linux - kernel/workqueue.c

7969 lines · 222.8 KiB

1	// SPDX-License-Identifier: GPL-2.0-only
2	/*
3	* kernel/workqueue.c - generic async execution with shared worker pool
4	*
5	* Copyright (C) 2002 Ingo Molnar
6	*
7	* Derived from the taskqueue/keventd code by:
8	* David Woodhouse <[email protected]>
9	* Andrew Morton
10	* Kai Petzke <[email protected]>
11	* Theodore Ts'o <[email protected]>
12	*
13	* Made to use alloc_percpu by Christoph Lameter.
14	*
15	* Copyright (C) 2010 SUSE Linux Products GmbH
16	* Copyright (C) 2010 Tejun Heo <[email protected]>
17	*
18	* This is the generic async execution mechanism. Work items as are
19	* executed in process context. The worker pool is shared and
20	* automatically managed. There are two worker pools for each CPU (one for
21	* normal work items and the other for high priority ones) and some extra
22	* pools for workqueues which are not bound to any specific CPU - the
23	* number of these backing pools is dynamic.
24	*
25	* Please read Documentation/core-api/workqueue.rst for details.
26	*/
27
28	#include <linux/export.h>
29	#include <linux/kernel.h>
30	#include <linux/sched.h>
31	#include <linux/init.h>
32	#include <linux/interrupt.h>
33	#include <linux/signal.h>
34	#include <linux/completion.h>
35	#include <linux/workqueue.h>
36	#include <linux/slab.h>
37	#include <linux/cpu.h>
38	#include <linux/notifier.h>
39	#include <linux/kthread.h>
40	#include <linux/hardirq.h>
41	#include <linux/mempolicy.h>
42	#include <linux/freezer.h>
43	#include <linux/debug_locks.h>
44	#include <linux/lockdep.h>
45	#include <linux/idr.h>
46	#include <linux/jhash.h>
47	#include <linux/hashtable.h>
48	#include <linux/rculist.h>
49	#include <linux/nodemask.h>
50	#include <linux/moduleparam.h>
51	#include <linux/uaccess.h>
52	#include <linux/sched/isolation.h>
53	#include <linux/sched/debug.h>
54	#include <linux/nmi.h>
55	#include <linux/kvm_para.h>
56	#include <linux/delay.h>
57	#include <linux/irq_work.h>
58
59	#include "workqueue_internal.h"
60
61	enum worker_pool_flags {
62	/*
63	* worker_pool flags
64	*
65	* A bound pool is either associated or disassociated with its CPU.
66	* While associated (!DISASSOCIATED), all workers are bound to the
67	* CPU and none has %WORKER_UNBOUND set and concurrency management
68	* is in effect.
69	*
70	* While DISASSOCIATED, the cpu may be offline and all workers have
71	* %WORKER_UNBOUND set and concurrency management disabled, and may
72	* be executing on any CPU. The pool behaves as an unbound one.
73	*
74	* Note that DISASSOCIATED should be flipped only while holding
75	* wq_pool_attach_mutex to avoid changing binding state while
76	* worker_attach_to_pool() is in progress.
77	*
78	* As there can only be one concurrent BH execution context per CPU, a
79	* BH pool is per-CPU and always DISASSOCIATED.
80	*/
81	POOL_BH = 1 << 0, /* is a BH pool */
82	POOL_MANAGER_ACTIVE = 1 << 1, /* being managed */
83	POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
84	POOL_BH_DRAINING = 1 << 3, /* draining after CPU offline */
85	};
86
87	enum worker_flags {
88	/* worker flags */
89	WORKER_DIE = 1 << 1, /* die die die */
90	WORKER_IDLE = 1 << 2, /* is idle */
91	WORKER_PREP = 1 << 3, /* preparing to run works */
92	WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
93	WORKER_UNBOUND = 1 << 7, /* worker is unbound */
94	WORKER_REBOUND = 1 << 8, /* worker was rebound */
95
96	WORKER_NOT_RUNNING = WORKER_PREP \| WORKER_CPU_INTENSIVE \|
97	WORKER_UNBOUND \| WORKER_REBOUND,
98	};
99
100	enum work_cancel_flags {
101	WORK_CANCEL_DELAYED = 1 << 0, /* canceling a delayed_work */
102	WORK_CANCEL_DISABLE = 1 << 1, /* canceling to disable */
103	};
104
105	enum wq_internal_consts {
106	NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
107
108	UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
109	BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
110
111	MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
112	IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
113
114	MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
115	/* call for help after 10ms
116	(min two ticks) */
117	MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
118	CREATE_COOLDOWN = HZ, /* time to breath after fail */
119
120	/*
121	* Rescue workers are used only on emergencies and shared by
122	* all cpus. Give MIN_NICE.
123	*/
124	RESCUER_NICE_LEVEL = MIN_NICE,
125	HIGHPRI_NICE_LEVEL = MIN_NICE,
126
127	WQ_NAME_LEN = 32,
128	WORKER_ID_LEN = 10 + WQ_NAME_LEN, /* "kworker/R-" + WQ_NAME_LEN */
129	};
130
131	/*
132	* We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and
133	* MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because
134	* msecs_to_jiffies() can't be an initializer.
135	*/
136	#define BH_WORKER_JIFFIES msecs_to_jiffies(2)
137	#define BH_WORKER_RESTARTS 10
138
139	/*
140	* Structure fields follow one of the following exclusion rules.
141	*
142	* I: Modifiable by initialization/destruction paths and read-only for
143	* everyone else.
144	*
145	* P: Preemption protected. Disabling preemption is enough and should
146	* only be modified and accessed from the local cpu.
147	*
148	* L: pool->lock protected. Access with pool->lock held.
149	*
150	* LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for
151	* reads.
152	*
153	* K: Only modified by worker while holding pool->lock. Can be safely read by
154	* self, while holding pool->lock or from IRQ context if %current is the
155	* kworker.
156	*
157	* S: Only modified by worker self.
158	*
159	* A: wq_pool_attach_mutex protected.
160	*
161	* PL: wq_pool_mutex protected.
162	*
163	* PR: wq_pool_mutex protected for writes. RCU protected for reads.
164	*
165	* PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
166	*
167	* PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
168	* RCU for reads.
169	*
170	* WQ: wq->mutex protected.
171	*
172	* WR: wq->mutex protected for writes. RCU protected for reads.
173	*
174	* WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read
175	* with READ_ONCE() without locking.
176	*
177	* MD: wq_mayday_lock protected.
178	*
179	* WD: Used internally by the watchdog.
180	*/
181
182	/* struct worker is defined in workqueue_internal.h */
183
184	struct worker_pool {
185	raw_spinlock_t lock; /* the pool lock */
186	int cpu; /* I: the associated cpu */
187	int node; /* I: the associated node ID */
188	int id; /* I: pool ID */
189	unsigned int flags; /* L: flags */
190
191	unsigned long watchdog_ts; /* L: watchdog timestamp */
192	bool cpu_stall; /* WD: stalled cpu bound pool */
193
194	/*
195	* The counter is incremented in a process context on the associated CPU
196	* w/ preemption disabled, and decremented or reset in the same context
197	* but w/ pool->lock held. The readers grab pool->lock and are
198	* guaranteed to see if the counter reached zero.
199	*/
200	int nr_running;
201
202	struct list_head worklist; /* L: list of pending works */
203
204	int nr_workers; /* L: total number of workers */
205	int nr_idle; /* L: currently idle workers */
206
207	struct list_head idle_list; /* L: list of idle workers */
208	struct timer_list idle_timer; /* L: worker idle timeout */
209	struct work_struct idle_cull_work; /* L: worker idle cleanup */
210
211	struct timer_list mayday_timer; /* L: SOS timer for workers */
212
213	/* a workers is either on busy_hash or idle_list, or the manager */
214	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
215	/* L: hash of busy workers */
216
217	struct worker manager; / L: purely informational */
218	struct list_head workers; /* A: attached workers */
219
220	struct ida worker_ida; /* worker IDs for task name */
221
222	struct workqueue_attrs attrs; / I: worker attributes */
223	struct hlist_node hash_node; /* PL: unbound_pool_hash node */
224	int refcnt; /* PL: refcnt for unbound pools */
225
226	/*
227	* Destruction of pool is RCU protected to allow dereferences
228	* from get_work_pool().
229	*/
230	struct rcu_head rcu;
231	};
232
233	/*
234	* Per-pool_workqueue statistics. These can be monitored using
235	* tools/workqueue/wq_monitor.py.
236	*/
237	enum pool_workqueue_stats {
238	PWQ_STAT_STARTED, /* work items started execution */
239	PWQ_STAT_COMPLETED, /* work items completed execution */
240	PWQ_STAT_CPU_TIME, /* total CPU time consumed */
241	PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
242	PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
243	PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
244	PWQ_STAT_MAYDAY, /* maydays to rescuer */
245	PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
246
247	PWQ_NR_STATS,
248	};
249
250	/*
251	* The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT
252	* of work_struct->data are used for flags and the remaining high bits
253	* point to the pwq; thus, pwqs need to be aligned at two's power of the
254	* number of flag bits.
255	*/
256	struct pool_workqueue {
257	struct worker_pool pool; / I: the associated pool */
258	struct workqueue_struct wq; / I: the owning workqueue */
259	int work_color; /* L: current color */
260	int flush_color; /* L: flushing color */
261	int refcnt; /* L: reference count */
262	int nr_in_flight[WORK_NR_COLORS];
263	/* L: nr of in_flight works */
264	bool plugged; /* L: execution suspended */
265
266	/*
267	* nr_active management and WORK_STRUCT_INACTIVE:
268	*
269	* When pwq->nr_active >= max_active, new work item is queued to
270	* pwq->inactive_works instead of pool->worklist and marked with
271	* WORK_STRUCT_INACTIVE.
272	*
273	* All work items marked with WORK_STRUCT_INACTIVE do not participate in
274	* nr_active and all work items in pwq->inactive_works are marked with
275	* WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are
276	* in pwq->inactive_works. Some of them are ready to run in
277	* pool->worklist or worker->scheduled. Those work itmes are only struct
278	* wq_barrier which is used for flush_work() and should not participate
279	* in nr_active. For non-barrier work item, it is marked with
280	* WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
281	*/
282	int nr_active; /* L: nr of active works */
283	struct list_head inactive_works; /* L: inactive works */
284	struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */
285	struct list_head pwqs_node; /* WR: node on wq->pwqs */
286	struct list_head mayday_node; /* MD: node on wq->maydays */
287
288	u64 stats[PWQ_NR_STATS];
289
290	/*
291	* Release of unbound pwq is punted to a kthread_worker. See put_pwq()
292	* and pwq_release_workfn() for details. pool_workqueue itself is also
293	* RCU protected so that the first pwq can be determined without
294	* grabbing wq->mutex.
295	*/
296	struct kthread_work release_work;
297	struct rcu_head rcu;
298	} __aligned(1 << WORK_STRUCT_PWQ_SHIFT);
299
300	/*
301	* Structure used to wait for workqueue flush.
302	*/
303	struct wq_flusher {
304	struct list_head list; /* WQ: list of flushers */
305	int flush_color; /* WQ: flush color waiting for */
306	struct completion done; /* flush completion */
307	};
308
309	struct wq_device;
310
311	/*
312	* Unlike in a per-cpu workqueue where max_active limits its concurrency level
313	* on each CPU, in an unbound workqueue, max_active applies to the whole system.
314	* As sharing a single nr_active across multiple sockets can be very expensive,
315	* the counting and enforcement is per NUMA node.
316	*
317	* The following struct is used to enforce per-node max_active. When a pwq wants
318	* to start executing a work item, it should increment ->nr using
319	* tryinc_node_nr_active(). If acquisition fails due to ->nr already being over
320	* ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish
321	* and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in
322	* round-robin order.
323	*/
324	struct wq_node_nr_active {
325	int max; /* per-node max_active */
326	atomic_t nr; /* per-node nr_active */
327	raw_spinlock_t lock; /* nests inside pool locks */
328	struct list_head pending_pwqs; /* LN: pwqs with inactive works */
329	};
330
331	/*
332	* The externally visible workqueue. It relays the issued work items to
333	* the appropriate worker_pool through its pool_workqueues.
334	*/
335	struct workqueue_struct {
336	struct list_head pwqs; /* WR: all pwqs of this wq */
337	struct list_head list; /* PR: list of all workqueues */
338
339	struct mutex mutex; /* protects this wq */
340	int work_color; /* WQ: current work color */
341	int flush_color; /* WQ: current flush color */
342	atomic_t nr_pwqs_to_flush; /* flush in progress */
343	struct wq_flusher first_flusher; / WQ: first flusher */
344	struct list_head flusher_queue; /* WQ: flush waiters */
345	struct list_head flusher_overflow; /* WQ: flush overflow list */
346
347	struct list_head maydays; /* MD: pwqs requesting rescue */
348	struct worker rescuer; / MD: rescue worker */
349
350	int nr_drainers; /* WQ: drain in progress */
351
352	/* See alloc_workqueue() function comment for info on min/max_active */
353	int max_active; /* WO: max active works */
354	int min_active; /* WO: min active works */
355	int saved_max_active; /* WQ: saved max_active */
356	int saved_min_active; /* WQ: saved min_active */
357
358	struct workqueue_attrs unbound_attrs; / PW: only for unbound wqs */
359	struct pool_workqueue __rcu dfl_pwq; / PW: only for unbound wqs */
360
361	#ifdef CONFIG_SYSFS
362	struct wq_device wq_dev; / I: for sysfs interface */
363	#endif
364	#ifdef CONFIG_LOCKDEP
365	char *lock_name;
366	struct lock_class_key key;
367	struct lockdep_map lockdep_map;
368	#endif
369	char name[WQ_NAME_LEN]; /* I: workqueue name */
370
371	/*
372	* Destruction of workqueue_struct is RCU protected to allow walking
373	* the workqueues list without grabbing wq_pool_mutex.
374	* This is used to dump all workqueues from sysrq.
375	*/
376	struct rcu_head rcu;
377
378	/* hot fields used during command issue, aligned to cacheline */
379	unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
380	struct pool_workqueue __rcu * __percpu cpu_pwq; / I: per-cpu pwqs */
381	struct wq_node_nr_active node_nr_active[]; / I: per-node nr_active */
382	};
383
384	/*
385	* Each pod type describes how CPUs should be grouped for unbound workqueues.
386	* See the comment above workqueue_attrs->affn_scope.
387	*/
388	struct wq_pod_type {
389	int nr_pods; /* number of pods */
390	cpumask_var_t pod_cpus; / pod -> cpus */
391	int pod_node; / pod -> node */
392	int cpu_pod; / cpu -> pod */
393	};
394
395	struct work_offq_data {
396	u32 pool_id;
397	u32 disable;
398	u32 flags;
399	};
400
401	static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
402	[WQ_AFFN_DFL] = "default",
403	[WQ_AFFN_CPU] = "cpu",
404	[WQ_AFFN_SMT] = "smt",
405	[WQ_AFFN_CACHE] = "cache",
406	[WQ_AFFN_NUMA] = "numa",
407	[WQ_AFFN_SYSTEM] = "system",
408	};
409
410	/*
411	* Per-cpu work items which run for longer than the following threshold are
412	* automatically considered CPU intensive and excluded from concurrency
413	* management to prevent them from noticeably delaying other per-cpu work items.
414	* ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
415	* The actual value is initialized in wq_cpu_intensive_thresh_init().
416	*/
417	static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
418	module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
419	#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
420	static unsigned int wq_cpu_intensive_warning_thresh = 4;
421	module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644);
422	#endif
423
424	/* see the comment above the definition of WQ_POWER_EFFICIENT */
425	static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
426	module_param_named(power_efficient, wq_power_efficient, bool, 0444);
427
428	static bool wq_online; /* can kworkers be created yet? */
429	static bool wq_topo_initialized __read_mostly = false;
430
431	static struct kmem_cache *pwq_cache;
432
433	static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
434	static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
435
436	/* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
437	static struct workqueue_attrs *unbound_wq_update_pwq_attrs_buf;
438
439	static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
440	static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
441	static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
442	/* wait for manager to go away */
443	static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
444
445	static LIST_HEAD(workqueues); /* PR: list of all workqueues */
446	static bool workqueue_freezing; /* PL: have wqs started freezing? */
447
448	/* PL: mirror the cpu_online_mask excluding the CPU in the midst of hotplugging */
449	static cpumask_var_t wq_online_cpumask;
450
451	/* PL&A: allowable cpus for unbound wqs and work items */
452	static cpumask_var_t wq_unbound_cpumask;
453
454	/* PL: user requested unbound cpumask via sysfs */
455	static cpumask_var_t wq_requested_unbound_cpumask;
456
457	/* PL: isolated cpumask to be excluded from unbound cpumask */
458	static cpumask_var_t wq_isolated_cpumask;
459
460	/* for further constrain wq_unbound_cpumask by cmdline parameter*/
461	static struct cpumask wq_cmdline_cpumask __initdata;
462
463	/* CPU where unbound work was last round robin scheduled from this CPU */
464	static DEFINE_PER_CPU(int, wq_rr_cpu_last);
465
466	/*
467	* Local execution of unbound work items is no longer guaranteed. The
468	* following always forces round-robin CPU selection on unbound work items
469	* to uncover usages which depend on it.
470	*/
471	#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
472	static bool wq_debug_force_rr_cpu = true;
473	#else
474	static bool wq_debug_force_rr_cpu = false;
475	#endif
476	module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
477
478	/* to raise softirq for the BH worker pools on other CPUs */
479	static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS],
480	bh_pool_irq_works);
481
482	/* the BH worker pools */
483	static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
484	bh_worker_pools);
485
486	/* the per-cpu worker pools */
487	static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
488	cpu_worker_pools);
489
490	static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
491
492	/* PL: hash of all unbound pools keyed by pool->attrs */
493	static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
494
495	/* I: attributes used when instantiating standard unbound pools on demand */
496	static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
497
498	/* I: attributes used when instantiating ordered pools on demand */
499	static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
500
501	/*
502	* I: kthread_worker to release pwq's. pwq release needs to be bounced to a
503	* process context while holding a pool lock. Bounce to a dedicated kthread
504	* worker to avoid A-A deadlocks.
505	*/
506	static struct kthread_worker *pwq_release_worker __ro_after_init;
507
508	struct workqueue_struct *system_wq __ro_after_init;
509	EXPORT_SYMBOL(system_wq);
510	struct workqueue_struct *system_highpri_wq __ro_after_init;
511	EXPORT_SYMBOL_GPL(system_highpri_wq);
512	struct workqueue_struct *system_long_wq __ro_after_init;
513	EXPORT_SYMBOL_GPL(system_long_wq);
514	struct workqueue_struct *system_unbound_wq __ro_after_init;
515	EXPORT_SYMBOL_GPL(system_unbound_wq);
516	struct workqueue_struct *system_freezable_wq __ro_after_init;
517	EXPORT_SYMBOL_GPL(system_freezable_wq);
518	struct workqueue_struct *system_power_efficient_wq __ro_after_init;
519	EXPORT_SYMBOL_GPL(system_power_efficient_wq);
520	struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
521	EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
522	struct workqueue_struct *system_bh_wq;
523	EXPORT_SYMBOL_GPL(system_bh_wq);
524	struct workqueue_struct *system_bh_highpri_wq;
525	EXPORT_SYMBOL_GPL(system_bh_highpri_wq);
526
527	static int worker_thread(void *__worker);
528	static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
529	static void show_pwq(struct pool_workqueue *pwq);
530	static void show_one_worker_pool(struct worker_pool *pool);
531
532	#define CREATE_TRACE_POINTS
533	#include <trace/events/workqueue.h>
534
535	#define assert_rcu_or_pool_mutex() \
536	RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
537	!lockdep_is_held(&wq_pool_mutex), \
538	"RCU or wq_pool_mutex should be held")
539
540	#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
541	RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
542	!lockdep_is_held(&wq->mutex) && \
543	!lockdep_is_held(&wq_pool_mutex), \
544	"RCU, wq->mutex or wq_pool_mutex should be held")
545
546	#define for_each_bh_worker_pool(pool, cpu) \
547	for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \
548	(pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
549	(pool)++)
550
551	#define for_each_cpu_worker_pool(pool, cpu) \
552	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
553	(pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
554	(pool)++)
555
556	/**
557	* for_each_pool - iterate through all worker_pools in the system
558	* @pool: iteration cursor
559	* @pi: integer used for iteration
560	*
561	* This must be called either with wq_pool_mutex held or RCU read
562	* locked. If the pool needs to be used beyond the locking in effect, the
563	* caller is responsible for guaranteeing that the pool stays online.
564	*
565	* The if/else clause exists only for the lockdep assertion and can be
566	* ignored.
567	*/
568	#define for_each_pool(pool, pi) \
569	idr_for_each_entry(&worker_pool_idr, pool, pi) \
570	if (({ assert_rcu_or_pool_mutex(); false; })) { } \
571	else
572
573	/**
574	* for_each_pool_worker - iterate through all workers of a worker_pool
575	* @worker: iteration cursor
576	* @pool: worker_pool to iterate workers of
577	*
578	* This must be called with wq_pool_attach_mutex.
579	*
580	* The if/else clause exists only for the lockdep assertion and can be
581	* ignored.
582	*/
583	#define for_each_pool_worker(worker, pool) \
584	list_for_each_entry((worker), &(pool)->workers, node) \
585	if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
586	else
587
588	/**
589	* for_each_pwq - iterate through all pool_workqueues of the specified workqueue
590	* @pwq: iteration cursor
591	* @wq: the target workqueue
592	*
593	* This must be called either with wq->mutex held or RCU read locked.
594	* If the pwq needs to be used beyond the locking in effect, the caller is
595	* responsible for guaranteeing that the pwq stays online.
596	*
597	* The if/else clause exists only for the lockdep assertion and can be
598	* ignored.
599	*/
600	#define for_each_pwq(pwq, wq) \
601	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
602	lockdep_is_held(&(wq->mutex)))
603
604	#ifdef CONFIG_DEBUG_OBJECTS_WORK
605
606	static const struct debug_obj_descr work_debug_descr;
607
608	static void work_debug_hint(void addr)
609	{
610	return ((struct work_struct *) addr)->func;
611	}
612
613	static bool work_is_static_object(void *addr)
614	{
615	struct work_struct *work = addr;
616
617	return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
618	}
619
620	/*
621	* fixup_init is called when:
622	* - an active object is initialized
623	*/
624	static bool work_fixup_init(void *addr, enum debug_obj_state state)
625	{
626	struct work_struct *work = addr;
627
628	switch (state) {
629	case ODEBUG_STATE_ACTIVE:
630	cancel_work_sync(work);
631	debug_object_init(work, &work_debug_descr);
632	return true;
633	default:
634	return false;
635	}
636	}
637
638	/*
639	* fixup_free is called when:
640	* - an active object is freed
641	*/
642	static bool work_fixup_free(void *addr, enum debug_obj_state state)
643	{
644	struct work_struct *work = addr;
645
646	switch (state) {
647	case ODEBUG_STATE_ACTIVE:
648	cancel_work_sync(work);
649	debug_object_free(work, &work_debug_descr);
650	return true;
651	default:
652	return false;
653	}
654	}
655
656	static const struct debug_obj_descr work_debug_descr = {
657	.name = "work_struct",
658	.debug_hint = work_debug_hint,
659	.is_static_object = work_is_static_object,
660	.fixup_init = work_fixup_init,
661	.fixup_free = work_fixup_free,
662	};
663
664	static inline void debug_work_activate(struct work_struct *work)
665	{
666	debug_object_activate(work, &work_debug_descr);
667	}
668
669	static inline void debug_work_deactivate(struct work_struct *work)
670	{
671	debug_object_deactivate(work, &work_debug_descr);
672	}
673
674	void __init_work(struct work_struct *work, int onstack)
675	{
676	if (onstack)
677	debug_object_init_on_stack(work, &work_debug_descr);
678	else
679	debug_object_init(work, &work_debug_descr);
680	}
681	EXPORT_SYMBOL_GPL(__init_work);
682
683	void destroy_work_on_stack(struct work_struct *work)
684	{
685	debug_object_free(work, &work_debug_descr);
686	}
687	EXPORT_SYMBOL_GPL(destroy_work_on_stack);
688
689	void destroy_delayed_work_on_stack(struct delayed_work *work)
690	{
691	destroy_timer_on_stack(&work->timer);
692	debug_object_free(&work->work, &work_debug_descr);
693	}
694	EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
695
696	#else
697	static inline void debug_work_activate(struct work_struct *work) { }
698	static inline void debug_work_deactivate(struct work_struct *work) { }
699	#endif
700
701	/**
702	* worker_pool_assign_id - allocate ID and assign it to @pool
703	* @pool: the pool pointer of interest
704	*
705	* Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
706	* successfully, -errno on failure.
707	*/
708	static int worker_pool_assign_id(struct worker_pool *pool)
709	{
710	int ret;
711
712	lockdep_assert_held(&wq_pool_mutex);
713
714	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
715	GFP_KERNEL);
716	if (ret >= 0) {
717	pool->id = ret;
718	return 0;
719	}
720	return ret;
721	}
722
723	static struct pool_workqueue __rcu **
724	unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
725	{
726	if (cpu >= 0)
727	return per_cpu_ptr(wq->cpu_pwq, cpu);
728	else
729	return &wq->dfl_pwq;
730	}
731
732	/* @cpu < 0 for dfl_pwq */
733	static struct pool_workqueue unbound_pwq(struct workqueue_struct wq, int cpu)
734	{
735	return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
736	lockdep_is_held(&wq_pool_mutex) \|\|
737	lockdep_is_held(&wq->mutex));
738	}
739
740	/**
741	* unbound_effective_cpumask - effective cpumask of an unbound workqueue
742	* @wq: workqueue of interest
743	*
744	* @wq->unbound_attrs->cpumask contains the cpumask requested by the user which
745	* is masked with wq_unbound_cpumask to determine the effective cpumask. The
746	* default pwq is always mapped to the pool with the current effective cpumask.
747	*/
748	static struct cpumask unbound_effective_cpumask(struct workqueue_struct wq)
749	{
750	return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask;
751	}
752
753	static unsigned int work_color_to_flags(int color)
754	{
755	return color << WORK_STRUCT_COLOR_SHIFT;
756	}
757
758	static int get_work_color(unsigned long work_data)
759	{
760	return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
761	((1 << WORK_STRUCT_COLOR_BITS) - 1);
762	}
763
764	static int work_next_color(int color)
765	{
766	return (color + 1) % WORK_NR_COLORS;
767	}
768
769	static unsigned long pool_offq_flags(struct worker_pool *pool)
770	{
771	return (pool->flags & POOL_BH) ? WORK_OFFQ_BH : 0;
772	}
773
774	/*
775	* While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
776	* contain the pointer to the queued pwq. Once execution starts, the flag
777	* is cleared and the high bits contain OFFQ flags and pool ID.
778	*
779	* set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling()
780	* can be used to set the pwq, pool or clear work->data. These functions should
781	* only be called while the work is owned - ie. while the PENDING bit is set.
782	*
783	* get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
784	* corresponding to a work. Pool is available once the work has been
785	* queued anywhere after initialization until it is sync canceled. pwq is
786	* available only while the work item is queued.
787	*/
788	static inline void set_work_data(struct work_struct *work, unsigned long data)
789	{
790	WARN_ON_ONCE(!work_pending(work));
791	atomic_long_set(&work->data, data \| work_static(work));
792	}
793
794	static void set_work_pwq(struct work_struct work, struct pool_workqueue pwq,
795	unsigned long flags)
796	{
797	set_work_data(work, (unsigned long)pwq \| WORK_STRUCT_PENDING \|
798	WORK_STRUCT_PWQ \| flags);
799	}
800
801	static void set_work_pool_and_keep_pending(struct work_struct *work,
802	int pool_id, unsigned long flags)
803	{
804	set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) \|
805	WORK_STRUCT_PENDING \| flags);
806	}
807
808	static void set_work_pool_and_clear_pending(struct work_struct *work,
809	int pool_id, unsigned long flags)
810	{
811	/*
812	* The following wmb is paired with the implied mb in
813	* test_and_set_bit(PENDING) and ensures all updates to @work made
814	* here are visible to and precede any updates by the next PENDING
815	* owner.
816	*/
817	smp_wmb();
818	set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) \|
819	flags);
820	/*
821	* The following mb guarantees that previous clear of a PENDING bit
822	* will not be reordered with any speculative LOADS or STORES from
823	* work->current_func, which is executed afterwards. This possible
824	* reordering can lead to a missed execution on attempt to queue
825	* the same @work. E.g. consider this case:
826	*
827	* CPU#0 CPU#1
828	* ---------------------------- --------------------------------
829	*
830	* 1 STORE event_indicated
831	* 2 queue_work_on() {
832	* 3 test_and_set_bit(PENDING)
833	* 4 } set_..._and_clear_pending() {
834	* 5 set_work_data() # clear bit
835	* 6 smp_mb()
836	* 7 work->current_func() {
837	* 8 LOAD event_indicated
838	* }
839	*
840	* Without an explicit full barrier speculative LOAD on line 8 can
841	* be executed before CPU#0 does STORE on line 1. If that happens,
842	* CPU#0 observes the PENDING bit is still set and new execution of
843	* a @work is not queued in a hope, that CPU#1 will eventually
844	* finish the queued @work. Meanwhile CPU#1 does not see
845	* event_indicated is set, because speculative LOAD was executed
846	* before actual STORE.
847	*/
848	smp_mb();
849	}
850
851	static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
852	{
853	return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK);
854	}
855
856	static struct pool_workqueue get_work_pwq(struct work_struct work)
857	{
858	unsigned long data = atomic_long_read(&work->data);
859
860	if (data & WORK_STRUCT_PWQ)
861	return work_struct_pwq(data);
862	else
863	return NULL;
864	}
865
866	/**
867	* get_work_pool - return the worker_pool a given work was associated with
868	* @work: the work item of interest
869	*
870	* Pools are created and destroyed under wq_pool_mutex, and allows read
871	* access under RCU read lock. As such, this function should be
872	* called under wq_pool_mutex or inside of a rcu_read_lock() region.
873	*
874	* All fields of the returned pool are accessible as long as the above
875	* mentioned locking is in effect. If the returned pool needs to be used
876	* beyond the critical section, the caller is responsible for ensuring the
877	* returned pool is and stays online.
878	*
879	* Return: The worker_pool @work was last associated with. %NULL if none.
880	*/
881	static struct worker_pool get_work_pool(struct work_struct work)
882	{
883	unsigned long data = atomic_long_read(&work->data);
884	int pool_id;
885
886	assert_rcu_or_pool_mutex();
887
888	if (data & WORK_STRUCT_PWQ)
889	return work_struct_pwq(data)->pool;
890
891	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
892	if (pool_id == WORK_OFFQ_POOL_NONE)
893	return NULL;
894
895	return idr_find(&worker_pool_idr, pool_id);
896	}
897
898	static unsigned long shift_and_mask(unsigned long v, u32 shift, u32 bits)
899	{
900	return (v >> shift) & ((1U << bits) - 1);
901	}
902
903	static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data)
904	{
905	WARN_ON_ONCE(data & WORK_STRUCT_PWQ);
906
907	offqd->pool_id = shift_and_mask(data, WORK_OFFQ_POOL_SHIFT,
908	WORK_OFFQ_POOL_BITS);
909	offqd->disable = shift_and_mask(data, WORK_OFFQ_DISABLE_SHIFT,
910	WORK_OFFQ_DISABLE_BITS);
911	offqd->flags = data & WORK_OFFQ_FLAG_MASK;
912	}
913
914	static unsigned long work_offqd_pack_flags(struct work_offq_data *offqd)
915	{
916	return ((unsigned long)offqd->disable << WORK_OFFQ_DISABLE_SHIFT) \|
917	((unsigned long)offqd->flags);
918	}
919
920	/*
921	* Policy functions. These define the policies on how the global worker
922	* pools are managed. Unless noted otherwise, these functions assume that
923	* they're being called with pool->lock held.
924	*/
925
926	/*
927	* Need to wake up a worker? Called from anything but currently
928	* running workers.
929	*
930	* Note that, because unbound workers never contribute to nr_running, this
931	* function will always return %true for unbound pools as long as the
932	* worklist isn't empty.
933	*/
934	static bool need_more_worker(struct worker_pool *pool)
935	{
936	return !list_empty(&pool->worklist) && !pool->nr_running;
937	}
938
939	/* Can I start working? Called from busy but !running workers. */
940	static bool may_start_working(struct worker_pool *pool)
941	{
942	return pool->nr_idle;
943	}
944
945	/* Do I need to keep working? Called from currently running workers. */
946	static bool keep_working(struct worker_pool *pool)
947	{
948	return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
949	}
950
951	/* Do we need a new worker? Called from manager. */
952	static bool need_to_create_worker(struct worker_pool *pool)
953	{
954	return need_more_worker(pool) && !may_start_working(pool);
955	}
956
957	/* Do we have too many workers and should some go away? */
958	static bool too_many_workers(struct worker_pool *pool)
959	{
960	bool managing = pool->flags & POOL_MANAGER_ACTIVE;
961	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
962	int nr_busy = pool->nr_workers - nr_idle;
963
964	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
965	}
966
967	/**
968	* worker_set_flags - set worker flags and adjust nr_running accordingly
969	* @worker: self
970	* @flags: flags to set
971	*
972	* Set @flags in @worker->flags and adjust nr_running accordingly.
973	*/
974	static inline void worker_set_flags(struct worker *worker, unsigned int flags)
975	{
976	struct worker_pool *pool = worker->pool;
977
978	lockdep_assert_held(&pool->lock);
979
980	/* If transitioning into NOT_RUNNING, adjust nr_running. */
981	if ((flags & WORKER_NOT_RUNNING) &&
982	!(worker->flags & WORKER_NOT_RUNNING)) {
983	pool->nr_running--;
984	}
985
986	worker->flags \|= flags;
987	}
988
989	/**
990	* worker_clr_flags - clear worker flags and adjust nr_running accordingly
991	* @worker: self
992	* @flags: flags to clear
993	*
994	* Clear @flags in @worker->flags and adjust nr_running accordingly.
995	*/
996	static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
997	{
998	struct worker_pool *pool = worker->pool;
999	unsigned int oflags = worker->flags;
1000
1001	lockdep_assert_held(&pool->lock);
1002
1003	worker->flags &= ~flags;
1004
1005	/*
1006	* If transitioning out of NOT_RUNNING, increment nr_running. Note
1007	* that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1008	* of multiple flags, not a single flag.
1009	*/
1010	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1011	if (!(worker->flags & WORKER_NOT_RUNNING))
1012	pool->nr_running++;
1013	}
1014
1015	/* Return the first idle worker. Called with pool->lock held. */
1016	static struct worker first_idle_worker(struct worker_pool pool)
1017	{
1018	if (unlikely(list_empty(&pool->idle_list)))
1019	return NULL;
1020
1021	return list_first_entry(&pool->idle_list, struct worker, entry);
1022	}
1023
1024	/**
1025	* worker_enter_idle - enter idle state
1026	* @worker: worker which is entering idle state
1027	*
1028	* @worker is entering idle state. Update stats and idle timer if
1029	* necessary.
1030	*
1031	* LOCKING:
1032	* raw_spin_lock_irq(pool->lock).
1033	*/
1034	static void worker_enter_idle(struct worker *worker)
1035	{
1036	struct worker_pool *pool = worker->pool;
1037
1038	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) \|\|
1039	WARN_ON_ONCE(!list_empty(&worker->entry) &&
1040	(worker->hentry.next \|\| worker->hentry.pprev)))
1041	return;
1042
1043	/* can't use worker_set_flags(), also called from create_worker() */
1044	worker->flags \|= WORKER_IDLE;
1045	pool->nr_idle++;
1046	worker->last_active = jiffies;
1047
1048	/* idle_list is LIFO */
1049	list_add(&worker->entry, &pool->idle_list);
1050
1051	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1052	mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1053
1054	/* Sanity check nr_running. */
1055	WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1056	}
1057
1058	/**
1059	* worker_leave_idle - leave idle state
1060	* @worker: worker which is leaving idle state
1061	*
1062	* @worker is leaving idle state. Update stats.
1063	*
1064	* LOCKING:
1065	* raw_spin_lock_irq(pool->lock).
1066	*/
1067	static void worker_leave_idle(struct worker *worker)
1068	{
1069	struct worker_pool *pool = worker->pool;
1070
1071	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1072	return;
1073	worker_clr_flags(worker, WORKER_IDLE);
1074	pool->nr_idle--;
1075	list_del_init(&worker->entry);
1076	}
1077
1078	/**
1079	* find_worker_executing_work - find worker which is executing a work
1080	* @pool: pool of interest
1081	* @work: work to find worker for
1082	*
1083	* Find a worker which is executing @work on @pool by searching
1084	* @pool->busy_hash which is keyed by the address of @work. For a worker
1085	* to match, its current execution should match the address of @work and
1086	* its work function. This is to avoid unwanted dependency between
1087	* unrelated work executions through a work item being recycled while still
1088	* being executed.
1089	*
1090	* This is a bit tricky. A work item may be freed once its execution
1091	* starts and nothing prevents the freed area from being recycled for
1092	* another work item. If the same work item address ends up being reused
1093	* before the original execution finishes, workqueue will identify the
1094	* recycled work item as currently executing and make it wait until the
1095	* current execution finishes, introducing an unwanted dependency.
1096	*
1097	* This function checks the work item address and work function to avoid
1098	* false positives. Note that this isn't complete as one may construct a
1099	* work function which can introduce dependency onto itself through a
1100	* recycled work item. Well, if somebody wants to shoot oneself in the
1101	* foot that badly, there's only so much we can do, and if such deadlock
1102	* actually occurs, it should be easy to locate the culprit work function.
1103	*
1104	* CONTEXT:
1105	* raw_spin_lock_irq(pool->lock).
1106	*
1107	* Return:
1108	* Pointer to worker which is executing @work if found, %NULL
1109	* otherwise.
1110	*/
1111	static struct worker find_worker_executing_work(struct worker_pool pool,
1112	struct work_struct *work)
1113	{
1114	struct worker *worker;
1115
1116	hash_for_each_possible(pool->busy_hash, worker, hentry,
1117	(unsigned long)work)
1118	if (worker->current_work == work &&
1119	worker->current_func == work->func)
1120	return worker;
1121
1122	return NULL;
1123	}
1124
1125	/**
1126	* move_linked_works - move linked works to a list
1127	* @work: start of series of works to be scheduled
1128	* @head: target list to append @work to
1129	* @nextp: out parameter for nested worklist walking
1130	*
1131	* Schedule linked works starting from @work to @head. Work series to be
1132	* scheduled starts at @work and includes any consecutive work with
1133	* WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1134	* @nextp.
1135	*
1136	* CONTEXT:
1137	* raw_spin_lock_irq(pool->lock).
1138	*/
1139	static void move_linked_works(struct work_struct work, struct list_head head,
1140	struct work_struct **nextp)
1141	{
1142	struct work_struct *n;
1143
1144	/*
1145	* Linked worklist will always end before the end of the list,
1146	* use NULL for list head.
1147	*/
1148	list_for_each_entry_safe_from(work, n, NULL, entry) {
1149	list_move_tail(&work->entry, head);
1150	if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1151	break;
1152	}
1153
1154	/*
1155	* If we're already inside safe list traversal and have moved
1156	* multiple works to the scheduled queue, the next position
1157	* needs to be updated.
1158	*/
1159	if (nextp)
1160	*nextp = n;
1161	}
1162
1163	/**
1164	* assign_work - assign a work item and its linked work items to a worker
1165	* @work: work to assign
1166	* @worker: worker to assign to
1167	* @nextp: out parameter for nested worklist walking
1168	*
1169	* Assign @work and its linked work items to @worker. If @work is already being
1170	* executed by another worker in the same pool, it'll be punted there.
1171	*
1172	* If @nextp is not NULL, it's updated to point to the next work of the last
1173	* scheduled work. This allows assign_work() to be nested inside
1174	* list_for_each_entry_safe().
1175	*
1176	* Returns %true if @work was successfully assigned to @worker. %false if @work
1177	* was punted to another worker already executing it.
1178	*/
1179	static bool assign_work(struct work_struct work, struct worker worker,
1180	struct work_struct **nextp)
1181	{
1182	struct worker_pool *pool = worker->pool;
1183	struct worker *collision;
1184
1185	lockdep_assert_held(&pool->lock);
1186
1187	/*
1188	* A single work shouldn't be executed concurrently by multiple workers.
1189	* __queue_work() ensures that @work doesn't jump to a different pool
1190	* while still running in the previous pool. Here, we should ensure that
1191	* @work is not executed concurrently by multiple workers from the same
1192	* pool. Check whether anyone is already processing the work. If so,
1193	* defer the work to the currently executing one.
1194	*/
1195	collision = find_worker_executing_work(pool, work);
1196	if (unlikely(collision)) {
1197	move_linked_works(work, &collision->scheduled, nextp);
1198	return false;
1199	}
1200
1201	move_linked_works(work, &worker->scheduled, nextp);
1202	return true;
1203	}
1204
1205	static struct irq_work bh_pool_irq_work(struct worker_pool pool)
1206	{
1207	int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0;
1208
1209	return &per_cpu(bh_pool_irq_works, pool->cpu)[high];
1210	}
1211
1212	static void kick_bh_pool(struct worker_pool *pool)
1213	{
1214	#ifdef CONFIG_SMP
1215	/* see drain_dead_softirq_workfn() for BH_DRAINING */
1216	if (unlikely(pool->cpu != smp_processor_id() &&
1217	!(pool->flags & POOL_BH_DRAINING))) {
1218	irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu);
1219	return;
1220	}
1221	#endif
1222	if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
1223	raise_softirq_irqoff(HI_SOFTIRQ);
1224	else
1225	raise_softirq_irqoff(TASKLET_SOFTIRQ);
1226	}
1227
1228	/**
1229	* kick_pool - wake up an idle worker if necessary
1230	* @pool: pool to kick
1231	*
1232	* @pool may have pending work items. Wake up worker if necessary. Returns
1233	* whether a worker was woken up.
1234	*/
1235	static bool kick_pool(struct worker_pool *pool)
1236	{
1237	struct worker *worker = first_idle_worker(pool);
1238	struct task_struct *p;
1239
1240	lockdep_assert_held(&pool->lock);
1241
1242	if (!need_more_worker(pool) \|\| !worker)
1243	return false;
1244
1245	if (pool->flags & POOL_BH) {
1246	kick_bh_pool(pool);
1247	return true;
1248	}
1249
1250	p = worker->task;
1251
1252	#ifdef CONFIG_SMP
1253	/*
1254	* Idle @worker is about to execute @work and waking up provides an
1255	* opportunity to migrate @worker at a lower cost by setting the task's
1256	* wake_cpu field. Let's see if we want to move @worker to improve
1257	* execution locality.
1258	*
1259	* We're waking the worker that went idle the latest and there's some
1260	* chance that @worker is marked idle but hasn't gone off CPU yet. If
1261	* so, setting the wake_cpu won't do anything. As this is a best-effort
1262	* optimization and the race window is narrow, let's leave as-is for
1263	* now. If this becomes pronounced, we can skip over workers which are
1264	* still on cpu when picking an idle worker.
1265	*
1266	* If @pool has non-strict affinity, @worker might have ended up outside
1267	* its affinity scope. Repatriate.
1268	*/
1269	if (!pool->attrs->affn_strict &&
1270	!cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1271	struct work_struct *work = list_first_entry(&pool->worklist,
1272	struct work_struct, entry);
1273	int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask,
1274	cpu_online_mask);
1275	if (wake_cpu < nr_cpu_ids) {
1276	p->wake_cpu = wake_cpu;
1277	get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1278	}
1279	}
1280	#endif
1281	wake_up_process(p);
1282	return true;
1283	}
1284
1285	#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1286
1287	/*
1288	* Concurrency-managed per-cpu work items that hog CPU for longer than
1289	* wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1290	* which prevents them from stalling other concurrency-managed work items. If a
1291	* work function keeps triggering this mechanism, it's likely that the work item
1292	* should be using an unbound workqueue instead.
1293	*
1294	* wq_cpu_intensive_report() tracks work functions which trigger such conditions
1295	* and report them so that they can be examined and converted to use unbound
1296	* workqueues as appropriate. To avoid flooding the console, each violating work
1297	* function is tracked and reported with exponential backoff.
1298	*/
1299	#define WCI_MAX_ENTS 128
1300
1301	struct wci_ent {
1302	work_func_t func;
1303	atomic64_t cnt;
1304	struct hlist_node hash_node;
1305	};
1306
1307	static struct wci_ent wci_ents[WCI_MAX_ENTS];
1308	static int wci_nr_ents;
1309	static DEFINE_RAW_SPINLOCK(wci_lock);
1310	static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1311
1312	static struct wci_ent *wci_find_ent(work_func_t func)
1313	{
1314	struct wci_ent *ent;
1315
1316	hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1317	(unsigned long)func) {
1318	if (ent->func == func)
1319	return ent;
1320	}
1321	return NULL;
1322	}
1323
1324	static void wq_cpu_intensive_report(work_func_t func)
1325	{
1326	struct wci_ent *ent;
1327
1328	restart:
1329	ent = wci_find_ent(func);
1330	if (ent) {
1331	u64 cnt;
1332
1333	/*
1334	* Start reporting from the warning_thresh and back off
1335	* exponentially.
1336	*/
1337	cnt = atomic64_inc_return_relaxed(&ent->cnt);
1338	if (wq_cpu_intensive_warning_thresh &&
1339	cnt >= wq_cpu_intensive_warning_thresh &&
1340	is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh))
1341	printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1342	ent->func, wq_cpu_intensive_thresh_us,
1343	atomic64_read(&ent->cnt));
1344	return;
1345	}
1346
1347	/*
1348	* @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1349	* is exhausted, something went really wrong and we probably made enough
1350	* noise already.
1351	*/
1352	if (wci_nr_ents >= WCI_MAX_ENTS)
1353	return;
1354
1355	raw_spin_lock(&wci_lock);
1356
1357	if (wci_nr_ents >= WCI_MAX_ENTS) {
1358	raw_spin_unlock(&wci_lock);
1359	return;
1360	}
1361
1362	if (wci_find_ent(func)) {
1363	raw_spin_unlock(&wci_lock);
1364	goto restart;
1365	}
1366
1367	ent = &wci_ents[wci_nr_ents++];
1368	ent->func = func;
1369	atomic64_set(&ent->cnt, 0);
1370	hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1371
1372	raw_spin_unlock(&wci_lock);
1373
1374	goto restart;
1375	}
1376
1377	#else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1378	static void wq_cpu_intensive_report(work_func_t func) {}
1379	#endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1380
1381	/**
1382	* wq_worker_running - a worker is running again
1383	* @task: task waking up
1384	*
1385	* This function is called when a worker returns from schedule()
1386	*/
1387	void wq_worker_running(struct task_struct *task)
1388	{
1389	struct worker *worker = kthread_data(task);
1390
1391	if (!READ_ONCE(worker->sleeping))
1392	return;
1393
1394	/*
1395	* If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1396	* and the nr_running increment below, we may ruin the nr_running reset
1397	* and leave with an unexpected pool->nr_running == 1 on the newly unbound
1398	* pool. Protect against such race.
1399	*/
1400	preempt_disable();
1401	if (!(worker->flags & WORKER_NOT_RUNNING))
1402	worker->pool->nr_running++;
1403	preempt_enable();
1404
1405	/*
1406	* CPU intensive auto-detection cares about how long a work item hogged
1407	* CPU without sleeping. Reset the starting timestamp on wakeup.
1408	*/
1409	worker->current_at = worker->task->se.sum_exec_runtime;
1410
1411	WRITE_ONCE(worker->sleeping, 0);
1412	}
1413
1414	/**
1415	* wq_worker_sleeping - a worker is going to sleep
1416	* @task: task going to sleep
1417	*
1418	* This function is called from schedule() when a busy worker is
1419	* going to sleep.
1420	*/
1421	void wq_worker_sleeping(struct task_struct *task)
1422	{
1423	struct worker *worker = kthread_data(task);
1424	struct worker_pool *pool;
1425
1426	/*
1427	* Rescuers, which may not have all the fields set up like normal
1428	* workers, also reach here, let's not access anything before
1429	* checking NOT_RUNNING.
1430	*/
1431	if (worker->flags & WORKER_NOT_RUNNING)
1432	return;
1433
1434	pool = worker->pool;
1435
1436	/* Return if preempted before wq_worker_running() was reached */
1437	if (READ_ONCE(worker->sleeping))
1438	return;
1439
1440	WRITE_ONCE(worker->sleeping, 1);
1441	raw_spin_lock_irq(&pool->lock);
1442
1443	/*
1444	* Recheck in case unbind_workers() preempted us. We don't
1445	* want to decrement nr_running after the worker is unbound
1446	* and nr_running has been reset.
1447	*/
1448	if (worker->flags & WORKER_NOT_RUNNING) {
1449	raw_spin_unlock_irq(&pool->lock);
1450	return;
1451	}
1452
1453	pool->nr_running--;
1454	if (kick_pool(pool))
1455	worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1456
1457	raw_spin_unlock_irq(&pool->lock);
1458	}
1459
1460	/**
1461	* wq_worker_tick - a scheduler tick occurred while a kworker is running
1462	* @task: task currently running
1463	*
1464	* Called from sched_tick(). We're in the IRQ context and the current
1465	* worker's fields which follow the 'K' locking rule can be accessed safely.
1466	*/
1467	void wq_worker_tick(struct task_struct *task)
1468	{
1469	struct worker *worker = kthread_data(task);
1470	struct pool_workqueue *pwq = worker->current_pwq;
1471	struct worker_pool *pool = worker->pool;
1472
1473	if (!pwq)
1474	return;
1475
1476	pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1477
1478	if (!wq_cpu_intensive_thresh_us)
1479	return;
1480
1481	/*
1482	* If the current worker is concurrency managed and hogged the CPU for
1483	* longer than wq_cpu_intensive_thresh_us, it's automatically marked
1484	* CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1485	*
1486	* Set @worker->sleeping means that @worker is in the process of
1487	* switching out voluntarily and won't be contributing to
1488	* @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1489	* decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1490	* double decrements. The task is releasing the CPU anyway. Let's skip.
1491	* We probably want to make this prettier in the future.
1492	*/
1493	if ((worker->flags & WORKER_NOT_RUNNING) \|\| READ_ONCE(worker->sleeping) \|\|
1494	worker->task->se.sum_exec_runtime - worker->current_at <
1495	wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1496	return;
1497
1498	raw_spin_lock(&pool->lock);
1499
1500	worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1501	wq_cpu_intensive_report(worker->current_func);
1502	pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1503
1504	if (kick_pool(pool))
1505	pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1506
1507	raw_spin_unlock(&pool->lock);
1508	}
1509
1510	/**
1511	* wq_worker_last_func - retrieve worker's last work function
1512	* @task: Task to retrieve last work function of.
1513	*
1514	* Determine the last function a worker executed. This is called from
1515	* the scheduler to get a worker's last known identity.
1516	*
1517	* CONTEXT:
1518	* raw_spin_lock_irq(rq->lock)
1519	*
1520	* This function is called during schedule() when a kworker is going
1521	* to sleep. It's used by psi to identify aggregation workers during
1522	* dequeuing, to allow periodic aggregation to shut-off when that
1523	* worker is the last task in the system or cgroup to go to sleep.
1524	*
1525	* As this function doesn't involve any workqueue-related locking, it
1526	* only returns stable values when called from inside the scheduler's
1527	* queuing and dequeuing paths, when @task, which must be a kworker,
1528	* is guaranteed to not be processing any works.
1529	*
1530	* Return:
1531	* The last work function %current executed as a worker, NULL if it
1532	* hasn't executed any work yet.
1533	*/
1534	work_func_t wq_worker_last_func(struct task_struct *task)
1535	{
1536	struct worker *worker = kthread_data(task);
1537
1538	return worker->last_func;
1539	}
1540
1541	/**
1542	* wq_node_nr_active - Determine wq_node_nr_active to use
1543	* @wq: workqueue of interest
1544	* @node: NUMA node, can be %NUMA_NO_NODE
1545	*
1546	* Determine wq_node_nr_active to use for @wq on @node. Returns:
1547	*
1548	* - %NULL for per-cpu workqueues as they don't need to use shared nr_active.
1549	*
1550	* - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1551	*
1552	* - Otherwise, node_nr_active[@node].
1553	*/
1554	static struct wq_node_nr_active wq_node_nr_active(struct workqueue_struct wq,
1555	int node)
1556	{
1557	if (!(wq->flags & WQ_UNBOUND))
1558	return NULL;
1559
1560	if (node == NUMA_NO_NODE)
1561	node = nr_node_ids;
1562
1563	return wq->node_nr_active[node];
1564	}
1565
1566	/**
1567	* wq_update_node_max_active - Update per-node max_actives to use
1568	* @wq: workqueue to update
1569	* @off_cpu: CPU that's going down, -1 if a CPU is not going down
1570	*
1571	* Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1572	* distributed among nodes according to the proportions of numbers of online
1573	* cpus. The result is always between @wq->min_active and max_active.
1574	*/
1575	static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1576	{
1577	struct cpumask *effective = unbound_effective_cpumask(wq);
1578	int min_active = READ_ONCE(wq->min_active);
1579	int max_active = READ_ONCE(wq->max_active);
1580	int total_cpus, node;
1581
1582	lockdep_assert_held(&wq->mutex);
1583
1584	if (!wq_topo_initialized)
1585	return;
1586
1587	if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective))
1588	off_cpu = -1;
1589
1590	total_cpus = cpumask_weight_and(effective, cpu_online_mask);
1591	if (off_cpu >= 0)
1592	total_cpus--;
1593
1594	/* If all CPUs of the wq get offline, use the default values */
1595	if (unlikely(!total_cpus)) {
1596	for_each_node(node)
1597	wq_node_nr_active(wq, node)->max = min_active;
1598
1599	wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1600	return;
1601	}
1602
1603	for_each_node(node) {
1604	int node_cpus;
1605
1606	node_cpus = cpumask_weight_and(effective, cpumask_of_node(node));
1607	if (off_cpu >= 0 && cpu_to_node(off_cpu) == node)
1608	node_cpus--;
1609
1610	wq_node_nr_active(wq, node)->max =
1611	clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1612	min_active, max_active);
1613	}
1614
1615	wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1616	}
1617
1618	/**
1619	* get_pwq - get an extra reference on the specified pool_workqueue
1620	* @pwq: pool_workqueue to get
1621	*
1622	* Obtain an extra reference on @pwq. The caller should guarantee that
1623	* @pwq has positive refcnt and be holding the matching pool->lock.
1624	*/
1625	static void get_pwq(struct pool_workqueue *pwq)
1626	{
1627	lockdep_assert_held(&pwq->pool->lock);
1628	WARN_ON_ONCE(pwq->refcnt <= 0);
1629	pwq->refcnt++;
1630	}
1631
1632	/**
1633	* put_pwq - put a pool_workqueue reference
1634	* @pwq: pool_workqueue to put
1635	*
1636	* Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1637	* destruction. The caller should be holding the matching pool->lock.
1638	*/
1639	static void put_pwq(struct pool_workqueue *pwq)
1640	{
1641	lockdep_assert_held(&pwq->pool->lock);
1642	if (likely(--pwq->refcnt))
1643	return;
1644	/*
1645	* @pwq can't be released under pool->lock, bounce to a dedicated
1646	* kthread_worker to avoid A-A deadlocks.
1647	*/
1648	kthread_queue_work(pwq_release_worker, &pwq->release_work);
1649	}
1650
1651	/**
1652	* put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1653	* @pwq: pool_workqueue to put (can be %NULL)
1654	*
1655	* put_pwq() with locking. This function also allows %NULL @pwq.
1656	*/
1657	static void put_pwq_unlocked(struct pool_workqueue *pwq)
1658	{
1659	if (pwq) {
1660	/*
1661	* As both pwqs and pools are RCU protected, the
1662	* following lock operations are safe.
1663	*/
1664	raw_spin_lock_irq(&pwq->pool->lock);
1665	put_pwq(pwq);
1666	raw_spin_unlock_irq(&pwq->pool->lock);
1667	}
1668	}
1669
1670	static bool pwq_is_empty(struct pool_workqueue *pwq)
1671	{
1672	return !pwq->nr_active && list_empty(&pwq->inactive_works);
1673	}
1674
1675	static void __pwq_activate_work(struct pool_workqueue *pwq,
1676	struct work_struct *work)
1677	{
1678	unsigned long *wdb = work_data_bits(work);
1679
1680	WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1681	trace_workqueue_activate_work(work);
1682	if (list_empty(&pwq->pool->worklist))
1683	pwq->pool->watchdog_ts = jiffies;
1684	move_linked_works(work, &pwq->pool->worklist, NULL);
1685	__clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1686	}
1687
1688	static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1689	{
1690	int max = READ_ONCE(nna->max);
1691
1692	while (true) {
1693	int old, tmp;
1694
1695	old = atomic_read(&nna->nr);
1696	if (old >= max)
1697	return false;
1698	tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1);
1699	if (tmp == old)
1700	return true;
1701	}
1702	}
1703
1704	/**
1705	* pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1706	* @pwq: pool_workqueue of interest
1707	* @fill: max_active may have increased, try to increase concurrency level
1708	*
1709	* Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1710	* successfully obtained. %false otherwise.
1711	*/
1712	static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1713	{
1714	struct workqueue_struct *wq = pwq->wq;
1715	struct worker_pool *pool = pwq->pool;
1716	struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node);
1717	bool obtained = false;
1718
1719	lockdep_assert_held(&pool->lock);
1720
1721	if (!nna) {
1722	/* BH or per-cpu workqueue, pwq->nr_active is sufficient */
1723	obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1724	goto out;
1725	}
1726
1727	if (unlikely(pwq->plugged))
1728	return false;
1729
1730	/*
1731	* Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1732	* already waiting on $nna, pwq_dec_nr_active() will maintain the
1733	* concurrency level. Don't jump the line.
1734	*
1735	* We need to ignore the pending test after max_active has increased as
1736	* pwq_dec_nr_active() can only maintain the concurrency level but not
1737	* increase it. This is indicated by @fill.
1738	*/
1739	if (!list_empty(&pwq->pending_node) && likely(!fill))
1740	goto out;
1741
1742	obtained = tryinc_node_nr_active(nna);
1743	if (obtained)
1744	goto out;
1745
1746	/*
1747	* Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1748	* and try again. The smp_mb() is paired with the implied memory barrier
1749	* of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1750	* we see the decremented $nna->nr or they see non-empty
1751	* $nna->pending_pwqs.
1752	*/
1753	raw_spin_lock(&nna->lock);
1754
1755	if (list_empty(&pwq->pending_node))
1756	list_add_tail(&pwq->pending_node, &nna->pending_pwqs);
1757	else if (likely(!fill))
1758	goto out_unlock;
1759
1760	smp_mb();
1761
1762	obtained = tryinc_node_nr_active(nna);
1763
1764	/*
1765	* If @fill, @pwq might have already been pending. Being spuriously
1766	* pending in cold paths doesn't affect anything. Let's leave it be.
1767	*/
1768	if (obtained && likely(!fill))
1769	list_del_init(&pwq->pending_node);
1770
1771	out_unlock:
1772	raw_spin_unlock(&nna->lock);
1773	out:
1774	if (obtained)
1775	pwq->nr_active++;
1776	return obtained;
1777	}
1778
1779	/**
1780	* pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1781	* @pwq: pool_workqueue of interest
1782	* @fill: max_active may have increased, try to increase concurrency level
1783	*
1784	* Activate the first inactive work item of @pwq if available and allowed by
1785	* max_active limit.
1786	*
1787	* Returns %true if an inactive work item has been activated. %false if no
1788	* inactive work item is found or max_active limit is reached.
1789	*/
1790	static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1791	{
1792	struct work_struct *work =
1793	list_first_entry_or_null(&pwq->inactive_works,
1794	struct work_struct, entry);
1795
1796	if (work && pwq_tryinc_nr_active(pwq, fill)) {
1797	__pwq_activate_work(pwq, work);
1798	return true;
1799	} else {
1800	return false;
1801	}
1802	}
1803
1804	/**
1805	* unplug_oldest_pwq - unplug the oldest pool_workqueue
1806	* @wq: workqueue_struct where its oldest pwq is to be unplugged
1807	*
1808	* This function should only be called for ordered workqueues where only the
1809	* oldest pwq is unplugged, the others are plugged to suspend execution to
1810	* ensure proper work item ordering::
1811	*
1812	* dfl_pwq --------------+ [P] - plugged
1813	* \|
1814	* v
1815	* pwqs -> A -> B [P] -> C [P] (newest)
1816	* \| \| \|
1817	* 1 3 5
1818	* \| \| \|
1819	* 2 4 6
1820	*
1821	* When the oldest pwq is drained and removed, this function should be called
1822	* to unplug the next oldest one to start its work item execution. Note that
1823	* pwq's are linked into wq->pwqs with the oldest first, so the first one in
1824	* the list is the oldest.
1825	*/
1826	static void unplug_oldest_pwq(struct workqueue_struct *wq)
1827	{
1828	struct pool_workqueue *pwq;
1829
1830	lockdep_assert_held(&wq->mutex);
1831
1832	/* Caller should make sure that pwqs isn't empty before calling */
1833	pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue,
1834	pwqs_node);
1835	raw_spin_lock_irq(&pwq->pool->lock);
1836	if (pwq->plugged) {
1837	pwq->plugged = false;
1838	if (pwq_activate_first_inactive(pwq, true))
1839	kick_pool(pwq->pool);
1840	}
1841	raw_spin_unlock_irq(&pwq->pool->lock);
1842	}
1843
1844	/**
1845	* node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1846	* @nna: wq_node_nr_active to activate a pending pwq for
1847	* @caller_pool: worker_pool the caller is locking
1848	*
1849	* Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1850	* @caller_pool may be unlocked and relocked to lock other worker_pools.
1851	*/
1852	static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1853	struct worker_pool *caller_pool)
1854	{
1855	struct worker_pool *locked_pool = caller_pool;
1856	struct pool_workqueue *pwq;
1857	struct work_struct *work;
1858
1859	lockdep_assert_held(&caller_pool->lock);
1860
1861	raw_spin_lock(&nna->lock);
1862	retry:
1863	pwq = list_first_entry_or_null(&nna->pending_pwqs,
1864	struct pool_workqueue, pending_node);
1865	if (!pwq)
1866	goto out_unlock;
1867
1868	/*
1869	* If @pwq is for a different pool than @locked_pool, we need to lock
1870	* @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1871	* / lock dance. For that, we also need to release @nna->lock as it's
1872	* nested inside pool locks.
1873	*/
1874	if (pwq->pool != locked_pool) {
1875	raw_spin_unlock(&locked_pool->lock);
1876	locked_pool = pwq->pool;
1877	if (!raw_spin_trylock(&locked_pool->lock)) {
1878	raw_spin_unlock(&nna->lock);
1879	raw_spin_lock(&locked_pool->lock);
1880	raw_spin_lock(&nna->lock);
1881	goto retry;
1882	}
1883	}
1884
1885	/*
1886	* $pwq may not have any inactive work items due to e.g. cancellations.
1887	* Drop it from pending_pwqs and see if there's another one.
1888	*/
1889	work = list_first_entry_or_null(&pwq->inactive_works,
1890	struct work_struct, entry);
1891	if (!work) {
1892	list_del_init(&pwq->pending_node);
1893	goto retry;
1894	}
1895
1896	/*
1897	* Acquire an nr_active count and activate the inactive work item. If
1898	* $pwq still has inactive work items, rotate it to the end of the
1899	* pending_pwqs so that we round-robin through them. This means that
1900	* inactive work items are not activated in queueing order which is fine
1901	* given that there has never been any ordering across different pwqs.
1902	*/
1903	if (likely(tryinc_node_nr_active(nna))) {
1904	pwq->nr_active++;
1905	__pwq_activate_work(pwq, work);
1906
1907	if (list_empty(&pwq->inactive_works))
1908	list_del_init(&pwq->pending_node);
1909	else
1910	list_move_tail(&pwq->pending_node, &nna->pending_pwqs);
1911
1912	/* if activating a foreign pool, make sure it's running */
1913	if (pwq->pool != caller_pool)
1914	kick_pool(pwq->pool);
1915	}
1916
1917	out_unlock:
1918	raw_spin_unlock(&nna->lock);
1919	if (locked_pool != caller_pool) {
1920	raw_spin_unlock(&locked_pool->lock);
1921	raw_spin_lock(&caller_pool->lock);
1922	}
1923	}
1924
1925	/**
1926	* pwq_dec_nr_active - Retire an active count
1927	* @pwq: pool_workqueue of interest
1928	*
1929	* Decrement @pwq's nr_active and try to activate the first inactive work item.
1930	* For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
1931	*/
1932	static void pwq_dec_nr_active(struct pool_workqueue *pwq)
1933	{
1934	struct worker_pool *pool = pwq->pool;
1935	struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node);
1936
1937	lockdep_assert_held(&pool->lock);
1938
1939	/*
1940	* @pwq->nr_active should be decremented for both percpu and unbound
1941	* workqueues.
1942	*/
1943	pwq->nr_active--;
1944
1945	/*
1946	* For a percpu workqueue, it's simple. Just need to kick the first
1947	* inactive work item on @pwq itself.
1948	*/
1949	if (!nna) {
1950	pwq_activate_first_inactive(pwq, false);
1951	return;
1952	}
1953
1954	/*
1955	* If @pwq is for an unbound workqueue, it's more complicated because
1956	* multiple pwqs and pools may be sharing the nr_active count. When a
1957	* pwq needs to wait for an nr_active count, it puts itself on
1958	* $nna->pending_pwqs. The following atomic_dec_return()'s implied
1959	* memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
1960	* guarantee that either we see non-empty pending_pwqs or they see
1961	* decremented $nna->nr.
1962	*
1963	* $nna->max may change as CPUs come online/offline and @pwq->wq's
1964	* max_active gets updated. However, it is guaranteed to be equal to or
1965	* larger than @pwq->wq->min_active which is above zero unless freezing.
1966	* This maintains the forward progress guarantee.
1967	*/
1968	if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max))
1969	return;
1970
1971	if (!list_empty(&nna->pending_pwqs))
1972	node_activate_pending_pwq(nna, pool);
1973	}
1974
1975	/**
1976	* pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1977	* @pwq: pwq of interest
1978	* @work_data: work_data of work which left the queue
1979	*
1980	* A work either has completed or is removed from pending queue,
1981	* decrement nr_in_flight of its pwq and handle workqueue flushing.
1982	*
1983	* NOTE:
1984	* For unbound workqueues, this function may temporarily drop @pwq->pool->lock
1985	* and thus should be called after all other state updates for the in-flight
1986	* work item is complete.
1987	*
1988	* CONTEXT:
1989	* raw_spin_lock_irq(pool->lock).
1990	*/
1991	static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1992	{
1993	int color = get_work_color(work_data);
1994
1995	if (!(work_data & WORK_STRUCT_INACTIVE))
1996	pwq_dec_nr_active(pwq);
1997
1998	pwq->nr_in_flight[color]--;
1999
2000	/* is flush in progress and are we at the flushing tip? */
2001	if (likely(pwq->flush_color != color))
2002	goto out_put;
2003
2004	/* are there still in-flight works? */
2005	if (pwq->nr_in_flight[color])
2006	goto out_put;
2007
2008	/* this pwq is done, clear flush_color */
2009	pwq->flush_color = -1;
2010
2011	/*
2012	* If this was the last pwq, wake up the first flusher. It
2013	* will handle the rest.
2014	*/
2015	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2016	complete(&pwq->wq->first_flusher->done);
2017	out_put:
2018	put_pwq(pwq);
2019	}
2020
2021	/**
2022	* try_to_grab_pending - steal work item from worklist and disable irq
2023	* @work: work item to steal
2024	* @cflags: %WORK_CANCEL_ flags
2025	* @irq_flags: place to store irq state
2026	*
2027	* Try to grab PENDING bit of @work. This function can handle @work in any
2028	* stable state - idle, on timer or on worklist.
2029	*
2030	* Return:
2031	*
2032	* ======== ================================================================
2033	* 1 if @work was pending and we successfully stole PENDING
2034	* 0 if @work was idle and we claimed PENDING
2035	* -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
2036	* ======== ================================================================
2037	*
2038	* Note:
2039	* On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2040	* interrupted while holding PENDING and @work off queue, irq must be
2041	* disabled on entry. This, combined with delayed_work->timer being
2042	* irqsafe, ensures that we return -EAGAIN for finite short period of time.
2043	*
2044	* On successful return, >= 0, irq is disabled and the caller is
2045	* responsible for releasing it using local_irq_restore(*@irq_flags).
2046	*
2047	* This function is safe to call from any context including IRQ handler.
2048	*/
2049	static int try_to_grab_pending(struct work_struct *work, u32 cflags,
2050	unsigned long *irq_flags)
2051	{
2052	struct worker_pool *pool;
2053	struct pool_workqueue *pwq;
2054
2055	local_irq_save(*irq_flags);
2056
2057	/* try to steal the timer if it exists */
2058	if (cflags & WORK_CANCEL_DELAYED) {
2059	struct delayed_work *dwork = to_delayed_work(work);
2060
2061	/*
2062	* dwork->timer is irqsafe. If del_timer() fails, it's
2063	* guaranteed that the timer is not queued anywhere and not
2064	* running on the local CPU.
2065	*/
2066	if (likely(del_timer(&dwork->timer)))
2067	return 1;
2068	}
2069
2070	/* try to claim PENDING the normal way */
2071	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2072	return 0;
2073
2074	rcu_read_lock();
2075	/*
2076	* The queueing is in progress, or it is already queued. Try to
2077	* steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2078	*/
2079	pool = get_work_pool(work);
2080	if (!pool)
2081	goto fail;
2082
2083	raw_spin_lock(&pool->lock);
2084	/*
2085	* work->data is guaranteed to point to pwq only while the work
2086	* item is queued on pwq->wq, and both updating work->data to point
2087	* to pwq on queueing and to pool on dequeueing are done under
2088	* pwq->pool->lock. This in turn guarantees that, if work->data
2089	* points to pwq which is associated with a locked pool, the work
2090	* item is currently queued on that pool.
2091	*/
2092	pwq = get_work_pwq(work);
2093	if (pwq && pwq->pool == pool) {
2094	unsigned long work_data = *work_data_bits(work);
2095
2096	debug_work_deactivate(work);
2097
2098	/*
2099	* A cancelable inactive work item must be in the
2100	* pwq->inactive_works since a queued barrier can't be
2101	* canceled (see the comments in insert_wq_barrier()).
2102	*
2103	* An inactive work item cannot be deleted directly because
2104	* it might have linked barrier work items which, if left
2105	* on the inactive_works list, will confuse pwq->nr_active
2106	* management later on and cause stall. Move the linked
2107	* barrier work items to the worklist when deleting the grabbed
2108	* item. Also keep WORK_STRUCT_INACTIVE in work_data, so that
2109	* it doesn't participate in nr_active management in later
2110	* pwq_dec_nr_in_flight().
2111	*/
2112	if (work_data & WORK_STRUCT_INACTIVE)
2113	move_linked_works(work, &pwq->pool->worklist, NULL);
2114
2115	list_del_init(&work->entry);
2116
2117	/*
2118	* work->data points to pwq iff queued. Let's point to pool. As
2119	* this destroys work->data needed by the next step, stash it.
2120	*/
2121	set_work_pool_and_keep_pending(work, pool->id,
2122	pool_offq_flags(pool));
2123
2124	/* must be the last step, see the function comment */
2125	pwq_dec_nr_in_flight(pwq, work_data);
2126
2127	raw_spin_unlock(&pool->lock);
2128	rcu_read_unlock();
2129	return 1;
2130	}
2131	raw_spin_unlock(&pool->lock);
2132	fail:
2133	rcu_read_unlock();
2134	local_irq_restore(*irq_flags);
2135	return -EAGAIN;
2136	}
2137
2138	/**
2139	* work_grab_pending - steal work item from worklist and disable irq
2140	* @work: work item to steal
2141	* @cflags: %WORK_CANCEL_ flags
2142	* @irq_flags: place to store IRQ state
2143	*
2144	* Grab PENDING bit of @work. @work can be in any stable state - idle, on timer
2145	* or on worklist.
2146	*
2147	* Can be called from any context. IRQ is disabled on return with IRQ state
2148	* stored in *@irq_flags. The caller is responsible for re-enabling it using
2149	* local_irq_restore().
2150	*
2151	* Returns %true if @work was pending. %false if idle.
2152	*/
2153	static bool work_grab_pending(struct work_struct *work, u32 cflags,
2154	unsigned long *irq_flags)
2155	{
2156	int ret;
2157
2158	while (true) {
2159	ret = try_to_grab_pending(work, cflags, irq_flags);
2160	if (ret >= 0)
2161	return ret;
2162	cpu_relax();
2163	}
2164	}
2165
2166	/**
2167	* insert_work - insert a work into a pool
2168	* @pwq: pwq @work belongs to
2169	* @work: work to insert
2170	* @head: insertion point
2171	* @extra_flags: extra WORK_STRUCT_* flags to set
2172	*
2173	* Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2174	* work_struct flags.
2175	*
2176	* CONTEXT:
2177	* raw_spin_lock_irq(pool->lock).
2178	*/
2179	static void insert_work(struct pool_workqueue pwq, struct work_struct work,
2180	struct list_head *head, unsigned int extra_flags)
2181	{
2182	debug_work_activate(work);
2183
2184	/* record the work call stack in order to print it in KASAN reports */
2185	kasan_record_aux_stack_noalloc(work);
2186
2187	/* we own @work, set data and link */
2188	set_work_pwq(work, pwq, extra_flags);
2189	list_add_tail(&work->entry, head);
2190	get_pwq(pwq);
2191	}
2192
2193	/*
2194	* Test whether @work is being queued from another work executing on the
2195	* same workqueue.
2196	*/
2197	static bool is_chained_work(struct workqueue_struct *wq)
2198	{
2199	struct worker *worker;
2200
2201	worker = current_wq_worker();
2202	/*
2203	* Return %true iff I'm a worker executing a work item on @wq. If
2204	* I'm @worker, it's safe to dereference it without locking.
2205	*/
2206	return worker && worker->current_pwq->wq == wq;
2207	}
2208
2209	/*
2210	* When queueing an unbound work item to a wq, prefer local CPU if allowed
2211	* by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2212	* avoid perturbing sensitive tasks.
2213	*/
2214	static int wq_select_unbound_cpu(int cpu)
2215	{
2216	int new_cpu;
2217
2218	if (likely(!wq_debug_force_rr_cpu)) {
2219	if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2220	return cpu;
2221	} else {
2222	pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2223	}
2224
2225	new_cpu = __this_cpu_read(wq_rr_cpu_last);
2226	new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2227	if (unlikely(new_cpu >= nr_cpu_ids)) {
2228	new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2229	if (unlikely(new_cpu >= nr_cpu_ids))
2230	return cpu;
2231	}
2232	__this_cpu_write(wq_rr_cpu_last, new_cpu);
2233
2234	return new_cpu;
2235	}
2236
2237	static void __queue_work(int cpu, struct workqueue_struct *wq,
2238	struct work_struct *work)
2239	{
2240	struct pool_workqueue *pwq;
2241	struct worker_pool last_pool, pool;
2242	unsigned int work_flags;
2243	unsigned int req_cpu = cpu;
2244
2245	/*
2246	* While a work item is PENDING && off queue, a task trying to
2247	* steal the PENDING will busy-loop waiting for it to either get
2248	* queued or lose PENDING. Grabbing PENDING and queueing should
2249	* happen with IRQ disabled.
2250	*/
2251	lockdep_assert_irqs_disabled();
2252
2253	/*
2254	* For a draining wq, only works from the same workqueue are
2255	* allowed. The __WQ_DESTROYING helps to spot the issue that
2256	* queues a new work item to a wq after destroy_workqueue(wq).
2257	*/
2258	if (unlikely(wq->flags & (__WQ_DESTROYING \| __WQ_DRAINING) &&
2259	WARN_ON_ONCE(!is_chained_work(wq))))
2260	return;
2261	rcu_read_lock();
2262	retry:
2263	/* pwq which will be used unless @work is executing elsewhere */
2264	if (req_cpu == WORK_CPU_UNBOUND) {
2265	if (wq->flags & WQ_UNBOUND)
2266	cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2267	else
2268	cpu = raw_smp_processor_id();
2269	}
2270
2271	pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2272	pool = pwq->pool;
2273
2274	/*
2275	* If @work was previously on a different pool, it might still be
2276	* running there, in which case the work needs to be queued on that
2277	* pool to guarantee non-reentrancy.
2278	*
2279	* For ordered workqueue, work items must be queued on the newest pwq
2280	* for accurate order management. Guaranteed order also guarantees
2281	* non-reentrancy. See the comments above unplug_oldest_pwq().
2282	*/
2283	last_pool = get_work_pool(work);
2284	if (last_pool && last_pool != pool && !(wq->flags & __WQ_ORDERED)) {
2285	struct worker *worker;
2286
2287	raw_spin_lock(&last_pool->lock);
2288
2289	worker = find_worker_executing_work(last_pool, work);
2290
2291	if (worker && worker->current_pwq->wq == wq) {
2292	pwq = worker->current_pwq;
2293	pool = pwq->pool;
2294	WARN_ON_ONCE(pool != last_pool);
2295	} else {
2296	/* meh... not running there, queue here */
2297	raw_spin_unlock(&last_pool->lock);
2298	raw_spin_lock(&pool->lock);
2299	}
2300	} else {
2301	raw_spin_lock(&pool->lock);
2302	}
2303
2304	/*
2305	* pwq is determined and locked. For unbound pools, we could have raced
2306	* with pwq release and it could already be dead. If its refcnt is zero,
2307	* repeat pwq selection. Note that unbound pwqs never die without
2308	* another pwq replacing it in cpu_pwq or while work items are executing
2309	* on it, so the retrying is guaranteed to make forward-progress.
2310	*/
2311	if (unlikely(!pwq->refcnt)) {
2312	if (wq->flags & WQ_UNBOUND) {
2313	raw_spin_unlock(&pool->lock);
2314	cpu_relax();
2315	goto retry;
2316	}
2317	/* oops */
2318	WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2319	wq->name, cpu);
2320	}
2321
2322	/* pwq determined, queue */
2323	trace_workqueue_queue_work(req_cpu, pwq, work);
2324
2325	if (WARN_ON(!list_empty(&work->entry)))
2326	goto out;
2327
2328	pwq->nr_in_flight[pwq->work_color]++;
2329	work_flags = work_color_to_flags(pwq->work_color);
2330
2331	/*
2332	* Limit the number of concurrently active work items to max_active.
2333	* @work must also queue behind existing inactive work items to maintain
2334	* ordering when max_active changes. See wq_adjust_max_active().
2335	*/
2336	if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) {
2337	if (list_empty(&pool->worklist))
2338	pool->watchdog_ts = jiffies;
2339
2340	trace_workqueue_activate_work(work);
2341	insert_work(pwq, work, &pool->worklist, work_flags);
2342	kick_pool(pool);
2343	} else {
2344	work_flags \|= WORK_STRUCT_INACTIVE;
2345	insert_work(pwq, work, &pwq->inactive_works, work_flags);
2346	}
2347
2348	out:
2349	raw_spin_unlock(&pool->lock);
2350	rcu_read_unlock();
2351	}
2352
2353	static bool clear_pending_if_disabled(struct work_struct *work)
2354	{
2355	unsigned long data = *work_data_bits(work);
2356	struct work_offq_data offqd;
2357
2358	if (likely((data & WORK_STRUCT_PWQ) \|\|
2359	!(data & WORK_OFFQ_DISABLE_MASK)))
2360	return false;
2361
2362	work_offqd_unpack(&offqd, data);
2363	set_work_pool_and_clear_pending(work, offqd.pool_id,
2364	work_offqd_pack_flags(&offqd));
2365	return true;
2366	}
2367
2368	/**
2369	* queue_work_on - queue work on specific cpu
2370	* @cpu: CPU number to execute work on
2371	* @wq: workqueue to use
2372	* @work: work to queue
2373	*
2374	* We queue the work to a specific CPU, the caller must ensure it
2375	* can't go away. Callers that fail to ensure that the specified
2376	* CPU cannot go away will execute on a randomly chosen CPU.
2377	* But note well that callers specifying a CPU that never has been
2378	* online will get a splat.
2379	*
2380	* Return: %false if @work was already on a queue, %true otherwise.
2381	*/
2382	bool queue_work_on(int cpu, struct workqueue_struct *wq,
2383	struct work_struct *work)
2384	{
2385	bool ret = false;
2386	unsigned long irq_flags;
2387
2388	local_irq_save(irq_flags);
2389
2390	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2391	!clear_pending_if_disabled(work)) {
2392	__queue_work(cpu, wq, work);
2393	ret = true;
2394	}
2395
2396	local_irq_restore(irq_flags);
2397	return ret;
2398	}
2399	EXPORT_SYMBOL(queue_work_on);
2400
2401	/**
2402	* select_numa_node_cpu - Select a CPU based on NUMA node
2403	* @node: NUMA node ID that we want to select a CPU from
2404	*
2405	* This function will attempt to find a "random" cpu available on a given
2406	* node. If there are no CPUs available on the given node it will return
2407	* WORK_CPU_UNBOUND indicating that we should just schedule to any
2408	* available CPU if we need to schedule this work.
2409	*/
2410	static int select_numa_node_cpu(int node)
2411	{
2412	int cpu;
2413
2414	/* Delay binding to CPU if node is not valid or online */
2415	if (node < 0 \|\| node >= MAX_NUMNODES \|\| !node_online(node))
2416	return WORK_CPU_UNBOUND;
2417
2418	/* Use local node/cpu if we are already there */
2419	cpu = raw_smp_processor_id();
2420	if (node == cpu_to_node(cpu))
2421	return cpu;
2422
2423	/* Use "random" otherwise know as "first" online CPU of node */
2424	cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2425
2426	/* If CPU is valid return that, otherwise just defer */
2427	return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2428	}
2429
2430	/**
2431	* queue_work_node - queue work on a "random" cpu for a given NUMA node
2432	* @node: NUMA node that we are targeting the work for
2433	* @wq: workqueue to use
2434	* @work: work to queue
2435	*
2436	* We queue the work to a "random" CPU within a given NUMA node. The basic
2437	* idea here is to provide a way to somehow associate work with a given
2438	* NUMA node.
2439	*
2440	* This function will only make a best effort attempt at getting this onto
2441	* the right NUMA node. If no node is requested or the requested node is
2442	* offline then we just fall back to standard queue_work behavior.
2443	*
2444	* Currently the "random" CPU ends up being the first available CPU in the
2445	* intersection of cpu_online_mask and the cpumask of the node, unless we
2446	* are running on the node. In that case we just use the current CPU.
2447	*
2448	* Return: %false if @work was already on a queue, %true otherwise.
2449	*/
2450	bool queue_work_node(int node, struct workqueue_struct *wq,
2451	struct work_struct *work)
2452	{
2453	unsigned long irq_flags;
2454	bool ret = false;
2455
2456	/*
2457	* This current implementation is specific to unbound workqueues.
2458	* Specifically we only return the first available CPU for a given
2459	* node instead of cycling through individual CPUs within the node.
2460	*
2461	* If this is used with a per-cpu workqueue then the logic in
2462	* workqueue_select_cpu_near would need to be updated to allow for
2463	* some round robin type logic.
2464	*/
2465	WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2466
2467	local_irq_save(irq_flags);
2468
2469	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2470	!clear_pending_if_disabled(work)) {
2471	int cpu = select_numa_node_cpu(node);
2472
2473	__queue_work(cpu, wq, work);
2474	ret = true;
2475	}
2476
2477	local_irq_restore(irq_flags);
2478	return ret;
2479	}
2480	EXPORT_SYMBOL_GPL(queue_work_node);
2481
2482	void delayed_work_timer_fn(struct timer_list *t)
2483	{
2484	struct delayed_work *dwork = from_timer(dwork, t, timer);
2485
2486	/* should have been called from irqsafe timer with irq already off */
2487	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
2488	}
2489	EXPORT_SYMBOL(delayed_work_timer_fn);
2490
2491	static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2492	struct delayed_work *dwork, unsigned long delay)
2493	{
2494	struct timer_list *timer = &dwork->timer;
2495	struct work_struct *work = &dwork->work;
2496
2497	WARN_ON_ONCE(!wq);
2498	WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2499	WARN_ON_ONCE(timer_pending(timer));
2500	WARN_ON_ONCE(!list_empty(&work->entry));
2501
2502	/*
2503	* If @delay is 0, queue @dwork->work immediately. This is for
2504	* both optimization and correctness. The earliest @timer can
2505	* expire is on the closest next tick and delayed_work users depend
2506	* on that there's no such delay when @delay is 0.
2507	*/
2508	if (!delay) {
2509	__queue_work(cpu, wq, &dwork->work);
2510	return;
2511	}
2512
2513	dwork->wq = wq;
2514	dwork->cpu = cpu;
2515	timer->expires = jiffies + delay;
2516
2517	if (housekeeping_enabled(HK_TYPE_TIMER)) {
2518	/* If the current cpu is a housekeeping cpu, use it. */
2519	cpu = smp_processor_id();
2520	if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER))
2521	cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
2522	add_timer_on(timer, cpu);
2523	} else {
2524	if (likely(cpu == WORK_CPU_UNBOUND))
2525	add_timer_global(timer);
2526	else
2527	add_timer_on(timer, cpu);
2528	}
2529	}
2530
2531	/**
2532	* queue_delayed_work_on - queue work on specific CPU after delay
2533	* @cpu: CPU number to execute work on
2534	* @wq: workqueue to use
2535	* @dwork: work to queue
2536	* @delay: number of jiffies to wait before queueing
2537	*
2538	* Return: %false if @work was already on a queue, %true otherwise. If
2539	* @delay is zero and @dwork is idle, it will be scheduled for immediate
2540	* execution.
2541	*/
2542	bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2543	struct delayed_work *dwork, unsigned long delay)
2544	{
2545	struct work_struct *work = &dwork->work;
2546	bool ret = false;
2547	unsigned long irq_flags;
2548
2549	/* read the comment in __queue_work() */
2550	local_irq_save(irq_flags);
2551
2552	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2553	!clear_pending_if_disabled(work)) {
2554	__queue_delayed_work(cpu, wq, dwork, delay);
2555	ret = true;
2556	}
2557
2558	local_irq_restore(irq_flags);
2559	return ret;
2560	}
2561	EXPORT_SYMBOL(queue_delayed_work_on);
2562
2563	/**
2564	* mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2565	* @cpu: CPU number to execute work on
2566	* @wq: workqueue to use
2567	* @dwork: work to queue
2568	* @delay: number of jiffies to wait before queueing
2569	*
2570	* If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2571	* modify @dwork's timer so that it expires after @delay. If @delay is
2572	* zero, @work is guaranteed to be scheduled immediately regardless of its
2573	* current state.
2574	*
2575	* Return: %false if @dwork was idle and queued, %true if @dwork was
2576	* pending and its timer was modified.
2577	*
2578	* This function is safe to call from any context including IRQ handler.
2579	* See try_to_grab_pending() for details.
2580	*/
2581	bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2582	struct delayed_work *dwork, unsigned long delay)
2583	{
2584	unsigned long irq_flags;
2585	bool ret;
2586
2587	ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags);
2588
2589	if (!clear_pending_if_disabled(&dwork->work))
2590	__queue_delayed_work(cpu, wq, dwork, delay);
2591
2592	local_irq_restore(irq_flags);
2593	return ret;
2594	}
2595	EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2596
2597	static void rcu_work_rcufn(struct rcu_head *rcu)
2598	{
2599	struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2600
2601	/* read the comment in __queue_work() */
2602	local_irq_disable();
2603	__queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2604	local_irq_enable();
2605	}
2606
2607	/**
2608	* queue_rcu_work - queue work after a RCU grace period
2609	* @wq: workqueue to use
2610	* @rwork: work to queue
2611	*
2612	* Return: %false if @rwork was already pending, %true otherwise. Note
2613	* that a full RCU grace period is guaranteed only after a %true return.
2614	* While @rwork is guaranteed to be executed after a %false return, the
2615	* execution may happen before a full RCU grace period has passed.
2616	*/
2617	bool queue_rcu_work(struct workqueue_struct wq, struct rcu_work rwork)
2618	{
2619	struct work_struct *work = &rwork->work;
2620
2621	/*
2622	* rcu_work can't be canceled or disabled. Warn if the user reached
2623	* inside @rwork and disabled the inner work.
2624	*/
2625	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2626	!WARN_ON_ONCE(clear_pending_if_disabled(work))) {
2627	rwork->wq = wq;
2628	call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2629	return true;
2630	}
2631
2632	return false;
2633	}
2634	EXPORT_SYMBOL(queue_rcu_work);
2635
2636	static struct worker *alloc_worker(int node)
2637	{
2638	struct worker *worker;
2639
2640	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2641	if (worker) {
2642	INIT_LIST_HEAD(&worker->entry);
2643	INIT_LIST_HEAD(&worker->scheduled);
2644	INIT_LIST_HEAD(&worker->node);
2645	/* on creation a worker is in !idle && prep state */
2646	worker->flags = WORKER_PREP;
2647	}
2648	return worker;
2649	}
2650
2651	static cpumask_t pool_allowed_cpus(struct worker_pool pool)
2652	{
2653	if (pool->cpu < 0 && pool->attrs->affn_strict)
2654	return pool->attrs->__pod_cpumask;
2655	else
2656	return pool->attrs->cpumask;
2657	}
2658
2659	/**
2660	* worker_attach_to_pool() - attach a worker to a pool
2661	* @worker: worker to be attached
2662	* @pool: the target pool
2663	*
2664	* Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2665	* cpu-binding of @worker are kept coordinated with the pool across
2666	* cpu-[un]hotplugs.
2667	*/
2668	static void worker_attach_to_pool(struct worker *worker,
2669	struct worker_pool *pool)
2670	{
2671	mutex_lock(&wq_pool_attach_mutex);
2672
2673	/*
2674	* The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable
2675	* across this function. See the comments above the flag definition for
2676	* details. BH workers are, while per-CPU, always DISASSOCIATED.
2677	*/
2678	if (pool->flags & POOL_DISASSOCIATED) {
2679	worker->flags \|= WORKER_UNBOUND;
2680	} else {
2681	WARN_ON_ONCE(pool->flags & POOL_BH);
2682	kthread_set_per_cpu(worker->task, pool->cpu);
2683	}
2684
2685	if (worker->rescue_wq)
2686	set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2687
2688	list_add_tail(&worker->node, &pool->workers);
2689	worker->pool = pool;
2690
2691	mutex_unlock(&wq_pool_attach_mutex);
2692	}
2693
2694	static void unbind_worker(struct worker *worker)
2695	{
2696	lockdep_assert_held(&wq_pool_attach_mutex);
2697
2698	kthread_set_per_cpu(worker->task, -1);
2699	if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2700	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2701	else
2702	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2703	}
2704
2705
2706	static void detach_worker(struct worker *worker)
2707	{
2708	lockdep_assert_held(&wq_pool_attach_mutex);
2709
2710	unbind_worker(worker);
2711	list_del(&worker->node);
2712	worker->pool = NULL;
2713	}
2714
2715	/**
2716	* worker_detach_from_pool() - detach a worker from its pool
2717	* @worker: worker which is attached to its pool
2718	*
2719	* Undo the attaching which had been done in worker_attach_to_pool(). The
2720	* caller worker shouldn't access to the pool after detached except it has
2721	* other reference to the pool.
2722	*/
2723	static void worker_detach_from_pool(struct worker *worker)
2724	{
2725	struct worker_pool *pool = worker->pool;
2726
2727	/* there is one permanent BH worker per CPU which should never detach */
2728	WARN_ON_ONCE(pool->flags & POOL_BH);
2729
2730	mutex_lock(&wq_pool_attach_mutex);
2731	detach_worker(worker);
2732	mutex_unlock(&wq_pool_attach_mutex);
2733
2734	/* clear leftover flags without pool->lock after it is detached */
2735	worker->flags &= ~(WORKER_UNBOUND \| WORKER_REBOUND);
2736	}
2737
2738	static int format_worker_id(char buf, size_t size, struct worker worker,
2739	struct worker_pool *pool)
2740	{
2741	if (worker->rescue_wq)
2742	return scnprintf(buf, size, "kworker/R-%s",
2743	worker->rescue_wq->name);
2744
2745	if (pool) {
2746	if (pool->cpu >= 0)
2747	return scnprintf(buf, size, "kworker/%d:%d%s",
2748	pool->cpu, worker->id,
2749	pool->attrs->nice < 0 ? "H" : "");
2750	else
2751	return scnprintf(buf, size, "kworker/u%d:%d",
2752	pool->id, worker->id);
2753	} else {
2754	return scnprintf(buf, size, "kworker/dying");
2755	}
2756	}
2757
2758	/**
2759	* create_worker - create a new workqueue worker
2760	* @pool: pool the new worker will belong to
2761	*
2762	* Create and start a new worker which is attached to @pool.
2763	*
2764	* CONTEXT:
2765	* Might sleep. Does GFP_KERNEL allocations.
2766	*
2767	* Return:
2768	* Pointer to the newly created worker.
2769	*/
2770	static struct worker create_worker(struct worker_pool pool)
2771	{
2772	struct worker *worker;
2773	int id;
2774
2775	/* ID is needed to determine kthread name */
2776	id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2777	if (id < 0) {
2778	pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2779	ERR_PTR(id));
2780	return NULL;
2781	}
2782
2783	worker = alloc_worker(pool->node);
2784	if (!worker) {
2785	pr_err_once("workqueue: Failed to allocate a worker\n");
2786	goto fail;
2787	}
2788
2789	worker->id = id;
2790
2791	if (!(pool->flags & POOL_BH)) {
2792	char id_buf[WORKER_ID_LEN];
2793
2794	format_worker_id(id_buf, sizeof(id_buf), worker, pool);
2795	worker->task = kthread_create_on_node(worker_thread, worker,
2796	pool->node, "%s", id_buf);
2797	if (IS_ERR(worker->task)) {
2798	if (PTR_ERR(worker->task) == -EINTR) {
2799	pr_err("workqueue: Interrupted when creating a worker thread \"%s\"\n",
2800	id_buf);
2801	} else {
2802	pr_err_once("workqueue: Failed to create a worker thread: %pe",
2803	worker->task);
2804	}
2805	goto fail;
2806	}
2807
2808	set_user_nice(worker->task, pool->attrs->nice);
2809	kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2810	}
2811
2812	/* successful, attach the worker to the pool */
2813	worker_attach_to_pool(worker, pool);
2814
2815	/* start the newly created worker */
2816	raw_spin_lock_irq(&pool->lock);
2817
2818	worker->pool->nr_workers++;
2819	worker_enter_idle(worker);
2820
2821	/*
2822	* @worker is waiting on a completion in kthread() and will trigger hung
2823	* check if not woken up soon. As kick_pool() is noop if @pool is empty,
2824	* wake it up explicitly.
2825	*/
2826	if (worker->task)
2827	wake_up_process(worker->task);
2828
2829	raw_spin_unlock_irq(&pool->lock);
2830
2831	return worker;
2832
2833	fail:
2834	ida_free(&pool->worker_ida, id);
2835	kfree(worker);
2836	return NULL;
2837	}
2838
2839	static void detach_dying_workers(struct list_head *cull_list)
2840	{
2841	struct worker *worker;
2842
2843	list_for_each_entry(worker, cull_list, entry)
2844	detach_worker(worker);
2845	}
2846
2847	static void reap_dying_workers(struct list_head *cull_list)
2848	{
2849	struct worker worker, tmp;
2850
2851	list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2852	list_del_init(&worker->entry);
2853	kthread_stop_put(worker->task);
2854	kfree(worker);
2855	}
2856	}
2857
2858	/**
2859	* set_worker_dying - Tag a worker for destruction
2860	* @worker: worker to be destroyed
2861	* @list: transfer worker away from its pool->idle_list and into list
2862	*
2863	* Tag @worker for destruction and adjust @pool stats accordingly. The worker
2864	* should be idle.
2865	*
2866	* CONTEXT:
2867	* raw_spin_lock_irq(pool->lock).
2868	*/
2869	static void set_worker_dying(struct worker worker, struct list_head list)
2870	{
2871	struct worker_pool *pool = worker->pool;
2872
2873	lockdep_assert_held(&pool->lock);
2874	lockdep_assert_held(&wq_pool_attach_mutex);
2875
2876	/* sanity check frenzy */
2877	if (WARN_ON(worker->current_work) \|\|
2878	WARN_ON(!list_empty(&worker->scheduled)) \|\|
2879	WARN_ON(!(worker->flags & WORKER_IDLE)))
2880	return;
2881
2882	pool->nr_workers--;
2883	pool->nr_idle--;
2884
2885	worker->flags \|= WORKER_DIE;
2886
2887	list_move(&worker->entry, list);
2888
2889	/* get an extra task struct reference for later kthread_stop_put() */
2890	get_task_struct(worker->task);
2891	}
2892
2893	/**
2894	* idle_worker_timeout - check if some idle workers can now be deleted.
2895	* @t: The pool's idle_timer that just expired
2896	*
2897	* The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2898	* worker_leave_idle(), as a worker flicking between idle and active while its
2899	* pool is at the too_many_workers() tipping point would cause too much timer
2900	* housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2901	* it expire and re-evaluate things from there.
2902	*/
2903	static void idle_worker_timeout(struct timer_list *t)
2904	{
2905	struct worker_pool *pool = from_timer(pool, t, idle_timer);
2906	bool do_cull = false;
2907
2908	if (work_pending(&pool->idle_cull_work))
2909	return;
2910
2911	raw_spin_lock_irq(&pool->lock);
2912
2913	if (too_many_workers(pool)) {
2914	struct worker *worker;
2915	unsigned long expires;
2916
2917	/* idle_list is kept in LIFO order, check the last one */
2918	worker = list_last_entry(&pool->idle_list, struct worker, entry);
2919	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2920	do_cull = !time_before(jiffies, expires);
2921
2922	if (!do_cull)
2923	mod_timer(&pool->idle_timer, expires);
2924	}
2925	raw_spin_unlock_irq(&pool->lock);
2926
2927	if (do_cull)
2928	queue_work(system_unbound_wq, &pool->idle_cull_work);
2929	}
2930
2931	/**
2932	* idle_cull_fn - cull workers that have been idle for too long.
2933	* @work: the pool's work for handling these idle workers
2934	*
2935	* This goes through a pool's idle workers and gets rid of those that have been
2936	* idle for at least IDLE_WORKER_TIMEOUT seconds.
2937	*
2938	* We don't want to disturb isolated CPUs because of a pcpu kworker being
2939	* culled, so this also resets worker affinity. This requires a sleepable
2940	* context, hence the split between timer callback and work item.
2941	*/
2942	static void idle_cull_fn(struct work_struct *work)
2943	{
2944	struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2945	LIST_HEAD(cull_list);
2946
2947	/*
2948	* Grabbing wq_pool_attach_mutex here ensures an already-running worker
2949	* cannot proceed beyong set_pf_worker() in its self-destruct path.
2950	* This is required as a previously-preempted worker could run after
2951	* set_worker_dying() has happened but before detach_dying_workers() did.
2952	*/
2953	mutex_lock(&wq_pool_attach_mutex);
2954	raw_spin_lock_irq(&pool->lock);
2955
2956	while (too_many_workers(pool)) {
2957	struct worker *worker;
2958	unsigned long expires;
2959
2960	worker = list_last_entry(&pool->idle_list, struct worker, entry);
2961	expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2962
2963	if (time_before(jiffies, expires)) {
2964	mod_timer(&pool->idle_timer, expires);
2965	break;
2966	}
2967
2968	set_worker_dying(worker, &cull_list);
2969	}
2970
2971	raw_spin_unlock_irq(&pool->lock);
2972	detach_dying_workers(&cull_list);
2973	mutex_unlock(&wq_pool_attach_mutex);
2974
2975	reap_dying_workers(&cull_list);
2976	}
2977
2978	static void send_mayday(struct work_struct *work)
2979	{
2980	struct pool_workqueue *pwq = get_work_pwq(work);
2981	struct workqueue_struct *wq = pwq->wq;
2982
2983	lockdep_assert_held(&wq_mayday_lock);
2984
2985	if (!wq->rescuer)
2986	return;
2987
2988	/* mayday mayday mayday */
2989	if (list_empty(&pwq->mayday_node)) {
2990	/*
2991	* If @pwq is for an unbound wq, its base ref may be put at
2992	* any time due to an attribute change. Pin @pwq until the
2993	* rescuer is done with it.
2994	*/
2995	get_pwq(pwq);
2996	list_add_tail(&pwq->mayday_node, &wq->maydays);
2997	wake_up_process(wq->rescuer->task);
2998	pwq->stats[PWQ_STAT_MAYDAY]++;
2999	}
3000	}
3001
3002	static void pool_mayday_timeout(struct timer_list *t)
3003	{
3004	struct worker_pool *pool = from_timer(pool, t, mayday_timer);
3005	struct work_struct *work;
3006
3007	raw_spin_lock_irq(&pool->lock);
3008	raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
3009
3010	if (need_to_create_worker(pool)) {
3011	/*
3012	* We've been trying to create a new worker but
3013	* haven't been successful. We might be hitting an
3014	* allocation deadlock. Send distress signals to
3015	* rescuers.
3016	*/
3017	list_for_each_entry(work, &pool->worklist, entry)
3018	send_mayday(work);
3019	}
3020
3021	raw_spin_unlock(&wq_mayday_lock);
3022	raw_spin_unlock_irq(&pool->lock);
3023
3024	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3025	}
3026
3027	/**
3028	* maybe_create_worker - create a new worker if necessary
3029	* @pool: pool to create a new worker for
3030	*
3031	* Create a new worker for @pool if necessary. @pool is guaranteed to
3032	* have at least one idle worker on return from this function. If
3033	* creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3034	* sent to all rescuers with works scheduled on @pool to resolve
3035	* possible allocation deadlock.
3036	*
3037	* On return, need_to_create_worker() is guaranteed to be %false and
3038	* may_start_working() %true.
3039	*
3040	* LOCKING:
3041	* raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3042	* multiple times. Does GFP_KERNEL allocations. Called only from
3043	* manager.
3044	*/
3045	static void maybe_create_worker(struct worker_pool *pool)
3046	__releases(&pool->lock)
3047	__acquires(&pool->lock)
3048	{
3049	restart:
3050	raw_spin_unlock_irq(&pool->lock);
3051
3052	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3053	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3054
3055	while (true) {
3056	if (create_worker(pool) \|\| !need_to_create_worker(pool))
3057	break;
3058
3059	schedule_timeout_interruptible(CREATE_COOLDOWN);
3060
3061	if (!need_to_create_worker(pool))
3062	break;
3063	}
3064
3065	del_timer_sync(&pool->mayday_timer);
3066	raw_spin_lock_irq(&pool->lock);
3067	/*
3068	* This is necessary even after a new worker was just successfully
3069	* created as @pool->lock was dropped and the new worker might have
3070	* already become busy.
3071	*/
3072	if (need_to_create_worker(pool))
3073	goto restart;
3074	}
3075
3076	/**
3077	* manage_workers - manage worker pool
3078	* @worker: self
3079	*
3080	* Assume the manager role and manage the worker pool @worker belongs
3081	* to. At any given time, there can be only zero or one manager per
3082	* pool. The exclusion is handled automatically by this function.
3083	*
3084	* The caller can safely start processing works on false return. On
3085	* true return, it's guaranteed that need_to_create_worker() is false
3086	* and may_start_working() is true.
3087	*
3088	* CONTEXT:
3089	* raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3090	* multiple times. Does GFP_KERNEL allocations.
3091	*
3092	* Return:
3093	* %false if the pool doesn't need management and the caller can safely
3094	* start processing works, %true if management function was performed and
3095	* the conditions that the caller verified before calling the function may
3096	* no longer be true.
3097	*/
3098	static bool manage_workers(struct worker *worker)
3099	{
3100	struct worker_pool *pool = worker->pool;
3101
3102	if (pool->flags & POOL_MANAGER_ACTIVE)
3103	return false;
3104
3105	pool->flags \|= POOL_MANAGER_ACTIVE;
3106	pool->manager = worker;
3107
3108	maybe_create_worker(pool);
3109
3110	pool->manager = NULL;
3111	pool->flags &= ~POOL_MANAGER_ACTIVE;
3112	rcuwait_wake_up(&manager_wait);
3113	return true;
3114	}
3115
3116	/**
3117	* process_one_work - process single work
3118	* @worker: self
3119	* @work: work to process
3120	*
3121	* Process @work. This function contains all the logics necessary to
3122	* process a single work including synchronization against and
3123	* interaction with other workers on the same cpu, queueing and
3124	* flushing. As long as context requirement is met, any worker can
3125	* call this function to process a work.
3126	*
3127	* CONTEXT:
3128	* raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3129	*/
3130	static void process_one_work(struct worker worker, struct work_struct work)
3131	__releases(&pool->lock)
3132	__acquires(&pool->lock)
3133	{
3134	struct pool_workqueue *pwq = get_work_pwq(work);
3135	struct worker_pool *pool = worker->pool;
3136	unsigned long work_data;
3137	int lockdep_start_depth, rcu_start_depth;
3138	bool bh_draining = pool->flags & POOL_BH_DRAINING;
3139	#ifdef CONFIG_LOCKDEP
3140	/*
3141	* It is permissible to free the struct work_struct from
3142	* inside the function that is called from it, this we need to
3143	* take into account for lockdep too. To avoid bogus "held
3144	* lock freed" warnings as well as problems when looking into
3145	* work->lockdep_map, make a copy and use that here.
3146	*/
3147	struct lockdep_map lockdep_map;
3148
3149	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3150	#endif
3151	/* ensure we're on the correct CPU */
3152	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3153	raw_smp_processor_id() != pool->cpu);
3154
3155	/* claim and dequeue */
3156	debug_work_deactivate(work);
3157	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3158	worker->current_work = work;
3159	worker->current_func = work->func;
3160	worker->current_pwq = pwq;
3161	if (worker->task)
3162	worker->current_at = worker->task->se.sum_exec_runtime;
3163	work_data = *work_data_bits(work);
3164	worker->current_color = get_work_color(work_data);
3165
3166	/*
3167	* Record wq name for cmdline and debug reporting, may get
3168	* overridden through set_worker_desc().
3169	*/
3170	strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3171
3172	list_del_init(&work->entry);
3173
3174	/*
3175	* CPU intensive works don't participate in concurrency management.
3176	* They're the scheduler's responsibility. This takes @worker out
3177	* of concurrency management and the next code block will chain
3178	* execution of the pending work items.
3179	*/
3180	if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3181	worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3182
3183	/*
3184	* Kick @pool if necessary. It's always noop for per-cpu worker pools
3185	* since nr_running would always be >= 1 at this point. This is used to
3186	* chain execution of the pending work items for WORKER_NOT_RUNNING
3187	* workers such as the UNBOUND and CPU_INTENSIVE ones.
3188	*/
3189	kick_pool(pool);
3190
3191	/*
3192	* Record the last pool and clear PENDING which should be the last
3193	* update to @work. Also, do this inside @pool->lock so that
3194	* PENDING and queued state changes happen together while IRQ is
3195	* disabled.
3196	*/
3197	set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool));
3198
3199	pwq->stats[PWQ_STAT_STARTED]++;
3200	raw_spin_unlock_irq(&pool->lock);
3201
3202	rcu_start_depth = rcu_preempt_depth();
3203	lockdep_start_depth = lockdep_depth(current);
3204	/* see drain_dead_softirq_workfn() */
3205	if (!bh_draining)
3206	lock_map_acquire(&pwq->wq->lockdep_map);
3207	lock_map_acquire(&lockdep_map);
3208	/*
3209	* Strictly speaking we should mark the invariant state without holding
3210	* any locks, that is, before these two lock_map_acquire()'s.
3211	*
3212	* However, that would result in:
3213	*
3214	* A(W1)
3215	* WFC(C)
3216	* A(W1)
3217	* C(C)
3218	*
3219	* Which would create W1->C->W1 dependencies, even though there is no
3220	* actual deadlock possible. There are two solutions, using a
3221	* read-recursive acquire on the work(queue) 'locks', but this will then
3222	* hit the lockdep limitation on recursive locks, or simply discard
3223	* these locks.
3224	*
3225	* AFAICT there is no possible deadlock scenario between the
3226	* flush_work() and complete() primitives (except for single-threaded
3227	* workqueues), so hiding them isn't a problem.
3228	*/
3229	lockdep_invariant_state(true);
3230	trace_workqueue_execute_start(work);
3231	worker->current_func(work);
3232	/*
3233	* While we must be careful to not use "work" after this, the trace
3234	* point will only record its address.
3235	*/
3236	trace_workqueue_execute_end(work, worker->current_func);
3237	pwq->stats[PWQ_STAT_COMPLETED]++;
3238	lock_map_release(&lockdep_map);
3239	if (!bh_draining)
3240	lock_map_release(&pwq->wq->lockdep_map);
3241
3242	if (unlikely((worker->task && in_atomic()) \|\|
3243	lockdep_depth(current) != lockdep_start_depth \|\|
3244	rcu_preempt_depth() != rcu_start_depth)) {
3245	pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n"
3246	" preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n",
3247	current->comm, task_pid_nr(current), preempt_count(),
3248	lockdep_start_depth, lockdep_depth(current),
3249	rcu_start_depth, rcu_preempt_depth(),
3250	worker->current_func);
3251	debug_show_held_locks(current);
3252	dump_stack();
3253	}
3254
3255	/*
3256	* The following prevents a kworker from hogging CPU on !PREEMPTION
3257	* kernels, where a requeueing work item waiting for something to
3258	* happen could deadlock with stop_machine as such work item could
3259	* indefinitely requeue itself while all other CPUs are trapped in
3260	* stop_machine. At the same time, report a quiescent RCU state so
3261	* the same condition doesn't freeze RCU.
3262	*/
3263	if (worker->task)
3264	cond_resched();
3265
3266	raw_spin_lock_irq(&pool->lock);
3267
3268	/*
3269	* In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3270	* CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3271	* wq_cpu_intensive_thresh_us. Clear it.
3272	*/
3273	worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3274
3275	/* tag the worker for identification in schedule() */
3276	worker->last_func = worker->current_func;
3277
3278	/* we're done with it, release */
3279	hash_del(&worker->hentry);
3280	worker->current_work = NULL;
3281	worker->current_func = NULL;
3282	worker->current_pwq = NULL;
3283	worker->current_color = INT_MAX;
3284
3285	/* must be the last step, see the function comment */
3286	pwq_dec_nr_in_flight(pwq, work_data);
3287	}
3288
3289	/**
3290	* process_scheduled_works - process scheduled works
3291	* @worker: self
3292	*
3293	* Process all scheduled works. Please note that the scheduled list
3294	* may change while processing a work, so this function repeatedly
3295	* fetches a work from the top and executes it.
3296	*
3297	* CONTEXT:
3298	* raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3299	* multiple times.
3300	*/
3301	static void process_scheduled_works(struct worker *worker)
3302	{
3303	struct work_struct *work;
3304	bool first = true;
3305
3306	while ((work = list_first_entry_or_null(&worker->scheduled,
3307	struct work_struct, entry))) {
3308	if (first) {
3309	worker->pool->watchdog_ts = jiffies;
3310	first = false;
3311	}
3312	process_one_work(worker, work);
3313	}
3314	}
3315
3316	static void set_pf_worker(bool val)
3317	{
3318	mutex_lock(&wq_pool_attach_mutex);
3319	if (val)
3320	current->flags \|= PF_WQ_WORKER;
3321	else
3322	current->flags &= ~PF_WQ_WORKER;
3323	mutex_unlock(&wq_pool_attach_mutex);
3324	}
3325
3326	/**
3327	* worker_thread - the worker thread function
3328	* @__worker: self
3329	*
3330	* The worker thread function. All workers belong to a worker_pool -
3331	* either a per-cpu one or dynamic unbound one. These workers process all
3332	* work items regardless of their specific target workqueue. The only
3333	* exception is work items which belong to workqueues with a rescuer which
3334	* will be explained in rescuer_thread().
3335	*
3336	* Return: 0
3337	*/
3338	static int worker_thread(void *__worker)
3339	{
3340	struct worker *worker = __worker;
3341	struct worker_pool *pool = worker->pool;
3342
3343	/* tell the scheduler that this is a workqueue worker */
3344	set_pf_worker(true);
3345	woke_up:
3346	raw_spin_lock_irq(&pool->lock);
3347
3348	/* am I supposed to die? */
3349	if (unlikely(worker->flags & WORKER_DIE)) {
3350	raw_spin_unlock_irq(&pool->lock);
3351	set_pf_worker(false);
3352
3353	ida_free(&pool->worker_ida, worker->id);
3354	return 0;
3355	}
3356
3357	worker_leave_idle(worker);
3358	recheck:
3359	/* no more worker necessary? */
3360	if (!need_more_worker(pool))
3361	goto sleep;
3362
3363	/* do we need to manage? */
3364	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3365	goto recheck;
3366
3367	/*
3368	* ->scheduled list can only be filled while a worker is
3369	* preparing to process a work or actually processing it.
3370	* Make sure nobody diddled with it while I was sleeping.
3371	*/
3372	WARN_ON_ONCE(!list_empty(&worker->scheduled));
3373
3374	/*
3375	* Finish PREP stage. We're guaranteed to have at least one idle
3376	* worker or that someone else has already assumed the manager
3377	* role. This is where @worker starts participating in concurrency
3378	* management if applicable and concurrency management is restored
3379	* after being rebound. See rebind_workers() for details.
3380	*/
3381	worker_clr_flags(worker, WORKER_PREP \| WORKER_REBOUND);
3382
3383	do {
3384	struct work_struct *work =
3385	list_first_entry(&pool->worklist,
3386	struct work_struct, entry);
3387
3388	if (assign_work(work, worker, NULL))
3389	process_scheduled_works(worker);
3390	} while (keep_working(pool));
3391
3392	worker_set_flags(worker, WORKER_PREP);
3393	sleep:
3394	/*
3395	* pool->lock is held and there's no work to process and no need to
3396	* manage, sleep. Workers are woken up only while holding
3397	* pool->lock or from local cpu, so setting the current state
3398	* before releasing pool->lock is enough to prevent losing any
3399	* event.
3400	*/
3401	worker_enter_idle(worker);
3402	__set_current_state(TASK_IDLE);
3403	raw_spin_unlock_irq(&pool->lock);
3404	schedule();
3405	goto woke_up;
3406	}
3407
3408	/**
3409	* rescuer_thread - the rescuer thread function
3410	* @__rescuer: self
3411	*
3412	* Workqueue rescuer thread function. There's one rescuer for each
3413	* workqueue which has WQ_MEM_RECLAIM set.
3414	*
3415	* Regular work processing on a pool may block trying to create a new
3416	* worker which uses GFP_KERNEL allocation which has slight chance of
3417	* developing into deadlock if some works currently on the same queue
3418	* need to be processed to satisfy the GFP_KERNEL allocation. This is
3419	* the problem rescuer solves.
3420	*
3421	* When such condition is possible, the pool summons rescuers of all
3422	* workqueues which have works queued on the pool and let them process
3423	* those works so that forward progress can be guaranteed.
3424	*
3425	* This should happen rarely.
3426	*
3427	* Return: 0
3428	*/
3429	static int rescuer_thread(void *__rescuer)
3430	{
3431	struct worker *rescuer = __rescuer;
3432	struct workqueue_struct *wq = rescuer->rescue_wq;
3433	bool should_stop;
3434
3435	set_user_nice(current, RESCUER_NICE_LEVEL);
3436
3437	/*
3438	* Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3439	* doesn't participate in concurrency management.
3440	*/
3441	set_pf_worker(true);
3442	repeat:
3443	set_current_state(TASK_IDLE);
3444
3445	/*
3446	* By the time the rescuer is requested to stop, the workqueue
3447	* shouldn't have any work pending, but @wq->maydays may still have
3448	* pwq(s) queued. This can happen by non-rescuer workers consuming
3449	* all the work items before the rescuer got to them. Go through
3450	* @wq->maydays processing before acting on should_stop so that the
3451	* list is always empty on exit.
3452	*/
3453	should_stop = kthread_should_stop();
3454
3455	/* see whether any pwq is asking for help */
3456	raw_spin_lock_irq(&wq_mayday_lock);
3457
3458	while (!list_empty(&wq->maydays)) {
3459	struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3460	struct pool_workqueue, mayday_node);
3461	struct worker_pool *pool = pwq->pool;
3462	struct work_struct work, n;
3463
3464	__set_current_state(TASK_RUNNING);
3465	list_del_init(&pwq->mayday_node);
3466
3467	raw_spin_unlock_irq(&wq_mayday_lock);
3468
3469	worker_attach_to_pool(rescuer, pool);
3470
3471	raw_spin_lock_irq(&pool->lock);
3472
3473	/*
3474	* Slurp in all works issued via this workqueue and
3475	* process'em.
3476	*/
3477	WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3478	list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3479	if (get_work_pwq(work) == pwq &&
3480	assign_work(work, rescuer, &n))
3481	pwq->stats[PWQ_STAT_RESCUED]++;
3482	}
3483
3484	if (!list_empty(&rescuer->scheduled)) {
3485	process_scheduled_works(rescuer);
3486
3487	/*
3488	* The above execution of rescued work items could
3489	* have created more to rescue through
3490	* pwq_activate_first_inactive() or chained
3491	* queueing. Let's put @pwq back on mayday list so
3492	* that such back-to-back work items, which may be
3493	* being used to relieve memory pressure, don't
3494	* incur MAYDAY_INTERVAL delay inbetween.
3495	*/
3496	if (pwq->nr_active && need_to_create_worker(pool)) {
3497	raw_spin_lock(&wq_mayday_lock);
3498	/*
3499	* Queue iff we aren't racing destruction
3500	* and somebody else hasn't queued it already.
3501	*/
3502	if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3503	get_pwq(pwq);
3504	list_add_tail(&pwq->mayday_node, &wq->maydays);
3505	}
3506	raw_spin_unlock(&wq_mayday_lock);
3507	}
3508	}
3509
3510	/*
3511	* Put the reference grabbed by send_mayday(). @pool won't
3512	* go away while we're still attached to it.
3513	*/
3514	put_pwq(pwq);
3515
3516	/*
3517	* Leave this pool. Notify regular workers; otherwise, we end up
3518	* with 0 concurrency and stalling the execution.
3519	*/
3520	kick_pool(pool);
3521
3522	raw_spin_unlock_irq(&pool->lock);
3523
3524	worker_detach_from_pool(rescuer);
3525
3526	raw_spin_lock_irq(&wq_mayday_lock);
3527	}
3528
3529	raw_spin_unlock_irq(&wq_mayday_lock);
3530
3531	if (should_stop) {
3532	__set_current_state(TASK_RUNNING);
3533	set_pf_worker(false);
3534	return 0;
3535	}
3536
3537	/* rescuers should never participate in concurrency management */
3538	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3539	schedule();
3540	goto repeat;
3541	}
3542
3543	static void bh_worker(struct worker *worker)
3544	{
3545	struct worker_pool *pool = worker->pool;
3546	int nr_restarts = BH_WORKER_RESTARTS;
3547	unsigned long end = jiffies + BH_WORKER_JIFFIES;
3548
3549	raw_spin_lock_irq(&pool->lock);
3550	worker_leave_idle(worker);
3551
3552	/*
3553	* This function follows the structure of worker_thread(). See there for
3554	* explanations on each step.
3555	*/
3556	if (!need_more_worker(pool))
3557	goto done;
3558
3559	WARN_ON_ONCE(!list_empty(&worker->scheduled));
3560	worker_clr_flags(worker, WORKER_PREP \| WORKER_REBOUND);
3561
3562	do {
3563	struct work_struct *work =
3564	list_first_entry(&pool->worklist,
3565	struct work_struct, entry);
3566
3567	if (assign_work(work, worker, NULL))
3568	process_scheduled_works(worker);
3569	} while (keep_working(pool) &&
3570	--nr_restarts && time_before(jiffies, end));
3571
3572	worker_set_flags(worker, WORKER_PREP);
3573	done:
3574	worker_enter_idle(worker);
3575	kick_pool(pool);
3576	raw_spin_unlock_irq(&pool->lock);
3577	}
3578
3579	/*
3580	* TODO: Convert all tasklet users to workqueue and use softirq directly.
3581	*
3582	* This is currently called from tasklet[_hi]action() and thus is also called
3583	* whenever there are tasklets to run. Let's do an early exit if there's nothing
3584	* queued. Once conversion from tasklet is complete, the need_more_worker() test
3585	* can be dropped.
3586	*
3587	* After full conversion, we'll add worker->softirq_action, directly use the
3588	* softirq action and obtain the worker pointer from the softirq_action pointer.
3589	*/
3590	void workqueue_softirq_action(bool highpri)
3591	{
3592	struct worker_pool *pool =
3593	&per_cpu(bh_worker_pools, smp_processor_id())[highpri];
3594	if (need_more_worker(pool))
3595	bh_worker(list_first_entry(&pool->workers, struct worker, node));
3596	}
3597
3598	struct wq_drain_dead_softirq_work {
3599	struct work_struct work;
3600	struct worker_pool *pool;
3601	struct completion done;
3602	};
3603
3604	static void drain_dead_softirq_workfn(struct work_struct *work)
3605	{
3606	struct wq_drain_dead_softirq_work *dead_work =
3607	container_of(work, struct wq_drain_dead_softirq_work, work);
3608	struct worker_pool *pool = dead_work->pool;
3609	bool repeat;
3610
3611	/*
3612	* @pool's CPU is dead and we want to execute its still pending work
3613	* items from this BH work item which is running on a different CPU. As
3614	* its CPU is dead, @pool can't be kicked and, as work execution path
3615	* will be nested, a lockdep annotation needs to be suppressed. Mark
3616	* @pool with %POOL_BH_DRAINING for the special treatments.
3617	*/
3618	raw_spin_lock_irq(&pool->lock);
3619	pool->flags \|= POOL_BH_DRAINING;
3620	raw_spin_unlock_irq(&pool->lock);
3621
3622	bh_worker(list_first_entry(&pool->workers, struct worker, node));
3623
3624	raw_spin_lock_irq(&pool->lock);
3625	pool->flags &= ~POOL_BH_DRAINING;
3626	repeat = need_more_worker(pool);
3627	raw_spin_unlock_irq(&pool->lock);
3628
3629	/*
3630	* bh_worker() might hit consecutive execution limit and bail. If there
3631	* still are pending work items, reschedule self and return so that we
3632	* don't hog this CPU's BH.
3633	*/
3634	if (repeat) {
3635	if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3636	queue_work(system_bh_highpri_wq, work);
3637	else
3638	queue_work(system_bh_wq, work);
3639	} else {
3640	complete(&dead_work->done);
3641	}
3642	}
3643
3644	/*
3645	* @cpu is dead. Drain the remaining BH work items on the current CPU. It's
3646	* possible to allocate dead_work per CPU and avoid flushing. However, then we
3647	* have to worry about draining overlapping with CPU coming back online or
3648	* nesting (one CPU's dead_work queued on another CPU which is also dead and so
3649	* on). Let's keep it simple and drain them synchronously. These are BH work
3650	* items which shouldn't be requeued on the same pool. Shouldn't take long.
3651	*/
3652	void workqueue_softirq_dead(unsigned int cpu)
3653	{
3654	int i;
3655
3656	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
3657	struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i];
3658	struct wq_drain_dead_softirq_work dead_work;
3659
3660	if (!need_more_worker(pool))
3661	continue;
3662
3663	INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn);
3664	dead_work.pool = pool;
3665	init_completion(&dead_work.done);
3666
3667	if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3668	queue_work(system_bh_highpri_wq, &dead_work.work);
3669	else
3670	queue_work(system_bh_wq, &dead_work.work);
3671
3672	wait_for_completion(&dead_work.done);
3673	destroy_work_on_stack(&dead_work.work);
3674	}
3675	}
3676
3677	/**
3678	* check_flush_dependency - check for flush dependency sanity
3679	* @target_wq: workqueue being flushed
3680	* @target_work: work item being flushed (NULL for workqueue flushes)
3681	*
3682	* %current is trying to flush the whole @target_wq or @target_work on it.
3683	* If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3684	* reclaiming memory or running on a workqueue which doesn't have
3685	* %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3686	* a deadlock.
3687	*/
3688	static void check_flush_dependency(struct workqueue_struct *target_wq,
3689	struct work_struct *target_work)
3690	{
3691	work_func_t target_func = target_work ? target_work->func : NULL;
3692	struct worker *worker;
3693
3694	if (target_wq->flags & WQ_MEM_RECLAIM)
3695	return;
3696
3697	worker = current_wq_worker();
3698
3699	WARN_ONCE(current->flags & PF_MEMALLOC,
3700	"workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3701	current->pid, current->comm, target_wq->name, target_func);
3702	WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3703	(WQ_MEM_RECLAIM \| __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3704	"workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3705	worker->current_pwq->wq->name, worker->current_func,
3706	target_wq->name, target_func);
3707	}
3708
3709	struct wq_barrier {
3710	struct work_struct work;
3711	struct completion done;
3712	struct task_struct task; / purely informational */
3713	};
3714
3715	static void wq_barrier_func(struct work_struct *work)
3716	{
3717	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3718	complete(&barr->done);
3719	}
3720
3721	/**
3722	* insert_wq_barrier - insert a barrier work
3723	* @pwq: pwq to insert barrier into
3724	* @barr: wq_barrier to insert
3725	* @target: target work to attach @barr to
3726	* @worker: worker currently executing @target, NULL if @target is not executing
3727	*
3728	* @barr is linked to @target such that @barr is completed only after
3729	* @target finishes execution. Please note that the ordering
3730	* guarantee is observed only with respect to @target and on the local
3731	* cpu.
3732	*
3733	* Currently, a queued barrier can't be canceled. This is because
3734	* try_to_grab_pending() can't determine whether the work to be
3735	* grabbed is at the head of the queue and thus can't clear LINKED
3736	* flag of the previous work while there must be a valid next work
3737	* after a work with LINKED flag set.
3738	*
3739	* Note that when @worker is non-NULL, @target may be modified
3740	* underneath us, so we can't reliably determine pwq from @target.
3741	*
3742	* CONTEXT:
3743	* raw_spin_lock_irq(pool->lock).
3744	*/
3745	static void insert_wq_barrier(struct pool_workqueue *pwq,
3746	struct wq_barrier *barr,
3747	struct work_struct target, struct worker worker)
3748	{
3749	static __maybe_unused struct lock_class_key bh_key, thr_key;
3750	unsigned int work_flags = 0;
3751	unsigned int work_color;
3752	struct list_head *head;
3753
3754	/*
3755	* debugobject calls are safe here even with pool->lock locked
3756	* as we know for sure that this will not trigger any of the
3757	* checks and call back into the fixup functions where we
3758	* might deadlock.
3759	*
3760	* BH and threaded workqueues need separate lockdep keys to avoid
3761	* spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W}
3762	* usage".
3763	*/
3764	INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func,
3765	(pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key);
3766	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3767
3768	init_completion_map(&barr->done, &target->lockdep_map);
3769
3770	barr->task = current;
3771
3772	/* The barrier work item does not participate in nr_active. */
3773	work_flags \|= WORK_STRUCT_INACTIVE;
3774
3775	/*
3776	* If @target is currently being executed, schedule the
3777	* barrier to the worker; otherwise, put it after @target.
3778	*/
3779	if (worker) {
3780	head = worker->scheduled.next;
3781	work_color = worker->current_color;
3782	} else {
3783	unsigned long *bits = work_data_bits(target);
3784
3785	head = target->entry.next;
3786	/* there can already be other linked works, inherit and set */
3787	work_flags \|= *bits & WORK_STRUCT_LINKED;
3788	work_color = get_work_color(*bits);
3789	__set_bit(WORK_STRUCT_LINKED_BIT, bits);
3790	}
3791
3792	pwq->nr_in_flight[work_color]++;
3793	work_flags \|= work_color_to_flags(work_color);
3794
3795	insert_work(pwq, &barr->work, head, work_flags);
3796	}
3797
3798	/**
3799	* flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3800	* @wq: workqueue being flushed
3801	* @flush_color: new flush color, < 0 for no-op
3802	* @work_color: new work color, < 0 for no-op
3803	*
3804	* Prepare pwqs for workqueue flushing.
3805	*
3806	* If @flush_color is non-negative, flush_color on all pwqs should be
3807	* -1. If no pwq has in-flight commands at the specified color, all
3808	* pwq->flush_color's stay at -1 and %false is returned. If any pwq
3809	* has in flight commands, its pwq->flush_color is set to
3810	* @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3811	* wakeup logic is armed and %true is returned.
3812	*
3813	* The caller should have initialized @wq->first_flusher prior to
3814	* calling this function with non-negative @flush_color. If
3815	* @flush_color is negative, no flush color update is done and %false
3816	* is returned.
3817	*
3818	* If @work_color is non-negative, all pwqs should have the same
3819	* work_color which is previous to @work_color and all will be
3820	* advanced to @work_color.
3821	*
3822	* CONTEXT:
3823	* mutex_lock(wq->mutex).
3824	*
3825	* Return:
3826	* %true if @flush_color >= 0 and there's something to flush. %false
3827	* otherwise.
3828	*/
3829	static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3830	int flush_color, int work_color)
3831	{
3832	bool wait = false;
3833	struct pool_workqueue *pwq;
3834
3835	if (flush_color >= 0) {
3836	WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3837	atomic_set(&wq->nr_pwqs_to_flush, 1);
3838	}
3839
3840	for_each_pwq(pwq, wq) {
3841	struct worker_pool *pool = pwq->pool;
3842
3843	raw_spin_lock_irq(&pool->lock);
3844
3845	if (flush_color >= 0) {
3846	WARN_ON_ONCE(pwq->flush_color != -1);
3847
3848	if (pwq->nr_in_flight[flush_color]) {
3849	pwq->flush_color = flush_color;
3850	atomic_inc(&wq->nr_pwqs_to_flush);
3851	wait = true;
3852	}
3853	}
3854
3855	if (work_color >= 0) {
3856	WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3857	pwq->work_color = work_color;
3858	}
3859
3860	raw_spin_unlock_irq(&pool->lock);
3861	}
3862
3863	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3864	complete(&wq->first_flusher->done);
3865
3866	return wait;
3867	}
3868
3869	static void touch_wq_lockdep_map(struct workqueue_struct *wq)
3870	{
3871	#ifdef CONFIG_LOCKDEP
3872	if (wq->flags & WQ_BH)
3873	local_bh_disable();
3874
3875	lock_map_acquire(&wq->lockdep_map);
3876	lock_map_release(&wq->lockdep_map);
3877
3878	if (wq->flags & WQ_BH)
3879	local_bh_enable();
3880	#endif
3881	}
3882
3883	static void touch_work_lockdep_map(struct work_struct *work,
3884	struct workqueue_struct *wq)
3885	{
3886	#ifdef CONFIG_LOCKDEP
3887	if (wq->flags & WQ_BH)
3888	local_bh_disable();
3889
3890	lock_map_acquire(&work->lockdep_map);
3891	lock_map_release(&work->lockdep_map);
3892
3893	if (wq->flags & WQ_BH)
3894	local_bh_enable();
3895	#endif
3896	}
3897
3898	/**
3899	* __flush_workqueue - ensure that any scheduled work has run to completion.
3900	* @wq: workqueue to flush
3901	*
3902	* This function sleeps until all work items which were queued on entry
3903	* have finished execution, but it is not livelocked by new incoming ones.
3904	*/
3905	void __flush_workqueue(struct workqueue_struct *wq)
3906	{
3907	struct wq_flusher this_flusher = {
3908	.list = LIST_HEAD_INIT(this_flusher.list),
3909	.flush_color = -1,
3910	.done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3911	};
3912	int next_color;
3913
3914	if (WARN_ON(!wq_online))
3915	return;
3916
3917	touch_wq_lockdep_map(wq);
3918
3919	mutex_lock(&wq->mutex);
3920
3921	/*
3922	* Start-to-wait phase
3923	*/
3924	next_color = work_next_color(wq->work_color);
3925
3926	if (next_color != wq->flush_color) {
3927	/*
3928	* Color space is not full. The current work_color
3929	* becomes our flush_color and work_color is advanced
3930	* by one.
3931	*/
3932	WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3933	this_flusher.flush_color = wq->work_color;
3934	wq->work_color = next_color;
3935
3936	if (!wq->first_flusher) {
3937	/* no flush in progress, become the first flusher */
3938	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3939
3940	wq->first_flusher = &this_flusher;
3941
3942	if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3943	wq->work_color)) {
3944	/* nothing to flush, done */
3945	wq->flush_color = next_color;
3946	wq->first_flusher = NULL;
3947	goto out_unlock;
3948	}
3949	} else {
3950	/* wait in queue */
3951	WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3952	list_add_tail(&this_flusher.list, &wq->flusher_queue);
3953	flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3954	}
3955	} else {
3956	/*
3957	* Oops, color space is full, wait on overflow queue.
3958	* The next flush completion will assign us
3959	* flush_color and transfer to flusher_queue.
3960	*/
3961	list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3962	}
3963
3964	check_flush_dependency(wq, NULL);
3965
3966	mutex_unlock(&wq->mutex);
3967
3968	wait_for_completion(&this_flusher.done);
3969
3970	/*
3971	* Wake-up-and-cascade phase
3972	*
3973	* First flushers are responsible for cascading flushes and
3974	* handling overflow. Non-first flushers can simply return.
3975	*/
3976	if (READ_ONCE(wq->first_flusher) != &this_flusher)
3977	return;
3978
3979	mutex_lock(&wq->mutex);
3980
3981	/* we might have raced, check again with mutex held */
3982	if (wq->first_flusher != &this_flusher)
3983	goto out_unlock;
3984
3985	WRITE_ONCE(wq->first_flusher, NULL);
3986
3987	WARN_ON_ONCE(!list_empty(&this_flusher.list));
3988	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3989
3990	while (true) {
3991	struct wq_flusher next, tmp;
3992
3993	/* complete all the flushers sharing the current flush color */
3994	list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3995	if (next->flush_color != wq->flush_color)
3996	break;
3997	list_del_init(&next->list);
3998	complete(&next->done);
3999	}
4000
4001	WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4002	wq->flush_color != work_next_color(wq->work_color));
4003
4004	/* this flush_color is finished, advance by one */
4005	wq->flush_color = work_next_color(wq->flush_color);
4006
4007	/* one color has been freed, handle overflow queue */
4008	if (!list_empty(&wq->flusher_overflow)) {
4009	/*
4010	* Assign the same color to all overflowed
4011	* flushers, advance work_color and append to
4012	* flusher_queue. This is the start-to-wait
4013	* phase for these overflowed flushers.
4014	*/
4015	list_for_each_entry(tmp, &wq->flusher_overflow, list)
4016	tmp->flush_color = wq->work_color;
4017
4018	wq->work_color = work_next_color(wq->work_color);
4019
4020	list_splice_tail_init(&wq->flusher_overflow,
4021	&wq->flusher_queue);
4022	flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4023	}
4024
4025	if (list_empty(&wq->flusher_queue)) {
4026	WARN_ON_ONCE(wq->flush_color != wq->work_color);
4027	break;
4028	}
4029
4030	/*
4031	* Need to flush more colors. Make the next flusher
4032	* the new first flusher and arm pwqs.
4033	*/
4034	WARN_ON_ONCE(wq->flush_color == wq->work_color);
4035	WARN_ON_ONCE(wq->flush_color != next->flush_color);
4036
4037	list_del_init(&next->list);
4038	wq->first_flusher = next;
4039
4040	if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4041	break;
4042
4043	/*
4044	* Meh... this color is already done, clear first
4045	* flusher and repeat cascading.
4046	*/
4047	wq->first_flusher = NULL;
4048	}
4049
4050	out_unlock:
4051	mutex_unlock(&wq->mutex);
4052	}
4053	EXPORT_SYMBOL(__flush_workqueue);
4054
4055	/**
4056	* drain_workqueue - drain a workqueue
4057	* @wq: workqueue to drain
4058	*
4059	* Wait until the workqueue becomes empty. While draining is in progress,
4060	* only chain queueing is allowed. IOW, only currently pending or running
4061	* work items on @wq can queue further work items on it. @wq is flushed
4062	* repeatedly until it becomes empty. The number of flushing is determined
4063	* by the depth of chaining and should be relatively short. Whine if it
4064	* takes too long.
4065	*/
4066	void drain_workqueue(struct workqueue_struct *wq)
4067	{
4068	unsigned int flush_cnt = 0;
4069	struct pool_workqueue *pwq;
4070
4071	/*
4072	* __queue_work() needs to test whether there are drainers, is much
4073	* hotter than drain_workqueue() and already looks at @wq->flags.
4074	* Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4075	*/
4076	mutex_lock(&wq->mutex);
4077	if (!wq->nr_drainers++)
4078	wq->flags \|= __WQ_DRAINING;
4079	mutex_unlock(&wq->mutex);
4080	reflush:
4081	__flush_workqueue(wq);
4082
4083	mutex_lock(&wq->mutex);
4084
4085	for_each_pwq(pwq, wq) {
4086	bool drained;
4087
4088	raw_spin_lock_irq(&pwq->pool->lock);
4089	drained = pwq_is_empty(pwq);
4090	raw_spin_unlock_irq(&pwq->pool->lock);
4091
4092	if (drained)
4093	continue;
4094
4095	if (++flush_cnt == 10 \|\|
4096	(flush_cnt % 100 == 0 && flush_cnt <= 1000))
4097	pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
4098	wq->name, __func__, flush_cnt);
4099
4100	mutex_unlock(&wq->mutex);
4101	goto reflush;
4102	}
4103
4104	if (!--wq->nr_drainers)
4105	wq->flags &= ~__WQ_DRAINING;
4106	mutex_unlock(&wq->mutex);
4107	}
4108	EXPORT_SYMBOL_GPL(drain_workqueue);
4109
4110	static bool start_flush_work(struct work_struct work, struct wq_barrier barr,
4111	bool from_cancel)
4112	{
4113	struct worker *worker = NULL;
4114	struct worker_pool *pool;
4115	struct pool_workqueue *pwq;
4116	struct workqueue_struct *wq;
4117
4118	rcu_read_lock();
4119	pool = get_work_pool(work);
4120	if (!pool) {
4121	rcu_read_unlock();
4122	return false;
4123	}
4124
4125	raw_spin_lock_irq(&pool->lock);
4126	/* see the comment in try_to_grab_pending() with the same code */
4127	pwq = get_work_pwq(work);
4128	if (pwq) {
4129	if (unlikely(pwq->pool != pool))
4130	goto already_gone;
4131	} else {
4132	worker = find_worker_executing_work(pool, work);
4133	if (!worker)
4134	goto already_gone;
4135	pwq = worker->current_pwq;
4136	}
4137
4138	wq = pwq->wq;
4139	check_flush_dependency(wq, work);
4140
4141	insert_wq_barrier(pwq, barr, work, worker);
4142	raw_spin_unlock_irq(&pool->lock);
4143
4144	touch_work_lockdep_map(work, wq);
4145
4146	/*
4147	* Force a lock recursion deadlock when using flush_work() inside a
4148	* single-threaded or rescuer equipped workqueue.
4149	*
4150	* For single threaded workqueues the deadlock happens when the work
4151	* is after the work issuing the flush_work(). For rescuer equipped
4152	* workqueues the deadlock happens when the rescuer stalls, blocking
4153	* forward progress.
4154	*/
4155	if (!from_cancel && (wq->saved_max_active == 1 \|\| wq->rescuer))
4156	touch_wq_lockdep_map(wq);
4157
4158	rcu_read_unlock();
4159	return true;
4160	already_gone:
4161	raw_spin_unlock_irq(&pool->lock);
4162	rcu_read_unlock();
4163	return false;
4164	}
4165
4166	static bool __flush_work(struct work_struct *work, bool from_cancel)
4167	{
4168	struct wq_barrier barr;
4169
4170	if (WARN_ON(!wq_online))
4171	return false;
4172
4173	if (WARN_ON(!work->func))
4174	return false;
4175
4176	if (!start_flush_work(work, &barr, from_cancel))
4177	return false;
4178
4179	/*
4180	* start_flush_work() returned %true. If @from_cancel is set, we know
4181	* that @work must have been executing during start_flush_work() and
4182	* can't currently be queued. Its data must contain OFFQ bits. If @work
4183	* was queued on a BH workqueue, we also know that it was running in the
4184	* BH context and thus can be busy-waited.
4185	*/
4186	if (from_cancel) {
4187	unsigned long data = *work_data_bits(work);
4188
4189	if (!WARN_ON_ONCE(data & WORK_STRUCT_PWQ) &&
4190	(data & WORK_OFFQ_BH)) {
4191	/*
4192	* On RT, prevent a live lock when %current preempted
4193	* soft interrupt processing or prevents ksoftirqd from
4194	* running by keeping flipping BH. If the BH work item
4195	* runs on a different CPU then this has no effect other
4196	* than doing the BH disable/enable dance for nothing.
4197	* This is copied from
4198	* kernel/softirq.c::tasklet_unlock_spin_wait().
4199	*/
4200	while (!try_wait_for_completion(&barr.done)) {
4201	if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
4202	local_bh_disable();
4203	local_bh_enable();
4204	} else {
4205	cpu_relax();
4206	}
4207	}
4208	goto out_destroy;
4209	}
4210	}
4211
4212	wait_for_completion(&barr.done);
4213
4214	out_destroy:
4215	destroy_work_on_stack(&barr.work);
4216	return true;
4217	}
4218
4219	/**
4220	* flush_work - wait for a work to finish executing the last queueing instance
4221	* @work: the work to flush
4222	*
4223	* Wait until @work has finished execution. @work is guaranteed to be idle
4224	* on return if it hasn't been requeued since flush started.
4225	*
4226	* Return:
4227	* %true if flush_work() waited for the work to finish execution,
4228	* %false if it was already idle.
4229	*/
4230	bool flush_work(struct work_struct *work)
4231	{
4232	might_sleep();
4233	return __flush_work(work, false);
4234	}
4235	EXPORT_SYMBOL_GPL(flush_work);
4236
4237	/**
4238	* flush_delayed_work - wait for a dwork to finish executing the last queueing
4239	* @dwork: the delayed work to flush
4240	*
4241	* Delayed timer is cancelled and the pending work is queued for
4242	* immediate execution. Like flush_work(), this function only
4243	* considers the last queueing instance of @dwork.
4244	*
4245	* Return:
4246	* %true if flush_work() waited for the work to finish execution,
4247	* %false if it was already idle.
4248	*/
4249	bool flush_delayed_work(struct delayed_work *dwork)
4250	{
4251	local_irq_disable();
4252	if (del_timer_sync(&dwork->timer))
4253	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
4254	local_irq_enable();
4255	return flush_work(&dwork->work);
4256	}
4257	EXPORT_SYMBOL(flush_delayed_work);
4258
4259	/**
4260	* flush_rcu_work - wait for a rwork to finish executing the last queueing
4261	* @rwork: the rcu work to flush
4262	*
4263	* Return:
4264	* %true if flush_rcu_work() waited for the work to finish execution,
4265	* %false if it was already idle.
4266	*/
4267	bool flush_rcu_work(struct rcu_work *rwork)
4268	{
4269	if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4270	rcu_barrier();
4271	flush_work(&rwork->work);
4272	return true;
4273	} else {
4274	return flush_work(&rwork->work);
4275	}
4276	}
4277	EXPORT_SYMBOL(flush_rcu_work);
4278
4279	static void work_offqd_disable(struct work_offq_data *offqd)
4280	{
4281	const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1;
4282
4283	if (likely(offqd->disable < max))
4284	offqd->disable++;
4285	else
4286	WARN_ONCE(true, "workqueue: work disable count overflowed\n");
4287	}
4288
4289	static void work_offqd_enable(struct work_offq_data *offqd)
4290	{
4291	if (likely(offqd->disable > 0))
4292	offqd->disable--;
4293	else
4294	WARN_ONCE(true, "workqueue: work disable count underflowed\n");
4295	}
4296
4297	static bool __cancel_work(struct work_struct *work, u32 cflags)
4298	{
4299	struct work_offq_data offqd;
4300	unsigned long irq_flags;
4301	int ret;
4302
4303	ret = work_grab_pending(work, cflags, &irq_flags);
4304
4305	work_offqd_unpack(&offqd, *work_data_bits(work));
4306
4307	if (cflags & WORK_CANCEL_DISABLE)
4308	work_offqd_disable(&offqd);
4309
4310	set_work_pool_and_clear_pending(work, offqd.pool_id,
4311	work_offqd_pack_flags(&offqd));
4312	local_irq_restore(irq_flags);
4313	return ret;
4314	}
4315
4316	static bool __cancel_work_sync(struct work_struct *work, u32 cflags)
4317	{
4318	bool ret;
4319
4320	ret = __cancel_work(work, cflags \| WORK_CANCEL_DISABLE);
4321
4322	if (*work_data_bits(work) & WORK_OFFQ_BH)
4323	WARN_ON_ONCE(in_hardirq());
4324	else
4325	might_sleep();
4326
4327	/*
4328	* Skip __flush_work() during early boot when we know that @work isn't
4329	* executing. This allows canceling during early boot.
4330	*/
4331	if (wq_online)
4332	__flush_work(work, true);
4333
4334	if (!(cflags & WORK_CANCEL_DISABLE))
4335	enable_work(work);
4336
4337	return ret;
4338	}
4339
4340	/*
4341	* See cancel_delayed_work()
4342	*/
4343	bool cancel_work(struct work_struct *work)
4344	{
4345	return __cancel_work(work, 0);
4346	}
4347	EXPORT_SYMBOL(cancel_work);
4348
4349	/**
4350	* cancel_work_sync - cancel a work and wait for it to finish
4351	* @work: the work to cancel
4352	*
4353	* Cancel @work and wait for its execution to finish. This function can be used
4354	* even if the work re-queues itself or migrates to another workqueue. On return
4355	* from this function, @work is guaranteed to be not pending or executing on any
4356	* CPU as long as there aren't racing enqueues.
4357	*
4358	* cancel_work_sync(&delayed_work->work) must not be used for delayed_work's.
4359	* Use cancel_delayed_work_sync() instead.
4360	*
4361	* Must be called from a sleepable context if @work was last queued on a non-BH
4362	* workqueue. Can also be called from non-hardirq atomic contexts including BH
4363	* if @work was last queued on a BH workqueue.
4364	*
4365	* Returns %true if @work was pending, %false otherwise.
4366	*/
4367	bool cancel_work_sync(struct work_struct *work)
4368	{
4369	return __cancel_work_sync(work, 0);
4370	}
4371	EXPORT_SYMBOL_GPL(cancel_work_sync);
4372
4373	/**
4374	* cancel_delayed_work - cancel a delayed work
4375	* @dwork: delayed_work to cancel
4376	*
4377	* Kill off a pending delayed_work.
4378	*
4379	* Return: %true if @dwork was pending and canceled; %false if it wasn't
4380	* pending.
4381	*
4382	* Note:
4383	* The work callback function may still be running on return, unless
4384	* it returns %true and the work doesn't re-arm itself. Explicitly flush or
4385	* use cancel_delayed_work_sync() to wait on it.
4386	*
4387	* This function is safe to call from any context including IRQ handler.
4388	*/
4389	bool cancel_delayed_work(struct delayed_work *dwork)
4390	{
4391	return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED);
4392	}
4393	EXPORT_SYMBOL(cancel_delayed_work);
4394
4395	/**
4396	* cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4397	* @dwork: the delayed work cancel
4398	*
4399	* This is cancel_work_sync() for delayed works.
4400	*
4401	* Return:
4402	* %true if @dwork was pending, %false otherwise.
4403	*/
4404	bool cancel_delayed_work_sync(struct delayed_work *dwork)
4405	{
4406	return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED);
4407	}
4408	EXPORT_SYMBOL(cancel_delayed_work_sync);
4409
4410	/**
4411	* disable_work - Disable and cancel a work item
4412	* @work: work item to disable
4413	*
4414	* Disable @work by incrementing its disable count and cancel it if currently
4415	* pending. As long as the disable count is non-zero, any attempt to queue @work
4416	* will fail and return %false. The maximum supported disable depth is 2 to the
4417	* power of %WORK_OFFQ_DISABLE_BITS, currently 65536.
4418	*
4419	* Can be called from any context. Returns %true if @work was pending, %false
4420	* otherwise.
4421	*/
4422	bool disable_work(struct work_struct *work)
4423	{
4424	return __cancel_work(work, WORK_CANCEL_DISABLE);
4425	}
4426	EXPORT_SYMBOL_GPL(disable_work);
4427
4428	/**
4429	* disable_work_sync - Disable, cancel and drain a work item
4430	* @work: work item to disable
4431	*
4432	* Similar to disable_work() but also wait for @work to finish if currently
4433	* executing.
4434	*
4435	* Must be called from a sleepable context if @work was last queued on a non-BH
4436	* workqueue. Can also be called from non-hardirq atomic contexts including BH
4437	* if @work was last queued on a BH workqueue.
4438	*
4439	* Returns %true if @work was pending, %false otherwise.
4440	*/
4441	bool disable_work_sync(struct work_struct *work)
4442	{
4443	return __cancel_work_sync(work, WORK_CANCEL_DISABLE);
4444	}
4445	EXPORT_SYMBOL_GPL(disable_work_sync);
4446
4447	/**
4448	* enable_work - Enable a work item
4449	* @work: work item to enable
4450	*
4451	* Undo disable_work[_sync]() by decrementing @work's disable count. @work can
4452	* only be queued if its disable count is 0.
4453	*
4454	* Can be called from any context. Returns %true if the disable count reached 0.
4455	* Otherwise, %false.
4456	*/
4457	bool enable_work(struct work_struct *work)
4458	{
4459	struct work_offq_data offqd;
4460	unsigned long irq_flags;
4461
4462	work_grab_pending(work, 0, &irq_flags);
4463
4464	work_offqd_unpack(&offqd, *work_data_bits(work));
4465	work_offqd_enable(&offqd);
4466	set_work_pool_and_clear_pending(work, offqd.pool_id,
4467	work_offqd_pack_flags(&offqd));
4468	local_irq_restore(irq_flags);
4469
4470	return !offqd.disable;
4471	}
4472	EXPORT_SYMBOL_GPL(enable_work);
4473
4474	/**
4475	* disable_delayed_work - Disable and cancel a delayed work item
4476	* @dwork: delayed work item to disable
4477	*
4478	* disable_work() for delayed work items.
4479	*/
4480	bool disable_delayed_work(struct delayed_work *dwork)
4481	{
4482	return __cancel_work(&dwork->work,
4483	WORK_CANCEL_DELAYED \| WORK_CANCEL_DISABLE);
4484	}
4485	EXPORT_SYMBOL_GPL(disable_delayed_work);
4486
4487	/**
4488	* disable_delayed_work_sync - Disable, cancel and drain a delayed work item
4489	* @dwork: delayed work item to disable
4490	*
4491	* disable_work_sync() for delayed work items.
4492	*/
4493	bool disable_delayed_work_sync(struct delayed_work *dwork)
4494	{
4495	return __cancel_work_sync(&dwork->work,
4496	WORK_CANCEL_DELAYED \| WORK_CANCEL_DISABLE);
4497	}
4498	EXPORT_SYMBOL_GPL(disable_delayed_work_sync);
4499
4500	/**
4501	* enable_delayed_work - Enable a delayed work item
4502	* @dwork: delayed work item to enable
4503	*
4504	* enable_work() for delayed work items.
4505	*/
4506	bool enable_delayed_work(struct delayed_work *dwork)
4507	{
4508	return enable_work(&dwork->work);
4509	}
4510	EXPORT_SYMBOL_GPL(enable_delayed_work);
4511
4512	/**
4513	* schedule_on_each_cpu - execute a function synchronously on each online CPU
4514	* @func: the function to call
4515	*
4516	* schedule_on_each_cpu() executes @func on each online CPU using the
4517	* system workqueue and blocks until all CPUs have completed.
4518	* schedule_on_each_cpu() is very slow.
4519	*
4520	* Return:
4521	* 0 on success, -errno on failure.
4522	*/
4523	int schedule_on_each_cpu(work_func_t func)
4524	{
4525	int cpu;
4526	struct work_struct __percpu *works;
4527
4528	works = alloc_percpu(struct work_struct);
4529	if (!works)
4530	return -ENOMEM;
4531
4532	cpus_read_lock();
4533
4534	for_each_online_cpu(cpu) {
4535	struct work_struct *work = per_cpu_ptr(works, cpu);
4536
4537	INIT_WORK(work, func);
4538	schedule_work_on(cpu, work);
4539	}
4540
4541	for_each_online_cpu(cpu)
4542	flush_work(per_cpu_ptr(works, cpu));
4543
4544	cpus_read_unlock();
4545	free_percpu(works);
4546	return 0;
4547	}
4548
4549	/**
4550	* execute_in_process_context - reliably execute the routine with user context
4551	* @fn: the function to execute
4552	* @ew: guaranteed storage for the execute work structure (must
4553	* be available when the work executes)
4554	*
4555	* Executes the function immediately if process context is available,
4556	* otherwise schedules the function for delayed execution.
4557	*
4558	* Return: 0 - function was executed
4559	* 1 - function was scheduled for execution
4560	*/
4561	int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4562	{
4563	if (!in_interrupt()) {
4564	fn(&ew->work);
4565	return 0;
4566	}
4567
4568	INIT_WORK(&ew->work, fn);
4569	schedule_work(&ew->work);
4570
4571	return 1;
4572	}
4573	EXPORT_SYMBOL_GPL(execute_in_process_context);
4574
4575	/**
4576	* free_workqueue_attrs - free a workqueue_attrs
4577	* @attrs: workqueue_attrs to free
4578	*
4579	* Undo alloc_workqueue_attrs().
4580	*/
4581	void free_workqueue_attrs(struct workqueue_attrs *attrs)
4582	{
4583	if (attrs) {
4584	free_cpumask_var(attrs->cpumask);
4585	free_cpumask_var(attrs->__pod_cpumask);
4586	kfree(attrs);
4587	}
4588	}
4589
4590	/**
4591	* alloc_workqueue_attrs - allocate a workqueue_attrs
4592	*
4593	* Allocate a new workqueue_attrs, initialize with default settings and
4594	* return it.
4595	*
4596	* Return: The allocated new workqueue_attr on success. %NULL on failure.
4597	*/
4598	struct workqueue_attrs *alloc_workqueue_attrs(void)
4599	{
4600	struct workqueue_attrs *attrs;
4601
4602	attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4603	if (!attrs)
4604	goto fail;
4605	if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4606	goto fail;
4607	if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4608	goto fail;
4609
4610	cpumask_copy(attrs->cpumask, cpu_possible_mask);
4611	attrs->affn_scope = WQ_AFFN_DFL;
4612	return attrs;
4613	fail:
4614	free_workqueue_attrs(attrs);
4615	return NULL;
4616	}
4617
4618	static void copy_workqueue_attrs(struct workqueue_attrs *to,
4619	const struct workqueue_attrs *from)
4620	{
4621	to->nice = from->nice;
4622	cpumask_copy(to->cpumask, from->cpumask);
4623	cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4624	to->affn_strict = from->affn_strict;
4625
4626	/*
4627	* Unlike hash and equality test, copying shouldn't ignore wq-only
4628	* fields as copying is used for both pool and wq attrs. Instead,
4629	* get_unbound_pool() explicitly clears the fields.
4630	*/
4631	to->affn_scope = from->affn_scope;
4632	to->ordered = from->ordered;
4633	}
4634
4635	/*
4636	* Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4637	* comments in 'struct workqueue_attrs' definition.
4638	*/
4639	static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4640	{
4641	attrs->affn_scope = WQ_AFFN_NR_TYPES;
4642	attrs->ordered = false;
4643	if (attrs->affn_strict)
4644	cpumask_copy(attrs->cpumask, cpu_possible_mask);
4645	}
4646
4647	/* hash value of the content of @attr */
4648	static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4649	{
4650	u32 hash = 0;
4651
4652	hash = jhash_1word(attrs->nice, hash);
4653	hash = jhash_1word(attrs->affn_strict, hash);
4654	hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4655	BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4656	if (!attrs->affn_strict)
4657	hash = jhash(cpumask_bits(attrs->cpumask),
4658	BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4659	return hash;
4660	}
4661
4662	/* content equality test */
4663	static bool wqattrs_equal(const struct workqueue_attrs *a,
4664	const struct workqueue_attrs *b)
4665	{
4666	if (a->nice != b->nice)
4667	return false;
4668	if (a->affn_strict != b->affn_strict)
4669	return false;
4670	if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4671	return false;
4672	if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask))
4673	return false;
4674	return true;
4675	}
4676
4677	/* Update @attrs with actually available CPUs */
4678	static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4679	const cpumask_t *unbound_cpumask)
4680	{
4681	/*
4682	* Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4683	* @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4684	* @unbound_cpumask.
4685	*/
4686	cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4687	if (unlikely(cpumask_empty(attrs->cpumask)))
4688	cpumask_copy(attrs->cpumask, unbound_cpumask);
4689	}
4690
4691	/* find wq_pod_type to use for @attrs */
4692	static const struct wq_pod_type *
4693	wqattrs_pod_type(const struct workqueue_attrs *attrs)
4694	{
4695	enum wq_affn_scope scope;
4696	struct wq_pod_type *pt;
4697
4698	/* to synchronize access to wq_affn_dfl */
4699	lockdep_assert_held(&wq_pool_mutex);
4700
4701	if (attrs->affn_scope == WQ_AFFN_DFL)
4702	scope = wq_affn_dfl;
4703	else
4704	scope = attrs->affn_scope;
4705
4706	pt = &wq_pod_types[scope];
4707
4708	if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4709	likely(pt->nr_pods))
4710	return pt;
4711
4712	/*
4713	* Before workqueue_init_topology(), only SYSTEM is available which is
4714	* initialized in workqueue_init_early().
4715	*/
4716	pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4717	BUG_ON(!pt->nr_pods);
4718	return pt;
4719	}
4720
4721	/**
4722	* init_worker_pool - initialize a newly zalloc'd worker_pool
4723	* @pool: worker_pool to initialize
4724	*
4725	* Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4726	*
4727	* Return: 0 on success, -errno on failure. Even on failure, all fields
4728	* inside @pool proper are initialized and put_unbound_pool() can be called
4729	* on @pool safely to release it.
4730	*/
4731	static int init_worker_pool(struct worker_pool *pool)
4732	{
4733	raw_spin_lock_init(&pool->lock);
4734	pool->id = -1;
4735	pool->cpu = -1;
4736	pool->node = NUMA_NO_NODE;
4737	pool->flags \|= POOL_DISASSOCIATED;
4738	pool->watchdog_ts = jiffies;
4739	INIT_LIST_HEAD(&pool->worklist);
4740	INIT_LIST_HEAD(&pool->idle_list);
4741	hash_init(pool->busy_hash);
4742
4743	timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4744	INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4745
4746	timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4747
4748	INIT_LIST_HEAD(&pool->workers);
4749
4750	ida_init(&pool->worker_ida);
4751	INIT_HLIST_NODE(&pool->hash_node);
4752	pool->refcnt = 1;
4753
4754	/* shouldn't fail above this point */
4755	pool->attrs = alloc_workqueue_attrs();
4756	if (!pool->attrs)
4757	return -ENOMEM;
4758
4759	wqattrs_clear_for_pool(pool->attrs);
4760
4761	return 0;
4762	}
4763
4764	#ifdef CONFIG_LOCKDEP
4765	static void wq_init_lockdep(struct workqueue_struct *wq)
4766	{
4767	char *lock_name;
4768
4769	lockdep_register_key(&wq->key);
4770	lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4771	if (!lock_name)
4772	lock_name = wq->name;
4773
4774	wq->lock_name = lock_name;
4775	lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
4776	}
4777
4778	static void wq_unregister_lockdep(struct workqueue_struct *wq)
4779	{
4780	lockdep_unregister_key(&wq->key);
4781	}
4782
4783	static void wq_free_lockdep(struct workqueue_struct *wq)
4784	{
4785	if (wq->lock_name != wq->name)
4786	kfree(wq->lock_name);
4787	}
4788	#else
4789	static void wq_init_lockdep(struct workqueue_struct *wq)
4790	{
4791	}
4792
4793	static void wq_unregister_lockdep(struct workqueue_struct *wq)
4794	{
4795	}
4796
4797	static void wq_free_lockdep(struct workqueue_struct *wq)
4798	{
4799	}
4800	#endif
4801
4802	static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
4803	{
4804	int node;
4805
4806	for_each_node(node) {
4807	kfree(nna_ar[node]);
4808	nna_ar[node] = NULL;
4809	}
4810
4811	kfree(nna_ar[nr_node_ids]);
4812	nna_ar[nr_node_ids] = NULL;
4813	}
4814
4815	static void init_node_nr_active(struct wq_node_nr_active *nna)
4816	{
4817	nna->max = WQ_DFL_MIN_ACTIVE;
4818	atomic_set(&nna->nr, 0);
4819	raw_spin_lock_init(&nna->lock);
4820	INIT_LIST_HEAD(&nna->pending_pwqs);
4821	}
4822
4823	/*
4824	* Each node's nr_active counter will be accessed mostly from its own node and
4825	* should be allocated in the node.
4826	*/
4827	static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
4828	{
4829	struct wq_node_nr_active *nna;
4830	int node;
4831
4832	for_each_node(node) {
4833	nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
4834	if (!nna)
4835	goto err_free;
4836	init_node_nr_active(nna);
4837	nna_ar[node] = nna;
4838	}
4839
4840	/* [nr_node_ids] is used as the fallback */
4841	nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
4842	if (!nna)
4843	goto err_free;
4844	init_node_nr_active(nna);
4845	nna_ar[nr_node_ids] = nna;
4846
4847	return 0;
4848
4849	err_free:
4850	free_node_nr_active(nna_ar);
4851	return -ENOMEM;
4852	}
4853
4854	static void rcu_free_wq(struct rcu_head *rcu)
4855	{
4856	struct workqueue_struct *wq =
4857	container_of(rcu, struct workqueue_struct, rcu);
4858
4859	if (wq->flags & WQ_UNBOUND)
4860	free_node_nr_active(wq->node_nr_active);
4861
4862	wq_free_lockdep(wq);
4863	free_percpu(wq->cpu_pwq);
4864	free_workqueue_attrs(wq->unbound_attrs);
4865	kfree(wq);
4866	}
4867
4868	static void rcu_free_pool(struct rcu_head *rcu)
4869	{
4870	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4871
4872	ida_destroy(&pool->worker_ida);
4873	free_workqueue_attrs(pool->attrs);
4874	kfree(pool);
4875	}
4876
4877	/**
4878	* put_unbound_pool - put a worker_pool
4879	* @pool: worker_pool to put
4880	*
4881	* Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4882	* safe manner. get_unbound_pool() calls this function on its failure path
4883	* and this function should be able to release pools which went through,
4884	* successfully or not, init_worker_pool().
4885	*
4886	* Should be called with wq_pool_mutex held.
4887	*/
4888	static void put_unbound_pool(struct worker_pool *pool)
4889	{
4890	struct worker *worker;
4891	LIST_HEAD(cull_list);
4892
4893	lockdep_assert_held(&wq_pool_mutex);
4894
4895	if (--pool->refcnt)
4896	return;
4897
4898	/* sanity checks */
4899	if (WARN_ON(!(pool->cpu < 0)) \|\|
4900	WARN_ON(!list_empty(&pool->worklist)))
4901	return;
4902
4903	/* release id and unhash */
4904	if (pool->id >= 0)
4905	idr_remove(&worker_pool_idr, pool->id);
4906	hash_del(&pool->hash_node);
4907
4908	/*
4909	* Become the manager and destroy all workers. This prevents
4910	* @pool's workers from blocking on attach_mutex. We're the last
4911	* manager and @pool gets freed with the flag set.
4912	*
4913	* Having a concurrent manager is quite unlikely to happen as we can
4914	* only get here with
4915	* pwq->refcnt == pool->refcnt == 0
4916	* which implies no work queued to the pool, which implies no worker can
4917	* become the manager. However a worker could have taken the role of
4918	* manager before the refcnts dropped to 0, since maybe_create_worker()
4919	* drops pool->lock
4920	*/
4921	while (true) {
4922	rcuwait_wait_event(&manager_wait,
4923	!(pool->flags & POOL_MANAGER_ACTIVE),
4924	TASK_UNINTERRUPTIBLE);
4925
4926	mutex_lock(&wq_pool_attach_mutex);
4927	raw_spin_lock_irq(&pool->lock);
4928	if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4929	pool->flags \|= POOL_MANAGER_ACTIVE;
4930	break;
4931	}
4932	raw_spin_unlock_irq(&pool->lock);
4933	mutex_unlock(&wq_pool_attach_mutex);
4934	}
4935
4936	while ((worker = first_idle_worker(pool)))
4937	set_worker_dying(worker, &cull_list);
4938	WARN_ON(pool->nr_workers \|\| pool->nr_idle);
4939	raw_spin_unlock_irq(&pool->lock);
4940
4941	detach_dying_workers(&cull_list);
4942
4943	mutex_unlock(&wq_pool_attach_mutex);
4944
4945	reap_dying_workers(&cull_list);
4946
4947	/* shut down the timers */
4948	del_timer_sync(&pool->idle_timer);
4949	cancel_work_sync(&pool->idle_cull_work);
4950	del_timer_sync(&pool->mayday_timer);
4951
4952	/* RCU protected to allow dereferences from get_work_pool() */
4953	call_rcu(&pool->rcu, rcu_free_pool);
4954	}
4955
4956	/**
4957	* get_unbound_pool - get a worker_pool with the specified attributes
4958	* @attrs: the attributes of the worker_pool to get
4959	*
4960	* Obtain a worker_pool which has the same attributes as @attrs, bump the
4961	* reference count and return it. If there already is a matching
4962	* worker_pool, it will be used; otherwise, this function attempts to
4963	* create a new one.
4964	*
4965	* Should be called with wq_pool_mutex held.
4966	*
4967	* Return: On success, a worker_pool with the same attributes as @attrs.
4968	* On failure, %NULL.
4969	*/
4970	static struct worker_pool get_unbound_pool(const struct workqueue_attrs attrs)
4971	{
4972	struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4973	u32 hash = wqattrs_hash(attrs);
4974	struct worker_pool *pool;
4975	int pod, node = NUMA_NO_NODE;
4976
4977	lockdep_assert_held(&wq_pool_mutex);
4978
4979	/* do we already have a matching pool? */
4980	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4981	if (wqattrs_equal(pool->attrs, attrs)) {
4982	pool->refcnt++;
4983	return pool;
4984	}
4985	}
4986
4987	/* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4988	for (pod = 0; pod < pt->nr_pods; pod++) {
4989	if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4990	node = pt->pod_node[pod];
4991	break;
4992	}
4993	}
4994
4995	/* nope, create a new one */
4996	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4997	if (!pool \|\| init_worker_pool(pool) < 0)
4998	goto fail;
4999
5000	pool->node = node;
5001	copy_workqueue_attrs(pool->attrs, attrs);
5002	wqattrs_clear_for_pool(pool->attrs);
5003
5004	if (worker_pool_assign_id(pool) < 0)
5005	goto fail;
5006
5007	/* create and start the initial worker */
5008	if (wq_online && !create_worker(pool))
5009	goto fail;
5010
5011	/* install */
5012	hash_add(unbound_pool_hash, &pool->hash_node, hash);
5013
5014	return pool;
5015	fail:
5016	if (pool)
5017	put_unbound_pool(pool);
5018	return NULL;
5019	}
5020
5021	/*
5022	* Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
5023	* refcnt and needs to be destroyed.
5024	*/
5025	static void pwq_release_workfn(struct kthread_work *work)
5026	{
5027	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
5028	release_work);
5029	struct workqueue_struct *wq = pwq->wq;
5030	struct worker_pool *pool = pwq->pool;
5031	bool is_last = false;
5032
5033	/*
5034	* When @pwq is not linked, it doesn't hold any reference to the
5035	* @wq, and @wq is invalid to access.
5036	*/
5037	if (!list_empty(&pwq->pwqs_node)) {
5038	mutex_lock(&wq->mutex);
5039	list_del_rcu(&pwq->pwqs_node);
5040	is_last = list_empty(&wq->pwqs);
5041
5042	/*
5043	* For ordered workqueue with a plugged dfl_pwq, restart it now.
5044	*/
5045	if (!is_last && (wq->flags & __WQ_ORDERED))
5046	unplug_oldest_pwq(wq);
5047
5048	mutex_unlock(&wq->mutex);
5049	}
5050
5051	if (wq->flags & WQ_UNBOUND) {
5052	mutex_lock(&wq_pool_mutex);
5053	put_unbound_pool(pool);
5054	mutex_unlock(&wq_pool_mutex);
5055	}
5056
5057	if (!list_empty(&pwq->pending_node)) {
5058	struct wq_node_nr_active *nna =
5059	wq_node_nr_active(pwq->wq, pwq->pool->node);
5060
5061	raw_spin_lock_irq(&nna->lock);
5062	list_del_init(&pwq->pending_node);
5063	raw_spin_unlock_irq(&nna->lock);
5064	}
5065
5066	kfree_rcu(pwq, rcu);
5067
5068	/*
5069	* If we're the last pwq going away, @wq is already dead and no one
5070	* is gonna access it anymore. Schedule RCU free.
5071	*/
5072	if (is_last) {
5073	wq_unregister_lockdep(wq);
5074	call_rcu(&wq->rcu, rcu_free_wq);
5075	}
5076	}
5077
5078	/* initialize newly allocated @pwq which is associated with @wq and @pool */
5079	static void init_pwq(struct pool_workqueue pwq, struct workqueue_struct wq,
5080	struct worker_pool *pool)
5081	{
5082	BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK);
5083
5084	memset(pwq, 0, sizeof(*pwq));
5085
5086	pwq->pool = pool;
5087	pwq->wq = wq;
5088	pwq->flush_color = -1;
5089	pwq->refcnt = 1;
5090	INIT_LIST_HEAD(&pwq->inactive_works);
5091	INIT_LIST_HEAD(&pwq->pending_node);
5092	INIT_LIST_HEAD(&pwq->pwqs_node);
5093	INIT_LIST_HEAD(&pwq->mayday_node);
5094	kthread_init_work(&pwq->release_work, pwq_release_workfn);
5095	}
5096
5097	/* sync @pwq with the current state of its associated wq and link it */
5098	static void link_pwq(struct pool_workqueue *pwq)
5099	{
5100	struct workqueue_struct *wq = pwq->wq;
5101
5102	lockdep_assert_held(&wq->mutex);
5103
5104	/* may be called multiple times, ignore if already linked */
5105	if (!list_empty(&pwq->pwqs_node))
5106	return;
5107
5108	/* set the matching work_color */
5109	pwq->work_color = wq->work_color;
5110
5111	/* link in @pwq */
5112	list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs);
5113	}
5114
5115	/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5116	static struct pool_workqueue alloc_unbound_pwq(struct workqueue_struct wq,
5117	const struct workqueue_attrs *attrs)
5118	{
5119	struct worker_pool *pool;
5120	struct pool_workqueue *pwq;
5121
5122	lockdep_assert_held(&wq_pool_mutex);
5123
5124	pool = get_unbound_pool(attrs);
5125	if (!pool)
5126	return NULL;
5127
5128	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5129	if (!pwq) {
5130	put_unbound_pool(pool);
5131	return NULL;
5132	}
5133
5134	init_pwq(pwq, wq, pool);
5135	return pwq;
5136	}
5137
5138	static void apply_wqattrs_lock(void)
5139	{
5140	mutex_lock(&wq_pool_mutex);
5141	}
5142
5143	static void apply_wqattrs_unlock(void)
5144	{
5145	mutex_unlock(&wq_pool_mutex);
5146	}
5147
5148	/**
5149	* wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
5150	* @attrs: the wq_attrs of the default pwq of the target workqueue
5151	* @cpu: the target CPU
5152	*
5153	* Calculate the cpumask a workqueue with @attrs should use on @pod.
5154	* The result is stored in @attrs->__pod_cpumask.
5155	*
5156	* If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
5157	* and @pod has online CPUs requested by @attrs, the returned cpumask is the
5158	* intersection of the possible CPUs of @pod and @attrs->cpumask.
5159	*
5160	* The caller is responsible for ensuring that the cpumask of @pod stays stable.
5161	*/
5162	static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu)
5163	{
5164	const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5165	int pod = pt->cpu_pod[cpu];
5166
5167	/* calculate possible CPUs in @pod that @attrs wants */
5168	cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
5169	/* does @pod have any online CPUs @attrs wants? */
5170	if (!cpumask_intersects(attrs->__pod_cpumask, wq_online_cpumask)) {
5171	cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
5172	return;
5173	}
5174	}
5175
5176	/* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
5177	static struct pool_workqueue install_unbound_pwq(struct workqueue_struct wq,
5178	int cpu, struct pool_workqueue *pwq)
5179	{
5180	struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
5181	struct pool_workqueue *old_pwq;
5182
5183	lockdep_assert_held(&wq_pool_mutex);
5184	lockdep_assert_held(&wq->mutex);
5185
5186	/* link_pwq() can handle duplicate calls */
5187	link_pwq(pwq);
5188
5189	old_pwq = rcu_access_pointer(*slot);
5190	rcu_assign_pointer(*slot, pwq);
5191	return old_pwq;
5192	}
5193
5194	/* context to store the prepared attrs & pwqs before applying */
5195	struct apply_wqattrs_ctx {
5196	struct workqueue_struct wq; / target workqueue */
5197	struct workqueue_attrs attrs; / attrs to apply */
5198	struct list_head list; /* queued for batching commit */
5199	struct pool_workqueue *dfl_pwq;
5200	struct pool_workqueue *pwq_tbl[];
5201	};
5202
5203	/* free the resources after success or abort */
5204	static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5205	{
5206	if (ctx) {
5207	int cpu;
5208
5209	for_each_possible_cpu(cpu)
5210	put_pwq_unlocked(ctx->pwq_tbl[cpu]);
5211	put_pwq_unlocked(ctx->dfl_pwq);
5212
5213	free_workqueue_attrs(ctx->attrs);
5214
5215	kfree(ctx);
5216	}
5217	}
5218
5219	/* allocate the attrs and pwqs for later installation */
5220	static struct apply_wqattrs_ctx *
5221	apply_wqattrs_prepare(struct workqueue_struct *wq,
5222	const struct workqueue_attrs *attrs,
5223	const cpumask_var_t unbound_cpumask)
5224	{
5225	struct apply_wqattrs_ctx *ctx;
5226	struct workqueue_attrs *new_attrs;
5227	int cpu;
5228
5229	lockdep_assert_held(&wq_pool_mutex);
5230
5231	if (WARN_ON(attrs->affn_scope < 0 \|\|
5232	attrs->affn_scope >= WQ_AFFN_NR_TYPES))
5233	return ERR_PTR(-EINVAL);
5234
5235	ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
5236
5237	new_attrs = alloc_workqueue_attrs();
5238	if (!ctx \|\| !new_attrs)
5239	goto out_free;
5240
5241	/*
5242	* If something goes wrong during CPU up/down, we'll fall back to
5243	* the default pwq covering whole @attrs->cpumask. Always create
5244	* it even if we don't use it immediately.
5245	*/
5246	copy_workqueue_attrs(new_attrs, attrs);
5247	wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
5248	cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5249	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
5250	if (!ctx->dfl_pwq)
5251	goto out_free;
5252
5253	for_each_possible_cpu(cpu) {
5254	if (new_attrs->ordered) {
5255	ctx->dfl_pwq->refcnt++;
5256	ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
5257	} else {
5258	wq_calc_pod_cpumask(new_attrs, cpu);
5259	ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
5260	if (!ctx->pwq_tbl[cpu])
5261	goto out_free;
5262	}
5263	}
5264
5265	/* save the user configured attrs and sanitize it. */
5266	copy_workqueue_attrs(new_attrs, attrs);
5267	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
5268	cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5269	ctx->attrs = new_attrs;
5270
5271	/*
5272	* For initialized ordered workqueues, there should only be one pwq
5273	* (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution
5274	* of newly queued work items until execution of older work items in
5275	* the old pwq's have completed.
5276	*/
5277	if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))
5278	ctx->dfl_pwq->plugged = true;
5279
5280	ctx->wq = wq;
5281	return ctx;
5282
5283	out_free:
5284	free_workqueue_attrs(new_attrs);
5285	apply_wqattrs_cleanup(ctx);
5286	return ERR_PTR(-ENOMEM);
5287	}
5288
5289	/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
5290	static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5291	{
5292	int cpu;
5293
5294	/* all pwqs have been created successfully, let's install'em */
5295	mutex_lock(&ctx->wq->mutex);
5296
5297	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
5298
5299	/* save the previous pwqs and install the new ones */
5300	for_each_possible_cpu(cpu)
5301	ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
5302	ctx->pwq_tbl[cpu]);
5303	ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq);
5304
5305	/* update node_nr_active->max */
5306	wq_update_node_max_active(ctx->wq, -1);
5307
5308	/* rescuer needs to respect wq cpumask changes */
5309	if (ctx->wq->rescuer)
5310	set_cpus_allowed_ptr(ctx->wq->rescuer->task,
5311	unbound_effective_cpumask(ctx->wq));
5312
5313	mutex_unlock(&ctx->wq->mutex);
5314	}
5315
5316	static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5317	const struct workqueue_attrs *attrs)
5318	{
5319	struct apply_wqattrs_ctx *ctx;
5320
5321	/* only unbound workqueues can change attributes */
5322	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5323	return -EINVAL;
5324
5325	ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
5326	if (IS_ERR(ctx))
5327	return PTR_ERR(ctx);
5328
5329	/* the ctx has been prepared successfully, let's commit it */
5330	apply_wqattrs_commit(ctx);
5331	apply_wqattrs_cleanup(ctx);
5332
5333	return 0;
5334	}
5335
5336	/**
5337	* apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5338	* @wq: the target workqueue
5339	* @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5340	*
5341	* Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
5342	* a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
5343	* work items are affine to the pod it was issued on. Older pwqs are released as
5344	* in-flight work items finish. Note that a work item which repeatedly requeues
5345	* itself back-to-back will stay on its current pwq.
5346	*
5347	* Performs GFP_KERNEL allocations.
5348	*
5349	* Return: 0 on success and -errno on failure.
5350	*/
5351	int apply_workqueue_attrs(struct workqueue_struct *wq,
5352	const struct workqueue_attrs *attrs)
5353	{
5354	int ret;
5355
5356	mutex_lock(&wq_pool_mutex);
5357	ret = apply_workqueue_attrs_locked(wq, attrs);
5358	mutex_unlock(&wq_pool_mutex);
5359
5360	return ret;
5361	}
5362
5363	/**
5364	* unbound_wq_update_pwq - update a pwq slot for CPU hot[un]plug
5365	* @wq: the target workqueue
5366	* @cpu: the CPU to update the pwq slot for
5367	*
5368	* This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5369	* %CPU_DOWN_FAILED. @cpu is in the same pod of the CPU being hot[un]plugged.
5370	*
5371	*
5372	* If pod affinity can't be adjusted due to memory allocation failure, it falls
5373	* back to @wq->dfl_pwq which may not be optimal but is always correct.
5374	*
5375	* Note that when the last allowed CPU of a pod goes offline for a workqueue
5376	* with a cpumask spanning multiple pods, the workers which were already
5377	* executing the work items for the workqueue will lose their CPU affinity and
5378	* may execute on any CPU. This is similar to how per-cpu workqueues behave on
5379	* CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
5380	* responsibility to flush the work item from CPU_DOWN_PREPARE.
5381	*/
5382	static void unbound_wq_update_pwq(struct workqueue_struct *wq, int cpu)
5383	{
5384	struct pool_workqueue old_pwq = NULL, pwq;
5385	struct workqueue_attrs *target_attrs;
5386
5387	lockdep_assert_held(&wq_pool_mutex);
5388
5389	if (!(wq->flags & WQ_UNBOUND) \|\| wq->unbound_attrs->ordered)
5390	return;
5391
5392	/*
5393	* We don't wanna alloc/free wq_attrs for each wq for each CPU.
5394	* Let's use a preallocated one. The following buf is protected by
5395	* CPU hotplug exclusion.
5396	*/
5397	target_attrs = unbound_wq_update_pwq_attrs_buf;
5398
5399	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5400	wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
5401
5402	/* nothing to do if the target cpumask matches the current pwq */
5403	wq_calc_pod_cpumask(target_attrs, cpu);
5404	if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs))
5405	return;
5406
5407	/* create a new pwq */
5408	pwq = alloc_unbound_pwq(wq, target_attrs);
5409	if (!pwq) {
5410	pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
5411	wq->name);
5412	goto use_dfl_pwq;
5413	}
5414
5415	/* Install the new pwq. */
5416	mutex_lock(&wq->mutex);
5417	old_pwq = install_unbound_pwq(wq, cpu, pwq);
5418	goto out_unlock;
5419
5420	use_dfl_pwq:
5421	mutex_lock(&wq->mutex);
5422	pwq = unbound_pwq(wq, -1);
5423	raw_spin_lock_irq(&pwq->pool->lock);
5424	get_pwq(pwq);
5425	raw_spin_unlock_irq(&pwq->pool->lock);
5426	old_pwq = install_unbound_pwq(wq, cpu, pwq);
5427	out_unlock:
5428	mutex_unlock(&wq->mutex);
5429	put_pwq_unlocked(old_pwq);
5430	}
5431
5432	static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5433	{
5434	bool highpri = wq->flags & WQ_HIGHPRI;
5435	int cpu, ret;
5436
5437	lockdep_assert_held(&wq_pool_mutex);
5438
5439	wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
5440	if (!wq->cpu_pwq)
5441	goto enomem;
5442
5443	if (!(wq->flags & WQ_UNBOUND)) {
5444	struct worker_pool __percpu *pools;
5445
5446	if (wq->flags & WQ_BH)
5447	pools = bh_worker_pools;
5448	else
5449	pools = cpu_worker_pools;
5450
5451	for_each_possible_cpu(cpu) {
5452	struct pool_workqueue **pwq_p;
5453	struct worker_pool *pool;
5454
5455	pool = &(per_cpu_ptr(pools, cpu)[highpri]);
5456	pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu);
5457
5458	*pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5459	pool->node);
5460	if (!*pwq_p)
5461	goto enomem;
5462
5463	init_pwq(*pwq_p, wq, pool);
5464
5465	mutex_lock(&wq->mutex);
5466	link_pwq(*pwq_p);
5467	mutex_unlock(&wq->mutex);
5468	}
5469	return 0;
5470	}
5471
5472	if (wq->flags & __WQ_ORDERED) {
5473	struct pool_workqueue *dfl_pwq;
5474
5475	ret = apply_workqueue_attrs_locked(wq, ordered_wq_attrs[highpri]);
5476	/* there should only be single pwq for ordering guarantee */
5477	dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5478	WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node \|\|
5479	wq->pwqs.prev != &dfl_pwq->pwqs_node),
5480	"ordering guarantee broken for workqueue %s\n", wq->name);
5481	} else {
5482	ret = apply_workqueue_attrs_locked(wq, unbound_std_wq_attrs[highpri]);
5483	}
5484
5485	return ret;
5486
5487	enomem:
5488	if (wq->cpu_pwq) {
5489	for_each_possible_cpu(cpu) {
5490	struct pool_workqueue pwq = per_cpu_ptr(wq->cpu_pwq, cpu);
5491
5492	if (pwq)
5493	kmem_cache_free(pwq_cache, pwq);
5494	}
5495	free_percpu(wq->cpu_pwq);
5496	wq->cpu_pwq = NULL;
5497	}
5498	return -ENOMEM;
5499	}
5500
5501	static int wq_clamp_max_active(int max_active, unsigned int flags,
5502	const char *name)
5503	{
5504	if (max_active < 1 \|\| max_active > WQ_MAX_ACTIVE)
5505	pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5506	max_active, name, 1, WQ_MAX_ACTIVE);
5507
5508	return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5509	}
5510
5511	/*
5512	* Workqueues which may be used during memory reclaim should have a rescuer
5513	* to guarantee forward progress.
5514	*/
5515	static int init_rescuer(struct workqueue_struct *wq)
5516	{
5517	struct worker *rescuer;
5518	char id_buf[WORKER_ID_LEN];
5519	int ret;
5520
5521	lockdep_assert_held(&wq_pool_mutex);
5522
5523	if (!(wq->flags & WQ_MEM_RECLAIM))
5524	return 0;
5525
5526	rescuer = alloc_worker(NUMA_NO_NODE);
5527	if (!rescuer) {
5528	pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5529	wq->name);
5530	return -ENOMEM;
5531	}
5532
5533	rescuer->rescue_wq = wq;
5534	format_worker_id(id_buf, sizeof(id_buf), rescuer, NULL);
5535
5536	rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", id_buf);
5537	if (IS_ERR(rescuer->task)) {
5538	ret = PTR_ERR(rescuer->task);
5539	pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5540	wq->name, ERR_PTR(ret));
5541	kfree(rescuer);
5542	return ret;
5543	}
5544
5545	wq->rescuer = rescuer;
5546	if (wq->flags & WQ_UNBOUND)
5547	kthread_bind_mask(rescuer->task, unbound_effective_cpumask(wq));
5548	else
5549	kthread_bind_mask(rescuer->task, cpu_possible_mask);
5550	wake_up_process(rescuer->task);
5551
5552	return 0;
5553	}
5554
5555	/**
5556	* wq_adjust_max_active - update a wq's max_active to the current setting
5557	* @wq: target workqueue
5558	*
5559	* If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5560	* activate inactive work items accordingly. If @wq is freezing, clear
5561	* @wq->max_active to zero.
5562	*/
5563	static void wq_adjust_max_active(struct workqueue_struct *wq)
5564	{
5565	bool activated;
5566	int new_max, new_min;
5567
5568	lockdep_assert_held(&wq->mutex);
5569
5570	if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5571	new_max = 0;
5572	new_min = 0;
5573	} else {
5574	new_max = wq->saved_max_active;
5575	new_min = wq->saved_min_active;
5576	}
5577
5578	if (wq->max_active == new_max && wq->min_active == new_min)
5579	return;
5580
5581	/*
5582	* Update @wq->max/min_active and then kick inactive work items if more
5583	* active work items are allowed. This doesn't break work item ordering
5584	* because new work items are always queued behind existing inactive
5585	* work items if there are any.
5586	*/
5587	WRITE_ONCE(wq->max_active, new_max);
5588	WRITE_ONCE(wq->min_active, new_min);
5589
5590	if (wq->flags & WQ_UNBOUND)
5591	wq_update_node_max_active(wq, -1);
5592
5593	if (new_max == 0)
5594	return;
5595
5596	/*
5597	* Round-robin through pwq's activating the first inactive work item
5598	* until max_active is filled.
5599	*/
5600	do {
5601	struct pool_workqueue *pwq;
5602
5603	activated = false;
5604	for_each_pwq(pwq, wq) {
5605	unsigned long irq_flags;
5606
5607	/* can be called during early boot w/ irq disabled */
5608	raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
5609	if (pwq_activate_first_inactive(pwq, true)) {
5610	activated = true;
5611	kick_pool(pwq->pool);
5612	}
5613	raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
5614	}
5615	} while (activated);
5616	}
5617
5618	__printf(1, 4)
5619	struct workqueue_struct alloc_workqueue(const char fmt,
5620	unsigned int flags,
5621	int max_active, ...)
5622	{
5623	va_list args;
5624	struct workqueue_struct *wq;
5625	size_t wq_size;
5626	int name_len;
5627
5628	if (flags & WQ_BH) {
5629	if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS))
5630	return NULL;
5631	if (WARN_ON_ONCE(max_active))
5632	return NULL;
5633	}
5634
5635	/* see the comment above the definition of WQ_POWER_EFFICIENT */
5636	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5637	flags \|= WQ_UNBOUND;
5638
5639	/* allocate wq and format name */
5640	if (flags & WQ_UNBOUND)
5641	wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5642	else
5643	wq_size = sizeof(*wq);
5644
5645	wq = kzalloc(wq_size, GFP_KERNEL);
5646	if (!wq)
5647	return NULL;
5648
5649	if (flags & WQ_UNBOUND) {
5650	wq->unbound_attrs = alloc_workqueue_attrs();
5651	if (!wq->unbound_attrs)
5652	goto err_free_wq;
5653	}
5654
5655	va_start(args, max_active);
5656	name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5657	va_end(args);
5658
5659	if (name_len >= WQ_NAME_LEN)
5660	pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5661	wq->name);
5662
5663	if (flags & WQ_BH) {
5664	/*
5665	* BH workqueues always share a single execution context per CPU
5666	* and don't impose any max_active limit.
5667	*/
5668	max_active = INT_MAX;
5669	} else {
5670	max_active = max_active ?: WQ_DFL_ACTIVE;
5671	max_active = wq_clamp_max_active(max_active, flags, wq->name);
5672	}
5673
5674	/* init wq */
5675	wq->flags = flags;
5676	wq->max_active = max_active;
5677	wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5678	wq->saved_max_active = wq->max_active;
5679	wq->saved_min_active = wq->min_active;
5680	mutex_init(&wq->mutex);
5681	atomic_set(&wq->nr_pwqs_to_flush, 0);
5682	INIT_LIST_HEAD(&wq->pwqs);
5683	INIT_LIST_HEAD(&wq->flusher_queue);
5684	INIT_LIST_HEAD(&wq->flusher_overflow);
5685	INIT_LIST_HEAD(&wq->maydays);
5686
5687	wq_init_lockdep(wq);
5688	INIT_LIST_HEAD(&wq->list);
5689
5690	if (flags & WQ_UNBOUND) {
5691	if (alloc_node_nr_active(wq->node_nr_active) < 0)
5692	goto err_unreg_lockdep;
5693	}
5694
5695	/*
5696	* wq_pool_mutex protects the workqueues list, allocations of PWQs,
5697	* and the global freeze state.
5698	*/
5699	apply_wqattrs_lock();
5700
5701	if (alloc_and_link_pwqs(wq) < 0)
5702	goto err_unlock_free_node_nr_active;
5703
5704	mutex_lock(&wq->mutex);
5705	wq_adjust_max_active(wq);
5706	mutex_unlock(&wq->mutex);
5707
5708	list_add_tail_rcu(&wq->list, &workqueues);
5709
5710	if (wq_online && init_rescuer(wq) < 0)
5711	goto err_unlock_destroy;
5712
5713	apply_wqattrs_unlock();
5714
5715	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5716	goto err_destroy;
5717
5718	return wq;
5719
5720	err_unlock_free_node_nr_active:
5721	apply_wqattrs_unlock();
5722	/*
5723	* Failed alloc_and_link_pwqs() may leave pending pwq->release_work,
5724	* flushing the pwq_release_worker ensures that the pwq_release_workfn()
5725	* completes before calling kfree(wq).
5726	*/
5727	if (wq->flags & WQ_UNBOUND) {
5728	kthread_flush_worker(pwq_release_worker);
5729	free_node_nr_active(wq->node_nr_active);
5730	}
5731	err_unreg_lockdep:
5732	wq_unregister_lockdep(wq);
5733	wq_free_lockdep(wq);
5734	err_free_wq:
5735	free_workqueue_attrs(wq->unbound_attrs);
5736	kfree(wq);
5737	return NULL;
5738	err_unlock_destroy:
5739	apply_wqattrs_unlock();
5740	err_destroy:
5741	destroy_workqueue(wq);
5742	return NULL;
5743	}
5744	EXPORT_SYMBOL_GPL(alloc_workqueue);
5745
5746	static bool pwq_busy(struct pool_workqueue *pwq)
5747	{
5748	int i;
5749
5750	for (i = 0; i < WORK_NR_COLORS; i++)
5751	if (pwq->nr_in_flight[i])
5752	return true;
5753
5754	if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
5755	return true;
5756	if (!pwq_is_empty(pwq))
5757	return true;
5758
5759	return false;
5760	}
5761
5762	/**
5763	* destroy_workqueue - safely terminate a workqueue
5764	* @wq: target workqueue
5765	*
5766	* Safely destroy a workqueue. All work currently pending will be done first.
5767	*/
5768	void destroy_workqueue(struct workqueue_struct *wq)
5769	{
5770	struct pool_workqueue *pwq;
5771	int cpu;
5772
5773	/*
5774	* Remove it from sysfs first so that sanity check failure doesn't
5775	* lead to sysfs name conflicts.
5776	*/
5777	workqueue_sysfs_unregister(wq);
5778
5779	/* mark the workqueue destruction is in progress */
5780	mutex_lock(&wq->mutex);
5781	wq->flags \|= __WQ_DESTROYING;
5782	mutex_unlock(&wq->mutex);
5783
5784	/* drain it before proceeding with destruction */
5785	drain_workqueue(wq);
5786
5787	/* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5788	if (wq->rescuer) {
5789	struct worker *rescuer = wq->rescuer;
5790
5791	/* this prevents new queueing */
5792	raw_spin_lock_irq(&wq_mayday_lock);
5793	wq->rescuer = NULL;
5794	raw_spin_unlock_irq(&wq_mayday_lock);
5795
5796	/* rescuer will empty maydays list before exiting */
5797	kthread_stop(rescuer->task);
5798	kfree(rescuer);
5799	}
5800
5801	/*
5802	* Sanity checks - grab all the locks so that we wait for all
5803	* in-flight operations which may do put_pwq().
5804	*/
5805	mutex_lock(&wq_pool_mutex);
5806	mutex_lock(&wq->mutex);
5807	for_each_pwq(pwq, wq) {
5808	raw_spin_lock_irq(&pwq->pool->lock);
5809	if (WARN_ON(pwq_busy(pwq))) {
5810	pr_warn("%s: %s has the following busy pwq\n",
5811	__func__, wq->name);
5812	show_pwq(pwq);
5813	raw_spin_unlock_irq(&pwq->pool->lock);
5814	mutex_unlock(&wq->mutex);
5815	mutex_unlock(&wq_pool_mutex);
5816	show_one_workqueue(wq);
5817	return;
5818	}
5819	raw_spin_unlock_irq(&pwq->pool->lock);
5820	}
5821	mutex_unlock(&wq->mutex);
5822
5823	/*
5824	* wq list is used to freeze wq, remove from list after
5825	* flushing is complete in case freeze races us.
5826	*/
5827	list_del_rcu(&wq->list);
5828	mutex_unlock(&wq_pool_mutex);
5829
5830	/*
5831	* We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5832	* to put the base refs. @wq will be auto-destroyed from the last
5833	* pwq_put. RCU read lock prevents @wq from going away from under us.
5834	*/
5835	rcu_read_lock();
5836
5837	for_each_possible_cpu(cpu) {
5838	put_pwq_unlocked(unbound_pwq(wq, cpu));
5839	RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
5840	}
5841
5842	put_pwq_unlocked(unbound_pwq(wq, -1));
5843	RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
5844
5845	rcu_read_unlock();
5846	}
5847	EXPORT_SYMBOL_GPL(destroy_workqueue);
5848
5849	/**
5850	* workqueue_set_max_active - adjust max_active of a workqueue
5851	* @wq: target workqueue
5852	* @max_active: new max_active value.
5853	*
5854	* Set max_active of @wq to @max_active. See the alloc_workqueue() function
5855	* comment.
5856	*
5857	* CONTEXT:
5858	* Don't call from IRQ context.
5859	*/
5860	void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5861	{
5862	/* max_active doesn't mean anything for BH workqueues */
5863	if (WARN_ON(wq->flags & WQ_BH))
5864	return;
5865	/* disallow meddling with max_active for ordered workqueues */
5866	if (WARN_ON(wq->flags & __WQ_ORDERED))
5867	return;
5868
5869	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5870
5871	mutex_lock(&wq->mutex);
5872
5873	wq->saved_max_active = max_active;
5874	if (wq->flags & WQ_UNBOUND)
5875	wq->saved_min_active = min(wq->saved_min_active, max_active);
5876
5877	wq_adjust_max_active(wq);
5878
5879	mutex_unlock(&wq->mutex);
5880	}
5881	EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5882
5883	/**
5884	* workqueue_set_min_active - adjust min_active of an unbound workqueue
5885	* @wq: target unbound workqueue
5886	* @min_active: new min_active value
5887	*
5888	* Set min_active of an unbound workqueue. Unlike other types of workqueues, an
5889	* unbound workqueue is not guaranteed to be able to process max_active
5890	* interdependent work items. Instead, an unbound workqueue is guaranteed to be
5891	* able to process min_active number of interdependent work items which is
5892	* %WQ_DFL_MIN_ACTIVE by default.
5893	*
5894	* Use this function to adjust the min_active value between 0 and the current
5895	* max_active.
5896	*/
5897	void workqueue_set_min_active(struct workqueue_struct *wq, int min_active)
5898	{
5899	/* min_active is only meaningful for non-ordered unbound workqueues */
5900	if (WARN_ON((wq->flags & (WQ_BH \| WQ_UNBOUND \| __WQ_ORDERED)) !=
5901	WQ_UNBOUND))
5902	return;
5903
5904	mutex_lock(&wq->mutex);
5905	wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active);
5906	wq_adjust_max_active(wq);
5907	mutex_unlock(&wq->mutex);
5908	}
5909
5910	/**
5911	* current_work - retrieve %current task's work struct
5912	*
5913	* Determine if %current task is a workqueue worker and what it's working on.
5914	* Useful to find out the context that the %current task is running in.
5915	*
5916	* Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5917	*/
5918	struct work_struct *current_work(void)
5919	{
5920	struct worker *worker = current_wq_worker();
5921
5922	return worker ? worker->current_work : NULL;
5923	}
5924	EXPORT_SYMBOL(current_work);
5925
5926	/**
5927	* current_is_workqueue_rescuer - is %current workqueue rescuer?
5928	*
5929	* Determine whether %current is a workqueue rescuer. Can be used from
5930	* work functions to determine whether it's being run off the rescuer task.
5931	*
5932	* Return: %true if %current is a workqueue rescuer. %false otherwise.
5933	*/
5934	bool current_is_workqueue_rescuer(void)
5935	{
5936	struct worker *worker = current_wq_worker();
5937
5938	return worker && worker->rescue_wq;
5939	}
5940
5941	/**
5942	* workqueue_congested - test whether a workqueue is congested
5943	* @cpu: CPU in question
5944	* @wq: target workqueue
5945	*
5946	* Test whether @wq's cpu workqueue for @cpu is congested. There is
5947	* no synchronization around this function and the test result is
5948	* unreliable and only useful as advisory hints or for debugging.
5949	*
5950	* If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
5951	*
5952	* With the exception of ordered workqueues, all workqueues have per-cpu
5953	* pool_workqueues, each with its own congested state. A workqueue being
5954	* congested on one CPU doesn't mean that the workqueue is contested on any
5955	* other CPUs.
5956	*
5957	* Return:
5958	* %true if congested, %false otherwise.
5959	*/
5960	bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5961	{
5962	struct pool_workqueue *pwq;
5963	bool ret;
5964
5965	rcu_read_lock();
5966	preempt_disable();
5967
5968	if (cpu == WORK_CPU_UNBOUND)
5969	cpu = smp_processor_id();
5970
5971	pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5972	ret = !list_empty(&pwq->inactive_works);
5973
5974	preempt_enable();
5975	rcu_read_unlock();
5976
5977	return ret;
5978	}
5979	EXPORT_SYMBOL_GPL(workqueue_congested);
5980
5981	/**
5982	* work_busy - test whether a work is currently pending or running
5983	* @work: the work to be tested
5984	*
5985	* Test whether @work is currently pending or running. There is no
5986	* synchronization around this function and the test result is
5987	* unreliable and only useful as advisory hints or for debugging.
5988	*
5989	* Return:
5990	* OR'd bitmask of WORK_BUSY_* bits.
5991	*/
5992	unsigned int work_busy(struct work_struct *work)
5993	{
5994	struct worker_pool *pool;
5995	unsigned long irq_flags;
5996	unsigned int ret = 0;
5997
5998	if (work_pending(work))
5999	ret \|= WORK_BUSY_PENDING;
6000
6001	rcu_read_lock();
6002	pool = get_work_pool(work);
6003	if (pool) {
6004	raw_spin_lock_irqsave(&pool->lock, irq_flags);
6005	if (find_worker_executing_work(pool, work))
6006	ret \|= WORK_BUSY_RUNNING;
6007	raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6008	}
6009	rcu_read_unlock();
6010
6011	return ret;
6012	}
6013	EXPORT_SYMBOL_GPL(work_busy);
6014
6015	/**
6016	* set_worker_desc - set description for the current work item
6017	* @fmt: printf-style format string
6018	* @...: arguments for the format string
6019	*
6020	* This function can be called by a running work function to describe what
6021	* the work item is about. If the worker task gets dumped, this
6022	* information will be printed out together to help debugging. The
6023	* description can be at most WORKER_DESC_LEN including the trailing '\0'.
6024	*/
6025	void set_worker_desc(const char *fmt, ...)
6026	{
6027	struct worker *worker = current_wq_worker();
6028	va_list args;
6029
6030	if (worker) {
6031	va_start(args, fmt);
6032	vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
6033	va_end(args);
6034	}
6035	}
6036	EXPORT_SYMBOL_GPL(set_worker_desc);
6037
6038	/**
6039	* print_worker_info - print out worker information and description
6040	* @log_lvl: the log level to use when printing
6041	* @task: target task
6042	*
6043	* If @task is a worker and currently executing a work item, print out the
6044	* name of the workqueue being serviced and worker description set with
6045	* set_worker_desc() by the currently executing work item.
6046	*
6047	* This function can be safely called on any task as long as the
6048	* task_struct itself is accessible. While safe, this function isn't
6049	* synchronized and may print out mixups or garbages of limited length.
6050	*/
6051	void print_worker_info(const char log_lvl, struct task_struct task)
6052	{
6053	work_func_t *fn = NULL;
6054	char name[WQ_NAME_LEN] = { };
6055	char desc[WORKER_DESC_LEN] = { };
6056	struct pool_workqueue *pwq = NULL;
6057	struct workqueue_struct *wq = NULL;
6058	struct worker *worker;
6059
6060	if (!(task->flags & PF_WQ_WORKER))
6061	return;
6062
6063	/*
6064	* This function is called without any synchronization and @task
6065	* could be in any state. Be careful with dereferences.
6066	*/
6067	worker = kthread_probe_data(task);
6068
6069	/*
6070	* Carefully copy the associated workqueue's workfn, name and desc.
6071	* Keep the original last '\0' in case the original is garbage.
6072	*/
6073	copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
6074	copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
6075	copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
6076	copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
6077	copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
6078
6079	if (fn \|\| name[0] \|\| desc[0]) {
6080	printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
6081	if (strcmp(name, desc))
6082	pr_cont(" (%s)", desc);
6083	pr_cont("\n");
6084	}
6085	}
6086
6087	static void pr_cont_pool_info(struct worker_pool *pool)
6088	{
6089	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
6090	if (pool->node != NUMA_NO_NODE)
6091	pr_cont(" node=%d", pool->node);
6092	pr_cont(" flags=0x%x", pool->flags);
6093	if (pool->flags & POOL_BH)
6094	pr_cont(" bh%s",
6095	pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6096	else
6097	pr_cont(" nice=%d", pool->attrs->nice);
6098	}
6099
6100	static void pr_cont_worker_id(struct worker *worker)
6101	{
6102	struct worker_pool *pool = worker->pool;
6103
6104	if (pool->flags & WQ_BH)
6105	pr_cont("bh%s",
6106	pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6107	else
6108	pr_cont("%d%s", task_pid_nr(worker->task),
6109	worker->rescue_wq ? "(RESCUER)" : "");
6110	}
6111
6112	struct pr_cont_work_struct {
6113	bool comma;
6114	work_func_t func;
6115	long ctr;
6116	};
6117
6118	static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
6119	{
6120	if (!pcwsp->ctr)
6121	goto out_record;
6122	if (func == pcwsp->func) {
6123	pcwsp->ctr++;
6124	return;
6125	}
6126	if (pcwsp->ctr == 1)
6127	pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
6128	else
6129	pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
6130	pcwsp->ctr = 0;
6131	out_record:
6132	if ((long)func == -1L)
6133	return;
6134	pcwsp->comma = comma;
6135	pcwsp->func = func;
6136	pcwsp->ctr = 1;
6137	}
6138
6139	static void pr_cont_work(bool comma, struct work_struct work, struct pr_cont_work_struct pcwsp)
6140	{
6141	if (work->func == wq_barrier_func) {
6142	struct wq_barrier *barr;
6143
6144	barr = container_of(work, struct wq_barrier, work);
6145
6146	pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6147	pr_cont("%s BAR(%d)", comma ? "," : "",
6148	task_pid_nr(barr->task));
6149	} else {
6150	if (!comma)
6151	pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6152	pr_cont_work_flush(comma, work->func, pcwsp);
6153	}
6154	}
6155
6156	static void show_pwq(struct pool_workqueue *pwq)
6157	{
6158	struct pr_cont_work_struct pcws = { .ctr = 0, };
6159	struct worker_pool *pool = pwq->pool;
6160	struct work_struct *work;
6161	struct worker *worker;
6162	bool has_in_flight = false, has_pending = false;
6163	int bkt;
6164
6165	pr_info(" pwq %d:", pool->id);
6166	pr_cont_pool_info(pool);
6167
6168	pr_cont(" active=%d refcnt=%d%s\n",
6169	pwq->nr_active, pwq->refcnt,
6170	!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6171
6172	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6173	if (worker->current_pwq == pwq) {
6174	has_in_flight = true;
6175	break;
6176	}
6177	}
6178	if (has_in_flight) {
6179	bool comma = false;
6180
6181	pr_info(" in-flight:");
6182	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6183	if (worker->current_pwq != pwq)
6184	continue;
6185
6186	pr_cont(" %s", comma ? "," : "");
6187	pr_cont_worker_id(worker);
6188	pr_cont(":%ps", worker->current_func);
6189	list_for_each_entry(work, &worker->scheduled, entry)
6190	pr_cont_work(false, work, &pcws);
6191	pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6192	comma = true;
6193	}
6194	pr_cont("\n");
6195	}
6196
6197	list_for_each_entry(work, &pool->worklist, entry) {
6198	if (get_work_pwq(work) == pwq) {
6199	has_pending = true;
6200	break;
6201	}
6202	}
6203	if (has_pending) {
6204	bool comma = false;
6205
6206	pr_info(" pending:");
6207	list_for_each_entry(work, &pool->worklist, entry) {
6208	if (get_work_pwq(work) != pwq)
6209	continue;
6210
6211	pr_cont_work(comma, work, &pcws);
6212	comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6213	}
6214	pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6215	pr_cont("\n");
6216	}
6217
6218	if (!list_empty(&pwq->inactive_works)) {
6219	bool comma = false;
6220
6221	pr_info(" inactive:");
6222	list_for_each_entry(work, &pwq->inactive_works, entry) {
6223	pr_cont_work(comma, work, &pcws);
6224	comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6225	}
6226	pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6227	pr_cont("\n");
6228	}
6229	}
6230
6231	/**
6232	* show_one_workqueue - dump state of specified workqueue
6233	* @wq: workqueue whose state will be printed
6234	*/
6235	void show_one_workqueue(struct workqueue_struct *wq)
6236	{
6237	struct pool_workqueue *pwq;
6238	bool idle = true;
6239	unsigned long irq_flags;
6240
6241	for_each_pwq(pwq, wq) {
6242	if (!pwq_is_empty(pwq)) {
6243	idle = false;
6244	break;
6245	}
6246	}
6247	if (idle) /* Nothing to print for idle workqueue */
6248	return;
6249
6250	pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6251
6252	for_each_pwq(pwq, wq) {
6253	raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
6254	if (!pwq_is_empty(pwq)) {
6255	/*
6256	* Defer printing to avoid deadlocks in console
6257	* drivers that queue work while holding locks
6258	* also taken in their write paths.
6259	*/
6260	printk_deferred_enter();
6261	show_pwq(pwq);
6262	printk_deferred_exit();
6263	}
6264	raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
6265	/*
6266	* We could be printing a lot from atomic context, e.g.
6267	* sysrq-t -> show_all_workqueues(). Avoid triggering
6268	* hard lockup.
6269	*/
6270	touch_nmi_watchdog();
6271	}
6272
6273	}
6274
6275	/**
6276	* show_one_worker_pool - dump state of specified worker pool
6277	* @pool: worker pool whose state will be printed
6278	*/
6279	static void show_one_worker_pool(struct worker_pool *pool)
6280	{
6281	struct worker *worker;
6282	bool first = true;
6283	unsigned long irq_flags;
6284	unsigned long hung = 0;
6285
6286	raw_spin_lock_irqsave(&pool->lock, irq_flags);
6287	if (pool->nr_workers == pool->nr_idle)
6288	goto next_pool;
6289
6290	/* How long the first pending work is waiting for a worker. */
6291	if (!list_empty(&pool->worklist))
6292	hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
6293
6294	/*
6295	* Defer printing to avoid deadlocks in console drivers that
6296	* queue work while holding locks also taken in their write
6297	* paths.
6298	*/
6299	printk_deferred_enter();
6300	pr_info("pool %d:", pool->id);
6301	pr_cont_pool_info(pool);
6302	pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
6303	if (pool->manager)
6304	pr_cont(" manager: %d",
6305	task_pid_nr(pool->manager->task));
6306	list_for_each_entry(worker, &pool->idle_list, entry) {
6307	pr_cont(" %s", first ? "idle: " : "");
6308	pr_cont_worker_id(worker);
6309	first = false;
6310	}
6311	pr_cont("\n");
6312	printk_deferred_exit();
6313	next_pool:
6314	raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6315	/*
6316	* We could be printing a lot from atomic context, e.g.
6317	* sysrq-t -> show_all_workqueues(). Avoid triggering
6318	* hard lockup.
6319	*/
6320	touch_nmi_watchdog();
6321
6322	}
6323
6324	/**
6325	* show_all_workqueues - dump workqueue state
6326	*
6327	* Called from a sysrq handler and prints out all busy workqueues and pools.
6328	*/
6329	void show_all_workqueues(void)
6330	{
6331	struct workqueue_struct *wq;
6332	struct worker_pool *pool;
6333	int pi;
6334
6335	rcu_read_lock();
6336
6337	pr_info("Showing busy workqueues and worker pools:\n");
6338
6339	list_for_each_entry_rcu(wq, &workqueues, list)
6340	show_one_workqueue(wq);
6341
6342	for_each_pool(pool, pi)
6343	show_one_worker_pool(pool);
6344
6345	rcu_read_unlock();
6346	}
6347
6348	/**
6349	* show_freezable_workqueues - dump freezable workqueue state
6350	*
6351	* Called from try_to_freeze_tasks() and prints out all freezable workqueues
6352	* still busy.
6353	*/
6354	void show_freezable_workqueues(void)
6355	{
6356	struct workqueue_struct *wq;
6357
6358	rcu_read_lock();
6359
6360	pr_info("Showing freezable workqueues that are still busy:\n");
6361
6362	list_for_each_entry_rcu(wq, &workqueues, list) {
6363	if (!(wq->flags & WQ_FREEZABLE))
6364	continue;
6365	show_one_workqueue(wq);
6366	}
6367
6368	rcu_read_unlock();
6369	}
6370
6371	/* used to show worker information through /proc/PID/{comm,stat,status} */
6372	void wq_worker_comm(char buf, size_t size, struct task_struct task)
6373	{
6374	/* stabilize PF_WQ_WORKER and worker pool association */
6375	mutex_lock(&wq_pool_attach_mutex);
6376
6377	if (task->flags & PF_WQ_WORKER) {
6378	struct worker *worker = kthread_data(task);
6379	struct worker_pool *pool = worker->pool;
6380	int off;
6381
6382	off = format_worker_id(buf, size, worker, pool);
6383
6384	if (pool) {
6385	raw_spin_lock_irq(&pool->lock);
6386	/*
6387	* ->desc tracks information (wq name or
6388	* set_worker_desc()) for the latest execution. If
6389	* current, prepend '+', otherwise '-'.
6390	*/
6391	if (worker->desc[0] != '\0') {
6392	if (worker->current_work)
6393	scnprintf(buf + off, size - off, "+%s",
6394	worker->desc);
6395	else
6396	scnprintf(buf + off, size - off, "-%s",
6397	worker->desc);
6398	}
6399	raw_spin_unlock_irq(&pool->lock);
6400	}
6401	} else {
6402	strscpy(buf, task->comm, size);
6403	}
6404
6405	mutex_unlock(&wq_pool_attach_mutex);
6406	}
6407
6408	#ifdef CONFIG_SMP
6409
6410	/*
6411	* CPU hotplug.
6412	*
6413	* There are two challenges in supporting CPU hotplug. Firstly, there
6414	* are a lot of assumptions on strong associations among work, pwq and
6415	* pool which make migrating pending and scheduled works very
6416	* difficult to implement without impacting hot paths. Secondly,
6417	* worker pools serve mix of short, long and very long running works making
6418	* blocked draining impractical.
6419	*
6420	* This is solved by allowing the pools to be disassociated from the CPU
6421	* running as an unbound one and allowing it to be reattached later if the
6422	* cpu comes back online.
6423	*/
6424
6425	static void unbind_workers(int cpu)
6426	{
6427	struct worker_pool *pool;
6428	struct worker *worker;
6429
6430	for_each_cpu_worker_pool(pool, cpu) {
6431	mutex_lock(&wq_pool_attach_mutex);
6432	raw_spin_lock_irq(&pool->lock);
6433
6434	/*
6435	* We've blocked all attach/detach operations. Make all workers
6436	* unbound and set DISASSOCIATED. Before this, all workers
6437	* must be on the cpu. After this, they may become diasporas.
6438	* And the preemption disabled section in their sched callbacks
6439	* are guaranteed to see WORKER_UNBOUND since the code here
6440	* is on the same cpu.
6441	*/
6442	for_each_pool_worker(worker, pool)
6443	worker->flags \|= WORKER_UNBOUND;
6444
6445	pool->flags \|= POOL_DISASSOCIATED;
6446
6447	/*
6448	* The handling of nr_running in sched callbacks are disabled
6449	* now. Zap nr_running. After this, nr_running stays zero and
6450	* need_more_worker() and keep_working() are always true as
6451	* long as the worklist is not empty. This pool now behaves as
6452	* an unbound (in terms of concurrency management) pool which
6453	* are served by workers tied to the pool.
6454	*/
6455	pool->nr_running = 0;
6456
6457	/*
6458	* With concurrency management just turned off, a busy
6459	* worker blocking could lead to lengthy stalls. Kick off
6460	* unbound chain execution of currently pending work items.
6461	*/
6462	kick_pool(pool);
6463
6464	raw_spin_unlock_irq(&pool->lock);
6465
6466	for_each_pool_worker(worker, pool)
6467	unbind_worker(worker);
6468
6469	mutex_unlock(&wq_pool_attach_mutex);
6470	}
6471	}
6472
6473	/**
6474	* rebind_workers - rebind all workers of a pool to the associated CPU
6475	* @pool: pool of interest
6476	*
6477	* @pool->cpu is coming online. Rebind all workers to the CPU.
6478	*/
6479	static void rebind_workers(struct worker_pool *pool)
6480	{
6481	struct worker *worker;
6482
6483	lockdep_assert_held(&wq_pool_attach_mutex);
6484
6485	/*
6486	* Restore CPU affinity of all workers. As all idle workers should
6487	* be on the run-queue of the associated CPU before any local
6488	* wake-ups for concurrency management happen, restore CPU affinity
6489	* of all workers first and then clear UNBOUND. As we're called
6490	* from CPU_ONLINE, the following shouldn't fail.
6491	*/
6492	for_each_pool_worker(worker, pool) {
6493	kthread_set_per_cpu(worker->task, pool->cpu);
6494	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6495	pool_allowed_cpus(pool)) < 0);
6496	}
6497
6498	raw_spin_lock_irq(&pool->lock);
6499
6500	pool->flags &= ~POOL_DISASSOCIATED;
6501
6502	for_each_pool_worker(worker, pool) {
6503	unsigned int worker_flags = worker->flags;
6504
6505	/*
6506	* We want to clear UNBOUND but can't directly call
6507	* worker_clr_flags() or adjust nr_running. Atomically
6508	* replace UNBOUND with another NOT_RUNNING flag REBOUND.
6509	* @worker will clear REBOUND using worker_clr_flags() when
6510	* it initiates the next execution cycle thus restoring
6511	* concurrency management. Note that when or whether
6512	* @worker clears REBOUND doesn't affect correctness.
6513	*
6514	* WRITE_ONCE() is necessary because @worker->flags may be
6515	* tested without holding any lock in
6516	* wq_worker_running(). Without it, NOT_RUNNING test may
6517	* fail incorrectly leading to premature concurrency
6518	* management operations.
6519	*/
6520	WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6521	worker_flags \|= WORKER_REBOUND;
6522	worker_flags &= ~WORKER_UNBOUND;
6523	WRITE_ONCE(worker->flags, worker_flags);
6524	}
6525
6526	raw_spin_unlock_irq(&pool->lock);
6527	}
6528
6529	/**
6530	* restore_unbound_workers_cpumask - restore cpumask of unbound workers
6531	* @pool: unbound pool of interest
6532	* @cpu: the CPU which is coming up
6533	*
6534	* An unbound pool may end up with a cpumask which doesn't have any online
6535	* CPUs. When a worker of such pool get scheduled, the scheduler resets
6536	* its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6537	* online CPU before, cpus_allowed of all its workers should be restored.
6538	*/
6539	static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6540	{
6541	static cpumask_t cpumask;
6542	struct worker *worker;
6543
6544	lockdep_assert_held(&wq_pool_attach_mutex);
6545
6546	/* is @cpu allowed for @pool? */
6547	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6548	return;
6549
6550	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6551
6552	/* as we're called from CPU_ONLINE, the following shouldn't fail */
6553	for_each_pool_worker(worker, pool)
6554	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6555	}
6556
6557	int workqueue_prepare_cpu(unsigned int cpu)
6558	{
6559	struct worker_pool *pool;
6560
6561	for_each_cpu_worker_pool(pool, cpu) {
6562	if (pool->nr_workers)
6563	continue;
6564	if (!create_worker(pool))
6565	return -ENOMEM;
6566	}
6567	return 0;
6568	}
6569
6570	int workqueue_online_cpu(unsigned int cpu)
6571	{
6572	struct worker_pool *pool;
6573	struct workqueue_struct *wq;
6574	int pi;
6575
6576	mutex_lock(&wq_pool_mutex);
6577
6578	cpumask_set_cpu(cpu, wq_online_cpumask);
6579
6580	for_each_pool(pool, pi) {
6581	/* BH pools aren't affected by hotplug */
6582	if (pool->flags & POOL_BH)
6583	continue;
6584
6585	mutex_lock(&wq_pool_attach_mutex);
6586	if (pool->cpu == cpu)
6587	rebind_workers(pool);
6588	else if (pool->cpu < 0)
6589	restore_unbound_workers_cpumask(pool, cpu);
6590	mutex_unlock(&wq_pool_attach_mutex);
6591	}
6592
6593	/* update pod affinity of unbound workqueues */
6594	list_for_each_entry(wq, &workqueues, list) {
6595	struct workqueue_attrs *attrs = wq->unbound_attrs;
6596
6597	if (attrs) {
6598	const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6599	int tcpu;
6600
6601	for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6602	unbound_wq_update_pwq(wq, tcpu);
6603
6604	mutex_lock(&wq->mutex);
6605	wq_update_node_max_active(wq, -1);
6606	mutex_unlock(&wq->mutex);
6607	}
6608	}
6609
6610	mutex_unlock(&wq_pool_mutex);
6611	return 0;
6612	}
6613
6614	int workqueue_offline_cpu(unsigned int cpu)
6615	{
6616	struct workqueue_struct *wq;
6617
6618	/* unbinding per-cpu workers should happen on the local CPU */
6619	if (WARN_ON(cpu != smp_processor_id()))
6620	return -1;
6621
6622	unbind_workers(cpu);
6623
6624	/* update pod affinity of unbound workqueues */
6625	mutex_lock(&wq_pool_mutex);
6626
6627	cpumask_clear_cpu(cpu, wq_online_cpumask);
6628
6629	list_for_each_entry(wq, &workqueues, list) {
6630	struct workqueue_attrs *attrs = wq->unbound_attrs;
6631
6632	if (attrs) {
6633	const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6634	int tcpu;
6635
6636	for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6637	unbound_wq_update_pwq(wq, tcpu);
6638
6639	mutex_lock(&wq->mutex);
6640	wq_update_node_max_active(wq, cpu);
6641	mutex_unlock(&wq->mutex);
6642	}
6643	}
6644	mutex_unlock(&wq_pool_mutex);
6645
6646	return 0;
6647	}
6648
6649	struct work_for_cpu {
6650	struct work_struct work;
6651	long (fn)(void );
6652	void *arg;
6653	long ret;
6654	};
6655
6656	static void work_for_cpu_fn(struct work_struct *work)
6657	{
6658	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6659
6660	wfc->ret = wfc->fn(wfc->arg);
6661	}
6662
6663	/**
6664	* work_on_cpu_key - run a function in thread context on a particular cpu
6665	* @cpu: the cpu to run on
6666	* @fn: the function to run
6667	* @arg: the function arg
6668	* @key: The lock class key for lock debugging purposes
6669	*
6670	* It is up to the caller to ensure that the cpu doesn't go offline.
6671	* The caller must not hold any locks which would prevent @fn from completing.
6672	*
6673	* Return: The value @fn returns.
6674	*/
6675	long work_on_cpu_key(int cpu, long (fn)(void ),
6676	void arg, struct lock_class_key key)
6677	{
6678	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6679
6680	INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6681	schedule_work_on(cpu, &wfc.work);
6682	flush_work(&wfc.work);
6683	destroy_work_on_stack(&wfc.work);
6684	return wfc.ret;
6685	}
6686	EXPORT_SYMBOL_GPL(work_on_cpu_key);
6687
6688	/**
6689	* work_on_cpu_safe_key - run a function in thread context on a particular cpu
6690	* @cpu: the cpu to run on
6691	* @fn: the function to run
6692	* @arg: the function argument
6693	* @key: The lock class key for lock debugging purposes
6694	*
6695	* Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6696	* any locks which would prevent @fn from completing.
6697	*
6698	* Return: The value @fn returns.
6699	*/
6700	long work_on_cpu_safe_key(int cpu, long (fn)(void ),
6701	void arg, struct lock_class_key key)
6702	{
6703	long ret = -ENODEV;
6704
6705	cpus_read_lock();
6706	if (cpu_online(cpu))
6707	ret = work_on_cpu_key(cpu, fn, arg, key);
6708	cpus_read_unlock();
6709	return ret;
6710	}
6711	EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
6712	#endif /* CONFIG_SMP */
6713
6714	#ifdef CONFIG_FREEZER
6715
6716	/**
6717	* freeze_workqueues_begin - begin freezing workqueues
6718	*
6719	* Start freezing workqueues. After this function returns, all freezable
6720	* workqueues will queue new works to their inactive_works list instead of
6721	* pool->worklist.
6722	*
6723	* CONTEXT:
6724	* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6725	*/
6726	void freeze_workqueues_begin(void)
6727	{
6728	struct workqueue_struct *wq;
6729
6730	mutex_lock(&wq_pool_mutex);
6731
6732	WARN_ON_ONCE(workqueue_freezing);
6733	workqueue_freezing = true;
6734
6735	list_for_each_entry(wq, &workqueues, list) {
6736	mutex_lock(&wq->mutex);
6737	wq_adjust_max_active(wq);
6738	mutex_unlock(&wq->mutex);
6739	}
6740
6741	mutex_unlock(&wq_pool_mutex);
6742	}
6743
6744	/**
6745	* freeze_workqueues_busy - are freezable workqueues still busy?
6746	*
6747	* Check whether freezing is complete. This function must be called
6748	* between freeze_workqueues_begin() and thaw_workqueues().
6749	*
6750	* CONTEXT:
6751	* Grabs and releases wq_pool_mutex.
6752	*
6753	* Return:
6754	* %true if some freezable workqueues are still busy. %false if freezing
6755	* is complete.
6756	*/
6757	bool freeze_workqueues_busy(void)
6758	{
6759	bool busy = false;
6760	struct workqueue_struct *wq;
6761	struct pool_workqueue *pwq;
6762
6763	mutex_lock(&wq_pool_mutex);
6764
6765	WARN_ON_ONCE(!workqueue_freezing);
6766
6767	list_for_each_entry(wq, &workqueues, list) {
6768	if (!(wq->flags & WQ_FREEZABLE))
6769	continue;
6770	/*
6771	* nr_active is monotonically decreasing. It's safe
6772	* to peek without lock.
6773	*/
6774	rcu_read_lock();
6775	for_each_pwq(pwq, wq) {
6776	WARN_ON_ONCE(pwq->nr_active < 0);
6777	if (pwq->nr_active) {
6778	busy = true;
6779	rcu_read_unlock();
6780	goto out_unlock;
6781	}
6782	}
6783	rcu_read_unlock();
6784	}
6785	out_unlock:
6786	mutex_unlock(&wq_pool_mutex);
6787	return busy;
6788	}
6789
6790	/**
6791	* thaw_workqueues - thaw workqueues
6792	*
6793	* Thaw workqueues. Normal queueing is restored and all collected
6794	* frozen works are transferred to their respective pool worklists.
6795	*
6796	* CONTEXT:
6797	* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6798	*/
6799	void thaw_workqueues(void)
6800	{
6801	struct workqueue_struct *wq;
6802
6803	mutex_lock(&wq_pool_mutex);
6804
6805	if (!workqueue_freezing)
6806	goto out_unlock;
6807
6808	workqueue_freezing = false;
6809
6810	/* restore max_active and repopulate worklist */
6811	list_for_each_entry(wq, &workqueues, list) {
6812	mutex_lock(&wq->mutex);
6813	wq_adjust_max_active(wq);
6814	mutex_unlock(&wq->mutex);
6815	}
6816
6817	out_unlock:
6818	mutex_unlock(&wq_pool_mutex);
6819	}
6820	#endif /* CONFIG_FREEZER */
6821
6822	static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6823	{
6824	LIST_HEAD(ctxs);
6825	int ret = 0;
6826	struct workqueue_struct *wq;
6827	struct apply_wqattrs_ctx ctx, n;
6828
6829	lockdep_assert_held(&wq_pool_mutex);
6830
6831	list_for_each_entry(wq, &workqueues, list) {
6832	if (!(wq->flags & WQ_UNBOUND) \|\| (wq->flags & __WQ_DESTROYING))
6833	continue;
6834
6835	ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
6836	if (IS_ERR(ctx)) {
6837	ret = PTR_ERR(ctx);
6838	break;
6839	}
6840
6841	list_add_tail(&ctx->list, &ctxs);
6842	}
6843
6844	list_for_each_entry_safe(ctx, n, &ctxs, list) {
6845	if (!ret)
6846	apply_wqattrs_commit(ctx);
6847	apply_wqattrs_cleanup(ctx);
6848	}
6849
6850	if (!ret) {
6851	mutex_lock(&wq_pool_attach_mutex);
6852	cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
6853	mutex_unlock(&wq_pool_attach_mutex);
6854	}
6855	return ret;
6856	}
6857
6858	/**
6859	* workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
6860	* @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
6861	*
6862	* This function can be called from cpuset code to provide a set of isolated
6863	* CPUs that should be excluded from wq_unbound_cpumask.
6864	*/
6865	int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
6866	{
6867	cpumask_var_t cpumask;
6868	int ret = 0;
6869
6870	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6871	return -ENOMEM;
6872
6873	mutex_lock(&wq_pool_mutex);
6874
6875	/*
6876	* If the operation fails, it will fall back to
6877	* wq_requested_unbound_cpumask which is initially set to
6878	* (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
6879	* by any subsequent write to workqueue/cpumask sysfs file.
6880	*/
6881	if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
6882	cpumask_copy(cpumask, wq_requested_unbound_cpumask);
6883	if (!cpumask_equal(cpumask, wq_unbound_cpumask))
6884	ret = workqueue_apply_unbound_cpumask(cpumask);
6885
6886	/* Save the current isolated cpumask & export it via sysfs */
6887	if (!ret)
6888	cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
6889
6890	mutex_unlock(&wq_pool_mutex);
6891	free_cpumask_var(cpumask);
6892	return ret;
6893	}
6894
6895	static int parse_affn_scope(const char *val)
6896	{
6897	int i;
6898
6899	for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6900	if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
6901	return i;
6902	}
6903	return -EINVAL;
6904	}
6905
6906	static int wq_affn_dfl_set(const char val, const struct kernel_param kp)
6907	{
6908	struct workqueue_struct *wq;
6909	int affn, cpu;
6910
6911	affn = parse_affn_scope(val);
6912	if (affn < 0)
6913	return affn;
6914	if (affn == WQ_AFFN_DFL)
6915	return -EINVAL;
6916
6917	cpus_read_lock();
6918	mutex_lock(&wq_pool_mutex);
6919
6920	wq_affn_dfl = affn;
6921
6922	list_for_each_entry(wq, &workqueues, list) {
6923	for_each_online_cpu(cpu)
6924	unbound_wq_update_pwq(wq, cpu);
6925	}
6926
6927	mutex_unlock(&wq_pool_mutex);
6928	cpus_read_unlock();
6929
6930	return 0;
6931	}
6932
6933	static int wq_affn_dfl_get(char buffer, const struct kernel_param kp)
6934	{
6935	return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
6936	}
6937
6938	static const struct kernel_param_ops wq_affn_dfl_ops = {
6939	.set = wq_affn_dfl_set,
6940	.get = wq_affn_dfl_get,
6941	};
6942
6943	module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
6944
6945	#ifdef CONFIG_SYSFS
6946	/*
6947	* Workqueues with WQ_SYSFS flag set is visible to userland via
6948	* /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
6949	* following attributes.
6950	*
6951	* per_cpu RO bool : whether the workqueue is per-cpu or unbound
6952	* max_active RW int : maximum number of in-flight work items
6953	*
6954	* Unbound workqueues have the following extra attributes.
6955	*
6956	* nice RW int : nice value of the workers
6957	* cpumask RW mask : bitmask of allowed CPUs for the workers
6958	* affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
6959	* affinity_strict RW bool : worker CPU affinity is strict
6960	*/
6961	struct wq_device {
6962	struct workqueue_struct *wq;
6963	struct device dev;
6964	};
6965
6966	static struct workqueue_struct dev_to_wq(struct device dev)
6967	{
6968	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6969
6970	return wq_dev->wq;
6971	}
6972
6973	static ssize_t per_cpu_show(struct device dev, struct device_attribute attr,
6974	char *buf)
6975	{
6976	struct workqueue_struct *wq = dev_to_wq(dev);
6977
6978	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
6979	}
6980	static DEVICE_ATTR_RO(per_cpu);
6981
6982	static ssize_t max_active_show(struct device *dev,
6983	struct device_attribute attr, char buf)
6984	{
6985	struct workqueue_struct *wq = dev_to_wq(dev);
6986
6987	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
6988	}
6989
6990	static ssize_t max_active_store(struct device *dev,
6991	struct device_attribute attr, const char buf,
6992	size_t count)
6993	{
6994	struct workqueue_struct *wq = dev_to_wq(dev);
6995	int val;
6996
6997	if (sscanf(buf, "%d", &val) != 1 \|\| val <= 0)
6998	return -EINVAL;
6999
7000	workqueue_set_max_active(wq, val);
7001	return count;
7002	}
7003	static DEVICE_ATTR_RW(max_active);
7004
7005	static struct attribute *wq_sysfs_attrs[] = {
7006	&dev_attr_per_cpu.attr,
7007	&dev_attr_max_active.attr,
7008	NULL,
7009	};
7010	ATTRIBUTE_GROUPS(wq_sysfs);
7011
7012	static ssize_t wq_nice_show(struct device dev, struct device_attribute attr,
7013	char *buf)
7014	{
7015	struct workqueue_struct *wq = dev_to_wq(dev);
7016	int written;
7017
7018	mutex_lock(&wq->mutex);
7019	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
7020	mutex_unlock(&wq->mutex);
7021
7022	return written;
7023	}
7024
7025	/* prepare workqueue_attrs for sysfs store operations */
7026	static struct workqueue_attrs wq_sysfs_prep_attrs(struct workqueue_struct wq)
7027	{
7028	struct workqueue_attrs *attrs;
7029
7030	lockdep_assert_held(&wq_pool_mutex);
7031
7032	attrs = alloc_workqueue_attrs();
7033	if (!attrs)
7034	return NULL;
7035
7036	copy_workqueue_attrs(attrs, wq->unbound_attrs);
7037	return attrs;
7038	}
7039
7040	static ssize_t wq_nice_store(struct device dev, struct device_attribute attr,
7041	const char *buf, size_t count)
7042	{
7043	struct workqueue_struct *wq = dev_to_wq(dev);
7044	struct workqueue_attrs *attrs;
7045	int ret = -ENOMEM;
7046
7047	apply_wqattrs_lock();
7048
7049	attrs = wq_sysfs_prep_attrs(wq);
7050	if (!attrs)
7051	goto out_unlock;
7052
7053	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
7054	attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
7055	ret = apply_workqueue_attrs_locked(wq, attrs);
7056	else
7057	ret = -EINVAL;
7058
7059	out_unlock:
7060	apply_wqattrs_unlock();
7061	free_workqueue_attrs(attrs);
7062	return ret ?: count;
7063	}
7064
7065	static ssize_t wq_cpumask_show(struct device *dev,
7066	struct device_attribute attr, char buf)
7067	{
7068	struct workqueue_struct *wq = dev_to_wq(dev);
7069	int written;
7070
7071	mutex_lock(&wq->mutex);
7072	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
7073	cpumask_pr_args(wq->unbound_attrs->cpumask));
7074	mutex_unlock(&wq->mutex);
7075	return written;
7076	}
7077
7078	static ssize_t wq_cpumask_store(struct device *dev,
7079	struct device_attribute *attr,
7080	const char *buf, size_t count)
7081	{
7082	struct workqueue_struct *wq = dev_to_wq(dev);
7083	struct workqueue_attrs *attrs;
7084	int ret = -ENOMEM;
7085
7086	apply_wqattrs_lock();
7087
7088	attrs = wq_sysfs_prep_attrs(wq);
7089	if (!attrs)
7090	goto out_unlock;
7091
7092	ret = cpumask_parse(buf, attrs->cpumask);
7093	if (!ret)
7094	ret = apply_workqueue_attrs_locked(wq, attrs);
7095
7096	out_unlock:
7097	apply_wqattrs_unlock();
7098	free_workqueue_attrs(attrs);
7099	return ret ?: count;
7100	}
7101
7102	static ssize_t wq_affn_scope_show(struct device *dev,
7103	struct device_attribute attr, char buf)
7104	{
7105	struct workqueue_struct *wq = dev_to_wq(dev);
7106	int written;
7107
7108	mutex_lock(&wq->mutex);
7109	if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
7110	written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
7111	wq_affn_names[WQ_AFFN_DFL],
7112	wq_affn_names[wq_affn_dfl]);
7113	else
7114	written = scnprintf(buf, PAGE_SIZE, "%s\n",
7115	wq_affn_names[wq->unbound_attrs->affn_scope]);
7116	mutex_unlock(&wq->mutex);
7117
7118	return written;
7119	}
7120
7121	static ssize_t wq_affn_scope_store(struct device *dev,
7122	struct device_attribute *attr,
7123	const char *buf, size_t count)
7124	{
7125	struct workqueue_struct *wq = dev_to_wq(dev);
7126	struct workqueue_attrs *attrs;
7127	int affn, ret = -ENOMEM;
7128
7129	affn = parse_affn_scope(buf);
7130	if (affn < 0)
7131	return affn;
7132
7133	apply_wqattrs_lock();
7134	attrs = wq_sysfs_prep_attrs(wq);
7135	if (attrs) {
7136	attrs->affn_scope = affn;
7137	ret = apply_workqueue_attrs_locked(wq, attrs);
7138	}
7139	apply_wqattrs_unlock();
7140	free_workqueue_attrs(attrs);
7141	return ret ?: count;
7142	}
7143
7144	static ssize_t wq_affinity_strict_show(struct device *dev,
7145	struct device_attribute attr, char buf)
7146	{
7147	struct workqueue_struct *wq = dev_to_wq(dev);
7148
7149	return scnprintf(buf, PAGE_SIZE, "%d\n",
7150	wq->unbound_attrs->affn_strict);
7151	}
7152
7153	static ssize_t wq_affinity_strict_store(struct device *dev,
7154	struct device_attribute *attr,
7155	const char *buf, size_t count)
7156	{
7157	struct workqueue_struct *wq = dev_to_wq(dev);
7158	struct workqueue_attrs *attrs;
7159	int v, ret = -ENOMEM;
7160
7161	if (sscanf(buf, "%d", &v) != 1)
7162	return -EINVAL;
7163
7164	apply_wqattrs_lock();
7165	attrs = wq_sysfs_prep_attrs(wq);
7166	if (attrs) {
7167	attrs->affn_strict = (bool)v;
7168	ret = apply_workqueue_attrs_locked(wq, attrs);
7169	}
7170	apply_wqattrs_unlock();
7171	free_workqueue_attrs(attrs);
7172	return ret ?: count;
7173	}
7174
7175	static struct device_attribute wq_sysfs_unbound_attrs[] = {
7176	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7177	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7178	__ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
7179	__ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
7180	__ATTR_NULL,
7181	};
7182
7183	static const struct bus_type wq_subsys = {
7184	.name = "workqueue",
7185	.dev_groups = wq_sysfs_groups,
7186	};
7187
7188	/**
7189	* workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
7190	* @cpumask: the cpumask to set
7191	*
7192	* The low-level workqueues cpumask is a global cpumask that limits
7193	* the affinity of all unbound workqueues. This function check the @cpumask
7194	* and apply it to all unbound workqueues and updates all pwqs of them.
7195	*
7196	* Return: 0 - Success
7197	* -EINVAL - Invalid @cpumask
7198	* -ENOMEM - Failed to allocate memory for attrs or pwqs.
7199	*/
7200	static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
7201	{
7202	int ret = -EINVAL;
7203
7204	/*
7205	* Not excluding isolated cpus on purpose.
7206	* If the user wishes to include them, we allow that.
7207	*/
7208	cpumask_and(cpumask, cpumask, cpu_possible_mask);
7209	if (!cpumask_empty(cpumask)) {
7210	ret = 0;
7211	apply_wqattrs_lock();
7212	if (!cpumask_equal(cpumask, wq_unbound_cpumask))
7213	ret = workqueue_apply_unbound_cpumask(cpumask);
7214	if (!ret)
7215	cpumask_copy(wq_requested_unbound_cpumask, cpumask);
7216	apply_wqattrs_unlock();
7217	}
7218
7219	return ret;
7220	}
7221
7222	static ssize_t __wq_cpumask_show(struct device *dev,
7223	struct device_attribute attr, char buf, cpumask_var_t mask)
7224	{
7225	int written;
7226
7227	mutex_lock(&wq_pool_mutex);
7228	written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
7229	mutex_unlock(&wq_pool_mutex);
7230
7231	return written;
7232	}
7233
7234	static ssize_t cpumask_requested_show(struct device *dev,
7235	struct device_attribute attr, char buf)
7236	{
7237	return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
7238	}
7239	static DEVICE_ATTR_RO(cpumask_requested);
7240
7241	static ssize_t cpumask_isolated_show(struct device *dev,
7242	struct device_attribute attr, char buf)
7243	{
7244	return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
7245	}
7246	static DEVICE_ATTR_RO(cpumask_isolated);
7247
7248	static ssize_t cpumask_show(struct device *dev,
7249	struct device_attribute attr, char buf)
7250	{
7251	return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
7252	}
7253
7254	static ssize_t cpumask_store(struct device *dev,
7255	struct device_attribute attr, const char buf, size_t count)
7256	{
7257	cpumask_var_t cpumask;
7258	int ret;
7259
7260	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7261	return -ENOMEM;
7262
7263	ret = cpumask_parse(buf, cpumask);
7264	if (!ret)
7265	ret = workqueue_set_unbound_cpumask(cpumask);
7266
7267	free_cpumask_var(cpumask);
7268	return ret ? ret : count;
7269	}
7270	static DEVICE_ATTR_RW(cpumask);
7271
7272	static struct attribute *wq_sysfs_cpumask_attrs[] = {
7273	&dev_attr_cpumask.attr,
7274	&dev_attr_cpumask_requested.attr,
7275	&dev_attr_cpumask_isolated.attr,
7276	NULL,
7277	};
7278	ATTRIBUTE_GROUPS(wq_sysfs_cpumask);
7279
7280	static int __init wq_sysfs_init(void)
7281	{
7282	return subsys_virtual_register(&wq_subsys, wq_sysfs_cpumask_groups);
7283	}
7284	core_initcall(wq_sysfs_init);
7285
7286	static void wq_device_release(struct device *dev)
7287	{
7288	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7289
7290	kfree(wq_dev);
7291	}
7292
7293	/**
7294	* workqueue_sysfs_register - make a workqueue visible in sysfs
7295	* @wq: the workqueue to register
7296	*
7297	* Expose @wq in sysfs under /sys/bus/workqueue/devices.
7298	* alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7299	* which is the preferred method.
7300	*
7301	* Workqueue user should use this function directly iff it wants to apply
7302	* workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7303	* apply_workqueue_attrs() may race against userland updating the
7304	* attributes.
7305	*
7306	* Return: 0 on success, -errno on failure.
7307	*/
7308	int workqueue_sysfs_register(struct workqueue_struct *wq)
7309	{
7310	struct wq_device *wq_dev;
7311	int ret;
7312
7313	/*
7314	* Adjusting max_active breaks ordering guarantee. Disallow exposing
7315	* ordered workqueues.
7316	*/
7317	if (WARN_ON(wq->flags & __WQ_ORDERED))
7318	return -EINVAL;
7319
7320	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7321	if (!wq_dev)
7322	return -ENOMEM;
7323
7324	wq_dev->wq = wq;
7325	wq_dev->dev.bus = &wq_subsys;
7326	wq_dev->dev.release = wq_device_release;
7327	dev_set_name(&wq_dev->dev, "%s", wq->name);
7328
7329	/*
7330	* unbound_attrs are created separately. Suppress uevent until
7331	* everything is ready.
7332	*/
7333	dev_set_uevent_suppress(&wq_dev->dev, true);
7334
7335	ret = device_register(&wq_dev->dev);
7336	if (ret) {
7337	put_device(&wq_dev->dev);
7338	wq->wq_dev = NULL;
7339	return ret;
7340	}
7341
7342	if (wq->flags & WQ_UNBOUND) {
7343	struct device_attribute *attr;
7344
7345	for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7346	ret = device_create_file(&wq_dev->dev, attr);
7347	if (ret) {
7348	device_unregister(&wq_dev->dev);
7349	wq->wq_dev = NULL;
7350	return ret;
7351	}
7352	}
7353	}
7354
7355	dev_set_uevent_suppress(&wq_dev->dev, false);
7356	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7357	return 0;
7358	}
7359
7360	/**
7361	* workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7362	* @wq: the workqueue to unregister
7363	*
7364	* If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7365	*/
7366	static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7367	{
7368	struct wq_device *wq_dev = wq->wq_dev;
7369
7370	if (!wq->wq_dev)
7371	return;
7372
7373	wq->wq_dev = NULL;
7374	device_unregister(&wq_dev->dev);
7375	}
7376	#else /* CONFIG_SYSFS */
7377	static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7378	#endif /* CONFIG_SYSFS */
7379
7380	/*
7381	* Workqueue watchdog.
7382	*
7383	* Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7384	* flush dependency, a concurrency managed work item which stays RUNNING
7385	* indefinitely. Workqueue stalls can be very difficult to debug as the
7386	* usual warning mechanisms don't trigger and internal workqueue state is
7387	* largely opaque.
7388	*
7389	* Workqueue watchdog monitors all worker pools periodically and dumps
7390	* state if some pools failed to make forward progress for a while where
7391	* forward progress is defined as the first item on ->worklist changing.
7392	*
7393	* This mechanism is controlled through the kernel parameter
7394	* "workqueue.watchdog_thresh" which can be updated at runtime through the
7395	* corresponding sysfs parameter file.
7396	*/
7397	#ifdef CONFIG_WQ_WATCHDOG
7398
7399	static unsigned long wq_watchdog_thresh = 30;
7400	static struct timer_list wq_watchdog_timer;
7401
7402	static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7403	static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7404
7405	/*
7406	* Show workers that might prevent the processing of pending work items.
7407	* The only candidates are CPU-bound workers in the running state.
7408	* Pending work items should be handled by another idle worker
7409	* in all other situations.
7410	*/
7411	static void show_cpu_pool_hog(struct worker_pool *pool)
7412	{
7413	struct worker *worker;
7414	unsigned long irq_flags;
7415	int bkt;
7416
7417	raw_spin_lock_irqsave(&pool->lock, irq_flags);
7418
7419	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
7420	if (task_is_running(worker->task)) {
7421	/*
7422	* Defer printing to avoid deadlocks in console
7423	* drivers that queue work while holding locks
7424	* also taken in their write paths.
7425	*/
7426	printk_deferred_enter();
7427
7428	pr_info("pool %d:\n", pool->id);
7429	sched_show_task(worker->task);
7430
7431	printk_deferred_exit();
7432	}
7433	}
7434
7435	raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
7436	}
7437
7438	static void show_cpu_pools_hogs(void)
7439	{
7440	struct worker_pool *pool;
7441	int pi;
7442
7443	pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
7444
7445	rcu_read_lock();
7446
7447	for_each_pool(pool, pi) {
7448	if (pool->cpu_stall)
7449	show_cpu_pool_hog(pool);
7450
7451	}
7452
7453	rcu_read_unlock();
7454	}
7455
7456	static void wq_watchdog_reset_touched(void)
7457	{
7458	int cpu;
7459
7460	wq_watchdog_touched = jiffies;
7461	for_each_possible_cpu(cpu)
7462	per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7463	}
7464
7465	static void wq_watchdog_timer_fn(struct timer_list *unused)
7466	{
7467	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7468	bool lockup_detected = false;
7469	bool cpu_pool_stall = false;
7470	unsigned long now = jiffies;
7471	struct worker_pool *pool;
7472	int pi;
7473
7474	if (!thresh)
7475	return;
7476
7477	rcu_read_lock();
7478
7479	for_each_pool(pool, pi) {
7480	unsigned long pool_ts, touched, ts;
7481
7482	pool->cpu_stall = false;
7483	if (list_empty(&pool->worklist))
7484	continue;
7485
7486	/*
7487	* If a virtual machine is stopped by the host it can look to
7488	* the watchdog like a stall.
7489	*/
7490	kvm_check_and_clear_guest_paused();
7491
7492	/* get the latest of pool and touched timestamps */
7493	if (pool->cpu >= 0)
7494	touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7495	else
7496	touched = READ_ONCE(wq_watchdog_touched);
7497	pool_ts = READ_ONCE(pool->watchdog_ts);
7498
7499	if (time_after(pool_ts, touched))
7500	ts = pool_ts;
7501	else
7502	ts = touched;
7503
7504	/* did we stall? */
7505	if (time_after(now, ts + thresh)) {
7506	lockup_detected = true;
7507	if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) {
7508	pool->cpu_stall = true;
7509	cpu_pool_stall = true;
7510	}
7511	pr_emerg("BUG: workqueue lockup - pool");
7512	pr_cont_pool_info(pool);
7513	pr_cont(" stuck for %us!\n",
7514	jiffies_to_msecs(now - pool_ts) / 1000);
7515	}
7516
7517
7518	}
7519
7520	rcu_read_unlock();
7521
7522	if (lockup_detected)
7523	show_all_workqueues();
7524
7525	if (cpu_pool_stall)
7526	show_cpu_pools_hogs();
7527
7528	wq_watchdog_reset_touched();
7529	mod_timer(&wq_watchdog_timer, jiffies + thresh);
7530	}
7531
7532	notrace void wq_watchdog_touch(int cpu)
7533	{
7534	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7535	unsigned long touch_ts = READ_ONCE(wq_watchdog_touched);
7536	unsigned long now = jiffies;
7537
7538	if (cpu >= 0)
7539	per_cpu(wq_watchdog_touched_cpu, cpu) = now;
7540	else
7541	WARN_ONCE(1, "%s should be called with valid CPU", __func__);
7542
7543	/* Don't unnecessarily store to global cacheline */
7544	if (time_after(now, touch_ts + thresh / 4))
7545	WRITE_ONCE(wq_watchdog_touched, jiffies);
7546	}
7547
7548	static void wq_watchdog_set_thresh(unsigned long thresh)
7549	{
7550	wq_watchdog_thresh = 0;
7551	del_timer_sync(&wq_watchdog_timer);
7552
7553	if (thresh) {
7554	wq_watchdog_thresh = thresh;
7555	wq_watchdog_reset_touched();
7556	mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7557	}
7558	}
7559
7560	static int wq_watchdog_param_set_thresh(const char *val,
7561	const struct kernel_param *kp)
7562	{
7563	unsigned long thresh;
7564	int ret;
7565
7566	ret = kstrtoul(val, 0, &thresh);
7567	if (ret)
7568	return ret;
7569
7570	if (system_wq)
7571	wq_watchdog_set_thresh(thresh);
7572	else
7573	wq_watchdog_thresh = thresh;
7574
7575	return 0;
7576	}
7577
7578	static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7579	.set = wq_watchdog_param_set_thresh,
7580	.get = param_get_ulong,
7581	};
7582
7583	module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7584	0644);
7585
7586	static void wq_watchdog_init(void)
7587	{
7588	timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7589	wq_watchdog_set_thresh(wq_watchdog_thresh);
7590	}
7591
7592	#else /* CONFIG_WQ_WATCHDOG */
7593
7594	static inline void wq_watchdog_init(void) { }
7595
7596	#endif /* CONFIG_WQ_WATCHDOG */
7597
7598	static void bh_pool_kick_normal(struct irq_work *irq_work)
7599	{
7600	raise_softirq_irqoff(TASKLET_SOFTIRQ);
7601	}
7602
7603	static void bh_pool_kick_highpri(struct irq_work *irq_work)
7604	{
7605	raise_softirq_irqoff(HI_SOFTIRQ);
7606	}
7607
7608	static void __init restrict_unbound_cpumask(const char name, const struct cpumask mask)
7609	{
7610	if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
7611	pr_warn("workqueue: Restricting unbound_cpumask (%pb) with %s (%pb) leaves no CPU, ignoring\n",
7612	cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7613	return;
7614	}
7615
7616	cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
7617	}
7618
7619	static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice)
7620	{
7621	BUG_ON(init_worker_pool(pool));
7622	pool->cpu = cpu;
7623	cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7624	cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
7625	pool->attrs->nice = nice;
7626	pool->attrs->affn_strict = true;
7627	pool->node = cpu_to_node(cpu);
7628
7629	/* alloc pool ID */
7630	mutex_lock(&wq_pool_mutex);
7631	BUG_ON(worker_pool_assign_id(pool));
7632	mutex_unlock(&wq_pool_mutex);
7633	}
7634
7635	/**
7636	* workqueue_init_early - early init for workqueue subsystem
7637	*
7638	* This is the first step of three-staged workqueue subsystem initialization and
7639	* invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7640	* up. It sets up all the data structures and system workqueues and allows early
7641	* boot code to create workqueues and queue/cancel work items. Actual work item
7642	* execution starts only after kthreads can be created and scheduled right
7643	* before early initcalls.
7644	*/
7645	void __init workqueue_init_early(void)
7646	{
7647	struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7648	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7649	void (irq_work_fns[2])(struct irq_work ) = { bh_pool_kick_normal,
7650	bh_pool_kick_highpri };
7651	int i, cpu;
7652
7653	BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7654
7655	BUG_ON(!alloc_cpumask_var(&wq_online_cpumask, GFP_KERNEL));
7656	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7657	BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
7658	BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
7659
7660	cpumask_copy(wq_online_cpumask, cpu_online_mask);
7661	cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
7662	restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
7663	restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
7664	if (!cpumask_empty(&wq_cmdline_cpumask))
7665	restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
7666
7667	cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
7668
7669	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7670
7671	unbound_wq_update_pwq_attrs_buf = alloc_workqueue_attrs();
7672	BUG_ON(!unbound_wq_update_pwq_attrs_buf);
7673
7674	/*
7675	* If nohz_full is enabled, set power efficient workqueue as unbound.
7676	* This allows workqueue items to be moved to HK CPUs.
7677	*/
7678	if (housekeeping_enabled(HK_TYPE_TICK))
7679	wq_power_efficient = true;
7680
7681	/* initialize WQ_AFFN_SYSTEM pods */
7682	pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7683	pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7684	pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7685	BUG_ON(!pt->pod_cpus \|\| !pt->pod_node \|\| !pt->cpu_pod);
7686
7687	BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7688
7689	pt->nr_pods = 1;
7690	cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
7691	pt->pod_node[0] = NUMA_NO_NODE;
7692	pt->cpu_pod[0] = 0;
7693
7694	/* initialize BH and CPU pools */
7695	for_each_possible_cpu(cpu) {
7696	struct worker_pool *pool;
7697
7698	i = 0;
7699	for_each_bh_worker_pool(pool, cpu) {
7700	init_cpu_worker_pool(pool, cpu, std_nice[i]);
7701	pool->flags \|= POOL_BH;
7702	init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]);
7703	i++;
7704	}
7705
7706	i = 0;
7707	for_each_cpu_worker_pool(pool, cpu)
7708	init_cpu_worker_pool(pool, cpu, std_nice[i++]);
7709	}
7710
7711	/* create default unbound and ordered wq attrs */
7712	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7713	struct workqueue_attrs *attrs;
7714
7715	BUG_ON(!(attrs = alloc_workqueue_attrs()));
7716	attrs->nice = std_nice[i];
7717	unbound_std_wq_attrs[i] = attrs;
7718
7719	/*
7720	* An ordered wq should have only one pwq as ordering is
7721	* guaranteed by max_active which is enforced by pwqs.
7722	*/
7723	BUG_ON(!(attrs = alloc_workqueue_attrs()));
7724	attrs->nice = std_nice[i];
7725	attrs->ordered = true;
7726	ordered_wq_attrs[i] = attrs;
7727	}
7728
7729	system_wq = alloc_workqueue("events", 0, 0);
7730	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7731	system_long_wq = alloc_workqueue("events_long", 0, 0);
7732	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7733	WQ_MAX_ACTIVE);
7734	system_freezable_wq = alloc_workqueue("events_freezable",
7735	WQ_FREEZABLE, 0);
7736	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7737	WQ_POWER_EFFICIENT, 0);
7738	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
7739	WQ_FREEZABLE \| WQ_POWER_EFFICIENT,
7740	0);
7741	system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0);
7742	system_bh_highpri_wq = alloc_workqueue("events_bh_highpri",
7743	WQ_BH \| WQ_HIGHPRI, 0);
7744	BUG_ON(!system_wq \|\| !system_highpri_wq \|\| !system_long_wq \|\|
7745	!system_unbound_wq \|\| !system_freezable_wq \|\|
7746	!system_power_efficient_wq \|\|
7747	!system_freezable_power_efficient_wq \|\|
7748	!system_bh_wq \|\| !system_bh_highpri_wq);
7749	}
7750
7751	static void __init wq_cpu_intensive_thresh_init(void)
7752	{
7753	unsigned long thresh;
7754	unsigned long bogo;
7755
7756	pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
7757	BUG_ON(IS_ERR(pwq_release_worker));
7758
7759	/* if the user set it to a specific value, keep it */
7760	if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7761	return;
7762
7763	/*
7764	* The default of 10ms is derived from the fact that most modern (as of
7765	* 2023) processors can do a lot in 10ms and that it's just below what
7766	* most consider human-perceivable. However, the kernel also runs on a
7767	* lot slower CPUs including microcontrollers where the threshold is way
7768	* too low.
7769	*
7770	* Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7771	* This is by no means accurate but it doesn't have to be. The mechanism
7772	* is still useful even when the threshold is fully scaled up. Also, as
7773	* the reports would usually be applicable to everyone, some machines
7774	* operating on longer thresholds won't significantly diminish their
7775	* usefulness.
7776	*/
7777	thresh = 10 * USEC_PER_MSEC;
7778
7779	/* see init/calibrate.c for lpj -> BogoMIPS calculation */
7780	bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7781	if (bogo < 4000)
7782	thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7783
7784	pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7785	loops_per_jiffy, bogo, thresh);
7786
7787	wq_cpu_intensive_thresh_us = thresh;
7788	}
7789
7790	/**
7791	* workqueue_init - bring workqueue subsystem fully online
7792	*
7793	* This is the second step of three-staged workqueue subsystem initialization
7794	* and invoked as soon as kthreads can be created and scheduled. Workqueues have
7795	* been created and work items queued on them, but there are no kworkers
7796	* executing the work items yet. Populate the worker pools with the initial
7797	* workers and enable future kworker creations.
7798	*/
7799	void __init workqueue_init(void)
7800	{
7801	struct workqueue_struct *wq;
7802	struct worker_pool *pool;
7803	int cpu, bkt;
7804
7805	wq_cpu_intensive_thresh_init();
7806
7807	mutex_lock(&wq_pool_mutex);
7808
7809	/*
7810	* Per-cpu pools created earlier could be missing node hint. Fix them
7811	* up. Also, create a rescuer for workqueues that requested it.
7812	*/
7813	for_each_possible_cpu(cpu) {
7814	for_each_bh_worker_pool(pool, cpu)
7815	pool->node = cpu_to_node(cpu);
7816	for_each_cpu_worker_pool(pool, cpu)
7817	pool->node = cpu_to_node(cpu);
7818	}
7819
7820	list_for_each_entry(wq, &workqueues, list) {
7821	WARN(init_rescuer(wq),
7822	"workqueue: failed to create early rescuer for %s",
7823	wq->name);
7824	}
7825
7826	mutex_unlock(&wq_pool_mutex);
7827
7828	/*
7829	* Create the initial workers. A BH pool has one pseudo worker that
7830	* represents the shared BH execution context and thus doesn't get
7831	* affected by hotplug events. Create the BH pseudo workers for all
7832	* possible CPUs here.
7833	*/
7834	for_each_possible_cpu(cpu)
7835	for_each_bh_worker_pool(pool, cpu)
7836	BUG_ON(!create_worker(pool));
7837
7838	for_each_online_cpu(cpu) {
7839	for_each_cpu_worker_pool(pool, cpu) {
7840	pool->flags &= ~POOL_DISASSOCIATED;
7841	BUG_ON(!create_worker(pool));
7842	}
7843	}
7844
7845	hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7846	BUG_ON(!create_worker(pool));
7847
7848	wq_online = true;
7849	wq_watchdog_init();
7850	}
7851
7852	/*
7853	* Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7854	* @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7855	* and consecutive pod ID. The rest of @pt is initialized accordingly.
7856	*/
7857	static void __init init_pod_type(struct wq_pod_type *pt,
7858	bool (*cpus_share_pod)(int, int))
7859	{
7860	int cur, pre, cpu, pod;
7861
7862	pt->nr_pods = 0;
7863
7864	/* init @pt->cpu_pod[] according to @cpus_share_pod() */
7865	pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7866	BUG_ON(!pt->cpu_pod);
7867
7868	for_each_possible_cpu(cur) {
7869	for_each_possible_cpu(pre) {
7870	if (pre >= cur) {
7871	pt->cpu_pod[cur] = pt->nr_pods++;
7872	break;
7873	}
7874	if (cpus_share_pod(cur, pre)) {
7875	pt->cpu_pod[cur] = pt->cpu_pod[pre];
7876	break;
7877	}
7878	}
7879	}
7880
7881	/* init the rest to match @pt->cpu_pod[] */
7882	pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7883	pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7884	BUG_ON(!pt->pod_cpus \|\| !pt->pod_node);
7885
7886	for (pod = 0; pod < pt->nr_pods; pod++)
7887	BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7888
7889	for_each_possible_cpu(cpu) {
7890	cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
7891	pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7892	}
7893	}
7894
7895	static bool __init cpus_dont_share(int cpu0, int cpu1)
7896	{
7897	return false;
7898	}
7899
7900	static bool __init cpus_share_smt(int cpu0, int cpu1)
7901	{
7902	#ifdef CONFIG_SCHED_SMT
7903	return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
7904	#else
7905	return false;
7906	#endif
7907	}
7908
7909	static bool __init cpus_share_numa(int cpu0, int cpu1)
7910	{
7911	return cpu_to_node(cpu0) == cpu_to_node(cpu1);
7912	}
7913
7914	/**
7915	* workqueue_init_topology - initialize CPU pods for unbound workqueues
7916	*
7917	* This is the third step of three-staged workqueue subsystem initialization and
7918	* invoked after SMP and topology information are fully initialized. It
7919	* initializes the unbound CPU pods accordingly.
7920	*/
7921	void __init workqueue_init_topology(void)
7922	{
7923	struct workqueue_struct *wq;
7924	int cpu;
7925
7926	init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
7927	init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
7928	init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
7929	init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
7930
7931	wq_topo_initialized = true;
7932
7933	mutex_lock(&wq_pool_mutex);
7934
7935	/*
7936	* Workqueues allocated earlier would have all CPUs sharing the default
7937	* worker pool. Explicitly call unbound_wq_update_pwq() on all workqueue
7938	* and CPU combinations to apply per-pod sharing.
7939	*/
7940	list_for_each_entry(wq, &workqueues, list) {
7941	for_each_online_cpu(cpu)
7942	unbound_wq_update_pwq(wq, cpu);
7943	if (wq->flags & WQ_UNBOUND) {
7944	mutex_lock(&wq->mutex);
7945	wq_update_node_max_active(wq, -1);
7946	mutex_unlock(&wq->mutex);
7947	}
7948	}
7949
7950	mutex_unlock(&wq_pool_mutex);
7951	}
7952
7953	void __warn_flushing_systemwide_wq(void)
7954	{
7955	pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
7956	dump_stack();
7957	}
7958	EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
7959
7960	static int __init workqueue_unbound_cpus_setup(char *str)
7961	{
7962	if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
7963	cpumask_clear(&wq_cmdline_cpumask);
7964	pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
7965	}
7966
7967	return 1;
7968	}
7969	__setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);