linux - kernel/workqueue.c

5618 lines · 154.4 KiB

1	/*
2	* kernel/workqueue.c - generic async execution with shared worker pool
3	*
4	* Copyright (C) 2002 Ingo Molnar
5	*
6	* Derived from the taskqueue/keventd code by:
7	* David Woodhouse <[email protected]>
8	* Andrew Morton
9	* Kai Petzke <[email protected]>
10	* Theodore Ts'o <[email protected]>
11	*
12	* Made to use alloc_percpu by Christoph Lameter.
13	*
14	* Copyright (C) 2010 SUSE Linux Products GmbH
15	* Copyright (C) 2010 Tejun Heo <[email protected]>
16	*
17	* This is the generic async execution mechanism. Work items as are
18	* executed in process context. The worker pool is shared and
19	* automatically managed. There are two worker pools for each CPU (one for
20	* normal work items and the other for high priority ones) and some extra
21	* pools for workqueues which are not bound to any specific CPU - the
22	* number of these backing pools is dynamic.
23	*
24	* Please read Documentation/workqueue.txt for details.
25	*/
26
27	#include <linux/export.h>
28	#include <linux/kernel.h>
29	#include <linux/sched.h>
30	#include <linux/init.h>
31	#include <linux/signal.h>
32	#include <linux/completion.h>
33	#include <linux/workqueue.h>
34	#include <linux/slab.h>
35	#include <linux/cpu.h>
36	#include <linux/notifier.h>
37	#include <linux/kthread.h>
38	#include <linux/hardirq.h>
39	#include <linux/mempolicy.h>
40	#include <linux/freezer.h>
41	#include <linux/kallsyms.h>
42	#include <linux/debug_locks.h>
43	#include <linux/lockdep.h>
44	#include <linux/idr.h>
45	#include <linux/jhash.h>
46	#include <linux/hashtable.h>
47	#include <linux/rculist.h>
48	#include <linux/nodemask.h>
49	#include <linux/moduleparam.h>
50	#include <linux/uaccess.h>
51
52	#include "workqueue_internal.h"
53
54	enum {
55	/*
56	* worker_pool flags
57	*
58	* A bound pool is either associated or disassociated with its CPU.
59	* While associated (!DISASSOCIATED), all workers are bound to the
60	* CPU and none has %WORKER_UNBOUND set and concurrency management
61	* is in effect.
62	*
63	* While DISASSOCIATED, the cpu may be offline and all workers have
64	* %WORKER_UNBOUND set and concurrency management disabled, and may
65	* be executing on any CPU. The pool behaves as an unbound one.
66	*
67	* Note that DISASSOCIATED should be flipped only while holding
68	* attach_mutex to avoid changing binding state while
69	* worker_attach_to_pool() is in progress.
70	*/
71	POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
72
73	/* worker flags */
74	WORKER_DIE = 1 << 1, /* die die die */
75	WORKER_IDLE = 1 << 2, /* is idle */
76	WORKER_PREP = 1 << 3, /* preparing to run works */
77	WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78	WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79	WORKER_REBOUND = 1 << 8, /* worker was rebound */
80
81	WORKER_NOT_RUNNING = WORKER_PREP \| WORKER_CPU_INTENSIVE \|
82	WORKER_UNBOUND \| WORKER_REBOUND,
83
84	NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
85
86	UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87	BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
88
89	MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90	IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91
92	MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93	/* call for help after 10ms
94	(min two ticks) */
95	MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96	CREATE_COOLDOWN = HZ, /* time to breath after fail */
97
98	/*
99	* Rescue workers are used only on emergencies and shared by
100	* all cpus. Give MIN_NICE.
101	*/
102	RESCUER_NICE_LEVEL = MIN_NICE,
103	HIGHPRI_NICE_LEVEL = MIN_NICE,
104
105	WQ_NAME_LEN = 24,
106	};
107
108	/*
109	* Structure fields follow one of the following exclusion rules.
110	*
111	* I: Modifiable by initialization/destruction paths and read-only for
112	* everyone else.
113	*
114	* P: Preemption protected. Disabling preemption is enough and should
115	* only be modified and accessed from the local cpu.
116	*
117	* L: pool->lock protected. Access with pool->lock held.
118	*
119	* X: During normal operation, modification requires pool->lock and should
120	* be done only from local cpu. Either disabling preemption on local
121	* cpu or grabbing pool->lock is enough for read access. If
122	* POOL_DISASSOCIATED is set, it's identical to L.
123	*
124	* A: pool->attach_mutex protected.
125	*
126	* PL: wq_pool_mutex protected.
127	*
128	* PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
129	*
130	* PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
131	*
132	* PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133	* sched-RCU for reads.
134	*
135	* WQ: wq->mutex protected.
136	*
137	* WR: wq->mutex protected for writes. Sched-RCU protected for reads.
138	*
139	* MD: wq_mayday_lock protected.
140	*/
141
142	/* struct worker is defined in workqueue_internal.h */
143
144	struct worker_pool {
145	spinlock_t lock; /* the pool lock */
146	int cpu; /* I: the associated cpu */
147	int node; /* I: the associated node ID */
148	int id; /* I: pool ID */
149	unsigned int flags; /* X: flags */
150
151	unsigned long watchdog_ts; /* L: watchdog timestamp */
152
153	struct list_head worklist; /* L: list of pending works */
154	int nr_workers; /* L: total number of workers */
155
156	/* nr_idle includes the ones off idle_list for rebinding */
157	int nr_idle; /* L: currently idle ones */
158
159	struct list_head idle_list; /* X: list of idle workers */
160	struct timer_list idle_timer; /* L: worker idle timeout */
161	struct timer_list mayday_timer; /* L: SOS timer for workers */
162
163	/* a workers is either on busy_hash or idle_list, or the manager */
164	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
165	/* L: hash of busy workers */
166
167	/* see manage_workers() for details on the two manager mutexes */
168	struct mutex manager_arb; /* manager arbitration */
169	struct worker manager; / L: purely informational */
170	struct mutex attach_mutex; /* attach/detach exclusion */
171	struct list_head workers; /* A: attached workers */
172	struct completion detach_completion; / all workers detached */
173
174	struct ida worker_ida; /* worker IDs for task name */
175
176	struct workqueue_attrs attrs; / I: worker attributes */
177	struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178	int refcnt; /* PL: refcnt for unbound pools */
179
180	/*
181	* The current concurrency level. As it's likely to be accessed
182	* from other CPUs during try_to_wake_up(), put it in a separate
183	* cacheline.
184	*/
185	atomic_t nr_running ____cacheline_aligned_in_smp;
186
187	/*
188	* Destruction of pool is sched-RCU protected to allow dereferences
189	* from get_work_pool().
190	*/
191	struct rcu_head rcu;
192	} ____cacheline_aligned_in_smp;
193
194	/*
195	* The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196	* of work_struct->data are used for flags and the remaining high bits
197	* point to the pwq; thus, pwqs need to be aligned at two's power of the
198	* number of flag bits.
199	*/
200	struct pool_workqueue {
201	struct worker_pool pool; / I: the associated pool */
202	struct workqueue_struct wq; / I: the owning workqueue */
203	int work_color; /* L: current color */
204	int flush_color; /* L: flushing color */
205	int refcnt; /* L: reference count */
206	int nr_in_flight[WORK_NR_COLORS];
207	/* L: nr of in_flight works */
208	int nr_active; /* L: nr of active works */
209	int max_active; /* L: max active works */
210	struct list_head delayed_works; /* L: delayed works */
211	struct list_head pwqs_node; /* WR: node on wq->pwqs */
212	struct list_head mayday_node; /* MD: node on wq->maydays */
213
214	/*
215	* Release of unbound pwq is punted to system_wq. See put_pwq()
216	* and pwq_unbound_release_workfn() for details. pool_workqueue
217	* itself is also sched-RCU protected so that the first pwq can be
218	* determined without grabbing wq->mutex.
219	*/
220	struct work_struct unbound_release_work;
221	struct rcu_head rcu;
222	} __aligned(1 << WORK_STRUCT_FLAG_BITS);
223
224	/*
225	* Structure used to wait for workqueue flush.
226	*/
227	struct wq_flusher {
228	struct list_head list; /* WQ: list of flushers */
229	int flush_color; /* WQ: flush color waiting for */
230	struct completion done; /* flush completion */
231	};
232
233	struct wq_device;
234
235	/*
236	* The externally visible workqueue. It relays the issued work items to
237	* the appropriate worker_pool through its pool_workqueues.
238	*/
239	struct workqueue_struct {
240	struct list_head pwqs; /* WR: all pwqs of this wq */
241	struct list_head list; /* PR: list of all workqueues */
242
243	struct mutex mutex; /* protects this wq */
244	int work_color; /* WQ: current work color */
245	int flush_color; /* WQ: current flush color */
246	atomic_t nr_pwqs_to_flush; /* flush in progress */
247	struct wq_flusher first_flusher; / WQ: first flusher */
248	struct list_head flusher_queue; /* WQ: flush waiters */
249	struct list_head flusher_overflow; /* WQ: flush overflow list */
250
251	struct list_head maydays; /* MD: pwqs requesting rescue */
252	struct worker rescuer; / I: rescue worker */
253
254	int nr_drainers; /* WQ: drain in progress */
255	int saved_max_active; /* WQ: saved pwq max_active */
256
257	struct workqueue_attrs unbound_attrs; / PW: only for unbound wqs */
258	struct pool_workqueue dfl_pwq; / PW: only for unbound wqs */
259
260	#ifdef CONFIG_SYSFS
261	struct wq_device wq_dev; / I: for sysfs interface */
262	#endif
263	#ifdef CONFIG_LOCKDEP
264	struct lockdep_map lockdep_map;
265	#endif
266	char name[WQ_NAME_LEN]; /* I: workqueue name */
267
268	/*
269	* Destruction of workqueue_struct is sched-RCU protected to allow
270	* walking the workqueues list without grabbing wq_pool_mutex.
271	* This is used to dump all workqueues from sysrq.
272	*/
273	struct rcu_head rcu;
274
275	/* hot fields used during command issue, aligned to cacheline */
276	unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277	struct pool_workqueue __percpu cpu_pwqs; / I: per-cpu pwqs */
278	struct pool_workqueue __rcu numa_pwq_tbl[]; / PWR: unbound pwqs indexed by node */
279	};
280
281	static struct kmem_cache *pwq_cache;
282
283	static cpumask_var_t *wq_numa_possible_cpumask;
284	/* possible CPUs of each node */
285
286	static bool wq_disable_numa;
287	module_param_named(disable_numa, wq_disable_numa, bool, 0444);
288
289	/* see the comment above the definition of WQ_POWER_EFFICIENT */
290	static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291	module_param_named(power_efficient, wq_power_efficient, bool, 0444);
292
293	static bool wq_online; /* can kworkers be created yet? */
294
295	static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
296
297	/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
298	static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
299
300	static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
301	static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
302
303	static LIST_HEAD(workqueues); /* PR: list of all workqueues */
304	static bool workqueue_freezing; /* PL: have wqs started freezing? */
305
306	/* PL: allowable cpus for unbound wqs and work items */
307	static cpumask_var_t wq_unbound_cpumask;
308
309	/* CPU where unbound work was last round robin scheduled from this CPU */
310	static DEFINE_PER_CPU(int, wq_rr_cpu_last);
311
312	/*
313	* Local execution of unbound work items is no longer guaranteed. The
314	* following always forces round-robin CPU selection on unbound work items
315	* to uncover usages which depend on it.
316	*/
317	#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
318	static bool wq_debug_force_rr_cpu = true;
319	#else
320	static bool wq_debug_force_rr_cpu = false;
321	#endif
322	module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
323
324	/* the per-cpu worker pools */
325	static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
326
327	static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
328
329	/* PL: hash of all unbound pools keyed by pool->attrs */
330	static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
331
332	/* I: attributes used when instantiating standard unbound pools on demand */
333	static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
334
335	/* I: attributes used when instantiating ordered pools on demand */
336	static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
337
338	struct workqueue_struct *system_wq __read_mostly;
339	EXPORT_SYMBOL(system_wq);
340	struct workqueue_struct *system_highpri_wq __read_mostly;
341	EXPORT_SYMBOL_GPL(system_highpri_wq);
342	struct workqueue_struct *system_long_wq __read_mostly;
343	EXPORT_SYMBOL_GPL(system_long_wq);
344	struct workqueue_struct *system_unbound_wq __read_mostly;
345	EXPORT_SYMBOL_GPL(system_unbound_wq);
346	struct workqueue_struct *system_freezable_wq __read_mostly;
347	EXPORT_SYMBOL_GPL(system_freezable_wq);
348	struct workqueue_struct *system_power_efficient_wq __read_mostly;
349	EXPORT_SYMBOL_GPL(system_power_efficient_wq);
350	struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
351	EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
352
353	static int worker_thread(void *__worker);
354	static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
355
356	#define CREATE_TRACE_POINTS
357	#include <trace/events/workqueue.h>
358
359	#define assert_rcu_or_pool_mutex() \
360	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
361	!lockdep_is_held(&wq_pool_mutex), \
362	"sched RCU or wq_pool_mutex should be held")
363
364	#define assert_rcu_or_wq_mutex(wq) \
365	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
366	!lockdep_is_held(&wq->mutex), \
367	"sched RCU or wq->mutex should be held")
368
369	#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
370	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
371	!lockdep_is_held(&wq->mutex) && \
372	!lockdep_is_held(&wq_pool_mutex), \
373	"sched RCU, wq->mutex or wq_pool_mutex should be held")
374
375	#define for_each_cpu_worker_pool(pool, cpu) \
376	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
377	(pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
378	(pool)++)
379
380	/**
381	* for_each_pool - iterate through all worker_pools in the system
382	* @pool: iteration cursor
383	* @pi: integer used for iteration
384	*
385	* This must be called either with wq_pool_mutex held or sched RCU read
386	* locked. If the pool needs to be used beyond the locking in effect, the
387	* caller is responsible for guaranteeing that the pool stays online.
388	*
389	* The if/else clause exists only for the lockdep assertion and can be
390	* ignored.
391	*/
392	#define for_each_pool(pool, pi) \
393	idr_for_each_entry(&worker_pool_idr, pool, pi) \
394	if (({ assert_rcu_or_pool_mutex(); false; })) { } \
395	else
396
397	/**
398	* for_each_pool_worker - iterate through all workers of a worker_pool
399	* @worker: iteration cursor
400	* @pool: worker_pool to iterate workers of
401	*
402	* This must be called with @pool->attach_mutex.
403	*
404	* The if/else clause exists only for the lockdep assertion and can be
405	* ignored.
406	*/
407	#define for_each_pool_worker(worker, pool) \
408	list_for_each_entry((worker), &(pool)->workers, node) \
409	if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
410	else
411
412	/**
413	* for_each_pwq - iterate through all pool_workqueues of the specified workqueue
414	* @pwq: iteration cursor
415	* @wq: the target workqueue
416	*
417	* This must be called either with wq->mutex held or sched RCU read locked.
418	* If the pwq needs to be used beyond the locking in effect, the caller is
419	* responsible for guaranteeing that the pwq stays online.
420	*
421	* The if/else clause exists only for the lockdep assertion and can be
422	* ignored.
423	*/
424	#define for_each_pwq(pwq, wq) \
425	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
426	if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
427	else
428
429	#ifdef CONFIG_DEBUG_OBJECTS_WORK
430
431	static struct debug_obj_descr work_debug_descr;
432
433	static void work_debug_hint(void addr)
434	{
435	return ((struct work_struct *) addr)->func;
436	}
437
438	static bool work_is_static_object(void *addr)
439	{
440	struct work_struct *work = addr;
441
442	return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
443	}
444
445	/*
446	* fixup_init is called when:
447	* - an active object is initialized
448	*/
449	static bool work_fixup_init(void *addr, enum debug_obj_state state)
450	{
451	struct work_struct *work = addr;
452
453	switch (state) {
454	case ODEBUG_STATE_ACTIVE:
455	cancel_work_sync(work);
456	debug_object_init(work, &work_debug_descr);
457	return true;
458	default:
459	return false;
460	}
461	}
462
463	/*
464	* fixup_free is called when:
465	* - an active object is freed
466	*/
467	static bool work_fixup_free(void *addr, enum debug_obj_state state)
468	{
469	struct work_struct *work = addr;
470
471	switch (state) {
472	case ODEBUG_STATE_ACTIVE:
473	cancel_work_sync(work);
474	debug_object_free(work, &work_debug_descr);
475	return true;
476	default:
477	return false;
478	}
479	}
480
481	static struct debug_obj_descr work_debug_descr = {
482	.name = "work_struct",
483	.debug_hint = work_debug_hint,
484	.is_static_object = work_is_static_object,
485	.fixup_init = work_fixup_init,
486	.fixup_free = work_fixup_free,
487	};
488
489	static inline void debug_work_activate(struct work_struct *work)
490	{
491	debug_object_activate(work, &work_debug_descr);
492	}
493
494	static inline void debug_work_deactivate(struct work_struct *work)
495	{
496	debug_object_deactivate(work, &work_debug_descr);
497	}
498
499	void __init_work(struct work_struct *work, int onstack)
500	{
501	if (onstack)
502	debug_object_init_on_stack(work, &work_debug_descr);
503	else
504	debug_object_init(work, &work_debug_descr);
505	}
506	EXPORT_SYMBOL_GPL(__init_work);
507
508	void destroy_work_on_stack(struct work_struct *work)
509	{
510	debug_object_free(work, &work_debug_descr);
511	}
512	EXPORT_SYMBOL_GPL(destroy_work_on_stack);
513
514	void destroy_delayed_work_on_stack(struct delayed_work *work)
515	{
516	destroy_timer_on_stack(&work->timer);
517	debug_object_free(&work->work, &work_debug_descr);
518	}
519	EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
520
521	#else
522	static inline void debug_work_activate(struct work_struct *work) { }
523	static inline void debug_work_deactivate(struct work_struct *work) { }
524	#endif
525
526	/**
527	* worker_pool_assign_id - allocate ID and assing it to @pool
528	* @pool: the pool pointer of interest
529	*
530	* Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
531	* successfully, -errno on failure.
532	*/
533	static int worker_pool_assign_id(struct worker_pool *pool)
534	{
535	int ret;
536
537	lockdep_assert_held(&wq_pool_mutex);
538
539	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
540	GFP_KERNEL);
541	if (ret >= 0) {
542	pool->id = ret;
543	return 0;
544	}
545	return ret;
546	}
547
548	/**
549	* unbound_pwq_by_node - return the unbound pool_workqueue for the given node
550	* @wq: the target workqueue
551	* @node: the node ID
552	*
553	* This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
554	* read locked.
555	* If the pwq needs to be used beyond the locking in effect, the caller is
556	* responsible for guaranteeing that the pwq stays online.
557	*
558	* Return: The unbound pool_workqueue for @node.
559	*/
560	static struct pool_workqueue unbound_pwq_by_node(struct workqueue_struct wq,
561	int node)
562	{
563	assert_rcu_or_wq_mutex_or_pool_mutex(wq);
564
565	/*
566	* XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
567	* delayed item is pending. The plan is to keep CPU -> NODE
568	* mapping valid and stable across CPU on/offlines. Once that
569	* happens, this workaround can be removed.
570	*/
571	if (unlikely(node == NUMA_NO_NODE))
572	return wq->dfl_pwq;
573
574	return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
575	}
576
577	static unsigned int work_color_to_flags(int color)
578	{
579	return color << WORK_STRUCT_COLOR_SHIFT;
580	}
581
582	static int get_work_color(struct work_struct *work)
583	{
584	return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
585	((1 << WORK_STRUCT_COLOR_BITS) - 1);
586	}
587
588	static int work_next_color(int color)
589	{
590	return (color + 1) % WORK_NR_COLORS;
591	}
592
593	/*
594	* While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
595	* contain the pointer to the queued pwq. Once execution starts, the flag
596	* is cleared and the high bits contain OFFQ flags and pool ID.
597	*
598	* set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
599	* and clear_work_data() can be used to set the pwq, pool or clear
600	* work->data. These functions should only be called while the work is
601	* owned - ie. while the PENDING bit is set.
602	*
603	* get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
604	* corresponding to a work. Pool is available once the work has been
605	* queued anywhere after initialization until it is sync canceled. pwq is
606	* available only while the work item is queued.
607	*
608	* %WORK_OFFQ_CANCELING is used to mark a work item which is being
609	* canceled. While being canceled, a work item may have its PENDING set
610	* but stay off timer and worklist for arbitrarily long and nobody should
611	* try to steal the PENDING bit.
612	*/
613	static inline void set_work_data(struct work_struct *work, unsigned long data,
614	unsigned long flags)
615	{
616	WARN_ON_ONCE(!work_pending(work));
617	atomic_long_set(&work->data, data \| flags \| work_static(work));
618	}
619
620	static void set_work_pwq(struct work_struct work, struct pool_workqueue pwq,
621	unsigned long extra_flags)
622	{
623	set_work_data(work, (unsigned long)pwq,
624	WORK_STRUCT_PENDING \| WORK_STRUCT_PWQ \| extra_flags);
625	}
626
627	static void set_work_pool_and_keep_pending(struct work_struct *work,
628	int pool_id)
629	{
630	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
631	WORK_STRUCT_PENDING);
632	}
633
634	static void set_work_pool_and_clear_pending(struct work_struct *work,
635	int pool_id)
636	{
637	/*
638	* The following wmb is paired with the implied mb in
639	* test_and_set_bit(PENDING) and ensures all updates to @work made
640	* here are visible to and precede any updates by the next PENDING
641	* owner.
642	*/
643	smp_wmb();
644	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
645	/*
646	* The following mb guarantees that previous clear of a PENDING bit
647	* will not be reordered with any speculative LOADS or STORES from
648	* work->current_func, which is executed afterwards. This possible
649	* reordering can lead to a missed execution on attempt to qeueue
650	* the same @work. E.g. consider this case:
651	*
652	* CPU#0 CPU#1
653	* ---------------------------- --------------------------------
654	*
655	* 1 STORE event_indicated
656	* 2 queue_work_on() {
657	* 3 test_and_set_bit(PENDING)
658	* 4 } set_..._and_clear_pending() {
659	* 5 set_work_data() # clear bit
660	* 6 smp_mb()
661	* 7 work->current_func() {
662	* 8 LOAD event_indicated
663	* }
664	*
665	* Without an explicit full barrier speculative LOAD on line 8 can
666	* be executed before CPU#0 does STORE on line 1. If that happens,
667	* CPU#0 observes the PENDING bit is still set and new execution of
668	* a @work is not queued in a hope, that CPU#1 will eventually
669	* finish the queued @work. Meanwhile CPU#1 does not see
670	* event_indicated is set, because speculative LOAD was executed
671	* before actual STORE.
672	*/
673	smp_mb();
674	}
675
676	static void clear_work_data(struct work_struct *work)
677	{
678	smp_wmb(); /* see set_work_pool_and_clear_pending() */
679	set_work_data(work, WORK_STRUCT_NO_POOL, 0);
680	}
681
682	static struct pool_workqueue get_work_pwq(struct work_struct work)
683	{
684	unsigned long data = atomic_long_read(&work->data);
685
686	if (data & WORK_STRUCT_PWQ)
687	return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
688	else
689	return NULL;
690	}
691
692	/**
693	* get_work_pool - return the worker_pool a given work was associated with
694	* @work: the work item of interest
695	*
696	* Pools are created and destroyed under wq_pool_mutex, and allows read
697	* access under sched-RCU read lock. As such, this function should be
698	* called under wq_pool_mutex or with preemption disabled.
699	*
700	* All fields of the returned pool are accessible as long as the above
701	* mentioned locking is in effect. If the returned pool needs to be used
702	* beyond the critical section, the caller is responsible for ensuring the
703	* returned pool is and stays online.
704	*
705	* Return: The worker_pool @work was last associated with. %NULL if none.
706	*/
707	static struct worker_pool get_work_pool(struct work_struct work)
708	{
709	unsigned long data = atomic_long_read(&work->data);
710	int pool_id;
711
712	assert_rcu_or_pool_mutex();
713
714	if (data & WORK_STRUCT_PWQ)
715	return ((struct pool_workqueue *)
716	(data & WORK_STRUCT_WQ_DATA_MASK))->pool;
717
718	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
719	if (pool_id == WORK_OFFQ_POOL_NONE)
720	return NULL;
721
722	return idr_find(&worker_pool_idr, pool_id);
723	}
724
725	/**
726	* get_work_pool_id - return the worker pool ID a given work is associated with
727	* @work: the work item of interest
728	*
729	* Return: The worker_pool ID @work was last associated with.
730	* %WORK_OFFQ_POOL_NONE if none.
731	*/
732	static int get_work_pool_id(struct work_struct *work)
733	{
734	unsigned long data = atomic_long_read(&work->data);
735
736	if (data & WORK_STRUCT_PWQ)
737	return ((struct pool_workqueue *)
738	(data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
739
740	return data >> WORK_OFFQ_POOL_SHIFT;
741	}
742
743	static void mark_work_canceling(struct work_struct *work)
744	{
745	unsigned long pool_id = get_work_pool_id(work);
746
747	pool_id <<= WORK_OFFQ_POOL_SHIFT;
748	set_work_data(work, pool_id \| WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
749	}
750
751	static bool work_is_canceling(struct work_struct *work)
752	{
753	unsigned long data = atomic_long_read(&work->data);
754
755	return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
756	}
757
758	/*
759	* Policy functions. These define the policies on how the global worker
760	* pools are managed. Unless noted otherwise, these functions assume that
761	* they're being called with pool->lock held.
762	*/
763
764	static bool __need_more_worker(struct worker_pool *pool)
765	{
766	return !atomic_read(&pool->nr_running);
767	}
768
769	/*
770	* Need to wake up a worker? Called from anything but currently
771	* running workers.
772	*
773	* Note that, because unbound workers never contribute to nr_running, this
774	* function will always return %true for unbound pools as long as the
775	* worklist isn't empty.
776	*/
777	static bool need_more_worker(struct worker_pool *pool)
778	{
779	return !list_empty(&pool->worklist) && __need_more_worker(pool);
780	}
781
782	/* Can I start working? Called from busy but !running workers. */
783	static bool may_start_working(struct worker_pool *pool)
784	{
785	return pool->nr_idle;
786	}
787
788	/* Do I need to keep working? Called from currently running workers. */
789	static bool keep_working(struct worker_pool *pool)
790	{
791	return !list_empty(&pool->worklist) &&
792	atomic_read(&pool->nr_running) <= 1;
793	}
794
795	/* Do we need a new worker? Called from manager. */
796	static bool need_to_create_worker(struct worker_pool *pool)
797	{
798	return need_more_worker(pool) && !may_start_working(pool);
799	}
800
801	/* Do we have too many workers and should some go away? */
802	static bool too_many_workers(struct worker_pool *pool)
803	{
804	bool managing = mutex_is_locked(&pool->manager_arb);
805	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
806	int nr_busy = pool->nr_workers - nr_idle;
807
808	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
809	}
810
811	/*
812	* Wake up functions.
813	*/
814
815	/* Return the first idle worker. Safe with preemption disabled */
816	static struct worker first_idle_worker(struct worker_pool pool)
817	{
818	if (unlikely(list_empty(&pool->idle_list)))
819	return NULL;
820
821	return list_first_entry(&pool->idle_list, struct worker, entry);
822	}
823
824	/**
825	* wake_up_worker - wake up an idle worker
826	* @pool: worker pool to wake worker from
827	*
828	* Wake up the first idle worker of @pool.
829	*
830	* CONTEXT:
831	* spin_lock_irq(pool->lock).
832	*/
833	static void wake_up_worker(struct worker_pool *pool)
834	{
835	struct worker *worker = first_idle_worker(pool);
836
837	if (likely(worker))
838	wake_up_process(worker->task);
839	}
840
841	/**
842	* wq_worker_waking_up - a worker is waking up
843	* @task: task waking up
844	* @cpu: CPU @task is waking up to
845	*
846	* This function is called during try_to_wake_up() when a worker is
847	* being awoken.
848	*
849	* CONTEXT:
850	* spin_lock_irq(rq->lock)
851	*/
852	void wq_worker_waking_up(struct task_struct *task, int cpu)
853	{
854	struct worker *worker = kthread_data(task);
855
856	if (!(worker->flags & WORKER_NOT_RUNNING)) {
857	WARN_ON_ONCE(worker->pool->cpu != cpu);
858	atomic_inc(&worker->pool->nr_running);
859	}
860	}
861
862	/**
863	* wq_worker_sleeping - a worker is going to sleep
864	* @task: task going to sleep
865	*
866	* This function is called during schedule() when a busy worker is
867	* going to sleep. Worker on the same cpu can be woken up by
868	* returning pointer to its task.
869	*
870	* CONTEXT:
871	* spin_lock_irq(rq->lock)
872	*
873	* Return:
874	* Worker task on @cpu to wake up, %NULL if none.
875	*/
876	struct task_struct wq_worker_sleeping(struct task_struct task)
877	{
878	struct worker worker = kthread_data(task), to_wakeup = NULL;
879	struct worker_pool *pool;
880
881	/*
882	* Rescuers, which may not have all the fields set up like normal
883	* workers, also reach here, let's not access anything before
884	* checking NOT_RUNNING.
885	*/
886	if (worker->flags & WORKER_NOT_RUNNING)
887	return NULL;
888
889	pool = worker->pool;
890
891	/* this can only happen on the local cpu */
892	if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
893	return NULL;
894
895	/*
896	* The counterpart of the following dec_and_test, implied mb,
897	* worklist not empty test sequence is in insert_work().
898	* Please read comment there.
899	*
900	* NOT_RUNNING is clear. This means that we're bound to and
901	* running on the local cpu w/ rq lock held and preemption
902	* disabled, which in turn means that none else could be
903	* manipulating idle_list, so dereferencing idle_list without pool
904	* lock is safe.
905	*/
906	if (atomic_dec_and_test(&pool->nr_running) &&
907	!list_empty(&pool->worklist))
908	to_wakeup = first_idle_worker(pool);
909	return to_wakeup ? to_wakeup->task : NULL;
910	}
911
912	/**
913	* worker_set_flags - set worker flags and adjust nr_running accordingly
914	* @worker: self
915	* @flags: flags to set
916	*
917	* Set @flags in @worker->flags and adjust nr_running accordingly.
918	*
919	* CONTEXT:
920	* spin_lock_irq(pool->lock)
921	*/
922	static inline void worker_set_flags(struct worker *worker, unsigned int flags)
923	{
924	struct worker_pool *pool = worker->pool;
925
926	WARN_ON_ONCE(worker->task != current);
927
928	/* If transitioning into NOT_RUNNING, adjust nr_running. */
929	if ((flags & WORKER_NOT_RUNNING) &&
930	!(worker->flags & WORKER_NOT_RUNNING)) {
931	atomic_dec(&pool->nr_running);
932	}
933
934	worker->flags \|= flags;
935	}
936
937	/**
938	* worker_clr_flags - clear worker flags and adjust nr_running accordingly
939	* @worker: self
940	* @flags: flags to clear
941	*
942	* Clear @flags in @worker->flags and adjust nr_running accordingly.
943	*
944	* CONTEXT:
945	* spin_lock_irq(pool->lock)
946	*/
947	static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
948	{
949	struct worker_pool *pool = worker->pool;
950	unsigned int oflags = worker->flags;
951
952	WARN_ON_ONCE(worker->task != current);
953
954	worker->flags &= ~flags;
955
956	/*
957	* If transitioning out of NOT_RUNNING, increment nr_running. Note
958	* that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
959	* of multiple flags, not a single flag.
960	*/
961	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
962	if (!(worker->flags & WORKER_NOT_RUNNING))
963	atomic_inc(&pool->nr_running);
964	}
965
966	/**
967	* find_worker_executing_work - find worker which is executing a work
968	* @pool: pool of interest
969	* @work: work to find worker for
970	*
971	* Find a worker which is executing @work on @pool by searching
972	* @pool->busy_hash which is keyed by the address of @work. For a worker
973	* to match, its current execution should match the address of @work and
974	* its work function. This is to avoid unwanted dependency between
975	* unrelated work executions through a work item being recycled while still
976	* being executed.
977	*
978	* This is a bit tricky. A work item may be freed once its execution
979	* starts and nothing prevents the freed area from being recycled for
980	* another work item. If the same work item address ends up being reused
981	* before the original execution finishes, workqueue will identify the
982	* recycled work item as currently executing and make it wait until the
983	* current execution finishes, introducing an unwanted dependency.
984	*
985	* This function checks the work item address and work function to avoid
986	* false positives. Note that this isn't complete as one may construct a
987	* work function which can introduce dependency onto itself through a
988	* recycled work item. Well, if somebody wants to shoot oneself in the
989	* foot that badly, there's only so much we can do, and if such deadlock
990	* actually occurs, it should be easy to locate the culprit work function.
991	*
992	* CONTEXT:
993	* spin_lock_irq(pool->lock).
994	*
995	* Return:
996	* Pointer to worker which is executing @work if found, %NULL
997	* otherwise.
998	*/
999	static struct worker find_worker_executing_work(struct worker_pool pool,
1000	struct work_struct *work)
1001	{
1002	struct worker *worker;
1003
1004	hash_for_each_possible(pool->busy_hash, worker, hentry,
1005	(unsigned long)work)
1006	if (worker->current_work == work &&
1007	worker->current_func == work->func)
1008	return worker;
1009
1010	return NULL;
1011	}
1012
1013	/**
1014	* move_linked_works - move linked works to a list
1015	* @work: start of series of works to be scheduled
1016	* @head: target list to append @work to
1017	* @nextp: out parameter for nested worklist walking
1018	*
1019	* Schedule linked works starting from @work to @head. Work series to
1020	* be scheduled starts at @work and includes any consecutive work with
1021	* WORK_STRUCT_LINKED set in its predecessor.
1022	*
1023	* If @nextp is not NULL, it's updated to point to the next work of
1024	* the last scheduled work. This allows move_linked_works() to be
1025	* nested inside outer list_for_each_entry_safe().
1026	*
1027	* CONTEXT:
1028	* spin_lock_irq(pool->lock).
1029	*/
1030	static void move_linked_works(struct work_struct work, struct list_head head,
1031	struct work_struct **nextp)
1032	{
1033	struct work_struct *n;
1034
1035	/*
1036	* Linked worklist will always end before the end of the list,
1037	* use NULL for list head.
1038	*/
1039	list_for_each_entry_safe_from(work, n, NULL, entry) {
1040	list_move_tail(&work->entry, head);
1041	if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1042	break;
1043	}
1044
1045	/*
1046	* If we're already inside safe list traversal and have moved
1047	* multiple works to the scheduled queue, the next position
1048	* needs to be updated.
1049	*/
1050	if (nextp)
1051	*nextp = n;
1052	}
1053
1054	/**
1055	* get_pwq - get an extra reference on the specified pool_workqueue
1056	* @pwq: pool_workqueue to get
1057	*
1058	* Obtain an extra reference on @pwq. The caller should guarantee that
1059	* @pwq has positive refcnt and be holding the matching pool->lock.
1060	*/
1061	static void get_pwq(struct pool_workqueue *pwq)
1062	{
1063	lockdep_assert_held(&pwq->pool->lock);
1064	WARN_ON_ONCE(pwq->refcnt <= 0);
1065	pwq->refcnt++;
1066	}
1067
1068	/**
1069	* put_pwq - put a pool_workqueue reference
1070	* @pwq: pool_workqueue to put
1071	*
1072	* Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1073	* destruction. The caller should be holding the matching pool->lock.
1074	*/
1075	static void put_pwq(struct pool_workqueue *pwq)
1076	{
1077	lockdep_assert_held(&pwq->pool->lock);
1078	if (likely(--pwq->refcnt))
1079	return;
1080	if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1081	return;
1082	/*
1083	* @pwq can't be released under pool->lock, bounce to
1084	* pwq_unbound_release_workfn(). This never recurses on the same
1085	* pool->lock as this path is taken only for unbound workqueues and
1086	* the release work item is scheduled on a per-cpu workqueue. To
1087	* avoid lockdep warning, unbound pool->locks are given lockdep
1088	* subclass of 1 in get_unbound_pool().
1089	*/
1090	schedule_work(&pwq->unbound_release_work);
1091	}
1092
1093	/**
1094	* put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1095	* @pwq: pool_workqueue to put (can be %NULL)
1096	*
1097	* put_pwq() with locking. This function also allows %NULL @pwq.
1098	*/
1099	static void put_pwq_unlocked(struct pool_workqueue *pwq)
1100	{
1101	if (pwq) {
1102	/*
1103	* As both pwqs and pools are sched-RCU protected, the
1104	* following lock operations are safe.
1105	*/
1106	spin_lock_irq(&pwq->pool->lock);
1107	put_pwq(pwq);
1108	spin_unlock_irq(&pwq->pool->lock);
1109	}
1110	}
1111
1112	static void pwq_activate_delayed_work(struct work_struct *work)
1113	{
1114	struct pool_workqueue *pwq = get_work_pwq(work);
1115
1116	trace_workqueue_activate_work(work);
1117	if (list_empty(&pwq->pool->worklist))
1118	pwq->pool->watchdog_ts = jiffies;
1119	move_linked_works(work, &pwq->pool->worklist, NULL);
1120	__clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1121	pwq->nr_active++;
1122	}
1123
1124	static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1125	{
1126	struct work_struct *work = list_first_entry(&pwq->delayed_works,
1127	struct work_struct, entry);
1128
1129	pwq_activate_delayed_work(work);
1130	}
1131
1132	/**
1133	* pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1134	* @pwq: pwq of interest
1135	* @color: color of work which left the queue
1136	*
1137	* A work either has completed or is removed from pending queue,
1138	* decrement nr_in_flight of its pwq and handle workqueue flushing.
1139	*
1140	* CONTEXT:
1141	* spin_lock_irq(pool->lock).
1142	*/
1143	static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1144	{
1145	/* uncolored work items don't participate in flushing or nr_active */
1146	if (color == WORK_NO_COLOR)
1147	goto out_put;
1148
1149	pwq->nr_in_flight[color]--;
1150
1151	pwq->nr_active--;
1152	if (!list_empty(&pwq->delayed_works)) {
1153	/* one down, submit a delayed one */
1154	if (pwq->nr_active < pwq->max_active)
1155	pwq_activate_first_delayed(pwq);
1156	}
1157
1158	/* is flush in progress and are we at the flushing tip? */
1159	if (likely(pwq->flush_color != color))
1160	goto out_put;
1161
1162	/* are there still in-flight works? */
1163	if (pwq->nr_in_flight[color])
1164	goto out_put;
1165
1166	/* this pwq is done, clear flush_color */
1167	pwq->flush_color = -1;
1168
1169	/*
1170	* If this was the last pwq, wake up the first flusher. It
1171	* will handle the rest.
1172	*/
1173	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1174	complete(&pwq->wq->first_flusher->done);
1175	out_put:
1176	put_pwq(pwq);
1177	}
1178
1179	/**
1180	* try_to_grab_pending - steal work item from worklist and disable irq
1181	* @work: work item to steal
1182	* @is_dwork: @work is a delayed_work
1183	* @flags: place to store irq state
1184	*
1185	* Try to grab PENDING bit of @work. This function can handle @work in any
1186	* stable state - idle, on timer or on worklist.
1187	*
1188	* Return:
1189	* 1 if @work was pending and we successfully stole PENDING
1190	* 0 if @work was idle and we claimed PENDING
1191	* -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1192	* -ENOENT if someone else is canceling @work, this state may persist
1193	* for arbitrarily long
1194	*
1195	* Note:
1196	* On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1197	* interrupted while holding PENDING and @work off queue, irq must be
1198	* disabled on entry. This, combined with delayed_work->timer being
1199	* irqsafe, ensures that we return -EAGAIN for finite short period of time.
1200	*
1201	* On successful return, >= 0, irq is disabled and the caller is
1202	* responsible for releasing it using local_irq_restore(*@flags).
1203	*
1204	* This function is safe to call from any context including IRQ handler.
1205	*/
1206	static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1207	unsigned long *flags)
1208	{
1209	struct worker_pool *pool;
1210	struct pool_workqueue *pwq;
1211
1212	local_irq_save(*flags);
1213
1214	/* try to steal the timer if it exists */
1215	if (is_dwork) {
1216	struct delayed_work *dwork = to_delayed_work(work);
1217
1218	/*
1219	* dwork->timer is irqsafe. If del_timer() fails, it's
1220	* guaranteed that the timer is not queued anywhere and not
1221	* running on the local CPU.
1222	*/
1223	if (likely(del_timer(&dwork->timer)))
1224	return 1;
1225	}
1226
1227	/* try to claim PENDING the normal way */
1228	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1229	return 0;
1230
1231	/*
1232	* The queueing is in progress, or it is already queued. Try to
1233	* steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1234	*/
1235	pool = get_work_pool(work);
1236	if (!pool)
1237	goto fail;
1238
1239	spin_lock(&pool->lock);
1240	/*
1241	* work->data is guaranteed to point to pwq only while the work
1242	* item is queued on pwq->wq, and both updating work->data to point
1243	* to pwq on queueing and to pool on dequeueing are done under
1244	* pwq->pool->lock. This in turn guarantees that, if work->data
1245	* points to pwq which is associated with a locked pool, the work
1246	* item is currently queued on that pool.
1247	*/
1248	pwq = get_work_pwq(work);
1249	if (pwq && pwq->pool == pool) {
1250	debug_work_deactivate(work);
1251
1252	/*
1253	* A delayed work item cannot be grabbed directly because
1254	* it might have linked NO_COLOR work items which, if left
1255	* on the delayed_list, will confuse pwq->nr_active
1256	* management later on and cause stall. Make sure the work
1257	* item is activated before grabbing.
1258	*/
1259	if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1260	pwq_activate_delayed_work(work);
1261
1262	list_del_init(&work->entry);
1263	pwq_dec_nr_in_flight(pwq, get_work_color(work));
1264
1265	/* work->data points to pwq iff queued, point to pool */
1266	set_work_pool_and_keep_pending(work, pool->id);
1267
1268	spin_unlock(&pool->lock);
1269	return 1;
1270	}
1271	spin_unlock(&pool->lock);
1272	fail:
1273	local_irq_restore(*flags);
1274	if (work_is_canceling(work))
1275	return -ENOENT;
1276	cpu_relax();
1277	return -EAGAIN;
1278	}
1279
1280	/**
1281	* insert_work - insert a work into a pool
1282	* @pwq: pwq @work belongs to
1283	* @work: work to insert
1284	* @head: insertion point
1285	* @extra_flags: extra WORK_STRUCT_* flags to set
1286	*
1287	* Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1288	* work_struct flags.
1289	*
1290	* CONTEXT:
1291	* spin_lock_irq(pool->lock).
1292	*/
1293	static void insert_work(struct pool_workqueue pwq, struct work_struct work,
1294	struct list_head *head, unsigned int extra_flags)
1295	{
1296	struct worker_pool *pool = pwq->pool;
1297
1298	/* we own @work, set data and link */
1299	set_work_pwq(work, pwq, extra_flags);
1300	list_add_tail(&work->entry, head);
1301	get_pwq(pwq);
1302
1303	/*
1304	* Ensure either wq_worker_sleeping() sees the above
1305	* list_add_tail() or we see zero nr_running to avoid workers lying
1306	* around lazily while there are works to be processed.
1307	*/
1308	smp_mb();
1309
1310	if (__need_more_worker(pool))
1311	wake_up_worker(pool);
1312	}
1313
1314	/*
1315	* Test whether @work is being queued from another work executing on the
1316	* same workqueue.
1317	*/
1318	static bool is_chained_work(struct workqueue_struct *wq)
1319	{
1320	struct worker *worker;
1321
1322	worker = current_wq_worker();
1323	/*
1324	* Return %true iff I'm a worker execuing a work item on @wq. If
1325	* I'm @worker, it's safe to dereference it without locking.
1326	*/
1327	return worker && worker->current_pwq->wq == wq;
1328	}
1329
1330	/*
1331	* When queueing an unbound work item to a wq, prefer local CPU if allowed
1332	* by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1333	* avoid perturbing sensitive tasks.
1334	*/
1335	static int wq_select_unbound_cpu(int cpu)
1336	{
1337	static bool printed_dbg_warning;
1338	int new_cpu;
1339
1340	if (likely(!wq_debug_force_rr_cpu)) {
1341	if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1342	return cpu;
1343	} else if (!printed_dbg_warning) {
1344	pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1345	printed_dbg_warning = true;
1346	}
1347
1348	if (cpumask_empty(wq_unbound_cpumask))
1349	return cpu;
1350
1351	new_cpu = __this_cpu_read(wq_rr_cpu_last);
1352	new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1353	if (unlikely(new_cpu >= nr_cpu_ids)) {
1354	new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1355	if (unlikely(new_cpu >= nr_cpu_ids))
1356	return cpu;
1357	}
1358	__this_cpu_write(wq_rr_cpu_last, new_cpu);
1359
1360	return new_cpu;
1361	}
1362
1363	static void __queue_work(int cpu, struct workqueue_struct *wq,
1364	struct work_struct *work)
1365	{
1366	struct pool_workqueue *pwq;
1367	struct worker_pool *last_pool;
1368	struct list_head *worklist;
1369	unsigned int work_flags;
1370	unsigned int req_cpu = cpu;
1371
1372	/*
1373	* While a work item is PENDING && off queue, a task trying to
1374	* steal the PENDING will busy-loop waiting for it to either get
1375	* queued or lose PENDING. Grabbing PENDING and queueing should
1376	* happen with IRQ disabled.
1377	*/
1378	WARN_ON_ONCE(!irqs_disabled());
1379
1380	debug_work_activate(work);
1381
1382	/* if draining, only works from the same workqueue are allowed */
1383	if (unlikely(wq->flags & __WQ_DRAINING) &&
1384	WARN_ON_ONCE(!is_chained_work(wq)))
1385	return;
1386	retry:
1387	if (req_cpu == WORK_CPU_UNBOUND)
1388	cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1389
1390	/* pwq which will be used unless @work is executing elsewhere */
1391	if (!(wq->flags & WQ_UNBOUND))
1392	pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1393	else
1394	pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1395
1396	/*
1397	* If @work was previously on a different pool, it might still be
1398	* running there, in which case the work needs to be queued on that
1399	* pool to guarantee non-reentrancy.
1400	*/
1401	last_pool = get_work_pool(work);
1402	if (last_pool && last_pool != pwq->pool) {
1403	struct worker *worker;
1404
1405	spin_lock(&last_pool->lock);
1406
1407	worker = find_worker_executing_work(last_pool, work);
1408
1409	if (worker && worker->current_pwq->wq == wq) {
1410	pwq = worker->current_pwq;
1411	} else {
1412	/* meh... not running there, queue here */
1413	spin_unlock(&last_pool->lock);
1414	spin_lock(&pwq->pool->lock);
1415	}
1416	} else {
1417	spin_lock(&pwq->pool->lock);
1418	}
1419
1420	/*
1421	* pwq is determined and locked. For unbound pools, we could have
1422	* raced with pwq release and it could already be dead. If its
1423	* refcnt is zero, repeat pwq selection. Note that pwqs never die
1424	* without another pwq replacing it in the numa_pwq_tbl or while
1425	* work items are executing on it, so the retrying is guaranteed to
1426	* make forward-progress.
1427	*/
1428	if (unlikely(!pwq->refcnt)) {
1429	if (wq->flags & WQ_UNBOUND) {
1430	spin_unlock(&pwq->pool->lock);
1431	cpu_relax();
1432	goto retry;
1433	}
1434	/* oops */
1435	WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1436	wq->name, cpu);
1437	}
1438
1439	/* pwq determined, queue */
1440	trace_workqueue_queue_work(req_cpu, pwq, work);
1441
1442	if (WARN_ON(!list_empty(&work->entry))) {
1443	spin_unlock(&pwq->pool->lock);
1444	return;
1445	}
1446
1447	pwq->nr_in_flight[pwq->work_color]++;
1448	work_flags = work_color_to_flags(pwq->work_color);
1449
1450	if (likely(pwq->nr_active < pwq->max_active)) {
1451	trace_workqueue_activate_work(work);
1452	pwq->nr_active++;
1453	worklist = &pwq->pool->worklist;
1454	if (list_empty(worklist))
1455	pwq->pool->watchdog_ts = jiffies;
1456	} else {
1457	work_flags \|= WORK_STRUCT_DELAYED;
1458	worklist = &pwq->delayed_works;
1459	}
1460
1461	insert_work(pwq, work, worklist, work_flags);
1462
1463	spin_unlock(&pwq->pool->lock);
1464	}
1465
1466	/**
1467	* queue_work_on - queue work on specific cpu
1468	* @cpu: CPU number to execute work on
1469	* @wq: workqueue to use
1470	* @work: work to queue
1471	*
1472	* We queue the work to a specific CPU, the caller must ensure it
1473	* can't go away.
1474	*
1475	* Return: %false if @work was already on a queue, %true otherwise.
1476	*/
1477	bool queue_work_on(int cpu, struct workqueue_struct *wq,
1478	struct work_struct *work)
1479	{
1480	bool ret = false;
1481	unsigned long flags;
1482
1483	local_irq_save(flags);
1484
1485	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1486	__queue_work(cpu, wq, work);
1487	ret = true;
1488	}
1489
1490	local_irq_restore(flags);
1491	return ret;
1492	}
1493	EXPORT_SYMBOL(queue_work_on);
1494
1495	void delayed_work_timer_fn(unsigned long __data)
1496	{
1497	struct delayed_work dwork = (struct delayed_work )__data;
1498
1499	/* should have been called from irqsafe timer with irq already off */
1500	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
1501	}
1502	EXPORT_SYMBOL(delayed_work_timer_fn);
1503
1504	static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1505	struct delayed_work *dwork, unsigned long delay)
1506	{
1507	struct timer_list *timer = &dwork->timer;
1508	struct work_struct *work = &dwork->work;
1509
1510	WARN_ON_ONCE(timer->function != delayed_work_timer_fn \|\|
1511	timer->data != (unsigned long)dwork);
1512	WARN_ON_ONCE(timer_pending(timer));
1513	WARN_ON_ONCE(!list_empty(&work->entry));
1514
1515	/*
1516	* If @delay is 0, queue @dwork->work immediately. This is for
1517	* both optimization and correctness. The earliest @timer can
1518	* expire is on the closest next tick and delayed_work users depend
1519	* on that there's no such delay when @delay is 0.
1520	*/
1521	if (!delay) {
1522	__queue_work(cpu, wq, &dwork->work);
1523	return;
1524	}
1525
1526	dwork->wq = wq;
1527	dwork->cpu = cpu;
1528	timer->expires = jiffies + delay;
1529
1530	if (unlikely(cpu != WORK_CPU_UNBOUND))
1531	add_timer_on(timer, cpu);
1532	else
1533	add_timer(timer);
1534	}
1535
1536	/**
1537	* queue_delayed_work_on - queue work on specific CPU after delay
1538	* @cpu: CPU number to execute work on
1539	* @wq: workqueue to use
1540	* @dwork: work to queue
1541	* @delay: number of jiffies to wait before queueing
1542	*
1543	* Return: %false if @work was already on a queue, %true otherwise. If
1544	* @delay is zero and @dwork is idle, it will be scheduled for immediate
1545	* execution.
1546	*/
1547	bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1548	struct delayed_work *dwork, unsigned long delay)
1549	{
1550	struct work_struct *work = &dwork->work;
1551	bool ret = false;
1552	unsigned long flags;
1553
1554	/* read the comment in __queue_work() */
1555	local_irq_save(flags);
1556
1557	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1558	__queue_delayed_work(cpu, wq, dwork, delay);
1559	ret = true;
1560	}
1561
1562	local_irq_restore(flags);
1563	return ret;
1564	}
1565	EXPORT_SYMBOL(queue_delayed_work_on);
1566
1567	/**
1568	* mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1569	* @cpu: CPU number to execute work on
1570	* @wq: workqueue to use
1571	* @dwork: work to queue
1572	* @delay: number of jiffies to wait before queueing
1573	*
1574	* If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1575	* modify @dwork's timer so that it expires after @delay. If @delay is
1576	* zero, @work is guaranteed to be scheduled immediately regardless of its
1577	* current state.
1578	*
1579	* Return: %false if @dwork was idle and queued, %true if @dwork was
1580	* pending and its timer was modified.
1581	*
1582	* This function is safe to call from any context including IRQ handler.
1583	* See try_to_grab_pending() for details.
1584	*/
1585	bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1586	struct delayed_work *dwork, unsigned long delay)
1587	{
1588	unsigned long flags;
1589	int ret;
1590
1591	do {
1592	ret = try_to_grab_pending(&dwork->work, true, &flags);
1593	} while (unlikely(ret == -EAGAIN));
1594
1595	if (likely(ret >= 0)) {
1596	__queue_delayed_work(cpu, wq, dwork, delay);
1597	local_irq_restore(flags);
1598	}
1599
1600	/* -ENOENT from try_to_grab_pending() becomes %true */
1601	return ret;
1602	}
1603	EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1604
1605	/**
1606	* worker_enter_idle - enter idle state
1607	* @worker: worker which is entering idle state
1608	*
1609	* @worker is entering idle state. Update stats and idle timer if
1610	* necessary.
1611	*
1612	* LOCKING:
1613	* spin_lock_irq(pool->lock).
1614	*/
1615	static void worker_enter_idle(struct worker *worker)
1616	{
1617	struct worker_pool *pool = worker->pool;
1618
1619	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) \|\|
1620	WARN_ON_ONCE(!list_empty(&worker->entry) &&
1621	(worker->hentry.next \|\| worker->hentry.pprev)))
1622	return;
1623
1624	/* can't use worker_set_flags(), also called from create_worker() */
1625	worker->flags \|= WORKER_IDLE;
1626	pool->nr_idle++;
1627	worker->last_active = jiffies;
1628
1629	/* idle_list is LIFO */
1630	list_add(&worker->entry, &pool->idle_list);
1631
1632	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1633	mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1634
1635	/*
1636	* Sanity check nr_running. Because wq_unbind_fn() releases
1637	* pool->lock between setting %WORKER_UNBOUND and zapping
1638	* nr_running, the warning may trigger spuriously. Check iff
1639	* unbind is not in progress.
1640	*/
1641	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1642	pool->nr_workers == pool->nr_idle &&
1643	atomic_read(&pool->nr_running));
1644	}
1645
1646	/**
1647	* worker_leave_idle - leave idle state
1648	* @worker: worker which is leaving idle state
1649	*
1650	* @worker is leaving idle state. Update stats.
1651	*
1652	* LOCKING:
1653	* spin_lock_irq(pool->lock).
1654	*/
1655	static void worker_leave_idle(struct worker *worker)
1656	{
1657	struct worker_pool *pool = worker->pool;
1658
1659	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1660	return;
1661	worker_clr_flags(worker, WORKER_IDLE);
1662	pool->nr_idle--;
1663	list_del_init(&worker->entry);
1664	}
1665
1666	static struct worker *alloc_worker(int node)
1667	{
1668	struct worker *worker;
1669
1670	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1671	if (worker) {
1672	INIT_LIST_HEAD(&worker->entry);
1673	INIT_LIST_HEAD(&worker->scheduled);
1674	INIT_LIST_HEAD(&worker->node);
1675	/* on creation a worker is in !idle && prep state */
1676	worker->flags = WORKER_PREP;
1677	}
1678	return worker;
1679	}
1680
1681	/**
1682	* worker_attach_to_pool() - attach a worker to a pool
1683	* @worker: worker to be attached
1684	* @pool: the target pool
1685	*
1686	* Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1687	* cpu-binding of @worker are kept coordinated with the pool across
1688	* cpu-[un]hotplugs.
1689	*/
1690	static void worker_attach_to_pool(struct worker *worker,
1691	struct worker_pool *pool)
1692	{
1693	mutex_lock(&pool->attach_mutex);
1694
1695	/*
1696	* set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1697	* online CPUs. It'll be re-applied when any of the CPUs come up.
1698	*/
1699	set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1700
1701	/*
1702	* The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1703	* stable across this function. See the comments above the
1704	* flag definition for details.
1705	*/
1706	if (pool->flags & POOL_DISASSOCIATED)
1707	worker->flags \|= WORKER_UNBOUND;
1708
1709	list_add_tail(&worker->node, &pool->workers);
1710
1711	mutex_unlock(&pool->attach_mutex);
1712	}
1713
1714	/**
1715	* worker_detach_from_pool() - detach a worker from its pool
1716	* @worker: worker which is attached to its pool
1717	* @pool: the pool @worker is attached to
1718	*
1719	* Undo the attaching which had been done in worker_attach_to_pool(). The
1720	* caller worker shouldn't access to the pool after detached except it has
1721	* other reference to the pool.
1722	*/
1723	static void worker_detach_from_pool(struct worker *worker,
1724	struct worker_pool *pool)
1725	{
1726	struct completion *detach_completion = NULL;
1727
1728	mutex_lock(&pool->attach_mutex);
1729	list_del(&worker->node);
1730	if (list_empty(&pool->workers))
1731	detach_completion = pool->detach_completion;
1732	mutex_unlock(&pool->attach_mutex);
1733
1734	/* clear leftover flags without pool->lock after it is detached */
1735	worker->flags &= ~(WORKER_UNBOUND \| WORKER_REBOUND);
1736
1737	if (detach_completion)
1738	complete(detach_completion);
1739	}
1740
1741	/**
1742	* create_worker - create a new workqueue worker
1743	* @pool: pool the new worker will belong to
1744	*
1745	* Create and start a new worker which is attached to @pool.
1746	*
1747	* CONTEXT:
1748	* Might sleep. Does GFP_KERNEL allocations.
1749	*
1750	* Return:
1751	* Pointer to the newly created worker.
1752	*/
1753	static struct worker create_worker(struct worker_pool pool)
1754	{
1755	struct worker *worker = NULL;
1756	int id = -1;
1757	char id_buf[16];
1758
1759	/* ID is needed to determine kthread name */
1760	id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1761	if (id < 0)
1762	goto fail;
1763
1764	worker = alloc_worker(pool->node);
1765	if (!worker)
1766	goto fail;
1767
1768	worker->pool = pool;
1769	worker->id = id;
1770
1771	if (pool->cpu >= 0)
1772	snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1773	pool->attrs->nice < 0 ? "H" : "");
1774	else
1775	snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1776
1777	worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1778	"kworker/%s", id_buf);
1779	if (IS_ERR(worker->task))
1780	goto fail;
1781
1782	set_user_nice(worker->task, pool->attrs->nice);
1783	kthread_bind_mask(worker->task, pool->attrs->cpumask);
1784
1785	/* successful, attach the worker to the pool */
1786	worker_attach_to_pool(worker, pool);
1787
1788	/* start the newly created worker */
1789	spin_lock_irq(&pool->lock);
1790	worker->pool->nr_workers++;
1791	worker_enter_idle(worker);
1792	wake_up_process(worker->task);
1793	spin_unlock_irq(&pool->lock);
1794
1795	return worker;
1796
1797	fail:
1798	if (id >= 0)
1799	ida_simple_remove(&pool->worker_ida, id);
1800	kfree(worker);
1801	return NULL;
1802	}
1803
1804	/**
1805	* destroy_worker - destroy a workqueue worker
1806	* @worker: worker to be destroyed
1807	*
1808	* Destroy @worker and adjust @pool stats accordingly. The worker should
1809	* be idle.
1810	*
1811	* CONTEXT:
1812	* spin_lock_irq(pool->lock).
1813	*/
1814	static void destroy_worker(struct worker *worker)
1815	{
1816	struct worker_pool *pool = worker->pool;
1817
1818	lockdep_assert_held(&pool->lock);
1819
1820	/* sanity check frenzy */
1821	if (WARN_ON(worker->current_work) \|\|
1822	WARN_ON(!list_empty(&worker->scheduled)) \|\|
1823	WARN_ON(!(worker->flags & WORKER_IDLE)))
1824	return;
1825
1826	pool->nr_workers--;
1827	pool->nr_idle--;
1828
1829	list_del_init(&worker->entry);
1830	worker->flags \|= WORKER_DIE;
1831	wake_up_process(worker->task);
1832	}
1833
1834	static void idle_worker_timeout(unsigned long __pool)
1835	{
1836	struct worker_pool pool = (void )__pool;
1837
1838	spin_lock_irq(&pool->lock);
1839
1840	while (too_many_workers(pool)) {
1841	struct worker *worker;
1842	unsigned long expires;
1843
1844	/* idle_list is kept in LIFO order, check the last one */
1845	worker = list_entry(pool->idle_list.prev, struct worker, entry);
1846	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1847
1848	if (time_before(jiffies, expires)) {
1849	mod_timer(&pool->idle_timer, expires);
1850	break;
1851	}
1852
1853	destroy_worker(worker);
1854	}
1855
1856	spin_unlock_irq(&pool->lock);
1857	}
1858
1859	static void send_mayday(struct work_struct *work)
1860	{
1861	struct pool_workqueue *pwq = get_work_pwq(work);
1862	struct workqueue_struct *wq = pwq->wq;
1863
1864	lockdep_assert_held(&wq_mayday_lock);
1865
1866	if (!wq->rescuer)
1867	return;
1868
1869	/* mayday mayday mayday */
1870	if (list_empty(&pwq->mayday_node)) {
1871	/*
1872	* If @pwq is for an unbound wq, its base ref may be put at
1873	* any time due to an attribute change. Pin @pwq until the
1874	* rescuer is done with it.
1875	*/
1876	get_pwq(pwq);
1877	list_add_tail(&pwq->mayday_node, &wq->maydays);
1878	wake_up_process(wq->rescuer->task);
1879	}
1880	}
1881
1882	static void pool_mayday_timeout(unsigned long __pool)
1883	{
1884	struct worker_pool pool = (void )__pool;
1885	struct work_struct *work;
1886
1887	spin_lock_irq(&pool->lock);
1888	spin_lock(&wq_mayday_lock); /* for wq->maydays */
1889
1890	if (need_to_create_worker(pool)) {
1891	/*
1892	* We've been trying to create a new worker but
1893	* haven't been successful. We might be hitting an
1894	* allocation deadlock. Send distress signals to
1895	* rescuers.
1896	*/
1897	list_for_each_entry(work, &pool->worklist, entry)
1898	send_mayday(work);
1899	}
1900
1901	spin_unlock(&wq_mayday_lock);
1902	spin_unlock_irq(&pool->lock);
1903
1904	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1905	}
1906
1907	/**
1908	* maybe_create_worker - create a new worker if necessary
1909	* @pool: pool to create a new worker for
1910	*
1911	* Create a new worker for @pool if necessary. @pool is guaranteed to
1912	* have at least one idle worker on return from this function. If
1913	* creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1914	* sent to all rescuers with works scheduled on @pool to resolve
1915	* possible allocation deadlock.
1916	*
1917	* On return, need_to_create_worker() is guaranteed to be %false and
1918	* may_start_working() %true.
1919	*
1920	* LOCKING:
1921	* spin_lock_irq(pool->lock) which may be released and regrabbed
1922	* multiple times. Does GFP_KERNEL allocations. Called only from
1923	* manager.
1924	*/
1925	static void maybe_create_worker(struct worker_pool *pool)
1926	__releases(&pool->lock)
1927	__acquires(&pool->lock)
1928	{
1929	restart:
1930	spin_unlock_irq(&pool->lock);
1931
1932	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1933	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1934
1935	while (true) {
1936	if (create_worker(pool) \|\| !need_to_create_worker(pool))
1937	break;
1938
1939	schedule_timeout_interruptible(CREATE_COOLDOWN);
1940
1941	if (!need_to_create_worker(pool))
1942	break;
1943	}
1944
1945	del_timer_sync(&pool->mayday_timer);
1946	spin_lock_irq(&pool->lock);
1947	/*
1948	* This is necessary even after a new worker was just successfully
1949	* created as @pool->lock was dropped and the new worker might have
1950	* already become busy.
1951	*/
1952	if (need_to_create_worker(pool))
1953	goto restart;
1954	}
1955
1956	/**
1957	* manage_workers - manage worker pool
1958	* @worker: self
1959	*
1960	* Assume the manager role and manage the worker pool @worker belongs
1961	* to. At any given time, there can be only zero or one manager per
1962	* pool. The exclusion is handled automatically by this function.
1963	*
1964	* The caller can safely start processing works on false return. On
1965	* true return, it's guaranteed that need_to_create_worker() is false
1966	* and may_start_working() is true.
1967	*
1968	* CONTEXT:
1969	* spin_lock_irq(pool->lock) which may be released and regrabbed
1970	* multiple times. Does GFP_KERNEL allocations.
1971	*
1972	* Return:
1973	* %false if the pool doesn't need management and the caller can safely
1974	* start processing works, %true if management function was performed and
1975	* the conditions that the caller verified before calling the function may
1976	* no longer be true.
1977	*/
1978	static bool manage_workers(struct worker *worker)
1979	{
1980	struct worker_pool *pool = worker->pool;
1981
1982	/*
1983	* Anyone who successfully grabs manager_arb wins the arbitration
1984	* and becomes the manager. mutex_trylock() on pool->manager_arb
1985	* failure while holding pool->lock reliably indicates that someone
1986	* else is managing the pool and the worker which failed trylock
1987	* can proceed to executing work items. This means that anyone
1988	* grabbing manager_arb is responsible for actually performing
1989	* manager duties. If manager_arb is grabbed and released without
1990	* actual management, the pool may stall indefinitely.
1991	*/
1992	if (!mutex_trylock(&pool->manager_arb))
1993	return false;
1994	pool->manager = worker;
1995
1996	maybe_create_worker(pool);
1997
1998	pool->manager = NULL;
1999	mutex_unlock(&pool->manager_arb);
2000	return true;
2001	}
2002
2003	/**
2004	* process_one_work - process single work
2005	* @worker: self
2006	* @work: work to process
2007	*
2008	* Process @work. This function contains all the logics necessary to
2009	* process a single work including synchronization against and
2010	* interaction with other workers on the same cpu, queueing and
2011	* flushing. As long as context requirement is met, any worker can
2012	* call this function to process a work.
2013	*
2014	* CONTEXT:
2015	* spin_lock_irq(pool->lock) which is released and regrabbed.
2016	*/
2017	static void process_one_work(struct worker worker, struct work_struct work)
2018	__releases(&pool->lock)
2019	__acquires(&pool->lock)
2020	{
2021	struct pool_workqueue *pwq = get_work_pwq(work);
2022	struct worker_pool *pool = worker->pool;
2023	bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2024	int work_color;
2025	struct worker *collision;
2026	#ifdef CONFIG_LOCKDEP
2027	/*
2028	* It is permissible to free the struct work_struct from
2029	* inside the function that is called from it, this we need to
2030	* take into account for lockdep too. To avoid bogus "held
2031	* lock freed" warnings as well as problems when looking into
2032	* work->lockdep_map, make a copy and use that here.
2033	*/
2034	struct lockdep_map lockdep_map;
2035
2036	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2037	#endif
2038	/* ensure we're on the correct CPU */
2039	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2040	raw_smp_processor_id() != pool->cpu);
2041
2042	/*
2043	* A single work shouldn't be executed concurrently by
2044	* multiple workers on a single cpu. Check whether anyone is
2045	* already processing the work. If so, defer the work to the
2046	* currently executing one.
2047	*/
2048	collision = find_worker_executing_work(pool, work);
2049	if (unlikely(collision)) {
2050	move_linked_works(work, &collision->scheduled, NULL);
2051	return;
2052	}
2053
2054	/* claim and dequeue */
2055	debug_work_deactivate(work);
2056	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2057	worker->current_work = work;
2058	worker->current_func = work->func;
2059	worker->current_pwq = pwq;
2060	work_color = get_work_color(work);
2061
2062	list_del_init(&work->entry);
2063
2064	/*
2065	* CPU intensive works don't participate in concurrency management.
2066	* They're the scheduler's responsibility. This takes @worker out
2067	* of concurrency management and the next code block will chain
2068	* execution of the pending work items.
2069	*/
2070	if (unlikely(cpu_intensive))
2071	worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2072
2073	/*
2074	* Wake up another worker if necessary. The condition is always
2075	* false for normal per-cpu workers since nr_running would always
2076	* be >= 1 at this point. This is used to chain execution of the
2077	* pending work items for WORKER_NOT_RUNNING workers such as the
2078	* UNBOUND and CPU_INTENSIVE ones.
2079	*/
2080	if (need_more_worker(pool))
2081	wake_up_worker(pool);
2082
2083	/*
2084	* Record the last pool and clear PENDING which should be the last
2085	* update to @work. Also, do this inside @pool->lock so that
2086	* PENDING and queued state changes happen together while IRQ is
2087	* disabled.
2088	*/
2089	set_work_pool_and_clear_pending(work, pool->id);
2090
2091	spin_unlock_irq(&pool->lock);
2092
2093	lock_map_acquire_read(&pwq->wq->lockdep_map);
2094	lock_map_acquire(&lockdep_map);
2095	trace_workqueue_execute_start(work);
2096	worker->current_func(work);
2097	/*
2098	* While we must be careful to not use "work" after this, the trace
2099	* point will only record its address.
2100	*/
2101	trace_workqueue_execute_end(work);
2102	lock_map_release(&lockdep_map);
2103	lock_map_release(&pwq->wq->lockdep_map);
2104
2105	if (unlikely(in_atomic() \|\| lockdep_depth(current) > 0)) {
2106	pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2107	" last function: %pf\n",
2108	current->comm, preempt_count(), task_pid_nr(current),
2109	worker->current_func);
2110	debug_show_held_locks(current);
2111	dump_stack();
2112	}
2113
2114	/*
2115	* The following prevents a kworker from hogging CPU on !PREEMPT
2116	* kernels, where a requeueing work item waiting for something to
2117	* happen could deadlock with stop_machine as such work item could
2118	* indefinitely requeue itself while all other CPUs are trapped in
2119	* stop_machine. At the same time, report a quiescent RCU state so
2120	* the same condition doesn't freeze RCU.
2121	*/
2122	cond_resched_rcu_qs();
2123
2124	spin_lock_irq(&pool->lock);
2125
2126	/* clear cpu intensive status */
2127	if (unlikely(cpu_intensive))
2128	worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2129
2130	/* we're done with it, release */
2131	hash_del(&worker->hentry);
2132	worker->current_work = NULL;
2133	worker->current_func = NULL;
2134	worker->current_pwq = NULL;
2135	worker->desc_valid = false;
2136	pwq_dec_nr_in_flight(pwq, work_color);
2137	}
2138
2139	/**
2140	* process_scheduled_works - process scheduled works
2141	* @worker: self
2142	*
2143	* Process all scheduled works. Please note that the scheduled list
2144	* may change while processing a work, so this function repeatedly
2145	* fetches a work from the top and executes it.
2146	*
2147	* CONTEXT:
2148	* spin_lock_irq(pool->lock) which may be released and regrabbed
2149	* multiple times.
2150	*/
2151	static void process_scheduled_works(struct worker *worker)
2152	{
2153	while (!list_empty(&worker->scheduled)) {
2154	struct work_struct *work = list_first_entry(&worker->scheduled,
2155	struct work_struct, entry);
2156	process_one_work(worker, work);
2157	}
2158	}
2159
2160	/**
2161	* worker_thread - the worker thread function
2162	* @__worker: self
2163	*
2164	* The worker thread function. All workers belong to a worker_pool -
2165	* either a per-cpu one or dynamic unbound one. These workers process all
2166	* work items regardless of their specific target workqueue. The only
2167	* exception is work items which belong to workqueues with a rescuer which
2168	* will be explained in rescuer_thread().
2169	*
2170	* Return: 0
2171	*/
2172	static int worker_thread(void *__worker)
2173	{
2174	struct worker *worker = __worker;
2175	struct worker_pool *pool = worker->pool;
2176
2177	/* tell the scheduler that this is a workqueue worker */
2178	worker->task->flags \|= PF_WQ_WORKER;
2179	woke_up:
2180	spin_lock_irq(&pool->lock);
2181
2182	/* am I supposed to die? */
2183	if (unlikely(worker->flags & WORKER_DIE)) {
2184	spin_unlock_irq(&pool->lock);
2185	WARN_ON_ONCE(!list_empty(&worker->entry));
2186	worker->task->flags &= ~PF_WQ_WORKER;
2187
2188	set_task_comm(worker->task, "kworker/dying");
2189	ida_simple_remove(&pool->worker_ida, worker->id);
2190	worker_detach_from_pool(worker, pool);
2191	kfree(worker);
2192	return 0;
2193	}
2194
2195	worker_leave_idle(worker);
2196	recheck:
2197	/* no more worker necessary? */
2198	if (!need_more_worker(pool))
2199	goto sleep;
2200
2201	/* do we need to manage? */
2202	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2203	goto recheck;
2204
2205	/*
2206	* ->scheduled list can only be filled while a worker is
2207	* preparing to process a work or actually processing it.
2208	* Make sure nobody diddled with it while I was sleeping.
2209	*/
2210	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2211
2212	/*
2213	* Finish PREP stage. We're guaranteed to have at least one idle
2214	* worker or that someone else has already assumed the manager
2215	* role. This is where @worker starts participating in concurrency
2216	* management if applicable and concurrency management is restored
2217	* after being rebound. See rebind_workers() for details.
2218	*/
2219	worker_clr_flags(worker, WORKER_PREP \| WORKER_REBOUND);
2220
2221	do {
2222	struct work_struct *work =
2223	list_first_entry(&pool->worklist,
2224	struct work_struct, entry);
2225
2226	pool->watchdog_ts = jiffies;
2227
2228	if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2229	/* optimization path, not strictly necessary */
2230	process_one_work(worker, work);
2231	if (unlikely(!list_empty(&worker->scheduled)))
2232	process_scheduled_works(worker);
2233	} else {
2234	move_linked_works(work, &worker->scheduled, NULL);
2235	process_scheduled_works(worker);
2236	}
2237	} while (keep_working(pool));
2238
2239	worker_set_flags(worker, WORKER_PREP);
2240	sleep:
2241	/*
2242	* pool->lock is held and there's no work to process and no need to
2243	* manage, sleep. Workers are woken up only while holding
2244	* pool->lock or from local cpu, so setting the current state
2245	* before releasing pool->lock is enough to prevent losing any
2246	* event.
2247	*/
2248	worker_enter_idle(worker);
2249	__set_current_state(TASK_INTERRUPTIBLE);
2250	spin_unlock_irq(&pool->lock);
2251	schedule();
2252	goto woke_up;
2253	}
2254
2255	/**
2256	* rescuer_thread - the rescuer thread function
2257	* @__rescuer: self
2258	*
2259	* Workqueue rescuer thread function. There's one rescuer for each
2260	* workqueue which has WQ_MEM_RECLAIM set.
2261	*
2262	* Regular work processing on a pool may block trying to create a new
2263	* worker which uses GFP_KERNEL allocation which has slight chance of
2264	* developing into deadlock if some works currently on the same queue
2265	* need to be processed to satisfy the GFP_KERNEL allocation. This is
2266	* the problem rescuer solves.
2267	*
2268	* When such condition is possible, the pool summons rescuers of all
2269	* workqueues which have works queued on the pool and let them process
2270	* those works so that forward progress can be guaranteed.
2271	*
2272	* This should happen rarely.
2273	*
2274	* Return: 0
2275	*/
2276	static int rescuer_thread(void *__rescuer)
2277	{
2278	struct worker *rescuer = __rescuer;
2279	struct workqueue_struct *wq = rescuer->rescue_wq;
2280	struct list_head *scheduled = &rescuer->scheduled;
2281	bool should_stop;
2282
2283	set_user_nice(current, RESCUER_NICE_LEVEL);
2284
2285	/*
2286	* Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2287	* doesn't participate in concurrency management.
2288	*/
2289	rescuer->task->flags \|= PF_WQ_WORKER;
2290	repeat:
2291	set_current_state(TASK_INTERRUPTIBLE);
2292
2293	/*
2294	* By the time the rescuer is requested to stop, the workqueue
2295	* shouldn't have any work pending, but @wq->maydays may still have
2296	* pwq(s) queued. This can happen by non-rescuer workers consuming
2297	* all the work items before the rescuer got to them. Go through
2298	* @wq->maydays processing before acting on should_stop so that the
2299	* list is always empty on exit.
2300	*/
2301	should_stop = kthread_should_stop();
2302
2303	/* see whether any pwq is asking for help */
2304	spin_lock_irq(&wq_mayday_lock);
2305
2306	while (!list_empty(&wq->maydays)) {
2307	struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2308	struct pool_workqueue, mayday_node);
2309	struct worker_pool *pool = pwq->pool;
2310	struct work_struct work, n;
2311	bool first = true;
2312
2313	__set_current_state(TASK_RUNNING);
2314	list_del_init(&pwq->mayday_node);
2315
2316	spin_unlock_irq(&wq_mayday_lock);
2317
2318	worker_attach_to_pool(rescuer, pool);
2319
2320	spin_lock_irq(&pool->lock);
2321	rescuer->pool = pool;
2322
2323	/*
2324	* Slurp in all works issued via this workqueue and
2325	* process'em.
2326	*/
2327	WARN_ON_ONCE(!list_empty(scheduled));
2328	list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2329	if (get_work_pwq(work) == pwq) {
2330	if (first)
2331	pool->watchdog_ts = jiffies;
2332	move_linked_works(work, scheduled, &n);
2333	}
2334	first = false;
2335	}
2336
2337	if (!list_empty(scheduled)) {
2338	process_scheduled_works(rescuer);
2339
2340	/*
2341	* The above execution of rescued work items could
2342	* have created more to rescue through
2343	* pwq_activate_first_delayed() or chained
2344	* queueing. Let's put @pwq back on mayday list so
2345	* that such back-to-back work items, which may be
2346	* being used to relieve memory pressure, don't
2347	* incur MAYDAY_INTERVAL delay inbetween.
2348	*/
2349	if (need_to_create_worker(pool)) {
2350	spin_lock(&wq_mayday_lock);
2351	get_pwq(pwq);
2352	list_move_tail(&pwq->mayday_node, &wq->maydays);
2353	spin_unlock(&wq_mayday_lock);
2354	}
2355	}
2356
2357	/*
2358	* Put the reference grabbed by send_mayday(). @pool won't
2359	* go away while we're still attached to it.
2360	*/
2361	put_pwq(pwq);
2362
2363	/*
2364	* Leave this pool. If need_more_worker() is %true, notify a
2365	* regular worker; otherwise, we end up with 0 concurrency
2366	* and stalling the execution.
2367	*/
2368	if (need_more_worker(pool))
2369	wake_up_worker(pool);
2370
2371	rescuer->pool = NULL;
2372	spin_unlock_irq(&pool->lock);
2373
2374	worker_detach_from_pool(rescuer, pool);
2375
2376	spin_lock_irq(&wq_mayday_lock);
2377	}
2378
2379	spin_unlock_irq(&wq_mayday_lock);
2380
2381	if (should_stop) {
2382	__set_current_state(TASK_RUNNING);
2383	rescuer->task->flags &= ~PF_WQ_WORKER;
2384	return 0;
2385	}
2386
2387	/* rescuers should never participate in concurrency management */
2388	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2389	schedule();
2390	goto repeat;
2391	}
2392
2393	/**
2394	* check_flush_dependency - check for flush dependency sanity
2395	* @target_wq: workqueue being flushed
2396	* @target_work: work item being flushed (NULL for workqueue flushes)
2397	*
2398	* %current is trying to flush the whole @target_wq or @target_work on it.
2399	* If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2400	* reclaiming memory or running on a workqueue which doesn't have
2401	* %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2402	* a deadlock.
2403	*/
2404	static void check_flush_dependency(struct workqueue_struct *target_wq,
2405	struct work_struct *target_work)
2406	{
2407	work_func_t target_func = target_work ? target_work->func : NULL;
2408	struct worker *worker;
2409
2410	if (target_wq->flags & WQ_MEM_RECLAIM)
2411	return;
2412
2413	worker = current_wq_worker();
2414
2415	WARN_ONCE(current->flags & PF_MEMALLOC,
2416	"workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2417	current->pid, current->comm, target_wq->name, target_func);
2418	WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2419	(WQ_MEM_RECLAIM \| __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2420	"workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2421	worker->current_pwq->wq->name, worker->current_func,
2422	target_wq->name, target_func);
2423	}
2424
2425	struct wq_barrier {
2426	struct work_struct work;
2427	struct completion done;
2428	struct task_struct task; / purely informational */
2429	};
2430
2431	static void wq_barrier_func(struct work_struct *work)
2432	{
2433	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2434	complete(&barr->done);
2435	}
2436
2437	/**
2438	* insert_wq_barrier - insert a barrier work
2439	* @pwq: pwq to insert barrier into
2440	* @barr: wq_barrier to insert
2441	* @target: target work to attach @barr to
2442	* @worker: worker currently executing @target, NULL if @target is not executing
2443	*
2444	* @barr is linked to @target such that @barr is completed only after
2445	* @target finishes execution. Please note that the ordering
2446	* guarantee is observed only with respect to @target and on the local
2447	* cpu.
2448	*
2449	* Currently, a queued barrier can't be canceled. This is because
2450	* try_to_grab_pending() can't determine whether the work to be
2451	* grabbed is at the head of the queue and thus can't clear LINKED
2452	* flag of the previous work while there must be a valid next work
2453	* after a work with LINKED flag set.
2454	*
2455	* Note that when @worker is non-NULL, @target may be modified
2456	* underneath us, so we can't reliably determine pwq from @target.
2457	*
2458	* CONTEXT:
2459	* spin_lock_irq(pool->lock).
2460	*/
2461	static void insert_wq_barrier(struct pool_workqueue *pwq,
2462	struct wq_barrier *barr,
2463	struct work_struct target, struct worker worker)
2464	{
2465	struct list_head *head;
2466	unsigned int linked = 0;
2467
2468	/*
2469	* debugobject calls are safe here even with pool->lock locked
2470	* as we know for sure that this will not trigger any of the
2471	* checks and call back into the fixup functions where we
2472	* might deadlock.
2473	*/
2474	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2475	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2476	init_completion(&barr->done);
2477	barr->task = current;
2478
2479	/*
2480	* If @target is currently being executed, schedule the
2481	* barrier to the worker; otherwise, put it after @target.
2482	*/
2483	if (worker)
2484	head = worker->scheduled.next;
2485	else {
2486	unsigned long *bits = work_data_bits(target);
2487
2488	head = target->entry.next;
2489	/* there can already be other linked works, inherit and set */
2490	linked = *bits & WORK_STRUCT_LINKED;
2491	__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2492	}
2493
2494	debug_work_activate(&barr->work);
2495	insert_work(pwq, &barr->work, head,
2496	work_color_to_flags(WORK_NO_COLOR) \| linked);
2497	}
2498
2499	/**
2500	* flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2501	* @wq: workqueue being flushed
2502	* @flush_color: new flush color, < 0 for no-op
2503	* @work_color: new work color, < 0 for no-op
2504	*
2505	* Prepare pwqs for workqueue flushing.
2506	*
2507	* If @flush_color is non-negative, flush_color on all pwqs should be
2508	* -1. If no pwq has in-flight commands at the specified color, all
2509	* pwq->flush_color's stay at -1 and %false is returned. If any pwq
2510	* has in flight commands, its pwq->flush_color is set to
2511	* @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2512	* wakeup logic is armed and %true is returned.
2513	*
2514	* The caller should have initialized @wq->first_flusher prior to
2515	* calling this function with non-negative @flush_color. If
2516	* @flush_color is negative, no flush color update is done and %false
2517	* is returned.
2518	*
2519	* If @work_color is non-negative, all pwqs should have the same
2520	* work_color which is previous to @work_color and all will be
2521	* advanced to @work_color.
2522	*
2523	* CONTEXT:
2524	* mutex_lock(wq->mutex).
2525	*
2526	* Return:
2527	* %true if @flush_color >= 0 and there's something to flush. %false
2528	* otherwise.
2529	*/
2530	static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2531	int flush_color, int work_color)
2532	{
2533	bool wait = false;
2534	struct pool_workqueue *pwq;
2535
2536	if (flush_color >= 0) {
2537	WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2538	atomic_set(&wq->nr_pwqs_to_flush, 1);
2539	}
2540
2541	for_each_pwq(pwq, wq) {
2542	struct worker_pool *pool = pwq->pool;
2543
2544	spin_lock_irq(&pool->lock);
2545
2546	if (flush_color >= 0) {
2547	WARN_ON_ONCE(pwq->flush_color != -1);
2548
2549	if (pwq->nr_in_flight[flush_color]) {
2550	pwq->flush_color = flush_color;
2551	atomic_inc(&wq->nr_pwqs_to_flush);
2552	wait = true;
2553	}
2554	}
2555
2556	if (work_color >= 0) {
2557	WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2558	pwq->work_color = work_color;
2559	}
2560
2561	spin_unlock_irq(&pool->lock);
2562	}
2563
2564	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2565	complete(&wq->first_flusher->done);
2566
2567	return wait;
2568	}
2569
2570	/**
2571	* flush_workqueue - ensure that any scheduled work has run to completion.
2572	* @wq: workqueue to flush
2573	*
2574	* This function sleeps until all work items which were queued on entry
2575	* have finished execution, but it is not livelocked by new incoming ones.
2576	*/
2577	void flush_workqueue(struct workqueue_struct *wq)
2578	{
2579	struct wq_flusher this_flusher = {
2580	.list = LIST_HEAD_INIT(this_flusher.list),
2581	.flush_color = -1,
2582	.done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2583	};
2584	int next_color;
2585
2586	if (WARN_ON(!wq_online))
2587	return;
2588
2589	lock_map_acquire(&wq->lockdep_map);
2590	lock_map_release(&wq->lockdep_map);
2591
2592	mutex_lock(&wq->mutex);
2593
2594	/*
2595	* Start-to-wait phase
2596	*/
2597	next_color = work_next_color(wq->work_color);
2598
2599	if (next_color != wq->flush_color) {
2600	/*
2601	* Color space is not full. The current work_color
2602	* becomes our flush_color and work_color is advanced
2603	* by one.
2604	*/
2605	WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2606	this_flusher.flush_color = wq->work_color;
2607	wq->work_color = next_color;
2608
2609	if (!wq->first_flusher) {
2610	/* no flush in progress, become the first flusher */
2611	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2612
2613	wq->first_flusher = &this_flusher;
2614
2615	if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2616	wq->work_color)) {
2617	/* nothing to flush, done */
2618	wq->flush_color = next_color;
2619	wq->first_flusher = NULL;
2620	goto out_unlock;
2621	}
2622	} else {
2623	/* wait in queue */
2624	WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2625	list_add_tail(&this_flusher.list, &wq->flusher_queue);
2626	flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2627	}
2628	} else {
2629	/*
2630	* Oops, color space is full, wait on overflow queue.
2631	* The next flush completion will assign us
2632	* flush_color and transfer to flusher_queue.
2633	*/
2634	list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2635	}
2636
2637	check_flush_dependency(wq, NULL);
2638
2639	mutex_unlock(&wq->mutex);
2640
2641	wait_for_completion(&this_flusher.done);
2642
2643	/*
2644	* Wake-up-and-cascade phase
2645	*
2646	* First flushers are responsible for cascading flushes and
2647	* handling overflow. Non-first flushers can simply return.
2648	*/
2649	if (wq->first_flusher != &this_flusher)
2650	return;
2651
2652	mutex_lock(&wq->mutex);
2653
2654	/* we might have raced, check again with mutex held */
2655	if (wq->first_flusher != &this_flusher)
2656	goto out_unlock;
2657
2658	wq->first_flusher = NULL;
2659
2660	WARN_ON_ONCE(!list_empty(&this_flusher.list));
2661	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2662
2663	while (true) {
2664	struct wq_flusher next, tmp;
2665
2666	/* complete all the flushers sharing the current flush color */
2667	list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2668	if (next->flush_color != wq->flush_color)
2669	break;
2670	list_del_init(&next->list);
2671	complete(&next->done);
2672	}
2673
2674	WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2675	wq->flush_color != work_next_color(wq->work_color));
2676
2677	/* this flush_color is finished, advance by one */
2678	wq->flush_color = work_next_color(wq->flush_color);
2679
2680	/* one color has been freed, handle overflow queue */
2681	if (!list_empty(&wq->flusher_overflow)) {
2682	/*
2683	* Assign the same color to all overflowed
2684	* flushers, advance work_color and append to
2685	* flusher_queue. This is the start-to-wait
2686	* phase for these overflowed flushers.
2687	*/
2688	list_for_each_entry(tmp, &wq->flusher_overflow, list)
2689	tmp->flush_color = wq->work_color;
2690
2691	wq->work_color = work_next_color(wq->work_color);
2692
2693	list_splice_tail_init(&wq->flusher_overflow,
2694	&wq->flusher_queue);
2695	flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2696	}
2697
2698	if (list_empty(&wq->flusher_queue)) {
2699	WARN_ON_ONCE(wq->flush_color != wq->work_color);
2700	break;
2701	}
2702
2703	/*
2704	* Need to flush more colors. Make the next flusher
2705	* the new first flusher and arm pwqs.
2706	*/
2707	WARN_ON_ONCE(wq->flush_color == wq->work_color);
2708	WARN_ON_ONCE(wq->flush_color != next->flush_color);
2709
2710	list_del_init(&next->list);
2711	wq->first_flusher = next;
2712
2713	if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2714	break;
2715
2716	/*
2717	* Meh... this color is already done, clear first
2718	* flusher and repeat cascading.
2719	*/
2720	wq->first_flusher = NULL;
2721	}
2722
2723	out_unlock:
2724	mutex_unlock(&wq->mutex);
2725	}
2726	EXPORT_SYMBOL(flush_workqueue);
2727
2728	/**
2729	* drain_workqueue - drain a workqueue
2730	* @wq: workqueue to drain
2731	*
2732	* Wait until the workqueue becomes empty. While draining is in progress,
2733	* only chain queueing is allowed. IOW, only currently pending or running
2734	* work items on @wq can queue further work items on it. @wq is flushed
2735	* repeatedly until it becomes empty. The number of flushing is determined
2736	* by the depth of chaining and should be relatively short. Whine if it
2737	* takes too long.
2738	*/
2739	void drain_workqueue(struct workqueue_struct *wq)
2740	{
2741	unsigned int flush_cnt = 0;
2742	struct pool_workqueue *pwq;
2743
2744	/*
2745	* __queue_work() needs to test whether there are drainers, is much
2746	* hotter than drain_workqueue() and already looks at @wq->flags.
2747	* Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2748	*/
2749	mutex_lock(&wq->mutex);
2750	if (!wq->nr_drainers++)
2751	wq->flags \|= __WQ_DRAINING;
2752	mutex_unlock(&wq->mutex);
2753	reflush:
2754	flush_workqueue(wq);
2755
2756	mutex_lock(&wq->mutex);
2757
2758	for_each_pwq(pwq, wq) {
2759	bool drained;
2760
2761	spin_lock_irq(&pwq->pool->lock);
2762	drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2763	spin_unlock_irq(&pwq->pool->lock);
2764
2765	if (drained)
2766	continue;
2767
2768	if (++flush_cnt == 10 \|\|
2769	(flush_cnt % 100 == 0 && flush_cnt <= 1000))
2770	pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2771	wq->name, flush_cnt);
2772
2773	mutex_unlock(&wq->mutex);
2774	goto reflush;
2775	}
2776
2777	if (!--wq->nr_drainers)
2778	wq->flags &= ~__WQ_DRAINING;
2779	mutex_unlock(&wq->mutex);
2780	}
2781	EXPORT_SYMBOL_GPL(drain_workqueue);
2782
2783	static bool start_flush_work(struct work_struct work, struct wq_barrier barr)
2784	{
2785	struct worker *worker = NULL;
2786	struct worker_pool *pool;
2787	struct pool_workqueue *pwq;
2788
2789	might_sleep();
2790
2791	local_irq_disable();
2792	pool = get_work_pool(work);
2793	if (!pool) {
2794	local_irq_enable();
2795	return false;
2796	}
2797
2798	spin_lock(&pool->lock);
2799	/* see the comment in try_to_grab_pending() with the same code */
2800	pwq = get_work_pwq(work);
2801	if (pwq) {
2802	if (unlikely(pwq->pool != pool))
2803	goto already_gone;
2804	} else {
2805	worker = find_worker_executing_work(pool, work);
2806	if (!worker)
2807	goto already_gone;
2808	pwq = worker->current_pwq;
2809	}
2810
2811	check_flush_dependency(pwq->wq, work);
2812
2813	insert_wq_barrier(pwq, barr, work, worker);
2814	spin_unlock_irq(&pool->lock);
2815
2816	/*
2817	* If @max_active is 1 or rescuer is in use, flushing another work
2818	* item on the same workqueue may lead to deadlock. Make sure the
2819	* flusher is not running on the same workqueue by verifying write
2820	* access.
2821	*/
2822	if (pwq->wq->saved_max_active == 1 \|\| pwq->wq->rescuer)
2823	lock_map_acquire(&pwq->wq->lockdep_map);
2824	else
2825	lock_map_acquire_read(&pwq->wq->lockdep_map);
2826	lock_map_release(&pwq->wq->lockdep_map);
2827
2828	return true;
2829	already_gone:
2830	spin_unlock_irq(&pool->lock);
2831	return false;
2832	}
2833
2834	/**
2835	* flush_work - wait for a work to finish executing the last queueing instance
2836	* @work: the work to flush
2837	*
2838	* Wait until @work has finished execution. @work is guaranteed to be idle
2839	* on return if it hasn't been requeued since flush started.
2840	*
2841	* Return:
2842	* %true if flush_work() waited for the work to finish execution,
2843	* %false if it was already idle.
2844	*/
2845	bool flush_work(struct work_struct *work)
2846	{
2847	struct wq_barrier barr;
2848
2849	if (WARN_ON(!wq_online))
2850	return false;
2851
2852	lock_map_acquire(&work->lockdep_map);
2853	lock_map_release(&work->lockdep_map);
2854
2855	if (start_flush_work(work, &barr)) {
2856	wait_for_completion(&barr.done);
2857	destroy_work_on_stack(&barr.work);
2858	return true;
2859	} else {
2860	return false;
2861	}
2862	}
2863	EXPORT_SYMBOL_GPL(flush_work);
2864
2865	struct cwt_wait {
2866	wait_queue_t wait;
2867	struct work_struct *work;
2868	};
2869
2870	static int cwt_wakefn(wait_queue_t wait, unsigned mode, int sync, void key)
2871	{
2872	struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2873
2874	if (cwait->work != key)
2875	return 0;
2876	return autoremove_wake_function(wait, mode, sync, key);
2877	}
2878
2879	static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2880	{
2881	static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2882	unsigned long flags;
2883	int ret;
2884
2885	do {
2886	ret = try_to_grab_pending(work, is_dwork, &flags);
2887	/*
2888	* If someone else is already canceling, wait for it to
2889	* finish. flush_work() doesn't work for PREEMPT_NONE
2890	* because we may get scheduled between @work's completion
2891	* and the other canceling task resuming and clearing
2892	* CANCELING - flush_work() will return false immediately
2893	* as @work is no longer busy, try_to_grab_pending() will
2894	* return -ENOENT as @work is still being canceled and the
2895	* other canceling task won't be able to clear CANCELING as
2896	* we're hogging the CPU.
2897	*
2898	* Let's wait for completion using a waitqueue. As this
2899	* may lead to the thundering herd problem, use a custom
2900	* wake function which matches @work along with exclusive
2901	* wait and wakeup.
2902	*/
2903	if (unlikely(ret == -ENOENT)) {
2904	struct cwt_wait cwait;
2905
2906	init_wait(&cwait.wait);
2907	cwait.wait.func = cwt_wakefn;
2908	cwait.work = work;
2909
2910	prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2911	TASK_UNINTERRUPTIBLE);
2912	if (work_is_canceling(work))
2913	schedule();
2914	finish_wait(&cancel_waitq, &cwait.wait);
2915	}
2916	} while (unlikely(ret < 0));
2917
2918	/* tell other tasks trying to grab @work to back off */
2919	mark_work_canceling(work);
2920	local_irq_restore(flags);
2921
2922	/*
2923	* This allows canceling during early boot. We know that @work
2924	* isn't executing.
2925	*/
2926	if (wq_online)
2927	flush_work(work);
2928
2929	clear_work_data(work);
2930
2931	/*
2932	* Paired with prepare_to_wait() above so that either
2933	* waitqueue_active() is visible here or !work_is_canceling() is
2934	* visible there.
2935	*/
2936	smp_mb();
2937	if (waitqueue_active(&cancel_waitq))
2938	__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2939
2940	return ret;
2941	}
2942
2943	/**
2944	* cancel_work_sync - cancel a work and wait for it to finish
2945	* @work: the work to cancel
2946	*
2947	* Cancel @work and wait for its execution to finish. This function
2948	* can be used even if the work re-queues itself or migrates to
2949	* another workqueue. On return from this function, @work is
2950	* guaranteed to be not pending or executing on any CPU.
2951	*
2952	* cancel_work_sync(&delayed_work->work) must not be used for
2953	* delayed_work's. Use cancel_delayed_work_sync() instead.
2954	*
2955	* The caller must ensure that the workqueue on which @work was last
2956	* queued can't be destroyed before this function returns.
2957	*
2958	* Return:
2959	* %true if @work was pending, %false otherwise.
2960	*/
2961	bool cancel_work_sync(struct work_struct *work)
2962	{
2963	return __cancel_work_timer(work, false);
2964	}
2965	EXPORT_SYMBOL_GPL(cancel_work_sync);
2966
2967	/**
2968	* flush_delayed_work - wait for a dwork to finish executing the last queueing
2969	* @dwork: the delayed work to flush
2970	*
2971	* Delayed timer is cancelled and the pending work is queued for
2972	* immediate execution. Like flush_work(), this function only
2973	* considers the last queueing instance of @dwork.
2974	*
2975	* Return:
2976	* %true if flush_work() waited for the work to finish execution,
2977	* %false if it was already idle.
2978	*/
2979	bool flush_delayed_work(struct delayed_work *dwork)
2980	{
2981	local_irq_disable();
2982	if (del_timer_sync(&dwork->timer))
2983	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
2984	local_irq_enable();
2985	return flush_work(&dwork->work);
2986	}
2987	EXPORT_SYMBOL(flush_delayed_work);
2988
2989	static bool __cancel_work(struct work_struct *work, bool is_dwork)
2990	{
2991	unsigned long flags;
2992	int ret;
2993
2994	do {
2995	ret = try_to_grab_pending(work, is_dwork, &flags);
2996	} while (unlikely(ret == -EAGAIN));
2997
2998	if (unlikely(ret < 0))
2999	return false;
3000
3001	set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3002	local_irq_restore(flags);
3003	return ret;
3004	}
3005
3006	/*
3007	* See cancel_delayed_work()
3008	*/
3009	bool cancel_work(struct work_struct *work)
3010	{
3011	return __cancel_work(work, false);
3012	}
3013
3014	/**
3015	* cancel_delayed_work - cancel a delayed work
3016	* @dwork: delayed_work to cancel
3017	*
3018	* Kill off a pending delayed_work.
3019	*
3020	* Return: %true if @dwork was pending and canceled; %false if it wasn't
3021	* pending.
3022	*
3023	* Note:
3024	* The work callback function may still be running on return, unless
3025	* it returns %true and the work doesn't re-arm itself. Explicitly flush or
3026	* use cancel_delayed_work_sync() to wait on it.
3027	*
3028	* This function is safe to call from any context including IRQ handler.
3029	*/
3030	bool cancel_delayed_work(struct delayed_work *dwork)
3031	{
3032	return __cancel_work(&dwork->work, true);
3033	}
3034	EXPORT_SYMBOL(cancel_delayed_work);
3035
3036	/**
3037	* cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3038	* @dwork: the delayed work cancel
3039	*
3040	* This is cancel_work_sync() for delayed works.
3041	*
3042	* Return:
3043	* %true if @dwork was pending, %false otherwise.
3044	*/
3045	bool cancel_delayed_work_sync(struct delayed_work *dwork)
3046	{
3047	return __cancel_work_timer(&dwork->work, true);
3048	}
3049	EXPORT_SYMBOL(cancel_delayed_work_sync);
3050
3051	/**
3052	* schedule_on_each_cpu - execute a function synchronously on each online CPU
3053	* @func: the function to call
3054	*
3055	* schedule_on_each_cpu() executes @func on each online CPU using the
3056	* system workqueue and blocks until all CPUs have completed.
3057	* schedule_on_each_cpu() is very slow.
3058	*
3059	* Return:
3060	* 0 on success, -errno on failure.
3061	*/
3062	int schedule_on_each_cpu(work_func_t func)
3063	{
3064	int cpu;
3065	struct work_struct __percpu *works;
3066
3067	works = alloc_percpu(struct work_struct);
3068	if (!works)
3069	return -ENOMEM;
3070
3071	get_online_cpus();
3072
3073	for_each_online_cpu(cpu) {
3074	struct work_struct *work = per_cpu_ptr(works, cpu);
3075
3076	INIT_WORK(work, func);
3077	schedule_work_on(cpu, work);
3078	}
3079
3080	for_each_online_cpu(cpu)
3081	flush_work(per_cpu_ptr(works, cpu));
3082
3083	put_online_cpus();
3084	free_percpu(works);
3085	return 0;
3086	}
3087
3088	/**
3089	* execute_in_process_context - reliably execute the routine with user context
3090	* @fn: the function to execute
3091	* @ew: guaranteed storage for the execute work structure (must
3092	* be available when the work executes)
3093	*
3094	* Executes the function immediately if process context is available,
3095	* otherwise schedules the function for delayed execution.
3096	*
3097	* Return: 0 - function was executed
3098	* 1 - function was scheduled for execution
3099	*/
3100	int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3101	{
3102	if (!in_interrupt()) {
3103	fn(&ew->work);
3104	return 0;
3105	}
3106
3107	INIT_WORK(&ew->work, fn);
3108	schedule_work(&ew->work);
3109
3110	return 1;
3111	}
3112	EXPORT_SYMBOL_GPL(execute_in_process_context);
3113
3114	/**
3115	* free_workqueue_attrs - free a workqueue_attrs
3116	* @attrs: workqueue_attrs to free
3117	*
3118	* Undo alloc_workqueue_attrs().
3119	*/
3120	void free_workqueue_attrs(struct workqueue_attrs *attrs)
3121	{
3122	if (attrs) {
3123	free_cpumask_var(attrs->cpumask);
3124	kfree(attrs);
3125	}
3126	}
3127
3128	/**
3129	* alloc_workqueue_attrs - allocate a workqueue_attrs
3130	* @gfp_mask: allocation mask to use
3131	*
3132	* Allocate a new workqueue_attrs, initialize with default settings and
3133	* return it.
3134	*
3135	* Return: The allocated new workqueue_attr on success. %NULL on failure.
3136	*/
3137	struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3138	{
3139	struct workqueue_attrs *attrs;
3140
3141	attrs = kzalloc(sizeof(*attrs), gfp_mask);
3142	if (!attrs)
3143	goto fail;
3144	if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3145	goto fail;
3146
3147	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3148	return attrs;
3149	fail:
3150	free_workqueue_attrs(attrs);
3151	return NULL;
3152	}
3153
3154	static void copy_workqueue_attrs(struct workqueue_attrs *to,
3155	const struct workqueue_attrs *from)
3156	{
3157	to->nice = from->nice;
3158	cpumask_copy(to->cpumask, from->cpumask);
3159	/*
3160	* Unlike hash and equality test, this function doesn't ignore
3161	* ->no_numa as it is used for both pool and wq attrs. Instead,
3162	* get_unbound_pool() explicitly clears ->no_numa after copying.
3163	*/
3164	to->no_numa = from->no_numa;
3165	}
3166
3167	/* hash value of the content of @attr */
3168	static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3169	{
3170	u32 hash = 0;
3171
3172	hash = jhash_1word(attrs->nice, hash);
3173	hash = jhash(cpumask_bits(attrs->cpumask),
3174	BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3175	return hash;
3176	}
3177
3178	/* content equality test */
3179	static bool wqattrs_equal(const struct workqueue_attrs *a,
3180	const struct workqueue_attrs *b)
3181	{
3182	if (a->nice != b->nice)
3183	return false;
3184	if (!cpumask_equal(a->cpumask, b->cpumask))
3185	return false;
3186	return true;
3187	}
3188
3189	/**
3190	* init_worker_pool - initialize a newly zalloc'd worker_pool
3191	* @pool: worker_pool to initialize
3192	*
3193	* Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3194	*
3195	* Return: 0 on success, -errno on failure. Even on failure, all fields
3196	* inside @pool proper are initialized and put_unbound_pool() can be called
3197	* on @pool safely to release it.
3198	*/
3199	static int init_worker_pool(struct worker_pool *pool)
3200	{
3201	spin_lock_init(&pool->lock);
3202	pool->id = -1;
3203	pool->cpu = -1;
3204	pool->node = NUMA_NO_NODE;
3205	pool->flags \|= POOL_DISASSOCIATED;
3206	pool->watchdog_ts = jiffies;
3207	INIT_LIST_HEAD(&pool->worklist);
3208	INIT_LIST_HEAD(&pool->idle_list);
3209	hash_init(pool->busy_hash);
3210
3211	init_timer_deferrable(&pool->idle_timer);
3212	pool->idle_timer.function = idle_worker_timeout;
3213	pool->idle_timer.data = (unsigned long)pool;
3214
3215	setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3216	(unsigned long)pool);
3217
3218	mutex_init(&pool->manager_arb);
3219	mutex_init(&pool->attach_mutex);
3220	INIT_LIST_HEAD(&pool->workers);
3221
3222	ida_init(&pool->worker_ida);
3223	INIT_HLIST_NODE(&pool->hash_node);
3224	pool->refcnt = 1;
3225
3226	/* shouldn't fail above this point */
3227	pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3228	if (!pool->attrs)
3229	return -ENOMEM;
3230	return 0;
3231	}
3232
3233	static void rcu_free_wq(struct rcu_head *rcu)
3234	{
3235	struct workqueue_struct *wq =
3236	container_of(rcu, struct workqueue_struct, rcu);
3237
3238	if (!(wq->flags & WQ_UNBOUND))
3239	free_percpu(wq->cpu_pwqs);
3240	else
3241	free_workqueue_attrs(wq->unbound_attrs);
3242
3243	kfree(wq->rescuer);
3244	kfree(wq);
3245	}
3246
3247	static void rcu_free_pool(struct rcu_head *rcu)
3248	{
3249	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3250
3251	ida_destroy(&pool->worker_ida);
3252	free_workqueue_attrs(pool->attrs);
3253	kfree(pool);
3254	}
3255
3256	/**
3257	* put_unbound_pool - put a worker_pool
3258	* @pool: worker_pool to put
3259	*
3260	* Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3261	* safe manner. get_unbound_pool() calls this function on its failure path
3262	* and this function should be able to release pools which went through,
3263	* successfully or not, init_worker_pool().
3264	*
3265	* Should be called with wq_pool_mutex held.
3266	*/
3267	static void put_unbound_pool(struct worker_pool *pool)
3268	{
3269	DECLARE_COMPLETION_ONSTACK(detach_completion);
3270	struct worker *worker;
3271
3272	lockdep_assert_held(&wq_pool_mutex);
3273
3274	if (--pool->refcnt)
3275	return;
3276
3277	/* sanity checks */
3278	if (WARN_ON(!(pool->cpu < 0)) \|\|
3279	WARN_ON(!list_empty(&pool->worklist)))
3280	return;
3281
3282	/* release id and unhash */
3283	if (pool->id >= 0)
3284	idr_remove(&worker_pool_idr, pool->id);
3285	hash_del(&pool->hash_node);
3286
3287	/*
3288	* Become the manager and destroy all workers. Grabbing
3289	* manager_arb prevents @pool's workers from blocking on
3290	* attach_mutex.
3291	*/
3292	mutex_lock(&pool->manager_arb);
3293
3294	spin_lock_irq(&pool->lock);
3295	while ((worker = first_idle_worker(pool)))
3296	destroy_worker(worker);
3297	WARN_ON(pool->nr_workers \|\| pool->nr_idle);
3298	spin_unlock_irq(&pool->lock);
3299
3300	mutex_lock(&pool->attach_mutex);
3301	if (!list_empty(&pool->workers))
3302	pool->detach_completion = &detach_completion;
3303	mutex_unlock(&pool->attach_mutex);
3304
3305	if (pool->detach_completion)
3306	wait_for_completion(pool->detach_completion);
3307
3308	mutex_unlock(&pool->manager_arb);
3309
3310	/* shut down the timers */
3311	del_timer_sync(&pool->idle_timer);
3312	del_timer_sync(&pool->mayday_timer);
3313
3314	/* sched-RCU protected to allow dereferences from get_work_pool() */
3315	call_rcu_sched(&pool->rcu, rcu_free_pool);
3316	}
3317
3318	/**
3319	* get_unbound_pool - get a worker_pool with the specified attributes
3320	* @attrs: the attributes of the worker_pool to get
3321	*
3322	* Obtain a worker_pool which has the same attributes as @attrs, bump the
3323	* reference count and return it. If there already is a matching
3324	* worker_pool, it will be used; otherwise, this function attempts to
3325	* create a new one.
3326	*
3327	* Should be called with wq_pool_mutex held.
3328	*
3329	* Return: On success, a worker_pool with the same attributes as @attrs.
3330	* On failure, %NULL.
3331	*/
3332	static struct worker_pool get_unbound_pool(const struct workqueue_attrs attrs)
3333	{
3334	u32 hash = wqattrs_hash(attrs);
3335	struct worker_pool *pool;
3336	int node;
3337	int target_node = NUMA_NO_NODE;
3338
3339	lockdep_assert_held(&wq_pool_mutex);
3340
3341	/* do we already have a matching pool? */
3342	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3343	if (wqattrs_equal(pool->attrs, attrs)) {
3344	pool->refcnt++;
3345	return pool;
3346	}
3347	}
3348
3349	/* if cpumask is contained inside a NUMA node, we belong to that node */
3350	if (wq_numa_enabled) {
3351	for_each_node(node) {
3352	if (cpumask_subset(attrs->cpumask,
3353	wq_numa_possible_cpumask[node])) {
3354	target_node = node;
3355	break;
3356	}
3357	}
3358	}
3359
3360	/* nope, create a new one */
3361	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3362	if (!pool \|\| init_worker_pool(pool) < 0)
3363	goto fail;
3364
3365	lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3366	copy_workqueue_attrs(pool->attrs, attrs);
3367	pool->node = target_node;
3368
3369	/*
3370	* no_numa isn't a worker_pool attribute, always clear it. See
3371	* 'struct workqueue_attrs' comments for detail.
3372	*/
3373	pool->attrs->no_numa = false;
3374
3375	if (worker_pool_assign_id(pool) < 0)
3376	goto fail;
3377
3378	/* create and start the initial worker */
3379	if (wq_online && !create_worker(pool))
3380	goto fail;
3381
3382	/* install */
3383	hash_add(unbound_pool_hash, &pool->hash_node, hash);
3384
3385	return pool;
3386	fail:
3387	if (pool)
3388	put_unbound_pool(pool);
3389	return NULL;
3390	}
3391
3392	static void rcu_free_pwq(struct rcu_head *rcu)
3393	{
3394	kmem_cache_free(pwq_cache,
3395	container_of(rcu, struct pool_workqueue, rcu));
3396	}
3397
3398	/*
3399	* Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3400	* and needs to be destroyed.
3401	*/
3402	static void pwq_unbound_release_workfn(struct work_struct *work)
3403	{
3404	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3405	unbound_release_work);
3406	struct workqueue_struct *wq = pwq->wq;
3407	struct worker_pool *pool = pwq->pool;
3408	bool is_last;
3409
3410	if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3411	return;
3412
3413	mutex_lock(&wq->mutex);
3414	list_del_rcu(&pwq->pwqs_node);
3415	is_last = list_empty(&wq->pwqs);
3416	mutex_unlock(&wq->mutex);
3417
3418	mutex_lock(&wq_pool_mutex);
3419	put_unbound_pool(pool);
3420	mutex_unlock(&wq_pool_mutex);
3421
3422	call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3423
3424	/*
3425	* If we're the last pwq going away, @wq is already dead and no one
3426	* is gonna access it anymore. Schedule RCU free.
3427	*/
3428	if (is_last)
3429	call_rcu_sched(&wq->rcu, rcu_free_wq);
3430	}
3431
3432	/**
3433	* pwq_adjust_max_active - update a pwq's max_active to the current setting
3434	* @pwq: target pool_workqueue
3435	*
3436	* If @pwq isn't freezing, set @pwq->max_active to the associated
3437	* workqueue's saved_max_active and activate delayed work items
3438	* accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3439	*/
3440	static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3441	{
3442	struct workqueue_struct *wq = pwq->wq;
3443	bool freezable = wq->flags & WQ_FREEZABLE;
3444	unsigned long flags;
3445
3446	/* for @wq->saved_max_active */
3447	lockdep_assert_held(&wq->mutex);
3448
3449	/* fast exit for non-freezable wqs */
3450	if (!freezable && pwq->max_active == wq->saved_max_active)
3451	return;
3452
3453	/* this function can be called during early boot w/ irq disabled */
3454	spin_lock_irqsave(&pwq->pool->lock, flags);
3455
3456	/*
3457	* During [un]freezing, the caller is responsible for ensuring that
3458	* this function is called at least once after @workqueue_freezing
3459	* is updated and visible.
3460	*/
3461	if (!freezable \|\| !workqueue_freezing) {
3462	pwq->max_active = wq->saved_max_active;
3463
3464	while (!list_empty(&pwq->delayed_works) &&
3465	pwq->nr_active < pwq->max_active)
3466	pwq_activate_first_delayed(pwq);
3467
3468	/*
3469	* Need to kick a worker after thawed or an unbound wq's
3470	* max_active is bumped. It's a slow path. Do it always.
3471	*/
3472	wake_up_worker(pwq->pool);
3473	} else {
3474	pwq->max_active = 0;
3475	}
3476
3477	spin_unlock_irqrestore(&pwq->pool->lock, flags);
3478	}
3479
3480	/* initialize newly alloced @pwq which is associated with @wq and @pool */
3481	static void init_pwq(struct pool_workqueue pwq, struct workqueue_struct wq,
3482	struct worker_pool *pool)
3483	{
3484	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3485
3486	memset(pwq, 0, sizeof(*pwq));
3487
3488	pwq->pool = pool;
3489	pwq->wq = wq;
3490	pwq->flush_color = -1;
3491	pwq->refcnt = 1;
3492	INIT_LIST_HEAD(&pwq->delayed_works);
3493	INIT_LIST_HEAD(&pwq->pwqs_node);
3494	INIT_LIST_HEAD(&pwq->mayday_node);
3495	INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3496	}
3497
3498	/* sync @pwq with the current state of its associated wq and link it */
3499	static void link_pwq(struct pool_workqueue *pwq)
3500	{
3501	struct workqueue_struct *wq = pwq->wq;
3502
3503	lockdep_assert_held(&wq->mutex);
3504
3505	/* may be called multiple times, ignore if already linked */
3506	if (!list_empty(&pwq->pwqs_node))
3507	return;
3508
3509	/* set the matching work_color */
3510	pwq->work_color = wq->work_color;
3511
3512	/* sync max_active to the current setting */
3513	pwq_adjust_max_active(pwq);
3514
3515	/* link in @pwq */
3516	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3517	}
3518
3519	/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3520	static struct pool_workqueue alloc_unbound_pwq(struct workqueue_struct wq,
3521	const struct workqueue_attrs *attrs)
3522	{
3523	struct worker_pool *pool;
3524	struct pool_workqueue *pwq;
3525
3526	lockdep_assert_held(&wq_pool_mutex);
3527
3528	pool = get_unbound_pool(attrs);
3529	if (!pool)
3530	return NULL;
3531
3532	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3533	if (!pwq) {
3534	put_unbound_pool(pool);
3535	return NULL;
3536	}
3537
3538	init_pwq(pwq, wq, pool);
3539	return pwq;
3540	}
3541
3542	/**
3543	* wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3544	* @attrs: the wq_attrs of the default pwq of the target workqueue
3545	* @node: the target NUMA node
3546	* @cpu_going_down: if >= 0, the CPU to consider as offline
3547	* @cpumask: outarg, the resulting cpumask
3548	*
3549	* Calculate the cpumask a workqueue with @attrs should use on @node. If
3550	* @cpu_going_down is >= 0, that cpu is considered offline during
3551	* calculation. The result is stored in @cpumask.
3552	*
3553	* If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3554	* enabled and @node has online CPUs requested by @attrs, the returned
3555	* cpumask is the intersection of the possible CPUs of @node and
3556	* @attrs->cpumask.
3557	*
3558	* The caller is responsible for ensuring that the cpumask of @node stays
3559	* stable.
3560	*
3561	* Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3562	* %false if equal.
3563	*/
3564	static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3565	int cpu_going_down, cpumask_t *cpumask)
3566	{
3567	if (!wq_numa_enabled \|\| attrs->no_numa)
3568	goto use_dfl;
3569
3570	/* does @node have any online CPUs @attrs wants? */
3571	cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3572	if (cpu_going_down >= 0)
3573	cpumask_clear_cpu(cpu_going_down, cpumask);
3574
3575	if (cpumask_empty(cpumask))
3576	goto use_dfl;
3577
3578	/* yeap, return possible CPUs in @node that @attrs wants */
3579	cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3580	return !cpumask_equal(cpumask, attrs->cpumask);
3581
3582	use_dfl:
3583	cpumask_copy(cpumask, attrs->cpumask);
3584	return false;
3585	}
3586
3587	/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3588	static struct pool_workqueue numa_pwq_tbl_install(struct workqueue_struct wq,
3589	int node,
3590	struct pool_workqueue *pwq)
3591	{
3592	struct pool_workqueue *old_pwq;
3593
3594	lockdep_assert_held(&wq_pool_mutex);
3595	lockdep_assert_held(&wq->mutex);
3596
3597	/* link_pwq() can handle duplicate calls */
3598	link_pwq(pwq);
3599
3600	old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3601	rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3602	return old_pwq;
3603	}
3604
3605	/* context to store the prepared attrs & pwqs before applying */
3606	struct apply_wqattrs_ctx {
3607	struct workqueue_struct wq; / target workqueue */
3608	struct workqueue_attrs attrs; / attrs to apply */
3609	struct list_head list; /* queued for batching commit */
3610	struct pool_workqueue *dfl_pwq;
3611	struct pool_workqueue *pwq_tbl[];
3612	};
3613
3614	/* free the resources after success or abort */
3615	static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3616	{
3617	if (ctx) {
3618	int node;
3619
3620	for_each_node(node)
3621	put_pwq_unlocked(ctx->pwq_tbl[node]);
3622	put_pwq_unlocked(ctx->dfl_pwq);
3623
3624	free_workqueue_attrs(ctx->attrs);
3625
3626	kfree(ctx);
3627	}
3628	}
3629
3630	/* allocate the attrs and pwqs for later installation */
3631	static struct apply_wqattrs_ctx *
3632	apply_wqattrs_prepare(struct workqueue_struct *wq,
3633	const struct workqueue_attrs *attrs)
3634	{
3635	struct apply_wqattrs_ctx *ctx;
3636	struct workqueue_attrs new_attrs, tmp_attrs;
3637	int node;
3638
3639	lockdep_assert_held(&wq_pool_mutex);
3640
3641	ctx = kzalloc(sizeof(ctx) + nr_node_ids sizeof(ctx->pwq_tbl[0]),
3642	GFP_KERNEL);
3643
3644	new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3645	tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3646	if (!ctx \|\| !new_attrs \|\| !tmp_attrs)
3647	goto out_free;
3648
3649	/*
3650	* Calculate the attrs of the default pwq.
3651	* If the user configured cpumask doesn't overlap with the
3652	* wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3653	*/
3654	copy_workqueue_attrs(new_attrs, attrs);
3655	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3656	if (unlikely(cpumask_empty(new_attrs->cpumask)))
3657	cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3658
3659	/*
3660	* We may create multiple pwqs with differing cpumasks. Make a
3661	* copy of @new_attrs which will be modified and used to obtain
3662	* pools.
3663	*/
3664	copy_workqueue_attrs(tmp_attrs, new_attrs);
3665
3666	/*
3667	* If something goes wrong during CPU up/down, we'll fall back to
3668	* the default pwq covering whole @attrs->cpumask. Always create
3669	* it even if we don't use it immediately.
3670	*/
3671	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3672	if (!ctx->dfl_pwq)
3673	goto out_free;
3674
3675	for_each_node(node) {
3676	if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3677	ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3678	if (!ctx->pwq_tbl[node])
3679	goto out_free;
3680	} else {
3681	ctx->dfl_pwq->refcnt++;
3682	ctx->pwq_tbl[node] = ctx->dfl_pwq;
3683	}
3684	}
3685
3686	/* save the user configured attrs and sanitize it. */
3687	copy_workqueue_attrs(new_attrs, attrs);
3688	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3689	ctx->attrs = new_attrs;
3690
3691	ctx->wq = wq;
3692	free_workqueue_attrs(tmp_attrs);
3693	return ctx;
3694
3695	out_free:
3696	free_workqueue_attrs(tmp_attrs);
3697	free_workqueue_attrs(new_attrs);
3698	apply_wqattrs_cleanup(ctx);
3699	return NULL;
3700	}
3701
3702	/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3703	static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3704	{
3705	int node;
3706
3707	/* all pwqs have been created successfully, let's install'em */
3708	mutex_lock(&ctx->wq->mutex);
3709
3710	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3711
3712	/* save the previous pwq and install the new one */
3713	for_each_node(node)
3714	ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3715	ctx->pwq_tbl[node]);
3716
3717	/* @dfl_pwq might not have been used, ensure it's linked */
3718	link_pwq(ctx->dfl_pwq);
3719	swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3720
3721	mutex_unlock(&ctx->wq->mutex);
3722	}
3723
3724	static void apply_wqattrs_lock(void)
3725	{
3726	/* CPUs should stay stable across pwq creations and installations */
3727	get_online_cpus();
3728	mutex_lock(&wq_pool_mutex);
3729	}
3730
3731	static void apply_wqattrs_unlock(void)
3732	{
3733	mutex_unlock(&wq_pool_mutex);
3734	put_online_cpus();
3735	}
3736
3737	static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3738	const struct workqueue_attrs *attrs)
3739	{
3740	struct apply_wqattrs_ctx *ctx;
3741
3742	/* only unbound workqueues can change attributes */
3743	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3744	return -EINVAL;
3745
3746	/* creating multiple pwqs breaks ordering guarantee */
3747	if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3748	return -EINVAL;
3749
3750	ctx = apply_wqattrs_prepare(wq, attrs);
3751	if (!ctx)
3752	return -ENOMEM;
3753
3754	/* the ctx has been prepared successfully, let's commit it */
3755	apply_wqattrs_commit(ctx);
3756	apply_wqattrs_cleanup(ctx);
3757
3758	return 0;
3759	}
3760
3761	/**
3762	* apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3763	* @wq: the target workqueue
3764	* @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3765	*
3766	* Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3767	* machines, this function maps a separate pwq to each NUMA node with
3768	* possibles CPUs in @attrs->cpumask so that work items are affine to the
3769	* NUMA node it was issued on. Older pwqs are released as in-flight work
3770	* items finish. Note that a work item which repeatedly requeues itself
3771	* back-to-back will stay on its current pwq.
3772	*
3773	* Performs GFP_KERNEL allocations.
3774	*
3775	* Return: 0 on success and -errno on failure.
3776	*/
3777	int apply_workqueue_attrs(struct workqueue_struct *wq,
3778	const struct workqueue_attrs *attrs)
3779	{
3780	int ret;
3781
3782	apply_wqattrs_lock();
3783	ret = apply_workqueue_attrs_locked(wq, attrs);
3784	apply_wqattrs_unlock();
3785
3786	return ret;
3787	}
3788
3789	/**
3790	* wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3791	* @wq: the target workqueue
3792	* @cpu: the CPU coming up or going down
3793	* @online: whether @cpu is coming up or going down
3794	*
3795	* This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3796	* %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3797	* @wq accordingly.
3798	*
3799	* If NUMA affinity can't be adjusted due to memory allocation failure, it
3800	* falls back to @wq->dfl_pwq which may not be optimal but is always
3801	* correct.
3802	*
3803	* Note that when the last allowed CPU of a NUMA node goes offline for a
3804	* workqueue with a cpumask spanning multiple nodes, the workers which were
3805	* already executing the work items for the workqueue will lose their CPU
3806	* affinity and may execute on any CPU. This is similar to how per-cpu
3807	* workqueues behave on CPU_DOWN. If a workqueue user wants strict
3808	* affinity, it's the user's responsibility to flush the work item from
3809	* CPU_DOWN_PREPARE.
3810	*/
3811	static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3812	bool online)
3813	{
3814	int node = cpu_to_node(cpu);
3815	int cpu_off = online ? -1 : cpu;
3816	struct pool_workqueue old_pwq = NULL, pwq;
3817	struct workqueue_attrs *target_attrs;
3818	cpumask_t *cpumask;
3819
3820	lockdep_assert_held(&wq_pool_mutex);
3821
3822	if (!wq_numa_enabled \|\| !(wq->flags & WQ_UNBOUND) \|\|
3823	wq->unbound_attrs->no_numa)
3824	return;
3825
3826	/*
3827	* We don't wanna alloc/free wq_attrs for each wq for each CPU.
3828	* Let's use a preallocated one. The following buf is protected by
3829	* CPU hotplug exclusion.
3830	*/
3831	target_attrs = wq_update_unbound_numa_attrs_buf;
3832	cpumask = target_attrs->cpumask;
3833
3834	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3835	pwq = unbound_pwq_by_node(wq, node);
3836
3837	/*
3838	* Let's determine what needs to be done. If the target cpumask is
3839	* different from the default pwq's, we need to compare it to @pwq's
3840	* and create a new one if they don't match. If the target cpumask
3841	* equals the default pwq's, the default pwq should be used.
3842	*/
3843	if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3844	if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3845	return;
3846	} else {
3847	goto use_dfl_pwq;
3848	}
3849
3850	/* create a new pwq */
3851	pwq = alloc_unbound_pwq(wq, target_attrs);
3852	if (!pwq) {
3853	pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3854	wq->name);
3855	goto use_dfl_pwq;
3856	}
3857
3858	/* Install the new pwq. */
3859	mutex_lock(&wq->mutex);
3860	old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3861	goto out_unlock;
3862
3863	use_dfl_pwq:
3864	mutex_lock(&wq->mutex);
3865	spin_lock_irq(&wq->dfl_pwq->pool->lock);
3866	get_pwq(wq->dfl_pwq);
3867	spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3868	old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3869	out_unlock:
3870	mutex_unlock(&wq->mutex);
3871	put_pwq_unlocked(old_pwq);
3872	}
3873
3874	static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3875	{
3876	bool highpri = wq->flags & WQ_HIGHPRI;
3877	int cpu, ret;
3878
3879	if (!(wq->flags & WQ_UNBOUND)) {
3880	wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3881	if (!wq->cpu_pwqs)
3882	return -ENOMEM;
3883
3884	for_each_possible_cpu(cpu) {
3885	struct pool_workqueue *pwq =
3886	per_cpu_ptr(wq->cpu_pwqs, cpu);
3887	struct worker_pool *cpu_pools =
3888	per_cpu(cpu_worker_pools, cpu);
3889
3890	init_pwq(pwq, wq, &cpu_pools[highpri]);
3891
3892	mutex_lock(&wq->mutex);
3893	link_pwq(pwq);
3894	mutex_unlock(&wq->mutex);
3895	}
3896	return 0;
3897	} else if (wq->flags & __WQ_ORDERED) {
3898	ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3899	/* there should only be single pwq for ordering guarantee */
3900	WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node \|\|
3901	wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3902	"ordering guarantee broken for workqueue %s\n", wq->name);
3903	return ret;
3904	} else {
3905	return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3906	}
3907	}
3908
3909	static int wq_clamp_max_active(int max_active, unsigned int flags,
3910	const char *name)
3911	{
3912	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3913
3914	if (max_active < 1 \|\| max_active > lim)
3915	pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3916	max_active, name, 1, lim);
3917
3918	return clamp_val(max_active, 1, lim);
3919	}
3920
3921	struct workqueue_struct __alloc_workqueue_key(const char fmt,
3922	unsigned int flags,
3923	int max_active,
3924	struct lock_class_key *key,
3925	const char *lock_name, ...)
3926	{
3927	size_t tbl_size = 0;
3928	va_list args;
3929	struct workqueue_struct *wq;
3930	struct pool_workqueue *pwq;
3931
3932	/* see the comment above the definition of WQ_POWER_EFFICIENT */
3933	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3934	flags \|= WQ_UNBOUND;
3935
3936	/* allocate wq and format name */
3937	if (flags & WQ_UNBOUND)
3938	tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3939
3940	wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3941	if (!wq)
3942	return NULL;
3943
3944	if (flags & WQ_UNBOUND) {
3945	wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3946	if (!wq->unbound_attrs)
3947	goto err_free_wq;
3948	}
3949
3950	va_start(args, lock_name);
3951	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3952	va_end(args);
3953
3954	max_active = max_active ?: WQ_DFL_ACTIVE;
3955	max_active = wq_clamp_max_active(max_active, flags, wq->name);
3956
3957	/* init wq */
3958	wq->flags = flags;
3959	wq->saved_max_active = max_active;
3960	mutex_init(&wq->mutex);
3961	atomic_set(&wq->nr_pwqs_to_flush, 0);
3962	INIT_LIST_HEAD(&wq->pwqs);
3963	INIT_LIST_HEAD(&wq->flusher_queue);
3964	INIT_LIST_HEAD(&wq->flusher_overflow);
3965	INIT_LIST_HEAD(&wq->maydays);
3966
3967	lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3968	INIT_LIST_HEAD(&wq->list);
3969
3970	if (alloc_and_link_pwqs(wq) < 0)
3971	goto err_free_wq;
3972
3973	/*
3974	* Workqueues which may be used during memory reclaim should
3975	* have a rescuer to guarantee forward progress.
3976	*/
3977	if (flags & WQ_MEM_RECLAIM) {
3978	struct worker *rescuer;
3979
3980	rescuer = alloc_worker(NUMA_NO_NODE);
3981	if (!rescuer)
3982	goto err_destroy;
3983
3984	rescuer->rescue_wq = wq;
3985	rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3986	wq->name);
3987	if (IS_ERR(rescuer->task)) {
3988	kfree(rescuer);
3989	goto err_destroy;
3990	}
3991
3992	wq->rescuer = rescuer;
3993	kthread_bind_mask(rescuer->task, cpu_possible_mask);
3994	wake_up_process(rescuer->task);
3995	}
3996
3997	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3998	goto err_destroy;
3999
4000	/*
4001	* wq_pool_mutex protects global freeze state and workqueues list.
4002	* Grab it, adjust max_active and add the new @wq to workqueues
4003	* list.
4004	*/
4005	mutex_lock(&wq_pool_mutex);
4006
4007	mutex_lock(&wq->mutex);
4008	for_each_pwq(pwq, wq)
4009	pwq_adjust_max_active(pwq);
4010	mutex_unlock(&wq->mutex);
4011
4012	list_add_tail_rcu(&wq->list, &workqueues);
4013
4014	mutex_unlock(&wq_pool_mutex);
4015
4016	return wq;
4017
4018	err_free_wq:
4019	free_workqueue_attrs(wq->unbound_attrs);
4020	kfree(wq);
4021	return NULL;
4022	err_destroy:
4023	destroy_workqueue(wq);
4024	return NULL;
4025	}
4026	EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4027
4028	/**
4029	* destroy_workqueue - safely terminate a workqueue
4030	* @wq: target workqueue
4031	*
4032	* Safely destroy a workqueue. All work currently pending will be done first.
4033	*/
4034	void destroy_workqueue(struct workqueue_struct *wq)
4035	{
4036	struct pool_workqueue *pwq;
4037	int node;
4038
4039	/* drain it before proceeding with destruction */
4040	drain_workqueue(wq);
4041
4042	/* sanity checks */
4043	mutex_lock(&wq->mutex);
4044	for_each_pwq(pwq, wq) {
4045	int i;
4046
4047	for (i = 0; i < WORK_NR_COLORS; i++) {
4048	if (WARN_ON(pwq->nr_in_flight[i])) {
4049	mutex_unlock(&wq->mutex);
4050	show_workqueue_state();
4051	return;
4052	}
4053	}
4054
4055	if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) \|\|
4056	WARN_ON(pwq->nr_active) \|\|
4057	WARN_ON(!list_empty(&pwq->delayed_works))) {
4058	mutex_unlock(&wq->mutex);
4059	show_workqueue_state();
4060	return;
4061	}
4062	}
4063	mutex_unlock(&wq->mutex);
4064
4065	/*
4066	* wq list is used to freeze wq, remove from list after
4067	* flushing is complete in case freeze races us.
4068	*/
4069	mutex_lock(&wq_pool_mutex);
4070	list_del_rcu(&wq->list);
4071	mutex_unlock(&wq_pool_mutex);
4072
4073	workqueue_sysfs_unregister(wq);
4074
4075	if (wq->rescuer)
4076	kthread_stop(wq->rescuer->task);
4077
4078	if (!(wq->flags & WQ_UNBOUND)) {
4079	/*
4080	* The base ref is never dropped on per-cpu pwqs. Directly
4081	* schedule RCU free.
4082	*/
4083	call_rcu_sched(&wq->rcu, rcu_free_wq);
4084	} else {
4085	/*
4086	* We're the sole accessor of @wq at this point. Directly
4087	* access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4088	* @wq will be freed when the last pwq is released.
4089	*/
4090	for_each_node(node) {
4091	pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4092	RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4093	put_pwq_unlocked(pwq);
4094	}
4095
4096	/*
4097	* Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4098	* put. Don't access it afterwards.
4099	*/
4100	pwq = wq->dfl_pwq;
4101	wq->dfl_pwq = NULL;
4102	put_pwq_unlocked(pwq);
4103	}
4104	}
4105	EXPORT_SYMBOL_GPL(destroy_workqueue);
4106
4107	/**
4108	* workqueue_set_max_active - adjust max_active of a workqueue
4109	* @wq: target workqueue
4110	* @max_active: new max_active value.
4111	*
4112	* Set max_active of @wq to @max_active.
4113	*
4114	* CONTEXT:
4115	* Don't call from IRQ context.
4116	*/
4117	void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4118	{
4119	struct pool_workqueue *pwq;
4120
4121	/* disallow meddling with max_active for ordered workqueues */
4122	if (WARN_ON(wq->flags & __WQ_ORDERED))
4123	return;
4124
4125	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4126
4127	mutex_lock(&wq->mutex);
4128
4129	wq->saved_max_active = max_active;
4130
4131	for_each_pwq(pwq, wq)
4132	pwq_adjust_max_active(pwq);
4133
4134	mutex_unlock(&wq->mutex);
4135	}
4136	EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4137
4138	/**
4139	* current_is_workqueue_rescuer - is %current workqueue rescuer?
4140	*
4141	* Determine whether %current is a workqueue rescuer. Can be used from
4142	* work functions to determine whether it's being run off the rescuer task.
4143	*
4144	* Return: %true if %current is a workqueue rescuer. %false otherwise.
4145	*/
4146	bool current_is_workqueue_rescuer(void)
4147	{
4148	struct worker *worker = current_wq_worker();
4149
4150	return worker && worker->rescue_wq;
4151	}
4152
4153	/**
4154	* workqueue_congested - test whether a workqueue is congested
4155	* @cpu: CPU in question
4156	* @wq: target workqueue
4157	*
4158	* Test whether @wq's cpu workqueue for @cpu is congested. There is
4159	* no synchronization around this function and the test result is
4160	* unreliable and only useful as advisory hints or for debugging.
4161	*
4162	* If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4163	* Note that both per-cpu and unbound workqueues may be associated with
4164	* multiple pool_workqueues which have separate congested states. A
4165	* workqueue being congested on one CPU doesn't mean the workqueue is also
4166	* contested on other CPUs / NUMA nodes.
4167	*
4168	* Return:
4169	* %true if congested, %false otherwise.
4170	*/
4171	bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4172	{
4173	struct pool_workqueue *pwq;
4174	bool ret;
4175
4176	rcu_read_lock_sched();
4177
4178	if (cpu == WORK_CPU_UNBOUND)
4179	cpu = smp_processor_id();
4180
4181	if (!(wq->flags & WQ_UNBOUND))
4182	pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4183	else
4184	pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4185
4186	ret = !list_empty(&pwq->delayed_works);
4187	rcu_read_unlock_sched();
4188
4189	return ret;
4190	}
4191	EXPORT_SYMBOL_GPL(workqueue_congested);
4192
4193	/**
4194	* work_busy - test whether a work is currently pending or running
4195	* @work: the work to be tested
4196	*
4197	* Test whether @work is currently pending or running. There is no
4198	* synchronization around this function and the test result is
4199	* unreliable and only useful as advisory hints or for debugging.
4200	*
4201	* Return:
4202	* OR'd bitmask of WORK_BUSY_* bits.
4203	*/
4204	unsigned int work_busy(struct work_struct *work)
4205	{
4206	struct worker_pool *pool;
4207	unsigned long flags;
4208	unsigned int ret = 0;
4209
4210	if (work_pending(work))
4211	ret \|= WORK_BUSY_PENDING;
4212
4213	local_irq_save(flags);
4214	pool = get_work_pool(work);
4215	if (pool) {
4216	spin_lock(&pool->lock);
4217	if (find_worker_executing_work(pool, work))
4218	ret \|= WORK_BUSY_RUNNING;
4219	spin_unlock(&pool->lock);
4220	}
4221	local_irq_restore(flags);
4222
4223	return ret;
4224	}
4225	EXPORT_SYMBOL_GPL(work_busy);
4226
4227	/**
4228	* set_worker_desc - set description for the current work item
4229	* @fmt: printf-style format string
4230	* @...: arguments for the format string
4231	*
4232	* This function can be called by a running work function to describe what
4233	* the work item is about. If the worker task gets dumped, this
4234	* information will be printed out together to help debugging. The
4235	* description can be at most WORKER_DESC_LEN including the trailing '\0'.
4236	*/
4237	void set_worker_desc(const char *fmt, ...)
4238	{
4239	struct worker *worker = current_wq_worker();
4240	va_list args;
4241
4242	if (worker) {
4243	va_start(args, fmt);
4244	vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4245	va_end(args);
4246	worker->desc_valid = true;
4247	}
4248	}
4249
4250	/**
4251	* print_worker_info - print out worker information and description
4252	* @log_lvl: the log level to use when printing
4253	* @task: target task
4254	*
4255	* If @task is a worker and currently executing a work item, print out the
4256	* name of the workqueue being serviced and worker description set with
4257	* set_worker_desc() by the currently executing work item.
4258	*
4259	* This function can be safely called on any task as long as the
4260	* task_struct itself is accessible. While safe, this function isn't
4261	* synchronized and may print out mixups or garbages of limited length.
4262	*/
4263	void print_worker_info(const char log_lvl, struct task_struct task)
4264	{
4265	work_func_t *fn = NULL;
4266	char name[WQ_NAME_LEN] = { };
4267	char desc[WORKER_DESC_LEN] = { };
4268	struct pool_workqueue *pwq = NULL;
4269	struct workqueue_struct *wq = NULL;
4270	bool desc_valid = false;
4271	struct worker *worker;
4272
4273	if (!(task->flags & PF_WQ_WORKER))
4274	return;
4275
4276	/*
4277	* This function is called without any synchronization and @task
4278	* could be in any state. Be careful with dereferences.
4279	*/
4280	worker = kthread_probe_data(task);
4281
4282	/*
4283	* Carefully copy the associated workqueue's workfn and name. Keep
4284	* the original last '\0' in case the original contains garbage.
4285	*/
4286	probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4287	probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4288	probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4289	probe_kernel_read(name, wq->name, sizeof(name) - 1);
4290
4291	/* copy worker description */
4292	probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4293	if (desc_valid)
4294	probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4295
4296	if (fn \|\| name[0] \|\| desc[0]) {
4297	printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4298	if (desc[0])
4299	pr_cont(" (%s)", desc);
4300	pr_cont("\n");
4301	}
4302	}
4303
4304	static void pr_cont_pool_info(struct worker_pool *pool)
4305	{
4306	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4307	if (pool->node != NUMA_NO_NODE)
4308	pr_cont(" node=%d", pool->node);
4309	pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4310	}
4311
4312	static void pr_cont_work(bool comma, struct work_struct *work)
4313	{
4314	if (work->func == wq_barrier_func) {
4315	struct wq_barrier *barr;
4316
4317	barr = container_of(work, struct wq_barrier, work);
4318
4319	pr_cont("%s BAR(%d)", comma ? "," : "",
4320	task_pid_nr(barr->task));
4321	} else {
4322	pr_cont("%s %pf", comma ? "," : "", work->func);
4323	}
4324	}
4325
4326	static void show_pwq(struct pool_workqueue *pwq)
4327	{
4328	struct worker_pool *pool = pwq->pool;
4329	struct work_struct *work;
4330	struct worker *worker;
4331	bool has_in_flight = false, has_pending = false;
4332	int bkt;
4333
4334	pr_info(" pwq %d:", pool->id);
4335	pr_cont_pool_info(pool);
4336
4337	pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4338	!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4339
4340	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4341	if (worker->current_pwq == pwq) {
4342	has_in_flight = true;
4343	break;
4344	}
4345	}
4346	if (has_in_flight) {
4347	bool comma = false;
4348
4349	pr_info(" in-flight:");
4350	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4351	if (worker->current_pwq != pwq)
4352	continue;
4353
4354	pr_cont("%s %d%s:%pf", comma ? "," : "",
4355	task_pid_nr(worker->task),
4356	worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4357	worker->current_func);
4358	list_for_each_entry(work, &worker->scheduled, entry)
4359	pr_cont_work(false, work);
4360	comma = true;
4361	}
4362	pr_cont("\n");
4363	}
4364
4365	list_for_each_entry(work, &pool->worklist, entry) {
4366	if (get_work_pwq(work) == pwq) {
4367	has_pending = true;
4368	break;
4369	}
4370	}
4371	if (has_pending) {
4372	bool comma = false;
4373
4374	pr_info(" pending:");
4375	list_for_each_entry(work, &pool->worklist, entry) {
4376	if (get_work_pwq(work) != pwq)
4377	continue;
4378
4379	pr_cont_work(comma, work);
4380	comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4381	}
4382	pr_cont("\n");
4383	}
4384
4385	if (!list_empty(&pwq->delayed_works)) {
4386	bool comma = false;
4387
4388	pr_info(" delayed:");
4389	list_for_each_entry(work, &pwq->delayed_works, entry) {
4390	pr_cont_work(comma, work);
4391	comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4392	}
4393	pr_cont("\n");
4394	}
4395	}
4396
4397	/**
4398	* show_workqueue_state - dump workqueue state
4399	*
4400	* Called from a sysrq handler or try_to_freeze_tasks() and prints out
4401	* all busy workqueues and pools.
4402	*/
4403	void show_workqueue_state(void)
4404	{
4405	struct workqueue_struct *wq;
4406	struct worker_pool *pool;
4407	unsigned long flags;
4408	int pi;
4409
4410	rcu_read_lock_sched();
4411
4412	pr_info("Showing busy workqueues and worker pools:\n");
4413
4414	list_for_each_entry_rcu(wq, &workqueues, list) {
4415	struct pool_workqueue *pwq;
4416	bool idle = true;
4417
4418	for_each_pwq(pwq, wq) {
4419	if (pwq->nr_active \|\| !list_empty(&pwq->delayed_works)) {
4420	idle = false;
4421	break;
4422	}
4423	}
4424	if (idle)
4425	continue;
4426
4427	pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4428
4429	for_each_pwq(pwq, wq) {
4430	spin_lock_irqsave(&pwq->pool->lock, flags);
4431	if (pwq->nr_active \|\| !list_empty(&pwq->delayed_works))
4432	show_pwq(pwq);
4433	spin_unlock_irqrestore(&pwq->pool->lock, flags);
4434	}
4435	}
4436
4437	for_each_pool(pool, pi) {
4438	struct worker *worker;
4439	bool first = true;
4440
4441	spin_lock_irqsave(&pool->lock, flags);
4442	if (pool->nr_workers == pool->nr_idle)
4443	goto next_pool;
4444
4445	pr_info("pool %d:", pool->id);
4446	pr_cont_pool_info(pool);
4447	pr_cont(" hung=%us workers=%d",
4448	jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4449	pool->nr_workers);
4450	if (pool->manager)
4451	pr_cont(" manager: %d",
4452	task_pid_nr(pool->manager->task));
4453	list_for_each_entry(worker, &pool->idle_list, entry) {
4454	pr_cont(" %s%d", first ? "idle: " : "",
4455	task_pid_nr(worker->task));
4456	first = false;
4457	}
4458	pr_cont("\n");
4459	next_pool:
4460	spin_unlock_irqrestore(&pool->lock, flags);
4461	}
4462
4463	rcu_read_unlock_sched();
4464	}
4465
4466	/*
4467	* CPU hotplug.
4468	*
4469	* There are two challenges in supporting CPU hotplug. Firstly, there
4470	* are a lot of assumptions on strong associations among work, pwq and
4471	* pool which make migrating pending and scheduled works very
4472	* difficult to implement without impacting hot paths. Secondly,
4473	* worker pools serve mix of short, long and very long running works making
4474	* blocked draining impractical.
4475	*
4476	* This is solved by allowing the pools to be disassociated from the CPU
4477	* running as an unbound one and allowing it to be reattached later if the
4478	* cpu comes back online.
4479	*/
4480
4481	static void wq_unbind_fn(struct work_struct *work)
4482	{
4483	int cpu = smp_processor_id();
4484	struct worker_pool *pool;
4485	struct worker *worker;
4486
4487	for_each_cpu_worker_pool(pool, cpu) {
4488	mutex_lock(&pool->attach_mutex);
4489	spin_lock_irq(&pool->lock);
4490
4491	/*
4492	* We've blocked all attach/detach operations. Make all workers
4493	* unbound and set DISASSOCIATED. Before this, all workers
4494	* except for the ones which are still executing works from
4495	* before the last CPU down must be on the cpu. After
4496	* this, they may become diasporas.
4497	*/
4498	for_each_pool_worker(worker, pool)
4499	worker->flags \|= WORKER_UNBOUND;
4500
4501	pool->flags \|= POOL_DISASSOCIATED;
4502
4503	spin_unlock_irq(&pool->lock);
4504	mutex_unlock(&pool->attach_mutex);
4505
4506	/*
4507	* Call schedule() so that we cross rq->lock and thus can
4508	* guarantee sched callbacks see the %WORKER_UNBOUND flag.
4509	* This is necessary as scheduler callbacks may be invoked
4510	* from other cpus.
4511	*/
4512	schedule();
4513
4514	/*
4515	* Sched callbacks are disabled now. Zap nr_running.
4516	* After this, nr_running stays zero and need_more_worker()
4517	* and keep_working() are always true as long as the
4518	* worklist is not empty. This pool now behaves as an
4519	* unbound (in terms of concurrency management) pool which
4520	* are served by workers tied to the pool.
4521	*/
4522	atomic_set(&pool->nr_running, 0);
4523
4524	/*
4525	* With concurrency management just turned off, a busy
4526	* worker blocking could lead to lengthy stalls. Kick off
4527	* unbound chain execution of currently pending work items.
4528	*/
4529	spin_lock_irq(&pool->lock);
4530	wake_up_worker(pool);
4531	spin_unlock_irq(&pool->lock);
4532	}
4533	}
4534
4535	/**
4536	* rebind_workers - rebind all workers of a pool to the associated CPU
4537	* @pool: pool of interest
4538	*
4539	* @pool->cpu is coming online. Rebind all workers to the CPU.
4540	*/
4541	static void rebind_workers(struct worker_pool *pool)
4542	{
4543	struct worker *worker;
4544
4545	lockdep_assert_held(&pool->attach_mutex);
4546
4547	/*
4548	* Restore CPU affinity of all workers. As all idle workers should
4549	* be on the run-queue of the associated CPU before any local
4550	* wake-ups for concurrency management happen, restore CPU affinity
4551	* of all workers first and then clear UNBOUND. As we're called
4552	* from CPU_ONLINE, the following shouldn't fail.
4553	*/
4554	for_each_pool_worker(worker, pool)
4555	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4556	pool->attrs->cpumask) < 0);
4557
4558	spin_lock_irq(&pool->lock);
4559
4560	/*
4561	* XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4562	* w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4563	* being reworked and this can go away in time.
4564	*/
4565	if (!(pool->flags & POOL_DISASSOCIATED)) {
4566	spin_unlock_irq(&pool->lock);
4567	return;
4568	}
4569
4570	pool->flags &= ~POOL_DISASSOCIATED;
4571
4572	for_each_pool_worker(worker, pool) {
4573	unsigned int worker_flags = worker->flags;
4574
4575	/*
4576	* A bound idle worker should actually be on the runqueue
4577	* of the associated CPU for local wake-ups targeting it to
4578	* work. Kick all idle workers so that they migrate to the
4579	* associated CPU. Doing this in the same loop as
4580	* replacing UNBOUND with REBOUND is safe as no worker will
4581	* be bound before @pool->lock is released.
4582	*/
4583	if (worker_flags & WORKER_IDLE)
4584	wake_up_process(worker->task);
4585
4586	/*
4587	* We want to clear UNBOUND but can't directly call
4588	* worker_clr_flags() or adjust nr_running. Atomically
4589	* replace UNBOUND with another NOT_RUNNING flag REBOUND.
4590	* @worker will clear REBOUND using worker_clr_flags() when
4591	* it initiates the next execution cycle thus restoring
4592	* concurrency management. Note that when or whether
4593	* @worker clears REBOUND doesn't affect correctness.
4594	*
4595	* ACCESS_ONCE() is necessary because @worker->flags may be
4596	* tested without holding any lock in
4597	* wq_worker_waking_up(). Without it, NOT_RUNNING test may
4598	* fail incorrectly leading to premature concurrency
4599	* management operations.
4600	*/
4601	WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4602	worker_flags \|= WORKER_REBOUND;
4603	worker_flags &= ~WORKER_UNBOUND;
4604	ACCESS_ONCE(worker->flags) = worker_flags;
4605	}
4606
4607	spin_unlock_irq(&pool->lock);
4608	}
4609
4610	/**
4611	* restore_unbound_workers_cpumask - restore cpumask of unbound workers
4612	* @pool: unbound pool of interest
4613	* @cpu: the CPU which is coming up
4614	*
4615	* An unbound pool may end up with a cpumask which doesn't have any online
4616	* CPUs. When a worker of such pool get scheduled, the scheduler resets
4617	* its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4618	* online CPU before, cpus_allowed of all its workers should be restored.
4619	*/
4620	static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4621	{
4622	static cpumask_t cpumask;
4623	struct worker *worker;
4624
4625	lockdep_assert_held(&pool->attach_mutex);
4626
4627	/* is @cpu allowed for @pool? */
4628	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4629	return;
4630
4631	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4632
4633	/* as we're called from CPU_ONLINE, the following shouldn't fail */
4634	for_each_pool_worker(worker, pool)
4635	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4636	}
4637
4638	int workqueue_prepare_cpu(unsigned int cpu)
4639	{
4640	struct worker_pool *pool;
4641
4642	for_each_cpu_worker_pool(pool, cpu) {
4643	if (pool->nr_workers)
4644	continue;
4645	if (!create_worker(pool))
4646	return -ENOMEM;
4647	}
4648	return 0;
4649	}
4650
4651	int workqueue_online_cpu(unsigned int cpu)
4652	{
4653	struct worker_pool *pool;
4654	struct workqueue_struct *wq;
4655	int pi;
4656
4657	mutex_lock(&wq_pool_mutex);
4658
4659	for_each_pool(pool, pi) {
4660	mutex_lock(&pool->attach_mutex);
4661
4662	if (pool->cpu == cpu)
4663	rebind_workers(pool);
4664	else if (pool->cpu < 0)
4665	restore_unbound_workers_cpumask(pool, cpu);
4666
4667	mutex_unlock(&pool->attach_mutex);
4668	}
4669
4670	/* update NUMA affinity of unbound workqueues */
4671	list_for_each_entry(wq, &workqueues, list)
4672	wq_update_unbound_numa(wq, cpu, true);
4673
4674	mutex_unlock(&wq_pool_mutex);
4675	return 0;
4676	}
4677
4678	int workqueue_offline_cpu(unsigned int cpu)
4679	{
4680	struct work_struct unbind_work;
4681	struct workqueue_struct *wq;
4682
4683	/* unbinding per-cpu workers should happen on the local CPU */
4684	INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4685	queue_work_on(cpu, system_highpri_wq, &unbind_work);
4686
4687	/* update NUMA affinity of unbound workqueues */
4688	mutex_lock(&wq_pool_mutex);
4689	list_for_each_entry(wq, &workqueues, list)
4690	wq_update_unbound_numa(wq, cpu, false);
4691	mutex_unlock(&wq_pool_mutex);
4692
4693	/* wait for per-cpu unbinding to finish */
4694	flush_work(&unbind_work);
4695	destroy_work_on_stack(&unbind_work);
4696	return 0;
4697	}
4698
4699	#ifdef CONFIG_SMP
4700
4701	struct work_for_cpu {
4702	struct work_struct work;
4703	long (fn)(void );
4704	void *arg;
4705	long ret;
4706	};
4707
4708	static void work_for_cpu_fn(struct work_struct *work)
4709	{
4710	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4711
4712	wfc->ret = wfc->fn(wfc->arg);
4713	}
4714
4715	/**
4716	* work_on_cpu - run a function in thread context on a particular cpu
4717	* @cpu: the cpu to run on
4718	* @fn: the function to run
4719	* @arg: the function arg
4720	*
4721	* It is up to the caller to ensure that the cpu doesn't go offline.
4722	* The caller must not hold any locks which would prevent @fn from completing.
4723	*
4724	* Return: The value @fn returns.
4725	*/
4726	long work_on_cpu(int cpu, long (fn)(void ), void *arg)
4727	{
4728	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4729
4730	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4731	schedule_work_on(cpu, &wfc.work);
4732	flush_work(&wfc.work);
4733	destroy_work_on_stack(&wfc.work);
4734	return wfc.ret;
4735	}
4736	EXPORT_SYMBOL_GPL(work_on_cpu);
4737	#endif /* CONFIG_SMP */
4738
4739	#ifdef CONFIG_FREEZER
4740
4741	/**
4742	* freeze_workqueues_begin - begin freezing workqueues
4743	*
4744	* Start freezing workqueues. After this function returns, all freezable
4745	* workqueues will queue new works to their delayed_works list instead of
4746	* pool->worklist.
4747	*
4748	* CONTEXT:
4749	* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4750	*/
4751	void freeze_workqueues_begin(void)
4752	{
4753	struct workqueue_struct *wq;
4754	struct pool_workqueue *pwq;
4755
4756	mutex_lock(&wq_pool_mutex);
4757
4758	WARN_ON_ONCE(workqueue_freezing);
4759	workqueue_freezing = true;
4760
4761	list_for_each_entry(wq, &workqueues, list) {
4762	mutex_lock(&wq->mutex);
4763	for_each_pwq(pwq, wq)
4764	pwq_adjust_max_active(pwq);
4765	mutex_unlock(&wq->mutex);
4766	}
4767
4768	mutex_unlock(&wq_pool_mutex);
4769	}
4770
4771	/**
4772	* freeze_workqueues_busy - are freezable workqueues still busy?
4773	*
4774	* Check whether freezing is complete. This function must be called
4775	* between freeze_workqueues_begin() and thaw_workqueues().
4776	*
4777	* CONTEXT:
4778	* Grabs and releases wq_pool_mutex.
4779	*
4780	* Return:
4781	* %true if some freezable workqueues are still busy. %false if freezing
4782	* is complete.
4783	*/
4784	bool freeze_workqueues_busy(void)
4785	{
4786	bool busy = false;
4787	struct workqueue_struct *wq;
4788	struct pool_workqueue *pwq;
4789
4790	mutex_lock(&wq_pool_mutex);
4791
4792	WARN_ON_ONCE(!workqueue_freezing);
4793
4794	list_for_each_entry(wq, &workqueues, list) {
4795	if (!(wq->flags & WQ_FREEZABLE))
4796	continue;
4797	/*
4798	* nr_active is monotonically decreasing. It's safe
4799	* to peek without lock.
4800	*/
4801	rcu_read_lock_sched();
4802	for_each_pwq(pwq, wq) {
4803	WARN_ON_ONCE(pwq->nr_active < 0);
4804	if (pwq->nr_active) {
4805	busy = true;
4806	rcu_read_unlock_sched();
4807	goto out_unlock;
4808	}
4809	}
4810	rcu_read_unlock_sched();
4811	}
4812	out_unlock:
4813	mutex_unlock(&wq_pool_mutex);
4814	return busy;
4815	}
4816
4817	/**
4818	* thaw_workqueues - thaw workqueues
4819	*
4820	* Thaw workqueues. Normal queueing is restored and all collected
4821	* frozen works are transferred to their respective pool worklists.
4822	*
4823	* CONTEXT:
4824	* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4825	*/
4826	void thaw_workqueues(void)
4827	{
4828	struct workqueue_struct *wq;
4829	struct pool_workqueue *pwq;
4830
4831	mutex_lock(&wq_pool_mutex);
4832
4833	if (!workqueue_freezing)
4834	goto out_unlock;
4835
4836	workqueue_freezing = false;
4837
4838	/* restore max_active and repopulate worklist */
4839	list_for_each_entry(wq, &workqueues, list) {
4840	mutex_lock(&wq->mutex);
4841	for_each_pwq(pwq, wq)
4842	pwq_adjust_max_active(pwq);
4843	mutex_unlock(&wq->mutex);
4844	}
4845
4846	out_unlock:
4847	mutex_unlock(&wq_pool_mutex);
4848	}
4849	#endif /* CONFIG_FREEZER */
4850
4851	static int workqueue_apply_unbound_cpumask(void)
4852	{
4853	LIST_HEAD(ctxs);
4854	int ret = 0;
4855	struct workqueue_struct *wq;
4856	struct apply_wqattrs_ctx ctx, n;
4857
4858	lockdep_assert_held(&wq_pool_mutex);
4859
4860	list_for_each_entry(wq, &workqueues, list) {
4861	if (!(wq->flags & WQ_UNBOUND))
4862	continue;
4863	/* creating multiple pwqs breaks ordering guarantee */
4864	if (wq->flags & __WQ_ORDERED)
4865	continue;
4866
4867	ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4868	if (!ctx) {
4869	ret = -ENOMEM;
4870	break;
4871	}
4872
4873	list_add_tail(&ctx->list, &ctxs);
4874	}
4875
4876	list_for_each_entry_safe(ctx, n, &ctxs, list) {
4877	if (!ret)
4878	apply_wqattrs_commit(ctx);
4879	apply_wqattrs_cleanup(ctx);
4880	}
4881
4882	return ret;
4883	}
4884
4885	/**
4886	* workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4887	* @cpumask: the cpumask to set
4888	*
4889	* The low-level workqueues cpumask is a global cpumask that limits
4890	* the affinity of all unbound workqueues. This function check the @cpumask
4891	* and apply it to all unbound workqueues and updates all pwqs of them.
4892	*
4893	* Retun: 0 - Success
4894	* -EINVAL - Invalid @cpumask
4895	* -ENOMEM - Failed to allocate memory for attrs or pwqs.
4896	*/
4897	int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4898	{
4899	int ret = -EINVAL;
4900	cpumask_var_t saved_cpumask;
4901
4902	if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4903	return -ENOMEM;
4904
4905	cpumask_and(cpumask, cpumask, cpu_possible_mask);
4906	if (!cpumask_empty(cpumask)) {
4907	apply_wqattrs_lock();
4908
4909	/* save the old wq_unbound_cpumask. */
4910	cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4911
4912	/* update wq_unbound_cpumask at first and apply it to wqs. */
4913	cpumask_copy(wq_unbound_cpumask, cpumask);
4914	ret = workqueue_apply_unbound_cpumask();
4915
4916	/* restore the wq_unbound_cpumask when failed. */
4917	if (ret < 0)
4918	cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4919
4920	apply_wqattrs_unlock();
4921	}
4922
4923	free_cpumask_var(saved_cpumask);
4924	return ret;
4925	}
4926
4927	#ifdef CONFIG_SYSFS
4928	/*
4929	* Workqueues with WQ_SYSFS flag set is visible to userland via
4930	* /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4931	* following attributes.
4932	*
4933	* per_cpu RO bool : whether the workqueue is per-cpu or unbound
4934	* max_active RW int : maximum number of in-flight work items
4935	*
4936	* Unbound workqueues have the following extra attributes.
4937	*
4938	* id RO int : the associated pool ID
4939	* nice RW int : nice value of the workers
4940	* cpumask RW mask : bitmask of allowed CPUs for the workers
4941	*/
4942	struct wq_device {
4943	struct workqueue_struct *wq;
4944	struct device dev;
4945	};
4946
4947	static struct workqueue_struct dev_to_wq(struct device dev)
4948	{
4949	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4950
4951	return wq_dev->wq;
4952	}
4953
4954	static ssize_t per_cpu_show(struct device dev, struct device_attribute attr,
4955	char *buf)
4956	{
4957	struct workqueue_struct *wq = dev_to_wq(dev);
4958
4959	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4960	}
4961	static DEVICE_ATTR_RO(per_cpu);
4962
4963	static ssize_t max_active_show(struct device *dev,
4964	struct device_attribute attr, char buf)
4965	{
4966	struct workqueue_struct *wq = dev_to_wq(dev);
4967
4968	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4969	}
4970
4971	static ssize_t max_active_store(struct device *dev,
4972	struct device_attribute attr, const char buf,
4973	size_t count)
4974	{
4975	struct workqueue_struct *wq = dev_to_wq(dev);
4976	int val;
4977
4978	if (sscanf(buf, "%d", &val) != 1 \|\| val <= 0)
4979	return -EINVAL;
4980
4981	workqueue_set_max_active(wq, val);
4982	return count;
4983	}
4984	static DEVICE_ATTR_RW(max_active);
4985
4986	static struct attribute *wq_sysfs_attrs[] = {
4987	&dev_attr_per_cpu.attr,
4988	&dev_attr_max_active.attr,
4989	NULL,
4990	};
4991	ATTRIBUTE_GROUPS(wq_sysfs);
4992
4993	static ssize_t wq_pool_ids_show(struct device *dev,
4994	struct device_attribute attr, char buf)
4995	{
4996	struct workqueue_struct *wq = dev_to_wq(dev);
4997	const char *delim = "";
4998	int node, written = 0;
4999
5000	rcu_read_lock_sched();
5001	for_each_node(node) {
5002	written += scnprintf(buf + written, PAGE_SIZE - written,
5003	"%s%d:%d", delim, node,
5004	unbound_pwq_by_node(wq, node)->pool->id);
5005	delim = " ";
5006	}
5007	written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5008	rcu_read_unlock_sched();
5009
5010	return written;
5011	}
5012
5013	static ssize_t wq_nice_show(struct device dev, struct device_attribute attr,
5014	char *buf)
5015	{
5016	struct workqueue_struct *wq = dev_to_wq(dev);
5017	int written;
5018
5019	mutex_lock(&wq->mutex);
5020	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5021	mutex_unlock(&wq->mutex);
5022
5023	return written;
5024	}
5025
5026	/* prepare workqueue_attrs for sysfs store operations */
5027	static struct workqueue_attrs wq_sysfs_prep_attrs(struct workqueue_struct wq)
5028	{
5029	struct workqueue_attrs *attrs;
5030
5031	lockdep_assert_held(&wq_pool_mutex);
5032
5033	attrs = alloc_workqueue_attrs(GFP_KERNEL);
5034	if (!attrs)
5035	return NULL;
5036
5037	copy_workqueue_attrs(attrs, wq->unbound_attrs);
5038	return attrs;
5039	}
5040
5041	static ssize_t wq_nice_store(struct device dev, struct device_attribute attr,
5042	const char *buf, size_t count)
5043	{
5044	struct workqueue_struct *wq = dev_to_wq(dev);
5045	struct workqueue_attrs *attrs;
5046	int ret = -ENOMEM;
5047
5048	apply_wqattrs_lock();
5049
5050	attrs = wq_sysfs_prep_attrs(wq);
5051	if (!attrs)
5052	goto out_unlock;
5053
5054	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5055	attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5056	ret = apply_workqueue_attrs_locked(wq, attrs);
5057	else
5058	ret = -EINVAL;
5059
5060	out_unlock:
5061	apply_wqattrs_unlock();
5062	free_workqueue_attrs(attrs);
5063	return ret ?: count;
5064	}
5065
5066	static ssize_t wq_cpumask_show(struct device *dev,
5067	struct device_attribute attr, char buf)
5068	{
5069	struct workqueue_struct *wq = dev_to_wq(dev);
5070	int written;
5071
5072	mutex_lock(&wq->mutex);
5073	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5074	cpumask_pr_args(wq->unbound_attrs->cpumask));
5075	mutex_unlock(&wq->mutex);
5076	return written;
5077	}
5078
5079	static ssize_t wq_cpumask_store(struct device *dev,
5080	struct device_attribute *attr,
5081	const char *buf, size_t count)
5082	{
5083	struct workqueue_struct *wq = dev_to_wq(dev);
5084	struct workqueue_attrs *attrs;
5085	int ret = -ENOMEM;
5086
5087	apply_wqattrs_lock();
5088
5089	attrs = wq_sysfs_prep_attrs(wq);
5090	if (!attrs)
5091	goto out_unlock;
5092
5093	ret = cpumask_parse(buf, attrs->cpumask);
5094	if (!ret)
5095	ret = apply_workqueue_attrs_locked(wq, attrs);
5096
5097	out_unlock:
5098	apply_wqattrs_unlock();
5099	free_workqueue_attrs(attrs);
5100	return ret ?: count;
5101	}
5102
5103	static ssize_t wq_numa_show(struct device dev, struct device_attribute attr,
5104	char *buf)
5105	{
5106	struct workqueue_struct *wq = dev_to_wq(dev);
5107	int written;
5108
5109	mutex_lock(&wq->mutex);
5110	written = scnprintf(buf, PAGE_SIZE, "%d\n",
5111	!wq->unbound_attrs->no_numa);
5112	mutex_unlock(&wq->mutex);
5113
5114	return written;
5115	}
5116
5117	static ssize_t wq_numa_store(struct device dev, struct device_attribute attr,
5118	const char *buf, size_t count)
5119	{
5120	struct workqueue_struct *wq = dev_to_wq(dev);
5121	struct workqueue_attrs *attrs;
5122	int v, ret = -ENOMEM;
5123
5124	apply_wqattrs_lock();
5125
5126	attrs = wq_sysfs_prep_attrs(wq);
5127	if (!attrs)
5128	goto out_unlock;
5129
5130	ret = -EINVAL;
5131	if (sscanf(buf, "%d", &v) == 1) {
5132	attrs->no_numa = !v;
5133	ret = apply_workqueue_attrs_locked(wq, attrs);
5134	}
5135
5136	out_unlock:
5137	apply_wqattrs_unlock();
5138	free_workqueue_attrs(attrs);
5139	return ret ?: count;
5140	}
5141
5142	static struct device_attribute wq_sysfs_unbound_attrs[] = {
5143	__ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5144	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5145	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5146	__ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5147	__ATTR_NULL,
5148	};
5149
5150	static struct bus_type wq_subsys = {
5151	.name = "workqueue",
5152	.dev_groups = wq_sysfs_groups,
5153	};
5154
5155	static ssize_t wq_unbound_cpumask_show(struct device *dev,
5156	struct device_attribute attr, char buf)
5157	{
5158	int written;
5159
5160	mutex_lock(&wq_pool_mutex);
5161	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5162	cpumask_pr_args(wq_unbound_cpumask));
5163	mutex_unlock(&wq_pool_mutex);
5164
5165	return written;
5166	}
5167
5168	static ssize_t wq_unbound_cpumask_store(struct device *dev,
5169	struct device_attribute attr, const char buf, size_t count)
5170	{
5171	cpumask_var_t cpumask;
5172	int ret;
5173
5174	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5175	return -ENOMEM;
5176
5177	ret = cpumask_parse(buf, cpumask);
5178	if (!ret)
5179	ret = workqueue_set_unbound_cpumask(cpumask);
5180
5181	free_cpumask_var(cpumask);
5182	return ret ? ret : count;
5183	}
5184
5185	static struct device_attribute wq_sysfs_cpumask_attr =
5186	__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5187	wq_unbound_cpumask_store);
5188
5189	static int __init wq_sysfs_init(void)
5190	{
5191	int err;
5192
5193	err = subsys_virtual_register(&wq_subsys, NULL);
5194	if (err)
5195	return err;
5196
5197	return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5198	}
5199	core_initcall(wq_sysfs_init);
5200
5201	static void wq_device_release(struct device *dev)
5202	{
5203	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5204
5205	kfree(wq_dev);
5206	}
5207
5208	/**
5209	* workqueue_sysfs_register - make a workqueue visible in sysfs
5210	* @wq: the workqueue to register
5211	*
5212	* Expose @wq in sysfs under /sys/bus/workqueue/devices.
5213	* alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5214	* which is the preferred method.
5215	*
5216	* Workqueue user should use this function directly iff it wants to apply
5217	* workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5218	* apply_workqueue_attrs() may race against userland updating the
5219	* attributes.
5220	*
5221	* Return: 0 on success, -errno on failure.
5222	*/
5223	int workqueue_sysfs_register(struct workqueue_struct *wq)
5224	{
5225	struct wq_device *wq_dev;
5226	int ret;
5227
5228	/*
5229	* Adjusting max_active or creating new pwqs by applying
5230	* attributes breaks ordering guarantee. Disallow exposing ordered
5231	* workqueues.
5232	*/
5233	if (WARN_ON(wq->flags & __WQ_ORDERED))
5234	return -EINVAL;
5235
5236	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5237	if (!wq_dev)
5238	return -ENOMEM;
5239
5240	wq_dev->wq = wq;
5241	wq_dev->dev.bus = &wq_subsys;
5242	wq_dev->dev.release = wq_device_release;
5243	dev_set_name(&wq_dev->dev, "%s", wq->name);
5244
5245	/*
5246	* unbound_attrs are created separately. Suppress uevent until
5247	* everything is ready.
5248	*/
5249	dev_set_uevent_suppress(&wq_dev->dev, true);
5250
5251	ret = device_register(&wq_dev->dev);
5252	if (ret) {
5253	kfree(wq_dev);
5254	wq->wq_dev = NULL;
5255	return ret;
5256	}
5257
5258	if (wq->flags & WQ_UNBOUND) {
5259	struct device_attribute *attr;
5260
5261	for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5262	ret = device_create_file(&wq_dev->dev, attr);
5263	if (ret) {
5264	device_unregister(&wq_dev->dev);
5265	wq->wq_dev = NULL;
5266	return ret;
5267	}
5268	}
5269	}
5270
5271	dev_set_uevent_suppress(&wq_dev->dev, false);
5272	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5273	return 0;
5274	}
5275
5276	/**
5277	* workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5278	* @wq: the workqueue to unregister
5279	*
5280	* If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5281	*/
5282	static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5283	{
5284	struct wq_device *wq_dev = wq->wq_dev;
5285
5286	if (!wq->wq_dev)
5287	return;
5288
5289	wq->wq_dev = NULL;
5290	device_unregister(&wq_dev->dev);
5291	}
5292	#else /* CONFIG_SYSFS */
5293	static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5294	#endif /* CONFIG_SYSFS */
5295
5296	/*
5297	* Workqueue watchdog.
5298	*
5299	* Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5300	* flush dependency, a concurrency managed work item which stays RUNNING
5301	* indefinitely. Workqueue stalls can be very difficult to debug as the
5302	* usual warning mechanisms don't trigger and internal workqueue state is
5303	* largely opaque.
5304	*
5305	* Workqueue watchdog monitors all worker pools periodically and dumps
5306	* state if some pools failed to make forward progress for a while where
5307	* forward progress is defined as the first item on ->worklist changing.
5308	*
5309	* This mechanism is controlled through the kernel parameter
5310	* "workqueue.watchdog_thresh" which can be updated at runtime through the
5311	* corresponding sysfs parameter file.
5312	*/
5313	#ifdef CONFIG_WQ_WATCHDOG
5314
5315	static void wq_watchdog_timer_fn(unsigned long data);
5316
5317	static unsigned long wq_watchdog_thresh = 30;
5318	static struct timer_list wq_watchdog_timer =
5319	TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5320
5321	static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5322	static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5323
5324	static void wq_watchdog_reset_touched(void)
5325	{
5326	int cpu;
5327
5328	wq_watchdog_touched = jiffies;
5329	for_each_possible_cpu(cpu)
5330	per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5331	}
5332
5333	static void wq_watchdog_timer_fn(unsigned long data)
5334	{
5335	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5336	bool lockup_detected = false;
5337	struct worker_pool *pool;
5338	int pi;
5339
5340	if (!thresh)
5341	return;
5342
5343	rcu_read_lock();
5344
5345	for_each_pool(pool, pi) {
5346	unsigned long pool_ts, touched, ts;
5347
5348	if (list_empty(&pool->worklist))
5349	continue;
5350
5351	/* get the latest of pool and touched timestamps */
5352	pool_ts = READ_ONCE(pool->watchdog_ts);
5353	touched = READ_ONCE(wq_watchdog_touched);
5354
5355	if (time_after(pool_ts, touched))
5356	ts = pool_ts;
5357	else
5358	ts = touched;
5359
5360	if (pool->cpu >= 0) {
5361	unsigned long cpu_touched =
5362	READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5363	pool->cpu));
5364	if (time_after(cpu_touched, ts))
5365	ts = cpu_touched;
5366	}
5367
5368	/* did we stall? */
5369	if (time_after(jiffies, ts + thresh)) {
5370	lockup_detected = true;
5371	pr_emerg("BUG: workqueue lockup - pool");
5372	pr_cont_pool_info(pool);
5373	pr_cont(" stuck for %us!\n",
5374	jiffies_to_msecs(jiffies - pool_ts) / 1000);
5375	}
5376	}
5377
5378	rcu_read_unlock();
5379
5380	if (lockup_detected)
5381	show_workqueue_state();
5382
5383	wq_watchdog_reset_touched();
5384	mod_timer(&wq_watchdog_timer, jiffies + thresh);
5385	}
5386
5387	void wq_watchdog_touch(int cpu)
5388	{
5389	if (cpu >= 0)
5390	per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5391	else
5392	wq_watchdog_touched = jiffies;
5393	}
5394
5395	static void wq_watchdog_set_thresh(unsigned long thresh)
5396	{
5397	wq_watchdog_thresh = 0;
5398	del_timer_sync(&wq_watchdog_timer);
5399
5400	if (thresh) {
5401	wq_watchdog_thresh = thresh;
5402	wq_watchdog_reset_touched();
5403	mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5404	}
5405	}
5406
5407	static int wq_watchdog_param_set_thresh(const char *val,
5408	const struct kernel_param *kp)
5409	{
5410	unsigned long thresh;
5411	int ret;
5412
5413	ret = kstrtoul(val, 0, &thresh);
5414	if (ret)
5415	return ret;
5416
5417	if (system_wq)
5418	wq_watchdog_set_thresh(thresh);
5419	else
5420	wq_watchdog_thresh = thresh;
5421
5422	return 0;
5423	}
5424
5425	static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5426	.set = wq_watchdog_param_set_thresh,
5427	.get = param_get_ulong,
5428	};
5429
5430	module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5431	0644);
5432
5433	static void wq_watchdog_init(void)
5434	{
5435	wq_watchdog_set_thresh(wq_watchdog_thresh);
5436	}
5437
5438	#else /* CONFIG_WQ_WATCHDOG */
5439
5440	static inline void wq_watchdog_init(void) { }
5441
5442	#endif /* CONFIG_WQ_WATCHDOG */
5443
5444	static void __init wq_numa_init(void)
5445	{
5446	cpumask_var_t *tbl;
5447	int node, cpu;
5448
5449	if (num_possible_nodes() <= 1)
5450	return;
5451
5452	if (wq_disable_numa) {
5453	pr_info("workqueue: NUMA affinity support disabled\n");
5454	return;
5455	}
5456
5457	wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5458	BUG_ON(!wq_update_unbound_numa_attrs_buf);
5459
5460	/*
5461	* We want masks of possible CPUs of each node which isn't readily
5462	* available. Build one from cpu_to_node() which should have been
5463	* fully initialized by now.
5464	*/
5465	tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5466	BUG_ON(!tbl);
5467
5468	for_each_node(node)
5469	BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5470	node_online(node) ? node : NUMA_NO_NODE));
5471
5472	for_each_possible_cpu(cpu) {
5473	node = cpu_to_node(cpu);
5474	if (WARN_ON(node == NUMA_NO_NODE)) {
5475	pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5476	/* happens iff arch is bonkers, let's just proceed */
5477	return;
5478	}
5479	cpumask_set_cpu(cpu, tbl[node]);
5480	}
5481
5482	wq_numa_possible_cpumask = tbl;
5483	wq_numa_enabled = true;
5484	}
5485
5486	/**
5487	* workqueue_init_early - early init for workqueue subsystem
5488	*
5489	* This is the first half of two-staged workqueue subsystem initialization
5490	* and invoked as soon as the bare basics - memory allocation, cpumasks and
5491	* idr are up. It sets up all the data structures and system workqueues
5492	* and allows early boot code to create workqueues and queue/cancel work
5493	* items. Actual work item execution starts only after kthreads can be
5494	* created and scheduled right before early initcalls.
5495	*/
5496	int __init workqueue_init_early(void)
5497	{
5498	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5499	int i, cpu;
5500
5501	WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5502
5503	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5504	cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5505
5506	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5507
5508	/* initialize CPU pools */
5509	for_each_possible_cpu(cpu) {
5510	struct worker_pool *pool;
5511
5512	i = 0;
5513	for_each_cpu_worker_pool(pool, cpu) {
5514	BUG_ON(init_worker_pool(pool));
5515	pool->cpu = cpu;
5516	cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5517	pool->attrs->nice = std_nice[i++];
5518	pool->node = cpu_to_node(cpu);
5519
5520	/* alloc pool ID */
5521	mutex_lock(&wq_pool_mutex);
5522	BUG_ON(worker_pool_assign_id(pool));
5523	mutex_unlock(&wq_pool_mutex);
5524	}
5525	}
5526
5527	/* create default unbound and ordered wq attrs */
5528	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5529	struct workqueue_attrs *attrs;
5530
5531	BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5532	attrs->nice = std_nice[i];
5533	unbound_std_wq_attrs[i] = attrs;
5534
5535	/*
5536	* An ordered wq should have only one pwq as ordering is
5537	* guaranteed by max_active which is enforced by pwqs.
5538	* Turn off NUMA so that dfl_pwq is used for all nodes.
5539	*/
5540	BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5541	attrs->nice = std_nice[i];
5542	attrs->no_numa = true;
5543	ordered_wq_attrs[i] = attrs;
5544	}
5545
5546	system_wq = alloc_workqueue("events", 0, 0);
5547	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5548	system_long_wq = alloc_workqueue("events_long", 0, 0);
5549	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5550	WQ_UNBOUND_MAX_ACTIVE);
5551	system_freezable_wq = alloc_workqueue("events_freezable",
5552	WQ_FREEZABLE, 0);
5553	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5554	WQ_POWER_EFFICIENT, 0);
5555	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5556	WQ_FREEZABLE \| WQ_POWER_EFFICIENT,
5557	0);
5558	BUG_ON(!system_wq \|\| !system_highpri_wq \|\| !system_long_wq \|\|
5559	!system_unbound_wq \|\| !system_freezable_wq \|\|
5560	!system_power_efficient_wq \|\|
5561	!system_freezable_power_efficient_wq);
5562
5563	return 0;
5564	}
5565
5566	/**
5567	* workqueue_init - bring workqueue subsystem fully online
5568	*
5569	* This is the latter half of two-staged workqueue subsystem initialization
5570	* and invoked as soon as kthreads can be created and scheduled.
5571	* Workqueues have been created and work items queued on them, but there
5572	* are no kworkers executing the work items yet. Populate the worker pools
5573	* with the initial workers and enable future kworker creations.
5574	*/
5575	int __init workqueue_init(void)
5576	{
5577	struct workqueue_struct *wq;
5578	struct worker_pool *pool;
5579	int cpu, bkt;
5580
5581	/*
5582	* It'd be simpler to initialize NUMA in workqueue_init_early() but
5583	* CPU to node mapping may not be available that early on some
5584	* archs such as power and arm64. As per-cpu pools created
5585	* previously could be missing node hint and unbound pools NUMA
5586	* affinity, fix them up.
5587	*/
5588	wq_numa_init();
5589
5590	mutex_lock(&wq_pool_mutex);
5591
5592	for_each_possible_cpu(cpu) {
5593	for_each_cpu_worker_pool(pool, cpu) {
5594	pool->node = cpu_to_node(cpu);
5595	}
5596	}
5597
5598	list_for_each_entry(wq, &workqueues, list)
5599	wq_update_unbound_numa(wq, smp_processor_id(), true);
5600
5601	mutex_unlock(&wq_pool_mutex);
5602
5603	/* create the initial workers */
5604	for_each_online_cpu(cpu) {
5605	for_each_cpu_worker_pool(pool, cpu) {
5606	pool->flags &= ~POOL_DISASSOCIATED;
5607	BUG_ON(!create_worker(pool));
5608	}
5609	}
5610
5611	hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5612	BUG_ON(!create_worker(pool));
5613
5614	wq_online = true;
5615	wq_watchdog_init();
5616
5617	return 0;
5618	}