path: root/kernel/workqueue.c (plain)
blob: 6a293cb09a917c8cf98467797ff1b9e37e60d813

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 <dwmw2@infradead.org>
8	* Andrew Morton
9	* Kai Petzke <wpp@marie.physik.tu-berlin.de>
10	* Theodore Ts'o <tytso@mit.edu>
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 <tj@kernel.org>
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	#include <linux/nmi.h>
52
53	#include "workqueue_internal.h"
54
55	enum {
56	/*
57	* worker_pool flags
58	*
59	* A bound pool is either associated or disassociated with its CPU.
60	* While associated (!DISASSOCIATED), all workers are bound to the
61	* CPU and none has %WORKER_UNBOUND set and concurrency management
62	* is in effect.
63	*
64	* While DISASSOCIATED, the cpu may be offline and all workers have
65	* %WORKER_UNBOUND set and concurrency management disabled, and may
66	* be executing on any CPU. The pool behaves as an unbound one.
67	*
68	* Note that DISASSOCIATED should be flipped only while holding
69	* attach_mutex to avoid changing binding state while
70	* worker_attach_to_pool() is in progress.
71	*/
72	POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
73	POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
74
75	/* worker flags */
76	WORKER_DIE = 1 << 1, /* die die die */
77	WORKER_IDLE = 1 << 2, /* is idle */
78	WORKER_PREP = 1 << 3, /* preparing to run works */
79	WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
80	WORKER_UNBOUND = 1 << 7, /* worker is unbound */
81	WORKER_REBOUND = 1 << 8, /* worker was rebound */
82
83	WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
84	WORKER_UNBOUND | WORKER_REBOUND,
85
86	NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
87
88	UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
89	BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
90
91	MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
92	IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
93
94	MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
95	/* call for help after 10ms
96	(min two ticks) */
97	MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
98	CREATE_COOLDOWN = HZ, /* time to breath after fail */
99
100	/*
101	* Rescue workers are used only on emergencies and shared by
102	* all cpus. Give MIN_NICE.
103	*/
104	RESCUER_NICE_LEVEL = MIN_NICE,
105	HIGHPRI_NICE_LEVEL = MIN_NICE,
106
107	WQ_NAME_LEN = 24,
108	};
109
110	/*
111	* Structure fields follow one of the following exclusion rules.
112	*
113	* I: Modifiable by initialization/destruction paths and read-only for
114	* everyone else.
115	*
116	* P: Preemption protected. Disabling preemption is enough and should
117	* only be modified and accessed from the local cpu.
118	*
119	* L: pool->lock protected. Access with pool->lock held.
120	*
121	* X: During normal operation, modification requires pool->lock and should
122	* be done only from local cpu. Either disabling preemption on local
123	* cpu or grabbing pool->lock is enough for read access. If
124	* POOL_DISASSOCIATED is set, it's identical to L.
125	*
126	* A: pool->attach_mutex protected.
127	*
128	* PL: wq_pool_mutex protected.
129	*
130	* PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
131	*
132	* PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
133	*
134	* PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
135	* sched-RCU for reads.
136	*
137	* WQ: wq->mutex protected.
138	*
139	* WR: wq->mutex protected for writes. Sched-RCU protected for reads.
140	*
141	* MD: wq_mayday_lock protected.
142	*/
143
144	/* struct worker is defined in workqueue_internal.h */
145
146	struct worker_pool {
147	spinlock_t lock; /* the pool lock */
148	int cpu; /* I: the associated cpu */
149	int node; /* I: the associated node ID */
150	int id; /* I: pool ID */
151	unsigned int flags; /* X: flags */
152
153	unsigned long watchdog_ts; /* L: watchdog timestamp */
154
155	struct list_head worklist; /* L: list of pending works */
156	int nr_workers; /* L: total number of workers */
157
158	/* nr_idle includes the ones off idle_list for rebinding */
159	int nr_idle; /* L: currently idle ones */
160
161	struct list_head idle_list; /* X: list of idle workers */
162	struct timer_list idle_timer; /* L: worker idle timeout */
163	struct timer_list mayday_timer; /* L: SOS timer for workers */
164
165	/* a workers is either on busy_hash or idle_list, or the manager */
166	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
167	/* L: hash of busy workers */
168
169	/* see manage_workers() for details on the two manager mutexes */
170	struct worker *manager; /* L: purely informational */
171	struct mutex attach_mutex; /* attach/detach exclusion */
172	struct list_head workers; /* A: attached workers */
173	struct completion *detach_completion; /* all workers detached */
174
175	struct ida worker_ida; /* worker IDs for task name */
176
177	struct workqueue_attrs *attrs; /* I: worker attributes */
178	struct hlist_node hash_node; /* PL: unbound_pool_hash node */
179	int refcnt; /* PL: refcnt for unbound pools */
180
181	/*
182	* The current concurrency level. As it's likely to be accessed
183	* from other CPUs during try_to_wake_up(), put it in a separate
184	* cacheline.
185	*/
186	atomic_t nr_running ____cacheline_aligned_in_smp;
187
188	/*
189	* Destruction of pool is sched-RCU protected to allow dereferences
190	* from get_work_pool().
191	*/
192	struct rcu_head rcu;
193	} ____cacheline_aligned_in_smp;
194
195	/*
196	* The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
197	* of work_struct->data are used for flags and the remaining high bits
198	* point to the pwq; thus, pwqs need to be aligned at two's power of the
199	* number of flag bits.
200	*/
201	struct pool_workqueue {
202	struct worker_pool *pool; /* I: the associated pool */
203	struct workqueue_struct *wq; /* I: the owning workqueue */
204	int work_color; /* L: current color */
205	int flush_color; /* L: flushing color */
206	int refcnt; /* L: reference count */
207	int nr_in_flight[WORK_NR_COLORS];
208	/* L: nr of in_flight works */
209	int nr_active; /* L: nr of active works */
210	int max_active; /* L: max active works */
211	struct list_head delayed_works; /* L: delayed works */
212	struct list_head pwqs_node; /* WR: node on wq->pwqs */
213	struct list_head mayday_node; /* MD: node on wq->maydays */
214
215	/*
216	* Release of unbound pwq is punted to system_wq. See put_pwq()
217	* and pwq_unbound_release_workfn() for details. pool_workqueue
218	* itself is also sched-RCU protected so that the first pwq can be
219	* determined without grabbing wq->mutex.
220	*/
221	struct work_struct unbound_release_work;
222	struct rcu_head rcu;
223	} __aligned(1 << WORK_STRUCT_FLAG_BITS);
224
225	/*
226	* Structure used to wait for workqueue flush.
227	*/
228	struct wq_flusher {
229	struct list_head list; /* WQ: list of flushers */
230	int flush_color; /* WQ: flush color waiting for */
231	struct completion done; /* flush completion */
232	};
233
234	struct wq_device;
235
236	/*
237	* The externally visible workqueue. It relays the issued work items to
238	* the appropriate worker_pool through its pool_workqueues.
239	*/
240	struct workqueue_struct {
241	struct list_head pwqs; /* WR: all pwqs of this wq */
242	struct list_head list; /* PR: list of all workqueues */
243
244	struct mutex mutex; /* protects this wq */
245	int work_color; /* WQ: current work color */
246	int flush_color; /* WQ: current flush color */
247	atomic_t nr_pwqs_to_flush; /* flush in progress */
248	struct wq_flusher *first_flusher; /* WQ: first flusher */
249	struct list_head flusher_queue; /* WQ: flush waiters */
250	struct list_head flusher_overflow; /* WQ: flush overflow list */
251
252	struct list_head maydays; /* MD: pwqs requesting rescue */
253	struct worker *rescuer; /* I: rescue worker */
254
255	int nr_drainers; /* WQ: drain in progress */
256	int saved_max_active; /* WQ: saved pwq max_active */
257
258	struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
259	struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
260
261	#ifdef CONFIG_SYSFS
262	struct wq_device *wq_dev; /* I: for sysfs interface */
263	#endif
264	#ifdef CONFIG_LOCKDEP
265	struct lockdep_map lockdep_map;
266	#endif
267	char name[WQ_NAME_LEN]; /* I: workqueue name */
268
269	/*
270	* Destruction of workqueue_struct is sched-RCU protected to allow
271	* walking the workqueues list without grabbing wq_pool_mutex.
272	* This is used to dump all workqueues from sysrq.
273	*/
274	struct rcu_head rcu;
275
276	/* hot fields used during command issue, aligned to cacheline */
277	unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
278	struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
279	struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
280	};
281
282	static struct kmem_cache *pwq_cache;
283
284	static cpumask_var_t *wq_numa_possible_cpumask;
285	/* possible CPUs of each node */
286
287	static bool wq_disable_numa;
288	module_param_named(disable_numa, wq_disable_numa, bool, 0444);
289
290	/* see the comment above the definition of WQ_POWER_EFFICIENT */
291	static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
292	module_param_named(power_efficient, wq_power_efficient, bool, 0444);
293
294	bool wq_online; /* can kworkers be created yet? */
295
296	static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
297
298	/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
299	static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
300
301	static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
302	static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
303	static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
304
305	static LIST_HEAD(workqueues); /* PR: list of all workqueues */
306	static bool workqueue_freezing; /* PL: have wqs started freezing? */
307
308	/* PL: allowable cpus for unbound wqs and work items */
309	static cpumask_var_t wq_unbound_cpumask;
310
311	/* CPU where unbound work was last round robin scheduled from this CPU */
312	static DEFINE_PER_CPU(int, wq_rr_cpu_last);
313
314	/*
315	* Local execution of unbound work items is no longer guaranteed. The
316	* following always forces round-robin CPU selection on unbound work items
317	* to uncover usages which depend on it.
318	*/
319	#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
320	static bool wq_debug_force_rr_cpu = true;
321	#else
322	static bool wq_debug_force_rr_cpu = false;
323	#endif
324	module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
325
326	/* the per-cpu worker pools */
327	static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
328
329	static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
330
331	/* PL: hash of all unbound pools keyed by pool->attrs */
332	static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
333
334	/* I: attributes used when instantiating standard unbound pools on demand */
335	static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
336
337	/* I: attributes used when instantiating ordered pools on demand */
338	static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
339
340	struct workqueue_struct *system_wq __read_mostly;
341	EXPORT_SYMBOL(system_wq);
342	struct workqueue_struct *system_highpri_wq __read_mostly;
343	EXPORT_SYMBOL_GPL(system_highpri_wq);
344	struct workqueue_struct *system_long_wq __read_mostly;
345	EXPORT_SYMBOL_GPL(system_long_wq);
346	struct workqueue_struct *system_unbound_wq __read_mostly;
347	EXPORT_SYMBOL_GPL(system_unbound_wq);
348	struct workqueue_struct *system_freezable_wq __read_mostly;
349	EXPORT_SYMBOL_GPL(system_freezable_wq);
350	struct workqueue_struct *system_power_efficient_wq __read_mostly;
351	EXPORT_SYMBOL_GPL(system_power_efficient_wq);
352	struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
353	EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
354
355	static int worker_thread(void *__worker);
356	static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
357
358	#define CREATE_TRACE_POINTS
359	#include <trace/events/workqueue.h>
360
361	#define assert_rcu_or_pool_mutex() \
362	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
363	!lockdep_is_held(&wq_pool_mutex), \
364	"sched RCU or wq_pool_mutex should be held")
365
366	#define assert_rcu_or_wq_mutex(wq) \
367	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
368	!lockdep_is_held(&wq->mutex), \
369	"sched RCU or wq->mutex should be held")
370
371	#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
372	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
373	!lockdep_is_held(&wq->mutex) && \
374	!lockdep_is_held(&wq_pool_mutex), \
375	"sched RCU, wq->mutex or wq_pool_mutex should be held")
376
377	#define for_each_cpu_worker_pool(pool, cpu) \
378	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
379	(pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
380	(pool)++)
381
382	/**
383	* for_each_pool - iterate through all worker_pools in the system
384	* @pool: iteration cursor
385	* @pi: integer used for iteration
386	*
387	* This must be called either with wq_pool_mutex held or sched RCU read
388	* locked. If the pool needs to be used beyond the locking in effect, the
389	* caller is responsible for guaranteeing that the pool stays online.
390	*
391	* The if/else clause exists only for the lockdep assertion and can be
392	* ignored.
393	*/
394	#define for_each_pool(pool, pi) \
395	idr_for_each_entry(&worker_pool_idr, pool, pi) \
396	if (({ assert_rcu_or_pool_mutex(); false; })) { } \
397	else
398
399	/**
400	* for_each_pool_worker - iterate through all workers of a worker_pool
401	* @worker: iteration cursor
402	* @pool: worker_pool to iterate workers of
403	*
404	* This must be called with @pool->attach_mutex.
405	*
406	* The if/else clause exists only for the lockdep assertion and can be
407	* ignored.
408	*/
409	#define for_each_pool_worker(worker, pool) \
410	list_for_each_entry((worker), &(pool)->workers, node) \
411	if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
412	else
413
414	/**
415	* for_each_pwq - iterate through all pool_workqueues of the specified workqueue
416	* @pwq: iteration cursor
417	* @wq: the target workqueue
418	*
419	* This must be called either with wq->mutex held or sched RCU read locked.
420	* If the pwq needs to be used beyond the locking in effect, the caller is
421	* responsible for guaranteeing that the pwq stays online.
422	*
423	* The if/else clause exists only for the lockdep assertion and can be
424	* ignored.
425	*/
426	#define for_each_pwq(pwq, wq) \
427	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
428	if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
429	else
430
431	#ifdef CONFIG_DEBUG_OBJECTS_WORK
432
433	static struct debug_obj_descr work_debug_descr;
434
435	static void *work_debug_hint(void *addr)
436	{
437	return ((struct work_struct *) addr)->func;
438	}
439
440	static bool work_is_static_object(void *addr)
441	{
442	struct work_struct *work = addr;
443
444	return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
445	}
446
447	/*
448	* fixup_init is called when:
449	* - an active object is initialized
450	*/
451	static bool work_fixup_init(void *addr, enum debug_obj_state state)
452	{
453	struct work_struct *work = addr;
454
455	switch (state) {
456	case ODEBUG_STATE_ACTIVE:
457	cancel_work_sync(work);
458	debug_object_init(work, &work_debug_descr);
459	return true;
460	default:
461	return false;
462	}
463	}
464
465	/*
466	* fixup_free is called when:
467	* - an active object is freed
468	*/
469	static bool work_fixup_free(void *addr, enum debug_obj_state state)
470	{
471	struct work_struct *work = addr;
472
473	switch (state) {
474	case ODEBUG_STATE_ACTIVE:
475	cancel_work_sync(work);
476	debug_object_free(work, &work_debug_descr);
477	return true;
478	default:
479	return false;
480	}
481	}
482
483	static struct debug_obj_descr work_debug_descr = {
484	.name = "work_struct",
485	.debug_hint = work_debug_hint,
486	.is_static_object = work_is_static_object,
487	.fixup_init = work_fixup_init,
488	.fixup_free = work_fixup_free,
489	};
490
491	static inline void debug_work_activate(struct work_struct *work)
492	{
493	debug_object_activate(work, &work_debug_descr);
494	}
495
496	static inline void debug_work_deactivate(struct work_struct *work)
497	{
498	debug_object_deactivate(work, &work_debug_descr);
499	}
500
501	void __init_work(struct work_struct *work, int onstack)
502	{
503	if (onstack)
504	debug_object_init_on_stack(work, &work_debug_descr);
505	else
506	debug_object_init(work, &work_debug_descr);
507	}
508	EXPORT_SYMBOL_GPL(__init_work);
509
510	void destroy_work_on_stack(struct work_struct *work)
511	{
512	debug_object_free(work, &work_debug_descr);
513	}
514	EXPORT_SYMBOL_GPL(destroy_work_on_stack);
515
516	void destroy_delayed_work_on_stack(struct delayed_work *work)
517	{
518	destroy_timer_on_stack(&work->timer);
519	debug_object_free(&work->work, &work_debug_descr);
520	}
521	EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
522
523	#else
524	static inline void debug_work_activate(struct work_struct *work) { }
525	static inline void debug_work_deactivate(struct work_struct *work) { }
526	#endif
527
528	/**
529	* worker_pool_assign_id - allocate ID and assing it to @pool
530	* @pool: the pool pointer of interest
531	*
532	* Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
533	* successfully, -errno on failure.
534	*/
535	static int worker_pool_assign_id(struct worker_pool *pool)
536	{
537	int ret;
538
539	lockdep_assert_held(&wq_pool_mutex);
540
541	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
542	GFP_KERNEL);
543	if (ret >= 0) {
544	pool->id = ret;
545	return 0;
546	}
547	return ret;
548	}
549
550	/**
551	* unbound_pwq_by_node - return the unbound pool_workqueue for the given node
552	* @wq: the target workqueue
553	* @node: the node ID
554	*
555	* This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
556	* read locked.
557	* If the pwq needs to be used beyond the locking in effect, the caller is
558	* responsible for guaranteeing that the pwq stays online.
559	*
560	* Return: The unbound pool_workqueue for @node.
561	*/
562	static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
563	int node)
564	{
565	assert_rcu_or_wq_mutex_or_pool_mutex(wq);
566
567	/*
568	* XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
569	* delayed item is pending. The plan is to keep CPU -> NODE
570	* mapping valid and stable across CPU on/offlines. Once that
571	* happens, this workaround can be removed.
572	*/
573	if (unlikely(node == NUMA_NO_NODE))
574	return wq->dfl_pwq;
575
576	return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
577	}
578
579	static unsigned int work_color_to_flags(int color)
580	{
581	return color << WORK_STRUCT_COLOR_SHIFT;
582	}
583
584	static int get_work_color(struct work_struct *work)
585	{
586	return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
587	((1 << WORK_STRUCT_COLOR_BITS) - 1);
588	}
589
590	static int work_next_color(int color)
591	{
592	return (color + 1) % WORK_NR_COLORS;
593	}
594
595	/*
596	* While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
597	* contain the pointer to the queued pwq. Once execution starts, the flag
598	* is cleared and the high bits contain OFFQ flags and pool ID.
599	*
600	* set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
601	* and clear_work_data() can be used to set the pwq, pool or clear
602	* work->data. These functions should only be called while the work is
603	* owned - ie. while the PENDING bit is set.
604	*
605	* get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
606	* corresponding to a work. Pool is available once the work has been
607	* queued anywhere after initialization until it is sync canceled. pwq is
608	* available only while the work item is queued.
609	*
610	* %WORK_OFFQ_CANCELING is used to mark a work item which is being
611	* canceled. While being canceled, a work item may have its PENDING set
612	* but stay off timer and worklist for arbitrarily long and nobody should
613	* try to steal the PENDING bit.
614	*/
615	static inline void set_work_data(struct work_struct *work, unsigned long data,
616	unsigned long flags)
617	{
618	WARN_ON_ONCE(!work_pending(work));
619	atomic_long_set(&work->data, data | flags | work_static(work));
620	}
621
622	static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
623	unsigned long extra_flags)
624	{
625	set_work_data(work, (unsigned long)pwq,
626	WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
627	}
628
629	static void set_work_pool_and_keep_pending(struct work_struct *work,
630	int pool_id)
631	{
632	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
633	WORK_STRUCT_PENDING);
634	}
635
636	static void set_work_pool_and_clear_pending(struct work_struct *work,
637	int pool_id)
638	{
639	/*
640	* The following wmb is paired with the implied mb in
641	* test_and_set_bit(PENDING) and ensures all updates to @work made
642	* here are visible to and precede any updates by the next PENDING
643	* owner.
644	*/
645	smp_wmb();
646	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
647	/*
648	* The following mb guarantees that previous clear of a PENDING bit
649	* will not be reordered with any speculative LOADS or STORES from
650	* work->current_func, which is executed afterwards. This possible
651	* reordering can lead to a missed execution on attempt to qeueue
652	* the same @work. E.g. consider this case:
653	*
654	* CPU#0 CPU#1
655	* ---------------------------- --------------------------------
656	*
657	* 1 STORE event_indicated
658	* 2 queue_work_on() {
659	* 3 test_and_set_bit(PENDING)
660	* 4 } set_..._and_clear_pending() {
661	* 5 set_work_data() # clear bit
662	* 6 smp_mb()
663	* 7 work->current_func() {
664	* 8 LOAD event_indicated
665	* }
666	*
667	* Without an explicit full barrier speculative LOAD on line 8 can
668	* be executed before CPU#0 does STORE on line 1. If that happens,
669	* CPU#0 observes the PENDING bit is still set and new execution of
670	* a @work is not queued in a hope, that CPU#1 will eventually
671	* finish the queued @work. Meanwhile CPU#1 does not see
672	* event_indicated is set, because speculative LOAD was executed
673	* before actual STORE.
674	*/
675	smp_mb();
676	}
677
678	static void clear_work_data(struct work_struct *work)
679	{
680	smp_wmb(); /* see set_work_pool_and_clear_pending() */
681	set_work_data(work, WORK_STRUCT_NO_POOL, 0);
682	}
683
684	static struct pool_workqueue *get_work_pwq(struct work_struct *work)
685	{
686	unsigned long data = atomic_long_read(&work->data);
687
688	if (data & WORK_STRUCT_PWQ)
689	return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
690	else
691	return NULL;
692	}
693
694	/**
695	* get_work_pool - return the worker_pool a given work was associated with
696	* @work: the work item of interest
697	*
698	* Pools are created and destroyed under wq_pool_mutex, and allows read
699	* access under sched-RCU read lock. As such, this function should be
700	* called under wq_pool_mutex or with preemption disabled.
701	*
702	* All fields of the returned pool are accessible as long as the above
703	* mentioned locking is in effect. If the returned pool needs to be used
704	* beyond the critical section, the caller is responsible for ensuring the
705	* returned pool is and stays online.
706	*
707	* Return: The worker_pool @work was last associated with. %NULL if none.
708	*/
709	static struct worker_pool *get_work_pool(struct work_struct *work)
710	{
711	unsigned long data = atomic_long_read(&work->data);
712	int pool_id;
713
714	assert_rcu_or_pool_mutex();
715
716	if (data & WORK_STRUCT_PWQ)
717	return ((struct pool_workqueue *)
718	(data & WORK_STRUCT_WQ_DATA_MASK))->pool;
719
720	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
721	if (pool_id == WORK_OFFQ_POOL_NONE)
722	return NULL;
723
724	return idr_find(&worker_pool_idr, pool_id);
725	}
726
727	/**
728	* get_work_pool_id - return the worker pool ID a given work is associated with
729	* @work: the work item of interest
730	*
731	* Return: The worker_pool ID @work was last associated with.
732	* %WORK_OFFQ_POOL_NONE if none.
733	*/
734	static int get_work_pool_id(struct work_struct *work)
735	{
736	unsigned long data = atomic_long_read(&work->data);
737
738	if (data & WORK_STRUCT_PWQ)
739	return ((struct pool_workqueue *)
740	(data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
741
742	return data >> WORK_OFFQ_POOL_SHIFT;
743	}
744
745	static void mark_work_canceling(struct work_struct *work)
746	{
747	unsigned long pool_id = get_work_pool_id(work);
748
749	pool_id <<= WORK_OFFQ_POOL_SHIFT;
750	set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
751	}
752
753	static bool work_is_canceling(struct work_struct *work)
754	{
755	unsigned long data = atomic_long_read(&work->data);
756
757	return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
758	}
759
760	/*
761	* Policy functions. These define the policies on how the global worker
762	* pools are managed. Unless noted otherwise, these functions assume that
763	* they're being called with pool->lock held.
764	*/
765
766	static bool __need_more_worker(struct worker_pool *pool)
767	{
768	return !atomic_read(&pool->nr_running);
769	}
770
771	/*
772	* Need to wake up a worker? Called from anything but currently
773	* running workers.
774	*
775	* Note that, because unbound workers never contribute to nr_running, this
776	* function will always return %true for unbound pools as long as the
777	* worklist isn't empty.
778	*/
779	static bool need_more_worker(struct worker_pool *pool)
780	{
781	return !list_empty(&pool->worklist) && __need_more_worker(pool);
782	}
783
784	/* Can I start working? Called from busy but !running workers. */
785	static bool may_start_working(struct worker_pool *pool)
786	{
787	return pool->nr_idle;
788	}
789
790	/* Do I need to keep working? Called from currently running workers. */
791	static bool keep_working(struct worker_pool *pool)
792	{
793	return !list_empty(&pool->worklist) &&
794	atomic_read(&pool->nr_running) <= 1;
795	}
796
797	/* Do we need a new worker? Called from manager. */
798	static bool need_to_create_worker(struct worker_pool *pool)
799	{
800	return need_more_worker(pool) && !may_start_working(pool);
801	}
802
803	/* Do we have too many workers and should some go away? */
804	static bool too_many_workers(struct worker_pool *pool)
805	{
806	bool managing = pool->flags & POOL_MANAGER_ACTIVE;
807	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
808	int nr_busy = pool->nr_workers - nr_idle;
809
810	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
811	}
812
813	/*
814	* Wake up functions.
815	*/
816
817	/* Return the first idle worker. Safe with preemption disabled */
818	static struct worker *first_idle_worker(struct worker_pool *pool)
819	{
820	if (unlikely(list_empty(&pool->idle_list)))
821	return NULL;
822
823	return list_first_entry(&pool->idle_list, struct worker, entry);
824	}
825
826	/**
827	* wake_up_worker - wake up an idle worker
828	* @pool: worker pool to wake worker from
829	*
830	* Wake up the first idle worker of @pool.
831	*
832	* CONTEXT:
833	* spin_lock_irq(pool->lock).
834	*/
835	static void wake_up_worker(struct worker_pool *pool)
836	{
837	struct worker *worker = first_idle_worker(pool);
838
839	if (likely(worker))
840	wake_up_process(worker->task);
841	}
842
843	/**
844	* wq_worker_waking_up - a worker is waking up
845	* @task: task waking up
846	* @cpu: CPU @task is waking up to
847	*
848	* This function is called during try_to_wake_up() when a worker is
849	* being awoken.
850	*
851	* CONTEXT:
852	* spin_lock_irq(rq->lock)
853	*/
854	void wq_worker_waking_up(struct task_struct *task, int cpu)
855	{
856	struct worker *worker = kthread_data(task);
857
858	if (!(worker->flags & WORKER_NOT_RUNNING)) {
859	WARN_ON_ONCE(worker->pool->cpu != cpu);
860	atomic_inc(&worker->pool->nr_running);
861	}
862	}
863
864	/**
865	* wq_worker_sleeping - a worker is going to sleep
866	* @task: task going to sleep
867	*
868	* This function is called during schedule() when a busy worker is
869	* going to sleep. Worker on the same cpu can be woken up by
870	* returning pointer to its task.
871	*
872	* CONTEXT:
873	* spin_lock_irq(rq->lock)
874	*
875	* Return:
876	* Worker task on @cpu to wake up, %NULL if none.
877	*/
878	struct task_struct *wq_worker_sleeping(struct task_struct *task)
879	{
880	struct worker *worker = kthread_data(task), *to_wakeup = NULL;
881	struct worker_pool *pool;
882
883	/*
884	* Rescuers, which may not have all the fields set up like normal
885	* workers, also reach here, let's not access anything before
886	* checking NOT_RUNNING.
887	*/
888	if (worker->flags & WORKER_NOT_RUNNING)
889	return NULL;
890
891	pool = worker->pool;
892
893	/* this can only happen on the local cpu */
894	if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
895	return NULL;
896
897	/*
898	* The counterpart of the following dec_and_test, implied mb,
899	* worklist not empty test sequence is in insert_work().
900	* Please read comment there.
901	*
902	* NOT_RUNNING is clear. This means that we're bound to and
903	* running on the local cpu w/ rq lock held and preemption
904	* disabled, which in turn means that none else could be
905	* manipulating idle_list, so dereferencing idle_list without pool
906	* lock is safe.
907	*/
908	if (atomic_dec_and_test(&pool->nr_running) &&
909	!list_empty(&pool->worklist))
910	to_wakeup = first_idle_worker(pool);
911	return to_wakeup ? to_wakeup->task : NULL;
912	}
913
914	/**
915	* wq_worker_last_func - retrieve worker's last work function
916	*
917	* Determine the last function a worker executed. This is called from
918	* the scheduler to get a worker's last known identity.
919	*
920	* CONTEXT:
921	* spin_lock_irq(rq->lock)
922	*
923	* Return:
924	* The last work function %current executed as a worker, NULL if it
925	* hasn't executed any work yet.
926	*/
927	work_func_t wq_worker_last_func(struct task_struct *task)
928	{
929	struct worker *worker = kthread_data(task);
930
931	return worker->last_func;
932	}
933
934	/**
935	* worker_set_flags - set worker flags and adjust nr_running accordingly
936	* @worker: self
937	* @flags: flags to set
938	*
939	* Set @flags in @worker->flags and adjust nr_running accordingly.
940	*
941	* CONTEXT:
942	* spin_lock_irq(pool->lock)
943	*/
944	static inline void worker_set_flags(struct worker *worker, unsigned int flags)
945	{
946	struct worker_pool *pool = worker->pool;
947
948	WARN_ON_ONCE(worker->task != current);
949
950	/* If transitioning into NOT_RUNNING, adjust nr_running. */
951	if ((flags & WORKER_NOT_RUNNING) &&
952	!(worker->flags & WORKER_NOT_RUNNING)) {
953	atomic_dec(&pool->nr_running);
954	}
955
956	worker->flags |= flags;
957	}
958
959	/**
960	* worker_clr_flags - clear worker flags and adjust nr_running accordingly
961	* @worker: self
962	* @flags: flags to clear
963	*
964	* Clear @flags in @worker->flags and adjust nr_running accordingly.
965	*
966	* CONTEXT:
967	* spin_lock_irq(pool->lock)
968	*/
969	static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
970	{
971	struct worker_pool *pool = worker->pool;
972	unsigned int oflags = worker->flags;
973
974	WARN_ON_ONCE(worker->task != current);
975
976	worker->flags &= ~flags;
977
978	/*
979	* If transitioning out of NOT_RUNNING, increment nr_running. Note
980	* that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
981	* of multiple flags, not a single flag.
982	*/
983	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
984	if (!(worker->flags & WORKER_NOT_RUNNING))
985	atomic_inc(&pool->nr_running);
986	}
987
988	/**
989	* find_worker_executing_work - find worker which is executing a work
990	* @pool: pool of interest
991	* @work: work to find worker for
992	*
993	* Find a worker which is executing @work on @pool by searching
994	* @pool->busy_hash which is keyed by the address of @work. For a worker
995	* to match, its current execution should match the address of @work and
996	* its work function. This is to avoid unwanted dependency between
997	* unrelated work executions through a work item being recycled while still
998	* being executed.
999	*
1000	* This is a bit tricky. A work item may be freed once its execution
1001	* starts and nothing prevents the freed area from being recycled for
1002	* another work item. If the same work item address ends up being reused
1003	* before the original execution finishes, workqueue will identify the
1004	* recycled work item as currently executing and make it wait until the
1005	* current execution finishes, introducing an unwanted dependency.
1006	*
1007	* This function checks the work item address and work function to avoid
1008	* false positives. Note that this isn't complete as one may construct a
1009	* work function which can introduce dependency onto itself through a
1010	* recycled work item. Well, if somebody wants to shoot oneself in the
1011	* foot that badly, there's only so much we can do, and if such deadlock
1012	* actually occurs, it should be easy to locate the culprit work function.
1013	*
1014	* CONTEXT:
1015	* spin_lock_irq(pool->lock).
1016	*
1017	* Return:
1018	* Pointer to worker which is executing @work if found, %NULL
1019	* otherwise.
1020	*/
1021	static struct worker *find_worker_executing_work(struct worker_pool *pool,
1022	struct work_struct *work)
1023	{
1024	struct worker *worker;
1025
1026	hash_for_each_possible(pool->busy_hash, worker, hentry,
1027	(unsigned long)work)
1028	if (worker->current_work == work &&
1029	worker->current_func == work->func)
1030	return worker;
1031
1032	return NULL;
1033	}
1034
1035	/**
1036	* move_linked_works - move linked works to a list
1037	* @work: start of series of works to be scheduled
1038	* @head: target list to append @work to
1039	* @nextp: out parameter for nested worklist walking
1040	*
1041	* Schedule linked works starting from @work to @head. Work series to
1042	* be scheduled starts at @work and includes any consecutive work with
1043	* WORK_STRUCT_LINKED set in its predecessor.
1044	*
1045	* If @nextp is not NULL, it's updated to point to the next work of
1046	* the last scheduled work. This allows move_linked_works() to be
1047	* nested inside outer list_for_each_entry_safe().
1048	*
1049	* CONTEXT:
1050	* spin_lock_irq(pool->lock).
1051	*/
1052	static void move_linked_works(struct work_struct *work, struct list_head *head,
1053	struct work_struct **nextp)
1054	{
1055	struct work_struct *n;
1056
1057	/*
1058	* Linked worklist will always end before the end of the list,
1059	* use NULL for list head.
1060	*/
1061	list_for_each_entry_safe_from(work, n, NULL, entry) {
1062	list_move_tail(&work->entry, head);
1063	if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1064	break;
1065	}
1066
1067	/*
1068	* If we're already inside safe list traversal and have moved
1069	* multiple works to the scheduled queue, the next position
1070	* needs to be updated.
1071	*/
1072	if (nextp)
1073	*nextp = n;
1074	}
1075
1076	/**
1077	* get_pwq - get an extra reference on the specified pool_workqueue
1078	* @pwq: pool_workqueue to get
1079	*
1080	* Obtain an extra reference on @pwq. The caller should guarantee that
1081	* @pwq has positive refcnt and be holding the matching pool->lock.
1082	*/
1083	static void get_pwq(struct pool_workqueue *pwq)
1084	{
1085	lockdep_assert_held(&pwq->pool->lock);
1086	WARN_ON_ONCE(pwq->refcnt <= 0);
1087	pwq->refcnt++;
1088	}
1089
1090	/**
1091	* put_pwq - put a pool_workqueue reference
1092	* @pwq: pool_workqueue to put
1093	*
1094	* Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1095	* destruction. The caller should be holding the matching pool->lock.
1096	*/
1097	static void put_pwq(struct pool_workqueue *pwq)
1098	{
1099	lockdep_assert_held(&pwq->pool->lock);
1100	if (likely(--pwq->refcnt))
1101	return;
1102	if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1103	return;
1104	/*
1105	* @pwq can't be released under pool->lock, bounce to
1106	* pwq_unbound_release_workfn(). This never recurses on the same
1107	* pool->lock as this path is taken only for unbound workqueues and
1108	* the release work item is scheduled on a per-cpu workqueue. To
1109	* avoid lockdep warning, unbound pool->locks are given lockdep
1110	* subclass of 1 in get_unbound_pool().
1111	*/
1112	schedule_work(&pwq->unbound_release_work);
1113	}
1114
1115	/**
1116	* put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1117	* @pwq: pool_workqueue to put (can be %NULL)
1118	*
1119	* put_pwq() with locking. This function also allows %NULL @pwq.
1120	*/
1121	static void put_pwq_unlocked(struct pool_workqueue *pwq)
1122	{
1123	if (pwq) {
1124	/*
1125	* As both pwqs and pools are sched-RCU protected, the
1126	* following lock operations are safe.
1127	*/
1128	spin_lock_irq(&pwq->pool->lock);
1129	put_pwq(pwq);
1130	spin_unlock_irq(&pwq->pool->lock);
1131	}
1132	}
1133
1134	static void pwq_activate_delayed_work(struct work_struct *work)
1135	{
1136	struct pool_workqueue *pwq = get_work_pwq(work);
1137
1138	trace_workqueue_activate_work(work);
1139	if (list_empty(&pwq->pool->worklist))
1140	pwq->pool->watchdog_ts = jiffies;
1141	move_linked_works(work, &pwq->pool->worklist, NULL);
1142	__clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1143	pwq->nr_active++;
1144	}
1145
1146	static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1147	{
1148	struct work_struct *work = list_first_entry(&pwq->delayed_works,
1149	struct work_struct, entry);
1150
1151	pwq_activate_delayed_work(work);
1152	}
1153
1154	/**
1155	* pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1156	* @pwq: pwq of interest
1157	* @color: color of work which left the queue
1158	*
1159	* A work either has completed or is removed from pending queue,
1160	* decrement nr_in_flight of its pwq and handle workqueue flushing.
1161	*
1162	* CONTEXT:
1163	* spin_lock_irq(pool->lock).
1164	*/
1165	static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1166	{
1167	/* uncolored work items don't participate in flushing or nr_active */
1168	if (color == WORK_NO_COLOR)
1169	goto out_put;
1170
1171	pwq->nr_in_flight[color]--;
1172
1173	pwq->nr_active--;
1174	if (!list_empty(&pwq->delayed_works)) {
1175	/* one down, submit a delayed one */
1176	if (pwq->nr_active < pwq->max_active)
1177	pwq_activate_first_delayed(pwq);
1178	}
1179
1180	/* is flush in progress and are we at the flushing tip? */
1181	if (likely(pwq->flush_color != color))
1182	goto out_put;
1183
1184	/* are there still in-flight works? */
1185	if (pwq->nr_in_flight[color])
1186	goto out_put;
1187
1188	/* this pwq is done, clear flush_color */
1189	pwq->flush_color = -1;
1190
1191	/*
1192	* If this was the last pwq, wake up the first flusher. It
1193	* will handle the rest.
1194	*/
1195	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1196	complete(&pwq->wq->first_flusher->done);
1197	out_put:
1198	put_pwq(pwq);
1199	}
1200
1201	/**
1202	* try_to_grab_pending - steal work item from worklist and disable irq
1203	* @work: work item to steal
1204	* @is_dwork: @work is a delayed_work
1205	* @flags: place to store irq state
1206	*
1207	* Try to grab PENDING bit of @work. This function can handle @work in any
1208	* stable state - idle, on timer or on worklist.
1209	*
1210	* Return:
1211	* 1 if @work was pending and we successfully stole PENDING
1212	* 0 if @work was idle and we claimed PENDING
1213	* -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1214	* -ENOENT if someone else is canceling @work, this state may persist
1215	* for arbitrarily long
1216	*
1217	* Note:
1218	* On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1219	* interrupted while holding PENDING and @work off queue, irq must be
1220	* disabled on entry. This, combined with delayed_work->timer being
1221	* irqsafe, ensures that we return -EAGAIN for finite short period of time.
1222	*
1223	* On successful return, >= 0, irq is disabled and the caller is
1224	* responsible for releasing it using local_irq_restore(*@flags).
1225	*
1226	* This function is safe to call from any context including IRQ handler.
1227	*/
1228	static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1229	unsigned long *flags)
1230	{
1231	struct worker_pool *pool;
1232	struct pool_workqueue *pwq;
1233
1234	local_irq_save(*flags);
1235
1236	/* try to steal the timer if it exists */
1237	if (is_dwork) {
1238	struct delayed_work *dwork = to_delayed_work(work);
1239
1240	/*
1241	* dwork->timer is irqsafe. If del_timer() fails, it's
1242	* guaranteed that the timer is not queued anywhere and not
1243	* running on the local CPU.
1244	*/
1245	if (likely(del_timer(&dwork->timer)))
1246	return 1;
1247	}
1248
1249	/* try to claim PENDING the normal way */
1250	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1251	return 0;
1252
1253	/*
1254	* The queueing is in progress, or it is already queued. Try to
1255	* steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1256	*/
1257	pool = get_work_pool(work);
1258	if (!pool)
1259	goto fail;
1260
1261	spin_lock(&pool->lock);
1262	/*
1263	* work->data is guaranteed to point to pwq only while the work
1264	* item is queued on pwq->wq, and both updating work->data to point
1265	* to pwq on queueing and to pool on dequeueing are done under
1266	* pwq->pool->lock. This in turn guarantees that, if work->data
1267	* points to pwq which is associated with a locked pool, the work
1268	* item is currently queued on that pool.
1269	*/
1270	pwq = get_work_pwq(work);
1271	if (pwq && pwq->pool == pool) {
1272	debug_work_deactivate(work);
1273
1274	/*
1275	* A delayed work item cannot be grabbed directly because
1276	* it might have linked NO_COLOR work items which, if left
1277	* on the delayed_list, will confuse pwq->nr_active
1278	* management later on and cause stall. Make sure the work
1279	* item is activated before grabbing.
1280	*/
1281	if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1282	pwq_activate_delayed_work(work);
1283
1284	list_del_init(&work->entry);
1285	pwq_dec_nr_in_flight(pwq, get_work_color(work));
1286
1287	/* work->data points to pwq iff queued, point to pool */
1288	set_work_pool_and_keep_pending(work, pool->id);
1289
1290	spin_unlock(&pool->lock);
1291	return 1;
1292	}
1293	spin_unlock(&pool->lock);
1294	fail:
1295	local_irq_restore(*flags);
1296	if (work_is_canceling(work))
1297	return -ENOENT;
1298	cpu_relax();
1299	return -EAGAIN;
1300	}
1301
1302	/**
1303	* insert_work - insert a work into a pool
1304	* @pwq: pwq @work belongs to
1305	* @work: work to insert
1306	* @head: insertion point
1307	* @extra_flags: extra WORK_STRUCT_* flags to set
1308	*
1309	* Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1310	* work_struct flags.
1311	*
1312	* CONTEXT:
1313	* spin_lock_irq(pool->lock).
1314	*/
1315	static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1316	struct list_head *head, unsigned int extra_flags)
1317	{
1318	struct worker_pool *pool = pwq->pool;
1319
1320	/* we own @work, set data and link */
1321	set_work_pwq(work, pwq, extra_flags);
1322	list_add_tail(&work->entry, head);
1323	get_pwq(pwq);
1324
1325	/*
1326	* Ensure either wq_worker_sleeping() sees the above
1327	* list_add_tail() or we see zero nr_running to avoid workers lying
1328	* around lazily while there are works to be processed.
1329	*/
1330	smp_mb();
1331
1332	if (__need_more_worker(pool))
1333	wake_up_worker(pool);
1334	}
1335
1336	/*
1337	* Test whether @work is being queued from another work executing on the
1338	* same workqueue.
1339	*/
1340	static bool is_chained_work(struct workqueue_struct *wq)
1341	{
1342	struct worker *worker;
1343
1344	worker = current_wq_worker();
1345	/*
1346	* Return %true iff I'm a worker execuing a work item on @wq. If
1347	* I'm @worker, it's safe to dereference it without locking.
1348	*/
1349	return worker && worker->current_pwq->wq == wq;
1350	}
1351
1352	/*
1353	* When queueing an unbound work item to a wq, prefer local CPU if allowed
1354	* by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1355	* avoid perturbing sensitive tasks.
1356	*/
1357	static int wq_select_unbound_cpu(int cpu)
1358	{
1359	static bool printed_dbg_warning;
1360	int new_cpu;
1361
1362	if (likely(!wq_debug_force_rr_cpu)) {
1363	if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1364	return cpu;
1365	} else if (!printed_dbg_warning) {
1366	pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1367	printed_dbg_warning = true;
1368	}
1369
1370	if (cpumask_empty(wq_unbound_cpumask))
1371	return cpu;
1372
1373	new_cpu = __this_cpu_read(wq_rr_cpu_last);
1374	new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1375	if (unlikely(new_cpu >= nr_cpu_ids)) {
1376	new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1377	if (unlikely(new_cpu >= nr_cpu_ids))
1378	return cpu;
1379	}
1380	__this_cpu_write(wq_rr_cpu_last, new_cpu);
1381
1382	return new_cpu;
1383	}
1384
1385	static void __queue_work(int cpu, struct workqueue_struct *wq,
1386	struct work_struct *work)
1387	{
1388	struct pool_workqueue *pwq;
1389	struct worker_pool *last_pool;
1390	struct list_head *worklist;
1391	unsigned int work_flags;
1392	unsigned int req_cpu = cpu;
1393
1394	/*
1395	* While a work item is PENDING && off queue, a task trying to
1396	* steal the PENDING will busy-loop waiting for it to either get
1397	* queued or lose PENDING. Grabbing PENDING and queueing should
1398	* happen with IRQ disabled.
1399	*/
1400	WARN_ON_ONCE(!irqs_disabled());
1401
1402	debug_work_activate(work);
1403
1404	/* if draining, only works from the same workqueue are allowed */
1405	if (unlikely(wq->flags & __WQ_DRAINING) &&
1406	WARN_ON_ONCE(!is_chained_work(wq)))
1407	return;
1408	retry:
1409	if (req_cpu == WORK_CPU_UNBOUND)
1410	cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1411
1412	/* pwq which will be used unless @work is executing elsewhere */
1413	if (!(wq->flags & WQ_UNBOUND))
1414	pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1415	else
1416	pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1417
1418	/*
1419	* If @work was previously on a different pool, it might still be
1420	* running there, in which case the work needs to be queued on that
1421	* pool to guarantee non-reentrancy.
1422	*/
1423	last_pool = get_work_pool(work);
1424	if (last_pool && last_pool != pwq->pool) {
1425	struct worker *worker;
1426
1427	spin_lock(&last_pool->lock);
1428
1429	worker = find_worker_executing_work(last_pool, work);
1430
1431	if (worker && worker->current_pwq->wq == wq) {
1432	pwq = worker->current_pwq;
1433	} else {
1434	/* meh... not running there, queue here */
1435	spin_unlock(&last_pool->lock);
1436	spin_lock(&pwq->pool->lock);
1437	}
1438	} else {
1439	spin_lock(&pwq->pool->lock);
1440	}
1441
1442	/*
1443	* pwq is determined and locked. For unbound pools, we could have
1444	* raced with pwq release and it could already be dead. If its
1445	* refcnt is zero, repeat pwq selection. Note that pwqs never die
1446	* without another pwq replacing it in the numa_pwq_tbl or while
1447	* work items are executing on it, so the retrying is guaranteed to
1448	* make forward-progress.
1449	*/
1450	if (unlikely(!pwq->refcnt)) {
1451	if (wq->flags & WQ_UNBOUND) {
1452	spin_unlock(&pwq->pool->lock);
1453	cpu_relax();
1454	goto retry;
1455	}
1456	/* oops */
1457	WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1458	wq->name, cpu);
1459	}
1460
1461	/* pwq determined, queue */
1462	trace_workqueue_queue_work(req_cpu, pwq, work);
1463
1464	if (WARN_ON(!list_empty(&work->entry))) {
1465	spin_unlock(&pwq->pool->lock);
1466	return;
1467	}
1468
1469	pwq->nr_in_flight[pwq->work_color]++;
1470	work_flags = work_color_to_flags(pwq->work_color);
1471
1472	if (likely(pwq->nr_active < pwq->max_active)) {
1473	trace_workqueue_activate_work(work);
1474	pwq->nr_active++;
1475	worklist = &pwq->pool->worklist;
1476	if (list_empty(worklist))
1477	pwq->pool->watchdog_ts = jiffies;
1478	} else {
1479	work_flags |= WORK_STRUCT_DELAYED;
1480	worklist = &pwq->delayed_works;
1481	}
1482
1483	insert_work(pwq, work, worklist, work_flags);
1484
1485	spin_unlock(&pwq->pool->lock);
1486	}
1487
1488	/**
1489	* queue_work_on - queue work on specific cpu
1490	* @cpu: CPU number to execute work on
1491	* @wq: workqueue to use
1492	* @work: work to queue
1493	*
1494	* We queue the work to a specific CPU, the caller must ensure it
1495	* can't go away.
1496	*
1497	* Return: %false if @work was already on a queue, %true otherwise.
1498	*/
1499	bool queue_work_on(int cpu, struct workqueue_struct *wq,
1500	struct work_struct *work)
1501	{
1502	bool ret = false;
1503	unsigned long flags;
1504
1505	local_irq_save(flags);
1506
1507	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1508	__queue_work(cpu, wq, work);
1509	ret = true;
1510	}
1511
1512	local_irq_restore(flags);
1513	return ret;
1514	}
1515	EXPORT_SYMBOL(queue_work_on);
1516
1517	void delayed_work_timer_fn(unsigned long __data)
1518	{
1519	struct delayed_work *dwork = (struct delayed_work *)__data;
1520
1521	/* should have been called from irqsafe timer with irq already off */
1522	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
1523	}
1524	EXPORT_SYMBOL(delayed_work_timer_fn);
1525
1526	static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1527	struct delayed_work *dwork, unsigned long delay)
1528	{
1529	struct timer_list *timer = &dwork->timer;
1530	struct work_struct *work = &dwork->work;
1531
1532	WARN_ON_ONCE(!wq);
1533	WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1534	timer->data != (unsigned long)dwork);
1535	WARN_ON_ONCE(timer_pending(timer));
1536	WARN_ON_ONCE(!list_empty(&work->entry));
1537
1538	/*
1539	* If @delay is 0, queue @dwork->work immediately. This is for
1540	* both optimization and correctness. The earliest @timer can
1541	* expire is on the closest next tick and delayed_work users depend
1542	* on that there's no such delay when @delay is 0.
1543	*/
1544	if (!delay) {
1545	__queue_work(cpu, wq, &dwork->work);
1546	return;
1547	}
1548
1549	timer_stats_timer_set_start_info(&dwork->timer);
1550
1551	dwork->wq = wq;
1552	dwork->cpu = cpu;
1553	timer->expires = jiffies + delay;
1554
1555	if (unlikely(cpu != WORK_CPU_UNBOUND))
1556	add_timer_on(timer, cpu);
1557	else
1558	add_timer(timer);
1559	}
1560
1561	/**
1562	* queue_delayed_work_on - queue work on specific CPU after delay
1563	* @cpu: CPU number to execute work on
1564	* @wq: workqueue to use
1565	* @dwork: work to queue
1566	* @delay: number of jiffies to wait before queueing
1567	*
1568	* Return: %false if @work was already on a queue, %true otherwise. If
1569	* @delay is zero and @dwork is idle, it will be scheduled for immediate
1570	* execution.
1571	*/
1572	bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1573	struct delayed_work *dwork, unsigned long delay)
1574	{
1575	struct work_struct *work = &dwork->work;
1576	bool ret = false;
1577	unsigned long flags;
1578
1579	/* read the comment in __queue_work() */
1580	local_irq_save(flags);
1581
1582	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1583	__queue_delayed_work(cpu, wq, dwork, delay);
1584	ret = true;
1585	}
1586
1587	local_irq_restore(flags);
1588	return ret;
1589	}
1590	EXPORT_SYMBOL(queue_delayed_work_on);
1591
1592	/**
1593	* mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1594	* @cpu: CPU number to execute work on
1595	* @wq: workqueue to use
1596	* @dwork: work to queue
1597	* @delay: number of jiffies to wait before queueing
1598	*
1599	* If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1600	* modify @dwork's timer so that it expires after @delay. If @delay is
1601	* zero, @work is guaranteed to be scheduled immediately regardless of its
1602	* current state.
1603	*
1604	* Return: %false if @dwork was idle and queued, %true if @dwork was
1605	* pending and its timer was modified.
1606	*
1607	* This function is safe to call from any context including IRQ handler.
1608	* See try_to_grab_pending() for details.
1609	*/
1610	bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1611	struct delayed_work *dwork, unsigned long delay)
1612	{
1613	unsigned long flags;
1614	int ret;
1615
1616	do {
1617	ret = try_to_grab_pending(&dwork->work, true, &flags);
1618	} while (unlikely(ret == -EAGAIN));
1619
1620	if (likely(ret >= 0)) {
1621	__queue_delayed_work(cpu, wq, dwork, delay);
1622	local_irq_restore(flags);
1623	}
1624
1625	/* -ENOENT from try_to_grab_pending() becomes %true */
1626	return ret;
1627	}
1628	EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1629
1630	/**
1631	* worker_enter_idle - enter idle state
1632	* @worker: worker which is entering idle state
1633	*
1634	* @worker is entering idle state. Update stats and idle timer if
1635	* necessary.
1636	*
1637	* LOCKING:
1638	* spin_lock_irq(pool->lock).
1639	*/
1640	static void worker_enter_idle(struct worker *worker)
1641	{
1642	struct worker_pool *pool = worker->pool;
1643
1644	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1645	WARN_ON_ONCE(!list_empty(&worker->entry) &&
1646	(worker->hentry.next || worker->hentry.pprev)))
1647	return;
1648
1649	/* can't use worker_set_flags(), also called from create_worker() */
1650	worker->flags |= WORKER_IDLE;
1651	pool->nr_idle++;
1652	worker->last_active = jiffies;
1653
1654	/* idle_list is LIFO */
1655	list_add(&worker->entry, &pool->idle_list);
1656
1657	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1658	mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1659
1660	/*
1661	* Sanity check nr_running. Because wq_unbind_fn() releases
1662	* pool->lock between setting %WORKER_UNBOUND and zapping
1663	* nr_running, the warning may trigger spuriously. Check iff
1664	* unbind is not in progress.
1665	*/
1666	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1667	pool->nr_workers == pool->nr_idle &&
1668	atomic_read(&pool->nr_running));
1669	}
1670
1671	/**
1672	* worker_leave_idle - leave idle state
1673	* @worker: worker which is leaving idle state
1674	*
1675	* @worker is leaving idle state. Update stats.
1676	*
1677	* LOCKING:
1678	* spin_lock_irq(pool->lock).
1679	*/
1680	static void worker_leave_idle(struct worker *worker)
1681	{
1682	struct worker_pool *pool = worker->pool;
1683
1684	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1685	return;
1686	worker_clr_flags(worker, WORKER_IDLE);
1687	pool->nr_idle--;
1688	list_del_init(&worker->entry);
1689	}
1690
1691	static struct worker *alloc_worker(int node)
1692	{
1693	struct worker *worker;
1694
1695	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1696	if (worker) {
1697	INIT_LIST_HEAD(&worker->entry);
1698	INIT_LIST_HEAD(&worker->scheduled);
1699	INIT_LIST_HEAD(&worker->node);
1700	/* on creation a worker is in !idle && prep state */
1701	worker->flags = WORKER_PREP;
1702	}
1703	return worker;
1704	}
1705
1706	/**
1707	* worker_attach_to_pool() - attach a worker to a pool
1708	* @worker: worker to be attached
1709	* @pool: the target pool
1710	*
1711	* Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1712	* cpu-binding of @worker are kept coordinated with the pool across
1713	* cpu-[un]hotplugs.
1714	*/
1715	static void worker_attach_to_pool(struct worker *worker,
1716	struct worker_pool *pool)
1717	{
1718	mutex_lock(&pool->attach_mutex);
1719
1720	/*
1721	* set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1722	* online CPUs. It'll be re-applied when any of the CPUs come up.
1723	*/
1724	set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1725
1726	/*
1727	* The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1728	* stable across this function. See the comments above the
1729	* flag definition for details.
1730	*/
1731	if (pool->flags & POOL_DISASSOCIATED)
1732	worker->flags |= WORKER_UNBOUND;
1733
1734	list_add_tail(&worker->node, &pool->workers);
1735
1736	mutex_unlock(&pool->attach_mutex);
1737	}
1738
1739	/**
1740	* worker_detach_from_pool() - detach a worker from its pool
1741	* @worker: worker which is attached to its pool
1742	* @pool: the pool @worker is attached to
1743	*
1744	* Undo the attaching which had been done in worker_attach_to_pool(). The
1745	* caller worker shouldn't access to the pool after detached except it has
1746	* other reference to the pool.
1747	*/
1748	static void worker_detach_from_pool(struct worker *worker,
1749	struct worker_pool *pool)
1750	{
1751	struct completion *detach_completion = NULL;
1752
1753	mutex_lock(&pool->attach_mutex);
1754	list_del(&worker->node);
1755	if (list_empty(&pool->workers))
1756	detach_completion = pool->detach_completion;
1757	mutex_unlock(&pool->attach_mutex);
1758
1759	/* clear leftover flags without pool->lock after it is detached */
1760	worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1761
1762	if (detach_completion)
1763	complete(detach_completion);
1764	}
1765
1766	/**
1767	* create_worker - create a new workqueue worker
1768	* @pool: pool the new worker will belong to
1769	*
1770	* Create and start a new worker which is attached to @pool.
1771	*
1772	* CONTEXT:
1773	* Might sleep. Does GFP_KERNEL allocations.
1774	*
1775	* Return:
1776	* Pointer to the newly created worker.
1777	*/
1778	static struct worker *create_worker(struct worker_pool *pool)
1779	{
1780	struct worker *worker = NULL;
1781	int id = -1;
1782	char id_buf[16];
1783
1784	/* ID is needed to determine kthread name */
1785	id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1786	if (id < 0)
1787	goto fail;
1788
1789	worker = alloc_worker(pool->node);
1790	if (!worker)
1791	goto fail;
1792
1793	worker->pool = pool;
1794	worker->id = id;
1795
1796	if (pool->cpu >= 0)
1797	snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1798	pool->attrs->nice < 0 ? "H" : "");
1799	else
1800	snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1801
1802	worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1803	"kworker/%s", id_buf);
1804	if (IS_ERR(worker->task))
1805	goto fail;
1806
1807	set_user_nice(worker->task, pool->attrs->nice);
1808	kthread_bind_mask(worker->task, pool->attrs->cpumask);
1809
1810	/* successful, attach the worker to the pool */
1811	worker_attach_to_pool(worker, pool);
1812
1813	/* start the newly created worker */
1814	spin_lock_irq(&pool->lock);
1815	worker->pool->nr_workers++;
1816	worker_enter_idle(worker);
1817	wake_up_process(worker->task);
1818	spin_unlock_irq(&pool->lock);
1819
1820	return worker;
1821
1822	fail:
1823	if (id >= 0)
1824	ida_simple_remove(&pool->worker_ida, id);
1825	kfree(worker);
1826	return NULL;
1827	}
1828
1829	/**
1830	* destroy_worker - destroy a workqueue worker
1831	* @worker: worker to be destroyed
1832	*
1833	* Destroy @worker and adjust @pool stats accordingly. The worker should
1834	* be idle.
1835	*
1836	* CONTEXT:
1837	* spin_lock_irq(pool->lock).
1838	*/
1839	static void destroy_worker(struct worker *worker)
1840	{
1841	struct worker_pool *pool = worker->pool;
1842
1843	lockdep_assert_held(&pool->lock);
1844
1845	/* sanity check frenzy */
1846	if (WARN_ON(worker->current_work) ||
1847	WARN_ON(!list_empty(&worker->scheduled)) ||
1848	WARN_ON(!(worker->flags & WORKER_IDLE)))
1849	return;
1850
1851	pool->nr_workers--;
1852	pool->nr_idle--;
1853
1854	list_del_init(&worker->entry);
1855	worker->flags |= WORKER_DIE;
1856	wake_up_process(worker->task);
1857	}
1858
1859	static void idle_worker_timeout(unsigned long __pool)
1860	{
1861	struct worker_pool *pool = (void *)__pool;
1862
1863	spin_lock_irq(&pool->lock);
1864
1865	while (too_many_workers(pool)) {
1866	struct worker *worker;
1867	unsigned long expires;
1868
1869	/* idle_list is kept in LIFO order, check the last one */
1870	worker = list_entry(pool->idle_list.prev, struct worker, entry);
1871	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1872
1873	if (time_before(jiffies, expires)) {
1874	mod_timer(&pool->idle_timer, expires);
1875	break;
1876	}
1877
1878	destroy_worker(worker);
1879	}
1880
1881	spin_unlock_irq(&pool->lock);
1882	}
1883
1884	static void send_mayday(struct work_struct *work)
1885	{
1886	struct pool_workqueue *pwq = get_work_pwq(work);
1887	struct workqueue_struct *wq = pwq->wq;
1888
1889	lockdep_assert_held(&wq_mayday_lock);
1890
1891	if (!wq->rescuer)
1892	return;
1893
1894	/* mayday mayday mayday */
1895	if (list_empty(&pwq->mayday_node)) {
1896	/*
1897	* If @pwq is for an unbound wq, its base ref may be put at
1898	* any time due to an attribute change. Pin @pwq until the
1899	* rescuer is done with it.
1900	*/
1901	get_pwq(pwq);
1902	list_add_tail(&pwq->mayday_node, &wq->maydays);
1903	wake_up_process(wq->rescuer->task);
1904	}
1905	}
1906
1907	static void pool_mayday_timeout(unsigned long __pool)
1908	{
1909	struct worker_pool *pool = (void *)__pool;
1910	struct work_struct *work;
1911
1912	spin_lock_irq(&pool->lock);
1913	spin_lock(&wq_mayday_lock); /* for wq->maydays */
1914
1915	if (need_to_create_worker(pool)) {
1916	/*
1917	* We've been trying to create a new worker but
1918	* haven't been successful. We might be hitting an
1919	* allocation deadlock. Send distress signals to
1920	* rescuers.
1921	*/
1922	list_for_each_entry(work, &pool->worklist, entry)
1923	send_mayday(work);
1924	}
1925
1926	spin_unlock(&wq_mayday_lock);
1927	spin_unlock_irq(&pool->lock);
1928
1929	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1930	}
1931
1932	/**
1933	* maybe_create_worker - create a new worker if necessary
1934	* @pool: pool to create a new worker for
1935	*
1936	* Create a new worker for @pool if necessary. @pool is guaranteed to
1937	* have at least one idle worker on return from this function. If
1938	* creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1939	* sent to all rescuers with works scheduled on @pool to resolve
1940	* possible allocation deadlock.
1941	*
1942	* On return, need_to_create_worker() is guaranteed to be %false and
1943	* may_start_working() %true.
1944	*
1945	* LOCKING:
1946	* spin_lock_irq(pool->lock) which may be released and regrabbed
1947	* multiple times. Does GFP_KERNEL allocations. Called only from
1948	* manager.
1949	*/
1950	static void maybe_create_worker(struct worker_pool *pool)
1951	__releases(&pool->lock)
1952	__acquires(&pool->lock)
1953	{
1954	restart:
1955	spin_unlock_irq(&pool->lock);
1956
1957	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1958	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1959
1960	while (true) {
1961	if (create_worker(pool) || !need_to_create_worker(pool))
1962	break;
1963
1964	schedule_timeout_interruptible(CREATE_COOLDOWN);
1965
1966	if (!need_to_create_worker(pool))
1967	break;
1968	}
1969
1970	del_timer_sync(&pool->mayday_timer);
1971	spin_lock_irq(&pool->lock);
1972	/*
1973	* This is necessary even after a new worker was just successfully
1974	* created as @pool->lock was dropped and the new worker might have
1975	* already become busy.
1976	*/
1977	if (need_to_create_worker(pool))
1978	goto restart;
1979	}
1980
1981	/**
1982	* manage_workers - manage worker pool
1983	* @worker: self
1984	*
1985	* Assume the manager role and manage the worker pool @worker belongs
1986	* to. At any given time, there can be only zero or one manager per
1987	* pool. The exclusion is handled automatically by this function.
1988	*
1989	* The caller can safely start processing works on false return. On
1990	* true return, it's guaranteed that need_to_create_worker() is false
1991	* and may_start_working() is true.
1992	*
1993	* CONTEXT:
1994	* spin_lock_irq(pool->lock) which may be released and regrabbed
1995	* multiple times. Does GFP_KERNEL allocations.
1996	*
1997	* Return:
1998	* %false if the pool doesn't need management and the caller can safely
1999	* start processing works, %true if management function was performed and
2000	* the conditions that the caller verified before calling the function may
2001	* no longer be true.
2002	*/
2003	static bool manage_workers(struct worker *worker)
2004	{
2005	struct worker_pool *pool = worker->pool;
2006
2007	if (pool->flags & POOL_MANAGER_ACTIVE)
2008	return false;
2009
2010	pool->flags |= POOL_MANAGER_ACTIVE;
2011	pool->manager = worker;
2012
2013	maybe_create_worker(pool);
2014
2015	pool->manager = NULL;
2016	pool->flags &= ~POOL_MANAGER_ACTIVE;
2017	wake_up(&wq_manager_wait);
2018	return true;
2019	}
2020
2021	/**
2022	* process_one_work - process single work
2023	* @worker: self
2024	* @work: work to process
2025	*
2026	* Process @work. This function contains all the logics necessary to
2027	* process a single work including synchronization against and
2028	* interaction with other workers on the same cpu, queueing and
2029	* flushing. As long as context requirement is met, any worker can
2030	* call this function to process a work.
2031	*
2032	* CONTEXT:
2033	* spin_lock_irq(pool->lock) which is released and regrabbed.
2034	*/
2035	static void process_one_work(struct worker *worker, struct work_struct *work)
2036	__releases(&pool->lock)
2037	__acquires(&pool->lock)
2038	{
2039	#ifdef CONFIG_AMLOGIC_MODIFY
2040	int work_color;
2041	struct worker *collision;
2042	bool cpu_intensive;
2043	struct pool_workqueue *pwq = get_work_pwq(work);
2044	struct worker_pool *pool = worker->pool;
2045	#else
2046	struct pool_workqueue *pwq = get_work_pwq(work);
2047	struct worker_pool *pool = worker->pool;
2048	bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2049	int work_color;
2050	struct worker *collision;
2051	#endif
2052
2053	#ifdef CONFIG_LOCKDEP
2054	/*
2055	* It is permissible to free the struct work_struct from
2056	* inside the function that is called from it, this we need to
2057	* take into account for lockdep too. To avoid bogus "held
2058	* lock freed" warnings as well as problems when looking into
2059	* work->lockdep_map, make a copy and use that here.
2060	*/
2061	struct lockdep_map lockdep_map;
2062
2063	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2064	#endif
2065	#ifdef CONFIG_AMLOGIC_MODIFY
2066	if (!pwq) {
2067	WARN_ONCE(1, "<%s> pwq_NULL <%lx> <%pf>, <%pf> %s\n",
2068	__func__, atomic_long_read(&work->data),
2069	work->func, worker->current_func, worker->desc);
2070	return;
2071	}
2072
2073	cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2074	#endif
2075	/* ensure we're on the correct CPU */
2076	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2077	raw_smp_processor_id() != pool->cpu);
2078
2079	/*
2080	* A single work shouldn't be executed concurrently by
2081	* multiple workers on a single cpu. Check whether anyone is
2082	* already processing the work. If so, defer the work to the
2083	* currently executing one.
2084	*/
2085	collision = find_worker_executing_work(pool, work);
2086	if (unlikely(collision)) {
2087	move_linked_works(work, &collision->scheduled, NULL);
2088	return;
2089	}
2090
2091	/* claim and dequeue */
2092	debug_work_deactivate(work);
2093	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2094	worker->current_work = work;
2095	worker->current_func = work->func;
2096	worker->current_pwq = pwq;
2097	work_color = get_work_color(work);
2098
2099	list_del_init(&work->entry);
2100
2101	/*
2102	* CPU intensive works don't participate in concurrency management.
2103	* They're the scheduler's responsibility. This takes @worker out
2104	* of concurrency management and the next code block will chain
2105	* execution of the pending work items.
2106	*/
2107	if (unlikely(cpu_intensive))
2108	worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2109
2110	/*
2111	* Wake up another worker if necessary. The condition is always
2112	* false for normal per-cpu workers since nr_running would always
2113	* be >= 1 at this point. This is used to chain execution of the
2114	* pending work items for WORKER_NOT_RUNNING workers such as the
2115	* UNBOUND and CPU_INTENSIVE ones.
2116	*/
2117	if (need_more_worker(pool))
2118	wake_up_worker(pool);
2119
2120	/*
2121	* Record the last pool and clear PENDING which should be the last
2122	* update to @work. Also, do this inside @pool->lock so that
2123	* PENDING and queued state changes happen together while IRQ is
2124	* disabled.
2125	*/
2126	set_work_pool_and_clear_pending(work, pool->id);
2127
2128	spin_unlock_irq(&pool->lock);
2129
2130	lock_map_acquire_read(&pwq->wq->lockdep_map);
2131	lock_map_acquire(&lockdep_map);
2132	trace_workqueue_execute_start(work);
2133	worker->current_func(work);
2134	/*
2135	* While we must be careful to not use "work" after this, the trace
2136	* point will only record its address.
2137	*/
2138	trace_workqueue_execute_end(work);
2139	lock_map_release(&lockdep_map);
2140	lock_map_release(&pwq->wq->lockdep_map);
2141
2142	if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2143	pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2144	" last function: %pf\n",
2145	current->comm, preempt_count(), task_pid_nr(current),
2146	worker->current_func);
2147	debug_show_held_locks(current);
2148	dump_stack();
2149	}
2150
2151	/*
2152	* The following prevents a kworker from hogging CPU on !PREEMPT
2153	* kernels, where a requeueing work item waiting for something to
2154	* happen could deadlock with stop_machine as such work item could
2155	* indefinitely requeue itself while all other CPUs are trapped in
2156	* stop_machine. At the same time, report a quiescent RCU state so
2157	* the same condition doesn't freeze RCU.
2158	*/
2159	cond_resched_rcu_qs();
2160
2161	spin_lock_irq(&pool->lock);
2162
2163	/* clear cpu intensive status */
2164	if (unlikely(cpu_intensive))
2165	worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2166
2167	/* tag the worker for identification in schedule() */
2168	worker->last_func = worker->current_func;
2169
2170	/* we're done with it, release */
2171	hash_del(&worker->hentry);
2172	worker->current_work = NULL;
2173	worker->current_func = NULL;
2174	worker->current_pwq = NULL;
2175	worker->desc_valid = false;
2176	pwq_dec_nr_in_flight(pwq, work_color);
2177	}
2178
2179	/**
2180	* process_scheduled_works - process scheduled works
2181	* @worker: self
2182	*
2183	* Process all scheduled works. Please note that the scheduled list
2184	* may change while processing a work, so this function repeatedly
2185	* fetches a work from the top and executes it.
2186	*
2187	* CONTEXT:
2188	* spin_lock_irq(pool->lock) which may be released and regrabbed
2189	* multiple times.
2190	*/
2191	static void process_scheduled_works(struct worker *worker)
2192	{
2193	while (!list_empty(&worker->scheduled)) {
2194	struct work_struct *work = list_first_entry(&worker->scheduled,
2195	struct work_struct, entry);
2196	process_one_work(worker, work);
2197	}
2198	}
2199
2200	/**
2201	* worker_thread - the worker thread function
2202	* @__worker: self
2203	*
2204	* The worker thread function. All workers belong to a worker_pool -
2205	* either a per-cpu one or dynamic unbound one. These workers process all
2206	* work items regardless of their specific target workqueue. The only
2207	* exception is work items which belong to workqueues with a rescuer which
2208	* will be explained in rescuer_thread().
2209	*
2210	* Return: 0
2211	*/
2212	static int worker_thread(void *__worker)
2213	{
2214	struct worker *worker = __worker;
2215	struct worker_pool *pool = worker->pool;
2216
2217	/* tell the scheduler that this is a workqueue worker */
2218	worker->task->flags |= PF_WQ_WORKER;
2219	woke_up:
2220	spin_lock_irq(&pool->lock);
2221
2222	/* am I supposed to die? */
2223	if (unlikely(worker->flags & WORKER_DIE)) {
2224	spin_unlock_irq(&pool->lock);
2225	WARN_ON_ONCE(!list_empty(&worker->entry));
2226	worker->task->flags &= ~PF_WQ_WORKER;
2227
2228	set_task_comm(worker->task, "kworker/dying");
2229	ida_simple_remove(&pool->worker_ida, worker->id);
2230	worker_detach_from_pool(worker, pool);
2231	kfree(worker);
2232	return 0;
2233	}
2234
2235	worker_leave_idle(worker);
2236	recheck:
2237	/* no more worker necessary? */
2238	if (!need_more_worker(pool))
2239	goto sleep;
2240
2241	/* do we need to manage? */
2242	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2243	goto recheck;
2244
2245	/*
2246	* ->scheduled list can only be filled while a worker is
2247	* preparing to process a work or actually processing it.
2248	* Make sure nobody diddled with it while I was sleeping.
2249	*/
2250	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2251
2252	/*
2253	* Finish PREP stage. We're guaranteed to have at least one idle
2254	* worker or that someone else has already assumed the manager
2255	* role. This is where @worker starts participating in concurrency
2256	* management if applicable and concurrency management is restored
2257	* after being rebound. See rebind_workers() for details.
2258	*/
2259	worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2260
2261	do {
2262	struct work_struct *work =
2263	list_first_entry(&pool->worklist,
2264	struct work_struct, entry);
2265
2266	pool->watchdog_ts = jiffies;
2267
2268	if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2269	/* optimization path, not strictly necessary */
2270	process_one_work(worker, work);
2271	if (unlikely(!list_empty(&worker->scheduled)))
2272	process_scheduled_works(worker);
2273	} else {
2274	move_linked_works(work, &worker->scheduled, NULL);
2275	process_scheduled_works(worker);
2276	}
2277	} while (keep_working(pool));
2278
2279	worker_set_flags(worker, WORKER_PREP);
2280	sleep:
2281	/*
2282	* pool->lock is held and there's no work to process and no need to
2283	* manage, sleep. Workers are woken up only while holding
2284	* pool->lock or from local cpu, so setting the current state
2285	* before releasing pool->lock is enough to prevent losing any
2286	* event.
2287	*/
2288	worker_enter_idle(worker);
2289	__set_current_state(TASK_INTERRUPTIBLE);
2290	spin_unlock_irq(&pool->lock);
2291	schedule();
2292	goto woke_up;
2293	}
2294
2295	/**
2296	* rescuer_thread - the rescuer thread function
2297	* @__rescuer: self
2298	*
2299	* Workqueue rescuer thread function. There's one rescuer for each
2300	* workqueue which has WQ_MEM_RECLAIM set.
2301	*
2302	* Regular work processing on a pool may block trying to create a new
2303	* worker which uses GFP_KERNEL allocation which has slight chance of
2304	* developing into deadlock if some works currently on the same queue
2305	* need to be processed to satisfy the GFP_KERNEL allocation. This is
2306	* the problem rescuer solves.
2307	*
2308	* When such condition is possible, the pool summons rescuers of all
2309	* workqueues which have works queued on the pool and let them process
2310	* those works so that forward progress can be guaranteed.
2311	*
2312	* This should happen rarely.
2313	*
2314	* Return: 0
2315	*/
2316	static int rescuer_thread(void *__rescuer)
2317	{
2318	struct worker *rescuer = __rescuer;
2319	struct workqueue_struct *wq = rescuer->rescue_wq;
2320	struct list_head *scheduled = &rescuer->scheduled;
2321	bool should_stop;
2322
2323	set_user_nice(current, RESCUER_NICE_LEVEL);
2324
2325	/*
2326	* Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2327	* doesn't participate in concurrency management.
2328	*/
2329	rescuer->task->flags |= PF_WQ_WORKER;
2330	repeat:
2331	set_current_state(TASK_INTERRUPTIBLE);
2332
2333	/*
2334	* By the time the rescuer is requested to stop, the workqueue
2335	* shouldn't have any work pending, but @wq->maydays may still have
2336	* pwq(s) queued. This can happen by non-rescuer workers consuming
2337	* all the work items before the rescuer got to them. Go through
2338	* @wq->maydays processing before acting on should_stop so that the
2339	* list is always empty on exit.
2340	*/
2341	should_stop = kthread_should_stop();
2342
2343	/* see whether any pwq is asking for help */
2344	spin_lock_irq(&wq_mayday_lock);
2345
2346	while (!list_empty(&wq->maydays)) {
2347	struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2348	struct pool_workqueue, mayday_node);
2349	struct worker_pool *pool = pwq->pool;
2350	struct work_struct *work, *n;
2351	bool first = true;
2352
2353	__set_current_state(TASK_RUNNING);
2354	list_del_init(&pwq->mayday_node);
2355
2356	spin_unlock_irq(&wq_mayday_lock);
2357
2358	worker_attach_to_pool(rescuer, pool);
2359
2360	spin_lock_irq(&pool->lock);
2361	rescuer->pool = pool;
2362
2363	/*
2364	* Slurp in all works issued via this workqueue and
2365	* process'em.
2366	*/
2367	WARN_ON_ONCE(!list_empty(scheduled));
2368	list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2369	if (get_work_pwq(work) == pwq) {
2370	if (first)
2371	pool->watchdog_ts = jiffies;
2372	move_linked_works(work, scheduled, &n);
2373	}
2374	first = false;
2375	}
2376
2377	if (!list_empty(scheduled)) {
2378	process_scheduled_works(rescuer);
2379
2380	/*
2381	* The above execution of rescued work items could
2382	* have created more to rescue through
2383	* pwq_activate_first_delayed() or chained
2384	* queueing. Let's put @pwq back on mayday list so
2385	* that such back-to-back work items, which may be
2386	* being used to relieve memory pressure, don't
2387	* incur MAYDAY_INTERVAL delay inbetween.
2388	*/
2389	if (need_to_create_worker(pool)) {
2390	spin_lock(&wq_mayday_lock);
2391	get_pwq(pwq);
2392	list_move_tail(&pwq->mayday_node, &wq->maydays);
2393	spin_unlock(&wq_mayday_lock);
2394	}
2395	}
2396
2397	/*
2398	* Put the reference grabbed by send_mayday(). @pool won't
2399	* go away while we're still attached to it.
2400	*/
2401	put_pwq(pwq);
2402
2403	/*
2404	* Leave this pool. If need_more_worker() is %true, notify a
2405	* regular worker; otherwise, we end up with 0 concurrency
2406	* and stalling the execution.
2407	*/
2408	if (need_more_worker(pool))
2409	wake_up_worker(pool);
2410
2411	rescuer->pool = NULL;
2412	spin_unlock_irq(&pool->lock);
2413
2414	worker_detach_from_pool(rescuer, pool);
2415
2416	spin_lock_irq(&wq_mayday_lock);
2417	}
2418
2419	spin_unlock_irq(&wq_mayday_lock);
2420
2421	if (should_stop) {
2422	__set_current_state(TASK_RUNNING);
2423	rescuer->task->flags &= ~PF_WQ_WORKER;
2424	return 0;
2425	}
2426
2427	/* rescuers should never participate in concurrency management */
2428	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2429	schedule();
2430	goto repeat;
2431	}
2432
2433	/**
2434	* check_flush_dependency - check for flush dependency sanity
2435	* @target_wq: workqueue being flushed
2436	* @target_work: work item being flushed (NULL for workqueue flushes)
2437	*
2438	* %current is trying to flush the whole @target_wq or @target_work on it.
2439	* If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2440	* reclaiming memory or running on a workqueue which doesn't have
2441	* %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2442	* a deadlock.
2443	*/
2444	static void check_flush_dependency(struct workqueue_struct *target_wq,
2445	struct work_struct *target_work)
2446	{
2447	work_func_t target_func = target_work ? target_work->func : NULL;
2448	struct worker *worker;
2449
2450	if (target_wq->flags & WQ_MEM_RECLAIM)
2451	return;
2452
2453	worker = current_wq_worker();
2454
2455	WARN_ONCE(current->flags & PF_MEMALLOC,
2456	"workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2457	current->pid, current->comm, target_wq->name, target_func);
2458	WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2459	(WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2460	"workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2461	worker->current_pwq->wq->name, worker->current_func,
2462	target_wq->name, target_func);
2463	}
2464
2465	struct wq_barrier {
2466	struct work_struct work;
2467	struct completion done;
2468	struct task_struct *task; /* purely informational */
2469	};
2470
2471	static void wq_barrier_func(struct work_struct *work)
2472	{
2473	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2474	complete(&barr->done);
2475	}
2476
2477	/**
2478	* insert_wq_barrier - insert a barrier work
2479	* @pwq: pwq to insert barrier into
2480	* @barr: wq_barrier to insert
2481	* @target: target work to attach @barr to
2482	* @worker: worker currently executing @target, NULL if @target is not executing
2483	*
2484	* @barr is linked to @target such that @barr is completed only after
2485	* @target finishes execution. Please note that the ordering
2486	* guarantee is observed only with respect to @target and on the local
2487	* cpu.
2488	*
2489	* Currently, a queued barrier can't be canceled. This is because
2490	* try_to_grab_pending() can't determine whether the work to be
2491	* grabbed is at the head of the queue and thus can't clear LINKED
2492	* flag of the previous work while there must be a valid next work
2493	* after a work with LINKED flag set.
2494	*
2495	* Note that when @worker is non-NULL, @target may be modified
2496	* underneath us, so we can't reliably determine pwq from @target.
2497	*
2498	* CONTEXT:
2499	* spin_lock_irq(pool->lock).
2500	*/
2501	static void insert_wq_barrier(struct pool_workqueue *pwq,
2502	struct wq_barrier *barr,
2503	struct work_struct *target, struct worker *worker)
2504	{
2505	struct list_head *head;
2506	unsigned int linked = 0;
2507
2508	/*
2509	* debugobject calls are safe here even with pool->lock locked
2510	* as we know for sure that this will not trigger any of the
2511	* checks and call back into the fixup functions where we
2512	* might deadlock.
2513	*/
2514	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2515	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2516	init_completion(&barr->done);
2517	barr->task = current;
2518
2519	/*
2520	* If @target is currently being executed, schedule the
2521	* barrier to the worker; otherwise, put it after @target.
2522	*/
2523	if (worker)
2524	head = worker->scheduled.next;
2525	else {
2526	unsigned long *bits = work_data_bits(target);
2527
2528	head = target->entry.next;
2529	/* there can already be other linked works, inherit and set */
2530	linked = *bits & WORK_STRUCT_LINKED;
2531	__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2532	}
2533
2534	debug_work_activate(&barr->work);
2535	insert_work(pwq, &barr->work, head,
2536	work_color_to_flags(WORK_NO_COLOR) | linked);
2537	}
2538
2539	/**
2540	* flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2541	* @wq: workqueue being flushed
2542	* @flush_color: new flush color, < 0 for no-op
2543	* @work_color: new work color, < 0 for no-op
2544	*
2545	* Prepare pwqs for workqueue flushing.
2546	*
2547	* If @flush_color is non-negative, flush_color on all pwqs should be
2548	* -1. If no pwq has in-flight commands at the specified color, all
2549	* pwq->flush_color's stay at -1 and %false is returned. If any pwq
2550	* has in flight commands, its pwq->flush_color is set to
2551	* @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2552	* wakeup logic is armed and %true is returned.
2553	*
2554	* The caller should have initialized @wq->first_flusher prior to
2555	* calling this function with non-negative @flush_color. If
2556	* @flush_color is negative, no flush color update is done and %false
2557	* is returned.
2558	*
2559	* If @work_color is non-negative, all pwqs should have the same
2560	* work_color which is previous to @work_color and all will be
2561	* advanced to @work_color.
2562	*
2563	* CONTEXT:
2564	* mutex_lock(wq->mutex).
2565	*
2566	* Return:
2567	* %true if @flush_color >= 0 and there's something to flush. %false
2568	* otherwise.
2569	*/
2570	static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2571	int flush_color, int work_color)
2572	{
2573	bool wait = false;
2574	struct pool_workqueue *pwq;
2575
2576	if (flush_color >= 0) {
2577	WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2578	atomic_set(&wq->nr_pwqs_to_flush, 1);
2579	}
2580
2581	for_each_pwq(pwq, wq) {
2582	struct worker_pool *pool = pwq->pool;
2583
2584	spin_lock_irq(&pool->lock);
2585
2586	if (flush_color >= 0) {
2587	WARN_ON_ONCE(pwq->flush_color != -1);
2588
2589	if (pwq->nr_in_flight[flush_color]) {
2590	pwq->flush_color = flush_color;
2591	atomic_inc(&wq->nr_pwqs_to_flush);
2592	wait = true;
2593	}
2594	}
2595
2596	if (work_color >= 0) {
2597	WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2598	pwq->work_color = work_color;
2599	}
2600
2601	spin_unlock_irq(&pool->lock);
2602	}
2603
2604	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2605	complete(&wq->first_flusher->done);
2606
2607	return wait;
2608	}
2609
2610	/**
2611	* flush_workqueue - ensure that any scheduled work has run to completion.
2612	* @wq: workqueue to flush
2613	*
2614	* This function sleeps until all work items which were queued on entry
2615	* have finished execution, but it is not livelocked by new incoming ones.
2616	*/
2617	void flush_workqueue(struct workqueue_struct *wq)
2618	{
2619	struct wq_flusher this_flusher = {
2620	.list = LIST_HEAD_INIT(this_flusher.list),
2621	.flush_color = -1,
2622	.done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2623	};
2624	int next_color;
2625
2626	if (WARN_ON(!wq_online))
2627	return;
2628
2629	lock_map_acquire(&wq->lockdep_map);
2630	lock_map_release(&wq->lockdep_map);
2631
2632	mutex_lock(&wq->mutex);
2633
2634	/*
2635	* Start-to-wait phase
2636	*/
2637	next_color = work_next_color(wq->work_color);
2638
2639	if (next_color != wq->flush_color) {
2640	/*
2641	* Color space is not full. The current work_color
2642	* becomes our flush_color and work_color is advanced
2643	* by one.
2644	*/
2645	WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2646	this_flusher.flush_color = wq->work_color;
2647	wq->work_color = next_color;
2648
2649	if (!wq->first_flusher) {
2650	/* no flush in progress, become the first flusher */
2651	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2652
2653	wq->first_flusher = &this_flusher;
2654
2655	if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2656	wq->work_color)) {
2657	/* nothing to flush, done */
2658	wq->flush_color = next_color;
2659	wq->first_flusher = NULL;
2660	goto out_unlock;
2661	}
2662	} else {
2663	/* wait in queue */
2664	WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2665	list_add_tail(&this_flusher.list, &wq->flusher_queue);
2666	flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2667	}
2668	} else {
2669	/*
2670	* Oops, color space is full, wait on overflow queue.
2671	* The next flush completion will assign us
2672	* flush_color and transfer to flusher_queue.
2673	*/
2674	list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2675	}
2676
2677	check_flush_dependency(wq, NULL);
2678
2679	mutex_unlock(&wq->mutex);
2680
2681	wait_for_completion(&this_flusher.done);
2682
2683	/*
2684	* Wake-up-and-cascade phase
2685	*
2686	* First flushers are responsible for cascading flushes and
2687	* handling overflow. Non-first flushers can simply return.
2688	*/
2689	if (wq->first_flusher != &this_flusher)
2690	return;
2691
2692	mutex_lock(&wq->mutex);
2693
2694	/* we might have raced, check again with mutex held */
2695	if (wq->first_flusher != &this_flusher)
2696	goto out_unlock;
2697
2698	wq->first_flusher = NULL;
2699
2700	WARN_ON_ONCE(!list_empty(&this_flusher.list));
2701	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2702
2703	while (true) {
2704	struct wq_flusher *next, *tmp;
2705
2706	/* complete all the flushers sharing the current flush color */
2707	list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2708	if (next->flush_color != wq->flush_color)
2709	break;
2710	list_del_init(&next->list);
2711	complete(&next->done);
2712	}
2713
2714	WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2715	wq->flush_color != work_next_color(wq->work_color));
2716
2717	/* this flush_color is finished, advance by one */
2718	wq->flush_color = work_next_color(wq->flush_color);
2719
2720	/* one color has been freed, handle overflow queue */
2721	if (!list_empty(&wq->flusher_overflow)) {
2722	/*
2723	* Assign the same color to all overflowed
2724	* flushers, advance work_color and append to
2725	* flusher_queue. This is the start-to-wait
2726	* phase for these overflowed flushers.
2727	*/
2728	list_for_each_entry(tmp, &wq->flusher_overflow, list)
2729	tmp->flush_color = wq->work_color;
2730
2731	wq->work_color = work_next_color(wq->work_color);
2732
2733	list_splice_tail_init(&wq->flusher_overflow,
2734	&wq->flusher_queue);
2735	flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2736	}
2737
2738	if (list_empty(&wq->flusher_queue)) {
2739	WARN_ON_ONCE(wq->flush_color != wq->work_color);
2740	break;
2741	}
2742
2743	/*
2744	* Need to flush more colors. Make the next flusher
2745	* the new first flusher and arm pwqs.
2746	*/
2747	WARN_ON_ONCE(wq->flush_color == wq->work_color);
2748	WARN_ON_ONCE(wq->flush_color != next->flush_color);
2749
2750	list_del_init(&next->list);
2751	wq->first_flusher = next;
2752
2753	if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2754	break;
2755
2756	/*
2757	* Meh... this color is already done, clear first
2758	* flusher and repeat cascading.
2759	*/
2760	wq->first_flusher = NULL;
2761	}
2762
2763	out_unlock:
2764	mutex_unlock(&wq->mutex);
2765	}
2766	EXPORT_SYMBOL(flush_workqueue);
2767
2768	/**
2769	* drain_workqueue - drain a workqueue
2770	* @wq: workqueue to drain
2771	*
2772	* Wait until the workqueue becomes empty. While draining is in progress,
2773	* only chain queueing is allowed. IOW, only currently pending or running
2774	* work items on @wq can queue further work items on it. @wq is flushed
2775	* repeatedly until it becomes empty. The number of flushing is determined
2776	* by the depth of chaining and should be relatively short. Whine if it
2777	* takes too long.
2778	*/
2779	void drain_workqueue(struct workqueue_struct *wq)
2780	{
2781	unsigned int flush_cnt = 0;
2782	struct pool_workqueue *pwq;
2783
2784	/*
2785	* __queue_work() needs to test whether there are drainers, is much
2786	* hotter than drain_workqueue() and already looks at @wq->flags.
2787	* Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2788	*/
2789	mutex_lock(&wq->mutex);
2790	if (!wq->nr_drainers++)
2791	wq->flags |= __WQ_DRAINING;
2792	mutex_unlock(&wq->mutex);
2793	reflush:
2794	flush_workqueue(wq);
2795
2796	mutex_lock(&wq->mutex);
2797
2798	for_each_pwq(pwq, wq) {
2799	bool drained;
2800
2801	spin_lock_irq(&pwq->pool->lock);
2802	drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2803	spin_unlock_irq(&pwq->pool->lock);
2804
2805	if (drained)
2806	continue;
2807
2808	if (++flush_cnt == 10 ||
2809	(flush_cnt % 100 == 0 && flush_cnt <= 1000))
2810	pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2811	wq->name, flush_cnt);
2812
2813	mutex_unlock(&wq->mutex);
2814	goto reflush;
2815	}
2816
2817	if (!--wq->nr_drainers)
2818	wq->flags &= ~__WQ_DRAINING;
2819	mutex_unlock(&wq->mutex);
2820	}
2821	EXPORT_SYMBOL_GPL(drain_workqueue);
2822
2823	static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2824	{
2825	struct worker *worker = NULL;
2826	struct worker_pool *pool;
2827	struct pool_workqueue *pwq;
2828
2829	might_sleep();
2830
2831	local_irq_disable();
2832	pool = get_work_pool(work);
2833	if (!pool) {
2834	local_irq_enable();
2835	return false;
2836	}
2837
2838	spin_lock(&pool->lock);
2839	/* see the comment in try_to_grab_pending() with the same code */
2840	pwq = get_work_pwq(work);
2841	if (pwq) {
2842	if (unlikely(pwq->pool != pool))
2843	goto already_gone;
2844	} else {
2845	worker = find_worker_executing_work(pool, work);
2846	if (!worker)
2847	goto already_gone;
2848	pwq = worker->current_pwq;
2849	}
2850
2851	check_flush_dependency(pwq->wq, work);
2852
2853	insert_wq_barrier(pwq, barr, work, worker);
2854	spin_unlock_irq(&pool->lock);
2855
2856	/*
2857	* If @max_active is 1 or rescuer is in use, flushing another work
2858	* item on the same workqueue may lead to deadlock. Make sure the
2859	* flusher is not running on the same workqueue by verifying write
2860	* access.
2861	*/
2862	if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2863	lock_map_acquire(&pwq->wq->lockdep_map);
2864	else
2865	lock_map_acquire_read(&pwq->wq->lockdep_map);
2866	lock_map_release(&pwq->wq->lockdep_map);
2867
2868	return true;
2869	already_gone:
2870	spin_unlock_irq(&pool->lock);
2871	return false;
2872	}
2873
2874	/**
2875	* flush_work - wait for a work to finish executing the last queueing instance
2876	* @work: the work to flush
2877	*
2878	* Wait until @work has finished execution. @work is guaranteed to be idle
2879	* on return if it hasn't been requeued since flush started.
2880	*
2881	* Return:
2882	* %true if flush_work() waited for the work to finish execution,
2883	* %false if it was already idle.
2884	*/
2885	bool flush_work(struct work_struct *work)
2886	{
2887	struct wq_barrier barr;
2888
2889	if (WARN_ON(!wq_online))
2890	return false;
2891
2892	lock_map_acquire(&work->lockdep_map);
2893	lock_map_release(&work->lockdep_map);
2894
2895	if (start_flush_work(work, &barr)) {
2896	wait_for_completion(&barr.done);
2897	destroy_work_on_stack(&barr.work);
2898	return true;
2899	} else {
2900	return false;
2901	}
2902	}
2903	EXPORT_SYMBOL_GPL(flush_work);
2904
2905	struct cwt_wait {
2906	wait_queue_t wait;
2907	struct work_struct *work;
2908	};
2909
2910	static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2911	{
2912	struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2913
2914	if (cwait->work != key)
2915	return 0;
2916	return autoremove_wake_function(wait, mode, sync, key);
2917	}
2918
2919	static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2920	{
2921	static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2922	unsigned long flags;
2923	int ret;
2924
2925	do {
2926	ret = try_to_grab_pending(work, is_dwork, &flags);
2927	/*
2928	* If someone else is already canceling, wait for it to
2929	* finish. flush_work() doesn't work for PREEMPT_NONE
2930	* because we may get scheduled between @work's completion
2931	* and the other canceling task resuming and clearing
2932	* CANCELING - flush_work() will return false immediately
2933	* as @work is no longer busy, try_to_grab_pending() will
2934	* return -ENOENT as @work is still being canceled and the
2935	* other canceling task won't be able to clear CANCELING as
2936	* we're hogging the CPU.
2937	*
2938	* Let's wait for completion using a waitqueue. As this
2939	* may lead to the thundering herd problem, use a custom
2940	* wake function which matches @work along with exclusive
2941	* wait and wakeup.
2942	*/
2943	if (unlikely(ret == -ENOENT)) {
2944	struct cwt_wait cwait;
2945
2946	init_wait(&cwait.wait);
2947	cwait.wait.func = cwt_wakefn;
2948	cwait.work = work;
2949
2950	prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2951	TASK_UNINTERRUPTIBLE);
2952	if (work_is_canceling(work))
2953	schedule();
2954	finish_wait(&cancel_waitq, &cwait.wait);
2955	}
2956	} while (unlikely(ret < 0));
2957
2958	/* tell other tasks trying to grab @work to back off */
2959	mark_work_canceling(work);
2960	local_irq_restore(flags);
2961
2962	/*
2963	* This allows canceling during early boot. We know that @work
2964	* isn't executing.
2965	*/
2966	if (wq_online)
2967	flush_work(work);
2968
2969	clear_work_data(work);
2970
2971	/*
2972	* Paired with prepare_to_wait() above so that either
2973	* waitqueue_active() is visible here or !work_is_canceling() is
2974	* visible there.
2975	*/
2976	smp_mb();
2977	if (waitqueue_active(&cancel_waitq))
2978	__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2979
2980	return ret;
2981	}
2982
2983	/**
2984	* cancel_work_sync - cancel a work and wait for it to finish
2985	* @work: the work to cancel
2986	*
2987	* Cancel @work and wait for its execution to finish. This function
2988	* can be used even if the work re-queues itself or migrates to
2989	* another workqueue. On return from this function, @work is
2990	* guaranteed to be not pending or executing on any CPU.
2991	*
2992	* cancel_work_sync(&delayed_work->work) must not be used for
2993	* delayed_work's. Use cancel_delayed_work_sync() instead.
2994	*
2995	* The caller must ensure that the workqueue on which @work was last
2996	* queued can't be destroyed before this function returns.
2997	*
2998	* Return:
2999	* %true if @work was pending, %false otherwise.
3000	*/
3001	bool cancel_work_sync(struct work_struct *work)
3002	{
3003	return __cancel_work_timer(work, false);
3004	}
3005	EXPORT_SYMBOL_GPL(cancel_work_sync);
3006
3007	/**
3008	* flush_delayed_work - wait for a dwork to finish executing the last queueing
3009	* @dwork: the delayed work to flush
3010	*
3011	* Delayed timer is cancelled and the pending work is queued for
3012	* immediate execution. Like flush_work(), this function only
3013	* considers the last queueing instance of @dwork.
3014	*
3015	* Return:
3016	* %true if flush_work() waited for the work to finish execution,
3017	* %false if it was already idle.
3018	*/
3019	bool flush_delayed_work(struct delayed_work *dwork)
3020	{
3021	local_irq_disable();
3022	if (del_timer_sync(&dwork->timer))
3023	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
3024	local_irq_enable();
3025	return flush_work(&dwork->work);
3026	}
3027	EXPORT_SYMBOL(flush_delayed_work);
3028
3029	static bool __cancel_work(struct work_struct *work, bool is_dwork)
3030	{
3031	unsigned long flags;
3032	int ret;
3033
3034	do {
3035	ret = try_to_grab_pending(work, is_dwork, &flags);
3036	} while (unlikely(ret == -EAGAIN));
3037
3038	if (unlikely(ret < 0))
3039	return false;
3040
3041	set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3042	local_irq_restore(flags);
3043	return ret;
3044	}
3045
3046	/*
3047	* See cancel_delayed_work()
3048	*/
3049	bool cancel_work(struct work_struct *work)
3050	{
3051	return __cancel_work(work, false);
3052	}
3053
3054	/**
3055	* cancel_delayed_work - cancel a delayed work
3056	* @dwork: delayed_work to cancel
3057	*
3058	* Kill off a pending delayed_work.
3059	*
3060	* Return: %true if @dwork was pending and canceled; %false if it wasn't
3061	* pending.
3062	*
3063	* Note:
3064	* The work callback function may still be running on return, unless
3065	* it returns %true and the work doesn't re-arm itself. Explicitly flush or
3066	* use cancel_delayed_work_sync() to wait on it.
3067	*
3068	* This function is safe to call from any context including IRQ handler.
3069	*/
3070	bool cancel_delayed_work(struct delayed_work *dwork)
3071	{
3072	return __cancel_work(&dwork->work, true);
3073	}
3074	EXPORT_SYMBOL(cancel_delayed_work);
3075
3076	/**
3077	* cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3078	* @dwork: the delayed work cancel
3079	*
3080	* This is cancel_work_sync() for delayed works.
3081	*
3082	* Return:
3083	* %true if @dwork was pending, %false otherwise.
3084	*/
3085	bool cancel_delayed_work_sync(struct delayed_work *dwork)
3086	{
3087	return __cancel_work_timer(&dwork->work, true);
3088	}
3089	EXPORT_SYMBOL(cancel_delayed_work_sync);
3090
3091	/**
3092	* schedule_on_each_cpu - execute a function synchronously on each online CPU
3093	* @func: the function to call
3094	*
3095	* schedule_on_each_cpu() executes @func on each online CPU using the
3096	* system workqueue and blocks until all CPUs have completed.
3097	* schedule_on_each_cpu() is very slow.
3098	*
3099	* Return:
3100	* 0 on success, -errno on failure.
3101	*/
3102	int schedule_on_each_cpu(work_func_t func)
3103	{
3104	int cpu;
3105	struct work_struct __percpu *works;
3106
3107	works = alloc_percpu(struct work_struct);
3108	if (!works)
3109	return -ENOMEM;
3110
3111	get_online_cpus();
3112
3113	for_each_online_cpu(cpu) {
3114	struct work_struct *work = per_cpu_ptr(works, cpu);
3115
3116	INIT_WORK(work, func);
3117	schedule_work_on(cpu, work);
3118	}
3119
3120	for_each_online_cpu(cpu)
3121	flush_work(per_cpu_ptr(works, cpu));
3122
3123	put_online_cpus();
3124	free_percpu(works);
3125	return 0;
3126	}
3127
3128	/**
3129	* execute_in_process_context - reliably execute the routine with user context
3130	* @fn: the function to execute
3131	* @ew: guaranteed storage for the execute work structure (must
3132	* be available when the work executes)
3133	*
3134	* Executes the function immediately if process context is available,
3135	* otherwise schedules the function for delayed execution.
3136	*
3137	* Return: 0 - function was executed
3138	* 1 - function was scheduled for execution
3139	*/
3140	int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3141	{
3142	if (!in_interrupt()) {
3143	fn(&ew->work);
3144	return 0;
3145	}
3146
3147	INIT_WORK(&ew->work, fn);
3148	schedule_work(&ew->work);
3149
3150	return 1;
3151	}
3152	EXPORT_SYMBOL_GPL(execute_in_process_context);
3153
3154	/**
3155	* free_workqueue_attrs - free a workqueue_attrs
3156	* @attrs: workqueue_attrs to free
3157	*
3158	* Undo alloc_workqueue_attrs().
3159	*/
3160	void free_workqueue_attrs(struct workqueue_attrs *attrs)
3161	{
3162	if (attrs) {
3163	free_cpumask_var(attrs->cpumask);
3164	kfree(attrs);
3165	}
3166	}
3167
3168	/**
3169	* alloc_workqueue_attrs - allocate a workqueue_attrs
3170	* @gfp_mask: allocation mask to use
3171	*
3172	* Allocate a new workqueue_attrs, initialize with default settings and
3173	* return it.
3174	*
3175	* Return: The allocated new workqueue_attr on success. %NULL on failure.
3176	*/
3177	struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3178	{
3179	struct workqueue_attrs *attrs;
3180
3181	attrs = kzalloc(sizeof(*attrs), gfp_mask);
3182	if (!attrs)
3183	goto fail;
3184	if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3185	goto fail;
3186
3187	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3188	return attrs;
3189	fail:
3190	free_workqueue_attrs(attrs);
3191	return NULL;
3192	}
3193
3194	static void copy_workqueue_attrs(struct workqueue_attrs *to,
3195	const struct workqueue_attrs *from)
3196	{
3197	to->nice = from->nice;
3198	cpumask_copy(to->cpumask, from->cpumask);
3199	/*
3200	* Unlike hash and equality test, this function doesn't ignore
3201	* ->no_numa as it is used for both pool and wq attrs. Instead,
3202	* get_unbound_pool() explicitly clears ->no_numa after copying.
3203	*/
3204	to->no_numa = from->no_numa;
3205	}
3206
3207	/* hash value of the content of @attr */
3208	static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3209	{
3210	u32 hash = 0;
3211
3212	hash = jhash_1word(attrs->nice, hash);
3213	hash = jhash(cpumask_bits(attrs->cpumask),
3214	BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3215	return hash;
3216	}
3217
3218	/* content equality test */
3219	static bool wqattrs_equal(const struct workqueue_attrs *a,
3220	const struct workqueue_attrs *b)
3221	{
3222	if (a->nice != b->nice)
3223	return false;
3224	if (!cpumask_equal(a->cpumask, b->cpumask))
3225	return false;
3226	return true;
3227	}
3228
3229	/**
3230	* init_worker_pool - initialize a newly zalloc'd worker_pool
3231	* @pool: worker_pool to initialize
3232	*
3233	* Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3234	*
3235	* Return: 0 on success, -errno on failure. Even on failure, all fields
3236	* inside @pool proper are initialized and put_unbound_pool() can be called
3237	* on @pool safely to release it.
3238	*/
3239	static int init_worker_pool(struct worker_pool *pool)
3240	{
3241	spin_lock_init(&pool->lock);
3242	pool->id = -1;
3243	pool->cpu = -1;
3244	pool->node = NUMA_NO_NODE;
3245	pool->flags |= POOL_DISASSOCIATED;
3246	pool->watchdog_ts = jiffies;
3247	INIT_LIST_HEAD(&pool->worklist);
3248	INIT_LIST_HEAD(&pool->idle_list);
3249	hash_init(pool->busy_hash);
3250
3251	init_timer_deferrable(&pool->idle_timer);
3252	pool->idle_timer.function = idle_worker_timeout;
3253	pool->idle_timer.data = (unsigned long)pool;
3254
3255	setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3256	(unsigned long)pool);
3257
3258	mutex_init(&pool->attach_mutex);
3259	INIT_LIST_HEAD(&pool->workers);
3260
3261	ida_init(&pool->worker_ida);
3262	INIT_HLIST_NODE(&pool->hash_node);
3263	pool->refcnt = 1;
3264
3265	/* shouldn't fail above this point */
3266	pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3267	if (!pool->attrs)
3268	return -ENOMEM;
3269	return 0;
3270	}
3271
3272	static void rcu_free_wq(struct rcu_head *rcu)
3273	{
3274	struct workqueue_struct *wq =
3275	container_of(rcu, struct workqueue_struct, rcu);
3276
3277	if (!(wq->flags & WQ_UNBOUND))
3278	free_percpu(wq->cpu_pwqs);
3279	else
3280	free_workqueue_attrs(wq->unbound_attrs);
3281
3282	kfree(wq->rescuer);
3283	kfree(wq);
3284	}
3285
3286	static void rcu_free_pool(struct rcu_head *rcu)
3287	{
3288	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3289
3290	ida_destroy(&pool->worker_ida);
3291	free_workqueue_attrs(pool->attrs);
3292	kfree(pool);
3293	}
3294
3295	/**
3296	* put_unbound_pool - put a worker_pool
3297	* @pool: worker_pool to put
3298	*
3299	* Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3300	* safe manner. get_unbound_pool() calls this function on its failure path
3301	* and this function should be able to release pools which went through,
3302	* successfully or not, init_worker_pool().
3303	*
3304	* Should be called with wq_pool_mutex held.
3305	*/
3306	static void put_unbound_pool(struct worker_pool *pool)
3307	{
3308	DECLARE_COMPLETION_ONSTACK(detach_completion);
3309	struct worker *worker;
3310
3311	lockdep_assert_held(&wq_pool_mutex);
3312
3313	if (--pool->refcnt)
3314	return;
3315
3316	/* sanity checks */
3317	if (WARN_ON(!(pool->cpu < 0)) ||
3318	WARN_ON(!list_empty(&pool->worklist)))
3319	return;
3320
3321	/* release id and unhash */
3322	if (pool->id >= 0)
3323	idr_remove(&worker_pool_idr, pool->id);
3324	hash_del(&pool->hash_node);
3325
3326	/*
3327	* Become the manager and destroy all workers. This prevents
3328	* @pool's workers from blocking on attach_mutex. We're the last
3329	* manager and @pool gets freed with the flag set.
3330	*/
3331	spin_lock_irq(&pool->lock);
3332	wait_event_lock_irq(wq_manager_wait,
3333	!(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3334	pool->flags |= POOL_MANAGER_ACTIVE;
3335
3336	while ((worker = first_idle_worker(pool)))
3337	destroy_worker(worker);
3338	WARN_ON(pool->nr_workers || pool->nr_idle);
3339	spin_unlock_irq(&pool->lock);
3340
3341	mutex_lock(&pool->attach_mutex);
3342	if (!list_empty(&pool->workers))
3343	pool->detach_completion = &detach_completion;
3344	mutex_unlock(&pool->attach_mutex);
3345
3346	if (pool->detach_completion)
3347	wait_for_completion(pool->detach_completion);
3348
3349	/* shut down the timers */
3350	del_timer_sync(&pool->idle_timer);
3351	del_timer_sync(&pool->mayday_timer);
3352
3353	/* sched-RCU protected to allow dereferences from get_work_pool() */
3354	call_rcu_sched(&pool->rcu, rcu_free_pool);
3355	}
3356
3357	/**
3358	* get_unbound_pool - get a worker_pool with the specified attributes
3359	* @attrs: the attributes of the worker_pool to get
3360	*
3361	* Obtain a worker_pool which has the same attributes as @attrs, bump the
3362	* reference count and return it. If there already is a matching
3363	* worker_pool, it will be used; otherwise, this function attempts to
3364	* create a new one.
3365	*
3366	* Should be called with wq_pool_mutex held.
3367	*
3368	* Return: On success, a worker_pool with the same attributes as @attrs.
3369	* On failure, %NULL.
3370	*/
3371	static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3372	{
3373	u32 hash = wqattrs_hash(attrs);
3374	struct worker_pool *pool;
3375	int node;
3376	int target_node = NUMA_NO_NODE;
3377
3378	lockdep_assert_held(&wq_pool_mutex);
3379
3380	/* do we already have a matching pool? */
3381	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3382	if (wqattrs_equal(pool->attrs, attrs)) {
3383	pool->refcnt++;
3384	return pool;
3385	}
3386	}
3387
3388	/* if cpumask is contained inside a NUMA node, we belong to that node */
3389	if (wq_numa_enabled) {
3390	for_each_node(node) {
3391	if (cpumask_subset(attrs->cpumask,
3392	wq_numa_possible_cpumask[node])) {
3393	target_node = node;
3394	break;
3395	}
3396	}
3397	}
3398
3399	/* nope, create a new one */
3400	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3401	if (!pool || init_worker_pool(pool) < 0)
3402	goto fail;
3403
3404	lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3405	copy_workqueue_attrs(pool->attrs, attrs);
3406	pool->node = target_node;
3407
3408	/*
3409	* no_numa isn't a worker_pool attribute, always clear it. See
3410	* 'struct workqueue_attrs' comments for detail.
3411	*/
3412	pool->attrs->no_numa = false;
3413
3414	if (worker_pool_assign_id(pool) < 0)
3415	goto fail;
3416
3417	/* create and start the initial worker */
3418	if (wq_online && !create_worker(pool))
3419	goto fail;
3420
3421	/* install */
3422	hash_add(unbound_pool_hash, &pool->hash_node, hash);
3423
3424	return pool;
3425	fail:
3426	if (pool)
3427	put_unbound_pool(pool);
3428	return NULL;
3429	}
3430
3431	static void rcu_free_pwq(struct rcu_head *rcu)
3432	{
3433	kmem_cache_free(pwq_cache,
3434	container_of(rcu, struct pool_workqueue, rcu));
3435	}
3436
3437	/*
3438	* Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3439	* and needs to be destroyed.
3440	*/
3441	static void pwq_unbound_release_workfn(struct work_struct *work)
3442	{
3443	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3444	unbound_release_work);
3445	struct workqueue_struct *wq = pwq->wq;
3446	struct worker_pool *pool = pwq->pool;
3447	bool is_last;
3448
3449	if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3450	return;
3451
3452	mutex_lock(&wq->mutex);
3453	list_del_rcu(&pwq->pwqs_node);
3454	is_last = list_empty(&wq->pwqs);
3455	mutex_unlock(&wq->mutex);
3456
3457	mutex_lock(&wq_pool_mutex);
3458	put_unbound_pool(pool);
3459	mutex_unlock(&wq_pool_mutex);
3460
3461	call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3462
3463	/*
3464	* If we're the last pwq going away, @wq is already dead and no one
3465	* is gonna access it anymore. Schedule RCU free.
3466	*/
3467	if (is_last)
3468	call_rcu_sched(&wq->rcu, rcu_free_wq);
3469	}
3470
3471	/**
3472	* pwq_adjust_max_active - update a pwq's max_active to the current setting
3473	* @pwq: target pool_workqueue
3474	*
3475	* If @pwq isn't freezing, set @pwq->max_active to the associated
3476	* workqueue's saved_max_active and activate delayed work items
3477	* accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3478	*/
3479	static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3480	{
3481	struct workqueue_struct *wq = pwq->wq;
3482	bool freezable = wq->flags & WQ_FREEZABLE;
3483	unsigned long flags;
3484
3485	/* for @wq->saved_max_active */
3486	lockdep_assert_held(&wq->mutex);
3487
3488	/* fast exit for non-freezable wqs */
3489	if (!freezable && pwq->max_active == wq->saved_max_active)
3490	return;
3491
3492	/* this function can be called during early boot w/ irq disabled */
3493	spin_lock_irqsave(&pwq->pool->lock, flags);
3494
3495	/*
3496	* During [un]freezing, the caller is responsible for ensuring that
3497	* this function is called at least once after @workqueue_freezing
3498	* is updated and visible.
3499	*/
3500	if (!freezable || !workqueue_freezing) {
3501	pwq->max_active = wq->saved_max_active;
3502
3503	while (!list_empty(&pwq->delayed_works) &&
3504	pwq->nr_active < pwq->max_active)
3505	pwq_activate_first_delayed(pwq);
3506
3507	/*
3508	* Need to kick a worker after thawed or an unbound wq's
3509	* max_active is bumped. It's a slow path. Do it always.
3510	*/
3511	wake_up_worker(pwq->pool);
3512	} else {
3513	pwq->max_active = 0;
3514	}
3515
3516	spin_unlock_irqrestore(&pwq->pool->lock, flags);
3517	}
3518
3519	/* initialize newly alloced @pwq which is associated with @wq and @pool */
3520	static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3521	struct worker_pool *pool)
3522	{
3523	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3524
3525	memset(pwq, 0, sizeof(*pwq));
3526
3527	pwq->pool = pool;
3528	pwq->wq = wq;
3529	pwq->flush_color = -1;
3530	pwq->refcnt = 1;
3531	INIT_LIST_HEAD(&pwq->delayed_works);
3532	INIT_LIST_HEAD(&pwq->pwqs_node);
3533	INIT_LIST_HEAD(&pwq->mayday_node);
3534	INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3535	}
3536
3537	/* sync @pwq with the current state of its associated wq and link it */
3538	static void link_pwq(struct pool_workqueue *pwq)
3539	{
3540	struct workqueue_struct *wq = pwq->wq;
3541
3542	lockdep_assert_held(&wq->mutex);
3543
3544	/* may be called multiple times, ignore if already linked */
3545	if (!list_empty(&pwq->pwqs_node))
3546	return;
3547
3548	/* set the matching work_color */
3549	pwq->work_color = wq->work_color;
3550
3551	/* sync max_active to the current setting */
3552	pwq_adjust_max_active(pwq);
3553
3554	/* link in @pwq */
3555	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3556	}
3557
3558	/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3559	static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3560	const struct workqueue_attrs *attrs)
3561	{
3562	struct worker_pool *pool;
3563	struct pool_workqueue *pwq;
3564
3565	lockdep_assert_held(&wq_pool_mutex);
3566
3567	pool = get_unbound_pool(attrs);
3568	if (!pool)
3569	return NULL;
3570
3571	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3572	if (!pwq) {
3573	put_unbound_pool(pool);
3574	return NULL;
3575	}
3576
3577	init_pwq(pwq, wq, pool);
3578	return pwq;
3579	}
3580
3581	/**
3582	* wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3583	* @attrs: the wq_attrs of the default pwq of the target workqueue
3584	* @node: the target NUMA node
3585	* @cpu_going_down: if >= 0, the CPU to consider as offline
3586	* @cpumask: outarg, the resulting cpumask
3587	*
3588	* Calculate the cpumask a workqueue with @attrs should use on @node. If
3589	* @cpu_going_down is >= 0, that cpu is considered offline during
3590	* calculation. The result is stored in @cpumask.
3591	*
3592	* If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3593	* enabled and @node has online CPUs requested by @attrs, the returned
3594	* cpumask is the intersection of the possible CPUs of @node and
3595	* @attrs->cpumask.
3596	*
3597	* The caller is responsible for ensuring that the cpumask of @node stays
3598	* stable.
3599	*
3600	* Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3601	* %false if equal.
3602	*/
3603	static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3604	int cpu_going_down, cpumask_t *cpumask)
3605	{
3606	if (!wq_numa_enabled || attrs->no_numa)
3607	goto use_dfl;
3608
3609	/* does @node have any online CPUs @attrs wants? */
3610	cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3611	if (cpu_going_down >= 0)
3612	cpumask_clear_cpu(cpu_going_down, cpumask);
3613
3614	if (cpumask_empty(cpumask))
3615	goto use_dfl;
3616
3617	/* yeap, return possible CPUs in @node that @attrs wants */
3618	cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3619	return !cpumask_equal(cpumask, attrs->cpumask);
3620
3621	use_dfl:
3622	cpumask_copy(cpumask, attrs->cpumask);
3623	return false;
3624	}
3625
3626	/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3627	static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3628	int node,
3629	struct pool_workqueue *pwq)
3630	{
3631	struct pool_workqueue *old_pwq;
3632
3633	lockdep_assert_held(&wq_pool_mutex);
3634	lockdep_assert_held(&wq->mutex);
3635
3636	/* link_pwq() can handle duplicate calls */
3637	link_pwq(pwq);
3638
3639	old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3640	rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3641	return old_pwq;
3642	}
3643
3644	/* context to store the prepared attrs & pwqs before applying */
3645	struct apply_wqattrs_ctx {
3646	struct workqueue_struct *wq; /* target workqueue */
3647	struct workqueue_attrs *attrs; /* attrs to apply */
3648	struct list_head list; /* queued for batching commit */
3649	struct pool_workqueue *dfl_pwq;
3650	struct pool_workqueue *pwq_tbl[];
3651	};
3652
3653	/* free the resources after success or abort */
3654	static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3655	{
3656	if (ctx) {
3657	int node;
3658
3659	for_each_node(node)
3660	put_pwq_unlocked(ctx->pwq_tbl[node]);
3661	put_pwq_unlocked(ctx->dfl_pwq);
3662
3663	free_workqueue_attrs(ctx->attrs);
3664
3665	kfree(ctx);
3666	}
3667	}
3668
3669	/* allocate the attrs and pwqs for later installation */
3670	static struct apply_wqattrs_ctx *
3671	apply_wqattrs_prepare(struct workqueue_struct *wq,
3672	const struct workqueue_attrs *attrs)
3673	{
3674	struct apply_wqattrs_ctx *ctx;
3675	struct workqueue_attrs *new_attrs, *tmp_attrs;
3676	int node;
3677
3678	lockdep_assert_held(&wq_pool_mutex);
3679
3680	ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3681	GFP_KERNEL);
3682
3683	new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3684	tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3685	if (!ctx || !new_attrs || !tmp_attrs)
3686	goto out_free;
3687
3688	/*
3689	* Calculate the attrs of the default pwq.
3690	* If the user configured cpumask doesn't overlap with the
3691	* wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3692	*/
3693	copy_workqueue_attrs(new_attrs, attrs);
3694	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3695	if (unlikely(cpumask_empty(new_attrs->cpumask)))
3696	cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3697
3698	/*
3699	* We may create multiple pwqs with differing cpumasks. Make a
3700	* copy of @new_attrs which will be modified and used to obtain
3701	* pools.
3702	*/
3703	copy_workqueue_attrs(tmp_attrs, new_attrs);
3704
3705	/*
3706	* If something goes wrong during CPU up/down, we'll fall back to
3707	* the default pwq covering whole @attrs->cpumask. Always create
3708	* it even if we don't use it immediately.
3709	*/
3710	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3711	if (!ctx->dfl_pwq)
3712	goto out_free;
3713
3714	for_each_node(node) {
3715	if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3716	ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3717	if (!ctx->pwq_tbl[node])
3718	goto out_free;
3719	} else {
3720	ctx->dfl_pwq->refcnt++;
3721	ctx->pwq_tbl[node] = ctx->dfl_pwq;
3722	}
3723	}
3724
3725	/* save the user configured attrs and sanitize it. */
3726	copy_workqueue_attrs(new_attrs, attrs);
3727	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3728	ctx->attrs = new_attrs;
3729
3730	ctx->wq = wq;
3731	free_workqueue_attrs(tmp_attrs);
3732	return ctx;
3733
3734	out_free:
3735	free_workqueue_attrs(tmp_attrs);
3736	free_workqueue_attrs(new_attrs);
3737	apply_wqattrs_cleanup(ctx);
3738	return NULL;
3739	}
3740
3741	/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3742	static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3743	{
3744	int node;
3745
3746	/* all pwqs have been created successfully, let's install'em */
3747	mutex_lock(&ctx->wq->mutex);
3748
3749	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3750
3751	/* save the previous pwq and install the new one */
3752	for_each_node(node)
3753	ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3754	ctx->pwq_tbl[node]);
3755
3756	/* @dfl_pwq might not have been used, ensure it's linked */
3757	link_pwq(ctx->dfl_pwq);
3758	swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3759
3760	mutex_unlock(&ctx->wq->mutex);
3761	}
3762
3763	static void apply_wqattrs_lock(void)
3764	{
3765	/* CPUs should stay stable across pwq creations and installations */
3766	get_online_cpus();
3767	mutex_lock(&wq_pool_mutex);
3768	}
3769
3770	static void apply_wqattrs_unlock(void)
3771	{
3772	mutex_unlock(&wq_pool_mutex);
3773	put_online_cpus();
3774	}
3775
3776	static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3777	const struct workqueue_attrs *attrs)
3778	{
3779	struct apply_wqattrs_ctx *ctx;
3780
3781	/* only unbound workqueues can change attributes */
3782	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3783	return -EINVAL;
3784
3785	/* creating multiple pwqs breaks ordering guarantee */
3786	if (!list_empty(&wq->pwqs)) {
3787	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3788	return -EINVAL;
3789
3790	wq->flags &= ~__WQ_ORDERED;
3791	}
3792
3793	ctx = apply_wqattrs_prepare(wq, attrs);
3794	if (!ctx)
3795	return -ENOMEM;
3796
3797	/* the ctx has been prepared successfully, let's commit it */
3798	apply_wqattrs_commit(ctx);
3799	apply_wqattrs_cleanup(ctx);
3800
3801	return 0;
3802	}
3803
3804	/**
3805	* apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3806	* @wq: the target workqueue
3807	* @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3808	*
3809	* Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3810	* machines, this function maps a separate pwq to each NUMA node with
3811	* possibles CPUs in @attrs->cpumask so that work items are affine to the
3812	* NUMA node it was issued on. Older pwqs are released as in-flight work
3813	* items finish. Note that a work item which repeatedly requeues itself
3814	* back-to-back will stay on its current pwq.
3815	*
3816	* Performs GFP_KERNEL allocations.
3817	*
3818	* Return: 0 on success and -errno on failure.
3819	*/
3820	int apply_workqueue_attrs(struct workqueue_struct *wq,
3821	const struct workqueue_attrs *attrs)
3822	{
3823	int ret;
3824
3825	apply_wqattrs_lock();
3826	ret = apply_workqueue_attrs_locked(wq, attrs);
3827	apply_wqattrs_unlock();
3828
3829	return ret;
3830	}
3831
3832	/**
3833	* wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3834	* @wq: the target workqueue
3835	* @cpu: the CPU coming up or going down
3836	* @online: whether @cpu is coming up or going down
3837	*
3838	* This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3839	* %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3840	* @wq accordingly.
3841	*
3842	* If NUMA affinity can't be adjusted due to memory allocation failure, it
3843	* falls back to @wq->dfl_pwq which may not be optimal but is always
3844	* correct.
3845	*
3846	* Note that when the last allowed CPU of a NUMA node goes offline for a
3847	* workqueue with a cpumask spanning multiple nodes, the workers which were
3848	* already executing the work items for the workqueue will lose their CPU
3849	* affinity and may execute on any CPU. This is similar to how per-cpu
3850	* workqueues behave on CPU_DOWN. If a workqueue user wants strict
3851	* affinity, it's the user's responsibility to flush the work item from
3852	* CPU_DOWN_PREPARE.
3853	*/
3854	static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3855	bool online)
3856	{
3857	int node = cpu_to_node(cpu);
3858	int cpu_off = online ? -1 : cpu;
3859	struct pool_workqueue *old_pwq = NULL, *pwq;
3860	struct workqueue_attrs *target_attrs;
3861	cpumask_t *cpumask;
3862
3863	lockdep_assert_held(&wq_pool_mutex);
3864
3865	if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3866	wq->unbound_attrs->no_numa)
3867	return;
3868
3869	/*
3870	* We don't wanna alloc/free wq_attrs for each wq for each CPU.
3871	* Let's use a preallocated one. The following buf is protected by
3872	* CPU hotplug exclusion.
3873	*/
3874	target_attrs = wq_update_unbound_numa_attrs_buf;
3875	cpumask = target_attrs->cpumask;
3876
3877	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3878	pwq = unbound_pwq_by_node(wq, node);
3879
3880	/*
3881	* Let's determine what needs to be done. If the target cpumask is
3882	* different from the default pwq's, we need to compare it to @pwq's
3883	* and create a new one if they don't match. If the target cpumask
3884	* equals the default pwq's, the default pwq should be used.
3885	*/
3886	if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3887	if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3888	return;
3889	} else {
3890	goto use_dfl_pwq;
3891	}
3892
3893	/* create a new pwq */
3894	pwq = alloc_unbound_pwq(wq, target_attrs);
3895	if (!pwq) {
3896	pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3897	wq->name);
3898	goto use_dfl_pwq;
3899	}
3900
3901	/* Install the new pwq. */
3902	mutex_lock(&wq->mutex);
3903	old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3904	goto out_unlock;
3905
3906	use_dfl_pwq:
3907	mutex_lock(&wq->mutex);
3908	spin_lock_irq(&wq->dfl_pwq->pool->lock);
3909	get_pwq(wq->dfl_pwq);
3910	spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3911	old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3912	out_unlock:
3913	mutex_unlock(&wq->mutex);
3914	put_pwq_unlocked(old_pwq);
3915	}
3916
3917	static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3918	{
3919	bool highpri = wq->flags & WQ_HIGHPRI;
3920	int cpu, ret;
3921
3922	if (!(wq->flags & WQ_UNBOUND)) {
3923	wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3924	if (!wq->cpu_pwqs)
3925	return -ENOMEM;
3926
3927	for_each_possible_cpu(cpu) {
3928	struct pool_workqueue *pwq =
3929	per_cpu_ptr(wq->cpu_pwqs, cpu);
3930	struct worker_pool *cpu_pools =
3931	per_cpu(cpu_worker_pools, cpu);
3932
3933	init_pwq(pwq, wq, &cpu_pools[highpri]);
3934
3935	mutex_lock(&wq->mutex);
3936	link_pwq(pwq);
3937	mutex_unlock(&wq->mutex);
3938	}
3939	return 0;
3940	} else if (wq->flags & __WQ_ORDERED) {
3941	ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3942	/* there should only be single pwq for ordering guarantee */
3943	WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3944	wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3945	"ordering guarantee broken for workqueue %s\n", wq->name);
3946	return ret;
3947	} else {
3948	return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3949	}
3950	}
3951
3952	static int wq_clamp_max_active(int max_active, unsigned int flags,
3953	const char *name)
3954	{
3955	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3956
3957	if (max_active < 1 || max_active > lim)
3958	pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3959	max_active, name, 1, lim);
3960
3961	return clamp_val(max_active, 1, lim);
3962	}
3963
3964	struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3965	unsigned int flags,
3966	int max_active,
3967	struct lock_class_key *key,
3968	const char *lock_name, ...)
3969	{
3970	size_t tbl_size = 0;
3971	va_list args;
3972	struct workqueue_struct *wq;
3973	struct pool_workqueue *pwq;
3974
3975	/*
3976	* Unbound && max_active == 1 used to imply ordered, which is no
3977	* longer the case on NUMA machines due to per-node pools. While
3978	* alloc_ordered_workqueue() is the right way to create an ordered
3979	* workqueue, keep the previous behavior to avoid subtle breakages
3980	* on NUMA.
3981	*/
3982	if ((flags & WQ_UNBOUND) && max_active == 1)
3983	flags |= __WQ_ORDERED;
3984
3985	/* see the comment above the definition of WQ_POWER_EFFICIENT */
3986	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3987	flags |= WQ_UNBOUND;
3988
3989	/* allocate wq and format name */
3990	if (flags & WQ_UNBOUND)
3991	tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3992
3993	wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3994	if (!wq)
3995	return NULL;
3996
3997	if (flags & WQ_UNBOUND) {
3998	wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3999	if (!wq->unbound_attrs)
4000	goto err_free_wq;
4001	}
4002
4003	va_start(args, lock_name);
4004	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4005	va_end(args);
4006
4007	max_active = max_active ?: WQ_DFL_ACTIVE;
4008	max_active = wq_clamp_max_active(max_active, flags, wq->name);
4009
4010	/* init wq */
4011	wq->flags = flags;
4012	wq->saved_max_active = max_active;
4013	mutex_init(&wq->mutex);
4014	atomic_set(&wq->nr_pwqs_to_flush, 0);
4015	INIT_LIST_HEAD(&wq->pwqs);
4016	INIT_LIST_HEAD(&wq->flusher_queue);
4017	INIT_LIST_HEAD(&wq->flusher_overflow);
4018	INIT_LIST_HEAD(&wq->maydays);
4019
4020	lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4021	INIT_LIST_HEAD(&wq->list);
4022
4023	if (alloc_and_link_pwqs(wq) < 0)
4024	goto err_free_wq;
4025
4026	/*
4027	* Workqueues which may be used during memory reclaim should
4028	* have a rescuer to guarantee forward progress.
4029	*/
4030	if (flags & WQ_MEM_RECLAIM) {
4031	struct worker *rescuer;
4032
4033	rescuer = alloc_worker(NUMA_NO_NODE);
4034	if (!rescuer)
4035	goto err_destroy;
4036
4037	rescuer->rescue_wq = wq;
4038	rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4039	wq->name);
4040	if (IS_ERR(rescuer->task)) {
4041	kfree(rescuer);
4042	goto err_destroy;
4043	}
4044
4045	wq->rescuer = rescuer;
4046	kthread_bind_mask(rescuer->task, cpu_possible_mask);
4047	wake_up_process(rescuer->task);
4048	}
4049
4050	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4051	goto err_destroy;
4052
4053	/*
4054	* wq_pool_mutex protects global freeze state and workqueues list.
4055	* Grab it, adjust max_active and add the new @wq to workqueues
4056	* list.
4057	*/
4058	mutex_lock(&wq_pool_mutex);
4059
4060	mutex_lock(&wq->mutex);
4061	for_each_pwq(pwq, wq)
4062	pwq_adjust_max_active(pwq);
4063	mutex_unlock(&wq->mutex);
4064
4065	list_add_tail_rcu(&wq->list, &workqueues);
4066
4067	mutex_unlock(&wq_pool_mutex);
4068
4069	return wq;
4070
4071	err_free_wq:
4072	free_workqueue_attrs(wq->unbound_attrs);
4073	kfree(wq);
4074	return NULL;
4075	err_destroy:
4076	destroy_workqueue(wq);
4077	return NULL;
4078	}
4079	EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4080
4081	/**
4082	* destroy_workqueue - safely terminate a workqueue
4083	* @wq: target workqueue
4084	*
4085	* Safely destroy a workqueue. All work currently pending will be done first.
4086	*/
4087	void destroy_workqueue(struct workqueue_struct *wq)
4088	{
4089	struct pool_workqueue *pwq;
4090	int node;
4091
4092	/* drain it before proceeding with destruction */
4093	drain_workqueue(wq);
4094
4095	/* sanity checks */
4096	mutex_lock(&wq->mutex);
4097	for_each_pwq(pwq, wq) {
4098	int i;
4099
4100	for (i = 0; i < WORK_NR_COLORS; i++) {
4101	if (WARN_ON(pwq->nr_in_flight[i])) {
4102	mutex_unlock(&wq->mutex);
4103	return;
4104	}
4105	}
4106
4107	if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4108	WARN_ON(pwq->nr_active) ||
4109	WARN_ON(!list_empty(&pwq->delayed_works))) {
4110	mutex_unlock(&wq->mutex);
4111	return;
4112	}
4113	}
4114	mutex_unlock(&wq->mutex);
4115
4116	/*
4117	* wq list is used to freeze wq, remove from list after
4118	* flushing is complete in case freeze races us.
4119	*/
4120	mutex_lock(&wq_pool_mutex);
4121	list_del_rcu(&wq->list);
4122	mutex_unlock(&wq_pool_mutex);
4123
4124	workqueue_sysfs_unregister(wq);
4125
4126	if (wq->rescuer)
4127	kthread_stop(wq->rescuer->task);
4128
4129	if (!(wq->flags & WQ_UNBOUND)) {
4130	/*
4131	* The base ref is never dropped on per-cpu pwqs. Directly
4132	* schedule RCU free.
4133	*/
4134	call_rcu_sched(&wq->rcu, rcu_free_wq);
4135	} else {
4136	/*
4137	* We're the sole accessor of @wq at this point. Directly
4138	* access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4139	* @wq will be freed when the last pwq is released.
4140	*/
4141	for_each_node(node) {
4142	pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4143	RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4144	put_pwq_unlocked(pwq);
4145	}
4146
4147	/*
4148	* Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4149	* put. Don't access it afterwards.
4150	*/
4151	pwq = wq->dfl_pwq;
4152	wq->dfl_pwq = NULL;
4153	put_pwq_unlocked(pwq);
4154	}
4155	}
4156	EXPORT_SYMBOL_GPL(destroy_workqueue);
4157
4158	/**
4159	* workqueue_set_max_active - adjust max_active of a workqueue
4160	* @wq: target workqueue
4161	* @max_active: new max_active value.
4162	*
4163	* Set max_active of @wq to @max_active.
4164	*
4165	* CONTEXT:
4166	* Don't call from IRQ context.
4167	*/
4168	void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4169	{
4170	struct pool_workqueue *pwq;
4171
4172	/* disallow meddling with max_active for ordered workqueues */
4173	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4174	return;
4175
4176	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4177
4178	mutex_lock(&wq->mutex);
4179
4180	wq->flags &= ~__WQ_ORDERED;
4181	wq->saved_max_active = max_active;
4182
4183	for_each_pwq(pwq, wq)
4184	pwq_adjust_max_active(pwq);
4185
4186	mutex_unlock(&wq->mutex);
4187	}
4188	EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4189
4190	/**
4191	* current_work - retrieve %current task's work struct
4192	*
4193	* Determine if %current task is a workqueue worker and what it's working on.
4194	* Useful to find out the context that the %current task is running in.
4195	*
4196	* Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4197	*/
4198	struct work_struct *current_work(void)
4199	{
4200	struct worker *worker = current_wq_worker();
4201
4202	return worker ? worker->current_work : NULL;
4203	}
4204	EXPORT_SYMBOL(current_work);
4205
4206	/**
4207	* current_is_workqueue_rescuer - is %current workqueue rescuer?
4208	*
4209	* Determine whether %current is a workqueue rescuer. Can be used from
4210	* work functions to determine whether it's being run off the rescuer task.
4211	*
4212	* Return: %true if %current is a workqueue rescuer. %false otherwise.
4213	*/
4214	bool current_is_workqueue_rescuer(void)
4215	{
4216	struct worker *worker = current_wq_worker();
4217
4218	return worker && worker->rescue_wq;
4219	}
4220
4221	/**
4222	* workqueue_congested - test whether a workqueue is congested
4223	* @cpu: CPU in question
4224	* @wq: target workqueue
4225	*
4226	* Test whether @wq's cpu workqueue for @cpu is congested. There is
4227	* no synchronization around this function and the test result is
4228	* unreliable and only useful as advisory hints or for debugging.
4229	*
4230	* If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4231	* Note that both per-cpu and unbound workqueues may be associated with
4232	* multiple pool_workqueues which have separate congested states. A
4233	* workqueue being congested on one CPU doesn't mean the workqueue is also
4234	* contested on other CPUs / NUMA nodes.
4235	*
4236	* Return:
4237	* %true if congested, %false otherwise.
4238	*/
4239	bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4240	{
4241	struct pool_workqueue *pwq;
4242	bool ret;
4243
4244	rcu_read_lock_sched();
4245
4246	if (cpu == WORK_CPU_UNBOUND)
4247	cpu = smp_processor_id();
4248
4249	if (!(wq->flags & WQ_UNBOUND))
4250	pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4251	else
4252	pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4253
4254	ret = !list_empty(&pwq->delayed_works);
4255	rcu_read_unlock_sched();
4256
4257	return ret;
4258	}
4259	EXPORT_SYMBOL_GPL(workqueue_congested);
4260
4261	/**
4262	* work_busy - test whether a work is currently pending or running
4263	* @work: the work to be tested
4264	*
4265	* Test whether @work is currently pending or running. There is no
4266	* synchronization around this function and the test result is
4267	* unreliable and only useful as advisory hints or for debugging.
4268	*
4269	* Return:
4270	* OR'd bitmask of WORK_BUSY_* bits.
4271	*/
4272	unsigned int work_busy(struct work_struct *work)
4273	{
4274	struct worker_pool *pool;
4275	unsigned long flags;
4276	unsigned int ret = 0;
4277
4278	if (work_pending(work))
4279	ret |= WORK_BUSY_PENDING;
4280
4281	local_irq_save(flags);
4282	pool = get_work_pool(work);
4283	if (pool) {
4284	spin_lock(&pool->lock);
4285	if (find_worker_executing_work(pool, work))
4286	ret |= WORK_BUSY_RUNNING;
4287	spin_unlock(&pool->lock);
4288	}
4289	local_irq_restore(flags);
4290
4291	return ret;
4292	}
4293	EXPORT_SYMBOL_GPL(work_busy);
4294
4295	/**
4296	* set_worker_desc - set description for the current work item
4297	* @fmt: printf-style format string
4298	* @...: arguments for the format string
4299	*
4300	* This function can be called by a running work function to describe what
4301	* the work item is about. If the worker task gets dumped, this
4302	* information will be printed out together to help debugging. The
4303	* description can be at most WORKER_DESC_LEN including the trailing '\0'.
4304	*/
4305	void set_worker_desc(const char *fmt, ...)
4306	{
4307	struct worker *worker = current_wq_worker();
4308	va_list args;
4309
4310	if (worker) {
4311	va_start(args, fmt);
4312	vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4313	va_end(args);
4314	worker->desc_valid = true;
4315	}
4316	}
4317
4318	/**
4319	* print_worker_info - print out worker information and description
4320	* @log_lvl: the log level to use when printing
4321	* @task: target task
4322	*
4323	* If @task is a worker and currently executing a work item, print out the
4324	* name of the workqueue being serviced and worker description set with
4325	* set_worker_desc() by the currently executing work item.
4326	*
4327	* This function can be safely called on any task as long as the
4328	* task_struct itself is accessible. While safe, this function isn't
4329	* synchronized and may print out mixups or garbages of limited length.
4330	*/
4331	void print_worker_info(const char *log_lvl, struct task_struct *task)
4332	{
4333	work_func_t *fn = NULL;
4334	char name[WQ_NAME_LEN] = { };
4335	char desc[WORKER_DESC_LEN] = { };
4336	struct pool_workqueue *pwq = NULL;
4337	struct workqueue_struct *wq = NULL;
4338	bool desc_valid = false;
4339	struct worker *worker;
4340
4341	if (!(task->flags & PF_WQ_WORKER))
4342	return;
4343
4344	/*
4345	* This function is called without any synchronization and @task
4346	* could be in any state. Be careful with dereferences.
4347	*/
4348	worker = kthread_probe_data(task);
4349
4350	/*
4351	* Carefully copy the associated workqueue's workfn and name. Keep
4352	* the original last '\0' in case the original contains garbage.
4353	*/
4354	probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4355	probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4356	probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4357	probe_kernel_read(name, wq->name, sizeof(name) - 1);
4358
4359	/* copy worker description */
4360	probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4361	if (desc_valid)
4362	probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4363
4364	if (fn || name[0] || desc[0]) {
4365	printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4366	if (desc[0])
4367	pr_cont(" (%s)", desc);
4368	pr_cont("\n");
4369	}
4370	}
4371
4372	static void pr_cont_pool_info(struct worker_pool *pool)
4373	{
4374	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4375	if (pool->node != NUMA_NO_NODE)
4376	pr_cont(" node=%d", pool->node);
4377	pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4378	}
4379
4380	static void pr_cont_work(bool comma, struct work_struct *work)
4381	{
4382	if (work->func == wq_barrier_func) {
4383	struct wq_barrier *barr;
4384
4385	barr = container_of(work, struct wq_barrier, work);
4386
4387	pr_cont("%s BAR(%d)", comma ? "," : "",
4388	task_pid_nr(barr->task));
4389	} else {
4390	pr_cont("%s %pf", comma ? "," : "", work->func);
4391	}
4392	}
4393
4394	static void show_pwq(struct pool_workqueue *pwq)
4395	{
4396	struct worker_pool *pool = pwq->pool;
4397	struct work_struct *work;
4398	struct worker *worker;
4399	bool has_in_flight = false, has_pending = false;
4400	int bkt;
4401
4402	pr_info(" pwq %d:", pool->id);
4403	pr_cont_pool_info(pool);
4404
4405	pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4406	!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4407
4408	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4409	if (worker->current_pwq == pwq) {
4410	has_in_flight = true;
4411	break;
4412	}
4413	}
4414	if (has_in_flight) {
4415	bool comma = false;
4416
4417	pr_info(" in-flight:");
4418	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4419	if (worker->current_pwq != pwq)
4420	continue;
4421
4422	pr_cont("%s %d%s:%pf", comma ? "," : "",
4423	task_pid_nr(worker->task),
4424	worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4425	worker->current_func);
4426	list_for_each_entry(work, &worker->scheduled, entry)
4427	pr_cont_work(false, work);
4428	comma = true;
4429	}
4430	pr_cont("\n");
4431	}
4432
4433	list_for_each_entry(work, &pool->worklist, entry) {
4434	if (get_work_pwq(work) == pwq) {
4435	has_pending = true;
4436	break;
4437	}
4438	}
4439	if (has_pending) {
4440	bool comma = false;
4441
4442	pr_info(" pending:");
4443	list_for_each_entry(work, &pool->worklist, entry) {
4444	if (get_work_pwq(work) != pwq)
4445	continue;
4446
4447	pr_cont_work(comma, work);
4448	comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4449	}
4450	pr_cont("\n");
4451	}
4452
4453	if (!list_empty(&pwq->delayed_works)) {
4454	bool comma = false;
4455
4456	pr_info(" delayed:");
4457	list_for_each_entry(work, &pwq->delayed_works, entry) {
4458	pr_cont_work(comma, work);
4459	comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4460	}
4461	pr_cont("\n");
4462	}
4463	}
4464
4465	/**
4466	* show_workqueue_state - dump workqueue state
4467	*
4468	* Called from a sysrq handler or try_to_freeze_tasks() and prints out
4469	* all busy workqueues and pools.
4470	*/
4471	void show_workqueue_state(void)
4472	{
4473	struct workqueue_struct *wq;
4474	struct worker_pool *pool;
4475	unsigned long flags;
4476	int pi;
4477
4478	rcu_read_lock_sched();
4479
4480	pr_info("Showing busy workqueues and worker pools:\n");
4481
4482	list_for_each_entry_rcu(wq, &workqueues, list) {
4483	struct pool_workqueue *pwq;
4484	bool idle = true;
4485
4486	for_each_pwq(pwq, wq) {
4487	if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4488	idle = false;
4489	break;
4490	}
4491	}
4492	if (idle)
4493	continue;
4494
4495	pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4496
4497	for_each_pwq(pwq, wq) {
4498	spin_lock_irqsave(&pwq->pool->lock, flags);
4499	if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4500	show_pwq(pwq);
4501	spin_unlock_irqrestore(&pwq->pool->lock, flags);
4502	/*
4503	* We could be printing a lot from atomic context, e.g.
4504	* sysrq-t -> show_workqueue_state(). Avoid triggering
4505	* hard lockup.
4506	*/
4507	touch_nmi_watchdog();
4508	}
4509	}
4510
4511	for_each_pool(pool, pi) {
4512	struct worker *worker;
4513	bool first = true;
4514
4515	spin_lock_irqsave(&pool->lock, flags);
4516	if (pool->nr_workers == pool->nr_idle)
4517	goto next_pool;
4518
4519	pr_info("pool %d:", pool->id);
4520	pr_cont_pool_info(pool);
4521	pr_cont(" hung=%us workers=%d",
4522	jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4523	pool->nr_workers);
4524	if (pool->manager)
4525	pr_cont(" manager: %d",
4526	task_pid_nr(pool->manager->task));
4527	list_for_each_entry(worker, &pool->idle_list, entry) {
4528	pr_cont(" %s%d", first ? "idle: " : "",
4529	task_pid_nr(worker->task));
4530	first = false;
4531	}
4532	pr_cont("\n");
4533	next_pool:
4534	spin_unlock_irqrestore(&pool->lock, flags);
4535	/*
4536	* We could be printing a lot from atomic context, e.g.
4537	* sysrq-t -> show_workqueue_state(). Avoid triggering
4538	* hard lockup.
4539	*/
4540	touch_nmi_watchdog();
4541	}
4542
4543	rcu_read_unlock_sched();
4544	}
4545
4546	/*
4547	* CPU hotplug.
4548	*
4549	* There are two challenges in supporting CPU hotplug. Firstly, there
4550	* are a lot of assumptions on strong associations among work, pwq and
4551	* pool which make migrating pending and scheduled works very
4552	* difficult to implement without impacting hot paths. Secondly,
4553	* worker pools serve mix of short, long and very long running works making
4554	* blocked draining impractical.
4555	*
4556	* This is solved by allowing the pools to be disassociated from the CPU
4557	* running as an unbound one and allowing it to be reattached later if the
4558	* cpu comes back online.
4559	*/
4560
4561	static void wq_unbind_fn(struct work_struct *work)
4562	{
4563	int cpu = smp_processor_id();
4564	struct worker_pool *pool;
4565	struct worker *worker;
4566
4567	for_each_cpu_worker_pool(pool, cpu) {
4568	mutex_lock(&pool->attach_mutex);
4569	spin_lock_irq(&pool->lock);
4570
4571	/*
4572	* We've blocked all attach/detach operations. Make all workers
4573	* unbound and set DISASSOCIATED. Before this, all workers
4574	* except for the ones which are still executing works from
4575	* before the last CPU down must be on the cpu. After
4576	* this, they may become diasporas.
4577	*/
4578	for_each_pool_worker(worker, pool)
4579	worker->flags |= WORKER_UNBOUND;
4580
4581	pool->flags |= POOL_DISASSOCIATED;
4582
4583	spin_unlock_irq(&pool->lock);
4584	mutex_unlock(&pool->attach_mutex);
4585
4586	/*
4587	* Call schedule() so that we cross rq->lock and thus can
4588	* guarantee sched callbacks see the %WORKER_UNBOUND flag.
4589	* This is necessary as scheduler callbacks may be invoked
4590	* from other cpus.
4591	*/
4592	schedule();
4593
4594	/*
4595	* Sched callbacks are disabled now. Zap nr_running.
4596	* After this, nr_running stays zero and need_more_worker()
4597	* and keep_working() are always true as long as the
4598	* worklist is not empty. This pool now behaves as an
4599	* unbound (in terms of concurrency management) pool which
4600	* are served by workers tied to the pool.
4601	*/
4602	atomic_set(&pool->nr_running, 0);
4603
4604	/*
4605	* With concurrency management just turned off, a busy
4606	* worker blocking could lead to lengthy stalls. Kick off
4607	* unbound chain execution of currently pending work items.
4608	*/
4609	spin_lock_irq(&pool->lock);
4610	wake_up_worker(pool);
4611	spin_unlock_irq(&pool->lock);
4612	}
4613	}
4614
4615	/**
4616	* rebind_workers - rebind all workers of a pool to the associated CPU
4617	* @pool: pool of interest
4618	*
4619	* @pool->cpu is coming online. Rebind all workers to the CPU.
4620	*/
4621	static void rebind_workers(struct worker_pool *pool)
4622	{
4623	struct worker *worker;
4624
4625	lockdep_assert_held(&pool->attach_mutex);
4626
4627	/*
4628	* Restore CPU affinity of all workers. As all idle workers should
4629	* be on the run-queue of the associated CPU before any local
4630	* wake-ups for concurrency management happen, restore CPU affinity
4631	* of all workers first and then clear UNBOUND. As we're called
4632	* from CPU_ONLINE, the following shouldn't fail.
4633	*/
4634	for_each_pool_worker(worker, pool)
4635	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4636	pool->attrs->cpumask) < 0);
4637
4638	spin_lock_irq(&pool->lock);
4639
4640	/*
4641	* XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4642	* w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4643	* being reworked and this can go away in time.
4644	*/
4645	if (!(pool->flags & POOL_DISASSOCIATED)) {
4646	spin_unlock_irq(&pool->lock);
4647	return;
4648	}
4649
4650	pool->flags &= ~POOL_DISASSOCIATED;
4651
4652	for_each_pool_worker(worker, pool) {
4653	unsigned int worker_flags = worker->flags;
4654
4655	/*
4656	* A bound idle worker should actually be on the runqueue
4657	* of the associated CPU for local wake-ups targeting it to
4658	* work. Kick all idle workers so that they migrate to the
4659	* associated CPU. Doing this in the same loop as
4660	* replacing UNBOUND with REBOUND is safe as no worker will
4661	* be bound before @pool->lock is released.
4662	*/
4663	if (worker_flags & WORKER_IDLE)
4664	wake_up_process(worker->task);
4665
4666	/*
4667	* We want to clear UNBOUND but can't directly call
4668	* worker_clr_flags() or adjust nr_running. Atomically
4669	* replace UNBOUND with another NOT_RUNNING flag REBOUND.
4670	* @worker will clear REBOUND using worker_clr_flags() when
4671	* it initiates the next execution cycle thus restoring
4672	* concurrency management. Note that when or whether
4673	* @worker clears REBOUND doesn't affect correctness.
4674	*
4675	* ACCESS_ONCE() is necessary because @worker->flags may be
4676	* tested without holding any lock in
4677	* wq_worker_waking_up(). Without it, NOT_RUNNING test may
4678	* fail incorrectly leading to premature concurrency
4679	* management operations.
4680	*/
4681	WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4682	worker_flags |= WORKER_REBOUND;
4683	worker_flags &= ~WORKER_UNBOUND;
4684	ACCESS_ONCE(worker->flags) = worker_flags;
4685	}
4686
4687	spin_unlock_irq(&pool->lock);
4688	}
4689
4690	/**
4691	* restore_unbound_workers_cpumask - restore cpumask of unbound workers
4692	* @pool: unbound pool of interest
4693	* @cpu: the CPU which is coming up
4694	*
4695	* An unbound pool may end up with a cpumask which doesn't have any online
4696	* CPUs. When a worker of such pool get scheduled, the scheduler resets
4697	* its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4698	* online CPU before, cpus_allowed of all its workers should be restored.
4699	*/
4700	static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4701	{
4702	static cpumask_t cpumask;
4703	struct worker *worker;
4704
4705	lockdep_assert_held(&pool->attach_mutex);
4706
4707	/* is @cpu allowed for @pool? */
4708	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4709	return;
4710
4711	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4712
4713	/* as we're called from CPU_ONLINE, the following shouldn't fail */
4714	for_each_pool_worker(worker, pool)
4715	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4716	}
4717
4718	int workqueue_prepare_cpu(unsigned int cpu)
4719	{
4720	struct worker_pool *pool;
4721
4722	for_each_cpu_worker_pool(pool, cpu) {
4723	if (pool->nr_workers)
4724	continue;
4725	if (!create_worker(pool))
4726	return -ENOMEM;
4727	}
4728	return 0;
4729	}
4730
4731	int workqueue_online_cpu(unsigned int cpu)
4732	{
4733	struct worker_pool *pool;
4734	struct workqueue_struct *wq;
4735	int pi;
4736
4737	mutex_lock(&wq_pool_mutex);
4738
4739	for_each_pool(pool, pi) {
4740	mutex_lock(&pool->attach_mutex);
4741
4742	if (pool->cpu == cpu)
4743	rebind_workers(pool);
4744	else if (pool->cpu < 0)
4745	restore_unbound_workers_cpumask(pool, cpu);
4746
4747	mutex_unlock(&pool->attach_mutex);
4748	}
4749
4750	/* update NUMA affinity of unbound workqueues */
4751	list_for_each_entry(wq, &workqueues, list)
4752	wq_update_unbound_numa(wq, cpu, true);
4753
4754	mutex_unlock(&wq_pool_mutex);
4755	return 0;
4756	}
4757
4758	int workqueue_offline_cpu(unsigned int cpu)
4759	{
4760	struct work_struct unbind_work;
4761	struct workqueue_struct *wq;
4762
4763	/* unbinding per-cpu workers should happen on the local CPU */
4764	INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4765	queue_work_on(cpu, system_highpri_wq, &unbind_work);
4766
4767	/* update NUMA affinity of unbound workqueues */
4768	mutex_lock(&wq_pool_mutex);
4769	list_for_each_entry(wq, &workqueues, list)
4770	wq_update_unbound_numa(wq, cpu, false);
4771	mutex_unlock(&wq_pool_mutex);
4772
4773	/* wait for per-cpu unbinding to finish */
4774	flush_work(&unbind_work);
4775	destroy_work_on_stack(&unbind_work);
4776	return 0;
4777	}
4778
4779	#ifdef CONFIG_SMP
4780
4781	struct work_for_cpu {
4782	struct work_struct work;
4783	long (*fn)(void *);
4784	void *arg;
4785	long ret;
4786	};
4787
4788	static void work_for_cpu_fn(struct work_struct *work)
4789	{
4790	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4791
4792	wfc->ret = wfc->fn(wfc->arg);
4793	}
4794
4795	/**
4796	* work_on_cpu - run a function in thread context on a particular cpu
4797	* @cpu: the cpu to run on
4798	* @fn: the function to run
4799	* @arg: the function arg
4800	*
4801	* It is up to the caller to ensure that the cpu doesn't go offline.
4802	* The caller must not hold any locks which would prevent @fn from completing.
4803	*
4804	* Return: The value @fn returns.
4805	*/
4806	long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4807	{
4808	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4809
4810	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4811	schedule_work_on(cpu, &wfc.work);
4812	flush_work(&wfc.work);
4813	destroy_work_on_stack(&wfc.work);
4814	return wfc.ret;
4815	}
4816	EXPORT_SYMBOL_GPL(work_on_cpu);
4817	#endif /* CONFIG_SMP */
4818
4819	#ifdef CONFIG_FREEZER
4820
4821	/**
4822	* freeze_workqueues_begin - begin freezing workqueues
4823	*
4824	* Start freezing workqueues. After this function returns, all freezable
4825	* workqueues will queue new works to their delayed_works list instead of
4826	* pool->worklist.
4827	*
4828	* CONTEXT:
4829	* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4830	*/
4831	void freeze_workqueues_begin(void)
4832	{
4833	struct workqueue_struct *wq;
4834	struct pool_workqueue *pwq;
4835
4836	mutex_lock(&wq_pool_mutex);
4837
4838	WARN_ON_ONCE(workqueue_freezing);
4839	workqueue_freezing = true;
4840
4841	list_for_each_entry(wq, &workqueues, list) {
4842	mutex_lock(&wq->mutex);
4843	for_each_pwq(pwq, wq)
4844	pwq_adjust_max_active(pwq);
4845	mutex_unlock(&wq->mutex);
4846	}
4847
4848	mutex_unlock(&wq_pool_mutex);
4849	}
4850
4851	/**
4852	* freeze_workqueues_busy - are freezable workqueues still busy?
4853	*
4854	* Check whether freezing is complete. This function must be called
4855	* between freeze_workqueues_begin() and thaw_workqueues().
4856	*
4857	* CONTEXT:
4858	* Grabs and releases wq_pool_mutex.
4859	*
4860	* Return:
4861	* %true if some freezable workqueues are still busy. %false if freezing
4862	* is complete.
4863	*/
4864	bool freeze_workqueues_busy(void)
4865	{
4866	bool busy = false;
4867	struct workqueue_struct *wq;
4868	struct pool_workqueue *pwq;
4869
4870	mutex_lock(&wq_pool_mutex);
4871
4872	WARN_ON_ONCE(!workqueue_freezing);
4873
4874	list_for_each_entry(wq, &workqueues, list) {
4875	if (!(wq->flags & WQ_FREEZABLE))
4876	continue;
4877	/*
4878	* nr_active is monotonically decreasing. It's safe
4879	* to peek without lock.
4880	*/
4881	rcu_read_lock_sched();
4882	for_each_pwq(pwq, wq) {
4883	WARN_ON_ONCE(pwq->nr_active < 0);
4884	if (pwq->nr_active) {
4885	busy = true;
4886	rcu_read_unlock_sched();
4887	goto out_unlock;
4888	}
4889	}
4890	rcu_read_unlock_sched();
4891	}
4892	out_unlock:
4893	mutex_unlock(&wq_pool_mutex);
4894	return busy;
4895	}
4896
4897	/**
4898	* thaw_workqueues - thaw workqueues
4899	*
4900	* Thaw workqueues. Normal queueing is restored and all collected
4901	* frozen works are transferred to their respective pool worklists.
4902	*
4903	* CONTEXT:
4904	* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4905	*/
4906	void thaw_workqueues(void)
4907	{
4908	struct workqueue_struct *wq;
4909	struct pool_workqueue *pwq;
4910
4911	mutex_lock(&wq_pool_mutex);
4912
4913	if (!workqueue_freezing)
4914	goto out_unlock;
4915
4916	workqueue_freezing = false;
4917
4918	/* restore max_active and repopulate worklist */
4919	list_for_each_entry(wq, &workqueues, list) {
4920	mutex_lock(&wq->mutex);
4921	for_each_pwq(pwq, wq)
4922	pwq_adjust_max_active(pwq);
4923	mutex_unlock(&wq->mutex);
4924	}
4925
4926	out_unlock:
4927	mutex_unlock(&wq_pool_mutex);
4928	}
4929	#endif /* CONFIG_FREEZER */
4930
4931	static int workqueue_apply_unbound_cpumask(void)
4932	{
4933	LIST_HEAD(ctxs);
4934	int ret = 0;
4935	struct workqueue_struct *wq;
4936	struct apply_wqattrs_ctx *ctx, *n;
4937
4938	lockdep_assert_held(&wq_pool_mutex);
4939
4940	list_for_each_entry(wq, &workqueues, list) {
4941	if (!(wq->flags & WQ_UNBOUND))
4942	continue;
4943	/* creating multiple pwqs breaks ordering guarantee */
4944	if (wq->flags & __WQ_ORDERED)
4945	continue;
4946
4947	ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4948	if (!ctx) {
4949	ret = -ENOMEM;
4950	break;
4951	}
4952
4953	list_add_tail(&ctx->list, &ctxs);
4954	}
4955
4956	list_for_each_entry_safe(ctx, n, &ctxs, list) {
4957	if (!ret)
4958	apply_wqattrs_commit(ctx);
4959	apply_wqattrs_cleanup(ctx);
4960	}
4961
4962	return ret;
4963	}
4964
4965	/**
4966	* workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4967	* @cpumask: the cpumask to set
4968	*
4969	* The low-level workqueues cpumask is a global cpumask that limits
4970	* the affinity of all unbound workqueues. This function check the @cpumask
4971	* and apply it to all unbound workqueues and updates all pwqs of them.
4972	*
4973	* Retun: 0 - Success
4974	* -EINVAL - Invalid @cpumask
4975	* -ENOMEM - Failed to allocate memory for attrs or pwqs.
4976	*/
4977	int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4978	{
4979	int ret = -EINVAL;
4980	cpumask_var_t saved_cpumask;
4981
4982	if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4983	return -ENOMEM;
4984
4985	cpumask_and(cpumask, cpumask, cpu_possible_mask);
4986	if (!cpumask_empty(cpumask)) {
4987	apply_wqattrs_lock();
4988
4989	/* save the old wq_unbound_cpumask. */
4990	cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4991
4992	/* update wq_unbound_cpumask at first and apply it to wqs. */
4993	cpumask_copy(wq_unbound_cpumask, cpumask);
4994	ret = workqueue_apply_unbound_cpumask();
4995
4996	/* restore the wq_unbound_cpumask when failed. */
4997	if (ret < 0)
4998	cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4999
5000	apply_wqattrs_unlock();
5001	}
5002
5003	free_cpumask_var(saved_cpumask);
5004	return ret;
5005	}
5006
5007	#ifdef CONFIG_SYSFS
5008	/*
5009	* Workqueues with WQ_SYSFS flag set is visible to userland via
5010	* /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5011	* following attributes.
5012	*
5013	* per_cpu RO bool : whether the workqueue is per-cpu or unbound
5014	* max_active RW int : maximum number of in-flight work items
5015	*
5016	* Unbound workqueues have the following extra attributes.
5017	*
5018	* id RO int : the associated pool ID
5019	* nice RW int : nice value of the workers
5020	* cpumask RW mask : bitmask of allowed CPUs for the workers
5021	*/
5022	struct wq_device {
5023	struct workqueue_struct *wq;
5024	struct device dev;
5025	};
5026
5027	static struct workqueue_struct *dev_to_wq(struct device *dev)
5028	{
5029	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5030
5031	return wq_dev->wq;
5032	}
5033
5034	static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5035	char *buf)
5036	{
5037	struct workqueue_struct *wq = dev_to_wq(dev);
5038
5039	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5040	}
5041	static DEVICE_ATTR_RO(per_cpu);
5042
5043	static ssize_t max_active_show(struct device *dev,
5044	struct device_attribute *attr, char *buf)
5045	{
5046	struct workqueue_struct *wq = dev_to_wq(dev);
5047
5048	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5049	}
5050
5051	static ssize_t max_active_store(struct device *dev,
5052	struct device_attribute *attr, const char *buf,
5053	size_t count)
5054	{
5055	struct workqueue_struct *wq = dev_to_wq(dev);
5056	int val;
5057
5058	if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5059	return -EINVAL;
5060
5061	workqueue_set_max_active(wq, val);
5062	return count;
5063	}
5064	static DEVICE_ATTR_RW(max_active);
5065
5066	static struct attribute *wq_sysfs_attrs[] = {
5067	&dev_attr_per_cpu.attr,
5068	&dev_attr_max_active.attr,
5069	NULL,
5070	};
5071	ATTRIBUTE_GROUPS(wq_sysfs);
5072
5073	static ssize_t wq_pool_ids_show(struct device *dev,
5074	struct device_attribute *attr, char *buf)
5075	{
5076	struct workqueue_struct *wq = dev_to_wq(dev);
5077	const char *delim = "";
5078	int node, written = 0;
5079
5080	rcu_read_lock_sched();
5081	for_each_node(node) {
5082	written += scnprintf(buf + written, PAGE_SIZE - written,
5083	"%s%d:%d", delim, node,
5084	unbound_pwq_by_node(wq, node)->pool->id);
5085	delim = " ";
5086	}
5087	written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5088	rcu_read_unlock_sched();
5089
5090	return written;
5091	}
5092
5093	static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5094	char *buf)
5095	{
5096	struct workqueue_struct *wq = dev_to_wq(dev);
5097	int written;
5098
5099	mutex_lock(&wq->mutex);
5100	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5101	mutex_unlock(&wq->mutex);
5102
5103	return written;
5104	}
5105
5106	/* prepare workqueue_attrs for sysfs store operations */
5107	static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5108	{
5109	struct workqueue_attrs *attrs;
5110
5111	lockdep_assert_held(&wq_pool_mutex);
5112
5113	attrs = alloc_workqueue_attrs(GFP_KERNEL);
5114	if (!attrs)
5115	return NULL;
5116
5117	copy_workqueue_attrs(attrs, wq->unbound_attrs);
5118	return attrs;
5119	}
5120
5121	static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5122	const char *buf, size_t count)
5123	{
5124	struct workqueue_struct *wq = dev_to_wq(dev);
5125	struct workqueue_attrs *attrs;
5126	int ret = -ENOMEM;
5127
5128	apply_wqattrs_lock();
5129
5130	attrs = wq_sysfs_prep_attrs(wq);
5131	if (!attrs)
5132	goto out_unlock;
5133
5134	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5135	attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5136	ret = apply_workqueue_attrs_locked(wq, attrs);
5137	else
5138	ret = -EINVAL;
5139
5140	out_unlock:
5141	apply_wqattrs_unlock();
5142	free_workqueue_attrs(attrs);
5143	return ret ?: count;
5144	}
5145
5146	static ssize_t wq_cpumask_show(struct device *dev,
5147	struct device_attribute *attr, char *buf)
5148	{
5149	struct workqueue_struct *wq = dev_to_wq(dev);
5150	int written;
5151
5152	mutex_lock(&wq->mutex);
5153	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5154	cpumask_pr_args(wq->unbound_attrs->cpumask));
5155	mutex_unlock(&wq->mutex);
5156	return written;
5157	}
5158
5159	static ssize_t wq_cpumask_store(struct device *dev,
5160	struct device_attribute *attr,
5161	const char *buf, size_t count)
5162	{
5163	struct workqueue_struct *wq = dev_to_wq(dev);
5164	struct workqueue_attrs *attrs;
5165	int ret = -ENOMEM;
5166
5167	apply_wqattrs_lock();
5168
5169	attrs = wq_sysfs_prep_attrs(wq);
5170	if (!attrs)
5171	goto out_unlock;
5172
5173	ret = cpumask_parse(buf, attrs->cpumask);
5174	if (!ret)
5175	ret = apply_workqueue_attrs_locked(wq, attrs);
5176
5177	out_unlock:
5178	apply_wqattrs_unlock();
5179	free_workqueue_attrs(attrs);
5180	return ret ?: count;
5181	}
5182
5183	static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5184	char *buf)
5185	{
5186	struct workqueue_struct *wq = dev_to_wq(dev);
5187	int written;
5188
5189	mutex_lock(&wq->mutex);
5190	written = scnprintf(buf, PAGE_SIZE, "%d\n",
5191	!wq->unbound_attrs->no_numa);
5192	mutex_unlock(&wq->mutex);
5193
5194	return written;
5195	}
5196
5197	static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5198	const char *buf, size_t count)
5199	{
5200	struct workqueue_struct *wq = dev_to_wq(dev);
5201	struct workqueue_attrs *attrs;
5202	int v, ret = -ENOMEM;
5203
5204	apply_wqattrs_lock();
5205
5206	attrs = wq_sysfs_prep_attrs(wq);
5207	if (!attrs)
5208	goto out_unlock;
5209
5210	ret = -EINVAL;
5211	if (sscanf(buf, "%d", &v) == 1) {
5212	attrs->no_numa = !v;
5213	ret = apply_workqueue_attrs_locked(wq, attrs);
5214	}
5215
5216	out_unlock:
5217	apply_wqattrs_unlock();
5218	free_workqueue_attrs(attrs);
5219	return ret ?: count;
5220	}
5221
5222	static struct device_attribute wq_sysfs_unbound_attrs[] = {
5223	__ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5224	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5225	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5226	__ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5227	__ATTR_NULL,
5228	};
5229
5230	static struct bus_type wq_subsys = {
5231	.name = "workqueue",
5232	.dev_groups = wq_sysfs_groups,
5233	};
5234
5235	static ssize_t wq_unbound_cpumask_show(struct device *dev,
5236	struct device_attribute *attr, char *buf)
5237	{
5238	int written;
5239
5240	mutex_lock(&wq_pool_mutex);
5241	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5242	cpumask_pr_args(wq_unbound_cpumask));
5243	mutex_unlock(&wq_pool_mutex);
5244
5245	return written;
5246	}
5247
5248	static ssize_t wq_unbound_cpumask_store(struct device *dev,
5249	struct device_attribute *attr, const char *buf, size_t count)
5250	{
5251	cpumask_var_t cpumask;
5252	int ret;
5253
5254	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5255	return -ENOMEM;
5256
5257	ret = cpumask_parse(buf, cpumask);
5258	if (!ret)
5259	ret = workqueue_set_unbound_cpumask(cpumask);
5260
5261	free_cpumask_var(cpumask);
5262	return ret ? ret : count;
5263	}
5264
5265	static struct device_attribute wq_sysfs_cpumask_attr =
5266	__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5267	wq_unbound_cpumask_store);
5268
5269	static int __init wq_sysfs_init(void)
5270	{
5271	int err;
5272
5273	err = subsys_virtual_register(&wq_subsys, NULL);
5274	if (err)
5275	return err;
5276
5277	return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5278	}
5279	core_initcall(wq_sysfs_init);
5280
5281	static void wq_device_release(struct device *dev)
5282	{
5283	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5284
5285	kfree(wq_dev);
5286	}
5287
5288	/**
5289	* workqueue_sysfs_register - make a workqueue visible in sysfs
5290	* @wq: the workqueue to register
5291	*
5292	* Expose @wq in sysfs under /sys/bus/workqueue/devices.
5293	* alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5294	* which is the preferred method.
5295	*
5296	* Workqueue user should use this function directly iff it wants to apply
5297	* workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5298	* apply_workqueue_attrs() may race against userland updating the
5299	* attributes.
5300	*
5301	* Return: 0 on success, -errno on failure.
5302	*/
5303	int workqueue_sysfs_register(struct workqueue_struct *wq)
5304	{
5305	struct wq_device *wq_dev;
5306	int ret;
5307
5308	/*
5309	* Adjusting max_active or creating new pwqs by applying
5310	* attributes breaks ordering guarantee. Disallow exposing ordered
5311	* workqueues.
5312	*/
5313	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5314	return -EINVAL;
5315
5316	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5317	if (!wq_dev)
5318	return -ENOMEM;
5319
5320	wq_dev->wq = wq;
5321	wq_dev->dev.bus = &wq_subsys;
5322	wq_dev->dev.release = wq_device_release;
5323	dev_set_name(&wq_dev->dev, "%s", wq->name);
5324
5325	/*
5326	* unbound_attrs are created separately. Suppress uevent until
5327	* everything is ready.
5328	*/
5329	dev_set_uevent_suppress(&wq_dev->dev, true);
5330
5331	ret = device_register(&wq_dev->dev);
5332	if (ret) {
5333	put_device(&wq_dev->dev);
5334	wq->wq_dev = NULL;
5335	return ret;
5336	}
5337
5338	if (wq->flags & WQ_UNBOUND) {
5339	struct device_attribute *attr;
5340
5341	for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5342	ret = device_create_file(&wq_dev->dev, attr);
5343	if (ret) {
5344	device_unregister(&wq_dev->dev);
5345	wq->wq_dev = NULL;
5346	return ret;
5347	}
5348	}
5349	}
5350
5351	dev_set_uevent_suppress(&wq_dev->dev, false);
5352	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5353	return 0;
5354	}
5355
5356	/**
5357	* workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5358	* @wq: the workqueue to unregister
5359	*
5360	* If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5361	*/
5362	static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5363	{
5364	struct wq_device *wq_dev = wq->wq_dev;
5365
5366	if (!wq->wq_dev)
5367	return;
5368
5369	wq->wq_dev = NULL;
5370	device_unregister(&wq_dev->dev);
5371	}
5372	#else /* CONFIG_SYSFS */
5373	static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5374	#endif /* CONFIG_SYSFS */
5375
5376	/*
5377	* Workqueue watchdog.
5378	*
5379	* Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5380	* flush dependency, a concurrency managed work item which stays RUNNING
5381	* indefinitely. Workqueue stalls can be very difficult to debug as the
5382	* usual warning mechanisms don't trigger and internal workqueue state is
5383	* largely opaque.
5384	*
5385	* Workqueue watchdog monitors all worker pools periodically and dumps
5386	* state if some pools failed to make forward progress for a while where
5387	* forward progress is defined as the first item on ->worklist changing.
5388	*
5389	* This mechanism is controlled through the kernel parameter
5390	* "workqueue.watchdog_thresh" which can be updated at runtime through the
5391	* corresponding sysfs parameter file.
5392	*/
5393	#ifdef CONFIG_WQ_WATCHDOG
5394
5395	static void wq_watchdog_timer_fn(unsigned long data);
5396
5397	static unsigned long wq_watchdog_thresh = 30;
5398	static struct timer_list wq_watchdog_timer =
5399	TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5400
5401	static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5402	static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5403
5404	static void wq_watchdog_reset_touched(void)
5405	{
5406	int cpu;
5407
5408	wq_watchdog_touched = jiffies;
5409	for_each_possible_cpu(cpu)
5410	per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5411	}
5412
5413	static void wq_watchdog_timer_fn(unsigned long data)
5414	{
5415	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5416	bool lockup_detected = false;
5417	struct worker_pool *pool;
5418	int pi;
5419
5420	if (!thresh)
5421	return;
5422
5423	rcu_read_lock();
5424
5425	for_each_pool(pool, pi) {
5426	unsigned long pool_ts, touched, ts;
5427
5428	if (list_empty(&pool->worklist))
5429	continue;
5430
5431	/* get the latest of pool and touched timestamps */
5432	pool_ts = READ_ONCE(pool->watchdog_ts);
5433	touched = READ_ONCE(wq_watchdog_touched);
5434
5435	if (time_after(pool_ts, touched))
5436	ts = pool_ts;
5437	else
5438	ts = touched;
5439
5440	if (pool->cpu >= 0) {
5441	unsigned long cpu_touched =
5442	READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5443	pool->cpu));
5444	if (time_after(cpu_touched, ts))
5445	ts = cpu_touched;
5446	}
5447
5448	/* did we stall? */
5449	if (time_after(jiffies, ts + thresh)) {
5450	lockup_detected = true;
5451	pr_emerg("BUG: workqueue lockup - pool");
5452	pr_cont_pool_info(pool);
5453	pr_cont(" stuck for %us!\n",
5454	jiffies_to_msecs(jiffies - pool_ts) / 1000);
5455	}
5456	}
5457
5458	rcu_read_unlock();
5459
5460	if (lockup_detected)
5461	show_workqueue_state();
5462
5463	wq_watchdog_reset_touched();
5464	mod_timer(&wq_watchdog_timer, jiffies + thresh);
5465	}
5466
5467	void wq_watchdog_touch(int cpu)
5468	{
5469	if (cpu >= 0)
5470	per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5471	else
5472	wq_watchdog_touched = jiffies;
5473	}
5474
5475	static void wq_watchdog_set_thresh(unsigned long thresh)
5476	{
5477	wq_watchdog_thresh = 0;
5478	del_timer_sync(&wq_watchdog_timer);
5479
5480	if (thresh) {
5481	wq_watchdog_thresh = thresh;
5482	wq_watchdog_reset_touched();
5483	mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5484	}
5485	}
5486
5487	static int wq_watchdog_param_set_thresh(const char *val,
5488	const struct kernel_param *kp)
5489	{
5490	unsigned long thresh;
5491	int ret;
5492
5493	ret = kstrtoul(val, 0, &thresh);
5494	if (ret)
5495	return ret;
5496
5497	if (system_wq)
5498	wq_watchdog_set_thresh(thresh);
5499	else
5500	wq_watchdog_thresh = thresh;
5501
5502	return 0;
5503	}
5504
5505	static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5506	.set = wq_watchdog_param_set_thresh,
5507	.get = param_get_ulong,
5508	};
5509
5510	module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5511	0644);
5512
5513	static void wq_watchdog_init(void)
5514	{
5515	wq_watchdog_set_thresh(wq_watchdog_thresh);
5516	}
5517
5518	#else /* CONFIG_WQ_WATCHDOG */
5519
5520	static inline void wq_watchdog_init(void) { }
5521
5522	#endif /* CONFIG_WQ_WATCHDOG */
5523
5524	static void __init wq_numa_init(void)
5525	{
5526	cpumask_var_t *tbl;
5527	int node, cpu;
5528
5529	if (num_possible_nodes() <= 1)
5530	return;
5531
5532	if (wq_disable_numa) {
5533	pr_info("workqueue: NUMA affinity support disabled\n");
5534	return;
5535	}
5536
5537	wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5538	BUG_ON(!wq_update_unbound_numa_attrs_buf);
5539
5540	/*
5541	* We want masks of possible CPUs of each node which isn't readily
5542	* available. Build one from cpu_to_node() which should have been
5543	* fully initialized by now.
5544	*/
5545	tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5546	BUG_ON(!tbl);
5547
5548	for_each_node(node)
5549	BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5550	node_online(node) ? node : NUMA_NO_NODE));
5551
5552	for_each_possible_cpu(cpu) {
5553	node = cpu_to_node(cpu);
5554	if (WARN_ON(node == NUMA_NO_NODE)) {
5555	pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5556	/* happens iff arch is bonkers, let's just proceed */
5557	return;
5558	}
5559	cpumask_set_cpu(cpu, tbl[node]);
5560	}
5561
5562	wq_numa_possible_cpumask = tbl;
5563	wq_numa_enabled = true;
5564	}
5565
5566	/**
5567	* workqueue_init_early - early init for workqueue subsystem
5568	*
5569	* This is the first half of two-staged workqueue subsystem initialization
5570	* and invoked as soon as the bare basics - memory allocation, cpumasks and
5571	* idr are up. It sets up all the data structures and system workqueues
5572	* and allows early boot code to create workqueues and queue/cancel work
5573	* items. Actual work item execution starts only after kthreads can be
5574	* created and scheduled right before early initcalls.
5575	*/
5576	int __init workqueue_init_early(void)
5577	{
5578	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5579	int i, cpu;
5580
5581	WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5582
5583	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5584	cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5585
5586	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5587
5588	wq_numa_init();
5589
5590	/* initialize CPU pools */
5591	for_each_possible_cpu(cpu) {
5592	struct worker_pool *pool;
5593
5594	i = 0;
5595	for_each_cpu_worker_pool(pool, cpu) {
5596	BUG_ON(init_worker_pool(pool));
5597	pool->cpu = cpu;
5598	cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5599	pool->attrs->nice = std_nice[i++];
5600	pool->node = cpu_to_node(cpu);
5601
5602	/* alloc pool ID */
5603	mutex_lock(&wq_pool_mutex);
5604	BUG_ON(worker_pool_assign_id(pool));
5605	mutex_unlock(&wq_pool_mutex);
5606	}
5607	}
5608
5609	/* create default unbound and ordered wq attrs */
5610	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5611	struct workqueue_attrs *attrs;
5612
5613	BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5614	attrs->nice = std_nice[i];
5615	unbound_std_wq_attrs[i] = attrs;
5616
5617	/*
5618	* An ordered wq should have only one pwq as ordering is
5619	* guaranteed by max_active which is enforced by pwqs.
5620	* Turn off NUMA so that dfl_pwq is used for all nodes.
5621	*/
5622	BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5623	attrs->nice = std_nice[i];
5624	attrs->no_numa = true;
5625	ordered_wq_attrs[i] = attrs;
5626	}
5627
5628	system_wq = alloc_workqueue("events", 0, 0);
5629	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5630	system_long_wq = alloc_workqueue("events_long", 0, 0);
5631	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5632	WQ_UNBOUND_MAX_ACTIVE);
5633	system_freezable_wq = alloc_workqueue("events_freezable",
5634	WQ_FREEZABLE, 0);
5635	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5636	WQ_POWER_EFFICIENT, 0);
5637	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5638	WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5639	0);
5640	BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5641	!system_unbound_wq || !system_freezable_wq ||
5642	!system_power_efficient_wq ||
5643	!system_freezable_power_efficient_wq);
5644
5645	return 0;
5646	}
5647
5648	/**
5649	* workqueue_init - bring workqueue subsystem fully online
5650	*
5651	* This is the latter half of two-staged workqueue subsystem initialization
5652	* and invoked as soon as kthreads can be created and scheduled.
5653	* Workqueues have been created and work items queued on them, but there
5654	* are no kworkers executing the work items yet. Populate the worker pools
5655	* with the initial workers and enable future kworker creations.
5656	*/
5657	int __init workqueue_init(void)
5658	{
5659	struct worker_pool *pool;
5660	int cpu, bkt;
5661
5662	/* create the initial workers */
5663	for_each_online_cpu(cpu) {
5664	for_each_cpu_worker_pool(pool, cpu) {
5665	pool->flags &= ~POOL_DISASSOCIATED;
5666	BUG_ON(!create_worker(pool));
5667	}
5668	}
5669
5670	hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5671	BUG_ON(!create_worker(pool));
5672
5673	wq_online = true;
5674	wq_watchdog_init();
5675
5676	return 0;
5677	}
5678