linux - kernel/workqueue.c

6805 lines · 188.5 KiB

1	// SPDX-License-Identifier: GPL-2.0-only
2	/*
3	* kernel/workqueue.c - generic async execution with shared worker pool
4	*
5	* Copyright (C) 2002 Ingo Molnar
6	*
7	* Derived from the taskqueue/keventd code by:
8	* David Woodhouse <[email protected]>
9	* Andrew Morton
10	* Kai Petzke <[email protected]>
11	* Theodore Ts'o <[email protected]>
12	*
13	* Made to use alloc_percpu by Christoph Lameter.
14	*
15	* Copyright (C) 2010 SUSE Linux Products GmbH
16	* Copyright (C) 2010 Tejun Heo <[email protected]>
17	*
18	* This is the generic async execution mechanism. Work items as are
19	* executed in process context. The worker pool is shared and
20	* automatically managed. There are two worker pools for each CPU (one for
21	* normal work items and the other for high priority ones) and some extra
22	* pools for workqueues which are not bound to any specific CPU - the
23	* number of these backing pools is dynamic.
24	*
25	* Please read Documentation/core-api/workqueue.rst for details.
26	*/
27
28	#include <linux/export.h>
29	#include <linux/kernel.h>
30	#include <linux/sched.h>
31	#include <linux/init.h>
32	#include <linux/signal.h>
33	#include <linux/completion.h>
34	#include <linux/workqueue.h>
35	#include <linux/slab.h>
36	#include <linux/cpu.h>
37	#include <linux/notifier.h>
38	#include <linux/kthread.h>
39	#include <linux/hardirq.h>
40	#include <linux/mempolicy.h>
41	#include <linux/freezer.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/sched/isolation.h>
52	#include <linux/sched/debug.h>
53	#include <linux/nmi.h>
54	#include <linux/kvm_para.h>
55	#include <linux/delay.h>
56
57	#include "workqueue_internal.h"
58
59	enum {
60	/*
61	* worker_pool flags
62	*
63	* A bound pool is either associated or disassociated with its CPU.
64	* While associated (!DISASSOCIATED), all workers are bound to the
65	* CPU and none has %WORKER_UNBOUND set and concurrency management
66	* is in effect.
67	*
68	* While DISASSOCIATED, the cpu may be offline and all workers have
69	* %WORKER_UNBOUND set and concurrency management disabled, and may
70	* be executing on any CPU. The pool behaves as an unbound one.
71	*
72	* Note that DISASSOCIATED should be flipped only while holding
73	* wq_pool_attach_mutex to avoid changing binding state while
74	* worker_attach_to_pool() is in progress.
75	*/
76	POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
77	POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
78
79	/* worker flags */
80	WORKER_DIE = 1 << 1, /* die die die */
81	WORKER_IDLE = 1 << 2, /* is idle */
82	WORKER_PREP = 1 << 3, /* preparing to run works */
83	WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
84	WORKER_UNBOUND = 1 << 7, /* worker is unbound */
85	WORKER_REBOUND = 1 << 8, /* worker was rebound */
86
87	WORKER_NOT_RUNNING = WORKER_PREP \| WORKER_CPU_INTENSIVE \|
88	WORKER_UNBOUND \| WORKER_REBOUND,
89
90	NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
91
92	UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
93	BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
94
95	MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
96	IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
97
98	MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
99	/* call for help after 10ms
100	(min two ticks) */
101	MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
102	CREATE_COOLDOWN = HZ, /* time to breath after fail */
103
104	/*
105	* Rescue workers are used only on emergencies and shared by
106	* all cpus. Give MIN_NICE.
107	*/
108	RESCUER_NICE_LEVEL = MIN_NICE,
109	HIGHPRI_NICE_LEVEL = MIN_NICE,
110
111	WQ_NAME_LEN = 24,
112	};
113
114	/*
115	* Structure fields follow one of the following exclusion rules.
116	*
117	* I: Modifiable by initialization/destruction paths and read-only for
118	* everyone else.
119	*
120	* P: Preemption protected. Disabling preemption is enough and should
121	* only be modified and accessed from the local cpu.
122	*
123	* L: pool->lock protected. Access with pool->lock held.
124	*
125	* K: Only modified by worker while holding pool->lock. Can be safely read by
126	* self, while holding pool->lock or from IRQ context if %current is the
127	* kworker.
128	*
129	* S: Only modified by worker self.
130	*
131	* A: wq_pool_attach_mutex protected.
132	*
133	* PL: wq_pool_mutex protected.
134	*
135	* PR: wq_pool_mutex protected for writes. RCU protected for reads.
136	*
137	* PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
138	*
139	* PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
140	* RCU for reads.
141	*
142	* WQ: wq->mutex protected.
143	*
144	* WR: wq->mutex protected for writes. RCU protected for reads.
145	*
146	* MD: wq_mayday_lock protected.
147	*
148	* WD: Used internally by the watchdog.
149	*/
150
151	/* struct worker is defined in workqueue_internal.h */
152
153	struct worker_pool {
154	raw_spinlock_t lock; /* the pool lock */
155	int cpu; /* I: the associated cpu */
156	int node; /* I: the associated node ID */
157	int id; /* I: pool ID */
158	unsigned int flags; /* L: flags */
159
160	unsigned long watchdog_ts; /* L: watchdog timestamp */
161	bool cpu_stall; /* WD: stalled cpu bound pool */
162
163	/*
164	* The counter is incremented in a process context on the associated CPU
165	* w/ preemption disabled, and decremented or reset in the same context
166	* but w/ pool->lock held. The readers grab pool->lock and are
167	* guaranteed to see if the counter reached zero.
168	*/
169	int nr_running;
170
171	struct list_head worklist; /* L: list of pending works */
172
173	int nr_workers; /* L: total number of workers */
174	int nr_idle; /* L: currently idle workers */
175
176	struct list_head idle_list; /* L: list of idle workers */
177	struct timer_list idle_timer; /* L: worker idle timeout */
178	struct work_struct idle_cull_work; /* L: worker idle cleanup */
179
180	struct timer_list mayday_timer; /* L: SOS timer for workers */
181
182	/* a workers is either on busy_hash or idle_list, or the manager */
183	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
184	/* L: hash of busy workers */
185
186	struct worker manager; / L: purely informational */
187	struct list_head workers; /* A: attached workers */
188	struct list_head dying_workers; /* A: workers about to die */
189	struct completion detach_completion; / all workers detached */
190
191	struct ida worker_ida; /* worker IDs for task name */
192
193	struct workqueue_attrs attrs; / I: worker attributes */
194	struct hlist_node hash_node; /* PL: unbound_pool_hash node */
195	int refcnt; /* PL: refcnt for unbound pools */
196
197	/*
198	* Destruction of pool is RCU protected to allow dereferences
199	* from get_work_pool().
200	*/
201	struct rcu_head rcu;
202	};
203
204	/*
205	* Per-pool_workqueue statistics. These can be monitored using
206	* tools/workqueue/wq_monitor.py.
207	*/
208	enum pool_workqueue_stats {
209	PWQ_STAT_STARTED, /* work items started execution */
210	PWQ_STAT_COMPLETED, /* work items completed execution */
211	PWQ_STAT_CPU_TIME, /* total CPU time consumed */
212	PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
213	PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
214	PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
215	PWQ_STAT_MAYDAY, /* maydays to rescuer */
216	PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
217
218	PWQ_NR_STATS,
219	};
220
221	/*
222	* The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
223	* of work_struct->data are used for flags and the remaining high bits
224	* point to the pwq; thus, pwqs need to be aligned at two's power of the
225	* number of flag bits.
226	*/
227	struct pool_workqueue {
228	struct worker_pool pool; / I: the associated pool */
229	struct workqueue_struct wq; / I: the owning workqueue */
230	int work_color; /* L: current color */
231	int flush_color; /* L: flushing color */
232	int refcnt; /* L: reference count */
233	int nr_in_flight[WORK_NR_COLORS];
234	/* L: nr of in_flight works */
235
236	/*
237	* nr_active management and WORK_STRUCT_INACTIVE:
238	*
239	* When pwq->nr_active >= max_active, new work item is queued to
240	* pwq->inactive_works instead of pool->worklist and marked with
241	* WORK_STRUCT_INACTIVE.
242	*
243	* All work items marked with WORK_STRUCT_INACTIVE do not participate
244	* in pwq->nr_active and all work items in pwq->inactive_works are
245	* marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE
246	* work items are in pwq->inactive_works. Some of them are ready to
247	* run in pool->worklist or worker->scheduled. Those work itmes are
248	* only struct wq_barrier which is used for flush_work() and should
249	* not participate in pwq->nr_active. For non-barrier work item, it
250	* is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
251	*/
252	int nr_active; /* L: nr of active works */
253	int max_active; /* L: max active works */
254	struct list_head inactive_works; /* L: inactive works */
255	struct list_head pwqs_node; /* WR: node on wq->pwqs */
256	struct list_head mayday_node; /* MD: node on wq->maydays */
257
258	u64 stats[PWQ_NR_STATS];
259
260	/*
261	* Release of unbound pwq is punted to a kthread_worker. See put_pwq()
262	* and pwq_release_workfn() for details. pool_workqueue itself is also
263	* RCU protected so that the first pwq can be determined without
264	* grabbing wq->mutex.
265	*/
266	struct kthread_work release_work;
267	struct rcu_head rcu;
268	} __aligned(1 << WORK_STRUCT_FLAG_BITS);
269
270	/*
271	* Structure used to wait for workqueue flush.
272	*/
273	struct wq_flusher {
274	struct list_head list; /* WQ: list of flushers */
275	int flush_color; /* WQ: flush color waiting for */
276	struct completion done; /* flush completion */
277	};
278
279	struct wq_device;
280
281	/*
282	* The externally visible workqueue. It relays the issued work items to
283	* the appropriate worker_pool through its pool_workqueues.
284	*/
285	struct workqueue_struct {
286	struct list_head pwqs; /* WR: all pwqs of this wq */
287	struct list_head list; /* PR: list of all workqueues */
288
289	struct mutex mutex; /* protects this wq */
290	int work_color; /* WQ: current work color */
291	int flush_color; /* WQ: current flush color */
292	atomic_t nr_pwqs_to_flush; /* flush in progress */
293	struct wq_flusher first_flusher; / WQ: first flusher */
294	struct list_head flusher_queue; /* WQ: flush waiters */
295	struct list_head flusher_overflow; /* WQ: flush overflow list */
296
297	struct list_head maydays; /* MD: pwqs requesting rescue */
298	struct worker rescuer; / MD: rescue worker */
299
300	int nr_drainers; /* WQ: drain in progress */
301	int saved_max_active; /* WQ: saved pwq max_active */
302
303	struct workqueue_attrs unbound_attrs; / PW: only for unbound wqs */
304	struct pool_workqueue dfl_pwq; / PW: only for unbound wqs */
305
306	#ifdef CONFIG_SYSFS
307	struct wq_device wq_dev; / I: for sysfs interface */
308	#endif
309	#ifdef CONFIG_LOCKDEP
310	char *lock_name;
311	struct lock_class_key key;
312	struct lockdep_map lockdep_map;
313	#endif
314	char name[WQ_NAME_LEN]; /* I: workqueue name */
315
316	/*
317	* Destruction of workqueue_struct is RCU protected to allow walking
318	* the workqueues list without grabbing wq_pool_mutex.
319	* This is used to dump all workqueues from sysrq.
320	*/
321	struct rcu_head rcu;
322
323	/* hot fields used during command issue, aligned to cacheline */
324	unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
325	struct pool_workqueue __percpu __rcu *cpu_pwq; / I: per-cpu pwqs */
326	};
327
328	static struct kmem_cache *pwq_cache;
329
330	/*
331	* Each pod type describes how CPUs should be grouped for unbound workqueues.
332	* See the comment above workqueue_attrs->affn_scope.
333	*/
334	struct wq_pod_type {
335	int nr_pods; /* number of pods */
336	cpumask_var_t pod_cpus; / pod -> cpus */
337	int pod_node; / pod -> node */
338	int cpu_pod; / cpu -> pod */
339	};
340
341	static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
342	static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
343
344	static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
345	[WQ_AFFN_DFL] = "default",
346	[WQ_AFFN_CPU] = "cpu",
347	[WQ_AFFN_SMT] = "smt",
348	[WQ_AFFN_CACHE] = "cache",
349	[WQ_AFFN_NUMA] = "numa",
350	[WQ_AFFN_SYSTEM] = "system",
351	};
352
353	/*
354	* Per-cpu work items which run for longer than the following threshold are
355	* automatically considered CPU intensive and excluded from concurrency
356	* management to prevent them from noticeably delaying other per-cpu work items.
357	* ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
358	* The actual value is initialized in wq_cpu_intensive_thresh_init().
359	*/
360	static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
361	module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
362
363	/* see the comment above the definition of WQ_POWER_EFFICIENT */
364	static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
365	module_param_named(power_efficient, wq_power_efficient, bool, 0444);
366
367	static bool wq_online; /* can kworkers be created yet? */
368
369	/* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
370	static struct workqueue_attrs *wq_update_pod_attrs_buf;
371
372	static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
373	static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
374	static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
375	/* wait for manager to go away */
376	static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
377
378	static LIST_HEAD(workqueues); /* PR: list of all workqueues */
379	static bool workqueue_freezing; /* PL: have wqs started freezing? */
380
381	/* PL&A: allowable cpus for unbound wqs and work items */
382	static cpumask_var_t wq_unbound_cpumask;
383
384	/* for further constrain wq_unbound_cpumask by cmdline parameter*/
385	static struct cpumask wq_cmdline_cpumask __initdata;
386
387	/* CPU where unbound work was last round robin scheduled from this CPU */
388	static DEFINE_PER_CPU(int, wq_rr_cpu_last);
389
390	/*
391	* Local execution of unbound work items is no longer guaranteed. The
392	* following always forces round-robin CPU selection on unbound work items
393	* to uncover usages which depend on it.
394	*/
395	#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
396	static bool wq_debug_force_rr_cpu = true;
397	#else
398	static bool wq_debug_force_rr_cpu = false;
399	#endif
400	module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
401
402	/* the per-cpu worker pools */
403	static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
404
405	static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
406
407	/* PL: hash of all unbound pools keyed by pool->attrs */
408	static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
409
410	/* I: attributes used when instantiating standard unbound pools on demand */
411	static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
412
413	/* I: attributes used when instantiating ordered pools on demand */
414	static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
415
416	/*
417	* I: kthread_worker to release pwq's. pwq release needs to be bounced to a
418	* process context while holding a pool lock. Bounce to a dedicated kthread
419	* worker to avoid A-A deadlocks.
420	*/
421	static struct kthread_worker *pwq_release_worker;
422
423	struct workqueue_struct *system_wq __read_mostly;
424	EXPORT_SYMBOL(system_wq);
425	struct workqueue_struct *system_highpri_wq __read_mostly;
426	EXPORT_SYMBOL_GPL(system_highpri_wq);
427	struct workqueue_struct *system_long_wq __read_mostly;
428	EXPORT_SYMBOL_GPL(system_long_wq);
429	struct workqueue_struct *system_unbound_wq __read_mostly;
430	EXPORT_SYMBOL_GPL(system_unbound_wq);
431	struct workqueue_struct *system_freezable_wq __read_mostly;
432	EXPORT_SYMBOL_GPL(system_freezable_wq);
433	struct workqueue_struct *system_power_efficient_wq __read_mostly;
434	EXPORT_SYMBOL_GPL(system_power_efficient_wq);
435	struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
436	EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
437
438	static int worker_thread(void *__worker);
439	static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
440	static void show_pwq(struct pool_workqueue *pwq);
441	static void show_one_worker_pool(struct worker_pool *pool);
442
443	#define CREATE_TRACE_POINTS
444	#include <trace/events/workqueue.h>
445
446	#define assert_rcu_or_pool_mutex() \
447	RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
448	!lockdep_is_held(&wq_pool_mutex), \
449	"RCU or wq_pool_mutex should be held")
450
451	#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
452	RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
453	!lockdep_is_held(&wq->mutex) && \
454	!lockdep_is_held(&wq_pool_mutex), \
455	"RCU, wq->mutex or wq_pool_mutex should be held")
456
457	#define for_each_cpu_worker_pool(pool, cpu) \
458	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
459	(pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
460	(pool)++)
461
462	/**
463	* for_each_pool - iterate through all worker_pools in the system
464	* @pool: iteration cursor
465	* @pi: integer used for iteration
466	*
467	* This must be called either with wq_pool_mutex held or RCU read
468	* locked. If the pool needs to be used beyond the locking in effect, the
469	* caller is responsible for guaranteeing that the pool stays online.
470	*
471	* The if/else clause exists only for the lockdep assertion and can be
472	* ignored.
473	*/
474	#define for_each_pool(pool, pi) \
475	idr_for_each_entry(&worker_pool_idr, pool, pi) \
476	if (({ assert_rcu_or_pool_mutex(); false; })) { } \
477	else
478
479	/**
480	* for_each_pool_worker - iterate through all workers of a worker_pool
481	* @worker: iteration cursor
482	* @pool: worker_pool to iterate workers of
483	*
484	* This must be called with wq_pool_attach_mutex.
485	*
486	* The if/else clause exists only for the lockdep assertion and can be
487	* ignored.
488	*/
489	#define for_each_pool_worker(worker, pool) \
490	list_for_each_entry((worker), &(pool)->workers, node) \
491	if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
492	else
493
494	/**
495	* for_each_pwq - iterate through all pool_workqueues of the specified workqueue
496	* @pwq: iteration cursor
497	* @wq: the target workqueue
498	*
499	* This must be called either with wq->mutex held or RCU read locked.
500	* If the pwq needs to be used beyond the locking in effect, the caller is
501	* responsible for guaranteeing that the pwq stays online.
502	*
503	* The if/else clause exists only for the lockdep assertion and can be
504	* ignored.
505	*/
506	#define for_each_pwq(pwq, wq) \
507	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
508	lockdep_is_held(&(wq->mutex)))
509
510	#ifdef CONFIG_DEBUG_OBJECTS_WORK
511
512	static const struct debug_obj_descr work_debug_descr;
513
514	static void work_debug_hint(void addr)
515	{
516	return ((struct work_struct *) addr)->func;
517	}
518
519	static bool work_is_static_object(void *addr)
520	{
521	struct work_struct *work = addr;
522
523	return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
524	}
525
526	/*
527	* fixup_init is called when:
528	* - an active object is initialized
529	*/
530	static bool work_fixup_init(void *addr, enum debug_obj_state state)
531	{
532	struct work_struct *work = addr;
533
534	switch (state) {
535	case ODEBUG_STATE_ACTIVE:
536	cancel_work_sync(work);
537	debug_object_init(work, &work_debug_descr);
538	return true;
539	default:
540	return false;
541	}
542	}
543
544	/*
545	* fixup_free is called when:
546	* - an active object is freed
547	*/
548	static bool work_fixup_free(void *addr, enum debug_obj_state state)
549	{
550	struct work_struct *work = addr;
551
552	switch (state) {
553	case ODEBUG_STATE_ACTIVE:
554	cancel_work_sync(work);
555	debug_object_free(work, &work_debug_descr);
556	return true;
557	default:
558	return false;
559	}
560	}
561
562	static const struct debug_obj_descr work_debug_descr = {
563	.name = "work_struct",
564	.debug_hint = work_debug_hint,
565	.is_static_object = work_is_static_object,
566	.fixup_init = work_fixup_init,
567	.fixup_free = work_fixup_free,
568	};
569
570	static inline void debug_work_activate(struct work_struct *work)
571	{
572	debug_object_activate(work, &work_debug_descr);
573	}
574
575	static inline void debug_work_deactivate(struct work_struct *work)
576	{
577	debug_object_deactivate(work, &work_debug_descr);
578	}
579
580	void __init_work(struct work_struct *work, int onstack)
581	{
582	if (onstack)
583	debug_object_init_on_stack(work, &work_debug_descr);
584	else
585	debug_object_init(work, &work_debug_descr);
586	}
587	EXPORT_SYMBOL_GPL(__init_work);
588
589	void destroy_work_on_stack(struct work_struct *work)
590	{
591	debug_object_free(work, &work_debug_descr);
592	}
593	EXPORT_SYMBOL_GPL(destroy_work_on_stack);
594
595	void destroy_delayed_work_on_stack(struct delayed_work *work)
596	{
597	destroy_timer_on_stack(&work->timer);
598	debug_object_free(&work->work, &work_debug_descr);
599	}
600	EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
601
602	#else
603	static inline void debug_work_activate(struct work_struct *work) { }
604	static inline void debug_work_deactivate(struct work_struct *work) { }
605	#endif
606
607	/**
608	* worker_pool_assign_id - allocate ID and assign it to @pool
609	* @pool: the pool pointer of interest
610	*
611	* Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
612	* successfully, -errno on failure.
613	*/
614	static int worker_pool_assign_id(struct worker_pool *pool)
615	{
616	int ret;
617
618	lockdep_assert_held(&wq_pool_mutex);
619
620	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
621	GFP_KERNEL);
622	if (ret >= 0) {
623	pool->id = ret;
624	return 0;
625	}
626	return ret;
627	}
628
629	static unsigned int work_color_to_flags(int color)
630	{
631	return color << WORK_STRUCT_COLOR_SHIFT;
632	}
633
634	static int get_work_color(unsigned long work_data)
635	{
636	return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
637	((1 << WORK_STRUCT_COLOR_BITS) - 1);
638	}
639
640	static int work_next_color(int color)
641	{
642	return (color + 1) % WORK_NR_COLORS;
643	}
644
645	/*
646	* While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
647	* contain the pointer to the queued pwq. Once execution starts, the flag
648	* is cleared and the high bits contain OFFQ flags and pool ID.
649	*
650	* set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
651	* and clear_work_data() can be used to set the pwq, pool or clear
652	* work->data. These functions should only be called while the work is
653	* owned - ie. while the PENDING bit is set.
654	*
655	* get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
656	* corresponding to a work. Pool is available once the work has been
657	* queued anywhere after initialization until it is sync canceled. pwq is
658	* available only while the work item is queued.
659	*
660	* %WORK_OFFQ_CANCELING is used to mark a work item which is being
661	* canceled. While being canceled, a work item may have its PENDING set
662	* but stay off timer and worklist for arbitrarily long and nobody should
663	* try to steal the PENDING bit.
664	*/
665	static inline void set_work_data(struct work_struct *work, unsigned long data,
666	unsigned long flags)
667	{
668	WARN_ON_ONCE(!work_pending(work));
669	atomic_long_set(&work->data, data \| flags \| work_static(work));
670	}
671
672	static void set_work_pwq(struct work_struct work, struct pool_workqueue pwq,
673	unsigned long extra_flags)
674	{
675	set_work_data(work, (unsigned long)pwq,
676	WORK_STRUCT_PENDING \| WORK_STRUCT_PWQ \| extra_flags);
677	}
678
679	static void set_work_pool_and_keep_pending(struct work_struct *work,
680	int pool_id)
681	{
682	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
683	WORK_STRUCT_PENDING);
684	}
685
686	static void set_work_pool_and_clear_pending(struct work_struct *work,
687	int pool_id)
688	{
689	/*
690	* The following wmb is paired with the implied mb in
691	* test_and_set_bit(PENDING) and ensures all updates to @work made
692	* here are visible to and precede any updates by the next PENDING
693	* owner.
694	*/
695	smp_wmb();
696	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
697	/*
698	* The following mb guarantees that previous clear of a PENDING bit
699	* will not be reordered with any speculative LOADS or STORES from
700	* work->current_func, which is executed afterwards. This possible
701	* reordering can lead to a missed execution on attempt to queue
702	* the same @work. E.g. consider this case:
703	*
704	* CPU#0 CPU#1
705	* ---------------------------- --------------------------------
706	*
707	* 1 STORE event_indicated
708	* 2 queue_work_on() {
709	* 3 test_and_set_bit(PENDING)
710	* 4 } set_..._and_clear_pending() {
711	* 5 set_work_data() # clear bit
712	* 6 smp_mb()
713	* 7 work->current_func() {
714	* 8 LOAD event_indicated
715	* }
716	*
717	* Without an explicit full barrier speculative LOAD on line 8 can
718	* be executed before CPU#0 does STORE on line 1. If that happens,
719	* CPU#0 observes the PENDING bit is still set and new execution of
720	* a @work is not queued in a hope, that CPU#1 will eventually
721	* finish the queued @work. Meanwhile CPU#1 does not see
722	* event_indicated is set, because speculative LOAD was executed
723	* before actual STORE.
724	*/
725	smp_mb();
726	}
727
728	static void clear_work_data(struct work_struct *work)
729	{
730	smp_wmb(); /* see set_work_pool_and_clear_pending() */
731	set_work_data(work, WORK_STRUCT_NO_POOL, 0);
732	}
733
734	static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
735	{
736	return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
737	}
738
739	static struct pool_workqueue get_work_pwq(struct work_struct work)
740	{
741	unsigned long data = atomic_long_read(&work->data);
742
743	if (data & WORK_STRUCT_PWQ)
744	return work_struct_pwq(data);
745	else
746	return NULL;
747	}
748
749	/**
750	* get_work_pool - return the worker_pool a given work was associated with
751	* @work: the work item of interest
752	*
753	* Pools are created and destroyed under wq_pool_mutex, and allows read
754	* access under RCU read lock. As such, this function should be
755	* called under wq_pool_mutex or inside of a rcu_read_lock() region.
756	*
757	* All fields of the returned pool are accessible as long as the above
758	* mentioned locking is in effect. If the returned pool needs to be used
759	* beyond the critical section, the caller is responsible for ensuring the
760	* returned pool is and stays online.
761	*
762	* Return: The worker_pool @work was last associated with. %NULL if none.
763	*/
764	static struct worker_pool get_work_pool(struct work_struct work)
765	{
766	unsigned long data = atomic_long_read(&work->data);
767	int pool_id;
768
769	assert_rcu_or_pool_mutex();
770
771	if (data & WORK_STRUCT_PWQ)
772	return work_struct_pwq(data)->pool;
773
774	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
775	if (pool_id == WORK_OFFQ_POOL_NONE)
776	return NULL;
777
778	return idr_find(&worker_pool_idr, pool_id);
779	}
780
781	/**
782	* get_work_pool_id - return the worker pool ID a given work is associated with
783	* @work: the work item of interest
784	*
785	* Return: The worker_pool ID @work was last associated with.
786	* %WORK_OFFQ_POOL_NONE if none.
787	*/
788	static int get_work_pool_id(struct work_struct *work)
789	{
790	unsigned long data = atomic_long_read(&work->data);
791
792	if (data & WORK_STRUCT_PWQ)
793	return work_struct_pwq(data)->pool->id;
794
795	return data >> WORK_OFFQ_POOL_SHIFT;
796	}
797
798	static void mark_work_canceling(struct work_struct *work)
799	{
800	unsigned long pool_id = get_work_pool_id(work);
801
802	pool_id <<= WORK_OFFQ_POOL_SHIFT;
803	set_work_data(work, pool_id \| WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
804	}
805
806	static bool work_is_canceling(struct work_struct *work)
807	{
808	unsigned long data = atomic_long_read(&work->data);
809
810	return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
811	}
812
813	/*
814	* Policy functions. These define the policies on how the global worker
815	* pools are managed. Unless noted otherwise, these functions assume that
816	* they're being called with pool->lock held.
817	*/
818
819	/*
820	* Need to wake up a worker? Called from anything but currently
821	* running workers.
822	*
823	* Note that, because unbound workers never contribute to nr_running, this
824	* function will always return %true for unbound pools as long as the
825	* worklist isn't empty.
826	*/
827	static bool need_more_worker(struct worker_pool *pool)
828	{
829	return !list_empty(&pool->worklist) && !pool->nr_running;
830	}
831
832	/* Can I start working? Called from busy but !running workers. */
833	static bool may_start_working(struct worker_pool *pool)
834	{
835	return pool->nr_idle;
836	}
837
838	/* Do I need to keep working? Called from currently running workers. */
839	static bool keep_working(struct worker_pool *pool)
840	{
841	return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
842	}
843
844	/* Do we need a new worker? Called from manager. */
845	static bool need_to_create_worker(struct worker_pool *pool)
846	{
847	return need_more_worker(pool) && !may_start_working(pool);
848	}
849
850	/* Do we have too many workers and should some go away? */
851	static bool too_many_workers(struct worker_pool *pool)
852	{
853	bool managing = pool->flags & POOL_MANAGER_ACTIVE;
854	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
855	int nr_busy = pool->nr_workers - nr_idle;
856
857	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
858	}
859
860	/**
861	* worker_set_flags - set worker flags and adjust nr_running accordingly
862	* @worker: self
863	* @flags: flags to set
864	*
865	* Set @flags in @worker->flags and adjust nr_running accordingly.
866	*/
867	static inline void worker_set_flags(struct worker *worker, unsigned int flags)
868	{
869	struct worker_pool *pool = worker->pool;
870
871	lockdep_assert_held(&pool->lock);
872
873	/* If transitioning into NOT_RUNNING, adjust nr_running. */
874	if ((flags & WORKER_NOT_RUNNING) &&
875	!(worker->flags & WORKER_NOT_RUNNING)) {
876	pool->nr_running--;
877	}
878
879	worker->flags \|= flags;
880	}
881
882	/**
883	* worker_clr_flags - clear worker flags and adjust nr_running accordingly
884	* @worker: self
885	* @flags: flags to clear
886	*
887	* Clear @flags in @worker->flags and adjust nr_running accordingly.
888	*/
889	static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
890	{
891	struct worker_pool *pool = worker->pool;
892	unsigned int oflags = worker->flags;
893
894	lockdep_assert_held(&pool->lock);
895
896	worker->flags &= ~flags;
897
898	/*
899	* If transitioning out of NOT_RUNNING, increment nr_running. Note
900	* that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
901	* of multiple flags, not a single flag.
902	*/
903	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
904	if (!(worker->flags & WORKER_NOT_RUNNING))
905	pool->nr_running++;
906	}
907
908	/* Return the first idle worker. Called with pool->lock held. */
909	static struct worker first_idle_worker(struct worker_pool pool)
910	{
911	if (unlikely(list_empty(&pool->idle_list)))
912	return NULL;
913
914	return list_first_entry(&pool->idle_list, struct worker, entry);
915	}
916
917	/**
918	* worker_enter_idle - enter idle state
919	* @worker: worker which is entering idle state
920	*
921	* @worker is entering idle state. Update stats and idle timer if
922	* necessary.
923	*
924	* LOCKING:
925	* raw_spin_lock_irq(pool->lock).
926	*/
927	static void worker_enter_idle(struct worker *worker)
928	{
929	struct worker_pool *pool = worker->pool;
930
931	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) \|\|
932	WARN_ON_ONCE(!list_empty(&worker->entry) &&
933	(worker->hentry.next \|\| worker->hentry.pprev)))
934	return;
935
936	/* can't use worker_set_flags(), also called from create_worker() */
937	worker->flags \|= WORKER_IDLE;
938	pool->nr_idle++;
939	worker->last_active = jiffies;
940
941	/* idle_list is LIFO */
942	list_add(&worker->entry, &pool->idle_list);
943
944	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
945	mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
946
947	/* Sanity check nr_running. */
948	WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
949	}
950
951	/**
952	* worker_leave_idle - leave idle state
953	* @worker: worker which is leaving idle state
954	*
955	* @worker is leaving idle state. Update stats.
956	*
957	* LOCKING:
958	* raw_spin_lock_irq(pool->lock).
959	*/
960	static void worker_leave_idle(struct worker *worker)
961	{
962	struct worker_pool *pool = worker->pool;
963
964	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
965	return;
966	worker_clr_flags(worker, WORKER_IDLE);
967	pool->nr_idle--;
968	list_del_init(&worker->entry);
969	}
970
971	/**
972	* find_worker_executing_work - find worker which is executing a work
973	* @pool: pool of interest
974	* @work: work to find worker for
975	*
976	* Find a worker which is executing @work on @pool by searching
977	* @pool->busy_hash which is keyed by the address of @work. For a worker
978	* to match, its current execution should match the address of @work and
979	* its work function. This is to avoid unwanted dependency between
980	* unrelated work executions through a work item being recycled while still
981	* being executed.
982	*
983	* This is a bit tricky. A work item may be freed once its execution
984	* starts and nothing prevents the freed area from being recycled for
985	* another work item. If the same work item address ends up being reused
986	* before the original execution finishes, workqueue will identify the
987	* recycled work item as currently executing and make it wait until the
988	* current execution finishes, introducing an unwanted dependency.
989	*
990	* This function checks the work item address and work function to avoid
991	* false positives. Note that this isn't complete as one may construct a
992	* work function which can introduce dependency onto itself through a
993	* recycled work item. Well, if somebody wants to shoot oneself in the
994	* foot that badly, there's only so much we can do, and if such deadlock
995	* actually occurs, it should be easy to locate the culprit work function.
996	*
997	* CONTEXT:
998	* raw_spin_lock_irq(pool->lock).
999	*
1000	* Return:
1001	* Pointer to worker which is executing @work if found, %NULL
1002	* otherwise.
1003	*/
1004	static struct worker find_worker_executing_work(struct worker_pool pool,
1005	struct work_struct *work)
1006	{
1007	struct worker *worker;
1008
1009	hash_for_each_possible(pool->busy_hash, worker, hentry,
1010	(unsigned long)work)
1011	if (worker->current_work == work &&
1012	worker->current_func == work->func)
1013	return worker;
1014
1015	return NULL;
1016	}
1017
1018	/**
1019	* move_linked_works - move linked works to a list
1020	* @work: start of series of works to be scheduled
1021	* @head: target list to append @work to
1022	* @nextp: out parameter for nested worklist walking
1023	*
1024	* Schedule linked works starting from @work to @head. Work series to be
1025	* scheduled starts at @work and includes any consecutive work with
1026	* WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1027	* @nextp.
1028	*
1029	* CONTEXT:
1030	* raw_spin_lock_irq(pool->lock).
1031	*/
1032	static void move_linked_works(struct work_struct work, struct list_head head,
1033	struct work_struct **nextp)
1034	{
1035	struct work_struct *n;
1036
1037	/*
1038	* Linked worklist will always end before the end of the list,
1039	* use NULL for list head.
1040	*/
1041	list_for_each_entry_safe_from(work, n, NULL, entry) {
1042	list_move_tail(&work->entry, head);
1043	if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1044	break;
1045	}
1046
1047	/*
1048	* If we're already inside safe list traversal and have moved
1049	* multiple works to the scheduled queue, the next position
1050	* needs to be updated.
1051	*/
1052	if (nextp)
1053	*nextp = n;
1054	}
1055
1056	/**
1057	* assign_work - assign a work item and its linked work items to a worker
1058	* @work: work to assign
1059	* @worker: worker to assign to
1060	* @nextp: out parameter for nested worklist walking
1061	*
1062	* Assign @work and its linked work items to @worker. If @work is already being
1063	* executed by another worker in the same pool, it'll be punted there.
1064	*
1065	* If @nextp is not NULL, it's updated to point to the next work of the last
1066	* scheduled work. This allows assign_work() to be nested inside
1067	* list_for_each_entry_safe().
1068	*
1069	* Returns %true if @work was successfully assigned to @worker. %false if @work
1070	* was punted to another worker already executing it.
1071	*/
1072	static bool assign_work(struct work_struct work, struct worker worker,
1073	struct work_struct **nextp)
1074	{
1075	struct worker_pool *pool = worker->pool;
1076	struct worker *collision;
1077
1078	lockdep_assert_held(&pool->lock);
1079
1080	/*
1081	* A single work shouldn't be executed concurrently by multiple workers.
1082	* __queue_work() ensures that @work doesn't jump to a different pool
1083	* while still running in the previous pool. Here, we should ensure that
1084	* @work is not executed concurrently by multiple workers from the same
1085	* pool. Check whether anyone is already processing the work. If so,
1086	* defer the work to the currently executing one.
1087	*/
1088	collision = find_worker_executing_work(pool, work);
1089	if (unlikely(collision)) {
1090	move_linked_works(work, &collision->scheduled, nextp);
1091	return false;
1092	}
1093
1094	move_linked_works(work, &worker->scheduled, nextp);
1095	return true;
1096	}
1097
1098	/**
1099	* kick_pool - wake up an idle worker if necessary
1100	* @pool: pool to kick
1101	*
1102	* @pool may have pending work items. Wake up worker if necessary. Returns
1103	* whether a worker was woken up.
1104	*/
1105	static bool kick_pool(struct worker_pool *pool)
1106	{
1107	struct worker *worker = first_idle_worker(pool);
1108	struct task_struct *p;
1109
1110	lockdep_assert_held(&pool->lock);
1111
1112	if (!need_more_worker(pool) \|\| !worker)
1113	return false;
1114
1115	p = worker->task;
1116
1117	#ifdef CONFIG_SMP
1118	/*
1119	* Idle @worker is about to execute @work and waking up provides an
1120	* opportunity to migrate @worker at a lower cost by setting the task's
1121	* wake_cpu field. Let's see if we want to move @worker to improve
1122	* execution locality.
1123	*
1124	* We're waking the worker that went idle the latest and there's some
1125	* chance that @worker is marked idle but hasn't gone off CPU yet. If
1126	* so, setting the wake_cpu won't do anything. As this is a best-effort
1127	* optimization and the race window is narrow, let's leave as-is for
1128	* now. If this becomes pronounced, we can skip over workers which are
1129	* still on cpu when picking an idle worker.
1130	*
1131	* If @pool has non-strict affinity, @worker might have ended up outside
1132	* its affinity scope. Repatriate.
1133	*/
1134	if (!pool->attrs->affn_strict &&
1135	!cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1136	struct work_struct *work = list_first_entry(&pool->worklist,
1137	struct work_struct, entry);
1138	p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask);
1139	get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1140	}
1141	#endif
1142	wake_up_process(p);
1143	return true;
1144	}
1145
1146	#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1147
1148	/*
1149	* Concurrency-managed per-cpu work items that hog CPU for longer than
1150	* wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1151	* which prevents them from stalling other concurrency-managed work items. If a
1152	* work function keeps triggering this mechanism, it's likely that the work item
1153	* should be using an unbound workqueue instead.
1154	*
1155	* wq_cpu_intensive_report() tracks work functions which trigger such conditions
1156	* and report them so that they can be examined and converted to use unbound
1157	* workqueues as appropriate. To avoid flooding the console, each violating work
1158	* function is tracked and reported with exponential backoff.
1159	*/
1160	#define WCI_MAX_ENTS 128
1161
1162	struct wci_ent {
1163	work_func_t func;
1164	atomic64_t cnt;
1165	struct hlist_node hash_node;
1166	};
1167
1168	static struct wci_ent wci_ents[WCI_MAX_ENTS];
1169	static int wci_nr_ents;
1170	static DEFINE_RAW_SPINLOCK(wci_lock);
1171	static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1172
1173	static struct wci_ent *wci_find_ent(work_func_t func)
1174	{
1175	struct wci_ent *ent;
1176
1177	hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1178	(unsigned long)func) {
1179	if (ent->func == func)
1180	return ent;
1181	}
1182	return NULL;
1183	}
1184
1185	static void wq_cpu_intensive_report(work_func_t func)
1186	{
1187	struct wci_ent *ent;
1188
1189	restart:
1190	ent = wci_find_ent(func);
1191	if (ent) {
1192	u64 cnt;
1193
1194	/*
1195	* Start reporting from the fourth time and back off
1196	* exponentially.
1197	*/
1198	cnt = atomic64_inc_return_relaxed(&ent->cnt);
1199	if (cnt >= 4 && is_power_of_2(cnt))
1200	printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1201	ent->func, wq_cpu_intensive_thresh_us,
1202	atomic64_read(&ent->cnt));
1203	return;
1204	}
1205
1206	/*
1207	* @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1208	* is exhausted, something went really wrong and we probably made enough
1209	* noise already.
1210	*/
1211	if (wci_nr_ents >= WCI_MAX_ENTS)
1212	return;
1213
1214	raw_spin_lock(&wci_lock);
1215
1216	if (wci_nr_ents >= WCI_MAX_ENTS) {
1217	raw_spin_unlock(&wci_lock);
1218	return;
1219	}
1220
1221	if (wci_find_ent(func)) {
1222	raw_spin_unlock(&wci_lock);
1223	goto restart;
1224	}
1225
1226	ent = &wci_ents[wci_nr_ents++];
1227	ent->func = func;
1228	atomic64_set(&ent->cnt, 1);
1229	hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1230
1231	raw_spin_unlock(&wci_lock);
1232	}
1233
1234	#else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1235	static void wq_cpu_intensive_report(work_func_t func) {}
1236	#endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1237
1238	/**
1239	* wq_worker_running - a worker is running again
1240	* @task: task waking up
1241	*
1242	* This function is called when a worker returns from schedule()
1243	*/
1244	void wq_worker_running(struct task_struct *task)
1245	{
1246	struct worker *worker = kthread_data(task);
1247
1248	if (!READ_ONCE(worker->sleeping))
1249	return;
1250
1251	/*
1252	* If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1253	* and the nr_running increment below, we may ruin the nr_running reset
1254	* and leave with an unexpected pool->nr_running == 1 on the newly unbound
1255	* pool. Protect against such race.
1256	*/
1257	preempt_disable();
1258	if (!(worker->flags & WORKER_NOT_RUNNING))
1259	worker->pool->nr_running++;
1260	preempt_enable();
1261
1262	/*
1263	* CPU intensive auto-detection cares about how long a work item hogged
1264	* CPU without sleeping. Reset the starting timestamp on wakeup.
1265	*/
1266	worker->current_at = worker->task->se.sum_exec_runtime;
1267
1268	WRITE_ONCE(worker->sleeping, 0);
1269	}
1270
1271	/**
1272	* wq_worker_sleeping - a worker is going to sleep
1273	* @task: task going to sleep
1274	*
1275	* This function is called from schedule() when a busy worker is
1276	* going to sleep.
1277	*/
1278	void wq_worker_sleeping(struct task_struct *task)
1279	{
1280	struct worker *worker = kthread_data(task);
1281	struct worker_pool *pool;
1282
1283	/*
1284	* Rescuers, which may not have all the fields set up like normal
1285	* workers, also reach here, let's not access anything before
1286	* checking NOT_RUNNING.
1287	*/
1288	if (worker->flags & WORKER_NOT_RUNNING)
1289	return;
1290
1291	pool = worker->pool;
1292
1293	/* Return if preempted before wq_worker_running() was reached */
1294	if (READ_ONCE(worker->sleeping))
1295	return;
1296
1297	WRITE_ONCE(worker->sleeping, 1);
1298	raw_spin_lock_irq(&pool->lock);
1299
1300	/*
1301	* Recheck in case unbind_workers() preempted us. We don't
1302	* want to decrement nr_running after the worker is unbound
1303	* and nr_running has been reset.
1304	*/
1305	if (worker->flags & WORKER_NOT_RUNNING) {
1306	raw_spin_unlock_irq(&pool->lock);
1307	return;
1308	}
1309
1310	pool->nr_running--;
1311	if (kick_pool(pool))
1312	worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1313
1314	raw_spin_unlock_irq(&pool->lock);
1315	}
1316
1317	/**
1318	* wq_worker_tick - a scheduler tick occurred while a kworker is running
1319	* @task: task currently running
1320	*
1321	* Called from scheduler_tick(). We're in the IRQ context and the current
1322	* worker's fields which follow the 'K' locking rule can be accessed safely.
1323	*/
1324	void wq_worker_tick(struct task_struct *task)
1325	{
1326	struct worker *worker = kthread_data(task);
1327	struct pool_workqueue *pwq = worker->current_pwq;
1328	struct worker_pool *pool = worker->pool;
1329
1330	if (!pwq)
1331	return;
1332
1333	pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1334
1335	if (!wq_cpu_intensive_thresh_us)
1336	return;
1337
1338	/*
1339	* If the current worker is concurrency managed and hogged the CPU for
1340	* longer than wq_cpu_intensive_thresh_us, it's automatically marked
1341	* CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1342	*
1343	* Set @worker->sleeping means that @worker is in the process of
1344	* switching out voluntarily and won't be contributing to
1345	* @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1346	* decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1347	* double decrements. The task is releasing the CPU anyway. Let's skip.
1348	* We probably want to make this prettier in the future.
1349	*/
1350	if ((worker->flags & WORKER_NOT_RUNNING) \|\| READ_ONCE(worker->sleeping) \|\|
1351	worker->task->se.sum_exec_runtime - worker->current_at <
1352	wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1353	return;
1354
1355	raw_spin_lock(&pool->lock);
1356
1357	worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1358	wq_cpu_intensive_report(worker->current_func);
1359	pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1360
1361	if (kick_pool(pool))
1362	pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1363
1364	raw_spin_unlock(&pool->lock);
1365	}
1366
1367	/**
1368	* wq_worker_last_func - retrieve worker's last work function
1369	* @task: Task to retrieve last work function of.
1370	*
1371	* Determine the last function a worker executed. This is called from
1372	* the scheduler to get a worker's last known identity.
1373	*
1374	* CONTEXT:
1375	* raw_spin_lock_irq(rq->lock)
1376	*
1377	* This function is called during schedule() when a kworker is going
1378	* to sleep. It's used by psi to identify aggregation workers during
1379	* dequeuing, to allow periodic aggregation to shut-off when that
1380	* worker is the last task in the system or cgroup to go to sleep.
1381	*
1382	* As this function doesn't involve any workqueue-related locking, it
1383	* only returns stable values when called from inside the scheduler's
1384	* queuing and dequeuing paths, when @task, which must be a kworker,
1385	* is guaranteed to not be processing any works.
1386	*
1387	* Return:
1388	* The last work function %current executed as a worker, NULL if it
1389	* hasn't executed any work yet.
1390	*/
1391	work_func_t wq_worker_last_func(struct task_struct *task)
1392	{
1393	struct worker *worker = kthread_data(task);
1394
1395	return worker->last_func;
1396	}
1397
1398	/**
1399	* get_pwq - get an extra reference on the specified pool_workqueue
1400	* @pwq: pool_workqueue to get
1401	*
1402	* Obtain an extra reference on @pwq. The caller should guarantee that
1403	* @pwq has positive refcnt and be holding the matching pool->lock.
1404	*/
1405	static void get_pwq(struct pool_workqueue *pwq)
1406	{
1407	lockdep_assert_held(&pwq->pool->lock);
1408	WARN_ON_ONCE(pwq->refcnt <= 0);
1409	pwq->refcnt++;
1410	}
1411
1412	/**
1413	* put_pwq - put a pool_workqueue reference
1414	* @pwq: pool_workqueue to put
1415	*
1416	* Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1417	* destruction. The caller should be holding the matching pool->lock.
1418	*/
1419	static void put_pwq(struct pool_workqueue *pwq)
1420	{
1421	lockdep_assert_held(&pwq->pool->lock);
1422	if (likely(--pwq->refcnt))
1423	return;
1424	/*
1425	* @pwq can't be released under pool->lock, bounce to a dedicated
1426	* kthread_worker to avoid A-A deadlocks.
1427	*/
1428	kthread_queue_work(pwq_release_worker, &pwq->release_work);
1429	}
1430
1431	/**
1432	* put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1433	* @pwq: pool_workqueue to put (can be %NULL)
1434	*
1435	* put_pwq() with locking. This function also allows %NULL @pwq.
1436	*/
1437	static void put_pwq_unlocked(struct pool_workqueue *pwq)
1438	{
1439	if (pwq) {
1440	/*
1441	* As both pwqs and pools are RCU protected, the
1442	* following lock operations are safe.
1443	*/
1444	raw_spin_lock_irq(&pwq->pool->lock);
1445	put_pwq(pwq);
1446	raw_spin_unlock_irq(&pwq->pool->lock);
1447	}
1448	}
1449
1450	static void pwq_activate_inactive_work(struct work_struct *work)
1451	{
1452	struct pool_workqueue *pwq = get_work_pwq(work);
1453
1454	trace_workqueue_activate_work(work);
1455	if (list_empty(&pwq->pool->worklist))
1456	pwq->pool->watchdog_ts = jiffies;
1457	move_linked_works(work, &pwq->pool->worklist, NULL);
1458	__clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work));
1459	pwq->nr_active++;
1460	}
1461
1462	static void pwq_activate_first_inactive(struct pool_workqueue *pwq)
1463	{
1464	struct work_struct *work = list_first_entry(&pwq->inactive_works,
1465	struct work_struct, entry);
1466
1467	pwq_activate_inactive_work(work);
1468	}
1469
1470	/**
1471	* pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1472	* @pwq: pwq of interest
1473	* @work_data: work_data of work which left the queue
1474	*
1475	* A work either has completed or is removed from pending queue,
1476	* decrement nr_in_flight of its pwq and handle workqueue flushing.
1477	*
1478	* CONTEXT:
1479	* raw_spin_lock_irq(pool->lock).
1480	*/
1481	static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1482	{
1483	int color = get_work_color(work_data);
1484
1485	if (!(work_data & WORK_STRUCT_INACTIVE)) {
1486	pwq->nr_active--;
1487	if (!list_empty(&pwq->inactive_works)) {
1488	/* one down, submit an inactive one */
1489	if (pwq->nr_active < pwq->max_active)
1490	pwq_activate_first_inactive(pwq);
1491	}
1492	}
1493
1494	pwq->nr_in_flight[color]--;
1495
1496	/* is flush in progress and are we at the flushing tip? */
1497	if (likely(pwq->flush_color != color))
1498	goto out_put;
1499
1500	/* are there still in-flight works? */
1501	if (pwq->nr_in_flight[color])
1502	goto out_put;
1503
1504	/* this pwq is done, clear flush_color */
1505	pwq->flush_color = -1;
1506
1507	/*
1508	* If this was the last pwq, wake up the first flusher. It
1509	* will handle the rest.
1510	*/
1511	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1512	complete(&pwq->wq->first_flusher->done);
1513	out_put:
1514	put_pwq(pwq);
1515	}
1516
1517	/**
1518	* try_to_grab_pending - steal work item from worklist and disable irq
1519	* @work: work item to steal
1520	* @is_dwork: @work is a delayed_work
1521	* @flags: place to store irq state
1522	*
1523	* Try to grab PENDING bit of @work. This function can handle @work in any
1524	* stable state - idle, on timer or on worklist.
1525	*
1526	* Return:
1527	*
1528	* ======== ================================================================
1529	* 1 if @work was pending and we successfully stole PENDING
1530	* 0 if @work was idle and we claimed PENDING
1531	* -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1532	* -ENOENT if someone else is canceling @work, this state may persist
1533	* for arbitrarily long
1534	* ======== ================================================================
1535	*
1536	* Note:
1537	* On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1538	* interrupted while holding PENDING and @work off queue, irq must be
1539	* disabled on entry. This, combined with delayed_work->timer being
1540	* irqsafe, ensures that we return -EAGAIN for finite short period of time.
1541	*
1542	* On successful return, >= 0, irq is disabled and the caller is
1543	* responsible for releasing it using local_irq_restore(*@flags).
1544	*
1545	* This function is safe to call from any context including IRQ handler.
1546	*/
1547	static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1548	unsigned long *flags)
1549	{
1550	struct worker_pool *pool;
1551	struct pool_workqueue *pwq;
1552
1553	local_irq_save(*flags);
1554
1555	/* try to steal the timer if it exists */
1556	if (is_dwork) {
1557	struct delayed_work *dwork = to_delayed_work(work);
1558
1559	/*
1560	* dwork->timer is irqsafe. If del_timer() fails, it's
1561	* guaranteed that the timer is not queued anywhere and not
1562	* running on the local CPU.
1563	*/
1564	if (likely(del_timer(&dwork->timer)))
1565	return 1;
1566	}
1567
1568	/* try to claim PENDING the normal way */
1569	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1570	return 0;
1571
1572	rcu_read_lock();
1573	/*
1574	* The queueing is in progress, or it is already queued. Try to
1575	* steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1576	*/
1577	pool = get_work_pool(work);
1578	if (!pool)
1579	goto fail;
1580
1581	raw_spin_lock(&pool->lock);
1582	/*
1583	* work->data is guaranteed to point to pwq only while the work
1584	* item is queued on pwq->wq, and both updating work->data to point
1585	* to pwq on queueing and to pool on dequeueing are done under
1586	* pwq->pool->lock. This in turn guarantees that, if work->data
1587	* points to pwq which is associated with a locked pool, the work
1588	* item is currently queued on that pool.
1589	*/
1590	pwq = get_work_pwq(work);
1591	if (pwq && pwq->pool == pool) {
1592	debug_work_deactivate(work);
1593
1594	/*
1595	* A cancelable inactive work item must be in the
1596	* pwq->inactive_works since a queued barrier can't be
1597	* canceled (see the comments in insert_wq_barrier()).
1598	*
1599	* An inactive work item cannot be grabbed directly because
1600	* it might have linked barrier work items which, if left
1601	* on the inactive_works list, will confuse pwq->nr_active
1602	* management later on and cause stall. Make sure the work
1603	* item is activated before grabbing.
1604	*/
1605	if (*work_data_bits(work) & WORK_STRUCT_INACTIVE)
1606	pwq_activate_inactive_work(work);
1607
1608	list_del_init(&work->entry);
1609	pwq_dec_nr_in_flight(pwq, *work_data_bits(work));
1610
1611	/* work->data points to pwq iff queued, point to pool */
1612	set_work_pool_and_keep_pending(work, pool->id);
1613
1614	raw_spin_unlock(&pool->lock);
1615	rcu_read_unlock();
1616	return 1;
1617	}
1618	raw_spin_unlock(&pool->lock);
1619	fail:
1620	rcu_read_unlock();
1621	local_irq_restore(*flags);
1622	if (work_is_canceling(work))
1623	return -ENOENT;
1624	cpu_relax();
1625	return -EAGAIN;
1626	}
1627
1628	/**
1629	* insert_work - insert a work into a pool
1630	* @pwq: pwq @work belongs to
1631	* @work: work to insert
1632	* @head: insertion point
1633	* @extra_flags: extra WORK_STRUCT_* flags to set
1634	*
1635	* Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1636	* work_struct flags.
1637	*
1638	* CONTEXT:
1639	* raw_spin_lock_irq(pool->lock).
1640	*/
1641	static void insert_work(struct pool_workqueue pwq, struct work_struct work,
1642	struct list_head *head, unsigned int extra_flags)
1643	{
1644	debug_work_activate(work);
1645
1646	/* record the work call stack in order to print it in KASAN reports */
1647	kasan_record_aux_stack_noalloc(work);
1648
1649	/* we own @work, set data and link */
1650	set_work_pwq(work, pwq, extra_flags);
1651	list_add_tail(&work->entry, head);
1652	get_pwq(pwq);
1653	}
1654
1655	/*
1656	* Test whether @work is being queued from another work executing on the
1657	* same workqueue.
1658	*/
1659	static bool is_chained_work(struct workqueue_struct *wq)
1660	{
1661	struct worker *worker;
1662
1663	worker = current_wq_worker();
1664	/*
1665	* Return %true iff I'm a worker executing a work item on @wq. If
1666	* I'm @worker, it's safe to dereference it without locking.
1667	*/
1668	return worker && worker->current_pwq->wq == wq;
1669	}
1670
1671	/*
1672	* When queueing an unbound work item to a wq, prefer local CPU if allowed
1673	* by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1674	* avoid perturbing sensitive tasks.
1675	*/
1676	static int wq_select_unbound_cpu(int cpu)
1677	{
1678	int new_cpu;
1679
1680	if (likely(!wq_debug_force_rr_cpu)) {
1681	if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1682	return cpu;
1683	} else {
1684	pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
1685	}
1686
1687	if (cpumask_empty(wq_unbound_cpumask))
1688	return cpu;
1689
1690	new_cpu = __this_cpu_read(wq_rr_cpu_last);
1691	new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1692	if (unlikely(new_cpu >= nr_cpu_ids)) {
1693	new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1694	if (unlikely(new_cpu >= nr_cpu_ids))
1695	return cpu;
1696	}
1697	__this_cpu_write(wq_rr_cpu_last, new_cpu);
1698
1699	return new_cpu;
1700	}
1701
1702	static void __queue_work(int cpu, struct workqueue_struct *wq,
1703	struct work_struct *work)
1704	{
1705	struct pool_workqueue *pwq;
1706	struct worker_pool last_pool, pool;
1707	unsigned int work_flags;
1708	unsigned int req_cpu = cpu;
1709
1710	/*
1711	* While a work item is PENDING && off queue, a task trying to
1712	* steal the PENDING will busy-loop waiting for it to either get
1713	* queued or lose PENDING. Grabbing PENDING and queueing should
1714	* happen with IRQ disabled.
1715	*/
1716	lockdep_assert_irqs_disabled();
1717
1718
1719	/*
1720	* For a draining wq, only works from the same workqueue are
1721	* allowed. The __WQ_DESTROYING helps to spot the issue that
1722	* queues a new work item to a wq after destroy_workqueue(wq).
1723	*/
1724	if (unlikely(wq->flags & (__WQ_DESTROYING \| __WQ_DRAINING) &&
1725	WARN_ON_ONCE(!is_chained_work(wq))))
1726	return;
1727	rcu_read_lock();
1728	retry:
1729	/* pwq which will be used unless @work is executing elsewhere */
1730	if (req_cpu == WORK_CPU_UNBOUND) {
1731	if (wq->flags & WQ_UNBOUND)
1732	cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1733	else
1734	cpu = raw_smp_processor_id();
1735	}
1736
1737	pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
1738	pool = pwq->pool;
1739
1740	/*
1741	* If @work was previously on a different pool, it might still be
1742	* running there, in which case the work needs to be queued on that
1743	* pool to guarantee non-reentrancy.
1744	*/
1745	last_pool = get_work_pool(work);
1746	if (last_pool && last_pool != pool) {
1747	struct worker *worker;
1748
1749	raw_spin_lock(&last_pool->lock);
1750
1751	worker = find_worker_executing_work(last_pool, work);
1752
1753	if (worker && worker->current_pwq->wq == wq) {
1754	pwq = worker->current_pwq;
1755	pool = pwq->pool;
1756	WARN_ON_ONCE(pool != last_pool);
1757	} else {
1758	/* meh... not running there, queue here */
1759	raw_spin_unlock(&last_pool->lock);
1760	raw_spin_lock(&pool->lock);
1761	}
1762	} else {
1763	raw_spin_lock(&pool->lock);
1764	}
1765
1766	/*
1767	* pwq is determined and locked. For unbound pools, we could have raced
1768	* with pwq release and it could already be dead. If its refcnt is zero,
1769	* repeat pwq selection. Note that unbound pwqs never die without
1770	* another pwq replacing it in cpu_pwq or while work items are executing
1771	* on it, so the retrying is guaranteed to make forward-progress.
1772	*/
1773	if (unlikely(!pwq->refcnt)) {
1774	if (wq->flags & WQ_UNBOUND) {
1775	raw_spin_unlock(&pool->lock);
1776	cpu_relax();
1777	goto retry;
1778	}
1779	/* oops */
1780	WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1781	wq->name, cpu);
1782	}
1783
1784	/* pwq determined, queue */
1785	trace_workqueue_queue_work(req_cpu, pwq, work);
1786
1787	if (WARN_ON(!list_empty(&work->entry)))
1788	goto out;
1789
1790	pwq->nr_in_flight[pwq->work_color]++;
1791	work_flags = work_color_to_flags(pwq->work_color);
1792
1793	if (likely(pwq->nr_active < pwq->max_active)) {
1794	if (list_empty(&pool->worklist))
1795	pool->watchdog_ts = jiffies;
1796
1797	trace_workqueue_activate_work(work);
1798	pwq->nr_active++;
1799	insert_work(pwq, work, &pool->worklist, work_flags);
1800	kick_pool(pool);
1801	} else {
1802	work_flags \|= WORK_STRUCT_INACTIVE;
1803	insert_work(pwq, work, &pwq->inactive_works, work_flags);
1804	}
1805
1806	out:
1807	raw_spin_unlock(&pool->lock);
1808	rcu_read_unlock();
1809	}
1810
1811	/**
1812	* queue_work_on - queue work on specific cpu
1813	* @cpu: CPU number to execute work on
1814	* @wq: workqueue to use
1815	* @work: work to queue
1816	*
1817	* We queue the work to a specific CPU, the caller must ensure it
1818	* can't go away. Callers that fail to ensure that the specified
1819	* CPU cannot go away will execute on a randomly chosen CPU.
1820	* But note well that callers specifying a CPU that never has been
1821	* online will get a splat.
1822	*
1823	* Return: %false if @work was already on a queue, %true otherwise.
1824	*/
1825	bool queue_work_on(int cpu, struct workqueue_struct *wq,
1826	struct work_struct *work)
1827	{
1828	bool ret = false;
1829	unsigned long flags;
1830
1831	local_irq_save(flags);
1832
1833	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1834	__queue_work(cpu, wq, work);
1835	ret = true;
1836	}
1837
1838	local_irq_restore(flags);
1839	return ret;
1840	}
1841	EXPORT_SYMBOL(queue_work_on);
1842
1843	/**
1844	* select_numa_node_cpu - Select a CPU based on NUMA node
1845	* @node: NUMA node ID that we want to select a CPU from
1846	*
1847	* This function will attempt to find a "random" cpu available on a given
1848	* node. If there are no CPUs available on the given node it will return
1849	* WORK_CPU_UNBOUND indicating that we should just schedule to any
1850	* available CPU if we need to schedule this work.
1851	*/
1852	static int select_numa_node_cpu(int node)
1853	{
1854	int cpu;
1855
1856	/* Delay binding to CPU if node is not valid or online */
1857	if (node < 0 \|\| node >= MAX_NUMNODES \|\| !node_online(node))
1858	return WORK_CPU_UNBOUND;
1859
1860	/* Use local node/cpu if we are already there */
1861	cpu = raw_smp_processor_id();
1862	if (node == cpu_to_node(cpu))
1863	return cpu;
1864
1865	/* Use "random" otherwise know as "first" online CPU of node */
1866	cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1867
1868	/* If CPU is valid return that, otherwise just defer */
1869	return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1870	}
1871
1872	/**
1873	* queue_work_node - queue work on a "random" cpu for a given NUMA node
1874	* @node: NUMA node that we are targeting the work for
1875	* @wq: workqueue to use
1876	* @work: work to queue
1877	*
1878	* We queue the work to a "random" CPU within a given NUMA node. The basic
1879	* idea here is to provide a way to somehow associate work with a given
1880	* NUMA node.
1881	*
1882	* This function will only make a best effort attempt at getting this onto
1883	* the right NUMA node. If no node is requested or the requested node is
1884	* offline then we just fall back to standard queue_work behavior.
1885	*
1886	* Currently the "random" CPU ends up being the first available CPU in the
1887	* intersection of cpu_online_mask and the cpumask of the node, unless we
1888	* are running on the node. In that case we just use the current CPU.
1889	*
1890	* Return: %false if @work was already on a queue, %true otherwise.
1891	*/
1892	bool queue_work_node(int node, struct workqueue_struct *wq,
1893	struct work_struct *work)
1894	{
1895	unsigned long flags;
1896	bool ret = false;
1897
1898	/*
1899	* This current implementation is specific to unbound workqueues.
1900	* Specifically we only return the first available CPU for a given
1901	* node instead of cycling through individual CPUs within the node.
1902	*
1903	* If this is used with a per-cpu workqueue then the logic in
1904	* workqueue_select_cpu_near would need to be updated to allow for
1905	* some round robin type logic.
1906	*/
1907	WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1908
1909	local_irq_save(flags);
1910
1911	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1912	int cpu = select_numa_node_cpu(node);
1913
1914	__queue_work(cpu, wq, work);
1915	ret = true;
1916	}
1917
1918	local_irq_restore(flags);
1919	return ret;
1920	}
1921	EXPORT_SYMBOL_GPL(queue_work_node);
1922
1923	void delayed_work_timer_fn(struct timer_list *t)
1924	{
1925	struct delayed_work *dwork = from_timer(dwork, t, timer);
1926
1927	/* should have been called from irqsafe timer with irq already off */
1928	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
1929	}
1930	EXPORT_SYMBOL(delayed_work_timer_fn);
1931
1932	static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1933	struct delayed_work *dwork, unsigned long delay)
1934	{
1935	struct timer_list *timer = &dwork->timer;
1936	struct work_struct *work = &dwork->work;
1937
1938	WARN_ON_ONCE(!wq);
1939	WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1940	WARN_ON_ONCE(timer_pending(timer));
1941	WARN_ON_ONCE(!list_empty(&work->entry));
1942
1943	/*
1944	* If @delay is 0, queue @dwork->work immediately. This is for
1945	* both optimization and correctness. The earliest @timer can
1946	* expire is on the closest next tick and delayed_work users depend
1947	* on that there's no such delay when @delay is 0.
1948	*/
1949	if (!delay) {
1950	__queue_work(cpu, wq, &dwork->work);
1951	return;
1952	}
1953
1954	dwork->wq = wq;
1955	dwork->cpu = cpu;
1956	timer->expires = jiffies + delay;
1957
1958	if (unlikely(cpu != WORK_CPU_UNBOUND))
1959	add_timer_on(timer, cpu);
1960	else
1961	add_timer(timer);
1962	}
1963
1964	/**
1965	* queue_delayed_work_on - queue work on specific CPU after delay
1966	* @cpu: CPU number to execute work on
1967	* @wq: workqueue to use
1968	* @dwork: work to queue
1969	* @delay: number of jiffies to wait before queueing
1970	*
1971	* Return: %false if @work was already on a queue, %true otherwise. If
1972	* @delay is zero and @dwork is idle, it will be scheduled for immediate
1973	* execution.
1974	*/
1975	bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1976	struct delayed_work *dwork, unsigned long delay)
1977	{
1978	struct work_struct *work = &dwork->work;
1979	bool ret = false;
1980	unsigned long flags;
1981
1982	/* read the comment in __queue_work() */
1983	local_irq_save(flags);
1984
1985	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1986	__queue_delayed_work(cpu, wq, dwork, delay);
1987	ret = true;
1988	}
1989
1990	local_irq_restore(flags);
1991	return ret;
1992	}
1993	EXPORT_SYMBOL(queue_delayed_work_on);
1994
1995	/**
1996	* mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1997	* @cpu: CPU number to execute work on
1998	* @wq: workqueue to use
1999	* @dwork: work to queue
2000	* @delay: number of jiffies to wait before queueing
2001	*
2002	* If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2003	* modify @dwork's timer so that it expires after @delay. If @delay is
2004	* zero, @work is guaranteed to be scheduled immediately regardless of its
2005	* current state.
2006	*
2007	* Return: %false if @dwork was idle and queued, %true if @dwork was
2008	* pending and its timer was modified.
2009	*
2010	* This function is safe to call from any context including IRQ handler.
2011	* See try_to_grab_pending() for details.
2012	*/
2013	bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2014	struct delayed_work *dwork, unsigned long delay)
2015	{
2016	unsigned long flags;
2017	int ret;
2018
2019	do {
2020	ret = try_to_grab_pending(&dwork->work, true, &flags);
2021	} while (unlikely(ret == -EAGAIN));
2022
2023	if (likely(ret >= 0)) {
2024	__queue_delayed_work(cpu, wq, dwork, delay);
2025	local_irq_restore(flags);
2026	}
2027
2028	/* -ENOENT from try_to_grab_pending() becomes %true */
2029	return ret;
2030	}
2031	EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2032
2033	static void rcu_work_rcufn(struct rcu_head *rcu)
2034	{
2035	struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2036
2037	/* read the comment in __queue_work() */
2038	local_irq_disable();
2039	__queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2040	local_irq_enable();
2041	}
2042
2043	/**
2044	* queue_rcu_work - queue work after a RCU grace period
2045	* @wq: workqueue to use
2046	* @rwork: work to queue
2047	*
2048	* Return: %false if @rwork was already pending, %true otherwise. Note
2049	* that a full RCU grace period is guaranteed only after a %true return.
2050	* While @rwork is guaranteed to be executed after a %false return, the
2051	* execution may happen before a full RCU grace period has passed.
2052	*/
2053	bool queue_rcu_work(struct workqueue_struct wq, struct rcu_work rwork)
2054	{
2055	struct work_struct *work = &rwork->work;
2056
2057	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2058	rwork->wq = wq;
2059	call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2060	return true;
2061	}
2062
2063	return false;
2064	}
2065	EXPORT_SYMBOL(queue_rcu_work);
2066
2067	static struct worker *alloc_worker(int node)
2068	{
2069	struct worker *worker;
2070
2071	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2072	if (worker) {
2073	INIT_LIST_HEAD(&worker->entry);
2074	INIT_LIST_HEAD(&worker->scheduled);
2075	INIT_LIST_HEAD(&worker->node);
2076	/* on creation a worker is in !idle && prep state */
2077	worker->flags = WORKER_PREP;
2078	}
2079	return worker;
2080	}
2081
2082	static cpumask_t pool_allowed_cpus(struct worker_pool pool)
2083	{
2084	if (pool->cpu < 0 && pool->attrs->affn_strict)
2085	return pool->attrs->__pod_cpumask;
2086	else
2087	return pool->attrs->cpumask;
2088	}
2089
2090	/**
2091	* worker_attach_to_pool() - attach a worker to a pool
2092	* @worker: worker to be attached
2093	* @pool: the target pool
2094	*
2095	* Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2096	* cpu-binding of @worker are kept coordinated with the pool across
2097	* cpu-[un]hotplugs.
2098	*/
2099	static void worker_attach_to_pool(struct worker *worker,
2100	struct worker_pool *pool)
2101	{
2102	mutex_lock(&wq_pool_attach_mutex);
2103
2104	/*
2105	* The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
2106	* stable across this function. See the comments above the flag
2107	* definition for details.
2108	*/
2109	if (pool->flags & POOL_DISASSOCIATED)
2110	worker->flags \|= WORKER_UNBOUND;
2111	else
2112	kthread_set_per_cpu(worker->task, pool->cpu);
2113
2114	if (worker->rescue_wq)
2115	set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2116
2117	list_add_tail(&worker->node, &pool->workers);
2118	worker->pool = pool;
2119
2120	mutex_unlock(&wq_pool_attach_mutex);
2121	}
2122
2123	/**
2124	* worker_detach_from_pool() - detach a worker from its pool
2125	* @worker: worker which is attached to its pool
2126	*
2127	* Undo the attaching which had been done in worker_attach_to_pool(). The
2128	* caller worker shouldn't access to the pool after detached except it has
2129	* other reference to the pool.
2130	*/
2131	static void worker_detach_from_pool(struct worker *worker)
2132	{
2133	struct worker_pool *pool = worker->pool;
2134	struct completion *detach_completion = NULL;
2135
2136	mutex_lock(&wq_pool_attach_mutex);
2137
2138	kthread_set_per_cpu(worker->task, -1);
2139	list_del(&worker->node);
2140	worker->pool = NULL;
2141
2142	if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
2143	detach_completion = pool->detach_completion;
2144	mutex_unlock(&wq_pool_attach_mutex);
2145
2146	/* clear leftover flags without pool->lock after it is detached */
2147	worker->flags &= ~(WORKER_UNBOUND \| WORKER_REBOUND);
2148
2149	if (detach_completion)
2150	complete(detach_completion);
2151	}
2152
2153	/**
2154	* create_worker - create a new workqueue worker
2155	* @pool: pool the new worker will belong to
2156	*
2157	* Create and start a new worker which is attached to @pool.
2158	*
2159	* CONTEXT:
2160	* Might sleep. Does GFP_KERNEL allocations.
2161	*
2162	* Return:
2163	* Pointer to the newly created worker.
2164	*/
2165	static struct worker create_worker(struct worker_pool pool)
2166	{
2167	struct worker *worker;
2168	int id;
2169	char id_buf[16];
2170
2171	/* ID is needed to determine kthread name */
2172	id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2173	if (id < 0) {
2174	pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2175	ERR_PTR(id));
2176	return NULL;
2177	}
2178
2179	worker = alloc_worker(pool->node);
2180	if (!worker) {
2181	pr_err_once("workqueue: Failed to allocate a worker\n");
2182	goto fail;
2183	}
2184
2185	worker->id = id;
2186
2187	if (pool->cpu >= 0)
2188	snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
2189	pool->attrs->nice < 0 ? "H" : "");
2190	else
2191	snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2192
2193	worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
2194	"kworker/%s", id_buf);
2195	if (IS_ERR(worker->task)) {
2196	if (PTR_ERR(worker->task) == -EINTR) {
2197	pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2198	id_buf);
2199	} else {
2200	pr_err_once("workqueue: Failed to create a worker thread: %pe",
2201	worker->task);
2202	}
2203	goto fail;
2204	}
2205
2206	set_user_nice(worker->task, pool->attrs->nice);
2207	kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2208
2209	/* successful, attach the worker to the pool */
2210	worker_attach_to_pool(worker, pool);
2211
2212	/* start the newly created worker */
2213	raw_spin_lock_irq(&pool->lock);
2214
2215	worker->pool->nr_workers++;
2216	worker_enter_idle(worker);
2217	kick_pool(pool);
2218
2219	/*
2220	* @worker is waiting on a completion in kthread() and will trigger hung
2221	* check if not woken up soon. As kick_pool() might not have waken it
2222	* up, wake it up explicitly once more.
2223	*/
2224	wake_up_process(worker->task);
2225
2226	raw_spin_unlock_irq(&pool->lock);
2227
2228	return worker;
2229
2230	fail:
2231	ida_free(&pool->worker_ida, id);
2232	kfree(worker);
2233	return NULL;
2234	}
2235
2236	static void unbind_worker(struct worker *worker)
2237	{
2238	lockdep_assert_held(&wq_pool_attach_mutex);
2239
2240	kthread_set_per_cpu(worker->task, -1);
2241	if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2242	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2243	else
2244	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2245	}
2246
2247	static void wake_dying_workers(struct list_head *cull_list)
2248	{
2249	struct worker worker, tmp;
2250
2251	list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2252	list_del_init(&worker->entry);
2253	unbind_worker(worker);
2254	/*
2255	* If the worker was somehow already running, then it had to be
2256	* in pool->idle_list when set_worker_dying() happened or we
2257	* wouldn't have gotten here.
2258	*
2259	* Thus, the worker must either have observed the WORKER_DIE
2260	* flag, or have set its state to TASK_IDLE. Either way, the
2261	* below will be observed by the worker and is safe to do
2262	* outside of pool->lock.
2263	*/
2264	wake_up_process(worker->task);
2265	}
2266	}
2267
2268	/**
2269	* set_worker_dying - Tag a worker for destruction
2270	* @worker: worker to be destroyed
2271	* @list: transfer worker away from its pool->idle_list and into list
2272	*
2273	* Tag @worker for destruction and adjust @pool stats accordingly. The worker
2274	* should be idle.
2275	*
2276	* CONTEXT:
2277	* raw_spin_lock_irq(pool->lock).
2278	*/
2279	static void set_worker_dying(struct worker worker, struct list_head list)
2280	{
2281	struct worker_pool *pool = worker->pool;
2282
2283	lockdep_assert_held(&pool->lock);
2284	lockdep_assert_held(&wq_pool_attach_mutex);
2285
2286	/* sanity check frenzy */
2287	if (WARN_ON(worker->current_work) \|\|
2288	WARN_ON(!list_empty(&worker->scheduled)) \|\|
2289	WARN_ON(!(worker->flags & WORKER_IDLE)))
2290	return;
2291
2292	pool->nr_workers--;
2293	pool->nr_idle--;
2294
2295	worker->flags \|= WORKER_DIE;
2296
2297	list_move(&worker->entry, list);
2298	list_move(&worker->node, &pool->dying_workers);
2299	}
2300
2301	/**
2302	* idle_worker_timeout - check if some idle workers can now be deleted.
2303	* @t: The pool's idle_timer that just expired
2304	*
2305	* The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2306	* worker_leave_idle(), as a worker flicking between idle and active while its
2307	* pool is at the too_many_workers() tipping point would cause too much timer
2308	* housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2309	* it expire and re-evaluate things from there.
2310	*/
2311	static void idle_worker_timeout(struct timer_list *t)
2312	{
2313	struct worker_pool *pool = from_timer(pool, t, idle_timer);
2314	bool do_cull = false;
2315
2316	if (work_pending(&pool->idle_cull_work))
2317	return;
2318
2319	raw_spin_lock_irq(&pool->lock);
2320
2321	if (too_many_workers(pool)) {
2322	struct worker *worker;
2323	unsigned long expires;
2324
2325	/* idle_list is kept in LIFO order, check the last one */
2326	worker = list_entry(pool->idle_list.prev, struct worker, entry);
2327	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2328	do_cull = !time_before(jiffies, expires);
2329
2330	if (!do_cull)
2331	mod_timer(&pool->idle_timer, expires);
2332	}
2333	raw_spin_unlock_irq(&pool->lock);
2334
2335	if (do_cull)
2336	queue_work(system_unbound_wq, &pool->idle_cull_work);
2337	}
2338
2339	/**
2340	* idle_cull_fn - cull workers that have been idle for too long.
2341	* @work: the pool's work for handling these idle workers
2342	*
2343	* This goes through a pool's idle workers and gets rid of those that have been
2344	* idle for at least IDLE_WORKER_TIMEOUT seconds.
2345	*
2346	* We don't want to disturb isolated CPUs because of a pcpu kworker being
2347	* culled, so this also resets worker affinity. This requires a sleepable
2348	* context, hence the split between timer callback and work item.
2349	*/
2350	static void idle_cull_fn(struct work_struct *work)
2351	{
2352	struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2353	LIST_HEAD(cull_list);
2354
2355	/*
2356	* Grabbing wq_pool_attach_mutex here ensures an already-running worker
2357	* cannot proceed beyong worker_detach_from_pool() in its self-destruct
2358	* path. This is required as a previously-preempted worker could run after
2359	* set_worker_dying() has happened but before wake_dying_workers() did.
2360	*/
2361	mutex_lock(&wq_pool_attach_mutex);
2362	raw_spin_lock_irq(&pool->lock);
2363
2364	while (too_many_workers(pool)) {
2365	struct worker *worker;
2366	unsigned long expires;
2367
2368	worker = list_entry(pool->idle_list.prev, struct worker, entry);
2369	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2370
2371	if (time_before(jiffies, expires)) {
2372	mod_timer(&pool->idle_timer, expires);
2373	break;
2374	}
2375
2376	set_worker_dying(worker, &cull_list);
2377	}
2378
2379	raw_spin_unlock_irq(&pool->lock);
2380	wake_dying_workers(&cull_list);
2381	mutex_unlock(&wq_pool_attach_mutex);
2382	}
2383
2384	static void send_mayday(struct work_struct *work)
2385	{
2386	struct pool_workqueue *pwq = get_work_pwq(work);
2387	struct workqueue_struct *wq = pwq->wq;
2388
2389	lockdep_assert_held(&wq_mayday_lock);
2390
2391	if (!wq->rescuer)
2392	return;
2393
2394	/* mayday mayday mayday */
2395	if (list_empty(&pwq->mayday_node)) {
2396	/*
2397	* If @pwq is for an unbound wq, its base ref may be put at
2398	* any time due to an attribute change. Pin @pwq until the
2399	* rescuer is done with it.
2400	*/
2401	get_pwq(pwq);
2402	list_add_tail(&pwq->mayday_node, &wq->maydays);
2403	wake_up_process(wq->rescuer->task);
2404	pwq->stats[PWQ_STAT_MAYDAY]++;
2405	}
2406	}
2407
2408	static void pool_mayday_timeout(struct timer_list *t)
2409	{
2410	struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2411	struct work_struct *work;
2412
2413	raw_spin_lock_irq(&pool->lock);
2414	raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
2415
2416	if (need_to_create_worker(pool)) {
2417	/*
2418	* We've been trying to create a new worker but
2419	* haven't been successful. We might be hitting an
2420	* allocation deadlock. Send distress signals to
2421	* rescuers.
2422	*/
2423	list_for_each_entry(work, &pool->worklist, entry)
2424	send_mayday(work);
2425	}
2426
2427	raw_spin_unlock(&wq_mayday_lock);
2428	raw_spin_unlock_irq(&pool->lock);
2429
2430	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2431	}
2432
2433	/**
2434	* maybe_create_worker - create a new worker if necessary
2435	* @pool: pool to create a new worker for
2436	*
2437	* Create a new worker for @pool if necessary. @pool is guaranteed to
2438	* have at least one idle worker on return from this function. If
2439	* creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2440	* sent to all rescuers with works scheduled on @pool to resolve
2441	* possible allocation deadlock.
2442	*
2443	* On return, need_to_create_worker() is guaranteed to be %false and
2444	* may_start_working() %true.
2445	*
2446	* LOCKING:
2447	* raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2448	* multiple times. Does GFP_KERNEL allocations. Called only from
2449	* manager.
2450	*/
2451	static void maybe_create_worker(struct worker_pool *pool)
2452	__releases(&pool->lock)
2453	__acquires(&pool->lock)
2454	{
2455	restart:
2456	raw_spin_unlock_irq(&pool->lock);
2457
2458	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2459	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2460
2461	while (true) {
2462	if (create_worker(pool) \|\| !need_to_create_worker(pool))
2463	break;
2464
2465	schedule_timeout_interruptible(CREATE_COOLDOWN);
2466
2467	if (!need_to_create_worker(pool))
2468	break;
2469	}
2470
2471	del_timer_sync(&pool->mayday_timer);
2472	raw_spin_lock_irq(&pool->lock);
2473	/*
2474	* This is necessary even after a new worker was just successfully
2475	* created as @pool->lock was dropped and the new worker might have
2476	* already become busy.
2477	*/
2478	if (need_to_create_worker(pool))
2479	goto restart;
2480	}
2481
2482	/**
2483	* manage_workers - manage worker pool
2484	* @worker: self
2485	*
2486	* Assume the manager role and manage the worker pool @worker belongs
2487	* to. At any given time, there can be only zero or one manager per
2488	* pool. The exclusion is handled automatically by this function.
2489	*
2490	* The caller can safely start processing works on false return. On
2491	* true return, it's guaranteed that need_to_create_worker() is false
2492	* and may_start_working() is true.
2493	*
2494	* CONTEXT:
2495	* raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2496	* multiple times. Does GFP_KERNEL allocations.
2497	*
2498	* Return:
2499	* %false if the pool doesn't need management and the caller can safely
2500	* start processing works, %true if management function was performed and
2501	* the conditions that the caller verified before calling the function may
2502	* no longer be true.
2503	*/
2504	static bool manage_workers(struct worker *worker)
2505	{
2506	struct worker_pool *pool = worker->pool;
2507
2508	if (pool->flags & POOL_MANAGER_ACTIVE)
2509	return false;
2510
2511	pool->flags \|= POOL_MANAGER_ACTIVE;
2512	pool->manager = worker;
2513
2514	maybe_create_worker(pool);
2515
2516	pool->manager = NULL;
2517	pool->flags &= ~POOL_MANAGER_ACTIVE;
2518	rcuwait_wake_up(&manager_wait);
2519	return true;
2520	}
2521
2522	/**
2523	* process_one_work - process single work
2524	* @worker: self
2525	* @work: work to process
2526	*
2527	* Process @work. This function contains all the logics necessary to
2528	* process a single work including synchronization against and
2529	* interaction with other workers on the same cpu, queueing and
2530	* flushing. As long as context requirement is met, any worker can
2531	* call this function to process a work.
2532	*
2533	* CONTEXT:
2534	* raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2535	*/
2536	static void process_one_work(struct worker worker, struct work_struct work)
2537	__releases(&pool->lock)
2538	__acquires(&pool->lock)
2539	{
2540	struct pool_workqueue *pwq = get_work_pwq(work);
2541	struct worker_pool *pool = worker->pool;
2542	unsigned long work_data;
2543	#ifdef CONFIG_LOCKDEP
2544	/*
2545	* It is permissible to free the struct work_struct from
2546	* inside the function that is called from it, this we need to
2547	* take into account for lockdep too. To avoid bogus "held
2548	* lock freed" warnings as well as problems when looking into
2549	* work->lockdep_map, make a copy and use that here.
2550	*/
2551	struct lockdep_map lockdep_map;
2552
2553	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2554	#endif
2555	/* ensure we're on the correct CPU */
2556	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2557	raw_smp_processor_id() != pool->cpu);
2558
2559	/* claim and dequeue */
2560	debug_work_deactivate(work);
2561	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2562	worker->current_work = work;
2563	worker->current_func = work->func;
2564	worker->current_pwq = pwq;
2565	worker->current_at = worker->task->se.sum_exec_runtime;
2566	work_data = *work_data_bits(work);
2567	worker->current_color = get_work_color(work_data);
2568
2569	/*
2570	* Record wq name for cmdline and debug reporting, may get
2571	* overridden through set_worker_desc().
2572	*/
2573	strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2574
2575	list_del_init(&work->entry);
2576
2577	/*
2578	* CPU intensive works don't participate in concurrency management.
2579	* They're the scheduler's responsibility. This takes @worker out
2580	* of concurrency management and the next code block will chain
2581	* execution of the pending work items.
2582	*/
2583	if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
2584	worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2585
2586	/*
2587	* Kick @pool if necessary. It's always noop for per-cpu worker pools
2588	* since nr_running would always be >= 1 at this point. This is used to
2589	* chain execution of the pending work items for WORKER_NOT_RUNNING
2590	* workers such as the UNBOUND and CPU_INTENSIVE ones.
2591	*/
2592	kick_pool(pool);
2593
2594	/*
2595	* Record the last pool and clear PENDING which should be the last
2596	* update to @work. Also, do this inside @pool->lock so that
2597	* PENDING and queued state changes happen together while IRQ is
2598	* disabled.
2599	*/
2600	set_work_pool_and_clear_pending(work, pool->id);
2601
2602	pwq->stats[PWQ_STAT_STARTED]++;
2603	raw_spin_unlock_irq(&pool->lock);
2604
2605	lock_map_acquire(&pwq->wq->lockdep_map);
2606	lock_map_acquire(&lockdep_map);
2607	/*
2608	* Strictly speaking we should mark the invariant state without holding
2609	* any locks, that is, before these two lock_map_acquire()'s.
2610	*
2611	* However, that would result in:
2612	*
2613	* A(W1)
2614	* WFC(C)
2615	* A(W1)
2616	* C(C)
2617	*
2618	* Which would create W1->C->W1 dependencies, even though there is no
2619	* actual deadlock possible. There are two solutions, using a
2620	* read-recursive acquire on the work(queue) 'locks', but this will then
2621	* hit the lockdep limitation on recursive locks, or simply discard
2622	* these locks.
2623	*
2624	* AFAICT there is no possible deadlock scenario between the
2625	* flush_work() and complete() primitives (except for single-threaded
2626	* workqueues), so hiding them isn't a problem.
2627	*/
2628	lockdep_invariant_state(true);
2629	trace_workqueue_execute_start(work);
2630	worker->current_func(work);
2631	/*
2632	* While we must be careful to not use "work" after this, the trace
2633	* point will only record its address.
2634	*/
2635	trace_workqueue_execute_end(work, worker->current_func);
2636	pwq->stats[PWQ_STAT_COMPLETED]++;
2637	lock_map_release(&lockdep_map);
2638	lock_map_release(&pwq->wq->lockdep_map);
2639
2640	if (unlikely(in_atomic() \|\| lockdep_depth(current) > 0)) {
2641	pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2642	" last function: %ps\n",
2643	current->comm, preempt_count(), task_pid_nr(current),
2644	worker->current_func);
2645	debug_show_held_locks(current);
2646	dump_stack();
2647	}
2648
2649	/*
2650	* The following prevents a kworker from hogging CPU on !PREEMPTION
2651	* kernels, where a requeueing work item waiting for something to
2652	* happen could deadlock with stop_machine as such work item could
2653	* indefinitely requeue itself while all other CPUs are trapped in
2654	* stop_machine. At the same time, report a quiescent RCU state so
2655	* the same condition doesn't freeze RCU.
2656	*/
2657	cond_resched();
2658
2659	raw_spin_lock_irq(&pool->lock);
2660
2661	/*
2662	* In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
2663	* CPU intensive by wq_worker_tick() if @work hogged CPU longer than
2664	* wq_cpu_intensive_thresh_us. Clear it.
2665	*/
2666	worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2667
2668	/* tag the worker for identification in schedule() */
2669	worker->last_func = worker->current_func;
2670
2671	/* we're done with it, release */
2672	hash_del(&worker->hentry);
2673	worker->current_work = NULL;
2674	worker->current_func = NULL;
2675	worker->current_pwq = NULL;
2676	worker->current_color = INT_MAX;
2677	pwq_dec_nr_in_flight(pwq, work_data);
2678	}
2679
2680	/**
2681	* process_scheduled_works - process scheduled works
2682	* @worker: self
2683	*
2684	* Process all scheduled works. Please note that the scheduled list
2685	* may change while processing a work, so this function repeatedly
2686	* fetches a work from the top and executes it.
2687	*
2688	* CONTEXT:
2689	* raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2690	* multiple times.
2691	*/
2692	static void process_scheduled_works(struct worker *worker)
2693	{
2694	struct work_struct *work;
2695	bool first = true;
2696
2697	while ((work = list_first_entry_or_null(&worker->scheduled,
2698	struct work_struct, entry))) {
2699	if (first) {
2700	worker->pool->watchdog_ts = jiffies;
2701	first = false;
2702	}
2703	process_one_work(worker, work);
2704	}
2705	}
2706
2707	static void set_pf_worker(bool val)
2708	{
2709	mutex_lock(&wq_pool_attach_mutex);
2710	if (val)
2711	current->flags \|= PF_WQ_WORKER;
2712	else
2713	current->flags &= ~PF_WQ_WORKER;
2714	mutex_unlock(&wq_pool_attach_mutex);
2715	}
2716
2717	/**
2718	* worker_thread - the worker thread function
2719	* @__worker: self
2720	*
2721	* The worker thread function. All workers belong to a worker_pool -
2722	* either a per-cpu one or dynamic unbound one. These workers process all
2723	* work items regardless of their specific target workqueue. The only
2724	* exception is work items which belong to workqueues with a rescuer which
2725	* will be explained in rescuer_thread().
2726	*
2727	* Return: 0
2728	*/
2729	static int worker_thread(void *__worker)
2730	{
2731	struct worker *worker = __worker;
2732	struct worker_pool *pool = worker->pool;
2733
2734	/* tell the scheduler that this is a workqueue worker */
2735	set_pf_worker(true);
2736	woke_up:
2737	raw_spin_lock_irq(&pool->lock);
2738
2739	/* am I supposed to die? */
2740	if (unlikely(worker->flags & WORKER_DIE)) {
2741	raw_spin_unlock_irq(&pool->lock);
2742	set_pf_worker(false);
2743
2744	set_task_comm(worker->task, "kworker/dying");
2745	ida_free(&pool->worker_ida, worker->id);
2746	worker_detach_from_pool(worker);
2747	WARN_ON_ONCE(!list_empty(&worker->entry));
2748	kfree(worker);
2749	return 0;
2750	}
2751
2752	worker_leave_idle(worker);
2753	recheck:
2754	/* no more worker necessary? */
2755	if (!need_more_worker(pool))
2756	goto sleep;
2757
2758	/* do we need to manage? */
2759	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2760	goto recheck;
2761
2762	/*
2763	* ->scheduled list can only be filled while a worker is
2764	* preparing to process a work or actually processing it.
2765	* Make sure nobody diddled with it while I was sleeping.
2766	*/
2767	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2768
2769	/*
2770	* Finish PREP stage. We're guaranteed to have at least one idle
2771	* worker or that someone else has already assumed the manager
2772	* role. This is where @worker starts participating in concurrency
2773	* management if applicable and concurrency management is restored
2774	* after being rebound. See rebind_workers() for details.
2775	*/
2776	worker_clr_flags(worker, WORKER_PREP \| WORKER_REBOUND);
2777
2778	do {
2779	struct work_struct *work =
2780	list_first_entry(&pool->worklist,
2781	struct work_struct, entry);
2782
2783	if (assign_work(work, worker, NULL))
2784	process_scheduled_works(worker);
2785	} while (keep_working(pool));
2786
2787	worker_set_flags(worker, WORKER_PREP);
2788	sleep:
2789	/*
2790	* pool->lock is held and there's no work to process and no need to
2791	* manage, sleep. Workers are woken up only while holding
2792	* pool->lock or from local cpu, so setting the current state
2793	* before releasing pool->lock is enough to prevent losing any
2794	* event.
2795	*/
2796	worker_enter_idle(worker);
2797	__set_current_state(TASK_IDLE);
2798	raw_spin_unlock_irq(&pool->lock);
2799	schedule();
2800	goto woke_up;
2801	}
2802
2803	/**
2804	* rescuer_thread - the rescuer thread function
2805	* @__rescuer: self
2806	*
2807	* Workqueue rescuer thread function. There's one rescuer for each
2808	* workqueue which has WQ_MEM_RECLAIM set.
2809	*
2810	* Regular work processing on a pool may block trying to create a new
2811	* worker which uses GFP_KERNEL allocation which has slight chance of
2812	* developing into deadlock if some works currently on the same queue
2813	* need to be processed to satisfy the GFP_KERNEL allocation. This is
2814	* the problem rescuer solves.
2815	*
2816	* When such condition is possible, the pool summons rescuers of all
2817	* workqueues which have works queued on the pool and let them process
2818	* those works so that forward progress can be guaranteed.
2819	*
2820	* This should happen rarely.
2821	*
2822	* Return: 0
2823	*/
2824	static int rescuer_thread(void *__rescuer)
2825	{
2826	struct worker *rescuer = __rescuer;
2827	struct workqueue_struct *wq = rescuer->rescue_wq;
2828	bool should_stop;
2829
2830	set_user_nice(current, RESCUER_NICE_LEVEL);
2831
2832	/*
2833	* Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2834	* doesn't participate in concurrency management.
2835	*/
2836	set_pf_worker(true);
2837	repeat:
2838	set_current_state(TASK_IDLE);
2839
2840	/*
2841	* By the time the rescuer is requested to stop, the workqueue
2842	* shouldn't have any work pending, but @wq->maydays may still have
2843	* pwq(s) queued. This can happen by non-rescuer workers consuming
2844	* all the work items before the rescuer got to them. Go through
2845	* @wq->maydays processing before acting on should_stop so that the
2846	* list is always empty on exit.
2847	*/
2848	should_stop = kthread_should_stop();
2849
2850	/* see whether any pwq is asking for help */
2851	raw_spin_lock_irq(&wq_mayday_lock);
2852
2853	while (!list_empty(&wq->maydays)) {
2854	struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2855	struct pool_workqueue, mayday_node);
2856	struct worker_pool *pool = pwq->pool;
2857	struct work_struct work, n;
2858
2859	__set_current_state(TASK_RUNNING);
2860	list_del_init(&pwq->mayday_node);
2861
2862	raw_spin_unlock_irq(&wq_mayday_lock);
2863
2864	worker_attach_to_pool(rescuer, pool);
2865
2866	raw_spin_lock_irq(&pool->lock);
2867
2868	/*
2869	* Slurp in all works issued via this workqueue and
2870	* process'em.
2871	*/
2872	WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2873	list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2874	if (get_work_pwq(work) == pwq &&
2875	assign_work(work, rescuer, &n))
2876	pwq->stats[PWQ_STAT_RESCUED]++;
2877	}
2878
2879	if (!list_empty(&rescuer->scheduled)) {
2880	process_scheduled_works(rescuer);
2881
2882	/*
2883	* The above execution of rescued work items could
2884	* have created more to rescue through
2885	* pwq_activate_first_inactive() or chained
2886	* queueing. Let's put @pwq back on mayday list so
2887	* that such back-to-back work items, which may be
2888	* being used to relieve memory pressure, don't
2889	* incur MAYDAY_INTERVAL delay inbetween.
2890	*/
2891	if (pwq->nr_active && need_to_create_worker(pool)) {
2892	raw_spin_lock(&wq_mayday_lock);
2893	/*
2894	* Queue iff we aren't racing destruction
2895	* and somebody else hasn't queued it already.
2896	*/
2897	if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2898	get_pwq(pwq);
2899	list_add_tail(&pwq->mayday_node, &wq->maydays);
2900	}
2901	raw_spin_unlock(&wq_mayday_lock);
2902	}
2903	}
2904
2905	/*
2906	* Put the reference grabbed by send_mayday(). @pool won't
2907	* go away while we're still attached to it.
2908	*/
2909	put_pwq(pwq);
2910
2911	/*
2912	* Leave this pool. Notify regular workers; otherwise, we end up
2913	* with 0 concurrency and stalling the execution.
2914	*/
2915	kick_pool(pool);
2916
2917	raw_spin_unlock_irq(&pool->lock);
2918
2919	worker_detach_from_pool(rescuer);
2920
2921	raw_spin_lock_irq(&wq_mayday_lock);
2922	}
2923
2924	raw_spin_unlock_irq(&wq_mayday_lock);
2925
2926	if (should_stop) {
2927	__set_current_state(TASK_RUNNING);
2928	set_pf_worker(false);
2929	return 0;
2930	}
2931
2932	/* rescuers should never participate in concurrency management */
2933	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2934	schedule();
2935	goto repeat;
2936	}
2937
2938	/**
2939	* check_flush_dependency - check for flush dependency sanity
2940	* @target_wq: workqueue being flushed
2941	* @target_work: work item being flushed (NULL for workqueue flushes)
2942	*
2943	* %current is trying to flush the whole @target_wq or @target_work on it.
2944	* If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2945	* reclaiming memory or running on a workqueue which doesn't have
2946	* %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2947	* a deadlock.
2948	*/
2949	static void check_flush_dependency(struct workqueue_struct *target_wq,
2950	struct work_struct *target_work)
2951	{
2952	work_func_t target_func = target_work ? target_work->func : NULL;
2953	struct worker *worker;
2954
2955	if (target_wq->flags & WQ_MEM_RECLAIM)
2956	return;
2957
2958	worker = current_wq_worker();
2959
2960	WARN_ONCE(current->flags & PF_MEMALLOC,
2961	"workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2962	current->pid, current->comm, target_wq->name, target_func);
2963	WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2964	(WQ_MEM_RECLAIM \| __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2965	"workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2966	worker->current_pwq->wq->name, worker->current_func,
2967	target_wq->name, target_func);
2968	}
2969
2970	struct wq_barrier {
2971	struct work_struct work;
2972	struct completion done;
2973	struct task_struct task; / purely informational */
2974	};
2975
2976	static void wq_barrier_func(struct work_struct *work)
2977	{
2978	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2979	complete(&barr->done);
2980	}
2981
2982	/**
2983	* insert_wq_barrier - insert a barrier work
2984	* @pwq: pwq to insert barrier into
2985	* @barr: wq_barrier to insert
2986	* @target: target work to attach @barr to
2987	* @worker: worker currently executing @target, NULL if @target is not executing
2988	*
2989	* @barr is linked to @target such that @barr is completed only after
2990	* @target finishes execution. Please note that the ordering
2991	* guarantee is observed only with respect to @target and on the local
2992	* cpu.
2993	*
2994	* Currently, a queued barrier can't be canceled. This is because
2995	* try_to_grab_pending() can't determine whether the work to be
2996	* grabbed is at the head of the queue and thus can't clear LINKED
2997	* flag of the previous work while there must be a valid next work
2998	* after a work with LINKED flag set.
2999	*
3000	* Note that when @worker is non-NULL, @target may be modified
3001	* underneath us, so we can't reliably determine pwq from @target.
3002	*
3003	* CONTEXT:
3004	* raw_spin_lock_irq(pool->lock).
3005	*/
3006	static void insert_wq_barrier(struct pool_workqueue *pwq,
3007	struct wq_barrier *barr,
3008	struct work_struct target, struct worker worker)
3009	{
3010	unsigned int work_flags = 0;
3011	unsigned int work_color;
3012	struct list_head *head;
3013
3014	/*
3015	* debugobject calls are safe here even with pool->lock locked
3016	* as we know for sure that this will not trigger any of the
3017	* checks and call back into the fixup functions where we
3018	* might deadlock.
3019	*/
3020	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
3021	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3022
3023	init_completion_map(&barr->done, &target->lockdep_map);
3024
3025	barr->task = current;
3026
3027	/* The barrier work item does not participate in pwq->nr_active. */
3028	work_flags \|= WORK_STRUCT_INACTIVE;
3029
3030	/*
3031	* If @target is currently being executed, schedule the
3032	* barrier to the worker; otherwise, put it after @target.
3033	*/
3034	if (worker) {
3035	head = worker->scheduled.next;
3036	work_color = worker->current_color;
3037	} else {
3038	unsigned long *bits = work_data_bits(target);
3039
3040	head = target->entry.next;
3041	/* there can already be other linked works, inherit and set */
3042	work_flags \|= *bits & WORK_STRUCT_LINKED;
3043	work_color = get_work_color(*bits);
3044	__set_bit(WORK_STRUCT_LINKED_BIT, bits);
3045	}
3046
3047	pwq->nr_in_flight[work_color]++;
3048	work_flags \|= work_color_to_flags(work_color);
3049
3050	insert_work(pwq, &barr->work, head, work_flags);
3051	}
3052
3053	/**
3054	* flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3055	* @wq: workqueue being flushed
3056	* @flush_color: new flush color, < 0 for no-op
3057	* @work_color: new work color, < 0 for no-op
3058	*
3059	* Prepare pwqs for workqueue flushing.
3060	*
3061	* If @flush_color is non-negative, flush_color on all pwqs should be
3062	* -1. If no pwq has in-flight commands at the specified color, all
3063	* pwq->flush_color's stay at -1 and %false is returned. If any pwq
3064	* has in flight commands, its pwq->flush_color is set to
3065	* @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3066	* wakeup logic is armed and %true is returned.
3067	*
3068	* The caller should have initialized @wq->first_flusher prior to
3069	* calling this function with non-negative @flush_color. If
3070	* @flush_color is negative, no flush color update is done and %false
3071	* is returned.
3072	*
3073	* If @work_color is non-negative, all pwqs should have the same
3074	* work_color which is previous to @work_color and all will be
3075	* advanced to @work_color.
3076	*
3077	* CONTEXT:
3078	* mutex_lock(wq->mutex).
3079	*
3080	* Return:
3081	* %true if @flush_color >= 0 and there's something to flush. %false
3082	* otherwise.
3083	*/
3084	static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3085	int flush_color, int work_color)
3086	{
3087	bool wait = false;
3088	struct pool_workqueue *pwq;
3089
3090	if (flush_color >= 0) {
3091	WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3092	atomic_set(&wq->nr_pwqs_to_flush, 1);
3093	}
3094
3095	for_each_pwq(pwq, wq) {
3096	struct worker_pool *pool = pwq->pool;
3097
3098	raw_spin_lock_irq(&pool->lock);
3099
3100	if (flush_color >= 0) {
3101	WARN_ON_ONCE(pwq->flush_color != -1);
3102
3103	if (pwq->nr_in_flight[flush_color]) {
3104	pwq->flush_color = flush_color;
3105	atomic_inc(&wq->nr_pwqs_to_flush);
3106	wait = true;
3107	}
3108	}
3109
3110	if (work_color >= 0) {
3111	WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3112	pwq->work_color = work_color;
3113	}
3114
3115	raw_spin_unlock_irq(&pool->lock);
3116	}
3117
3118	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3119	complete(&wq->first_flusher->done);
3120
3121	return wait;
3122	}
3123
3124	/**
3125	* __flush_workqueue - ensure that any scheduled work has run to completion.
3126	* @wq: workqueue to flush
3127	*
3128	* This function sleeps until all work items which were queued on entry
3129	* have finished execution, but it is not livelocked by new incoming ones.
3130	*/
3131	void __flush_workqueue(struct workqueue_struct *wq)
3132	{
3133	struct wq_flusher this_flusher = {
3134	.list = LIST_HEAD_INIT(this_flusher.list),
3135	.flush_color = -1,
3136	.done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3137	};
3138	int next_color;
3139
3140	if (WARN_ON(!wq_online))
3141	return;
3142
3143	lock_map_acquire(&wq->lockdep_map);
3144	lock_map_release(&wq->lockdep_map);
3145
3146	mutex_lock(&wq->mutex);
3147
3148	/*
3149	* Start-to-wait phase
3150	*/
3151	next_color = work_next_color(wq->work_color);
3152
3153	if (next_color != wq->flush_color) {
3154	/*
3155	* Color space is not full. The current work_color
3156	* becomes our flush_color and work_color is advanced
3157	* by one.
3158	*/
3159	WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3160	this_flusher.flush_color = wq->work_color;
3161	wq->work_color = next_color;
3162
3163	if (!wq->first_flusher) {
3164	/* no flush in progress, become the first flusher */
3165	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3166
3167	wq->first_flusher = &this_flusher;
3168
3169	if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3170	wq->work_color)) {
3171	/* nothing to flush, done */
3172	wq->flush_color = next_color;
3173	wq->first_flusher = NULL;
3174	goto out_unlock;
3175	}
3176	} else {
3177	/* wait in queue */
3178	WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3179	list_add_tail(&this_flusher.list, &wq->flusher_queue);
3180	flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3181	}
3182	} else {
3183	/*
3184	* Oops, color space is full, wait on overflow queue.
3185	* The next flush completion will assign us
3186	* flush_color and transfer to flusher_queue.
3187	*/
3188	list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3189	}
3190
3191	check_flush_dependency(wq, NULL);
3192
3193	mutex_unlock(&wq->mutex);
3194
3195	wait_for_completion(&this_flusher.done);
3196
3197	/*
3198	* Wake-up-and-cascade phase
3199	*
3200	* First flushers are responsible for cascading flushes and
3201	* handling overflow. Non-first flushers can simply return.
3202	*/
3203	if (READ_ONCE(wq->first_flusher) != &this_flusher)
3204	return;
3205
3206	mutex_lock(&wq->mutex);
3207
3208	/* we might have raced, check again with mutex held */
3209	if (wq->first_flusher != &this_flusher)
3210	goto out_unlock;
3211
3212	WRITE_ONCE(wq->first_flusher, NULL);
3213
3214	WARN_ON_ONCE(!list_empty(&this_flusher.list));
3215	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3216
3217	while (true) {
3218	struct wq_flusher next, tmp;
3219
3220	/* complete all the flushers sharing the current flush color */
3221	list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3222	if (next->flush_color != wq->flush_color)
3223	break;
3224	list_del_init(&next->list);
3225	complete(&next->done);
3226	}
3227
3228	WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
3229	wq->flush_color != work_next_color(wq->work_color));
3230
3231	/* this flush_color is finished, advance by one */
3232	wq->flush_color = work_next_color(wq->flush_color);
3233
3234	/* one color has been freed, handle overflow queue */
3235	if (!list_empty(&wq->flusher_overflow)) {
3236	/*
3237	* Assign the same color to all overflowed
3238	* flushers, advance work_color and append to
3239	* flusher_queue. This is the start-to-wait
3240	* phase for these overflowed flushers.
3241	*/
3242	list_for_each_entry(tmp, &wq->flusher_overflow, list)
3243	tmp->flush_color = wq->work_color;
3244
3245	wq->work_color = work_next_color(wq->work_color);
3246
3247	list_splice_tail_init(&wq->flusher_overflow,
3248	&wq->flusher_queue);
3249	flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3250	}
3251
3252	if (list_empty(&wq->flusher_queue)) {
3253	WARN_ON_ONCE(wq->flush_color != wq->work_color);
3254	break;
3255	}
3256
3257	/*
3258	* Need to flush more colors. Make the next flusher
3259	* the new first flusher and arm pwqs.
3260	*/
3261	WARN_ON_ONCE(wq->flush_color == wq->work_color);
3262	WARN_ON_ONCE(wq->flush_color != next->flush_color);
3263
3264	list_del_init(&next->list);
3265	wq->first_flusher = next;
3266
3267	if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
3268	break;
3269
3270	/*
3271	* Meh... this color is already done, clear first
3272	* flusher and repeat cascading.
3273	*/
3274	wq->first_flusher = NULL;
3275	}
3276
3277	out_unlock:
3278	mutex_unlock(&wq->mutex);
3279	}
3280	EXPORT_SYMBOL(__flush_workqueue);
3281
3282	/**
3283	* drain_workqueue - drain a workqueue
3284	* @wq: workqueue to drain
3285	*
3286	* Wait until the workqueue becomes empty. While draining is in progress,
3287	* only chain queueing is allowed. IOW, only currently pending or running
3288	* work items on @wq can queue further work items on it. @wq is flushed
3289	* repeatedly until it becomes empty. The number of flushing is determined
3290	* by the depth of chaining and should be relatively short. Whine if it
3291	* takes too long.
3292	*/
3293	void drain_workqueue(struct workqueue_struct *wq)
3294	{
3295	unsigned int flush_cnt = 0;
3296	struct pool_workqueue *pwq;
3297
3298	/*
3299	* __queue_work() needs to test whether there are drainers, is much
3300	* hotter than drain_workqueue() and already looks at @wq->flags.
3301	* Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
3302	*/
3303	mutex_lock(&wq->mutex);
3304	if (!wq->nr_drainers++)
3305	wq->flags \|= __WQ_DRAINING;
3306	mutex_unlock(&wq->mutex);
3307	reflush:
3308	__flush_workqueue(wq);
3309
3310	mutex_lock(&wq->mutex);
3311
3312	for_each_pwq(pwq, wq) {
3313	bool drained;
3314
3315	raw_spin_lock_irq(&pwq->pool->lock);
3316	drained = !pwq->nr_active && list_empty(&pwq->inactive_works);
3317	raw_spin_unlock_irq(&pwq->pool->lock);
3318
3319	if (drained)
3320	continue;
3321
3322	if (++flush_cnt == 10 \|\|
3323	(flush_cnt % 100 == 0 && flush_cnt <= 1000))
3324	pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
3325	wq->name, __func__, flush_cnt);
3326
3327	mutex_unlock(&wq->mutex);
3328	goto reflush;
3329	}
3330
3331	if (!--wq->nr_drainers)
3332	wq->flags &= ~__WQ_DRAINING;
3333	mutex_unlock(&wq->mutex);
3334	}
3335	EXPORT_SYMBOL_GPL(drain_workqueue);
3336
3337	static bool start_flush_work(struct work_struct work, struct wq_barrier barr,
3338	bool from_cancel)
3339	{
3340	struct worker *worker = NULL;
3341	struct worker_pool *pool;
3342	struct pool_workqueue *pwq;
3343
3344	might_sleep();
3345
3346	rcu_read_lock();
3347	pool = get_work_pool(work);
3348	if (!pool) {
3349	rcu_read_unlock();
3350	return false;
3351	}
3352
3353	raw_spin_lock_irq(&pool->lock);
3354	/* see the comment in try_to_grab_pending() with the same code */
3355	pwq = get_work_pwq(work);
3356	if (pwq) {
3357	if (unlikely(pwq->pool != pool))
3358	goto already_gone;
3359	} else {
3360	worker = find_worker_executing_work(pool, work);
3361	if (!worker)
3362	goto already_gone;
3363	pwq = worker->current_pwq;
3364	}
3365
3366	check_flush_dependency(pwq->wq, work);
3367
3368	insert_wq_barrier(pwq, barr, work, worker);
3369	raw_spin_unlock_irq(&pool->lock);
3370
3371	/*
3372	* Force a lock recursion deadlock when using flush_work() inside a
3373	* single-threaded or rescuer equipped workqueue.
3374	*
3375	* For single threaded workqueues the deadlock happens when the work
3376	* is after the work issuing the flush_work(). For rescuer equipped
3377	* workqueues the deadlock happens when the rescuer stalls, blocking
3378	* forward progress.
3379	*/
3380	if (!from_cancel &&
3381	(pwq->wq->saved_max_active == 1 \|\| pwq->wq->rescuer)) {
3382	lock_map_acquire(&pwq->wq->lockdep_map);
3383	lock_map_release(&pwq->wq->lockdep_map);
3384	}
3385	rcu_read_unlock();
3386	return true;
3387	already_gone:
3388	raw_spin_unlock_irq(&pool->lock);
3389	rcu_read_unlock();
3390	return false;
3391	}
3392
3393	static bool __flush_work(struct work_struct *work, bool from_cancel)
3394	{
3395	struct wq_barrier barr;
3396
3397	if (WARN_ON(!wq_online))
3398	return false;
3399
3400	if (WARN_ON(!work->func))
3401	return false;
3402
3403	lock_map_acquire(&work->lockdep_map);
3404	lock_map_release(&work->lockdep_map);
3405
3406	if (start_flush_work(work, &barr, from_cancel)) {
3407	wait_for_completion(&barr.done);
3408	destroy_work_on_stack(&barr.work);
3409	return true;
3410	} else {
3411	return false;
3412	}
3413	}
3414
3415	/**
3416	* flush_work - wait for a work to finish executing the last queueing instance
3417	* @work: the work to flush
3418	*
3419	* Wait until @work has finished execution. @work is guaranteed to be idle
3420	* on return if it hasn't been requeued since flush started.
3421	*
3422	* Return:
3423	* %true if flush_work() waited for the work to finish execution,
3424	* %false if it was already idle.
3425	*/
3426	bool flush_work(struct work_struct *work)
3427	{
3428	return __flush_work(work, false);
3429	}
3430	EXPORT_SYMBOL_GPL(flush_work);
3431
3432	struct cwt_wait {
3433	wait_queue_entry_t wait;
3434	struct work_struct *work;
3435	};
3436
3437	static int cwt_wakefn(wait_queue_entry_t wait, unsigned mode, int sync, void key)
3438	{
3439	struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3440
3441	if (cwait->work != key)
3442	return 0;
3443	return autoremove_wake_function(wait, mode, sync, key);
3444	}
3445
3446	static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3447	{
3448	static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3449	unsigned long flags;
3450	int ret;
3451
3452	do {
3453	ret = try_to_grab_pending(work, is_dwork, &flags);
3454	/*
3455	* If someone else is already canceling, wait for it to
3456	* finish. flush_work() doesn't work for PREEMPT_NONE
3457	* because we may get scheduled between @work's completion
3458	* and the other canceling task resuming and clearing
3459	* CANCELING - flush_work() will return false immediately
3460	* as @work is no longer busy, try_to_grab_pending() will
3461	* return -ENOENT as @work is still being canceled and the
3462	* other canceling task won't be able to clear CANCELING as
3463	* we're hogging the CPU.
3464	*
3465	* Let's wait for completion using a waitqueue. As this
3466	* may lead to the thundering herd problem, use a custom
3467	* wake function which matches @work along with exclusive
3468	* wait and wakeup.
3469	*/
3470	if (unlikely(ret == -ENOENT)) {
3471	struct cwt_wait cwait;
3472
3473	init_wait(&cwait.wait);
3474	cwait.wait.func = cwt_wakefn;
3475	cwait.work = work;
3476
3477	prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3478	TASK_UNINTERRUPTIBLE);
3479	if (work_is_canceling(work))
3480	schedule();
3481	finish_wait(&cancel_waitq, &cwait.wait);
3482	}
3483	} while (unlikely(ret < 0));
3484
3485	/* tell other tasks trying to grab @work to back off */
3486	mark_work_canceling(work);
3487	local_irq_restore(flags);
3488
3489	/*
3490	* This allows canceling during early boot. We know that @work
3491	* isn't executing.
3492	*/
3493	if (wq_online)
3494	__flush_work(work, true);
3495
3496	clear_work_data(work);
3497
3498	/*
3499	* Paired with prepare_to_wait() above so that either
3500	* waitqueue_active() is visible here or !work_is_canceling() is
3501	* visible there.
3502	*/
3503	smp_mb();
3504	if (waitqueue_active(&cancel_waitq))
3505	__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3506
3507	return ret;
3508	}
3509
3510	/**
3511	* cancel_work_sync - cancel a work and wait for it to finish
3512	* @work: the work to cancel
3513	*
3514	* Cancel @work and wait for its execution to finish. This function
3515	* can be used even if the work re-queues itself or migrates to
3516	* another workqueue. On return from this function, @work is
3517	* guaranteed to be not pending or executing on any CPU.
3518	*
3519	* cancel_work_sync(&delayed_work->work) must not be used for
3520	* delayed_work's. Use cancel_delayed_work_sync() instead.
3521	*
3522	* The caller must ensure that the workqueue on which @work was last
3523	* queued can't be destroyed before this function returns.
3524	*
3525	* Return:
3526	* %true if @work was pending, %false otherwise.
3527	*/
3528	bool cancel_work_sync(struct work_struct *work)
3529	{
3530	return __cancel_work_timer(work, false);
3531	}
3532	EXPORT_SYMBOL_GPL(cancel_work_sync);
3533
3534	/**
3535	* flush_delayed_work - wait for a dwork to finish executing the last queueing
3536	* @dwork: the delayed work to flush
3537	*
3538	* Delayed timer is cancelled and the pending work is queued for
3539	* immediate execution. Like flush_work(), this function only
3540	* considers the last queueing instance of @dwork.
3541	*
3542	* Return:
3543	* %true if flush_work() waited for the work to finish execution,
3544	* %false if it was already idle.
3545	*/
3546	bool flush_delayed_work(struct delayed_work *dwork)
3547	{
3548	local_irq_disable();
3549	if (del_timer_sync(&dwork->timer))
3550	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
3551	local_irq_enable();
3552	return flush_work(&dwork->work);
3553	}
3554	EXPORT_SYMBOL(flush_delayed_work);
3555
3556	/**
3557	* flush_rcu_work - wait for a rwork to finish executing the last queueing
3558	* @rwork: the rcu work to flush
3559	*
3560	* Return:
3561	* %true if flush_rcu_work() waited for the work to finish execution,
3562	* %false if it was already idle.
3563	*/
3564	bool flush_rcu_work(struct rcu_work *rwork)
3565	{
3566	if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3567	rcu_barrier();
3568	flush_work(&rwork->work);
3569	return true;
3570	} else {
3571	return flush_work(&rwork->work);
3572	}
3573	}
3574	EXPORT_SYMBOL(flush_rcu_work);
3575
3576	static bool __cancel_work(struct work_struct *work, bool is_dwork)
3577	{
3578	unsigned long flags;
3579	int ret;
3580
3581	do {
3582	ret = try_to_grab_pending(work, is_dwork, &flags);
3583	} while (unlikely(ret == -EAGAIN));
3584
3585	if (unlikely(ret < 0))
3586	return false;
3587
3588	set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3589	local_irq_restore(flags);
3590	return ret;
3591	}
3592
3593	/*
3594	* See cancel_delayed_work()
3595	*/
3596	bool cancel_work(struct work_struct *work)
3597	{
3598	return __cancel_work(work, false);
3599	}
3600	EXPORT_SYMBOL(cancel_work);
3601
3602	/**
3603	* cancel_delayed_work - cancel a delayed work
3604	* @dwork: delayed_work to cancel
3605	*
3606	* Kill off a pending delayed_work.
3607	*
3608	* Return: %true if @dwork was pending and canceled; %false if it wasn't
3609	* pending.
3610	*
3611	* Note:
3612	* The work callback function may still be running on return, unless
3613	* it returns %true and the work doesn't re-arm itself. Explicitly flush or
3614	* use cancel_delayed_work_sync() to wait on it.
3615	*
3616	* This function is safe to call from any context including IRQ handler.
3617	*/
3618	bool cancel_delayed_work(struct delayed_work *dwork)
3619	{
3620	return __cancel_work(&dwork->work, true);
3621	}
3622	EXPORT_SYMBOL(cancel_delayed_work);
3623
3624	/**
3625	* cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3626	* @dwork: the delayed work cancel
3627	*
3628	* This is cancel_work_sync() for delayed works.
3629	*
3630	* Return:
3631	* %true if @dwork was pending, %false otherwise.
3632	*/
3633	bool cancel_delayed_work_sync(struct delayed_work *dwork)
3634	{
3635	return __cancel_work_timer(&dwork->work, true);
3636	}
3637	EXPORT_SYMBOL(cancel_delayed_work_sync);
3638
3639	/**
3640	* schedule_on_each_cpu - execute a function synchronously on each online CPU
3641	* @func: the function to call
3642	*
3643	* schedule_on_each_cpu() executes @func on each online CPU using the
3644	* system workqueue and blocks until all CPUs have completed.
3645	* schedule_on_each_cpu() is very slow.
3646	*
3647	* Return:
3648	* 0 on success, -errno on failure.
3649	*/
3650	int schedule_on_each_cpu(work_func_t func)
3651	{
3652	int cpu;
3653	struct work_struct __percpu *works;
3654
3655	works = alloc_percpu(struct work_struct);
3656	if (!works)
3657	return -ENOMEM;
3658
3659	cpus_read_lock();
3660
3661	for_each_online_cpu(cpu) {
3662	struct work_struct *work = per_cpu_ptr(works, cpu);
3663
3664	INIT_WORK(work, func);
3665	schedule_work_on(cpu, work);
3666	}
3667
3668	for_each_online_cpu(cpu)
3669	flush_work(per_cpu_ptr(works, cpu));
3670
3671	cpus_read_unlock();
3672	free_percpu(works);
3673	return 0;
3674	}
3675
3676	/**
3677	* execute_in_process_context - reliably execute the routine with user context
3678	* @fn: the function to execute
3679	* @ew: guaranteed storage for the execute work structure (must
3680	* be available when the work executes)
3681	*
3682	* Executes the function immediately if process context is available,
3683	* otherwise schedules the function for delayed execution.
3684	*
3685	* Return: 0 - function was executed
3686	* 1 - function was scheduled for execution
3687	*/
3688	int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3689	{
3690	if (!in_interrupt()) {
3691	fn(&ew->work);
3692	return 0;
3693	}
3694
3695	INIT_WORK(&ew->work, fn);
3696	schedule_work(&ew->work);
3697
3698	return 1;
3699	}
3700	EXPORT_SYMBOL_GPL(execute_in_process_context);
3701
3702	/**
3703	* free_workqueue_attrs - free a workqueue_attrs
3704	* @attrs: workqueue_attrs to free
3705	*
3706	* Undo alloc_workqueue_attrs().
3707	*/
3708	void free_workqueue_attrs(struct workqueue_attrs *attrs)
3709	{
3710	if (attrs) {
3711	free_cpumask_var(attrs->cpumask);
3712	free_cpumask_var(attrs->__pod_cpumask);
3713	kfree(attrs);
3714	}
3715	}
3716
3717	/**
3718	* alloc_workqueue_attrs - allocate a workqueue_attrs
3719	*
3720	* Allocate a new workqueue_attrs, initialize with default settings and
3721	* return it.
3722	*
3723	* Return: The allocated new workqueue_attr on success. %NULL on failure.
3724	*/
3725	struct workqueue_attrs *alloc_workqueue_attrs(void)
3726	{
3727	struct workqueue_attrs *attrs;
3728
3729	attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3730	if (!attrs)
3731	goto fail;
3732	if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3733	goto fail;
3734	if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
3735	goto fail;
3736
3737	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3738	attrs->affn_scope = WQ_AFFN_DFL;
3739	return attrs;
3740	fail:
3741	free_workqueue_attrs(attrs);
3742	return NULL;
3743	}
3744
3745	static void copy_workqueue_attrs(struct workqueue_attrs *to,
3746	const struct workqueue_attrs *from)
3747	{
3748	to->nice = from->nice;
3749	cpumask_copy(to->cpumask, from->cpumask);
3750	cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
3751	to->affn_strict = from->affn_strict;
3752
3753	/*
3754	* Unlike hash and equality test, copying shouldn't ignore wq-only
3755	* fields as copying is used for both pool and wq attrs. Instead,
3756	* get_unbound_pool() explicitly clears the fields.
3757	*/
3758	to->affn_scope = from->affn_scope;
3759	to->ordered = from->ordered;
3760	}
3761
3762	/*
3763	* Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
3764	* comments in 'struct workqueue_attrs' definition.
3765	*/
3766	static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
3767	{
3768	attrs->affn_scope = WQ_AFFN_NR_TYPES;
3769	attrs->ordered = false;
3770	}
3771
3772	/* hash value of the content of @attr */
3773	static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3774	{
3775	u32 hash = 0;
3776
3777	hash = jhash_1word(attrs->nice, hash);
3778	hash = jhash(cpumask_bits(attrs->cpumask),
3779	BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3780	hash = jhash(cpumask_bits(attrs->__pod_cpumask),
3781	BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3782	hash = jhash_1word(attrs->affn_strict, hash);
3783	return hash;
3784	}
3785
3786	/* content equality test */
3787	static bool wqattrs_equal(const struct workqueue_attrs *a,
3788	const struct workqueue_attrs *b)
3789	{
3790	if (a->nice != b->nice)
3791	return false;
3792	if (!cpumask_equal(a->cpumask, b->cpumask))
3793	return false;
3794	if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
3795	return false;
3796	if (a->affn_strict != b->affn_strict)
3797	return false;
3798	return true;
3799	}
3800
3801	/* Update @attrs with actually available CPUs */
3802	static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
3803	const cpumask_t *unbound_cpumask)
3804	{
3805	/*
3806	* Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
3807	* @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
3808	* @unbound_cpumask.
3809	*/
3810	cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
3811	if (unlikely(cpumask_empty(attrs->cpumask)))
3812	cpumask_copy(attrs->cpumask, unbound_cpumask);
3813	}
3814
3815	/* find wq_pod_type to use for @attrs */
3816	static const struct wq_pod_type *
3817	wqattrs_pod_type(const struct workqueue_attrs *attrs)
3818	{
3819	enum wq_affn_scope scope;
3820	struct wq_pod_type *pt;
3821
3822	/* to synchronize access to wq_affn_dfl */
3823	lockdep_assert_held(&wq_pool_mutex);
3824
3825	if (attrs->affn_scope == WQ_AFFN_DFL)
3826	scope = wq_affn_dfl;
3827	else
3828	scope = attrs->affn_scope;
3829
3830	pt = &wq_pod_types[scope];
3831
3832	if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
3833	likely(pt->nr_pods))
3834	return pt;
3835
3836	/*
3837	* Before workqueue_init_topology(), only SYSTEM is available which is
3838	* initialized in workqueue_init_early().
3839	*/
3840	pt = &wq_pod_types[WQ_AFFN_SYSTEM];
3841	BUG_ON(!pt->nr_pods);
3842	return pt;
3843	}
3844
3845	/**
3846	* init_worker_pool - initialize a newly zalloc'd worker_pool
3847	* @pool: worker_pool to initialize
3848	*
3849	* Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3850	*
3851	* Return: 0 on success, -errno on failure. Even on failure, all fields
3852	* inside @pool proper are initialized and put_unbound_pool() can be called
3853	* on @pool safely to release it.
3854	*/
3855	static int init_worker_pool(struct worker_pool *pool)
3856	{
3857	raw_spin_lock_init(&pool->lock);
3858	pool->id = -1;
3859	pool->cpu = -1;
3860	pool->node = NUMA_NO_NODE;
3861	pool->flags \|= POOL_DISASSOCIATED;
3862	pool->watchdog_ts = jiffies;
3863	INIT_LIST_HEAD(&pool->worklist);
3864	INIT_LIST_HEAD(&pool->idle_list);
3865	hash_init(pool->busy_hash);
3866
3867	timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3868	INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
3869
3870	timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3871
3872	INIT_LIST_HEAD(&pool->workers);
3873	INIT_LIST_HEAD(&pool->dying_workers);
3874
3875	ida_init(&pool->worker_ida);
3876	INIT_HLIST_NODE(&pool->hash_node);
3877	pool->refcnt = 1;
3878
3879	/* shouldn't fail above this point */
3880	pool->attrs = alloc_workqueue_attrs();
3881	if (!pool->attrs)
3882	return -ENOMEM;
3883
3884	wqattrs_clear_for_pool(pool->attrs);
3885
3886	return 0;
3887	}
3888
3889	#ifdef CONFIG_LOCKDEP
3890	static void wq_init_lockdep(struct workqueue_struct *wq)
3891	{
3892	char *lock_name;
3893
3894	lockdep_register_key(&wq->key);
3895	lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3896	if (!lock_name)
3897	lock_name = wq->name;
3898
3899	wq->lock_name = lock_name;
3900	lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3901	}
3902
3903	static void wq_unregister_lockdep(struct workqueue_struct *wq)
3904	{
3905	lockdep_unregister_key(&wq->key);
3906	}
3907
3908	static void wq_free_lockdep(struct workqueue_struct *wq)
3909	{
3910	if (wq->lock_name != wq->name)
3911	kfree(wq->lock_name);
3912	}
3913	#else
3914	static void wq_init_lockdep(struct workqueue_struct *wq)
3915	{
3916	}
3917
3918	static void wq_unregister_lockdep(struct workqueue_struct *wq)
3919	{
3920	}
3921
3922	static void wq_free_lockdep(struct workqueue_struct *wq)
3923	{
3924	}
3925	#endif
3926
3927	static void rcu_free_wq(struct rcu_head *rcu)
3928	{
3929	struct workqueue_struct *wq =
3930	container_of(rcu, struct workqueue_struct, rcu);
3931
3932	wq_free_lockdep(wq);
3933	free_percpu(wq->cpu_pwq);
3934	free_workqueue_attrs(wq->unbound_attrs);
3935	kfree(wq);
3936	}
3937
3938	static void rcu_free_pool(struct rcu_head *rcu)
3939	{
3940	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3941
3942	ida_destroy(&pool->worker_ida);
3943	free_workqueue_attrs(pool->attrs);
3944	kfree(pool);
3945	}
3946
3947	/**
3948	* put_unbound_pool - put a worker_pool
3949	* @pool: worker_pool to put
3950	*
3951	* Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3952	* safe manner. get_unbound_pool() calls this function on its failure path
3953	* and this function should be able to release pools which went through,
3954	* successfully or not, init_worker_pool().
3955	*
3956	* Should be called with wq_pool_mutex held.
3957	*/
3958	static void put_unbound_pool(struct worker_pool *pool)
3959	{
3960	DECLARE_COMPLETION_ONSTACK(detach_completion);
3961	struct worker *worker;
3962	LIST_HEAD(cull_list);
3963
3964	lockdep_assert_held(&wq_pool_mutex);
3965
3966	if (--pool->refcnt)
3967	return;
3968
3969	/* sanity checks */
3970	if (WARN_ON(!(pool->cpu < 0)) \|\|
3971	WARN_ON(!list_empty(&pool->worklist)))
3972	return;
3973
3974	/* release id and unhash */
3975	if (pool->id >= 0)
3976	idr_remove(&worker_pool_idr, pool->id);
3977	hash_del(&pool->hash_node);
3978
3979	/*
3980	* Become the manager and destroy all workers. This prevents
3981	* @pool's workers from blocking on attach_mutex. We're the last
3982	* manager and @pool gets freed with the flag set.
3983	*
3984	* Having a concurrent manager is quite unlikely to happen as we can
3985	* only get here with
3986	* pwq->refcnt == pool->refcnt == 0
3987	* which implies no work queued to the pool, which implies no worker can
3988	* become the manager. However a worker could have taken the role of
3989	* manager before the refcnts dropped to 0, since maybe_create_worker()
3990	* drops pool->lock
3991	*/
3992	while (true) {
3993	rcuwait_wait_event(&manager_wait,
3994	!(pool->flags & POOL_MANAGER_ACTIVE),
3995	TASK_UNINTERRUPTIBLE);
3996
3997	mutex_lock(&wq_pool_attach_mutex);
3998	raw_spin_lock_irq(&pool->lock);
3999	if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4000	pool->flags \|= POOL_MANAGER_ACTIVE;
4001	break;
4002	}
4003	raw_spin_unlock_irq(&pool->lock);
4004	mutex_unlock(&wq_pool_attach_mutex);
4005	}
4006
4007	while ((worker = first_idle_worker(pool)))
4008	set_worker_dying(worker, &cull_list);
4009	WARN_ON(pool->nr_workers \|\| pool->nr_idle);
4010	raw_spin_unlock_irq(&pool->lock);
4011
4012	wake_dying_workers(&cull_list);
4013
4014	if (!list_empty(&pool->workers) \|\| !list_empty(&pool->dying_workers))
4015	pool->detach_completion = &detach_completion;
4016	mutex_unlock(&wq_pool_attach_mutex);
4017
4018	if (pool->detach_completion)
4019	wait_for_completion(pool->detach_completion);
4020
4021	/* shut down the timers */
4022	del_timer_sync(&pool->idle_timer);
4023	cancel_work_sync(&pool->idle_cull_work);
4024	del_timer_sync(&pool->mayday_timer);
4025
4026	/* RCU protected to allow dereferences from get_work_pool() */
4027	call_rcu(&pool->rcu, rcu_free_pool);
4028	}
4029
4030	/**
4031	* get_unbound_pool - get a worker_pool with the specified attributes
4032	* @attrs: the attributes of the worker_pool to get
4033	*
4034	* Obtain a worker_pool which has the same attributes as @attrs, bump the
4035	* reference count and return it. If there already is a matching
4036	* worker_pool, it will be used; otherwise, this function attempts to
4037	* create a new one.
4038	*
4039	* Should be called with wq_pool_mutex held.
4040	*
4041	* Return: On success, a worker_pool with the same attributes as @attrs.
4042	* On failure, %NULL.
4043	*/
4044	static struct worker_pool get_unbound_pool(const struct workqueue_attrs attrs)
4045	{
4046	struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4047	u32 hash = wqattrs_hash(attrs);
4048	struct worker_pool *pool;
4049	int pod, node = NUMA_NO_NODE;
4050
4051	lockdep_assert_held(&wq_pool_mutex);
4052
4053	/* do we already have a matching pool? */
4054	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4055	if (wqattrs_equal(pool->attrs, attrs)) {
4056	pool->refcnt++;
4057	return pool;
4058	}
4059	}
4060
4061	/* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4062	for (pod = 0; pod < pt->nr_pods; pod++) {
4063	if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4064	node = pt->pod_node[pod];
4065	break;
4066	}
4067	}
4068
4069	/* nope, create a new one */
4070	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4071	if (!pool \|\| init_worker_pool(pool) < 0)
4072	goto fail;
4073
4074	pool->node = node;
4075	copy_workqueue_attrs(pool->attrs, attrs);
4076	wqattrs_clear_for_pool(pool->attrs);
4077
4078	if (worker_pool_assign_id(pool) < 0)
4079	goto fail;
4080
4081	/* create and start the initial worker */
4082	if (wq_online && !create_worker(pool))
4083	goto fail;
4084
4085	/* install */
4086	hash_add(unbound_pool_hash, &pool->hash_node, hash);
4087
4088	return pool;
4089	fail:
4090	if (pool)
4091	put_unbound_pool(pool);
4092	return NULL;
4093	}
4094
4095	static void rcu_free_pwq(struct rcu_head *rcu)
4096	{
4097	kmem_cache_free(pwq_cache,
4098	container_of(rcu, struct pool_workqueue, rcu));
4099	}
4100
4101	/*
4102	* Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
4103	* refcnt and needs to be destroyed.
4104	*/
4105	static void pwq_release_workfn(struct kthread_work *work)
4106	{
4107	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
4108	release_work);
4109	struct workqueue_struct *wq = pwq->wq;
4110	struct worker_pool *pool = pwq->pool;
4111	bool is_last = false;
4112
4113	/*
4114	* When @pwq is not linked, it doesn't hold any reference to the
4115	* @wq, and @wq is invalid to access.
4116	*/
4117	if (!list_empty(&pwq->pwqs_node)) {
4118	mutex_lock(&wq->mutex);
4119	list_del_rcu(&pwq->pwqs_node);
4120	is_last = list_empty(&wq->pwqs);
4121	mutex_unlock(&wq->mutex);
4122	}
4123
4124	if (wq->flags & WQ_UNBOUND) {
4125	mutex_lock(&wq_pool_mutex);
4126	put_unbound_pool(pool);
4127	mutex_unlock(&wq_pool_mutex);
4128	}
4129
4130	call_rcu(&pwq->rcu, rcu_free_pwq);
4131
4132	/*
4133	* If we're the last pwq going away, @wq is already dead and no one
4134	* is gonna access it anymore. Schedule RCU free.
4135	*/
4136	if (is_last) {
4137	wq_unregister_lockdep(wq);
4138	call_rcu(&wq->rcu, rcu_free_wq);
4139	}
4140	}
4141
4142	/**
4143	* pwq_adjust_max_active - update a pwq's max_active to the current setting
4144	* @pwq: target pool_workqueue
4145	*
4146	* If @pwq isn't freezing, set @pwq->max_active to the associated
4147	* workqueue's saved_max_active and activate inactive work items
4148	* accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
4149	*/
4150	static void pwq_adjust_max_active(struct pool_workqueue *pwq)
4151	{
4152	struct workqueue_struct *wq = pwq->wq;
4153	bool freezable = wq->flags & WQ_FREEZABLE;
4154	unsigned long flags;
4155
4156	/* for @wq->saved_max_active */
4157	lockdep_assert_held(&wq->mutex);
4158
4159	/* fast exit for non-freezable wqs */
4160	if (!freezable && pwq->max_active == wq->saved_max_active)
4161	return;
4162
4163	/* this function can be called during early boot w/ irq disabled */
4164	raw_spin_lock_irqsave(&pwq->pool->lock, flags);
4165
4166	/*
4167	* During [un]freezing, the caller is responsible for ensuring that
4168	* this function is called at least once after @workqueue_freezing
4169	* is updated and visible.
4170	*/
4171	if (!freezable \|\| !workqueue_freezing) {
4172	pwq->max_active = wq->saved_max_active;
4173
4174	while (!list_empty(&pwq->inactive_works) &&
4175	pwq->nr_active < pwq->max_active)
4176	pwq_activate_first_inactive(pwq);
4177
4178	kick_pool(pwq->pool);
4179	} else {
4180	pwq->max_active = 0;
4181	}
4182
4183	raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
4184	}
4185
4186	/* initialize newly allocated @pwq which is associated with @wq and @pool */
4187	static void init_pwq(struct pool_workqueue pwq, struct workqueue_struct wq,
4188	struct worker_pool *pool)
4189	{
4190	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
4191
4192	memset(pwq, 0, sizeof(*pwq));
4193
4194	pwq->pool = pool;
4195	pwq->wq = wq;
4196	pwq->flush_color = -1;
4197	pwq->refcnt = 1;
4198	INIT_LIST_HEAD(&pwq->inactive_works);
4199	INIT_LIST_HEAD(&pwq->pwqs_node);
4200	INIT_LIST_HEAD(&pwq->mayday_node);
4201	kthread_init_work(&pwq->release_work, pwq_release_workfn);
4202	}
4203
4204	/* sync @pwq with the current state of its associated wq and link it */
4205	static void link_pwq(struct pool_workqueue *pwq)
4206	{
4207	struct workqueue_struct *wq = pwq->wq;
4208
4209	lockdep_assert_held(&wq->mutex);
4210
4211	/* may be called multiple times, ignore if already linked */
4212	if (!list_empty(&pwq->pwqs_node))
4213	return;
4214
4215	/* set the matching work_color */
4216	pwq->work_color = wq->work_color;
4217
4218	/* sync max_active to the current setting */
4219	pwq_adjust_max_active(pwq);
4220
4221	/* link in @pwq */
4222	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
4223	}
4224
4225	/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
4226	static struct pool_workqueue alloc_unbound_pwq(struct workqueue_struct wq,
4227	const struct workqueue_attrs *attrs)
4228	{
4229	struct worker_pool *pool;
4230	struct pool_workqueue *pwq;
4231
4232	lockdep_assert_held(&wq_pool_mutex);
4233
4234	pool = get_unbound_pool(attrs);
4235	if (!pool)
4236	return NULL;
4237
4238	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
4239	if (!pwq) {
4240	put_unbound_pool(pool);
4241	return NULL;
4242	}
4243
4244	init_pwq(pwq, wq, pool);
4245	return pwq;
4246	}
4247
4248	/**
4249	* wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
4250	* @attrs: the wq_attrs of the default pwq of the target workqueue
4251	* @cpu: the target CPU
4252	* @cpu_going_down: if >= 0, the CPU to consider as offline
4253	*
4254	* Calculate the cpumask a workqueue with @attrs should use on @pod. If
4255	* @cpu_going_down is >= 0, that cpu is considered offline during calculation.
4256	* The result is stored in @attrs->__pod_cpumask.
4257	*
4258	* If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
4259	* and @pod has online CPUs requested by @attrs, the returned cpumask is the
4260	* intersection of the possible CPUs of @pod and @attrs->cpumask.
4261	*
4262	* The caller is responsible for ensuring that the cpumask of @pod stays stable.
4263	*/
4264	static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
4265	int cpu_going_down)
4266	{
4267	const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
4268	int pod = pt->cpu_pod[cpu];
4269
4270	/* does @pod have any online CPUs @attrs wants? */
4271	cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
4272	cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
4273	if (cpu_going_down >= 0)
4274	cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
4275
4276	if (cpumask_empty(attrs->__pod_cpumask)) {
4277	cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
4278	return;
4279	}
4280
4281	/* yeap, return possible CPUs in @pod that @attrs wants */
4282	cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
4283
4284	if (cpumask_empty(attrs->__pod_cpumask))
4285	pr_warn_once("WARNING: workqueue cpumask: online intersect > "
4286	"possible intersect\n");
4287	}
4288
4289	/* install @pwq into @wq's cpu_pwq and return the old pwq */
4290	static struct pool_workqueue install_unbound_pwq(struct workqueue_struct wq,
4291	int cpu, struct pool_workqueue *pwq)
4292	{
4293	struct pool_workqueue *old_pwq;
4294
4295	lockdep_assert_held(&wq_pool_mutex);
4296	lockdep_assert_held(&wq->mutex);
4297
4298	/* link_pwq() can handle duplicate calls */
4299	link_pwq(pwq);
4300
4301	old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
4302	rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq);
4303	return old_pwq;
4304	}
4305
4306	/* context to store the prepared attrs & pwqs before applying */
4307	struct apply_wqattrs_ctx {
4308	struct workqueue_struct wq; / target workqueue */
4309	struct workqueue_attrs attrs; / attrs to apply */
4310	struct list_head list; /* queued for batching commit */
4311	struct pool_workqueue *dfl_pwq;
4312	struct pool_workqueue *pwq_tbl[];
4313	};
4314
4315	/* free the resources after success or abort */
4316	static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
4317	{
4318	if (ctx) {
4319	int cpu;
4320
4321	for_each_possible_cpu(cpu)
4322	put_pwq_unlocked(ctx->pwq_tbl[cpu]);
4323	put_pwq_unlocked(ctx->dfl_pwq);
4324
4325	free_workqueue_attrs(ctx->attrs);
4326
4327	kfree(ctx);
4328	}
4329	}
4330
4331	/* allocate the attrs and pwqs for later installation */
4332	static struct apply_wqattrs_ctx *
4333	apply_wqattrs_prepare(struct workqueue_struct *wq,
4334	const struct workqueue_attrs *attrs,
4335	const cpumask_var_t unbound_cpumask)
4336	{
4337	struct apply_wqattrs_ctx *ctx;
4338	struct workqueue_attrs *new_attrs;
4339	int cpu;
4340
4341	lockdep_assert_held(&wq_pool_mutex);
4342
4343	if (WARN_ON(attrs->affn_scope < 0 \|\|
4344	attrs->affn_scope >= WQ_AFFN_NR_TYPES))
4345	return ERR_PTR(-EINVAL);
4346
4347	ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
4348
4349	new_attrs = alloc_workqueue_attrs();
4350	if (!ctx \|\| !new_attrs)
4351	goto out_free;
4352
4353	/*
4354	* If something goes wrong during CPU up/down, we'll fall back to
4355	* the default pwq covering whole @attrs->cpumask. Always create
4356	* it even if we don't use it immediately.
4357	*/
4358	copy_workqueue_attrs(new_attrs, attrs);
4359	wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
4360	cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4361	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
4362	if (!ctx->dfl_pwq)
4363	goto out_free;
4364
4365	for_each_possible_cpu(cpu) {
4366	if (new_attrs->ordered) {
4367	ctx->dfl_pwq->refcnt++;
4368	ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
4369	} else {
4370	wq_calc_pod_cpumask(new_attrs, cpu, -1);
4371	ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
4372	if (!ctx->pwq_tbl[cpu])
4373	goto out_free;
4374	}
4375	}
4376
4377	/* save the user configured attrs and sanitize it. */
4378	copy_workqueue_attrs(new_attrs, attrs);
4379	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
4380	cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4381	ctx->attrs = new_attrs;
4382
4383	ctx->wq = wq;
4384	return ctx;
4385
4386	out_free:
4387	free_workqueue_attrs(new_attrs);
4388	apply_wqattrs_cleanup(ctx);
4389	return ERR_PTR(-ENOMEM);
4390	}
4391
4392	/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
4393	static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4394	{
4395	int cpu;
4396
4397	/* all pwqs have been created successfully, let's install'em */
4398	mutex_lock(&ctx->wq->mutex);
4399
4400	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4401
4402	/* save the previous pwq and install the new one */
4403	for_each_possible_cpu(cpu)
4404	ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
4405	ctx->pwq_tbl[cpu]);
4406
4407	/* @dfl_pwq might not have been used, ensure it's linked */
4408	link_pwq(ctx->dfl_pwq);
4409	swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
4410
4411	mutex_unlock(&ctx->wq->mutex);
4412	}
4413
4414	static void apply_wqattrs_lock(void)
4415	{
4416	/* CPUs should stay stable across pwq creations and installations */
4417	cpus_read_lock();
4418	mutex_lock(&wq_pool_mutex);
4419	}
4420
4421	static void apply_wqattrs_unlock(void)
4422	{
4423	mutex_unlock(&wq_pool_mutex);
4424	cpus_read_unlock();
4425	}
4426
4427	static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4428	const struct workqueue_attrs *attrs)
4429	{
4430	struct apply_wqattrs_ctx *ctx;
4431
4432	/* only unbound workqueues can change attributes */
4433	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4434	return -EINVAL;
4435
4436	/* creating multiple pwqs breaks ordering guarantee */
4437	if (!list_empty(&wq->pwqs)) {
4438	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4439	return -EINVAL;
4440
4441	wq->flags &= ~__WQ_ORDERED;
4442	}
4443
4444	ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
4445	if (IS_ERR(ctx))
4446	return PTR_ERR(ctx);
4447
4448	/* the ctx has been prepared successfully, let's commit it */
4449	apply_wqattrs_commit(ctx);
4450	apply_wqattrs_cleanup(ctx);
4451
4452	return 0;
4453	}
4454
4455	/**
4456	* apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4457	* @wq: the target workqueue
4458	* @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4459	*
4460	* Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
4461	* a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
4462	* work items are affine to the pod it was issued on. Older pwqs are released as
4463	* in-flight work items finish. Note that a work item which repeatedly requeues
4464	* itself back-to-back will stay on its current pwq.
4465	*
4466	* Performs GFP_KERNEL allocations.
4467	*
4468	* Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
4469	*
4470	* Return: 0 on success and -errno on failure.
4471	*/
4472	int apply_workqueue_attrs(struct workqueue_struct *wq,
4473	const struct workqueue_attrs *attrs)
4474	{
4475	int ret;
4476
4477	lockdep_assert_cpus_held();
4478
4479	mutex_lock(&wq_pool_mutex);
4480	ret = apply_workqueue_attrs_locked(wq, attrs);
4481	mutex_unlock(&wq_pool_mutex);
4482
4483	return ret;
4484	}
4485
4486	/**
4487	* wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
4488	* @wq: the target workqueue
4489	* @cpu: the CPU to update pool association for
4490	* @hotplug_cpu: the CPU coming up or going down
4491	* @online: whether @cpu is coming up or going down
4492	*
4493	* This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4494	* %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of
4495	* @wq accordingly.
4496	*
4497	*
4498	* If pod affinity can't be adjusted due to memory allocation failure, it falls
4499	* back to @wq->dfl_pwq which may not be optimal but is always correct.
4500	*
4501	* Note that when the last allowed CPU of a pod goes offline for a workqueue
4502	* with a cpumask spanning multiple pods, the workers which were already
4503	* executing the work items for the workqueue will lose their CPU affinity and
4504	* may execute on any CPU. This is similar to how per-cpu workqueues behave on
4505	* CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
4506	* responsibility to flush the work item from CPU_DOWN_PREPARE.
4507	*/
4508	static void wq_update_pod(struct workqueue_struct *wq, int cpu,
4509	int hotplug_cpu, bool online)
4510	{
4511	int off_cpu = online ? -1 : hotplug_cpu;
4512	struct pool_workqueue old_pwq = NULL, pwq;
4513	struct workqueue_attrs *target_attrs;
4514
4515	lockdep_assert_held(&wq_pool_mutex);
4516
4517	if (!(wq->flags & WQ_UNBOUND) \|\| wq->unbound_attrs->ordered)
4518	return;
4519
4520	/*
4521	* We don't wanna alloc/free wq_attrs for each wq for each CPU.
4522	* Let's use a preallocated one. The following buf is protected by
4523	* CPU hotplug exclusion.
4524	*/
4525	target_attrs = wq_update_pod_attrs_buf;
4526
4527	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4528	wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
4529
4530	/* nothing to do if the target cpumask matches the current pwq */
4531	wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
4532	pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu),
4533	lockdep_is_held(&wq_pool_mutex));
4534	if (wqattrs_equal(target_attrs, pwq->pool->attrs))
4535	return;
4536
4537	/* create a new pwq */
4538	pwq = alloc_unbound_pwq(wq, target_attrs);
4539	if (!pwq) {
4540	pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
4541	wq->name);
4542	goto use_dfl_pwq;
4543	}
4544
4545	/* Install the new pwq. */
4546	mutex_lock(&wq->mutex);
4547	old_pwq = install_unbound_pwq(wq, cpu, pwq);
4548	goto out_unlock;
4549
4550	use_dfl_pwq:
4551	mutex_lock(&wq->mutex);
4552	raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
4553	get_pwq(wq->dfl_pwq);
4554	raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4555	old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq);
4556	out_unlock:
4557	mutex_unlock(&wq->mutex);
4558	put_pwq_unlocked(old_pwq);
4559	}
4560
4561	static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4562	{
4563	bool highpri = wq->flags & WQ_HIGHPRI;
4564	int cpu, ret;
4565
4566	wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
4567	if (!wq->cpu_pwq)
4568	goto enomem;
4569
4570	if (!(wq->flags & WQ_UNBOUND)) {
4571	for_each_possible_cpu(cpu) {
4572	struct pool_workqueue **pwq_p =
4573	per_cpu_ptr(wq->cpu_pwq, cpu);
4574	struct worker_pool *pool =
4575	&(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]);
4576
4577	*pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
4578	pool->node);
4579	if (!*pwq_p)
4580	goto enomem;
4581
4582	init_pwq(*pwq_p, wq, pool);
4583
4584	mutex_lock(&wq->mutex);
4585	link_pwq(*pwq_p);
4586	mutex_unlock(&wq->mutex);
4587	}
4588	return 0;
4589	}
4590
4591	cpus_read_lock();
4592	if (wq->flags & __WQ_ORDERED) {
4593	ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4594	/* there should only be single pwq for ordering guarantee */
4595	WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node \|\|
4596	wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4597	"ordering guarantee broken for workqueue %s\n", wq->name);
4598	} else {
4599	ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4600	}
4601	cpus_read_unlock();
4602
4603	return ret;
4604
4605	enomem:
4606	if (wq->cpu_pwq) {
4607	for_each_possible_cpu(cpu)
4608	kfree(*per_cpu_ptr(wq->cpu_pwq, cpu));
4609	free_percpu(wq->cpu_pwq);
4610	wq->cpu_pwq = NULL;
4611	}
4612	return -ENOMEM;
4613	}
4614
4615	static int wq_clamp_max_active(int max_active, unsigned int flags,
4616	const char *name)
4617	{
4618	if (max_active < 1 \|\| max_active > WQ_MAX_ACTIVE)
4619	pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4620	max_active, name, 1, WQ_MAX_ACTIVE);
4621
4622	return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
4623	}
4624
4625	/*
4626	* Workqueues which may be used during memory reclaim should have a rescuer
4627	* to guarantee forward progress.
4628	*/
4629	static int init_rescuer(struct workqueue_struct *wq)
4630	{
4631	struct worker *rescuer;
4632	int ret;
4633
4634	if (!(wq->flags & WQ_MEM_RECLAIM))
4635	return 0;
4636
4637	rescuer = alloc_worker(NUMA_NO_NODE);
4638	if (!rescuer) {
4639	pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
4640	wq->name);
4641	return -ENOMEM;
4642	}
4643
4644	rescuer->rescue_wq = wq;
4645	rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
4646	if (IS_ERR(rescuer->task)) {
4647	ret = PTR_ERR(rescuer->task);
4648	pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
4649	wq->name, ERR_PTR(ret));
4650	kfree(rescuer);
4651	return ret;
4652	}
4653
4654	wq->rescuer = rescuer;
4655	kthread_bind_mask(rescuer->task, cpu_possible_mask);
4656	wake_up_process(rescuer->task);
4657
4658	return 0;
4659	}
4660
4661	__printf(1, 4)
4662	struct workqueue_struct alloc_workqueue(const char fmt,
4663	unsigned int flags,
4664	int max_active, ...)
4665	{
4666	va_list args;
4667	struct workqueue_struct *wq;
4668	struct pool_workqueue *pwq;
4669
4670	/*
4671	* Unbound && max_active == 1 used to imply ordered, which is no longer
4672	* the case on many machines due to per-pod pools. While
4673	* alloc_ordered_workqueue() is the right way to create an ordered
4674	* workqueue, keep the previous behavior to avoid subtle breakages.
4675	*/
4676	if ((flags & WQ_UNBOUND) && max_active == 1)
4677	flags \|= __WQ_ORDERED;
4678
4679	/* see the comment above the definition of WQ_POWER_EFFICIENT */
4680	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4681	flags \|= WQ_UNBOUND;
4682
4683	/* allocate wq and format name */
4684	wq = kzalloc(sizeof(*wq), GFP_KERNEL);
4685	if (!wq)
4686	return NULL;
4687
4688	if (flags & WQ_UNBOUND) {
4689	wq->unbound_attrs = alloc_workqueue_attrs();
4690	if (!wq->unbound_attrs)
4691	goto err_free_wq;
4692	}
4693
4694	va_start(args, max_active);
4695	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4696	va_end(args);
4697
4698	max_active = max_active ?: WQ_DFL_ACTIVE;
4699	max_active = wq_clamp_max_active(max_active, flags, wq->name);
4700
4701	/* init wq */
4702	wq->flags = flags;
4703	wq->saved_max_active = max_active;
4704	mutex_init(&wq->mutex);
4705	atomic_set(&wq->nr_pwqs_to_flush, 0);
4706	INIT_LIST_HEAD(&wq->pwqs);
4707	INIT_LIST_HEAD(&wq->flusher_queue);
4708	INIT_LIST_HEAD(&wq->flusher_overflow);
4709	INIT_LIST_HEAD(&wq->maydays);
4710
4711	wq_init_lockdep(wq);
4712	INIT_LIST_HEAD(&wq->list);
4713
4714	if (alloc_and_link_pwqs(wq) < 0)
4715	goto err_unreg_lockdep;
4716
4717	if (wq_online && init_rescuer(wq) < 0)
4718	goto err_destroy;
4719
4720	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4721	goto err_destroy;
4722
4723	/*
4724	* wq_pool_mutex protects global freeze state and workqueues list.
4725	* Grab it, adjust max_active and add the new @wq to workqueues
4726	* list.
4727	*/
4728	mutex_lock(&wq_pool_mutex);
4729
4730	mutex_lock(&wq->mutex);
4731	for_each_pwq(pwq, wq)
4732	pwq_adjust_max_active(pwq);
4733	mutex_unlock(&wq->mutex);
4734
4735	list_add_tail_rcu(&wq->list, &workqueues);
4736
4737	mutex_unlock(&wq_pool_mutex);
4738
4739	return wq;
4740
4741	err_unreg_lockdep:
4742	wq_unregister_lockdep(wq);
4743	wq_free_lockdep(wq);
4744	err_free_wq:
4745	free_workqueue_attrs(wq->unbound_attrs);
4746	kfree(wq);
4747	return NULL;
4748	err_destroy:
4749	destroy_workqueue(wq);
4750	return NULL;
4751	}
4752	EXPORT_SYMBOL_GPL(alloc_workqueue);
4753
4754	static bool pwq_busy(struct pool_workqueue *pwq)
4755	{
4756	int i;
4757
4758	for (i = 0; i < WORK_NR_COLORS; i++)
4759	if (pwq->nr_in_flight[i])
4760	return true;
4761
4762	if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4763	return true;
4764	if (pwq->nr_active \|\| !list_empty(&pwq->inactive_works))
4765	return true;
4766
4767	return false;
4768	}
4769
4770	/**
4771	* destroy_workqueue - safely terminate a workqueue
4772	* @wq: target workqueue
4773	*
4774	* Safely destroy a workqueue. All work currently pending will be done first.
4775	*/
4776	void destroy_workqueue(struct workqueue_struct *wq)
4777	{
4778	struct pool_workqueue *pwq;
4779	int cpu;
4780
4781	/*
4782	* Remove it from sysfs first so that sanity check failure doesn't
4783	* lead to sysfs name conflicts.
4784	*/
4785	workqueue_sysfs_unregister(wq);
4786
4787	/* mark the workqueue destruction is in progress */
4788	mutex_lock(&wq->mutex);
4789	wq->flags \|= __WQ_DESTROYING;
4790	mutex_unlock(&wq->mutex);
4791
4792	/* drain it before proceeding with destruction */
4793	drain_workqueue(wq);
4794
4795	/* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4796	if (wq->rescuer) {
4797	struct worker *rescuer = wq->rescuer;
4798
4799	/* this prevents new queueing */
4800	raw_spin_lock_irq(&wq_mayday_lock);
4801	wq->rescuer = NULL;
4802	raw_spin_unlock_irq(&wq_mayday_lock);
4803
4804	/* rescuer will empty maydays list before exiting */
4805	kthread_stop(rescuer->task);
4806	kfree(rescuer);
4807	}
4808
4809	/*
4810	* Sanity checks - grab all the locks so that we wait for all
4811	* in-flight operations which may do put_pwq().
4812	*/
4813	mutex_lock(&wq_pool_mutex);
4814	mutex_lock(&wq->mutex);
4815	for_each_pwq(pwq, wq) {
4816	raw_spin_lock_irq(&pwq->pool->lock);
4817	if (WARN_ON(pwq_busy(pwq))) {
4818	pr_warn("%s: %s has the following busy pwq\n",
4819	__func__, wq->name);
4820	show_pwq(pwq);
4821	raw_spin_unlock_irq(&pwq->pool->lock);
4822	mutex_unlock(&wq->mutex);
4823	mutex_unlock(&wq_pool_mutex);
4824	show_one_workqueue(wq);
4825	return;
4826	}
4827	raw_spin_unlock_irq(&pwq->pool->lock);
4828	}
4829	mutex_unlock(&wq->mutex);
4830
4831	/*
4832	* wq list is used to freeze wq, remove from list after
4833	* flushing is complete in case freeze races us.
4834	*/
4835	list_del_rcu(&wq->list);
4836	mutex_unlock(&wq_pool_mutex);
4837
4838	/*
4839	* We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
4840	* to put the base refs. @wq will be auto-destroyed from the last
4841	* pwq_put. RCU read lock prevents @wq from going away from under us.
4842	*/
4843	rcu_read_lock();
4844
4845	for_each_possible_cpu(cpu) {
4846	pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
4847	RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL);
4848	put_pwq_unlocked(pwq);
4849	}
4850
4851	put_pwq_unlocked(wq->dfl_pwq);
4852	wq->dfl_pwq = NULL;
4853
4854	rcu_read_unlock();
4855	}
4856	EXPORT_SYMBOL_GPL(destroy_workqueue);
4857
4858	/**
4859	* workqueue_set_max_active - adjust max_active of a workqueue
4860	* @wq: target workqueue
4861	* @max_active: new max_active value.
4862	*
4863	* Set max_active of @wq to @max_active.
4864	*
4865	* CONTEXT:
4866	* Don't call from IRQ context.
4867	*/
4868	void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4869	{
4870	struct pool_workqueue *pwq;
4871
4872	/* disallow meddling with max_active for ordered workqueues */
4873	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4874	return;
4875
4876	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4877
4878	mutex_lock(&wq->mutex);
4879
4880	wq->flags &= ~__WQ_ORDERED;
4881	wq->saved_max_active = max_active;
4882
4883	for_each_pwq(pwq, wq)
4884	pwq_adjust_max_active(pwq);
4885
4886	mutex_unlock(&wq->mutex);
4887	}
4888	EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4889
4890	/**
4891	* current_work - retrieve %current task's work struct
4892	*
4893	* Determine if %current task is a workqueue worker and what it's working on.
4894	* Useful to find out the context that the %current task is running in.
4895	*
4896	* Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4897	*/
4898	struct work_struct *current_work(void)
4899	{
4900	struct worker *worker = current_wq_worker();
4901
4902	return worker ? worker->current_work : NULL;
4903	}
4904	EXPORT_SYMBOL(current_work);
4905
4906	/**
4907	* current_is_workqueue_rescuer - is %current workqueue rescuer?
4908	*
4909	* Determine whether %current is a workqueue rescuer. Can be used from
4910	* work functions to determine whether it's being run off the rescuer task.
4911	*
4912	* Return: %true if %current is a workqueue rescuer. %false otherwise.
4913	*/
4914	bool current_is_workqueue_rescuer(void)
4915	{
4916	struct worker *worker = current_wq_worker();
4917
4918	return worker && worker->rescue_wq;
4919	}
4920
4921	/**
4922	* workqueue_congested - test whether a workqueue is congested
4923	* @cpu: CPU in question
4924	* @wq: target workqueue
4925	*
4926	* Test whether @wq's cpu workqueue for @cpu is congested. There is
4927	* no synchronization around this function and the test result is
4928	* unreliable and only useful as advisory hints or for debugging.
4929	*
4930	* If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4931	*
4932	* With the exception of ordered workqueues, all workqueues have per-cpu
4933	* pool_workqueues, each with its own congested state. A workqueue being
4934	* congested on one CPU doesn't mean that the workqueue is contested on any
4935	* other CPUs.
4936	*
4937	* Return:
4938	* %true if congested, %false otherwise.
4939	*/
4940	bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4941	{
4942	struct pool_workqueue *pwq;
4943	bool ret;
4944
4945	rcu_read_lock();
4946	preempt_disable();
4947
4948	if (cpu == WORK_CPU_UNBOUND)
4949	cpu = smp_processor_id();
4950
4951	pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
4952	ret = !list_empty(&pwq->inactive_works);
4953
4954	preempt_enable();
4955	rcu_read_unlock();
4956
4957	return ret;
4958	}
4959	EXPORT_SYMBOL_GPL(workqueue_congested);
4960
4961	/**
4962	* work_busy - test whether a work is currently pending or running
4963	* @work: the work to be tested
4964	*
4965	* Test whether @work is currently pending or running. There is no
4966	* synchronization around this function and the test result is
4967	* unreliable and only useful as advisory hints or for debugging.
4968	*
4969	* Return:
4970	* OR'd bitmask of WORK_BUSY_* bits.
4971	*/
4972	unsigned int work_busy(struct work_struct *work)
4973	{
4974	struct worker_pool *pool;
4975	unsigned long flags;
4976	unsigned int ret = 0;
4977
4978	if (work_pending(work))
4979	ret \|= WORK_BUSY_PENDING;
4980
4981	rcu_read_lock();
4982	pool = get_work_pool(work);
4983	if (pool) {
4984	raw_spin_lock_irqsave(&pool->lock, flags);
4985	if (find_worker_executing_work(pool, work))
4986	ret \|= WORK_BUSY_RUNNING;
4987	raw_spin_unlock_irqrestore(&pool->lock, flags);
4988	}
4989	rcu_read_unlock();
4990
4991	return ret;
4992	}
4993	EXPORT_SYMBOL_GPL(work_busy);
4994
4995	/**
4996	* set_worker_desc - set description for the current work item
4997	* @fmt: printf-style format string
4998	* @...: arguments for the format string
4999	*
5000	* This function can be called by a running work function to describe what
5001	* the work item is about. If the worker task gets dumped, this
5002	* information will be printed out together to help debugging. The
5003	* description can be at most WORKER_DESC_LEN including the trailing '\0'.
5004	*/
5005	void set_worker_desc(const char *fmt, ...)
5006	{
5007	struct worker *worker = current_wq_worker();
5008	va_list args;
5009
5010	if (worker) {
5011	va_start(args, fmt);
5012	vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
5013	va_end(args);
5014	}
5015	}
5016	EXPORT_SYMBOL_GPL(set_worker_desc);
5017
5018	/**
5019	* print_worker_info - print out worker information and description
5020	* @log_lvl: the log level to use when printing
5021	* @task: target task
5022	*
5023	* If @task is a worker and currently executing a work item, print out the
5024	* name of the workqueue being serviced and worker description set with
5025	* set_worker_desc() by the currently executing work item.
5026	*
5027	* This function can be safely called on any task as long as the
5028	* task_struct itself is accessible. While safe, this function isn't
5029	* synchronized and may print out mixups or garbages of limited length.
5030	*/
5031	void print_worker_info(const char log_lvl, struct task_struct task)
5032	{
5033	work_func_t *fn = NULL;
5034	char name[WQ_NAME_LEN] = { };
5035	char desc[WORKER_DESC_LEN] = { };
5036	struct pool_workqueue *pwq = NULL;
5037	struct workqueue_struct *wq = NULL;
5038	struct worker *worker;
5039
5040	if (!(task->flags & PF_WQ_WORKER))
5041	return;
5042
5043	/*
5044	* This function is called without any synchronization and @task
5045	* could be in any state. Be careful with dereferences.
5046	*/
5047	worker = kthread_probe_data(task);
5048
5049	/*
5050	* Carefully copy the associated workqueue's workfn, name and desc.
5051	* Keep the original last '\0' in case the original is garbage.
5052	*/
5053	copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
5054	copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
5055	copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
5056	copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
5057	copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
5058
5059	if (fn \|\| name[0] \|\| desc[0]) {
5060	printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
5061	if (strcmp(name, desc))
5062	pr_cont(" (%s)", desc);
5063	pr_cont("\n");
5064	}
5065	}
5066
5067	static void pr_cont_pool_info(struct worker_pool *pool)
5068	{
5069	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
5070	if (pool->node != NUMA_NO_NODE)
5071	pr_cont(" node=%d", pool->node);
5072	pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
5073	}
5074
5075	struct pr_cont_work_struct {
5076	bool comma;
5077	work_func_t func;
5078	long ctr;
5079	};
5080
5081	static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
5082	{
5083	if (!pcwsp->ctr)
5084	goto out_record;
5085	if (func == pcwsp->func) {
5086	pcwsp->ctr++;
5087	return;
5088	}
5089	if (pcwsp->ctr == 1)
5090	pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
5091	else
5092	pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
5093	pcwsp->ctr = 0;
5094	out_record:
5095	if ((long)func == -1L)
5096	return;
5097	pcwsp->comma = comma;
5098	pcwsp->func = func;
5099	pcwsp->ctr = 1;
5100	}
5101
5102	static void pr_cont_work(bool comma, struct work_struct work, struct pr_cont_work_struct pcwsp)
5103	{
5104	if (work->func == wq_barrier_func) {
5105	struct wq_barrier *barr;
5106
5107	barr = container_of(work, struct wq_barrier, work);
5108
5109	pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5110	pr_cont("%s BAR(%d)", comma ? "," : "",
5111	task_pid_nr(barr->task));
5112	} else {
5113	if (!comma)
5114	pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5115	pr_cont_work_flush(comma, work->func, pcwsp);
5116	}
5117	}
5118
5119	static void show_pwq(struct pool_workqueue *pwq)
5120	{
5121	struct pr_cont_work_struct pcws = { .ctr = 0, };
5122	struct worker_pool *pool = pwq->pool;
5123	struct work_struct *work;
5124	struct worker *worker;
5125	bool has_in_flight = false, has_pending = false;
5126	int bkt;
5127
5128	pr_info(" pwq %d:", pool->id);
5129	pr_cont_pool_info(pool);
5130
5131	pr_cont(" active=%d/%d refcnt=%d%s\n",
5132	pwq->nr_active, pwq->max_active, pwq->refcnt,
5133	!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
5134
5135	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5136	if (worker->current_pwq == pwq) {
5137	has_in_flight = true;
5138	break;
5139	}
5140	}
5141	if (has_in_flight) {
5142	bool comma = false;
5143
5144	pr_info(" in-flight:");
5145	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5146	if (worker->current_pwq != pwq)
5147	continue;
5148
5149	pr_cont("%s %d%s:%ps", comma ? "," : "",
5150	task_pid_nr(worker->task),
5151	worker->rescue_wq ? "(RESCUER)" : "",
5152	worker->current_func);
5153	list_for_each_entry(work, &worker->scheduled, entry)
5154	pr_cont_work(false, work, &pcws);
5155	pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5156	comma = true;
5157	}
5158	pr_cont("\n");
5159	}
5160
5161	list_for_each_entry(work, &pool->worklist, entry) {
5162	if (get_work_pwq(work) == pwq) {
5163	has_pending = true;
5164	break;
5165	}
5166	}
5167	if (has_pending) {
5168	bool comma = false;
5169
5170	pr_info(" pending:");
5171	list_for_each_entry(work, &pool->worklist, entry) {
5172	if (get_work_pwq(work) != pwq)
5173	continue;
5174
5175	pr_cont_work(comma, work, &pcws);
5176	comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5177	}
5178	pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5179	pr_cont("\n");
5180	}
5181
5182	if (!list_empty(&pwq->inactive_works)) {
5183	bool comma = false;
5184
5185	pr_info(" inactive:");
5186	list_for_each_entry(work, &pwq->inactive_works, entry) {
5187	pr_cont_work(comma, work, &pcws);
5188	comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5189	}
5190	pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5191	pr_cont("\n");
5192	}
5193	}
5194
5195	/**
5196	* show_one_workqueue - dump state of specified workqueue
5197	* @wq: workqueue whose state will be printed
5198	*/
5199	void show_one_workqueue(struct workqueue_struct *wq)
5200	{
5201	struct pool_workqueue *pwq;
5202	bool idle = true;
5203	unsigned long flags;
5204
5205	for_each_pwq(pwq, wq) {
5206	if (pwq->nr_active \|\| !list_empty(&pwq->inactive_works)) {
5207	idle = false;
5208	break;
5209	}
5210	}
5211	if (idle) /* Nothing to print for idle workqueue */
5212	return;
5213
5214	pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
5215
5216	for_each_pwq(pwq, wq) {
5217	raw_spin_lock_irqsave(&pwq->pool->lock, flags);
5218	if (pwq->nr_active \|\| !list_empty(&pwq->inactive_works)) {
5219	/*
5220	* Defer printing to avoid deadlocks in console
5221	* drivers that queue work while holding locks
5222	* also taken in their write paths.
5223	*/
5224	printk_deferred_enter();
5225	show_pwq(pwq);
5226	printk_deferred_exit();
5227	}
5228	raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
5229	/*
5230	* We could be printing a lot from atomic context, e.g.
5231	* sysrq-t -> show_all_workqueues(). Avoid triggering
5232	* hard lockup.
5233	*/
5234	touch_nmi_watchdog();
5235	}
5236
5237	}
5238
5239	/**
5240	* show_one_worker_pool - dump state of specified worker pool
5241	* @pool: worker pool whose state will be printed
5242	*/
5243	static void show_one_worker_pool(struct worker_pool *pool)
5244	{
5245	struct worker *worker;
5246	bool first = true;
5247	unsigned long flags;
5248	unsigned long hung = 0;
5249
5250	raw_spin_lock_irqsave(&pool->lock, flags);
5251	if (pool->nr_workers == pool->nr_idle)
5252	goto next_pool;
5253
5254	/* How long the first pending work is waiting for a worker. */
5255	if (!list_empty(&pool->worklist))
5256	hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
5257
5258	/*
5259	* Defer printing to avoid deadlocks in console drivers that
5260	* queue work while holding locks also taken in their write
5261	* paths.
5262	*/
5263	printk_deferred_enter();
5264	pr_info("pool %d:", pool->id);
5265	pr_cont_pool_info(pool);
5266	pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
5267	if (pool->manager)
5268	pr_cont(" manager: %d",
5269	task_pid_nr(pool->manager->task));
5270	list_for_each_entry(worker, &pool->idle_list, entry) {
5271	pr_cont(" %s%d", first ? "idle: " : "",
5272	task_pid_nr(worker->task));
5273	first = false;
5274	}
5275	pr_cont("\n");
5276	printk_deferred_exit();
5277	next_pool:
5278	raw_spin_unlock_irqrestore(&pool->lock, flags);
5279	/*
5280	* We could be printing a lot from atomic context, e.g.
5281	* sysrq-t -> show_all_workqueues(). Avoid triggering
5282	* hard lockup.
5283	*/
5284	touch_nmi_watchdog();
5285
5286	}
5287
5288	/**
5289	* show_all_workqueues - dump workqueue state
5290	*
5291	* Called from a sysrq handler and prints out all busy workqueues and pools.
5292	*/
5293	void show_all_workqueues(void)
5294	{
5295	struct workqueue_struct *wq;
5296	struct worker_pool *pool;
5297	int pi;
5298
5299	rcu_read_lock();
5300
5301	pr_info("Showing busy workqueues and worker pools:\n");
5302
5303	list_for_each_entry_rcu(wq, &workqueues, list)
5304	show_one_workqueue(wq);
5305
5306	for_each_pool(pool, pi)
5307	show_one_worker_pool(pool);
5308
5309	rcu_read_unlock();
5310	}
5311
5312	/**
5313	* show_freezable_workqueues - dump freezable workqueue state
5314	*
5315	* Called from try_to_freeze_tasks() and prints out all freezable workqueues
5316	* still busy.
5317	*/
5318	void show_freezable_workqueues(void)
5319	{
5320	struct workqueue_struct *wq;
5321
5322	rcu_read_lock();
5323
5324	pr_info("Showing freezable workqueues that are still busy:\n");
5325
5326	list_for_each_entry_rcu(wq, &workqueues, list) {
5327	if (!(wq->flags & WQ_FREEZABLE))
5328	continue;
5329	show_one_workqueue(wq);
5330	}
5331
5332	rcu_read_unlock();
5333	}
5334
5335	/* used to show worker information through /proc/PID/{comm,stat,status} */
5336	void wq_worker_comm(char buf, size_t size, struct task_struct task)
5337	{
5338	int off;
5339
5340	/* always show the actual comm */
5341	off = strscpy(buf, task->comm, size);
5342	if (off < 0)
5343	return;
5344
5345	/* stabilize PF_WQ_WORKER and worker pool association */
5346	mutex_lock(&wq_pool_attach_mutex);
5347
5348	if (task->flags & PF_WQ_WORKER) {
5349	struct worker *worker = kthread_data(task);
5350	struct worker_pool *pool = worker->pool;
5351
5352	if (pool) {
5353	raw_spin_lock_irq(&pool->lock);
5354	/*
5355	* ->desc tracks information (wq name or
5356	* set_worker_desc()) for the latest execution. If
5357	* current, prepend '+', otherwise '-'.
5358	*/
5359	if (worker->desc[0] != '\0') {
5360	if (worker->current_work)
5361	scnprintf(buf + off, size - off, "+%s",
5362	worker->desc);
5363	else
5364	scnprintf(buf + off, size - off, "-%s",
5365	worker->desc);
5366	}
5367	raw_spin_unlock_irq(&pool->lock);
5368	}
5369	}
5370
5371	mutex_unlock(&wq_pool_attach_mutex);
5372	}
5373
5374	#ifdef CONFIG_SMP
5375
5376	/*
5377	* CPU hotplug.
5378	*
5379	* There are two challenges in supporting CPU hotplug. Firstly, there
5380	* are a lot of assumptions on strong associations among work, pwq and
5381	* pool which make migrating pending and scheduled works very
5382	* difficult to implement without impacting hot paths. Secondly,
5383	* worker pools serve mix of short, long and very long running works making
5384	* blocked draining impractical.
5385	*
5386	* This is solved by allowing the pools to be disassociated from the CPU
5387	* running as an unbound one and allowing it to be reattached later if the
5388	* cpu comes back online.
5389	*/
5390
5391	static void unbind_workers(int cpu)
5392	{
5393	struct worker_pool *pool;
5394	struct worker *worker;
5395
5396	for_each_cpu_worker_pool(pool, cpu) {
5397	mutex_lock(&wq_pool_attach_mutex);
5398	raw_spin_lock_irq(&pool->lock);
5399
5400	/*
5401	* We've blocked all attach/detach operations. Make all workers
5402	* unbound and set DISASSOCIATED. Before this, all workers
5403	* must be on the cpu. After this, they may become diasporas.
5404	* And the preemption disabled section in their sched callbacks
5405	* are guaranteed to see WORKER_UNBOUND since the code here
5406	* is on the same cpu.
5407	*/
5408	for_each_pool_worker(worker, pool)
5409	worker->flags \|= WORKER_UNBOUND;
5410
5411	pool->flags \|= POOL_DISASSOCIATED;
5412
5413	/*
5414	* The handling of nr_running in sched callbacks are disabled
5415	* now. Zap nr_running. After this, nr_running stays zero and
5416	* need_more_worker() and keep_working() are always true as
5417	* long as the worklist is not empty. This pool now behaves as
5418	* an unbound (in terms of concurrency management) pool which
5419	* are served by workers tied to the pool.
5420	*/
5421	pool->nr_running = 0;
5422
5423	/*
5424	* With concurrency management just turned off, a busy
5425	* worker blocking could lead to lengthy stalls. Kick off
5426	* unbound chain execution of currently pending work items.
5427	*/
5428	kick_pool(pool);
5429
5430	raw_spin_unlock_irq(&pool->lock);
5431
5432	for_each_pool_worker(worker, pool)
5433	unbind_worker(worker);
5434
5435	mutex_unlock(&wq_pool_attach_mutex);
5436	}
5437	}
5438
5439	/**
5440	* rebind_workers - rebind all workers of a pool to the associated CPU
5441	* @pool: pool of interest
5442	*
5443	* @pool->cpu is coming online. Rebind all workers to the CPU.
5444	*/
5445	static void rebind_workers(struct worker_pool *pool)
5446	{
5447	struct worker *worker;
5448
5449	lockdep_assert_held(&wq_pool_attach_mutex);
5450
5451	/*
5452	* Restore CPU affinity of all workers. As all idle workers should
5453	* be on the run-queue of the associated CPU before any local
5454	* wake-ups for concurrency management happen, restore CPU affinity
5455	* of all workers first and then clear UNBOUND. As we're called
5456	* from CPU_ONLINE, the following shouldn't fail.
5457	*/
5458	for_each_pool_worker(worker, pool) {
5459	kthread_set_per_cpu(worker->task, pool->cpu);
5460	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
5461	pool_allowed_cpus(pool)) < 0);
5462	}
5463
5464	raw_spin_lock_irq(&pool->lock);
5465
5466	pool->flags &= ~POOL_DISASSOCIATED;
5467
5468	for_each_pool_worker(worker, pool) {
5469	unsigned int worker_flags = worker->flags;
5470
5471	/*
5472	* We want to clear UNBOUND but can't directly call
5473	* worker_clr_flags() or adjust nr_running. Atomically
5474	* replace UNBOUND with another NOT_RUNNING flag REBOUND.
5475	* @worker will clear REBOUND using worker_clr_flags() when
5476	* it initiates the next execution cycle thus restoring
5477	* concurrency management. Note that when or whether
5478	* @worker clears REBOUND doesn't affect correctness.
5479	*
5480	* WRITE_ONCE() is necessary because @worker->flags may be
5481	* tested without holding any lock in
5482	* wq_worker_running(). Without it, NOT_RUNNING test may
5483	* fail incorrectly leading to premature concurrency
5484	* management operations.
5485	*/
5486	WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
5487	worker_flags \|= WORKER_REBOUND;
5488	worker_flags &= ~WORKER_UNBOUND;
5489	WRITE_ONCE(worker->flags, worker_flags);
5490	}
5491
5492	raw_spin_unlock_irq(&pool->lock);
5493	}
5494
5495	/**
5496	* restore_unbound_workers_cpumask - restore cpumask of unbound workers
5497	* @pool: unbound pool of interest
5498	* @cpu: the CPU which is coming up
5499	*
5500	* An unbound pool may end up with a cpumask which doesn't have any online
5501	* CPUs. When a worker of such pool get scheduled, the scheduler resets
5502	* its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5503	* online CPU before, cpus_allowed of all its workers should be restored.
5504	*/
5505	static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5506	{
5507	static cpumask_t cpumask;
5508	struct worker *worker;
5509
5510	lockdep_assert_held(&wq_pool_attach_mutex);
5511
5512	/* is @cpu allowed for @pool? */
5513	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5514	return;
5515
5516	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5517
5518	/* as we're called from CPU_ONLINE, the following shouldn't fail */
5519	for_each_pool_worker(worker, pool)
5520	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5521	}
5522
5523	int workqueue_prepare_cpu(unsigned int cpu)
5524	{
5525	struct worker_pool *pool;
5526
5527	for_each_cpu_worker_pool(pool, cpu) {
5528	if (pool->nr_workers)
5529	continue;
5530	if (!create_worker(pool))
5531	return -ENOMEM;
5532	}
5533	return 0;
5534	}
5535
5536	int workqueue_online_cpu(unsigned int cpu)
5537	{
5538	struct worker_pool *pool;
5539	struct workqueue_struct *wq;
5540	int pi;
5541
5542	mutex_lock(&wq_pool_mutex);
5543
5544	for_each_pool(pool, pi) {
5545	mutex_lock(&wq_pool_attach_mutex);
5546
5547	if (pool->cpu == cpu)
5548	rebind_workers(pool);
5549	else if (pool->cpu < 0)
5550	restore_unbound_workers_cpumask(pool, cpu);
5551
5552	mutex_unlock(&wq_pool_attach_mutex);
5553	}
5554
5555	/* update pod affinity of unbound workqueues */
5556	list_for_each_entry(wq, &workqueues, list) {
5557	struct workqueue_attrs *attrs = wq->unbound_attrs;
5558
5559	if (attrs) {
5560	const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5561	int tcpu;
5562
5563	for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
5564	wq_update_pod(wq, tcpu, cpu, true);
5565	}
5566	}
5567
5568	mutex_unlock(&wq_pool_mutex);
5569	return 0;
5570	}
5571
5572	int workqueue_offline_cpu(unsigned int cpu)
5573	{
5574	struct workqueue_struct *wq;
5575
5576	/* unbinding per-cpu workers should happen on the local CPU */
5577	if (WARN_ON(cpu != smp_processor_id()))
5578	return -1;
5579
5580	unbind_workers(cpu);
5581
5582	/* update pod affinity of unbound workqueues */
5583	mutex_lock(&wq_pool_mutex);
5584	list_for_each_entry(wq, &workqueues, list) {
5585	struct workqueue_attrs *attrs = wq->unbound_attrs;
5586
5587	if (attrs) {
5588	const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5589	int tcpu;
5590
5591	for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
5592	wq_update_pod(wq, tcpu, cpu, false);
5593	}
5594	}
5595	mutex_unlock(&wq_pool_mutex);
5596
5597	return 0;
5598	}
5599
5600	struct work_for_cpu {
5601	struct work_struct work;
5602	long (fn)(void );
5603	void *arg;
5604	long ret;
5605	};
5606
5607	static void work_for_cpu_fn(struct work_struct *work)
5608	{
5609	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5610
5611	wfc->ret = wfc->fn(wfc->arg);
5612	}
5613
5614	/**
5615	* work_on_cpu - run a function in thread context on a particular cpu
5616	* @cpu: the cpu to run on
5617	* @fn: the function to run
5618	* @arg: the function arg
5619	*
5620	* It is up to the caller to ensure that the cpu doesn't go offline.
5621	* The caller must not hold any locks which would prevent @fn from completing.
5622	*
5623	* Return: The value @fn returns.
5624	*/
5625	long work_on_cpu(int cpu, long (fn)(void ), void *arg)
5626	{
5627	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5628
5629	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5630	schedule_work_on(cpu, &wfc.work);
5631	flush_work(&wfc.work);
5632	destroy_work_on_stack(&wfc.work);
5633	return wfc.ret;
5634	}
5635	EXPORT_SYMBOL_GPL(work_on_cpu);
5636
5637	/**
5638	* work_on_cpu_safe - run a function in thread context on a particular cpu
5639	* @cpu: the cpu to run on
5640	* @fn: the function to run
5641	* @arg: the function argument
5642	*
5643	* Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5644	* any locks which would prevent @fn from completing.
5645	*
5646	* Return: The value @fn returns.
5647	*/
5648	long work_on_cpu_safe(int cpu, long (fn)(void ), void *arg)
5649	{
5650	long ret = -ENODEV;
5651
5652	cpus_read_lock();
5653	if (cpu_online(cpu))
5654	ret = work_on_cpu(cpu, fn, arg);
5655	cpus_read_unlock();
5656	return ret;
5657	}
5658	EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5659	#endif /* CONFIG_SMP */
5660
5661	#ifdef CONFIG_FREEZER
5662
5663	/**
5664	* freeze_workqueues_begin - begin freezing workqueues
5665	*
5666	* Start freezing workqueues. After this function returns, all freezable
5667	* workqueues will queue new works to their inactive_works list instead of
5668	* pool->worklist.
5669	*
5670	* CONTEXT:
5671	* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5672	*/
5673	void freeze_workqueues_begin(void)
5674	{
5675	struct workqueue_struct *wq;
5676	struct pool_workqueue *pwq;
5677
5678	mutex_lock(&wq_pool_mutex);
5679
5680	WARN_ON_ONCE(workqueue_freezing);
5681	workqueue_freezing = true;
5682
5683	list_for_each_entry(wq, &workqueues, list) {
5684	mutex_lock(&wq->mutex);
5685	for_each_pwq(pwq, wq)
5686	pwq_adjust_max_active(pwq);
5687	mutex_unlock(&wq->mutex);
5688	}
5689
5690	mutex_unlock(&wq_pool_mutex);
5691	}
5692
5693	/**
5694	* freeze_workqueues_busy - are freezable workqueues still busy?
5695	*
5696	* Check whether freezing is complete. This function must be called
5697	* between freeze_workqueues_begin() and thaw_workqueues().
5698	*
5699	* CONTEXT:
5700	* Grabs and releases wq_pool_mutex.
5701	*
5702	* Return:
5703	* %true if some freezable workqueues are still busy. %false if freezing
5704	* is complete.
5705	*/
5706	bool freeze_workqueues_busy(void)
5707	{
5708	bool busy = false;
5709	struct workqueue_struct *wq;
5710	struct pool_workqueue *pwq;
5711
5712	mutex_lock(&wq_pool_mutex);
5713
5714	WARN_ON_ONCE(!workqueue_freezing);
5715
5716	list_for_each_entry(wq, &workqueues, list) {
5717	if (!(wq->flags & WQ_FREEZABLE))
5718	continue;
5719	/*
5720	* nr_active is monotonically decreasing. It's safe
5721	* to peek without lock.
5722	*/
5723	rcu_read_lock();
5724	for_each_pwq(pwq, wq) {
5725	WARN_ON_ONCE(pwq->nr_active < 0);
5726	if (pwq->nr_active) {
5727	busy = true;
5728	rcu_read_unlock();
5729	goto out_unlock;
5730	}
5731	}
5732	rcu_read_unlock();
5733	}
5734	out_unlock:
5735	mutex_unlock(&wq_pool_mutex);
5736	return busy;
5737	}
5738
5739	/**
5740	* thaw_workqueues - thaw workqueues
5741	*
5742	* Thaw workqueues. Normal queueing is restored and all collected
5743	* frozen works are transferred to their respective pool worklists.
5744	*
5745	* CONTEXT:
5746	* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5747	*/
5748	void thaw_workqueues(void)
5749	{
5750	struct workqueue_struct *wq;
5751	struct pool_workqueue *pwq;
5752
5753	mutex_lock(&wq_pool_mutex);
5754
5755	if (!workqueue_freezing)
5756	goto out_unlock;
5757
5758	workqueue_freezing = false;
5759
5760	/* restore max_active and repopulate worklist */
5761	list_for_each_entry(wq, &workqueues, list) {
5762	mutex_lock(&wq->mutex);
5763	for_each_pwq(pwq, wq)
5764	pwq_adjust_max_active(pwq);
5765	mutex_unlock(&wq->mutex);
5766	}
5767
5768	out_unlock:
5769	mutex_unlock(&wq_pool_mutex);
5770	}
5771	#endif /* CONFIG_FREEZER */
5772
5773	static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
5774	{
5775	LIST_HEAD(ctxs);
5776	int ret = 0;
5777	struct workqueue_struct *wq;
5778	struct apply_wqattrs_ctx ctx, n;
5779
5780	lockdep_assert_held(&wq_pool_mutex);
5781
5782	list_for_each_entry(wq, &workqueues, list) {
5783	if (!(wq->flags & WQ_UNBOUND))
5784	continue;
5785	/* creating multiple pwqs breaks ordering guarantee */
5786	if (wq->flags & __WQ_ORDERED)
5787	continue;
5788
5789	ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
5790	if (IS_ERR(ctx)) {
5791	ret = PTR_ERR(ctx);
5792	break;
5793	}
5794
5795	list_add_tail(&ctx->list, &ctxs);
5796	}
5797
5798	list_for_each_entry_safe(ctx, n, &ctxs, list) {
5799	if (!ret)
5800	apply_wqattrs_commit(ctx);
5801	apply_wqattrs_cleanup(ctx);
5802	}
5803
5804	if (!ret) {
5805	mutex_lock(&wq_pool_attach_mutex);
5806	cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
5807	mutex_unlock(&wq_pool_attach_mutex);
5808	}
5809	return ret;
5810	}
5811
5812	/**
5813	* workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5814	* @cpumask: the cpumask to set
5815	*
5816	* The low-level workqueues cpumask is a global cpumask that limits
5817	* the affinity of all unbound workqueues. This function check the @cpumask
5818	* and apply it to all unbound workqueues and updates all pwqs of them.
5819	*
5820	* Return: 0 - Success
5821	* -EINVAL - Invalid @cpumask
5822	* -ENOMEM - Failed to allocate memory for attrs or pwqs.
5823	*/
5824	int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5825	{
5826	int ret = -EINVAL;
5827
5828	/*
5829	* Not excluding isolated cpus on purpose.
5830	* If the user wishes to include them, we allow that.
5831	*/
5832	cpumask_and(cpumask, cpumask, cpu_possible_mask);
5833	if (!cpumask_empty(cpumask)) {
5834	apply_wqattrs_lock();
5835	if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
5836	ret = 0;
5837	goto out_unlock;
5838	}
5839
5840	ret = workqueue_apply_unbound_cpumask(cpumask);
5841
5842	out_unlock:
5843	apply_wqattrs_unlock();
5844	}
5845
5846	return ret;
5847	}
5848
5849	static int parse_affn_scope(const char *val)
5850	{
5851	int i;
5852
5853	for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
5854	if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
5855	return i;
5856	}
5857	return -EINVAL;
5858	}
5859
5860	static int wq_affn_dfl_set(const char val, const struct kernel_param kp)
5861	{
5862	struct workqueue_struct *wq;
5863	int affn, cpu;
5864
5865	affn = parse_affn_scope(val);
5866	if (affn < 0)
5867	return affn;
5868	if (affn == WQ_AFFN_DFL)
5869	return -EINVAL;
5870
5871	cpus_read_lock();
5872	mutex_lock(&wq_pool_mutex);
5873
5874	wq_affn_dfl = affn;
5875
5876	list_for_each_entry(wq, &workqueues, list) {
5877	for_each_online_cpu(cpu) {
5878	wq_update_pod(wq, cpu, cpu, true);
5879	}
5880	}
5881
5882	mutex_unlock(&wq_pool_mutex);
5883	cpus_read_unlock();
5884
5885	return 0;
5886	}
5887
5888	static int wq_affn_dfl_get(char buffer, const struct kernel_param kp)
5889	{
5890	return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
5891	}
5892
5893	static const struct kernel_param_ops wq_affn_dfl_ops = {
5894	.set = wq_affn_dfl_set,
5895	.get = wq_affn_dfl_get,
5896	};
5897
5898	module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
5899
5900	#ifdef CONFIG_SYSFS
5901	/*
5902	* Workqueues with WQ_SYSFS flag set is visible to userland via
5903	* /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5904	* following attributes.
5905	*
5906	* per_cpu RO bool : whether the workqueue is per-cpu or unbound
5907	* max_active RW int : maximum number of in-flight work items
5908	*
5909	* Unbound workqueues have the following extra attributes.
5910	*
5911	* nice RW int : nice value of the workers
5912	* cpumask RW mask : bitmask of allowed CPUs for the workers
5913	* affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
5914	* affinity_strict RW bool : worker CPU affinity is strict
5915	*/
5916	struct wq_device {
5917	struct workqueue_struct *wq;
5918	struct device dev;
5919	};
5920
5921	static struct workqueue_struct dev_to_wq(struct device dev)
5922	{
5923	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5924
5925	return wq_dev->wq;
5926	}
5927
5928	static ssize_t per_cpu_show(struct device dev, struct device_attribute attr,
5929	char *buf)
5930	{
5931	struct workqueue_struct *wq = dev_to_wq(dev);
5932
5933	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5934	}
5935	static DEVICE_ATTR_RO(per_cpu);
5936
5937	static ssize_t max_active_show(struct device *dev,
5938	struct device_attribute attr, char buf)
5939	{
5940	struct workqueue_struct *wq = dev_to_wq(dev);
5941
5942	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5943	}
5944
5945	static ssize_t max_active_store(struct device *dev,
5946	struct device_attribute attr, const char buf,
5947	size_t count)
5948	{
5949	struct workqueue_struct *wq = dev_to_wq(dev);
5950	int val;
5951
5952	if (sscanf(buf, "%d", &val) != 1 \|\| val <= 0)
5953	return -EINVAL;
5954
5955	workqueue_set_max_active(wq, val);
5956	return count;
5957	}
5958	static DEVICE_ATTR_RW(max_active);
5959
5960	static struct attribute *wq_sysfs_attrs[] = {
5961	&dev_attr_per_cpu.attr,
5962	&dev_attr_max_active.attr,
5963	NULL,
5964	};
5965	ATTRIBUTE_GROUPS(wq_sysfs);
5966
5967	static ssize_t wq_nice_show(struct device dev, struct device_attribute attr,
5968	char *buf)
5969	{
5970	struct workqueue_struct *wq = dev_to_wq(dev);
5971	int written;
5972
5973	mutex_lock(&wq->mutex);
5974	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5975	mutex_unlock(&wq->mutex);
5976
5977	return written;
5978	}
5979
5980	/* prepare workqueue_attrs for sysfs store operations */
5981	static struct workqueue_attrs wq_sysfs_prep_attrs(struct workqueue_struct wq)
5982	{
5983	struct workqueue_attrs *attrs;
5984
5985	lockdep_assert_held(&wq_pool_mutex);
5986
5987	attrs = alloc_workqueue_attrs();
5988	if (!attrs)
5989	return NULL;
5990
5991	copy_workqueue_attrs(attrs, wq->unbound_attrs);
5992	return attrs;
5993	}
5994
5995	static ssize_t wq_nice_store(struct device dev, struct device_attribute attr,
5996	const char *buf, size_t count)
5997	{
5998	struct workqueue_struct *wq = dev_to_wq(dev);
5999	struct workqueue_attrs *attrs;
6000	int ret = -ENOMEM;
6001
6002	apply_wqattrs_lock();
6003
6004	attrs = wq_sysfs_prep_attrs(wq);
6005	if (!attrs)
6006	goto out_unlock;
6007
6008	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
6009	attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
6010	ret = apply_workqueue_attrs_locked(wq, attrs);
6011	else
6012	ret = -EINVAL;
6013
6014	out_unlock:
6015	apply_wqattrs_unlock();
6016	free_workqueue_attrs(attrs);
6017	return ret ?: count;
6018	}
6019
6020	static ssize_t wq_cpumask_show(struct device *dev,
6021	struct device_attribute attr, char buf)
6022	{
6023	struct workqueue_struct *wq = dev_to_wq(dev);
6024	int written;
6025
6026	mutex_lock(&wq->mutex);
6027	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6028	cpumask_pr_args(wq->unbound_attrs->cpumask));
6029	mutex_unlock(&wq->mutex);
6030	return written;
6031	}
6032
6033	static ssize_t wq_cpumask_store(struct device *dev,
6034	struct device_attribute *attr,
6035	const char *buf, size_t count)
6036	{
6037	struct workqueue_struct *wq = dev_to_wq(dev);
6038	struct workqueue_attrs *attrs;
6039	int ret = -ENOMEM;
6040
6041	apply_wqattrs_lock();
6042
6043	attrs = wq_sysfs_prep_attrs(wq);
6044	if (!attrs)
6045	goto out_unlock;
6046
6047	ret = cpumask_parse(buf, attrs->cpumask);
6048	if (!ret)
6049	ret = apply_workqueue_attrs_locked(wq, attrs);
6050
6051	out_unlock:
6052	apply_wqattrs_unlock();
6053	free_workqueue_attrs(attrs);
6054	return ret ?: count;
6055	}
6056
6057	static ssize_t wq_affn_scope_show(struct device *dev,
6058	struct device_attribute attr, char buf)
6059	{
6060	struct workqueue_struct *wq = dev_to_wq(dev);
6061	int written;
6062
6063	mutex_lock(&wq->mutex);
6064	if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
6065	written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
6066	wq_affn_names[WQ_AFFN_DFL],
6067	wq_affn_names[wq_affn_dfl]);
6068	else
6069	written = scnprintf(buf, PAGE_SIZE, "%s\n",
6070	wq_affn_names[wq->unbound_attrs->affn_scope]);
6071	mutex_unlock(&wq->mutex);
6072
6073	return written;
6074	}
6075
6076	static ssize_t wq_affn_scope_store(struct device *dev,
6077	struct device_attribute *attr,
6078	const char *buf, size_t count)
6079	{
6080	struct workqueue_struct *wq = dev_to_wq(dev);
6081	struct workqueue_attrs *attrs;
6082	int affn, ret = -ENOMEM;
6083
6084	affn = parse_affn_scope(buf);
6085	if (affn < 0)
6086	return affn;
6087
6088	apply_wqattrs_lock();
6089	attrs = wq_sysfs_prep_attrs(wq);
6090	if (attrs) {
6091	attrs->affn_scope = affn;
6092	ret = apply_workqueue_attrs_locked(wq, attrs);
6093	}
6094	apply_wqattrs_unlock();
6095	free_workqueue_attrs(attrs);
6096	return ret ?: count;
6097	}
6098
6099	static ssize_t wq_affinity_strict_show(struct device *dev,
6100	struct device_attribute attr, char buf)
6101	{
6102	struct workqueue_struct *wq = dev_to_wq(dev);
6103
6104	return scnprintf(buf, PAGE_SIZE, "%d\n",
6105	wq->unbound_attrs->affn_strict);
6106	}
6107
6108	static ssize_t wq_affinity_strict_store(struct device *dev,
6109	struct device_attribute *attr,
6110	const char *buf, size_t count)
6111	{
6112	struct workqueue_struct *wq = dev_to_wq(dev);
6113	struct workqueue_attrs *attrs;
6114	int v, ret = -ENOMEM;
6115
6116	if (sscanf(buf, "%d", &v) != 1)
6117	return -EINVAL;
6118
6119	apply_wqattrs_lock();
6120	attrs = wq_sysfs_prep_attrs(wq);
6121	if (attrs) {
6122	attrs->affn_strict = (bool)v;
6123	ret = apply_workqueue_attrs_locked(wq, attrs);
6124	}
6125	apply_wqattrs_unlock();
6126	free_workqueue_attrs(attrs);
6127	return ret ?: count;
6128	}
6129
6130	static struct device_attribute wq_sysfs_unbound_attrs[] = {
6131	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
6132	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
6133	__ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
6134	__ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
6135	__ATTR_NULL,
6136	};
6137
6138	static struct bus_type wq_subsys = {
6139	.name = "workqueue",
6140	.dev_groups = wq_sysfs_groups,
6141	};
6142
6143	static ssize_t wq_unbound_cpumask_show(struct device *dev,
6144	struct device_attribute attr, char buf)
6145	{
6146	int written;
6147
6148	mutex_lock(&wq_pool_mutex);
6149	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6150	cpumask_pr_args(wq_unbound_cpumask));
6151	mutex_unlock(&wq_pool_mutex);
6152
6153	return written;
6154	}
6155
6156	static ssize_t wq_unbound_cpumask_store(struct device *dev,
6157	struct device_attribute attr, const char buf, size_t count)
6158	{
6159	cpumask_var_t cpumask;
6160	int ret;
6161
6162	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6163	return -ENOMEM;
6164
6165	ret = cpumask_parse(buf, cpumask);
6166	if (!ret)
6167	ret = workqueue_set_unbound_cpumask(cpumask);
6168
6169	free_cpumask_var(cpumask);
6170	return ret ? ret : count;
6171	}
6172
6173	static struct device_attribute wq_sysfs_cpumask_attr =
6174	__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
6175	wq_unbound_cpumask_store);
6176
6177	static int __init wq_sysfs_init(void)
6178	{
6179	struct device *dev_root;
6180	int err;
6181
6182	err = subsys_virtual_register(&wq_subsys, NULL);
6183	if (err)
6184	return err;
6185
6186	dev_root = bus_get_dev_root(&wq_subsys);
6187	if (dev_root) {
6188	err = device_create_file(dev_root, &wq_sysfs_cpumask_attr);
6189	put_device(dev_root);
6190	}
6191	return err;
6192	}
6193	core_initcall(wq_sysfs_init);
6194
6195	static void wq_device_release(struct device *dev)
6196	{
6197	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6198
6199	kfree(wq_dev);
6200	}
6201
6202	/**
6203	* workqueue_sysfs_register - make a workqueue visible in sysfs
6204	* @wq: the workqueue to register
6205	*
6206	* Expose @wq in sysfs under /sys/bus/workqueue/devices.
6207	* alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
6208	* which is the preferred method.
6209	*
6210	* Workqueue user should use this function directly iff it wants to apply
6211	* workqueue_attrs before making the workqueue visible in sysfs; otherwise,
6212	* apply_workqueue_attrs() may race against userland updating the
6213	* attributes.
6214	*
6215	* Return: 0 on success, -errno on failure.
6216	*/
6217	int workqueue_sysfs_register(struct workqueue_struct *wq)
6218	{
6219	struct wq_device *wq_dev;
6220	int ret;
6221
6222	/*
6223	* Adjusting max_active or creating new pwqs by applying
6224	* attributes breaks ordering guarantee. Disallow exposing ordered
6225	* workqueues.
6226	*/
6227	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
6228	return -EINVAL;
6229
6230	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
6231	if (!wq_dev)
6232	return -ENOMEM;
6233
6234	wq_dev->wq = wq;
6235	wq_dev->dev.bus = &wq_subsys;
6236	wq_dev->dev.release = wq_device_release;
6237	dev_set_name(&wq_dev->dev, "%s", wq->name);
6238
6239	/*
6240	* unbound_attrs are created separately. Suppress uevent until
6241	* everything is ready.
6242	*/
6243	dev_set_uevent_suppress(&wq_dev->dev, true);
6244
6245	ret = device_register(&wq_dev->dev);
6246	if (ret) {
6247	put_device(&wq_dev->dev);
6248	wq->wq_dev = NULL;
6249	return ret;
6250	}
6251
6252	if (wq->flags & WQ_UNBOUND) {
6253	struct device_attribute *attr;
6254
6255	for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
6256	ret = device_create_file(&wq_dev->dev, attr);
6257	if (ret) {
6258	device_unregister(&wq_dev->dev);
6259	wq->wq_dev = NULL;
6260	return ret;
6261	}
6262	}
6263	}
6264
6265	dev_set_uevent_suppress(&wq_dev->dev, false);
6266	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
6267	return 0;
6268	}
6269
6270	/**
6271	* workqueue_sysfs_unregister - undo workqueue_sysfs_register()
6272	* @wq: the workqueue to unregister
6273	*
6274	* If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
6275	*/
6276	static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
6277	{
6278	struct wq_device *wq_dev = wq->wq_dev;
6279
6280	if (!wq->wq_dev)
6281	return;
6282
6283	wq->wq_dev = NULL;
6284	device_unregister(&wq_dev->dev);
6285	}
6286	#else /* CONFIG_SYSFS */
6287	static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
6288	#endif /* CONFIG_SYSFS */
6289
6290	/*
6291	* Workqueue watchdog.
6292	*
6293	* Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
6294	* flush dependency, a concurrency managed work item which stays RUNNING
6295	* indefinitely. Workqueue stalls can be very difficult to debug as the
6296	* usual warning mechanisms don't trigger and internal workqueue state is
6297	* largely opaque.
6298	*
6299	* Workqueue watchdog monitors all worker pools periodically and dumps
6300	* state if some pools failed to make forward progress for a while where
6301	* forward progress is defined as the first item on ->worklist changing.
6302	*
6303	* This mechanism is controlled through the kernel parameter
6304	* "workqueue.watchdog_thresh" which can be updated at runtime through the
6305	* corresponding sysfs parameter file.
6306	*/
6307	#ifdef CONFIG_WQ_WATCHDOG
6308
6309	static unsigned long wq_watchdog_thresh = 30;
6310	static struct timer_list wq_watchdog_timer;
6311
6312	static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
6313	static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
6314
6315	/*
6316	* Show workers that might prevent the processing of pending work items.
6317	* The only candidates are CPU-bound workers in the running state.
6318	* Pending work items should be handled by another idle worker
6319	* in all other situations.
6320	*/
6321	static void show_cpu_pool_hog(struct worker_pool *pool)
6322	{
6323	struct worker *worker;
6324	unsigned long flags;
6325	int bkt;
6326
6327	raw_spin_lock_irqsave(&pool->lock, flags);
6328
6329	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6330	if (task_is_running(worker->task)) {
6331	/*
6332	* Defer printing to avoid deadlocks in console
6333	* drivers that queue work while holding locks
6334	* also taken in their write paths.
6335	*/
6336	printk_deferred_enter();
6337
6338	pr_info("pool %d:\n", pool->id);
6339	sched_show_task(worker->task);
6340
6341	printk_deferred_exit();
6342	}
6343	}
6344
6345	raw_spin_unlock_irqrestore(&pool->lock, flags);
6346	}
6347
6348	static void show_cpu_pools_hogs(void)
6349	{
6350	struct worker_pool *pool;
6351	int pi;
6352
6353	pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
6354
6355	rcu_read_lock();
6356
6357	for_each_pool(pool, pi) {
6358	if (pool->cpu_stall)
6359	show_cpu_pool_hog(pool);
6360
6361	}
6362
6363	rcu_read_unlock();
6364	}
6365
6366	static void wq_watchdog_reset_touched(void)
6367	{
6368	int cpu;
6369
6370	wq_watchdog_touched = jiffies;
6371	for_each_possible_cpu(cpu)
6372	per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
6373	}
6374
6375	static void wq_watchdog_timer_fn(struct timer_list *unused)
6376	{
6377	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
6378	bool lockup_detected = false;
6379	bool cpu_pool_stall = false;
6380	unsigned long now = jiffies;
6381	struct worker_pool *pool;
6382	int pi;
6383
6384	if (!thresh)
6385	return;
6386
6387	rcu_read_lock();
6388
6389	for_each_pool(pool, pi) {
6390	unsigned long pool_ts, touched, ts;
6391
6392	pool->cpu_stall = false;
6393	if (list_empty(&pool->worklist))
6394	continue;
6395
6396	/*
6397	* If a virtual machine is stopped by the host it can look to
6398	* the watchdog like a stall.
6399	*/
6400	kvm_check_and_clear_guest_paused();
6401
6402	/* get the latest of pool and touched timestamps */
6403	if (pool->cpu >= 0)
6404	touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
6405	else
6406	touched = READ_ONCE(wq_watchdog_touched);
6407	pool_ts = READ_ONCE(pool->watchdog_ts);
6408
6409	if (time_after(pool_ts, touched))
6410	ts = pool_ts;
6411	else
6412	ts = touched;
6413
6414	/* did we stall? */
6415	if (time_after(now, ts + thresh)) {
6416	lockup_detected = true;
6417	if (pool->cpu >= 0) {
6418	pool->cpu_stall = true;
6419	cpu_pool_stall = true;
6420	}
6421	pr_emerg("BUG: workqueue lockup - pool");
6422	pr_cont_pool_info(pool);
6423	pr_cont(" stuck for %us!\n",
6424	jiffies_to_msecs(now - pool_ts) / 1000);
6425	}
6426
6427
6428	}
6429
6430	rcu_read_unlock();
6431
6432	if (lockup_detected)
6433	show_all_workqueues();
6434
6435	if (cpu_pool_stall)
6436	show_cpu_pools_hogs();
6437
6438	wq_watchdog_reset_touched();
6439	mod_timer(&wq_watchdog_timer, jiffies + thresh);
6440	}
6441
6442	notrace void wq_watchdog_touch(int cpu)
6443	{
6444	if (cpu >= 0)
6445	per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
6446
6447	wq_watchdog_touched = jiffies;
6448	}
6449
6450	static void wq_watchdog_set_thresh(unsigned long thresh)
6451	{
6452	wq_watchdog_thresh = 0;
6453	del_timer_sync(&wq_watchdog_timer);
6454
6455	if (thresh) {
6456	wq_watchdog_thresh = thresh;
6457	wq_watchdog_reset_touched();
6458	mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
6459	}
6460	}
6461
6462	static int wq_watchdog_param_set_thresh(const char *val,
6463	const struct kernel_param *kp)
6464	{
6465	unsigned long thresh;
6466	int ret;
6467
6468	ret = kstrtoul(val, 0, &thresh);
6469	if (ret)
6470	return ret;
6471
6472	if (system_wq)
6473	wq_watchdog_set_thresh(thresh);
6474	else
6475	wq_watchdog_thresh = thresh;
6476
6477	return 0;
6478	}
6479
6480	static const struct kernel_param_ops wq_watchdog_thresh_ops = {
6481	.set = wq_watchdog_param_set_thresh,
6482	.get = param_get_ulong,
6483	};
6484
6485	module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
6486	0644);
6487
6488	static void wq_watchdog_init(void)
6489	{
6490	timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
6491	wq_watchdog_set_thresh(wq_watchdog_thresh);
6492	}
6493
6494	#else /* CONFIG_WQ_WATCHDOG */
6495
6496	static inline void wq_watchdog_init(void) { }
6497
6498	#endif /* CONFIG_WQ_WATCHDOG */
6499
6500	/**
6501	* workqueue_init_early - early init for workqueue subsystem
6502	*
6503	* This is the first step of three-staged workqueue subsystem initialization and
6504	* invoked as soon as the bare basics - memory allocation, cpumasks and idr are
6505	* up. It sets up all the data structures and system workqueues and allows early
6506	* boot code to create workqueues and queue/cancel work items. Actual work item
6507	* execution starts only after kthreads can be created and scheduled right
6508	* before early initcalls.
6509	*/
6510	void __init workqueue_init_early(void)
6511	{
6512	struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
6513	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
6514	int i, cpu;
6515
6516	BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
6517
6518	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
6519	cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ));
6520	cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN));
6521
6522	if (!cpumask_empty(&wq_cmdline_cpumask))
6523	cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, &wq_cmdline_cpumask);
6524
6525	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
6526
6527	wq_update_pod_attrs_buf = alloc_workqueue_attrs();
6528	BUG_ON(!wq_update_pod_attrs_buf);
6529
6530	/* initialize WQ_AFFN_SYSTEM pods */
6531	pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
6532	pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
6533	pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
6534	BUG_ON(!pt->pod_cpus \|\| !pt->pod_node \|\| !pt->cpu_pod);
6535
6536	BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
6537
6538	wq_update_pod_attrs_buf = alloc_workqueue_attrs();
6539	BUG_ON(!wq_update_pod_attrs_buf);
6540
6541	pt->nr_pods = 1;
6542	cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
6543	pt->pod_node[0] = NUMA_NO_NODE;
6544	pt->cpu_pod[0] = 0;
6545
6546	/* initialize CPU pools */
6547	for_each_possible_cpu(cpu) {
6548	struct worker_pool *pool;
6549
6550	i = 0;
6551	for_each_cpu_worker_pool(pool, cpu) {
6552	BUG_ON(init_worker_pool(pool));
6553	pool->cpu = cpu;
6554	cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
6555	cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
6556	pool->attrs->nice = std_nice[i++];
6557	pool->attrs->affn_strict = true;
6558	pool->node = cpu_to_node(cpu);
6559
6560	/* alloc pool ID */
6561	mutex_lock(&wq_pool_mutex);
6562	BUG_ON(worker_pool_assign_id(pool));
6563	mutex_unlock(&wq_pool_mutex);
6564	}
6565	}
6566
6567	/* create default unbound and ordered wq attrs */
6568	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
6569	struct workqueue_attrs *attrs;
6570
6571	BUG_ON(!(attrs = alloc_workqueue_attrs()));
6572	attrs->nice = std_nice[i];
6573	unbound_std_wq_attrs[i] = attrs;
6574
6575	/*
6576	* An ordered wq should have only one pwq as ordering is
6577	* guaranteed by max_active which is enforced by pwqs.
6578	*/
6579	BUG_ON(!(attrs = alloc_workqueue_attrs()));
6580	attrs->nice = std_nice[i];
6581	attrs->ordered = true;
6582	ordered_wq_attrs[i] = attrs;
6583	}
6584
6585	system_wq = alloc_workqueue("events", 0, 0);
6586	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
6587	system_long_wq = alloc_workqueue("events_long", 0, 0);
6588	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
6589	WQ_MAX_ACTIVE);
6590	system_freezable_wq = alloc_workqueue("events_freezable",
6591	WQ_FREEZABLE, 0);
6592	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
6593	WQ_POWER_EFFICIENT, 0);
6594	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
6595	WQ_FREEZABLE \| WQ_POWER_EFFICIENT,
6596	0);
6597	BUG_ON(!system_wq \|\| !system_highpri_wq \|\| !system_long_wq \|\|
6598	!system_unbound_wq \|\| !system_freezable_wq \|\|
6599	!system_power_efficient_wq \|\|
6600	!system_freezable_power_efficient_wq);
6601	}
6602
6603	static void __init wq_cpu_intensive_thresh_init(void)
6604	{
6605	unsigned long thresh;
6606	unsigned long bogo;
6607
6608	/* if the user set it to a specific value, keep it */
6609	if (wq_cpu_intensive_thresh_us != ULONG_MAX)
6610	return;
6611
6612	pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
6613	BUG_ON(IS_ERR(pwq_release_worker));
6614
6615	/*
6616	* The default of 10ms is derived from the fact that most modern (as of
6617	* 2023) processors can do a lot in 10ms and that it's just below what
6618	* most consider human-perceivable. However, the kernel also runs on a
6619	* lot slower CPUs including microcontrollers where the threshold is way
6620	* too low.
6621	*
6622	* Let's scale up the threshold upto 1 second if BogoMips is below 4000.
6623	* This is by no means accurate but it doesn't have to be. The mechanism
6624	* is still useful even when the threshold is fully scaled up. Also, as
6625	* the reports would usually be applicable to everyone, some machines
6626	* operating on longer thresholds won't significantly diminish their
6627	* usefulness.
6628	*/
6629	thresh = 10 * USEC_PER_MSEC;
6630
6631	/* see init/calibrate.c for lpj -> BogoMIPS calculation */
6632	bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
6633	if (bogo < 4000)
6634	thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
6635
6636	pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
6637	loops_per_jiffy, bogo, thresh);
6638
6639	wq_cpu_intensive_thresh_us = thresh;
6640	}
6641
6642	/**
6643	* workqueue_init - bring workqueue subsystem fully online
6644	*
6645	* This is the second step of three-staged workqueue subsystem initialization
6646	* and invoked as soon as kthreads can be created and scheduled. Workqueues have
6647	* been created and work items queued on them, but there are no kworkers
6648	* executing the work items yet. Populate the worker pools with the initial
6649	* workers and enable future kworker creations.
6650	*/
6651	void __init workqueue_init(void)
6652	{
6653	struct workqueue_struct *wq;
6654	struct worker_pool *pool;
6655	int cpu, bkt;
6656
6657	wq_cpu_intensive_thresh_init();
6658
6659	mutex_lock(&wq_pool_mutex);
6660
6661	/*
6662	* Per-cpu pools created earlier could be missing node hint. Fix them
6663	* up. Also, create a rescuer for workqueues that requested it.
6664	*/
6665	for_each_possible_cpu(cpu) {
6666	for_each_cpu_worker_pool(pool, cpu) {
6667	pool->node = cpu_to_node(cpu);
6668	}
6669	}
6670
6671	list_for_each_entry(wq, &workqueues, list) {
6672	WARN(init_rescuer(wq),
6673	"workqueue: failed to create early rescuer for %s",
6674	wq->name);
6675	}
6676
6677	mutex_unlock(&wq_pool_mutex);
6678
6679	/* create the initial workers */
6680	for_each_online_cpu(cpu) {
6681	for_each_cpu_worker_pool(pool, cpu) {
6682	pool->flags &= ~POOL_DISASSOCIATED;
6683	BUG_ON(!create_worker(pool));
6684	}
6685	}
6686
6687	hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6688	BUG_ON(!create_worker(pool));
6689
6690	wq_online = true;
6691	wq_watchdog_init();
6692	}
6693
6694	/*
6695	* Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
6696	* @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
6697	* and consecutive pod ID. The rest of @pt is initialized accordingly.
6698	*/
6699	static void __init init_pod_type(struct wq_pod_type *pt,
6700	bool (*cpus_share_pod)(int, int))
6701	{
6702	int cur, pre, cpu, pod;
6703
6704	pt->nr_pods = 0;
6705
6706	/* init @pt->cpu_pod[] according to @cpus_share_pod() */
6707	pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
6708	BUG_ON(!pt->cpu_pod);
6709
6710	for_each_possible_cpu(cur) {
6711	for_each_possible_cpu(pre) {
6712	if (pre >= cur) {
6713	pt->cpu_pod[cur] = pt->nr_pods++;
6714	break;
6715	}
6716	if (cpus_share_pod(cur, pre)) {
6717	pt->cpu_pod[cur] = pt->cpu_pod[pre];
6718	break;
6719	}
6720	}
6721	}
6722
6723	/* init the rest to match @pt->cpu_pod[] */
6724	pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
6725	pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
6726	BUG_ON(!pt->pod_cpus \|\| !pt->pod_node);
6727
6728	for (pod = 0; pod < pt->nr_pods; pod++)
6729	BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
6730
6731	for_each_possible_cpu(cpu) {
6732	cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
6733	pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
6734	}
6735	}
6736
6737	static bool __init cpus_dont_share(int cpu0, int cpu1)
6738	{
6739	return false;
6740	}
6741
6742	static bool __init cpus_share_smt(int cpu0, int cpu1)
6743	{
6744	#ifdef CONFIG_SCHED_SMT
6745	return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
6746	#else
6747	return false;
6748	#endif
6749	}
6750
6751	static bool __init cpus_share_numa(int cpu0, int cpu1)
6752	{
6753	return cpu_to_node(cpu0) == cpu_to_node(cpu1);
6754	}
6755
6756	/**
6757	* workqueue_init_topology - initialize CPU pods for unbound workqueues
6758	*
6759	* This is the third step of there-staged workqueue subsystem initialization and
6760	* invoked after SMP and topology information are fully initialized. It
6761	* initializes the unbound CPU pods accordingly.
6762	*/
6763	void __init workqueue_init_topology(void)
6764	{
6765	struct workqueue_struct *wq;
6766	int cpu;
6767
6768	init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
6769	init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
6770	init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
6771	init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
6772
6773	mutex_lock(&wq_pool_mutex);
6774
6775	/*
6776	* Workqueues allocated earlier would have all CPUs sharing the default
6777	* worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
6778	* combinations to apply per-pod sharing.
6779	*/
6780	list_for_each_entry(wq, &workqueues, list) {
6781	for_each_online_cpu(cpu) {
6782	wq_update_pod(wq, cpu, cpu, true);
6783	}
6784	}
6785
6786	mutex_unlock(&wq_pool_mutex);
6787	}
6788
6789	void __warn_flushing_systemwide_wq(void)
6790	{
6791	pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
6792	dump_stack();
6793	}
6794	EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
6795
6796	static int __init workqueue_unbound_cpus_setup(char *str)
6797	{
6798	if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
6799	cpumask_clear(&wq_cmdline_cpumask);
6800	pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
6801	}
6802
6803	return 1;
6804	}
6805	__setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);