
1105	put_pwq(pwq);
1106	spin_unlock_irq(&pwq->pool->lock);
1107	}
1108	}
1109
1110	static void pwq_activate_delayed_work(struct work_struct *work)
1111	{
1112	struct pool_workqueue *pwq = get_work_pwq(work);
1113
1114	trace_workqueue_activate_work(work);
1115	if (list_empty(&pwq->pool->worklist))
1116	pwq->pool->watchdog_ts = jiffies;
1117	move_linked_works(work, &pwq->pool->worklist, NULL);
1118	__clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1119	pwq->nr_active++;
1120	}
1121
1122	static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1123	{
1124	struct work_struct *work = list_first_entry(&pwq->delayed_works,
1125	struct work_struct, entry);
1126
1127	pwq_activate_delayed_work(work);
1128	}
1129
1130	/**
1131	* pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1132	* @pwq: pwq of interest
1133	* @color: color of work which left the queue
1134	*
1135	* A work either has completed or is removed from pending queue,
1136	* decrement nr_in_flight of its pwq and handle workqueue flushing.
1137	*
1138	* CONTEXT:
1139	* spin_lock_irq(pool->lock).
1140	*/
1141	static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1142	{
1143	/* uncolored work items don't participate in flushing or nr_active */
1144	if (color == WORK_NO_COLOR)
1145	goto out_put;
1146
1147	pwq->nr_in_flight[color]--;
1148
1149	pwq->nr_active--;
1150	if (!list_empty(&pwq->delayed_works)) {
1151	/* one down, submit a delayed one */
1152	if (pwq->nr_active < pwq->max_active)
1153	pwq_activate_first_delayed(pwq);
1154	}
1155
1156	/* is flush in progress and are we at the flushing tip? */
1157	if (likely(pwq->flush_color != color))
1158	goto out_put;
1159
1160	/* are there still in-flight works? */
1161	if (pwq->nr_in_flight[color])
1162	goto out_put;
1163
1164	/* this pwq is done, clear flush_color */
1165	pwq->flush_color = -1;
1166
1167	/*
1168	* If this was the last pwq, wake up the first flusher. It
1169	* will handle the rest.
1170	*/
1171	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1172	complete(&pwq->wq->first_flusher->done);
1173	out_put:
1174	put_pwq(pwq);
1175	}
1176
1177	/**
1178	* try_to_grab_pending - steal work item from worklist and disable irq
1179	* @work: work item to steal
1180	* @is_dwork: @work is a delayed_work
1181	* @flags: place to store irq state
1182	*
1183	* Try to grab PENDING bit of @work. This function can handle @work in any
1184	* stable state - idle, on timer or on worklist.
1185	*
1186	* Return:
1187	* 1 if @work was pending and we successfully stole PENDING
1188	* 0 if @work was idle and we claimed PENDING
1189	* -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1190	* -ENOENT if someone else is canceling @work, this state may persist
1191	* for arbitrarily long
1192	*
1193	* Note:
1194	* On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1195	* interrupted while holding PENDING and @work off queue, irq must be
1196	* disabled on entry. This, combined with delayed_work->timer being
1197	* irqsafe, ensures that we return -EAGAIN for finite short period of time.
1198	*
1199	* On successful return, >= 0, irq is disabled and the caller is
1200	* responsible for releasing it using local_irq_restore(*@flags).
1201	*
1202	* This function is safe to call from any context including IRQ handler.
1203	*/
1204	static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1205	unsigned long *flags)
1206	{
1207	struct worker_pool *pool;
1208	struct pool_workqueue *pwq;
1209
1210	local_irq_save(*flags);
1211
1212	/* try to steal the timer if it exists */
1213	if (is_dwork) {
1214	struct delayed_work *dwork = to_delayed_work(work);
1215
1216	/*
1217	* dwork->timer is irqsafe. If del_timer() fails, it's
1218	* guaranteed that the timer is not queued anywhere and not
1219	* running on the local CPU.
1220	*/
1221	if (likely(del_timer(&dwork->timer)))
1222	return 1;
1223	}
1224
1225	/* try to claim PENDING the normal way */
1226	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1227	return 0;
1228
1229	/*
1230	* The queueing is in progress, or it is already queued. Try to
1231	* steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1232	*/
1233	pool = get_work_pool(work);
1234	if (!pool)
1235	goto fail;
1236
1237	spin_lock(&pool->lock);
1238	/*
1239	* work->data is guaranteed to point to pwq only while the work
1240	* item is queued on pwq->wq, and both updating work->data to point
1241	* to pwq on queueing and to pool on dequeueing are done under
1242	* pwq->pool->lock. This in turn guarantees that, if work->data
1243	* points to pwq which is associated with a locked pool, the work
1244	* item is currently queued on that pool.
1245	*/
1246	pwq = get_work_pwq(work);
1247	if (pwq && pwq->pool == pool) {
1248	debug_work_deactivate(work);
1249
1250	/*
1251	* A delayed work item cannot be grabbed directly because
1252	* it might have linked NO_COLOR work items which, if left
1253	* on the delayed_list, will confuse pwq->nr_active
1254	* management later on and cause stall. Make sure the work
1255	* item is activated before grabbing.
1256	*/
1257	if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1258	pwq_activate_delayed_work(work);
1259
1260	list_del_init(&work->entry);
1261	pwq_dec_nr_in_flight(pwq, get_work_color(work));
1262
1263	/* work->data points to pwq iff queued, point to pool */
1264	set_work_pool_and_keep_pending(work, pool->id);
1265
1266	spin_unlock(&pool->lock);
1267	return 1;
1268	}
1269	spin_unlock(&pool->lock);
1270	fail:
1271	local_irq_restore(*flags);
1272	if (work_is_canceling(work))
1273	return -ENOENT;
1274	cpu_relax();
1275	return -EAGAIN;
1276	}
1277
1278	/**
1279	* insert_work - insert a work into a pool
1280	* @pwq: pwq @work belongs to
1281	* @work: work to insert
1282	* @head: insertion point
1283	* @extra_flags: extra WORK_STRUCT_* flags to set
1284	*
1285	* Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1286	* work_struct flags.
1287	*
1288	* CONTEXT:
1289	* spin_lock_irq(pool->lock).
1290	*/
1291	static void insert_work(struct pool_workqueue pwq, struct work_struct work,
1292	struct list_head *head, unsigned int extra_flags)
1293	{
1294	struct worker_pool *pool = pwq->pool;
1295
1296	/* we own @work, set data and link */
1297	set_work_pwq(work, pwq, extra_flags);
1298	list_add_tail(&work->entry, head);
1299	get_pwq(pwq);
1300
1301	/*
1302	* Ensure either wq_worker_sleeping() sees the above
1303	* list_add_tail() or we see zero nr_running to avoid workers lying
1304	* around lazily while there are works to be processed.
1305	*/
1306	smp_mb();
1307
1308	if (__need_more_worker(pool))
1309	wake_up_worker(pool);
1310	}
1311
1312	/*
1313	* Test whether @work is being queued from another work executing on the
1314	* same workqueue.
1315	*/
1316	static bool is_chained_work(struct workqueue_struct *wq)
1317	{
1318	struct worker *worker;
1319
1320	worker = current_wq_worker();
1321	/*
1322	* Return %true iff I'm a worker execuing a work item on @wq. If
1323	* I'm @worker, it's safe to dereference it without locking.
1324	*/
1325	return worker && worker->current_pwq->wq == wq;
1326	}
1327
1328	/*
1329	* When queueing an unbound work item to a wq, prefer local CPU if allowed
1330	* by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1331	* avoid perturbing sensitive tasks.
1332	*/
1333	static int wq_select_unbound_cpu(int cpu)
1334	{
1335	static bool printed_dbg_warning;
1336	int new_cpu;
1337
1338	if (likely(!wq_debug_force_rr_cpu)) {
1339	if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1340	return cpu;
1341	} else if (!printed_dbg_warning) {
1342	pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1343	printed_dbg_warning = true;
1344	}
1345
1346	if (cpumask_empty(wq_unbound_cpumask))
1347	return cpu;
1348
1349	new_cpu = __this_cpu_read(wq_rr_cpu_last);
1350	new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1351	if (unlikely(new_cpu >= nr_cpu_ids)) {
1352	new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1353	if (unlikely(new_cpu >= nr_cpu_ids))
1354	return cpu;
1355	}
1356	__this_cpu_write(wq_rr_cpu_last, new_cpu);
1357
1358	return new_cpu;
1359	}
1360
1361	static void __queue_work(int cpu, struct workqueue_struct *wq,
1362	struct work_struct *work)
1363	{
1364	struct pool_workqueue *pwq;
1365	struct worker_pool *last_pool;
1366	struct list_head *worklist;
1367	unsigned int work_flags;
1368	unsigned int req_cpu = cpu;
1369
1370	/*
1371	* While a work item is PENDING && off queue, a task trying to
1372	* steal the PENDING will busy-loop waiting for it to either get
1373	* queued or lose PENDING. Grabbing PENDING and queueing should
1374	* happen with IRQ disabled.
1375	*/
1376	WARN_ON_ONCE(!irqs_disabled());
1377
1378	debug_work_activate(work);
1379
1380	/* if draining, only works from the same workqueue are allowed */
1381	if (unlikely(wq->flags & __WQ_DRAINING) &&
1382	WARN_ON_ONCE(!is_chained_work(wq)))
1383	return;
1384	retry:
1385	if (req_cpu == WORK_CPU_UNBOUND)
1386	cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1387
1388	/* pwq which will be used unless @work is executing elsewhere */
1389	if (!(wq->flags & WQ_UNBOUND))
1390	pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1391	else
1392	pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1393
1394	/*
1395	* If @work was previously on a different pool, it might still be
1396	* running there, in which case the work needs to be queued on that
1397	* pool to guarantee non-reentrancy.
1398	*/
1399	last_pool = get_work_pool(work);
1400	if (last_pool && last_pool != pwq->pool) {
1401	struct worker *worker;
1402
1403	spin_lock(&last_pool->lock);
1404
1405	worker = find_worker_executing_work(last_pool, work);
1406
1407	if (worker && worker->current_pwq->wq == wq) {
1408	pwq = worker->current_pwq;
1409	} else {
1410	/* meh... not running there, queue here */
1411	spin_unlock(&last_pool->lock);
1412	spin_lock(&pwq->pool->lock);
1413	}
1414	} else {
1415	spin_lock(&pwq->pool->lock);
1416	}
1417
1418	/*
1419	* pwq is determined and locked. For unbound pools, we could have
1420	* raced with pwq release and it could already be dead. If its
1421	* refcnt is zero, repeat pwq selection. Note that pwqs never die
1422	* without another pwq replacing it in the numa_pwq_tbl or while
1423	* work items are executing on it, so the retrying is guaranteed to
1424	* make forward-progress.
1425	*/
1426	if (unlikely(!pwq->refcnt)) {
1427	if (wq->flags & WQ_UNBOUND) {
1428	spin_unlock(&pwq->pool->lock);
1429	cpu_relax();
1430	goto retry;
1431	}
1432	/* oops */
1433	WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1434	wq->name, cpu);
1435	}
1436
1437	/* pwq determined, queue */
1438	trace_workqueue_queue_work(req_cpu, pwq, work);
1439
1440	if (WARN_ON(!list_empty(&work->entry))) {
1441	spin_unlock(&pwq->pool->lock);
1442	return;
1443	}
1444
1445	pwq->nr_in_flight[pwq->work_color]++;
1446	work_flags = work_color_to_flags(pwq->work_color);
1447
1448	if (likely(pwq->nr_active < pwq->max_active)) {
1449	trace_workqueue_activate_work(work);
1450	pwq->nr_active++;
1451	worklist = &pwq->pool->worklist;
1452	if (list_empty(worklist))
1453	pwq->pool->watchdog_ts = jiffies;
1454	} else {
1455	work_flags \|= WORK_STRUCT_DELAYED;
1456	worklist = &pwq->delayed_works;
1457	}
1458
1459	insert_work(pwq, work, worklist, work_flags);
1460
1461	spin_unlock(&pwq->pool->lock);
1462	}
1463
1464	/**
1465	* queue_work_on - queue work on specific cpu
1466	* @cpu: CPU number to execute work on
1467	* @wq: workqueue to use
1468	* @work: work to queue
1469	*
1470	* We queue the work to a specific CPU, the caller must ensure it
1471	* can't go away.
1472	*
1473	* Return: %false if @work was already on a queue, %true otherwise.
1474	*/
1475	bool queue_work_on(int cpu, struct workqueue_struct *wq,
1476	struct work_struct *work)
1477	{
1478	bool ret = false;
1479	unsigned long flags;
1480
1481	local_irq_save(flags);
1482
1483	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1484	__queue_work(cpu, wq, work);
1485	ret = true;
1486	}
1487
1488	local_irq_restore(flags);
1489	return ret;
1490	}
1491	EXPORT_SYMBOL(queue_work_on);
1492
1493	void delayed_work_timer_fn(unsigned long __data)
1494	{
1495	struct delayed_work dwork = (struct delayed_work )__data;
1496
1497	/* should have been called from irqsafe timer with irq already off */
1498	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
1499	}
1500	EXPORT_SYMBOL(delayed_work_timer_fn);
1501
1502	static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1503	struct delayed_work *dwork, unsigned long delay)
1504	{
1505	struct timer_list *timer = &dwork->timer;
1506	struct work_struct *work = &dwork->work;
1507
1508	WARN_ON_ONCE(timer->function != delayed_work_timer_fn \|\|
1509	timer->data != (unsigned long)dwork);
1510	WARN_ON_ONCE(timer_pending(timer));
1511	WARN_ON_ONCE(!list_empty(&work->entry));
1512
1513	/*
1514	* If @delay is 0, queue @dwork->work immediately. This is for
1515	* both optimization and correctness. The earliest @timer can
1516	* expire is on the closest next tick and delayed_work users depend
1517	* on that there's no such delay when @delay is 0.
1518	*/
1519	if (!delay) {
1520	__queue_work(cpu, wq, &dwork->work);
1521	return;
1522	}
1523
1524	timer_stats_timer_set_start_info(&dwork->timer);
1525
1526	dwork->wq = wq;
1527	dwork->cpu = cpu;
1528	timer->expires = jiffies + delay;
1529
1530	if (unlikely(cpu != WORK_CPU_UNBOUND))
1531	add_timer_on(timer, cpu);
1532	else
1533	add_timer(timer);
1534	}
1535
1536	/**
1537	* queue_delayed_work_on - queue work on specific CPU after delay
1538	* @cpu: CPU number to execute work on
1539	* @wq: workqueue to use
1540	* @dwork: work to queue
1541	* @delay: number of jiffies to wait before queueing
1542	*
1543	* Return: %false if @work was already on a queue, %true otherwise. If
1544	* @delay is zero and @dwork is idle, it will be scheduled for immediate
1545	* execution.
1546	*/
1547	bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1548	struct delayed_work *dwork, unsigned long delay)
1549	{
1550	struct work_struct *work = &dwork->work;
1551	bool ret = false;
1552	unsigned long flags;
1553
1554	/* read the comment in __queue_work() */
1555	local_irq_save(flags);
1556
1557	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1558	__queue_delayed_work(cpu, wq, dwork, delay);
1559	ret = true;
1560	}
1561
1562	local_irq_restore(flags);
1563	return ret;
1564	}
1565	EXPORT_SYMBOL(queue_delayed_work_on);
1566
1567	/**
1568	* mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1569	* @cpu: CPU number to execute work on
1570	* @wq: workqueue to use
1571	* @dwork: work to queue
1572	* @delay: number of jiffies to wait before queueing
1573	*
1574	* If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1575	* modify @dwork's timer so that it expires after @delay. If @delay is
1576	* zero, @work is guaranteed to be scheduled immediately regardless of its
1577	* current state.
1578	*
1579	* Return: %false if @dwork was idle and queued, %true if @dwork was
1580	* pending and its timer was modified.
1581	*
1582	* This function is safe to call from any context including IRQ handler.
1583	* See try_to_grab_pending() for details.
1584	*/
1585	bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1586	struct delayed_work *dwork, unsigned long delay)
1587	{
1588	unsigned long flags;
1589	int ret;
1590
1591	do {
1592	ret = try_to_grab_pending(&dwork->work, true, &flags);
1593	} while (unlikely(ret == -EAGAIN));
1594
1595	if (likely(ret >= 0)) {
1596	__queue_delayed_work(cpu, wq, dwork, delay);
1597	local_irq_restore(flags);
1598	}
1599
1600	/* -ENOENT from try_to_grab_pending() becomes %true */
1601	return ret;
1602	}
1603	EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1604
1605	/**
1606	* worker_enter_idle - enter idle state
1607	* @worker: worker which is entering idle state
1608	*
1609	* @worker is entering idle state. Update stats and idle timer if
1610	* necessary.
1611	*
1612	* LOCKING:
1613	* spin_lock_irq(pool->lock).
1614	*/
1615	static void worker_enter_idle(struct worker *worker)
1616	{
1617	struct worker_pool *pool = worker->pool;
1618
1619	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) \|\|
1620	WARN_ON_ONCE(!list_empty(&worker->entry) &&
1621	(worker->hentry.next \|\| worker->hentry.pprev)))
1622	return;
1623
1624	/* can't use worker_set_flags(), also called from create_worker() */
1625	worker->flags \|= WORKER_IDLE;
1626	pool->nr_idle++;
1627	worker->last_active = jiffies;
1628
1629	/* idle_list is LIFO */
1630	list_add(&worker->entry, &pool->idle_list);
1631
1632	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1633	mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1634
1635	/*
1636	* Sanity check nr_running. Because wq_unbind_fn() releases
1637	* pool->lock between setting %WORKER_UNBOUND and zapping
1638	* nr_running, the warning may trigger spuriously. Check iff
1639	* unbind is not in progress.
1640	*/
1641	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1642	pool->nr_workers == pool->nr_idle &&
1643	atomic_read(&pool->nr_running));
1644	}
1645
1646	/**
1647	* worker_leave_idle - leave idle state
1648	* @worker: worker which is leaving idle state
1649	*
1650	* @worker is leaving idle state. Update stats.
1651	*
1652	* LOCKING:
1653	* spin_lock_irq(pool->lock).
1654	*/
1655	static void worker_leave_idle(struct worker *worker)
1656	{
1657	struct worker_pool *pool = worker->pool;
1658
1659	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1660	return;
1661	worker_clr_flags(worker, WORKER_IDLE);
1662	pool->nr_idle--;
1663	list_del_init(&worker->entry);
1664	}
1665
1666	static struct worker *alloc_worker(int node)
1667	{
1668	struct worker *worker;
1669
1670	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1671	if (worker) {
1672	INIT_LIST_HEAD(&worker->entry);
1673	INIT_LIST_HEAD(&worker->scheduled);
1674	INIT_LIST_HEAD(&worker->node);
1675	/* on creation a worker is in !idle && prep state */
1676	worker->flags = WORKER_PREP;
1677	}
1678	return worker;
1679	}
1680
1681	/**
1682	* worker_attach_to_pool() - attach a worker to a pool
1683	* @worker: worker to be attached
1684	* @pool: the target pool
1685	*
1686	* Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1687	* cpu-binding of @worker are kept coordinated with the pool across
1688	* cpu-[un]hotplugs.
1689	*/
1690	static void worker_attach_to_pool(struct worker *worker,
1691	struct worker_pool *pool)
1692	{
1693	mutex_lock(&pool->attach_mutex);
1694
1695	/*
1696	* set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1697	* online CPUs. It'll be re-applied when any of the CPUs come up.
1698	*/
1699	set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1700
1701	/*
1702	* The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1703	* stable across this function. See the comments above the
1704	* flag definition for details.
1705	*/
1706	if (pool->flags & POOL_DISASSOCIATED)
1707	worker->flags \|= WORKER_UNBOUND;
1708
1709	list_add_tail(&worker->node, &pool->workers);
1710
1711	mutex_unlock(&pool->attach_mutex);
1712	}
1713
1714	/**
1715	* worker_detach_from_pool() - detach a worker from its pool
1716	* @worker: worker which is attached to its pool
1717	* @pool: the pool @worker is attached to
1718	*
1719	* Undo the attaching which had been done in worker_attach_to_pool(). The
1720	* caller worker shouldn't access to the pool after detached except it has
1721	* other reference to the pool.
1722	*/
1723	static void worker_detach_from_pool(struct worker *worker,
1724	struct worker_pool *pool)
1725	{
1726	struct completion *detach_completion = NULL;
1727
1728	mutex_lock(&pool->attach_mutex);
1729	list_del(&worker->node);
1730	if (list_empty(&pool->workers))
1731	detach_completion = pool->detach_completion;
1732	mutex_unlock(&pool->attach_mutex);
1733
1734	/* clear leftover flags without pool->lock after it is detached */
1735	worker->flags &= ~(WORKER_UNBOUND \| WORKER_REBOUND);
1736
1737	if (detach_completion)
1738	complete(detach_completion);
1739	}
1740
1741	/**
1742	* create_worker - create a new workqueue worker
1743	* @pool: pool the new worker will belong to
1744	*
1745	* Create and start a new worker which is attached to @pool.
1746	*
1747	* CONTEXT:
1748	* Might sleep. Does GFP_KERNEL allocations.
1749	*
1750	* Return:
1751	* Pointer to the newly created worker.
1752	*/
1753	static struct worker create_worker(struct worker_pool pool)
1754	{
1755	struct worker *worker = NULL;
1756	int id = -1;
1757	char id_buf[16];
1758
1759	/* ID is needed to determine kthread name */
1760	id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1761	if (id < 0)
1762	goto fail;
1763
1764	worker = alloc_worker(pool->node);
1765	if (!worker)
1766	goto fail;
1767
1768	worker->pool = pool;
1769	worker->id = id;
1770
1771	if (pool->cpu >= 0)
1772	snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1773	pool->attrs->nice < 0 ? "H" : "");
1774	else
1775	snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1776
1777	worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1778	"kworker/%s", id_buf);
1779	if (IS_ERR(worker->task))
1780	goto fail;
1781
1782	set_user_nice(worker->task, pool->attrs->nice);
1783	kthread_bind_mask(worker->task, pool->attrs->cpumask);
1784
1785	/* successful, attach the worker to the pool */
1786	worker_attach_to_pool(worker, pool);
1787
1788	/* start the newly created worker */
1789	spin_lock_irq(&pool->lock);
1790	worker->pool->nr_workers++;
1791	worker_enter_idle(worker);
1792	wake_up_process(worker->task);
1793	spin_unlock_irq(&pool->lock);
1794
1795	return worker;
1796
1797	fail:
1798	if (id >= 0)
1799	ida_simple_remove(&pool->worker_ida, id);
1800	kfree(worker);
1801	return NULL;
1802	}
1803
1804	/**
1805	* destroy_worker - destroy a workqueue worker
1806	* @worker: worker to be destroyed
1807	*
1808	* Destroy @worker and adjust @pool stats accordingly. The worker should
1809	* be idle.
1810	*
1811	* CONTEXT:
1812	* spin_lock_irq(pool->lock).
1813	*/
1814	static void destroy_worker(struct worker *worker)
1815	{
1816	struct worker_pool *pool = worker->pool;
1817
1818	lockdep_assert_held(&pool->lock);
1819
1820	/* sanity check frenzy */
1821	if (WARN_ON(worker->current_work) \|\|
1822	WARN_ON(!list_empty(&worker->scheduled)) \|\|
1823	WARN_ON(!(worker->flags & WORKER_IDLE)))
1824	return;
1825
1826	pool->nr_workers--;
1827	pool->nr_idle--;
1828
1829	list_del_init(&worker->entry);
1830	worker->flags \|= WORKER_DIE;
1831	wake_up_process(worker->task);
1832	}
1833
1834	static void idle_worker_timeout(unsigned long __pool)
1835	{
1836	struct worker_pool pool = (void )__pool;
1837
1838	spin_lock_irq(&pool->lock);
1839
1840	while (too_many_workers(pool)) {
1841	struct worker *worker;
1842	unsigned long expires;
1843
1844	/* idle_list is kept in LIFO order, check the last one */
1845	worker = list_entry(pool->idle_list.prev, struct worker, entry);
1846	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1847
1848	if (time_before(jiffies, expires)) {
1849	mod_timer(&pool->idle_timer, expires);
1850	break;
1851	}
1852
1853	destroy_worker(worker);
1854	}
1855
1856	spin_unlock_irq(&pool->lock);
1857	}
1858
1859	static void send_mayday(struct work_struct *work)
1860	{
1861	struct pool_workqueue *pwq = get_work_pwq(work);
1862	struct workqueue_struct *wq = pwq->wq;
1863
1864	lockdep_assert_held(&wq_mayday_lock);
1865
1866	if (!wq->rescuer)
1867	return;
1868
1869	/* mayday mayday mayday */
1870	if (list_empty(&pwq->mayday_node)) {
1871	/*
1872	* If @pwq is for an unbound wq, its base ref may be put at
1873	* any time due to an attribute change. Pin @pwq until the
1874	* rescuer is done with it.
1875	*/
1876	get_pwq(pwq);
1877	list_add_tail(&pwq->mayday_node, &wq->maydays);
1878	wake_up_process(wq->rescuer->task);
1879	}
1880	}
1881
1882	static void pool_mayday_timeout(unsigned long __pool)
1883	{
1884	struct worker_pool pool = (void )__pool;
1885	struct work_struct *work;
1886
1887	spin_lock_irq(&pool->lock);
1888	spin_lock(&wq_mayday_lock); /* for wq->maydays */
1889
1890	if (need_to_create_worker(pool)) {
1891	/*
1892	* We've been trying to create a new worker but
1893	* haven't been successful. We might be hitting an
1894	* allocation deadlock. Send distress signals to
1895	* rescuers.
1896	*/
1897	list_for_each_entry(work, &pool->worklist, entry)
1898	send_mayday(work);
1899	}
1900
1901	spin_unlock(&wq_mayday_lock);
1902	spin_unlock_irq(&pool->lock);
1903
1904	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1905	}
1906
1907	/**
1908	* maybe_create_worker - create a new worker if necessary
1909	* @pool: pool to create a new worker for
1910	*
1911	* Create a new worker for @pool if necessary. @pool is guaranteed to
1912	* have at least one idle worker on return from this function. If
1913	* creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1914	* sent to all rescuers with works scheduled on @pool to resolve
1915	* possible allocation deadlock.
1916	*
1917	* On return, need_to_create_worker() is guaranteed to be %false and
1918	* may_start_working() %true.
1919	*
1920	* LOCKING:
1921	* spin_lock_irq(pool->lock) which may be released and regrabbed
1922	* multiple times. Does GFP_KERNEL allocations. Called only from
1923	* manager.
1924	*/
1925	static void maybe_create_worker(struct worker_pool *pool)
1926	__releases(&pool->lock)
1927	__acquires(&pool->lock)
1928	{
1929	restart:
1930	spin_unlock_irq(&pool->lock);
1931
1932	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1933	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1934
1935	while (true) {
1936	if (create_worker(pool) \|\| !need_to_create_worker(pool))
1937	break;
1938
1939	schedule_timeout_interruptible(CREATE_COOLDOWN);
1940
1941	if (!need_to_create_worker(pool))
1942	break;
1943	}
1944
1945	del_timer_sync(&pool->mayday_timer);
1946	spin_lock_irq(&pool->lock);
1947	/*
1948	* This is necessary even after a new worker was just successfully
1949	* created as @pool->lock was dropped and the new worker might have
1950	* already become busy.
1951	*/
1952	if (need_to_create_worker(pool))
1953	goto restart;
1954	}
1955
1956	/**
1957	* manage_workers - manage worker pool
1958	* @worker: self
1959	*
1960	* Assume the manager role and manage the worker pool @worker belongs
1961	* to. At any given time, there can be only zero or one manager per
1962	* pool. The exclusion is handled automatically by this function.
1963	*
1964	* The caller can safely start processing works on false return. On
1965	* true return, it's guaranteed that need_to_create_worker() is false
1966	* and may_start_working() is true.
1967	*
1968	* CONTEXT:
1969	* spin_lock_irq(pool->lock) which may be released and regrabbed
1970	* multiple times. Does GFP_KERNEL allocations.
1971	*
1972	* Return:
1973	* %false if the pool doesn't need management and the caller can safely
1974	* start processing works, %true if management function was performed and
1975	* the conditions that the caller verified before calling the function may
1976	* no longer be true.
1977	*/
1978	static bool manage_workers(struct worker *worker)
1979	{
1980	struct worker_pool *pool = worker->pool;
1981
1982	/*
1983	* Anyone who successfully grabs manager_arb wins the arbitration
1984	* and becomes the manager. mutex_trylock() on pool->manager_arb
1985	* failure while holding pool->lock reliably indicates that someone
1986	* else is managing the pool and the worker which failed trylock
1987	* can proceed to executing work items. This means that anyone
1988	* grabbing manager_arb is responsible for actually performing
1989	* manager duties. If manager_arb is grabbed and released without
1990	* actual management, the pool may stall indefinitely.
1991	*/
1992	if (!mutex_trylock(&pool->manager_arb))
1993	return false;
1994	pool->manager = worker;
1995
1996	maybe_create_worker(pool);
1997
1998	pool->manager = NULL;
1999	mutex_unlock(&pool->manager_arb);
2000	return true;
2001	}
2002
2003	/**
2004	* process_one_work - process single work
2005	* @worker: self
2006	* @work: work to process
2007	*
2008	* Process @work. This function contains all the logics necessary to
2009	* process a single work including synchronization against and
2010	* interaction with other workers on the same cpu, queueing and
2011	* flushing. As long as context requirement is met, any worker can
2012	* call this function to process a work.
2013	*
2014	* CONTEXT:
2015	* spin_lock_irq(pool->lock) which is released and regrabbed.
2016	*/
2017	static void process_one_work(struct worker worker, struct work_struct work)
2018	__releases(&pool->lock)
2019	__acquires(&pool->lock)
2020	{
2021	struct pool_workqueue *pwq = get_work_pwq(work);
2022	struct worker_pool *pool = worker->pool;
2023	bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2024	int work_color;
2025	struct worker *collision;
2026	#ifdef CONFIG_LOCKDEP
2027	/*
2028	* It is permissible to free the struct work_struct from
2029	* inside the function that is called from it, this we need to
2030	* take into account for lockdep too. To avoid bogus "held
2031	* lock freed" warnings as well as problems when looking into
2032	* work->lockdep_map, make a copy and use that here.
2033	*/
2034	struct lockdep_map lockdep_map;
2035
2036	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2037	#endif
2038	/* ensure we're on the correct CPU */
2039	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2040	raw_smp_processor_id() != pool->cpu);
2041
2042	/*
2043	* A single work shouldn't be executed concurrently by
2044	* multiple workers on a single cpu. Check whether anyone is
2045	* already processing the work. If so, defer the work to the
2046	* currently executing one.
2047	*/
2048	collision = find_worker_executing_work(pool, work);
2049	if (unlikely(collision)) {
2050	move_linked_works(work, &collision->scheduled, NULL);
2051	return;
2052	}
2053
2054	/* claim and dequeue */
2055	debug_work_deactivate(work);
2056	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2057	worker->current_work = work;
2058	worker->current_func = work->func;
2059	worker->current_pwq = pwq;
2060	work_color = get_work_color(work);
2061
2062	list_del_init(&work->entry);
2063
2064	/*
2065	* CPU intensive works don't participate in concurrency management.
2066	* They're the scheduler's responsibility. This takes @worker out
2067	* of concurrency management and the next code block will chain
2068	* execution of the pending work items.
2069	*/
2070	if (unlikely(cpu_intensive))
2071	worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2072
2073	/*
2074	* Wake up another worker if necessary. The condition is always
2075	* false for normal per-cpu workers since nr_running would always
2076	* be >= 1 at this point. This is used to chain execution of the
2077	* pending work items for WORKER_NOT_RUNNING workers such as the
2078	* UNBOUND and CPU_INTENSIVE ones.
2079	*/
2080	if (need_more_worker(pool))
2081	wake_up_worker(pool);
2082
2083	/*
2084	* Record the last pool and clear PENDING which should be the last
2085	* update to @work. Also, do this inside @pool->lock so that
2086	* PENDING and queued state changes happen together while IRQ is
2087	* disabled.
2088	*/
2089	set_work_pool_and_clear_pending(work, pool->id);
2090
2091	spin_unlock_irq(&pool->lock);
2092
2093	lock_map_acquire_read(&pwq->wq->lockdep_map);
2094	lock_map_acquire(&lockdep_map);
2095	trace_workqueue_execute_start(work);
2096	worker->current_func(work);
2097	/*
2098	* While we must be careful to not use "work" after this, the trace
2099	* point will only record its address.
2100	*/
2101	trace_workqueue_execute_end(work);
2102	lock_map_release(&lockdep_map);
2103	lock_map_release(&pwq->wq->lockdep_map);
2104
2105	if (unlikely(in_atomic() \|\| lockdep_depth(current) > 0)) {
2106	pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2107	" last function: %pf\n",
2108	current->comm, preempt_count(), task_pid_nr(current),
2109	worker->current_func);
2110	debug_show_held_locks(current);
2111	dump_stack();
2112	}
2113
2114	/*
2115	* The following prevents a kworker from hogging CPU on !PREEMPT
2116	* kernels, where a requeueing work item waiting for something to
2117	* happen could deadlock with stop_machine as such work item could
2118	* indefinitely requeue itself while all other CPUs are trapped in
2119	* stop_machine. At the same time, report a quiescent RCU state so
2120	* the same condition doesn't freeze RCU.
2121	*/
2122	cond_resched_rcu_qs();
2123
2124	spin_lock_irq(&pool->lock);
2125
2126	/* clear cpu intensive status */
2127	if (unlikely(cpu_intensive))
2128	worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2129
2130	/* we're done with it, release */
2131	hash_del(&worker->hentry);
2132	worker->current_work = NULL;
2133	worker->current_func = NULL;
2134	worker->current_pwq = NULL;
2135	worker->desc_valid = false;
2136	pwq_dec_nr_in_flight(pwq, work_color);
2137	}
2138
2139	/**
2140	* process_scheduled_works - process scheduled works
2141	* @worker: self
2142	*
2143	* Process all scheduled works. Please note that the scheduled list
2144	* may change while processing a work, so this function repeatedly
2145	* fetches a work from the top and executes it.
2146	*
2147	* CONTEXT:
2148	* spin_lock_irq(pool->lock) which may be released and regrabbed
2149	* multiple times.
2150	*/
2151	static void process_scheduled_works(struct worker *worker)
2152	{
2153	while (!list_empty(&worker->scheduled)) {
2154	struct work_struct *work = list_first_entry(&worker->scheduled,
2155	struct work_struct, entry);
2156	process_one_work(worker, work);
2157	}
2158	}
2159
2160	/**
2161	* worker_thread - the worker thread function
2162	* @__worker: self
2163	*
2164	* The worker thread function. All workers belong to a worker_pool -
2165	* either a per-cpu one or dynamic unbound one. These workers process all
2166	* work items regardless of their specific target workqueue. The only
2167	* exception is work items which belong to workqueues with a rescuer which
2168	* will be explained in rescuer_thread().
2169	*
2170	* Return: 0
2171	*/
2172	static int worker_thread(void *__worker)
2173	{
2174	struct worker *worker = __worker;
2175	struct worker_pool *pool = worker->pool;
2176
2177	/* tell the scheduler that this is a workqueue worker */
2178	worker->task->flags \|= PF_WQ_WORKER;
2179	woke_up:
2180	spin_lock_irq(&pool->lock);
2181
2182	/* am I supposed to die? */
2183	if (unlikely(worker->flags & WORKER_DIE)) {
2184	spin_unlock_irq(&pool->lock);
2185	WARN_ON_ONCE(!list_empty(&worker->entry));
2186	worker->task->flags &= ~PF_WQ_WORKER;
2187
2188	set_task_comm(worker->task, "kworker/dying");
2189	ida_simple_remove(&pool->worker_ida, worker->id);
2190	worker_detach_from_pool(worker, pool);
2191	kfree(worker);
2192	return 0;
2193	}
2194
2195	worker_leave_idle(worker);
2196	recheck:
2197	/* no more worker necessary? */
2198	if (!need_more_worker(pool))
2199	goto sleep;
2200
2201	/* do we need to manage? */
2202	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2203	goto recheck;
2204
2205	/*
2206	* ->scheduled list can only be filled while a worker is
2207	* preparing to process a work or actually processing it.
2208	* Make sure nobody diddled with it while I was sleeping.
2209	*/
2210	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2211
2212	/*
2213	* Finish PREP stage. We're guaranteed to have at least one idle
2214	* worker or that someone else has already assumed the manager
2215	* role. This is where @worker starts participating in concurrency
2216	* management if applicable and concurrency management is restored
2217	* after being rebound. See rebind_workers() for details.
2218	*/
2219	worker_clr_flags(worker, WORKER_PREP \| WORKER_REBOUND);
2220
2221	do {
2222	struct work_struct *work =
2223	list_first_entry(&pool->worklist,
2224	struct work_struct, entry);
2225
2226	pool->watchdog_ts = jiffies;
2227
2228	if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2229	/* optimization path, not strictly necessary */
2230	process_one_work(worker, work);
2231	if (unlikely(!list_empty(&worker->scheduled)))
2232	process_scheduled_works(worker);
2233	} else {
2234	move_linked_works(work, &worker->scheduled, NULL);
2235	process_scheduled_works(worker);
2236	}
2237	} while (keep_working(pool));
2238
2239	worker_set_flags(worker, WORKER_PREP);
2240	sleep:
2241	/*
2242	* pool->lock is held and there's no work to process and no need to
2243	* manage, sleep. Workers are woken up only while holding
2244	* pool->lock or from local cpu, so setting the current state
2245	* before releasing pool->lock is enough to prevent losing any
2246	* event.
2247	*/
2248	worker_enter_idle(worker);
2249	__set_current_state(TASK_INTERRUPTIBLE);
2250	spin_unlock_irq(&pool->lock);
2251	schedule();
2252	goto woke_up;
2253	}
2254
2255	/**
2256	* rescuer_thread - the rescuer thread function
2257	* @__rescuer: self
2258	*
2259	* Workqueue rescuer thread function. There's one rescuer for each
2260	* workqueue which has WQ_MEM_RECLAIM set.
2261	*
2262	* Regular work processing on a pool may block trying to create a new
2263	* worker which uses GFP_KERNEL allocation which has slight chance of
2264	* developing into deadlock if some works currently on the same queue
2265	* need to be processed to satisfy the GFP_KERNEL allocation. This is
2266	* the problem rescuer solves.
2267	*
2268	* When such condition is possible, the pool summons rescuers of all
2269	* workqueues which have works queued on the pool and let them process
2270	* those works so that forward progress can be guaranteed.
2271	*
2272	* This should happen rarely.
2273	*
2274	* Return: 0
2275	*/
2276	static int rescuer_thread(void *__rescuer)
2277	{
2278	struct worker *rescuer = __rescuer;
2279	struct workqueue_struct *wq = rescuer->rescue_wq;
2280	struct list_head *scheduled = &rescuer->scheduled;
2281	bool should_stop;
2282
2283	set_user_nice(current, RESCUER_NICE_LEVEL);
2284
2285	/*
2286	* Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2287	* doesn't participate in concurrency management.
2288	*/
2289	rescuer->task->flags \|= PF_WQ_WORKER;
2290	repeat:
2291	set_current_state(TASK_INTERRUPTIBLE);
2292
2293	/*
2294	* By the time the rescuer is requested to stop, the workqueue
2295	* shouldn't have any work pending, but @wq->maydays may still have
2296	* pwq(s) queued. This can happen by non-rescuer workers consuming
2297	* all the work items before the rescuer got to them. Go through
2298	* @wq->maydays processing before acting on should_stop so that the
2299	* list is always empty on exit.
2300	*/
2301	should_stop = kthread_should_stop();
2302
2303	/* see whether any pwq is asking for help */
2304	spin_lock_irq(&wq_mayday_lock);
2305
2306	while (!list_empty(&wq->maydays)) {
2307	struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2308	struct pool_workqueue, mayday_node);
2309	struct worker_pool *pool = pwq->pool;
2310	struct work_struct work, n;
2311	bool first = true;
2312
2313	__set_current_state(TASK_RUNNING);
2314	list_del_init(&pwq->mayday_node);
2315
2316	spin_unlock_irq(&wq_mayday_lock);
2317
2318	worker_attach_to_pool(rescuer, pool);
2319
2320	spin_lock_irq(&pool->lock);
2321	rescuer->pool = pool;
2322
2323	/*
2324	* Slurp in all works issued via this workqueue and
2325	* process'em.
2326	*/
2327	WARN_ON_ONCE(!list_empty(scheduled));
2328	list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2329	if (get_work_pwq(work) == pwq) {
2330	if (first)
2331	pool->watchdog_ts = jiffies;
2332	move_linked_works(work, scheduled, &n);
2333	}
2334	first = false;
2335	}
2336
2337	if (!list_empty(scheduled)) {
2338	process_scheduled_works(rescuer);
2339
2340	/*
2341	* The above execution of rescued work items could
2342	* have created more to rescue through
2343	* pwq_activate_first_delayed() or chained
2344	* queueing. Let's put @pwq back on mayday list so
2345	* that such back-to-back work items, which may be
2346	* being used to relieve memory pressure, don't
2347	* incur MAYDAY_INTERVAL delay inbetween.
2348	*/
2349	if (need_to_create_worker(pool)) {
2350	spin_lock(&wq_mayday_lock);
2351	get_pwq(pwq);
2352	list_move_tail(&pwq->mayday_node, &wq->maydays);
2353	spin_unlock(&wq_mayday_lock);
2354	}
2355	}
2356
2357	/*
2358	* Put the reference grabbed by send_mayday(). @pool won't
2359	* go away while we're still attached to it.
2360	*/
2361	put_pwq(pwq);
2362
2363	/*
2364	* Leave this pool. If need_more_worker() is %true, notify a
2365	* regular worker; otherwise, we end up with 0 concurrency
2366	* and stalling the execution.
2367	*/
2368	if (need_more_worker(pool))
2369	wake_up_worker(pool);
2370
2371	rescuer->pool = NULL;
2372	spin_unlock_irq(&pool->lock);
2373
2374	worker_detach_from_pool(rescuer, pool);
2375
2376	spin_lock_irq(&wq_mayday_lock);
2377	}
2378
2379	spin_unlock_irq(&wq_mayday_lock);
2380
2381	if (should_stop) {
2382	__set_current_state(TASK_RUNNING);
2383	rescuer->task->flags &= ~PF_WQ_WORKER;
2384	return 0;
2385	}
2386
2387	/* rescuers should never participate in concurrency management */
2388	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2389	schedule();
2390	goto repeat;
2391	}
2392
2393	/**
2394	* check_flush_dependency - check for flush dependency sanity
2395	* @target_wq: workqueue being flushed
2396	* @target_work: work item being flushed (NULL for workqueue flushes)
2397	*
2398	* %current is trying to flush the whole @target_wq or @target_work on it.
2399	* If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2400	* reclaiming memory or running on a workqueue which doesn't have
2401	* %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2402	* a deadlock.
2403	*/
2404	static void check_flush_dependency(struct workqueue_struct *target_wq,
2405	struct work_struct *target_work)
2406	{
2407	work_func_t target_func = target_work ? target_work->func : NULL;
2408	struct worker *worker;
2409
2410	if (target_wq->flags & WQ_MEM_RECLAIM)
2411	return;
2412
2413	worker = current_wq_worker();
2414
2415	WARN_ONCE(current->flags & PF_MEMALLOC,
2416	"workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2417	current->pid, current->comm, target_wq->name, target_func);
2418	WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2419	(WQ_MEM_RECLAIM \| __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2420	"workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2421	worker->current_pwq->wq->name, worker->current_func,
2422	target_wq->name, target_func);
2423	}
2424
2425	struct wq_barrier {
2426	struct work_struct work;
2427	struct completion done;
2428	struct task_struct task; / purely informational */
2429	};
2430
2431	static void wq_barrier_func(struct work_struct *work)
2432	{
2433	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2434	complete(&barr->done);
2435	}
2436
2437	/**
2438	* insert_wq_barrier - insert a barrier work
2439	* @pwq: pwq to insert barrier into
2440	* @barr: wq_barrier to insert
2441	* @target: target work to attach @barr to
2442	* @worker: worker currently executing @target, NULL if @target is not executing
2443	*
2444	* @barr is linked to @target such that @barr is completed only after
2445	* @target finishes execution. Please note that the ordering
2446	* guarantee is observed only with respect to @target and on the local
2447	* cpu.
2448	*
2449	* Currently, a queued barrier can't be canceled. This is because
2450	* try_to_grab_pending() can't determine whether the work to be
2451	* grabbed is at the head of the queue and thus can't clear LINKED
2452	* flag of the previous work while there must be a valid next work
2453	* after a work with LINKED flag set.
2454	*
2455	* Note that when @worker is non-NULL, @target may be modified
2456	* underneath us, so we can't reliably determine pwq from @target.
2457	*
2458	* CONTEXT:
2459	* spin_lock_irq(pool->lock).
2460	*/
2461	static void insert_wq_barrier(struct pool_workqueue *pwq,
2462	struct wq_barrier *barr,
2463	struct work_struct target, struct worker worker)
2464	{
2465	struct list_head *head;
2466	unsigned int linked = 0;
2467
2468	/*
2469	* debugobject calls are safe here even with pool->lock locked
2470	* as we know for sure that this will not trigger any of the
2471	* checks and call back into the fixup functions where we
2472	* might deadlock.
2473	*/
2474	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2475	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2476	init_completion(&barr->done);
2477	barr->task = current;
2478
2479	/*
2480	* If @target is currently being executed, schedule the
2481	* barrier to the worker; otherwise, put it after @target.
2482	*/
2483	if (worker)
2484	head = worker->scheduled.next;
2485	else {
2486	unsigned long *bits = work_data_bits(target);
2487
2488	head = target->entry.next;
2489	/* there can already be other linked works, inherit and set */
2490	linked = *bits & WORK_STRUCT_LINKED;
2491	__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2492	}
2493
2494	debug_work_activate(&barr->work);
2495	insert_work(pwq, &barr->work, head,
2496	work_color_to_flags(WORK_NO_COLOR) \| linked);
2497	}
2498
2499	/**
2500	* flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2501	* @wq: workqueue being flushed
2502	* @flush_color: new flush color, < 0 for no-op
2503	* @work_color: new work color, < 0 for no-op
2504	*
2505	* Prepare pwqs for workqueue flushing.
2506	*
2507	* If @flush_color is non-negative, flush_color on all pwqs should be
2508	* -1. If no pwq has in-flight commands at the specified color, all
2509	* pwq->flush_color's stay at -1 and %false is returned. If any pwq
2510	* has in flight commands, its pwq->flush_color is set to
2511	* @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2512	* wakeup logic is armed and %true is returned.
2513	*
2514	* The caller should have initialized @wq->first_flusher prior to
2515	* calling this function with non-negative @flush_color. If
2516	* @flush_color is negative, no flush color update is done and %false
2517	* is returned.
2518	*
2519	* If @work_color is non-negative, all pwqs should have the same
2520	* work_color which is previous to @work_color and all will be
2521	* advanced to @work_color.
2522	*
2523	* CONTEXT:
2524	* mutex_lock(wq->mutex).
2525	*
2526	* Return:
2527	* %true if @flush_color >= 0 and there's something to flush. %false
2528	* otherwise.
2529	*/
2530	static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2531	int flush_color, int work_color)
2532	{
2533	bool wait = false;
2534	struct pool_workqueue *pwq;
2535
2536	if (flush_color >= 0) {
2537	WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2538	atomic_set(&wq->nr_pwqs_to_flush, 1);
2539	}
2540
2541	for_each_pwq(pwq, wq) {
2542	struct worker_pool *pool = pwq->pool;
2543
2544	spin_lock_irq(&pool->lock);
2545
2546	if (flush_color >= 0) {
2547	WARN_ON_ONCE(pwq->flush_color != -1);
2548
2549	if (pwq->nr_in_flight[flush_color]) {
2550	pwq->flush_color = flush_color;
2551	atomic_inc(&wq->nr_pwqs_to_flush);
2552	wait = true;
2553	}
2554	}
2555
2556	if (work_color >= 0) {
2557	WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2558	pwq->work_color = work_color;
2559	}
2560
2561	spin_unlock_irq(&pool->lock);
2562	}
2563
2564	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2565	complete(&wq->first_flusher->done);
2566
2567	return wait;
2568	}
2569
2570	/**
2571	* flush_workqueue - ensure that any scheduled work has run to completion.
2572	* @wq: workqueue to flush
2573	*
2574	* This function sleeps until all work items which were queued on entry
2575	* have finished execution, but it is not livelocked by new incoming ones.
2576	*/
2577	void flush_workqueue(struct workqueue_struct *wq)
2578	{
2579	struct wq_flusher this_flusher = {
2580	.list = LIST_HEAD_INIT(this_flusher.list),
2581	.flush_color = -1,
2582	.done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2583	};
2584	int next_color;
2585
2586	lock_map_acquire(&wq->lockdep_map);
2587	lock_map_release(&wq->lockdep_map);
2588
2589	mutex_lock(&wq->mutex);
2590
2591	/*
2592	* Start-to-wait phase
2593	*/
2594	next_color = work_next_color(wq->work_color);
2595
2596	if (next_color != wq->flush_color) {
2597	/*
2598	* Color space is not full. The current work_color
2599	* becomes our flush_color and work_color is advanced
2600	* by one.
2601	*/
2602	WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2603	this_flusher.flush_color = wq->work_color;
2604	wq->work_color = next_color;
2605
2606	if (!wq->first_flusher) {
2607	/* no flush in progress, become the first flusher */
2608	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2609
2610	wq->first_flusher = &this_flusher;
2611
2612	if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2613	wq->work_color)) {
2614	/* nothing to flush, done */
2615	wq->flush_color = next_color;
2616	wq->first_flusher = NULL;
2617	goto out_unlock;
2618	}
2619	} else {
2620	/* wait in queue */
2621	WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2622	list_add_tail(&this_flusher.list, &wq->flusher_queue);
2623	flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2624	}
2625	} else {
2626	/*
2627	* Oops, color space is full, wait on overflow queue.
2628	* The next flush completion will assign us
2629	* flush_color and transfer to flusher_queue.
2630	*/
2631	list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2632	}
2633
2634	check_flush_dependency(wq, NULL);
2635
2636	mutex_unlock(&wq->mutex);
2637
2638	wait_for_completion(&this_flusher.done);
2639
2640	/*
2641	* Wake-up-and-cascade phase
2642	*
2643	* First flushers are responsible for cascading flushes and
2644	* handling overflow. Non-first flushers can simply return.
2645	*/
2646	if (wq->first_flusher != &this_flusher)
2647	return;
2648
2649	mutex_lock(&wq->mutex);
2650
2651	/* we might have raced, check again with mutex held */
2652	if (wq->first_flusher != &this_flusher)
2653	goto out_unlock;
2654
2655	wq->first_flusher = NULL;
2656
2657	WARN_ON_ONCE(!list_empty(&this_flusher.list));
2658	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2659
2660	while (true) {
2661	struct wq_flusher next, tmp;
2662
2663	/* complete all the flushers sharing the current flush color */
2664	list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2665	if (next->flush_color != wq->flush_color)
2666	break;
2667	list_del_init(&next->list);
2668	complete(&next->done);
2669	}
2670
2671	WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2672	wq->flush_color != work_next_color(wq->work_color));
2673
2674	/* this flush_color is finished, advance by one */
2675	wq->flush_color = work_next_color(wq->flush_color);
2676
2677	/* one color has been freed, handle overflow queue */
2678	if (!list_empty(&wq->flusher_overflow)) {
2679	/*
2680	* Assign the same color to all overflowed
2681	* flushers, advance work_color and append to
2682	* flusher_queue. This is the start-to-wait
2683	* phase for these overflowed flushers.
2684	*/
2685	list_for_each_entry(tmp, &wq->flusher_overflow, list)
2686	tmp->flush_color = wq->work_color;
2687
2688	wq->work_color = work_next_color(wq->work_color);
2689
2690	list_splice_tail_init(&wq->flusher_overflow,
2691	&wq->flusher_queue);
2692	flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2693	}
2694
2695	if (list_empty(&wq->flusher_queue)) {
2696	WARN_ON_ONCE(wq->flush_color != wq->work_color);
2697	break;
2698	}
2699
2700	/*
2701	* Need to flush more colors. Make the next flusher
2702	* the new first flusher and arm pwqs.
2703	*/
2704	WARN_ON_ONCE(wq->flush_color == wq->work_color);
2705	WARN_ON_ONCE(wq->flush_color != next->flush_color);
2706
2707	list_del_init(&next->list);
2708	wq->first_flusher = next;
2709
2710	if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2711	break;
2712
2713	/*
2714	* Meh... this color is already done, clear first
2715	* flusher and repeat cascading.
2716	*/
2717	wq->first_flusher = NULL;
2718	}
2719
2720	out_unlock:
2721	mutex_unlock(&wq->mutex);
2722	}
2723	EXPORT_SYMBOL(flush_workqueue);
2724
2725	/**
2726	* drain_workqueue - drain a workqueue
2727	* @wq: workqueue to drain
2728	*
2729	* Wait until the workqueue becomes empty. While draining is in progress,
2730	* only chain queueing is allowed. IOW, only currently pending or running
2731	* work items on @wq can queue further work items on it. @wq is flushed
2732	* repeatedly until it becomes empty. The number of flushing is determined
2733	* by the depth of chaining and should be relatively short. Whine if it
2734	* takes too long.
2735	*/
2736	void drain_workqueue(struct workqueue_struct *wq)
2737	{
2738	unsigned int flush_cnt = 0;
2739	struct pool_workqueue *pwq;
2740
2741	/*
2742	* __queue_work() needs to test whether there are drainers, is much
2743	* hotter than drain_workqueue() and already looks at @wq->flags.
2744	* Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2745	*/
2746	mutex_lock(&wq->mutex);
2747	if (!wq->nr_drainers++)
2748	wq->flags \|= __WQ_DRAINING;
2749	mutex_unlock(&wq->mutex);
2750	reflush:
2751	flush_workqueue(wq);
2752
2753	mutex_lock(&wq->mutex);
2754
2755	for_each_pwq(pwq, wq) {
2756	bool drained;
2757
2758	spin_lock_irq(&pwq->pool->lock);
2759	drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2760	spin_unlock_irq(&pwq->pool->lock);
2761
2762	if (drained)
2763	continue;
2764
2765	if (++flush_cnt == 10 \|\|
2766	(flush_cnt % 100 == 0 && flush_cnt <= 1000))
2767	pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2768	wq->name, flush_cnt);
2769
2770	mutex_unlock(&wq->mutex);
2771	goto reflush;
2772	}
2773
2774	if (!--wq->nr_drainers)
2775	wq->flags &= ~__WQ_DRAINING;
2776	mutex_unlock(&wq->mutex);
2777	}
2778	EXPORT_SYMBOL_GPL(drain_workqueue);
2779
2780	static bool start_flush_work(struct work_struct work, struct wq_barrier barr)
2781	{
2782	struct worker *worker = NULL;
2783	struct worker_pool *pool;
2784	struct pool_workqueue *pwq;
2785
2786	might_sleep();
2787
2788	local_irq_disable();
2789	pool = get_work_pool(work);
2790	if (!pool) {
2791	local_irq_enable();
2792	return false;
2793	}
2794
2795	spin_lock(&pool->lock);
2796	/* see the comment in try_to_grab_pending() with the same code */
2797	pwq = get_work_pwq(work);
2798	if (pwq) {
2799	if (unlikely(pwq->pool != pool))
2800	goto already_gone;
2801	} else {
2802	worker = find_worker_executing_work(pool, work);
2803	if (!worker)
2804	goto already_gone;
2805	pwq = worker->current_pwq;
2806	}
2807
2808	check_flush_dependency(pwq->wq, work);
2809
2810	insert_wq_barrier(pwq, barr, work, worker);
2811	spin_unlock_irq(&pool->lock);
2812
2813	/*
2814	* If @max_active is 1 or rescuer is in use, flushing another work
2815	* item on the same workqueue may lead to deadlock. Make sure the
2816	* flusher is not running on the same workqueue by verifying write
2817	* access.
2818	*/
2819	if (pwq->wq->saved_max_active == 1 \|\| pwq->wq->rescuer)
2820	lock_map_acquire(&pwq->wq->lockdep_map);
2821	else
2822	lock_map_acquire_read(&pwq->wq->lockdep_map);
2823	lock_map_release(&pwq->wq->lockdep_map);
2824
2825	return true;
2826	already_gone:
2827	spin_unlock_irq(&pool->lock);
2828	return false;
2829	}
2830
2831	/**
2832	* flush_work - wait for a work to finish executing the last queueing instance
2833	* @work: the work to flush
2834	*
2835	* Wait until @work has finished execution. @work is guaranteed to be idle
2836	* on return if it hasn't been requeued since flush started.
2837	*
2838	* Return:
2839	* %true if flush_work() waited for the work to finish execution,
2840	* %false if it was already idle.
2841	*/
2842	bool flush_work(struct work_struct *work)
2843	{
2844	struct wq_barrier barr;
2845
2846	lock_map_acquire(&work->lockdep_map);
2847	lock_map_release(&work->lockdep_map);
2848
2849	if (start_flush_work(work, &barr)) {
2850	wait_for_completion(&barr.done);
2851	destroy_work_on_stack(&barr.work);
2852	return true;
2853	} else {
2854	return false;
2855	}
2856	}
2857	EXPORT_SYMBOL_GPL(flush_work);
2858
2859	struct cwt_wait {
2860	wait_queue_t wait;
2861	struct work_struct *work;
2862	};
2863
2864	static int cwt_wakefn(wait_queue_t wait, unsigned mode, int sync, void key)
2865	{
2866	struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2867
2868	if (cwait->work != key)
2869	return 0;
2870	return autoremove_wake_function(wait, mode, sync, key);
2871	}
2872
2873	static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2874	{
2875	static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2876	unsigned long flags;
2877	int ret;
2878
2879	do {
2880	ret = try_to_grab_pending(work, is_dwork, &flags);
2881	/*
2882	* If someone else is already canceling, wait for it to
2883	* finish. flush_work() doesn't work for PREEMPT_NONE
2884	* because we may get scheduled between @work's completion
2885	* and the other canceling task resuming and clearing
2886	* CANCELING - flush_work() will return false immediately
2887	* as @work is no longer busy, try_to_grab_pending() will
2888	* return -ENOENT as @work is still being canceled and the
2889	* other canceling task won't be able to clear CANCELING as
2890	* we're hogging the CPU.
2891	*
2892	* Let's wait for completion using a waitqueue. As this
2893	* may lead to the thundering herd problem, use a custom
2894	* wake function which matches @work along with exclusive
2895	* wait and wakeup.
2896	*/
2897	if (unlikely(ret == -ENOENT)) {
2898	struct cwt_wait cwait;
2899
2900	init_wait(&cwait.wait);
2901	cwait.wait.func = cwt_wakefn;
2902	cwait.work = work;
2903
2904	prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2905	TASK_UNINTERRUPTIBLE);
2906	if (work_is_canceling(work))
2907	schedule();
2908	finish_wait(&cancel_waitq, &cwait.wait);
2909	}
2910	} while (unlikely(ret < 0));
2911
2912	/* tell other tasks trying to grab @work to back off */
2913	mark_work_canceling(work);
2914	local_irq_restore(flags);
2915
2916	flush_work(work);
2917	clear_work_data(work);
2918
2919	/*
2920	* Paired with prepare_to_wait() above so that either
2921	* waitqueue_active() is visible here or !work_is_canceling() is
2922	* visible there.
2923	*/
2924	smp_mb();
2925	if (waitqueue_active(&cancel_waitq))
2926	__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2927
2928	return ret;
2929	}
2930
2931	/**
2932	* cancel_work_sync - cancel a work and wait for it to finish
2933	* @work: the work to cancel
2934	*
2935	* Cancel @work and wait for its execution to finish. This function
2936	* can be used even if the work re-queues itself or migrates to
2937	* another workqueue. On return from this function, @work is
2938	* guaranteed to be not pending or executing on any CPU.
2939	*
2940	* cancel_work_sync(&delayed_work->work) must not be used for
2941	* delayed_work's. Use cancel_delayed_work_sync() instead.
2942	*
2943	* The caller must ensure that the workqueue on which @work was last
2944	* queued can't be destroyed before this function returns.
2945	*
2946	* Return:
2947	* %true if @work was pending, %false otherwise.
2948	*/
2949	bool cancel_work_sync(struct work_struct *work)
2950	{
2951	return __cancel_work_timer(work, false);
2952	}
2953	EXPORT_SYMBOL_GPL(cancel_work_sync);
2954
2955	/**
2956	* flush_delayed_work - wait for a dwork to finish executing the last queueing
2957	* @dwork: the delayed work to flush
2958	*
2959	* Delayed timer is cancelled and the pending work is queued for
2960	* immediate execution. Like flush_work(), this function only
2961	* considers the last queueing instance of @dwork.
2962	*
2963	* Return:
2964	* %true if flush_work() waited for the work to finish execution,
2965	* %false if it was already idle.
2966	*/
2967	bool flush_delayed_work(struct delayed_work *dwork)
2968	{
2969	local_irq_disable();
2970	if (del_timer_sync(&dwork->timer))
2971	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
2972	local_irq_enable();
2973	return flush_work(&dwork->work);
2974	}
2975	EXPORT_SYMBOL(flush_delayed_work);
2976
2977	/**
2978	* cancel_delayed_work - cancel a delayed work
2979	* @dwork: delayed_work to cancel
2980	*
2981	* Kill off a pending delayed_work.
2982	*
2983	* Return: %true if @dwork was pending and canceled; %false if it wasn't
2984	* pending.
2985	*
2986	* Note:
2987	* The work callback function may still be running on return, unless
2988	* it returns %true and the work doesn't re-arm itself. Explicitly flush or
2989	* use cancel_delayed_work_sync() to wait on it.
2990	*
2991	* This function is safe to call from any context including IRQ handler.
2992	*/
2993	bool cancel_delayed_work(struct delayed_work *dwork)
2994	{
2995	unsigned long flags;
2996	int ret;
2997
2998	do {
2999	ret = try_to_grab_pending(&dwork->work, true, &flags);
3000	} while (unlikely(ret == -EAGAIN));
3001
3002	if (unlikely(ret < 0))
3003	return false;
3004
3005	set_work_pool_and_clear_pending(&dwork->work,
3006	get_work_pool_id(&dwork->work));
3007	local_irq_restore(flags);
3008	return ret;
3009	}
3010	EXPORT_SYMBOL(cancel_delayed_work);
3011
3012	/**
3013	* cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3014	* @dwork: the delayed work cancel
3015	*
3016	* This is cancel_work_sync() for delayed works.
3017	*
3018	* Return:
3019	* %true if @dwork was pending, %false otherwise.
3020	*/
3021	bool cancel_delayed_work_sync(struct delayed_work *dwork)
3022	{
3023	return __cancel_work_timer(&dwork->work, true);
3024	}
3025	EXPORT_SYMBOL(cancel_delayed_work_sync);
3026
3027	/**
3028	* schedule_on_each_cpu - execute a function synchronously on each online CPU
3029	* @func: the function to call
3030	*
3031	* schedule_on_each_cpu() executes @func on each online CPU using the
3032	* system workqueue and blocks until all CPUs have completed.
3033	* schedule_on_each_cpu() is very slow.
3034	*
3035	* Return:
3036	* 0 on success, -errno on failure.
3037	*/
3038	int schedule_on_each_cpu(work_func_t func)
3039	{
3040	int cpu;
3041	struct work_struct __percpu *works;
3042
3043	works = alloc_percpu(struct work_struct);
3044	if (!works)
3045	return -ENOMEM;
3046
3047	get_online_cpus();
3048
3049	for_each_online_cpu(cpu) {
3050	struct work_struct *work = per_cpu_ptr(works, cpu);
3051
3052	INIT_WORK(work, func);
3053	schedule_work_on(cpu, work);
3054	}
3055
3056	for_each_online_cpu(cpu)
3057	flush_work(per_cpu_ptr(works, cpu));
3058
3059	put_online_cpus();
3060	free_percpu(works);
3061	return 0;
3062	}
3063
3064	/**
3065	* execute_in_process_context - reliably execute the routine with user context
3066	* @fn: the function to execute
3067	* @ew: guaranteed storage for the execute work structure (must
3068	* be available when the work executes)
3069	*
3070	* Executes the function immediately if process context is available,
3071	* otherwise schedules the function for delayed execution.
3072	*
3073	* Return: 0 - function was executed
3074	* 1 - function was scheduled for execution
3075	*/
3076	int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3077	{
3078	if (!in_interrupt()) {
3079	fn(&ew->work);
3080	return 0;
3081	}
3082
3083	INIT_WORK(&ew->work, fn);
3084	schedule_work(&ew->work);
3085
3086	return 1;
3087	}
3088	EXPORT_SYMBOL_GPL(execute_in_process_context);
3089
3090	/**
3091	* free_workqueue_attrs - free a workqueue_attrs
3092	* @attrs: workqueue_attrs to free
3093	*
3094	* Undo alloc_workqueue_attrs().
3095	*/
3096	void free_workqueue_attrs(struct workqueue_attrs *attrs)
3097	{
3098	if (attrs) {
3099	free_cpumask_var(attrs->cpumask);
3100	kfree(attrs);
3101	}
3102	}
3103
3104	/**
3105	* alloc_workqueue_attrs - allocate a workqueue_attrs
3106	* @gfp_mask: allocation mask to use
3107	*
3108	* Allocate a new workqueue_attrs, initialize with default settings and
3109	* return it.
3110	*
3111	* Return: The allocated new workqueue_attr on success. %NULL on failure.
3112	*/
3113	struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3114	{
3115	struct workqueue_attrs *attrs;
3116
3117	attrs = kzalloc(sizeof(*attrs), gfp_mask);
3118	if (!attrs)
3119	goto fail;
3120	if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3121	goto fail;
3122
3123	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3124	return attrs;
3125	fail:
3126	free_workqueue_attrs(attrs);
3127	return NULL;
3128	}
3129
3130	static void copy_workqueue_attrs(struct workqueue_attrs *to,
3131	const struct workqueue_attrs *from)
3132	{
3133	to->nice = from->nice;
3134	cpumask_copy(to->cpumask, from->cpumask);
3135	/*
3136	* Unlike hash and equality test, this function doesn't ignore
3137	* ->no_numa as it is used for both pool and wq attrs. Instead,
3138	* get_unbound_pool() explicitly clears ->no_numa after copying.
3139	*/
3140	to->no_numa = from->no_numa;
3141	}
3142
3143	/* hash value of the content of @attr */
3144	static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3145	{
3146	u32 hash = 0;
3147
3148	hash = jhash_1word(attrs->nice, hash);
3149	hash = jhash(cpumask_bits(attrs->cpumask),
3150	BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3151	return hash;
3152	}
3153
3154	/* content equality test */
3155	static bool wqattrs_equal(const struct workqueue_attrs *a,
3156	const struct workqueue_attrs *b)
3157	{
3158	if (a->nice != b->nice)
3159	return false;
3160	if (!cpumask_equal(a->cpumask, b->cpumask))
3161	return false;
3162	return true;
3163	}
3164
3165	/**
3166	* init_worker_pool - initialize a newly zalloc'd worker_pool
3167	* @pool: worker_pool to initialize
3168	*
3169	* Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3170	*
3171	* Return: 0 on success, -errno on failure. Even on failure, all fields
3172	* inside @pool proper are initialized and put_unbound_pool() can be called
3173	* on @pool safely to release it.
3174	*/
3175	static int init_worker_pool(struct worker_pool *pool)
3176	{
3177	spin_lock_init(&pool->lock);
3178	pool->id = -1;
3179	pool->cpu = -1;
3180	pool->node = NUMA_NO_NODE;
3181	pool->flags \|= POOL_DISASSOCIATED;
3182	pool->watchdog_ts = jiffies;
3183	INIT_LIST_HEAD(&pool->worklist);
3184	INIT_LIST_HEAD(&pool->idle_list);
3185	hash_init(pool->busy_hash);
3186
3187	init_timer_deferrable(&pool->idle_timer);
3188	pool->idle_timer.function = idle_worker_timeout;
3189	pool->idle_timer.data = (unsigned long)pool;
3190
3191	setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3192	(unsigned long)pool);
3193
3194	mutex_init(&pool->manager_arb);
3195	mutex_init(&pool->attach_mutex);
3196	INIT_LIST_HEAD(&pool->workers);
3197
3198	ida_init(&pool->worker_ida);
3199	INIT_HLIST_NODE(&pool->hash_node);
3200	pool->refcnt = 1;
3201
3202	/* shouldn't fail above this point */
3203	pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3204	if (!pool->attrs)
3205	return -ENOMEM;
3206	return 0;
3207	}
3208
3209	static void rcu_free_wq(struct rcu_head *rcu)
3210	{
3211	struct workqueue_struct *wq =
3212	container_of(rcu, struct workqueue_struct, rcu);
3213
3214	if (!(wq->flags & WQ_UNBOUND))
3215	free_percpu(wq->cpu_pwqs);
3216	else
3217	free_workqueue_attrs(wq->unbound_attrs);
3218
3219	kfree(wq->rescuer);
3220	kfree(wq);
3221	}
3222
3223	static void rcu_free_pool(struct rcu_head *rcu)
3224	{
3225	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3226
3227	ida_destroy(&pool->worker_ida);
3228	free_workqueue_attrs(pool->attrs);
3229	kfree(pool);
3230	}
3231
3232	/**
3233	* put_unbound_pool - put a worker_pool
3234	* @pool: worker_pool to put
3235	*
3236	* Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3237	* safe manner. get_unbound_pool() calls this function on its failure path
3238	* and this function should be able to release pools which went through,
3239	* successfully or not, init_worker_pool().
3240	*
3241	* Should be called with wq_pool_mutex held.
3242	*/
3243	static void put_unbound_pool(struct worker_pool *pool)
3244	{
3245	DECLARE_COMPLETION_ONSTACK(detach_completion);
3246	struct worker *worker;
3247
3248	lockdep_assert_held(&wq_pool_mutex);
3249
3250	if (--pool->refcnt)
3251	return;
3252
3253	/* sanity checks */
3254	if (WARN_ON(!(pool->cpu < 0)) \|\|
3255	WARN_ON(!list_empty(&pool->worklist)))
3256	return;
3257
3258	/* release id and unhash */
3259	if (pool->id >= 0)
3260	idr_remove(&worker_pool_idr, pool->id);
3261	hash_del(&pool->hash_node);
3262
3263	/*
3264	* Become the manager and destroy all workers. Grabbing
3265	* manager_arb prevents @pool's workers from blocking on
3266	* attach_mutex.
3267	*/
3268	mutex_lock(&pool->manager_arb);
3269
3270	spin_lock_irq(&pool->lock);
3271	while ((worker = first_idle_worker(pool)))
3272	destroy_worker(worker);
3273	WARN_ON(pool->nr_workers \|\| pool->nr_idle);
3274	spin_unlock_irq(&pool->lock);
3275
3276	mutex_lock(&pool->attach_mutex);
3277	if (!list_empty(&pool->workers))
3278	pool->detach_completion = &detach_completion;
3279	mutex_unlock(&pool->attach_mutex);
3280
3281	if (pool->detach_completion)
3282	wait_for_completion(pool->detach_completion);
3283
3284	mutex_unlock(&pool->manager_arb);
3285
3286	/* shut down the timers */
3287	del_timer_sync(&pool->idle_timer);
3288	del_timer_sync(&pool->mayday_timer);
3289
3290	/* sched-RCU protected to allow dereferences from get_work_pool() */
3291	call_rcu_sched(&pool->rcu, rcu_free_pool);
3292	}
3293
3294	/**
3295	* get_unbound_pool - get a worker_pool with the specified attributes
3296	* @attrs: the attributes of the worker_pool to get
3297	*
3298	* Obtain a worker_pool which has the same attributes as @attrs, bump the
3299	* reference count and return it. If there already is a matching
3300	* worker_pool, it will be used; otherwise, this function attempts to
3301	* create a new one.
3302	*
3303	* Should be called with wq_pool_mutex held.
3304	*
3305	* Return: On success, a worker_pool with the same attributes as @attrs.
3306	* On failure, %NULL.
3307	*/
3308	static struct worker_pool get_unbound_pool(const struct workqueue_attrs attrs)
3309	{
3310	u32 hash = wqattrs_hash(attrs);
3311	struct worker_pool *pool;
3312	int node;
3313	int target_node = NUMA_NO_NODE;
3314
3315	lockdep_assert_held(&wq_pool_mutex);
3316
3317	/* do we already have a matching pool? */
3318	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3319	if (wqattrs_equal(pool->attrs, attrs)) {
3320	pool->refcnt++;
3321	return pool;
3322	}
3323	}
3324
3325	/* if cpumask is contained inside a NUMA node, we belong to that node */
3326	if (wq_numa_enabled) {
3327	for_each_node(node) {
3328	if (cpumask_subset(attrs->cpumask,
3329	wq_numa_possible_cpumask[node])) {
3330	target_node = node;
3331	break;
3332	}
3333	}
3334	}
3335
3336	/* nope, create a new one */
3337	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3338	if (!pool \|\| init_worker_pool(pool) < 0)
3339	goto fail;
3340
3341	lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3342	copy_workqueue_attrs(pool->attrs, attrs);
3343	pool->node = target_node;
3344
3345	/*
3346	* no_numa isn't a worker_pool attribute, always clear it. See
3347	* 'struct workqueue_attrs' comments for detail.
3348	*/
3349	pool->attrs->no_numa = false;
3350
3351	if (worker_pool_assign_id(pool) < 0)
3352	goto fail;
3353
3354	/* create and start the initial worker */
3355	if (!create_worker(pool))
3356	goto fail;
3357
3358	/* install */
3359	hash_add(unbound_pool_hash, &pool->hash_node, hash);
3360
3361	return pool;
3362	fail:
3363	if (pool)
3364	put_unbound_pool(pool);
3365	return NULL;
3366	}
3367
3368	static void rcu_free_pwq(struct rcu_head *rcu)
3369	{
3370	kmem_cache_free(pwq_cache,
3371	container_of(rcu, struct pool_workqueue, rcu));
3372	}
3373
3374	/*
3375	* Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3376	* and needs to be destroyed.
3377	*/
3378	static void pwq_unbound_release_workfn(struct work_struct *work)
3379	{
3380	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3381	unbound_release_work);
3382	struct workqueue_struct *wq = pwq->wq;
3383	struct worker_pool *pool = pwq->pool;
3384	bool is_last;
3385
3386	if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3387	return;
3388
3389	mutex_lock(&wq->mutex);
3390	list_del_rcu(&pwq->pwqs_node);
3391	is_last = list_empty(&wq->pwqs);
3392	mutex_unlock(&wq->mutex);
3393
3394	mutex_lock(&wq_pool_mutex);
3395	put_unbound_pool(pool);
3396	mutex_unlock(&wq_pool_mutex);
3397
3398	call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3399
3400	/*
3401	* If we're the last pwq going away, @wq is already dead and no one
3402	* is gonna access it anymore. Schedule RCU free.
3403	*/
3404	if (is_last)
3405	call_rcu_sched(&wq->rcu, rcu_free_wq);
3406	}
3407
3408	/**
3409	* pwq_adjust_max_active - update a pwq's max_active to the current setting
3410	* @pwq: target pool_workqueue
3411	*
3412	* If @pwq isn't freezing, set @pwq->max_active to the associated
3413	* workqueue's saved_max_active and activate delayed work items
3414	* accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3415	*/
3416	static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3417	{
3418	struct workqueue_struct *wq = pwq->wq;
3419	bool freezable = wq->flags & WQ_FREEZABLE;
3420
3421	/* for @wq->saved_max_active */
3422	lockdep_assert_held(&wq->mutex);
3423
3424	/* fast exit for non-freezable wqs */
3425	if (!freezable && pwq->max_active == wq->saved_max_active)
3426	return;
3427
3428	spin_lock_irq(&pwq->pool->lock);
3429
3430	/*
3431	* During [un]freezing, the caller is responsible for ensuring that
3432	* this function is called at least once after @workqueue_freezing
3433	* is updated and visible.
3434	*/
3435	if (!freezable \|\| !workqueue_freezing) {
3436	pwq->max_active = wq->saved_max_active;
3437
3438	while (!list_empty(&pwq->delayed_works) &&
3439	pwq->nr_active < pwq->max_active)
3440	pwq_activate_first_delayed(pwq);
3441
3442	/*
3443	* Need to kick a worker after thawed or an unbound wq's
3444	* max_active is bumped. It's a slow path. Do it always.
3445	*/
3446	wake_up_worker(pwq->pool);
3447	} else {
3448	pwq->max_active = 0;
3449	}
3450
3451	spin_unlock_irq(&pwq->pool->lock);
3452	}
3453
3454	/* initialize newly alloced @pwq which is associated with @wq and @pool */
3455	static void init_pwq(struct pool_workqueue pwq, struct workqueue_struct wq,
3456	struct worker_pool *pool)
3457	{
3458	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3459
3460	memset(pwq, 0, sizeof(*pwq));
3461
3462	pwq->pool = pool;
3463	pwq->wq = wq;
3464	pwq->flush_color = -1;
3465	pwq->refcnt = 1;
3466	INIT_LIST_HEAD(&pwq->delayed_works);
3467	INIT_LIST_HEAD(&pwq->pwqs_node);
3468	INIT_LIST_HEAD(&pwq->mayday_node);
3469	INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3470	}
3471
3472	/* sync @pwq with the current state of its associated wq and link it */
3473	static void link_pwq(struct pool_workqueue *pwq)
3474	{
3475	struct workqueue_struct *wq = pwq->wq;
3476
3477	lockdep_assert_held(&wq->mutex);
3478
3479	/* may be called multiple times, ignore if already linked */
3480	if (!list_empty(&pwq->pwqs_node))
3481	return;
3482
3483	/* set the matching work_color */
3484	pwq->work_color = wq->work_color;
3485
3486	/* sync max_active to the current setting */
3487	pwq_adjust_max_active(pwq);
3488
3489	/* link in @pwq */
3490	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3491	}
3492
3493	/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3494	static struct pool_workqueue alloc_unbound_pwq(struct workqueue_struct wq,
3495	const struct workqueue_attrs *attrs)
3496	{
3497	struct worker_pool *pool;
3498	struct pool_workqueue *pwq;
3499
3500	lockdep_assert_held(&wq_pool_mutex);
3501
3502	pool = get_unbound_pool(attrs);
3503	if (!pool)
3504	return NULL;
3505
3506	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3507	if (!pwq) {
3508	put_unbound_pool(pool);
3509	return NULL;
3510	}
3511
3512	init_pwq(pwq, wq, pool);
3513	return pwq;
3514	}
3515
3516	/**
3517	* wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3518	* @attrs: the wq_attrs of the default pwq of the target workqueue
3519	* @node: the target NUMA node
3520	* @cpu_going_down: if >= 0, the CPU to consider as offline
3521	* @cpumask: outarg, the resulting cpumask
3522	*
3523	* Calculate the cpumask a workqueue with @attrs should use on @node. If
3524	* @cpu_going_down is >= 0, that cpu is considered offline during
3525	* calculation. The result is stored in @cpumask.
3526	*
3527	* If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3528	* enabled and @node has online CPUs requested by @attrs, the returned
3529	* cpumask is the intersection of the possible CPUs of @node and
3530	* @attrs->cpumask.
3531	*
3532	* The caller is responsible for ensuring that the cpumask of @node stays
3533	* stable.
3534	*
3535	* Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3536	* %false if equal.
3537	*/
3538	static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3539	int cpu_going_down, cpumask_t *cpumask)
3540	{
3541	if (!wq_numa_enabled \|\| attrs->no_numa)
3542	goto use_dfl;
3543
3544	/* does @node have any online CPUs @attrs wants? */
3545	cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3546	if (cpu_going_down >= 0)
3547	cpumask_clear_cpu(cpu_going_down, cpumask);
3548
3549	if (cpumask_empty(cpumask))
3550	goto use_dfl;
3551
3552	/* yeap, return possible CPUs in @node that @attrs wants */
3553	cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3554	return !cpumask_equal(cpumask, attrs->cpumask);
3555
3556	use_dfl:
3557	cpumask_copy(cpumask, attrs->cpumask);
3558	return false;
3559	}
3560
3561	/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3562	static struct pool_workqueue numa_pwq_tbl_install(struct workqueue_struct wq,
3563	int node,
3564	struct pool_workqueue *pwq)
3565	{
3566	struct pool_workqueue *old_pwq;
3567
3568	lockdep_assert_held(&wq_pool_mutex);
3569	lockdep_assert_held(&wq->mutex);
3570
3571	/* link_pwq() can handle duplicate calls */
3572	link_pwq(pwq);
3573
3574	old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3575	rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3576	return old_pwq;
3577	}
3578
3579	/* context to store the prepared attrs & pwqs before applying */
3580	struct apply_wqattrs_ctx {
3581	struct workqueue_struct wq; / target workqueue */
3582	struct workqueue_attrs attrs; / attrs to apply */
3583	struct list_head list; /* queued for batching commit */
3584	struct pool_workqueue *dfl_pwq;
3585	struct pool_workqueue *pwq_tbl[];
3586	};
3587
3588	/* free the resources after success or abort */
3589	static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3590	{
3591	if (ctx) {
3592	int node;
3593
3594	for_each_node(node)
3595	put_pwq_unlocked(ctx->pwq_tbl[node]);
3596	put_pwq_unlocked(ctx->dfl_pwq);
3597
3598	free_workqueue_attrs(ctx->attrs);
3599
3600	kfree(ctx);
3601	}
3602	}
3603
3604	/* allocate the attrs and pwqs for later installation */
3605	static struct apply_wqattrs_ctx *
3606	apply_wqattrs_prepare(struct workqueue_struct *wq,
3607	const struct workqueue_attrs *attrs)
3608	{
3609	struct apply_wqattrs_ctx *ctx;
3610	struct workqueue_attrs new_attrs, tmp_attrs;
3611	int node;
3612
3613	lockdep_assert_held(&wq_pool_mutex);
3614
3615	ctx = kzalloc(sizeof(ctx) + nr_node_ids sizeof(ctx->pwq_tbl[0]),
3616	GFP_KERNEL);
3617
3618	new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3619	tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3620	if (!ctx \|\| !new_attrs \|\| !tmp_attrs)
3621	goto out_free;
3622
3623	/*
3624	* Calculate the attrs of the default pwq.
3625	* If the user configured cpumask doesn't overlap with the
3626	* wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3627	*/
3628	copy_workqueue_attrs(new_attrs, attrs);
3629	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3630	if (unlikely(cpumask_empty(new_attrs->cpumask)))
3631	cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3632
3633	/*
3634	* We may create multiple pwqs with differing cpumasks. Make a
3635	* copy of @new_attrs which will be modified and used to obtain
3636	* pools.
3637	*/
3638	copy_workqueue_attrs(tmp_attrs, new_attrs);
3639
3640	/*
3641	* If something goes wrong during CPU up/down, we'll fall back to
3642	* the default pwq covering whole @attrs->cpumask. Always create
3643	* it even if we don't use it immediately.
3644	*/
3645	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3646	if (!ctx->dfl_pwq)
3647	goto out_free;
3648
3649	for_each_node(node) {
3650	if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3651	ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3652	if (!ctx->pwq_tbl[node])
3653	goto out_free;
3654	} else {
3655	ctx->dfl_pwq->refcnt++;
3656	ctx->pwq_tbl[node] = ctx->dfl_pwq;
3657	}
3658	}
3659
3660	/* save the user configured attrs and sanitize it. */
3661	copy_workqueue_attrs(new_attrs, attrs);
3662	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3663	ctx->attrs = new_attrs;
3664
3665	ctx->wq = wq;
3666	free_workqueue_attrs(tmp_attrs);
3667	return ctx;
3668
3669	out_free:
3670	free_workqueue_attrs(tmp_attrs);
3671	free_workqueue_attrs(new_attrs);
3672	apply_wqattrs_cleanup(ctx);
3673	return NULL;
3674	}
3675
3676	/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3677	static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3678	{
3679	int node;
3680
3681	/* all pwqs have been created successfully, let's install'em */
3682	mutex_lock(&ctx->wq->mutex);
3683
3684	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3685
3686	/* save the previous pwq and install the new one */
3687	for_each_node(node)
3688	ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3689	ctx->pwq_tbl[node]);
3690
3691	/* @dfl_pwq might not have been used, ensure it's linked */
3692	link_pwq(ctx->dfl_pwq);
3693	swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3694
3695	mutex_unlock(&ctx->wq->mutex);
3696	}
3697
3698	static void apply_wqattrs_lock(void)
3699	{
3700	/* CPUs should stay stable across pwq creations and installations */
3701	get_online_cpus();
3702	mutex_lock(&wq_pool_mutex);
3703	}
3704
3705	static void apply_wqattrs_unlock(void)
3706	{
3707	mutex_unlock(&wq_pool_mutex);
3708	put_online_cpus();
3709	}
3710
3711	static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3712	const struct workqueue_attrs *attrs)
3713	{
3714	struct apply_wqattrs_ctx *ctx;
3715
3716	/* only unbound workqueues can change attributes */
3717	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3718	return -EINVAL;
3719
3720	/* creating multiple pwqs breaks ordering guarantee */
3721	if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3722	return -EINVAL;
3723
3724	ctx = apply_wqattrs_prepare(wq, attrs);
3725	if (!ctx)
3726	return -ENOMEM;
3727
3728	/* the ctx has been prepared successfully, let's commit it */
3729	apply_wqattrs_commit(ctx);
3730	apply_wqattrs_cleanup(ctx);
3731
3732	return 0;
3733	}
3734
3735	/**
3736	* apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3737	* @wq: the target workqueue
3738	* @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3739	*
3740	* Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3741	* machines, this function maps a separate pwq to each NUMA node with
3742	* possibles CPUs in @attrs->cpumask so that work items are affine to the
3743	* NUMA node it was issued on. Older pwqs are released as in-flight work
3744	* items finish. Note that a work item which repeatedly requeues itself
3745	* back-to-back will stay on its current pwq.
3746	*
3747	* Performs GFP_KERNEL allocations.
3748	*
3749	* Return: 0 on success and -errno on failure.
3750	*/
3751	int apply_workqueue_attrs(struct workqueue_struct *wq,
3752	const struct workqueue_attrs *attrs)
3753	{
3754	int ret;
3755
3756	apply_wqattrs_lock();
3757	ret = apply_workqueue_attrs_locked(wq, attrs);
3758	apply_wqattrs_unlock();
3759
3760	return ret;
3761	}
3762
3763	/**
3764	* wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3765	* @wq: the target workqueue
3766	* @cpu: the CPU coming up or going down
3767	* @online: whether @cpu is coming up or going down
3768	*
3769	* This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3770	* %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3771	* @wq accordingly.
3772	*
3773	* If NUMA affinity can't be adjusted due to memory allocation failure, it
3774	* falls back to @wq->dfl_pwq which may not be optimal but is always
3775	* correct.
3776	*
3777	* Note that when the last allowed CPU of a NUMA node goes offline for a
3778	* workqueue with a cpumask spanning multiple nodes, the workers which were
3779	* already executing the work items for the workqueue will lose their CPU
3780	* affinity and may execute on any CPU. This is similar to how per-cpu
3781	* workqueues behave on CPU_DOWN. If a workqueue user wants strict
3782	* affinity, it's the user's responsibility to flush the work item from
3783	* CPU_DOWN_PREPARE.
3784	*/
3785	static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3786	bool online)
3787	{
3788	int node = cpu_to_node(cpu);
3789	int cpu_off = online ? -1 : cpu;
3790	struct pool_workqueue old_pwq = NULL, pwq;
3791	struct workqueue_attrs *target_attrs;
3792	cpumask_t *cpumask;
3793
3794	lockdep_assert_held(&wq_pool_mutex);
3795
3796	if (!wq_numa_enabled \|\| !(wq->flags & WQ_UNBOUND) \|\|
3797	wq->unbound_attrs->no_numa)
3798	return;
3799
3800	/*
3801	* We don't wanna alloc/free wq_attrs for each wq for each CPU.
3802	* Let's use a preallocated one. The following buf is protected by
3803	* CPU hotplug exclusion.
3804	*/
3805	target_attrs = wq_update_unbound_numa_attrs_buf;
3806	cpumask = target_attrs->cpumask;
3807
3808	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3809	pwq = unbound_pwq_by_node(wq, node);
3810
3811	/*
3812	* Let's determine what needs to be done. If the target cpumask is
3813	* different from the default pwq's, we need to compare it to @pwq's
3814	* and create a new one if they don't match. If the target cpumask
3815	* equals the default pwq's, the default pwq should be used.
3816	*/
3817	if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3818	if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3819	return;
3820	} else {
3821	goto use_dfl_pwq;
3822	}
3823
3824	/* create a new pwq */
3825	pwq = alloc_unbound_pwq(wq, target_attrs);
3826	if (!pwq) {
3827	pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3828	wq->name);
3829	goto use_dfl_pwq;
3830	}
3831
3832	/* Install the new pwq. */
3833	mutex_lock(&wq->mutex);
3834	old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3835	goto out_unlock;
3836
3837	use_dfl_pwq:
3838	mutex_lock(&wq->mutex);
3839	spin_lock_irq(&wq->dfl_pwq->pool->lock);
3840	get_pwq(wq->dfl_pwq);
3841	spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3842	old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3843	out_unlock:
3844	mutex_unlock(&wq->mutex);
3845	put_pwq_unlocked(old_pwq);
3846	}
3847
3848	static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3849	{
3850	bool highpri = wq->flags & WQ_HIGHPRI;
3851	int cpu, ret;
3852
3853	if (!(wq->flags & WQ_UNBOUND)) {
3854	wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3855	if (!wq->cpu_pwqs)
3856	return -ENOMEM;
3857
3858	for_each_possible_cpu(cpu) {
3859	struct pool_workqueue *pwq =
3860	per_cpu_ptr(wq->cpu_pwqs, cpu);
3861	struct worker_pool *cpu_pools =
3862	per_cpu(cpu_worker_pools, cpu);
3863
3864	init_pwq(pwq, wq, &cpu_pools[highpri]);
3865
3866	mutex_lock(&wq->mutex);
3867	link_pwq(pwq);
3868	mutex_unlock(&wq->mutex);
3869	}
3870	return 0;
3871	} else if (wq->flags & __WQ_ORDERED) {
3872	ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3873	/* there should only be single pwq for ordering guarantee */
3874	WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node \|\|
3875	wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3876	"ordering guarantee broken for workqueue %s\n", wq->name);
3877	return ret;
3878	} else {
3879	return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3880	}
3881	}
3882
3883	static int wq_clamp_max_active(int max_active, unsigned int flags,
3884	const char *name)
3885	{
3886	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3887
3888	if (max_active < 1 \|\| max_active > lim)
3889	pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3890	max_active, name, 1, lim);
3891
3892	return clamp_val(max_active, 1, lim);
3893	}
3894
3895	struct workqueue_struct __alloc_workqueue_key(const char fmt,
3896	unsigned int flags,
3897	int max_active,
3898	struct lock_class_key *key,
3899	const char *lock_name, ...)
3900	{
3901	size_t tbl_size = 0;
3902	va_list args;
3903	struct workqueue_struct *wq;
3904	struct pool_workqueue *pwq;
3905
3906	/* see the comment above the definition of WQ_POWER_EFFICIENT */
3907	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3908	flags \|= WQ_UNBOUND;
3909
3910	/* allocate wq and format name */
3911	if (flags & WQ_UNBOUND)
3912	tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3913
3914	wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3915	if (!wq)
3916	return NULL;
3917
3918	if (flags & WQ_UNBOUND) {
3919	wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3920	if (!wq->unbound_attrs)
3921	goto err_free_wq;
3922	}
3923
3924	va_start(args, lock_name);
3925	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3926	va_end(args);
3927
3928	max_active = max_active ?: WQ_DFL_ACTIVE;
3929	max_active = wq_clamp_max_active(max_active, flags, wq->name);
3930
3931	/* init wq */
3932	wq->flags = flags;
3933	wq->saved_max_active = max_active;
3934	mutex_init(&wq->mutex);
3935	atomic_set(&wq->nr_pwqs_to_flush, 0);
3936	INIT_LIST_HEAD(&wq->pwqs);
3937	INIT_LIST_HEAD(&wq->flusher_queue);
3938	INIT_LIST_HEAD(&wq->flusher_overflow);
3939	INIT_LIST_HEAD(&wq->maydays);
3940
3941	lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3942	INIT_LIST_HEAD(&wq->list);
3943
3944	if (alloc_and_link_pwqs(wq) < 0)
3945	goto err_free_wq;
3946
3947	/*
3948	* Workqueues which may be used during memory reclaim should
3949	* have a rescuer to guarantee forward progress.
3950	*/
3951	if (flags & WQ_MEM_RECLAIM) {
3952	struct worker *rescuer;
3953
3954	rescuer = alloc_worker(NUMA_NO_NODE);
3955	if (!rescuer)
3956	goto err_destroy;
3957
3958	rescuer->rescue_wq = wq;
3959	rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3960	wq->name);
3961	if (IS_ERR(rescuer->task)) {
3962	kfree(rescuer);
3963	goto err_destroy;
3964	}
3965
3966	wq->rescuer = rescuer;
3967	kthread_bind_mask(rescuer->task, cpu_possible_mask);
3968	wake_up_process(rescuer->task);
3969	}
3970
3971	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3972	goto err_destroy;
3973
3974	/*
3975	* wq_pool_mutex protects global freeze state and workqueues list.
3976	* Grab it, adjust max_active and add the new @wq to workqueues
3977	* list.
3978	*/
3979	mutex_lock(&wq_pool_mutex);
3980
3981	mutex_lock(&wq->mutex);
3982	for_each_pwq(pwq, wq)
3983	pwq_adjust_max_active(pwq);
3984	mutex_unlock(&wq->mutex);
3985
3986	list_add_tail_rcu(&wq->list, &workqueues);
3987
3988	mutex_unlock(&wq_pool_mutex);
3989
3990	return wq;
3991
3992	err_free_wq:
3993	free_workqueue_attrs(wq->unbound_attrs);
3994	kfree(wq);
3995	return NULL;
3996	err_destroy:
3997	destroy_workqueue(wq);
3998	return NULL;
3999	}
4000	EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4001
4002	/**
4003	* destroy_workqueue - safely terminate a workqueue
4004	* @wq: target workqueue
4005	*
4006	* Safely destroy a workqueue. All work currently pending will be done first.
4007	*/
4008	void destroy_workqueue(struct workqueue_struct *wq)
4009	{
4010	struct pool_workqueue *pwq;
4011	int node;
4012
4013	/* drain it before proceeding with destruction */
4014	drain_workqueue(wq);
4015
4016	/* sanity checks */
4017	mutex_lock(&wq->mutex);
4018	for_each_pwq(pwq, wq) {
4019	int i;
4020
4021	for (i = 0; i < WORK_NR_COLORS; i++) {
4022	if (WARN_ON(pwq->nr_in_flight[i])) {
4023	mutex_unlock(&wq->mutex);
4024	return;
4025	}
4026	}
4027
4028	if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) \|\|
4029	WARN_ON(pwq->nr_active) \|\|
4030	WARN_ON(!list_empty(&pwq->delayed_works))) {
4031	mutex_unlock(&wq->mutex);
4032	return;
4033	}
4034	}
4035	mutex_unlock(&wq->mutex);
4036
4037	/*
4038	* wq list is used to freeze wq, remove from list after
4039	* flushing is complete in case freeze races us.
4040	*/
4041	mutex_lock(&wq_pool_mutex);
4042	list_del_rcu(&wq->list);
4043	mutex_unlock(&wq_pool_mutex);
4044
4045	workqueue_sysfs_unregister(wq);
4046
4047	if (wq->rescuer)
4048	kthread_stop(wq->rescuer->task);
4049
4050	if (!(wq->flags & WQ_UNBOUND)) {
4051	/*
4052	* The base ref is never dropped on per-cpu pwqs. Directly
4053	* schedule RCU free.
4054	*/
4055	call_rcu_sched(&wq->rcu, rcu_free_wq);
4056	} else {
4057	/*
4058	* We're the sole accessor of @wq at this point. Directly
4059	* access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4060	* @wq will be freed when the last pwq is released.
4061	*/
4062	for_each_node(node) {
4063	pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4064	RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4065	put_pwq_unlocked(pwq);
4066	}
4067
4068	/*
4069	* Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4070	* put. Don't access it afterwards.
4071	*/
4072	pwq = wq->dfl_pwq;
4073	wq->dfl_pwq = NULL;
4074	put_pwq_unlocked(pwq);
4075	}
4076	}
4077	EXPORT_SYMBOL_GPL(destroy_workqueue);
4078
4079	/**
4080	* workqueue_set_max_active - adjust max_active of a workqueue
4081	* @wq: target workqueue
4082	* @max_active: new max_active value.
4083	*
4084	* Set max_active of @wq to @max_active.
4085	*
4086	* CONTEXT:
4087	* Don't call from IRQ context.
4088	*/
4089	void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4090	{
4091	struct pool_workqueue *pwq;
4092
4093	/* disallow meddling with max_active for ordered workqueues */
4094	if (WARN_ON(wq->flags & __WQ_ORDERED))
4095	return;
4096
4097	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4098
4099	mutex_lock(&wq->mutex);
4100
4101	wq->saved_max_active = max_active;
4102
4103	for_each_pwq(pwq, wq)
4104	pwq_adjust_max_active(pwq);
4105
4106	mutex_unlock(&wq->mutex);
4107	}
4108	EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4109
4110	/**
4111	* current_is_workqueue_rescuer - is %current workqueue rescuer?
4112	*
4113	* Determine whether %current is a workqueue rescuer. Can be used from
4114	* work functions to determine whether it's being run off the rescuer task.
4115	*
4116	* Return: %true if %current is a workqueue rescuer. %false otherwise.
4117	*/
4118	bool current_is_workqueue_rescuer(void)
4119	{
4120	struct worker *worker = current_wq_worker();
4121
4122	return worker && worker->rescue_wq;
4123	}
4124
4125	/**
4126	* workqueue_congested - test whether a workqueue is congested
4127	* @cpu: CPU in question
4128	* @wq: target workqueue
4129	*
4130	* Test whether @wq's cpu workqueue for @cpu is congested. There is
4131	* no synchronization around this function and the test result is
4132	* unreliable and only useful as advisory hints or for debugging.
4133	*
4134	* If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4135	* Note that both per-cpu and unbound workqueues may be associated with
4136	* multiple pool_workqueues which have separate congested states. A
4137	* workqueue being congested on one CPU doesn't mean the workqueue is also
4138	* contested on other CPUs / NUMA nodes.
4139	*
4140	* Return:
4141	* %true if congested, %false otherwise.
4142	*/
4143	bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4144	{
4145	struct pool_workqueue *pwq;
4146	bool ret;
4147
4148	rcu_read_lock_sched();
4149
4150	if (cpu == WORK_CPU_UNBOUND)
4151	cpu = smp_processor_id();
4152
4153	if (!(wq->flags & WQ_UNBOUND))
4154	pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4155	else
4156	pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4157
4158	ret = !list_empty(&pwq->delayed_works);
4159	rcu_read_unlock_sched();
4160
4161	return ret;
4162	}
4163	EXPORT_SYMBOL_GPL(workqueue_congested);
4164
4165	/**
4166	* work_busy - test whether a work is currently pending or running
4167	* @work: the work to be tested
4168	*
4169	* Test whether @work is currently pending or running. There is no
4170	* synchronization around this function and the test result is
4171	* unreliable and only useful as advisory hints or for debugging.
4172	*
4173	* Return:
4174	* OR'd bitmask of WORK_BUSY_* bits.
4175	*/
4176	unsigned int work_busy(struct work_struct *work)
4177	{
4178	struct worker_pool *pool;
4179	unsigned long flags;
4180	unsigned int ret = 0;
4181
4182	if (work_pending(work))
4183	ret \|= WORK_BUSY_PENDING;
4184
4185	local_irq_save(flags);
4186	pool = get_work_pool(work);
4187	if (pool) {
4188	spin_lock(&pool->lock);
4189	if (find_worker_executing_work(pool, work))
4190	ret \|= WORK_BUSY_RUNNING;
4191	spin_unlock(&pool->lock);
4192	}
4193	local_irq_restore(flags);
4194
4195	return ret;
4196	}
4197	EXPORT_SYMBOL_GPL(work_busy);
4198
4199	/**
4200	* set_worker_desc - set description for the current work item
4201	* @fmt: printf-style format string
4202	* @...: arguments for the format string
4203	*
4204	* This function can be called by a running work function to describe what
4205	* the work item is about. If the worker task gets dumped, this
4206	* information will be printed out together to help debugging. The
4207	* description can be at most WORKER_DESC_LEN including the trailing '\0'.
4208	*/
4209	void set_worker_desc(const char *fmt, ...)
4210	{
4211	struct worker *worker = current_wq_worker();
4212	va_list args;
4213
4214	if (worker) {
4215	va_start(args, fmt);
4216	vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4217	va_end(args);
4218	worker->desc_valid = true;
4219	}
4220	}
4221
4222	/**
4223	* print_worker_info - print out worker information and description
4224	* @log_lvl: the log level to use when printing
4225	* @task: target task
4226	*
4227	* If @task is a worker and currently executing a work item, print out the
4228	* name of the workqueue being serviced and worker description set with
4229	* set_worker_desc() by the currently executing work item.
4230	*
4231	* This function can be safely called on any task as long as the
4232	* task_struct itself is accessible. While safe, this function isn't
4233	* synchronized and may print out mixups or garbages of limited length.
4234	*/
4235	void print_worker_info(const char log_lvl, struct task_struct task)
4236	{
4237	work_func_t *fn = NULL;
4238	char name[WQ_NAME_LEN] = { };
4239	char desc[WORKER_DESC_LEN] = { };
4240	struct pool_workqueue *pwq = NULL;
4241	struct workqueue_struct *wq = NULL;
4242	bool desc_valid = false;
4243	struct worker *worker;
4244
4245	if (!(task->flags & PF_WQ_WORKER))
4246	return;
4247
4248	/*
4249	* This function is called without any synchronization and @task
4250	* could be in any state. Be careful with dereferences.
4251	*/
4252	worker = probe_kthread_data(task);
4253
4254	/*
4255	* Carefully copy the associated workqueue's workfn and name. Keep
4256	* the original last '\0' in case the original contains garbage.
4257	*/
4258	probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4259	probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4260	probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4261	probe_kernel_read(name, wq->name, sizeof(name) - 1);
4262
4263	/* copy worker description */
4264	probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4265	if (desc_valid)
4266	probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4267
4268	if (fn \|\| name[0] \|\| desc[0]) {
4269	printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4270	if (desc[0])
4271	pr_cont(" (%s)", desc);
4272	pr_cont("\n");
4273	}
4274	}
4275
4276	static void pr_cont_pool_info(struct worker_pool *pool)
4277	{
4278	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4279	if (pool->node != NUMA_NO_NODE)
4280	pr_cont(" node=%d", pool->node);
4281	pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4282	}
4283
4284	static void pr_cont_work(bool comma, struct work_struct *work)
4285	{
4286	if (work->func == wq_barrier_func) {
4287	struct wq_barrier *barr;
4288
4289	barr = container_of(work, struct wq_barrier, work);
4290
4291	pr_cont("%s BAR(%d)", comma ? "," : "",
4292	task_pid_nr(barr->task));
4293	} else {
4294	pr_cont("%s %pf", comma ? "," : "", work->func);
4295	}
4296	}
4297
4298	static void show_pwq(struct pool_workqueue *pwq)
4299	{
4300	struct worker_pool *pool = pwq->pool;
4301	struct work_struct *work;
4302	struct worker *worker;
4303	bool has_in_flight = false, has_pending = false;
4304	int bkt;
4305
4306	pr_info(" pwq %d:", pool->id);
4307	pr_cont_pool_info(pool);
4308
4309	pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4310	!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4311
4312	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4313	if (worker->current_pwq == pwq) {
4314	has_in_flight = true;
4315	break;
4316	}
4317	}
4318	if (has_in_flight) {
4319	bool comma = false;
4320
4321	pr_info(" in-flight:");
4322	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4323	if (worker->current_pwq != pwq)
4324	continue;
4325
4326	pr_cont("%s %d%s:%pf", comma ? "," : "",
4327	task_pid_nr(worker->task),
4328	worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4329	worker->current_func);
4330	list_for_each_entry(work, &worker->scheduled, entry)
4331	pr_cont_work(false, work);
4332	comma = true;
4333	}
4334	pr_cont("\n");
4335	}
4336
4337	list_for_each_entry(work, &pool->worklist, entry) {
4338	if (get_work_pwq(work) == pwq) {
4339	has_pending = true;
4340	break;
4341	}
4342	}
4343	if (has_pending) {
4344	bool comma = false;
4345
4346	pr_info(" pending:");
4347	list_for_each_entry(work, &pool->worklist, entry) {
4348	if (get_work_pwq(work) != pwq)
4349	continue;
4350
4351	pr_cont_work(comma, work);
4352	comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4353	}
4354	pr_cont("\n");
4355	}
4356
4357	if (!list_empty(&pwq->delayed_works)) {
4358	bool comma = false;
4359
4360	pr_info(" delayed:");
4361	list_for_each_entry(work, &pwq->delayed_works, entry) {
4362	pr_cont_work(comma, work);
4363	comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4364	}
4365	pr_cont("\n");
4366	}
4367	}
4368
4369	/**
4370	* show_workqueue_state - dump workqueue state
4371	*
4372	* Called from a sysrq handler or try_to_freeze_tasks() and prints out
4373	* all busy workqueues and pools.
4374	*/
4375	void show_workqueue_state(void)
4376	{
4377	struct workqueue_struct *wq;
4378	struct worker_pool *pool;
4379	unsigned long flags;
4380	int pi;
4381
4382	rcu_read_lock_sched();
4383
4384	pr_info("Showing busy workqueues and worker pools:\n");
4385
4386	list_for_each_entry_rcu(wq, &workqueues, list) {
4387	struct pool_workqueue *pwq;
4388	bool idle = true;
4389
4390	for_each_pwq(pwq, wq) {
4391	if (pwq->nr_active \|\| !list_empty(&pwq->delayed_works)) {
4392	idle = false;
4393	break;
4394	}
4395	}
4396	if (idle)
4397	continue;
4398
4399	pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4400
4401	for_each_pwq(pwq, wq) {
4402	spin_lock_irqsave(&pwq->pool->lock, flags);
4403	if (pwq->nr_active \|\| !list_empty(&pwq->delayed_works))
4404	show_pwq(pwq);
4405	spin_unlock_irqrestore(&pwq->pool->lock, flags);
4406	}
4407	}
4408
4409	for_each_pool(pool, pi) {
4410	struct worker *worker;
4411	bool first = true;
4412
4413	spin_lock_irqsave(&pool->lock, flags);
4414	if (pool->nr_workers == pool->nr_idle)
4415	goto next_pool;
4416
4417	pr_info("pool %d:", pool->id);
4418	pr_cont_pool_info(pool);
4419	pr_cont(" hung=%us workers=%d",
4420	jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4421	pool->nr_workers);
4422	if (pool->manager)
4423	pr_cont(" manager: %d",
4424	task_pid_nr(pool->manager->task));
4425	list_for_each_entry(worker, &pool->idle_list, entry) {
4426	pr_cont(" %s%d", first ? "idle: " : "",
4427	task_pid_nr(worker->task));
4428	first = false;
4429	}
4430	pr_cont("\n");
4431	next_pool:
4432	spin_unlock_irqrestore(&pool->lock, flags);
4433	}
4434
4435	rcu_read_unlock_sched();
4436	}
4437
4438	/*
4439	* CPU hotplug.
4440	*
4441	* There are two challenges in supporting CPU hotplug. Firstly, there
4442	* are a lot of assumptions on strong associations among work, pwq and
4443	* pool which make migrating pending and scheduled works very
4444	* difficult to implement without impacting hot paths. Secondly,
4445	* worker pools serve mix of short, long and very long running works making
4446	* blocked draining impractical.
4447	*
4448	* This is solved by allowing the pools to be disassociated from the CPU
4449	* running as an unbound one and allowing it to be reattached later if the
4450	* cpu comes back online.
4451	*/
4452
4453	static void wq_unbind_fn(struct work_struct *work)
4454	{
4455	int cpu = smp_processor_id();
4456	struct worker_pool *pool;
4457	struct worker *worker;
4458
4459	for_each_cpu_worker_pool(pool, cpu) {
4460	mutex_lock(&pool->attach_mutex);
4461	spin_lock_irq(&pool->lock);
4462
4463	/*
4464	* We've blocked all attach/detach operations. Make all workers
4465	* unbound and set DISASSOCIATED. Before this, all workers
4466	* except for the ones which are still executing works from
4467	* before the last CPU down must be on the cpu. After
4468	* this, they may become diasporas.
4469	*/
4470	for_each_pool_worker(worker, pool)
4471	worker->flags \|= WORKER_UNBOUND;
4472
4473	pool->flags \|= POOL_DISASSOCIATED;
4474
4475	spin_unlock_irq(&pool->lock);
4476	mutex_unlock(&pool->attach_mutex);
4477
4478	/*
4479	* Call schedule() so that we cross rq->lock and thus can
4480	* guarantee sched callbacks see the %WORKER_UNBOUND flag.
4481	* This is necessary as scheduler callbacks may be invoked
4482	* from other cpus.
4483	*/
4484	schedule();
4485
4486	/*
4487	* Sched callbacks are disabled now. Zap nr_running.
4488	* After this, nr_running stays zero and need_more_worker()
4489	* and keep_working() are always true as long as the
4490	* worklist is not empty. This pool now behaves as an
4491	* unbound (in terms of concurrency management) pool which
4492	* are served by workers tied to the pool.
4493	*/
4494	atomic_set(&pool->nr_running, 0);
4495
4496	/*
4497	* With concurrency management just turned off, a busy
4498	* worker blocking could lead to lengthy stalls. Kick off
4499	* unbound chain execution of currently pending work items.
4500	*/
4501	spin_lock_irq(&pool->lock);
4502	wake_up_worker(pool);
4503	spin_unlock_irq(&pool->lock);
4504	}
4505	}
4506
4507	/**
4508	* rebind_workers - rebind all workers of a pool to the associated CPU
4509	* @pool: pool of interest
4510	*
4511	* @pool->cpu is coming online. Rebind all workers to the CPU.
4512	*/
4513	static void rebind_workers(struct worker_pool *pool)
4514	{
4515	struct worker *worker;
4516
4517	lockdep_assert_held(&pool->attach_mutex);
4518
4519	/*
4520	* Restore CPU affinity of all workers. As all idle workers should
4521	* be on the run-queue of the associated CPU before any local
4522	* wake-ups for concurrency management happen, restore CPU affinity
4523	* of all workers first and then clear UNBOUND. As we're called
4524	* from CPU_ONLINE, the following shouldn't fail.
4525	*/
4526	for_each_pool_worker(worker, pool)
4527	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4528	pool->attrs->cpumask) < 0);
4529
4530	spin_lock_irq(&pool->lock);
4531
4532	/*
4533	* XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4534	* w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4535	* being reworked and this can go away in time.
4536	*/
4537	if (!(pool->flags & POOL_DISASSOCIATED)) {
4538	spin_unlock_irq(&pool->lock);
4539	return;
4540	}
4541
4542	pool->flags &= ~POOL_DISASSOCIATED;
4543
4544	for_each_pool_worker(worker, pool) {
4545	unsigned int worker_flags = worker->flags;
4546
4547	/*
4548	* A bound idle worker should actually be on the runqueue
4549	* of the associated CPU for local wake-ups targeting it to
4550	* work. Kick all idle workers so that they migrate to the
4551	* associated CPU. Doing this in the same loop as
4552	* replacing UNBOUND with REBOUND is safe as no worker will
4553	* be bound before @pool->lock is released.
4554	*/
4555	if (worker_flags & WORKER_IDLE)
4556	wake_up_process(worker->task);
4557
4558	/*
4559	* We want to clear UNBOUND but can't directly call
4560	* worker_clr_flags() or adjust nr_running. Atomically
4561	* replace UNBOUND with another NOT_RUNNING flag REBOUND.
4562	* @worker will clear REBOUND using worker_clr_flags() when
4563	* it initiates the next execution cycle thus restoring
4564	* concurrency management. Note that when or whether
4565	* @worker clears REBOUND doesn't affect correctness.
4566	*
4567	* ACCESS_ONCE() is necessary because @worker->flags may be
4568	* tested without holding any lock in
4569	* wq_worker_waking_up(). Without it, NOT_RUNNING test may
4570	* fail incorrectly leading to premature concurrency
4571	* management operations.
4572	*/
4573	WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4574	worker_flags \|= WORKER_REBOUND;
4575	worker_flags &= ~WORKER_UNBOUND;
4576	ACCESS_ONCE(worker->flags) = worker_flags;
4577	}
4578
4579	spin_unlock_irq(&pool->lock);
4580	}
4581
4582	/**
4583	* restore_unbound_workers_cpumask - restore cpumask of unbound workers
4584	* @pool: unbound pool of interest
4585	* @cpu: the CPU which is coming up
4586	*
4587	* An unbound pool may end up with a cpumask which doesn't have any online
4588	* CPUs. When a worker of such pool get scheduled, the scheduler resets
4589	* its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4590	* online CPU before, cpus_allowed of all its workers should be restored.
4591	*/
4592	static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4593	{
4594	static cpumask_t cpumask;
4595	struct worker *worker;
4596
4597	lockdep_assert_held(&pool->attach_mutex);
4598
4599	/* is @cpu allowed for @pool? */
4600	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4601	return;
4602
4603	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4604
4605	/* as we're called from CPU_ONLINE, the following shouldn't fail */
4606	for_each_pool_worker(worker, pool)
4607	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4608	}
4609
4610	int workqueue_prepare_cpu(unsigned int cpu)
4611	{
4612	struct worker_pool *pool;
4613
4614	for_each_cpu_worker_pool(pool, cpu) {
4615	if (pool->nr_workers)
4616	continue;
4617	if (!create_worker(pool))
4618	return -ENOMEM;
4619	}
4620	return 0;
4621	}
4622
4623	int workqueue_online_cpu(unsigned int cpu)
4624	{
4625	struct worker_pool *pool;
4626	struct workqueue_struct *wq;
4627	int pi;
4628
4629	mutex_lock(&wq_pool_mutex);
4630
4631	for_each_pool(pool, pi) {
4632	mutex_lock(&pool->attach_mutex);
4633
4634	if (pool->cpu == cpu)
4635	rebind_workers(pool);
4636	else if (pool->cpu < 0)
4637	restore_unbound_workers_cpumask(pool, cpu);
4638
4639	mutex_unlock(&pool->attach_mutex);
4640	}
4641
4642	/* update NUMA affinity of unbound workqueues */
4643	list_for_each_entry(wq, &workqueues, list)
4644	wq_update_unbound_numa(wq, cpu, true);
4645
4646	mutex_unlock(&wq_pool_mutex);
4647	return 0;
4648	}
4649
4650	int workqueue_offline_cpu(unsigned int cpu)
4651	{
4652	struct work_struct unbind_work;
4653	struct workqueue_struct *wq;
4654
4655	/* unbinding per-cpu workers should happen on the local CPU */
4656	INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4657	queue_work_on(cpu, system_highpri_wq, &unbind_work);
4658
4659	/* update NUMA affinity of unbound workqueues */
4660	mutex_lock(&wq_pool_mutex);
4661	list_for_each_entry(wq, &workqueues, list)
4662	wq_update_unbound_numa(wq, cpu, false);
4663	mutex_unlock(&wq_pool_mutex);
4664
4665	/* wait for per-cpu unbinding to finish */
4666	flush_work(&unbind_work);
4667	destroy_work_on_stack(&unbind_work);
4668	return 0;
4669	}
4670
4671	#ifdef CONFIG_SMP
4672
4673	struct work_for_cpu {
4674	struct work_struct work;
4675	long (fn)(void );
4676	void *arg;
4677	long ret;
4678	};
4679
4680	static void work_for_cpu_fn(struct work_struct *work)
4681	{
4682	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4683
4684	wfc->ret = wfc->fn(wfc->arg);
4685	}
4686
4687	/**
4688	* work_on_cpu - run a function in thread context on a particular cpu
4689	* @cpu: the cpu to run on
4690	* @fn: the function to run
4691	* @arg: the function arg
4692	*
4693	* It is up to the caller to ensure that the cpu doesn't go offline.
4694	* The caller must not hold any locks which would prevent @fn from completing.
4695	*
4696	* Return: The value @fn returns.
4697	*/
4698	long work_on_cpu(int cpu, long (fn)(void ), void *arg)
4699	{
4700	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4701
4702	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4703	schedule_work_on(cpu, &wfc.work);
4704	flush_work(&wfc.work);
4705	destroy_work_on_stack(&wfc.work);
4706	return wfc.ret;
4707	}
4708	EXPORT_SYMBOL_GPL(work_on_cpu);
4709	#endif /* CONFIG_SMP */
4710
4711	#ifdef CONFIG_FREEZER
4712
4713	/**
4714	* freeze_workqueues_begin - begin freezing workqueues
4715	*
4716	* Start freezing workqueues. After this function returns, all freezable
4717	* workqueues will queue new works to their delayed_works list instead of
4718	* pool->worklist.
4719	*
4720	* CONTEXT:
4721	* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4722	*/
4723	void freeze_workqueues_begin(void)
4724	{
4725	struct workqueue_struct *wq;
4726	struct pool_workqueue *pwq;
4727
4728	mutex_lock(&wq_pool_mutex);
4729
4730	WARN_ON_ONCE(workqueue_freezing);
4731	workqueue_freezing = true;
4732
4733	list_for_each_entry(wq, &workqueues, list) {
4734	mutex_lock(&wq->mutex);
4735	for_each_pwq(pwq, wq)
4736	pwq_adjust_max_active(pwq);
4737	mutex_unlock(&wq->mutex);
4738	}
4739
4740	mutex_unlock(&wq_pool_mutex);
4741	}
4742
4743	/**
4744	* freeze_workqueues_busy - are freezable workqueues still busy?
4745	*
4746	* Check whether freezing is complete. This function must be called
4747	* between freeze_workqueues_begin() and thaw_workqueues().
4748	*
4749	* CONTEXT:
4750	* Grabs and releases wq_pool_mutex.
4751	*
4752	* Return:
4753	* %true if some freezable workqueues are still busy. %false if freezing
4754	* is complete.
4755	*/
4756	bool freeze_workqueues_busy(void)
4757	{
4758	bool busy = false;
4759	struct workqueue_struct *wq;
4760	struct pool_workqueue *pwq;
4761
4762	mutex_lock(&wq_pool_mutex);
4763
4764	WARN_ON_ONCE(!workqueue_freezing);
4765
4766	list_for_each_entry(wq, &workqueues, list) {
4767	if (!(wq->flags & WQ_FREEZABLE))
4768	continue;
4769	/*
4770	* nr_active is monotonically decreasing. It's safe
4771	* to peek without lock.
4772	*/
4773	rcu_read_lock_sched();
4774	for_each_pwq(pwq, wq) {
4775	WARN_ON_ONCE(pwq->nr_active < 0);
4776	if (pwq->nr_active) {
4777	busy = true;
4778	rcu_read_unlock_sched();
4779	goto out_unlock;
4780	}
4781	}
4782	rcu_read_unlock_sched();
4783	}
4784	out_unlock:
4785	mutex_unlock(&wq_pool_mutex);
4786	return busy;
4787	}
4788
4789	/**
4790	* thaw_workqueues - thaw workqueues
4791	*
4792	* Thaw workqueues. Normal queueing is restored and all collected
4793	* frozen works are transferred to their respective pool worklists.
4794	*
4795	* CONTEXT:
4796	* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4797	*/
4798	void thaw_workqueues(void)
4799	{
4800	struct workqueue_struct *wq;
4801	struct pool_workqueue *pwq;
4802
4803	mutex_lock(&wq_pool_mutex);
4804
4805	if (!workqueue_freezing)
4806	goto out_unlock;
4807
4808	workqueue_freezing = false;
4809
4810	/* restore max_active and repopulate worklist */
4811	list_for_each_entry(wq, &workqueues, list) {
4812	mutex_lock(&wq->mutex);
4813	for_each_pwq(pwq, wq)
4814	pwq_adjust_max_active(pwq);
4815	mutex_unlock(&wq->mutex);
4816	}
4817
4818	out_unlock:
4819	mutex_unlock(&wq_pool_mutex);
4820	}
4821	#endif /* CONFIG_FREEZER */
4822
4823	static int workqueue_apply_unbound_cpumask(void)
4824	{
4825	LIST_HEAD(ctxs);
4826	int ret = 0;
4827	struct workqueue_struct *wq;
4828	struct apply_wqattrs_ctx ctx, n;
4829
4830	lockdep_assert_held(&wq_pool_mutex);
4831
4832	list_for_each_entry(wq, &workqueues, list) {
4833	if (!(wq->flags & WQ_UNBOUND))
4834	continue;
4835	/* creating multiple pwqs breaks ordering guarantee */
4836	if (wq->flags & __WQ_ORDERED)
4837	continue;
4838
4839	ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4840	if (!ctx) {
4841	ret = -ENOMEM;
4842	break;
4843	}
4844
4845	list_add_tail(&ctx->list, &ctxs);
4846	}
4847
4848	list_for_each_entry_safe(ctx, n, &ctxs, list) {
4849	if (!ret)
4850	apply_wqattrs_commit(ctx);
4851	apply_wqattrs_cleanup(ctx);
4852	}
4853
4854	return ret;
4855	}
4856
4857	/**
4858	* workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4859	* @cpumask: the cpumask to set
4860	*
4861	* The low-level workqueues cpumask is a global cpumask that limits
4862	* the affinity of all unbound workqueues. This function check the @cpumask
4863	* and apply it to all unbound workqueues and updates all pwqs of them.
4864	*
4865	* Retun: 0 - Success
4866	* -EINVAL - Invalid @cpumask
4867	* -ENOMEM - Failed to allocate memory for attrs or pwqs.
4868	*/
4869	int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4870	{
4871	int ret = -EINVAL;
4872	cpumask_var_t saved_cpumask;
4873
4874	if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4875	return -ENOMEM;
4876
4877	cpumask_and(cpumask, cpumask, cpu_possible_mask);
4878	if (!cpumask_empty(cpumask)) {
4879	apply_wqattrs_lock();
4880
4881	/* save the old wq_unbound_cpumask. */
4882	cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4883
4884	/* update wq_unbound_cpumask at first and apply it to wqs. */
4885	cpumask_copy(wq_unbound_cpumask, cpumask);
4886	ret = workqueue_apply_unbound_cpumask();
4887
4888	/* restore the wq_unbound_cpumask when failed. */
4889	if (ret < 0)
4890	cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4891
4892	apply_wqattrs_unlock();
4893	}
4894
4895	free_cpumask_var(saved_cpumask);
4896	return ret;
4897	}
4898
4899	#ifdef CONFIG_SYSFS
4900	/*
4901	* Workqueues with WQ_SYSFS flag set is visible to userland via
4902	* /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4903	* following attributes.
4904	*
4905	* per_cpu RO bool : whether the workqueue is per-cpu or unbound
4906	* max_active RW int : maximum number of in-flight work items
4907	*
4908	* Unbound workqueues have the following extra attributes.
4909	*
4910	* id RO int : the associated pool ID
4911	* nice RW int : nice value of the workers
4912	* cpumask RW mask : bitmask of allowed CPUs for the workers
4913	*/
4914	struct wq_device {
4915	struct workqueue_struct *wq;
4916	struct device dev;
4917	};
4918
4919	static struct workqueue_struct dev_to_wq(struct device dev)
4920	{
4921	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4922
4923	return wq_dev->wq;
4924	}
4925
4926	static ssize_t per_cpu_show(struct device dev, struct device_attribute attr,
4927	char *buf)
4928	{
4929	struct workqueue_struct *wq = dev_to_wq(dev);
4930
4931	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4932	}
4933	static DEVICE_ATTR_RO(per_cpu);
4934
4935	static ssize_t max_active_show(struct device *dev,
4936	struct device_attribute attr, char buf)
4937	{
4938	struct workqueue_struct *wq = dev_to_wq(dev);
4939
4940	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4941	}
4942
4943	static ssize_t max_active_store(struct device *dev,
4944	struct device_attribute attr, const char buf,
4945	size_t count)
4946	{
4947	struct workqueue_struct *wq = dev_to_wq(dev);
4948	int val;
4949
4950	if (sscanf(buf, "%d", &val) != 1 \|\| val <= 0)
4951	return -EINVAL;
4952
4953	workqueue_set_max_active(wq, val);
4954	return count;
4955	}
4956	static DEVICE_ATTR_RW(max_active);
4957
4958	static struct attribute *wq_sysfs_attrs[] = {
4959	&dev_attr_per_cpu.attr,
4960	&dev_attr_max_active.attr,
4961	NULL,
4962	};
4963	ATTRIBUTE_GROUPS(wq_sysfs);
4964
4965	static ssize_t wq_pool_ids_show(struct device *dev,
4966	struct device_attribute attr, char buf)
4967	{
4968	struct workqueue_struct *wq = dev_to_wq(dev);
4969	const char *delim = "";
4970	int node, written = 0;
4971
4972	rcu_read_lock_sched();
4973	for_each_node(node) {
4974	written += scnprintf(buf + written, PAGE_SIZE - written,
4975	"%s%d:%d", delim, node,
4976	unbound_pwq_by_node(wq, node)->pool->id);
4977	delim = " ";
4978	}
4979	written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4980	rcu_read_unlock_sched();
4981
4982	return written;
4983	}
4984
4985	static ssize_t wq_nice_show(struct device dev, struct device_attribute attr,
4986	char *buf)
4987	{
4988	struct workqueue_struct *wq = dev_to_wq(dev);
4989	int written;
4990
4991	mutex_lock(&wq->mutex);
4992	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
4993	mutex_unlock(&wq->mutex);
4994
4995	return written;
4996	}
4997
4998	/* prepare workqueue_attrs for sysfs store operations */
4999	static struct workqueue_attrs wq_sysfs_prep_attrs(struct workqueue_struct wq)
5000	{
5001	struct workqueue_attrs *attrs;
5002
5003	lockdep_assert_held(&wq_pool_mutex);
5004
5005	attrs = alloc_workqueue_attrs(GFP_KERNEL);
5006	if (!attrs)
5007	return NULL;
5008
5009	copy_workqueue_attrs(attrs, wq->unbound_attrs);
5010	return attrs;
5011	}
5012
5013	static ssize_t wq_nice_store(struct device dev, struct device_attribute attr,
5014	const char *buf, size_t count)
5015	{
5016	struct workqueue_struct *wq = dev_to_wq(dev);
5017	struct workqueue_attrs *attrs;
5018	int ret = -ENOMEM;
5019
5020	apply_wqattrs_lock();
5021
5022	attrs = wq_sysfs_prep_attrs(wq);
5023	if (!attrs)
5024	goto out_unlock;
5025
5026	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5027	attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5028	ret = apply_workqueue_attrs_locked(wq, attrs);
5029	else
5030	ret = -EINVAL;
5031
5032	out_unlock:
5033	apply_wqattrs_unlock();
5034	free_workqueue_attrs(attrs);
5035	return ret ?: count;
5036	}
5037
5038	static ssize_t wq_cpumask_show(struct device *dev,
5039	struct device_attribute attr, char buf)
5040	{
5041	struct workqueue_struct *wq = dev_to_wq(dev);
5042	int written;
5043
5044	mutex_lock(&wq->mutex);
5045	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5046	cpumask_pr_args(wq->unbound_attrs->cpumask));
5047	mutex_unlock(&wq->mutex);
5048	return written;
5049	}
5050
5051	static ssize_t wq_cpumask_store(struct device *dev,
5052	struct device_attribute *attr,
5053	const char *buf, size_t count)
5054	{
5055	struct workqueue_struct *wq = dev_to_wq(dev);
5056	struct workqueue_attrs *attrs;
5057	int ret = -ENOMEM;
5058
5059	apply_wqattrs_lock();
5060
5061	attrs = wq_sysfs_prep_attrs(wq);
5062	if (!attrs)
5063	goto out_unlock;
5064
5065	ret = cpumask_parse(buf, attrs->cpumask);
5066	if (!ret)
5067	ret = apply_workqueue_attrs_locked(wq, attrs);
5068
5069	out_unlock:
5070	apply_wqattrs_unlock();
5071	free_workqueue_attrs(attrs);
5072	return ret ?: count;
5073	}
5074
5075	static ssize_t wq_numa_show(struct device dev, struct device_attribute attr,
5076	char *buf)
5077	{
5078	struct workqueue_struct *wq = dev_to_wq(dev);
5079	int written;
5080
5081	mutex_lock(&wq->mutex);
5082	written = scnprintf(buf, PAGE_SIZE, "%d\n",
5083	!wq->unbound_attrs->no_numa);
5084	mutex_unlock(&wq->mutex);
5085
5086	return written;
5087	}
5088
5089	static ssize_t wq_numa_store(struct device dev, struct device_attribute attr,
5090	const char *buf, size_t count)
5091	{
5092	struct workqueue_struct *wq = dev_to_wq(dev);
5093	struct workqueue_attrs *attrs;
5094	int v, ret = -ENOMEM;
5095
5096	apply_wqattrs_lock();
5097
5098	attrs = wq_sysfs_prep_attrs(wq);
5099	if (!attrs)
5100	goto out_unlock;
5101
5102	ret = -EINVAL;
5103	if (sscanf(buf, "%d", &v) == 1) {
5104	attrs->no_numa = !v;
5105	ret = apply_workqueue_attrs_locked(wq, attrs);
5106	}
5107
5108	out_unlock:
5109	apply_wqattrs_unlock();
5110	free_workqueue_attrs(attrs);
5111	return ret ?: count;
5112	}
5113
5114	static struct device_attribute wq_sysfs_unbound_attrs[] = {
5115	__ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5116	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5117	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5118	__ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5119	__ATTR_NULL,
5120	};
5121
5122	static struct bus_type wq_subsys = {
5123	.name = "workqueue",
5124	.dev_groups = wq_sysfs_groups,
5125	};
5126
5127	static ssize_t wq_unbound_cpumask_show(struct device *dev,
5128	struct device_attribute attr, char buf)
5129	{
5130	int written;
5131
5132	mutex_lock(&wq_pool_mutex);
5133	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5134	cpumask_pr_args(wq_unbound_cpumask));
5135	mutex_unlock(&wq_pool_mutex);
5136
5137	return written;
5138	}
5139
5140	static ssize_t wq_unbound_cpumask_store(struct device *dev,
5141	struct device_attribute attr, const char buf, size_t count)
5142	{
5143	cpumask_var_t cpumask;
5144	int ret;
5145
5146	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5147	return -ENOMEM;
5148
5149	ret = cpumask_parse(buf, cpumask);
5150	if (!ret)
5151	ret = workqueue_set_unbound_cpumask(cpumask);
5152
5153	free_cpumask_var(cpumask);
5154	return ret ? ret : count;
5155	}
5156
5157	static struct device_attribute wq_sysfs_cpumask_attr =
5158	__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5159	wq_unbound_cpumask_store);
5160
5161	static int __init wq_sysfs_init(void)
5162	{
5163	int err;
5164
5165	err = subsys_virtual_register(&wq_subsys, NULL);
5166	if (err)
5167	return err;
5168
5169	return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5170	}
5171	core_initcall(wq_sysfs_init);
5172
5173	static void wq_device_release(struct device *dev)
5174	{
5175	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5176
5177	kfree(wq_dev);
5178	}
5179
5180	/**
5181	* workqueue_sysfs_register - make a workqueue visible in sysfs
5182	* @wq: the workqueue to register
5183	*
5184	* Expose @wq in sysfs under /sys/bus/workqueue/devices.
5185	* alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5186	* which is the preferred method.
5187	*
5188	* Workqueue user should use this function directly iff it wants to apply
5189	* workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5190	* apply_workqueue_attrs() may race against userland updating the
5191	* attributes.
5192	*
5193	* Return: 0 on success, -errno on failure.
5194	*/
5195	int workqueue_sysfs_register(struct workqueue_struct *wq)
5196	{
5197	struct wq_device *wq_dev;
5198	int ret;
5199
5200	/*
5201	* Adjusting max_active or creating new pwqs by applying
5202	* attributes breaks ordering guarantee. Disallow exposing ordered
5203	* workqueues.
5204	*/
5205	if (WARN_ON(wq->flags & __WQ_ORDERED))
5206	return -EINVAL;
5207
5208	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5209	if (!wq_dev)
5210	return -ENOMEM;
5211
5212	wq_dev->wq = wq;
5213	wq_dev->dev.bus = &wq_subsys;
5214	wq_dev->dev.release = wq_device_release;
5215	dev_set_name(&wq_dev->dev, "%s", wq->name);
5216
5217	/*
5218	* unbound_attrs are created separately. Suppress uevent until
5219	* everything is ready.
5220	*/
5221	dev_set_uevent_suppress(&wq_dev->dev, true);
5222
5223	ret = device_register(&wq_dev->dev);
5224	if (ret) {
5225	kfree(wq_dev);
5226	wq->wq_dev = NULL;
5227	return ret;
5228	}
5229
5230	if (wq->flags & WQ_UNBOUND) {
5231	struct device_attribute *attr;
5232
5233	for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5234	ret = device_create_file(&wq_dev->dev, attr);
5235	if (ret) {
5236	device_unregister(&wq_dev->dev);
5237	wq->wq_dev = NULL;
5238	return ret;
5239	}
5240	}
5241	}
5242
5243	dev_set_uevent_suppress(&wq_dev->dev, false);
5244	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5245	return 0;
5246	}
5247
5248	/**
5249	* workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5250	* @wq: the workqueue to unregister
5251	*
5252	* If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5253	*/
5254	static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5255	{
5256	struct wq_device *wq_dev = wq->wq_dev;
5257
5258	if (!wq->wq_dev)
5259	return;
5260
5261	wq->wq_dev = NULL;
5262	device_unregister(&wq_dev->dev);
5263	}
5264	#else /* CONFIG_SYSFS */
5265	static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5266	#endif /* CONFIG_SYSFS */
5267
5268	/*
5269	* Workqueue watchdog.
5270	*
5271	* Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5272	* flush dependency, a concurrency managed work item which stays RUNNING
5273	* indefinitely. Workqueue stalls can be very difficult to debug as the
5274	* usual warning mechanisms don't trigger and internal workqueue state is
5275	* largely opaque.
5276	*
5277	* Workqueue watchdog monitors all worker pools periodically and dumps
5278	* state if some pools failed to make forward progress for a while where
5279	* forward progress is defined as the first item on ->worklist changing.
5280	*
5281	* This mechanism is controlled through the kernel parameter
5282	* "workqueue.watchdog_thresh" which can be updated at runtime through the
5283	* corresponding sysfs parameter file.
5284	*/
5285	#ifdef CONFIG_WQ_WATCHDOG
5286
5287	static void wq_watchdog_timer_fn(unsigned long data);
5288
5289	static unsigned long wq_watchdog_thresh = 30;
5290	static struct timer_list wq_watchdog_timer =
5291	TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5292
5293	static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5294	static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5295
5296	static void wq_watchdog_reset_touched(void)
5297	{
5298	int cpu;
5299
5300	wq_watchdog_touched = jiffies;
5301	for_each_possible_cpu(cpu)
5302	per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5303	}
5304
5305	static void wq_watchdog_timer_fn(unsigned long data)
5306	{
5307	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5308	bool lockup_detected = false;
5309	struct worker_pool *pool;
5310	int pi;
5311
5312	if (!thresh)
5313	return;
5314
5315	rcu_read_lock();
5316
5317	for_each_pool(pool, pi) {
5318	unsigned long pool_ts, touched, ts;
5319
5320	if (list_empty(&pool->worklist))
5321	continue;
5322
5323	/* get the latest of pool and touched timestamps */
5324	pool_ts = READ_ONCE(pool->watchdog_ts);
5325	touched = READ_ONCE(wq_watchdog_touched);
5326
5327	if (time_after(pool_ts, touched))
5328	ts = pool_ts;
5329	else
5330	ts = touched;
5331
5332	if (pool->cpu >= 0) {
5333	unsigned long cpu_touched =
5334	READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5335	pool->cpu));
5336	if (time_after(cpu_touched, ts))
5337	ts = cpu_touched;
5338	}
5339
5340	/* did we stall? */
5341	if (time_after(jiffies, ts + thresh)) {
5342	lockup_detected = true;
5343	pr_emerg("BUG: workqueue lockup - pool");
5344	pr_cont_pool_info(pool);
5345	pr_cont(" stuck for %us!\n",
5346	jiffies_to_msecs(jiffies - pool_ts) / 1000);
5347	}
5348	}
5349
5350	rcu_read_unlock();
5351
5352	if (lockup_detected)
5353	show_workqueue_state();
5354
5355	wq_watchdog_reset_touched();
5356	mod_timer(&wq_watchdog_timer, jiffies + thresh);
5357	}
5358
5359	void wq_watchdog_touch(int cpu)
5360	{
5361	if (cpu >= 0)
5362	per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5363	else
5364	wq_watchdog_touched = jiffies;
5365	}
5366
5367	static void wq_watchdog_set_thresh(unsigned long thresh)
5368	{
5369	wq_watchdog_thresh = 0;
5370	del_timer_sync(&wq_watchdog_timer);
5371
5372	if (thresh) {
5373	wq_watchdog_thresh = thresh;
5374	wq_watchdog_reset_touched();
5375	mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5376	}
5377	}
5378
5379	static int wq_watchdog_param_set_thresh(const char *val,
5380	const struct kernel_param *kp)
5381	{
5382	unsigned long thresh;
5383	int ret;
5384
5385	ret = kstrtoul(val, 0, &thresh);
5386	if (ret)
5387	return ret;
5388
5389	if (system_wq)
5390	wq_watchdog_set_thresh(thresh);
5391	else
5392	wq_watchdog_thresh = thresh;
5393
5394	return 0;
5395	}
5396
5397	static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5398	.set = wq_watchdog_param_set_thresh,
5399	.get = param_get_ulong,
5400	};
5401
5402	module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5403	0644);
5404
5405	static void wq_watchdog_init(void)
5406	{
5407	wq_watchdog_set_thresh(wq_watchdog_thresh);
5408	}
5409
5410	#else /* CONFIG_WQ_WATCHDOG */
5411
5412	static inline void wq_watchdog_init(void) { }
5413
5414	#endif /* CONFIG_WQ_WATCHDOG */
5415
5416	static void __init wq_numa_init(void)
5417	{
5418	cpumask_var_t *tbl;
5419	int node, cpu;
5420
5421	if (num_possible_nodes() <= 1)
5422	return;
5423
5424	if (wq_disable_numa) {
5425	pr_info("workqueue: NUMA affinity support disabled\n");
5426	return;
5427	}
5428
5429	wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5430	BUG_ON(!wq_update_unbound_numa_attrs_buf);
5431
5432	/*
5433	* We want masks of possible CPUs of each node which isn't readily
5434	* available. Build one from cpu_to_node() which should have been
5435	* fully initialized by now.
5436	*/
5437	tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5438	BUG_ON(!tbl);
5439
5440	for_each_node(node)
5441	BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5442	node_online(node) ? node : NUMA_NO_NODE));
5443
5444	for_each_possible_cpu(cpu) {
5445	node = cpu_to_node(cpu);
5446	if (WARN_ON(node == NUMA_NO_NODE)) {
5447	pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5448	/* happens iff arch is bonkers, let's just proceed */
5449	return;
5450	}
5451	cpumask_set_cpu(cpu, tbl[node]);
5452	}
5453
5454	wq_numa_possible_cpumask = tbl;
5455	wq_numa_enabled = true;
5456	}
5457
5458	static int __init init_workqueues(void)
5459	{
5460	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5461	int i, cpu;
5462
5463	WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5464
5465	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5466	cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5467
5468	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5469
5470	wq_numa_init();
5471
5472	/* initialize CPU pools */
5473	for_each_possible_cpu(cpu) {
5474	struct worker_pool *pool;
5475
5476	i = 0;
5477	for_each_cpu_worker_pool(pool, cpu) {
5478	BUG_ON(init_worker_pool(pool));
5479	pool->cpu = cpu;
5480	cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5481	pool->attrs->nice = std_nice[i++];
5482	pool->node = cpu_to_node(cpu);
5483
5484	/* alloc pool ID */
5485	mutex_lock(&wq_pool_mutex);
5486	BUG_ON(worker_pool_assign_id(pool));
5487	mutex_unlock(&wq_pool_mutex);
5488	}
5489	}
5490
5491	/* create the initial worker */
5492	for_each_online_cpu(cpu) {
5493	struct worker_pool *pool;
5494
5495	for_each_cpu_worker_pool(pool, cpu) {
5496	pool->flags &= ~POOL_DISASSOCIATED;
5497	BUG_ON(!create_worker(pool));
5498	}
5499	}
5500
5501	/* create default unbound and ordered wq attrs */
5502	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5503	struct workqueue_attrs *attrs;
5504
5505	BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5506	attrs->nice = std_nice[i];
5507	unbound_std_wq_attrs[i] = attrs;
5508
5509	/*
5510	* An ordered wq should have only one pwq as ordering is
5511	* guaranteed by max_active which is enforced by pwqs.
5512	* Turn off NUMA so that dfl_pwq is used for all nodes.
5513	*/
5514	BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5515	attrs->nice = std_nice[i];
5516	attrs->no_numa = true;
5517	ordered_wq_attrs[i] = attrs;
5518	}
5519
5520	system_wq = alloc_workqueue("events", 0, 0);
5521	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5522	system_long_wq = alloc_workqueue("events_long", 0, 0);
5523	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5524	WQ_UNBOUND_MAX_ACTIVE);
5525	system_freezable_wq = alloc_workqueue("events_freezable",
5526	WQ_FREEZABLE, 0);
5527	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5528	WQ_POWER_EFFICIENT, 0);
5529	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5530	WQ_FREEZABLE \| WQ_POWER_EFFICIENT,
5531	0);
5532	BUG_ON(!system_wq \|\| !system_highpri_wq \|\| !system_long_wq \|\|
5533	!system_unbound_wq \|\| !system_freezable_wq \|\|
5534	!system_power_efficient_wq \|\|
5535	!system_freezable_power_efficient_wq);
5536
5537	wq_watchdog_init();
5538
5539	return 0;
5540	}
5541	early_initcall(init_workqueues);