xref: /linux-6.15/kernel/workqueue.c (revision 8639eceb)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * kernel/workqueue.c - generic async execution with shared worker pool
4  *
5  * Copyright (C) 2002		Ingo Molnar
6  *
7  *   Derived from the taskqueue/keventd code by:
8  *     David Woodhouse <[email protected]>
9  *     Andrew Morton
10  *     Kai Petzke <[email protected]>
11  *     Theodore Ts'o <[email protected]>
12  *
13  * Made to use alloc_percpu by Christoph Lameter.
14  *
15  * Copyright (C) 2010		SUSE Linux Products GmbH
16  * Copyright (C) 2010		Tejun Heo <[email protected]>
17  *
18  * This is the generic async execution mechanism.  Work items as are
19  * executed in process context.  The worker pool is shared and
20  * automatically managed.  There are two worker pools for each CPU (one for
21  * normal work items and the other for high priority ones) and some extra
22  * pools for workqueues which are not bound to any specific CPU - the
23  * number of these backing pools is dynamic.
24  *
25  * Please read Documentation/core-api/workqueue.rst for details.
26  */
27 
28 #include <linux/export.h>
29 #include <linux/kernel.h>
30 #include <linux/sched.h>
31 #include <linux/init.h>
32 #include <linux/signal.h>
33 #include <linux/completion.h>
34 #include <linux/workqueue.h>
35 #include <linux/slab.h>
36 #include <linux/cpu.h>
37 #include <linux/notifier.h>
38 #include <linux/kthread.h>
39 #include <linux/hardirq.h>
40 #include <linux/mempolicy.h>
41 #include <linux/freezer.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 #include <linux/sched/isolation.h>
52 #include <linux/sched/debug.h>
53 #include <linux/nmi.h>
54 #include <linux/kvm_para.h>
55 #include <linux/delay.h>
56 
57 #include "workqueue_internal.h"
58 
59 enum {
60 	/*
61 	 * worker_pool flags
62 	 *
63 	 * A bound pool is either associated or disassociated with its CPU.
64 	 * While associated (!DISASSOCIATED), all workers are bound to the
65 	 * CPU and none has %WORKER_UNBOUND set and concurrency management
66 	 * is in effect.
67 	 *
68 	 * While DISASSOCIATED, the cpu may be offline and all workers have
69 	 * %WORKER_UNBOUND set and concurrency management disabled, and may
70 	 * be executing on any CPU.  The pool behaves as an unbound one.
71 	 *
72 	 * Note that DISASSOCIATED should be flipped only while holding
73 	 * wq_pool_attach_mutex to avoid changing binding state while
74 	 * worker_attach_to_pool() is in progress.
75 	 */
76 	POOL_MANAGER_ACTIVE	= 1 << 0,	/* being managed */
77 	POOL_DISASSOCIATED	= 1 << 2,	/* cpu can't serve workers */
78 
79 	/* worker flags */
80 	WORKER_DIE		= 1 << 1,	/* die die die */
81 	WORKER_IDLE		= 1 << 2,	/* is idle */
82 	WORKER_PREP		= 1 << 3,	/* preparing to run works */
83 	WORKER_CPU_INTENSIVE	= 1 << 6,	/* cpu intensive */
84 	WORKER_UNBOUND		= 1 << 7,	/* worker is unbound */
85 	WORKER_REBOUND		= 1 << 8,	/* worker was rebound */
86 
87 	WORKER_NOT_RUNNING	= WORKER_PREP | WORKER_CPU_INTENSIVE |
88 				  WORKER_UNBOUND | WORKER_REBOUND,
89 
90 	NR_STD_WORKER_POOLS	= 2,		/* # standard pools per cpu */
91 
92 	UNBOUND_POOL_HASH_ORDER	= 6,		/* hashed by pool->attrs */
93 	BUSY_WORKER_HASH_ORDER	= 6,		/* 64 pointers */
94 
95 	MAX_IDLE_WORKERS_RATIO	= 4,		/* 1/4 of busy can be idle */
96 	IDLE_WORKER_TIMEOUT	= 300 * HZ,	/* keep idle ones for 5 mins */
97 
98 	MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
99 						/* call for help after 10ms
100 						   (min two ticks) */
101 	MAYDAY_INTERVAL		= HZ / 10,	/* and then every 100ms */
102 	CREATE_COOLDOWN		= HZ,		/* time to breath after fail */
103 
104 	/*
105 	 * Rescue workers are used only on emergencies and shared by
106 	 * all cpus.  Give MIN_NICE.
107 	 */
108 	RESCUER_NICE_LEVEL	= MIN_NICE,
109 	HIGHPRI_NICE_LEVEL	= MIN_NICE,
110 
111 	WQ_NAME_LEN		= 24,
112 };
113 
114 /*
115  * Structure fields follow one of the following exclusion rules.
116  *
117  * I: Modifiable by initialization/destruction paths and read-only for
118  *    everyone else.
119  *
120  * P: Preemption protected.  Disabling preemption is enough and should
121  *    only be modified and accessed from the local cpu.
122  *
123  * L: pool->lock protected.  Access with pool->lock held.
124  *
125  * K: Only modified by worker while holding pool->lock. Can be safely read by
126  *    self, while holding pool->lock or from IRQ context if %current is the
127  *    kworker.
128  *
129  * S: Only modified by worker self.
130  *
131  * A: wq_pool_attach_mutex protected.
132  *
133  * PL: wq_pool_mutex protected.
134  *
135  * PR: wq_pool_mutex protected for writes.  RCU protected for reads.
136  *
137  * PW: wq_pool_mutex and wq->mutex protected for writes.  Either for reads.
138  *
139  * PWR: wq_pool_mutex and wq->mutex protected for writes.  Either or
140  *      RCU for reads.
141  *
142  * WQ: wq->mutex protected.
143  *
144  * WR: wq->mutex protected for writes.  RCU protected for reads.
145  *
146  * MD: wq_mayday_lock protected.
147  *
148  * WD: Used internally by the watchdog.
149  */
150 
151 /* struct worker is defined in workqueue_internal.h */
152 
153 struct worker_pool {
154 	raw_spinlock_t		lock;		/* the pool lock */
155 	int			cpu;		/* I: the associated cpu */
156 	int			node;		/* I: the associated node ID */
157 	int			id;		/* I: pool ID */
158 	unsigned int		flags;		/* L: flags */
159 
160 	unsigned long		watchdog_ts;	/* L: watchdog timestamp */
161 	bool			cpu_stall;	/* WD: stalled cpu bound pool */
162 
163 	/*
164 	 * The counter is incremented in a process context on the associated CPU
165 	 * w/ preemption disabled, and decremented or reset in the same context
166 	 * but w/ pool->lock held. The readers grab pool->lock and are
167 	 * guaranteed to see if the counter reached zero.
168 	 */
169 	int			nr_running;
170 
171 	struct list_head	worklist;	/* L: list of pending works */
172 
173 	int			nr_workers;	/* L: total number of workers */
174 	int			nr_idle;	/* L: currently idle workers */
175 
176 	struct list_head	idle_list;	/* L: list of idle workers */
177 	struct timer_list	idle_timer;	/* L: worker idle timeout */
178 	struct work_struct      idle_cull_work; /* L: worker idle cleanup */
179 
180 	struct timer_list	mayday_timer;	  /* L: SOS timer for workers */
181 
182 	/* a workers is either on busy_hash or idle_list, or the manager */
183 	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
184 						/* L: hash of busy workers */
185 
186 	struct worker		*manager;	/* L: purely informational */
187 	struct list_head	workers;	/* A: attached workers */
188 	struct list_head        dying_workers;  /* A: workers about to die */
189 	struct completion	*detach_completion; /* all workers detached */
190 
191 	struct ida		worker_ida;	/* worker IDs for task name */
192 
193 	struct workqueue_attrs	*attrs;		/* I: worker attributes */
194 	struct hlist_node	hash_node;	/* PL: unbound_pool_hash node */
195 	int			refcnt;		/* PL: refcnt for unbound pools */
196 
197 	/*
198 	 * Destruction of pool is RCU protected to allow dereferences
199 	 * from get_work_pool().
200 	 */
201 	struct rcu_head		rcu;
202 };
203 
204 /*
205  * Per-pool_workqueue statistics. These can be monitored using
206  * tools/workqueue/wq_monitor.py.
207  */
208 enum pool_workqueue_stats {
209 	PWQ_STAT_STARTED,	/* work items started execution */
210 	PWQ_STAT_COMPLETED,	/* work items completed execution */
211 	PWQ_STAT_CPU_TIME,	/* total CPU time consumed */
212 	PWQ_STAT_CPU_INTENSIVE,	/* wq_cpu_intensive_thresh_us violations */
213 	PWQ_STAT_CM_WAKEUP,	/* concurrency-management worker wakeups */
214 	PWQ_STAT_REPATRIATED,	/* unbound workers brought back into scope */
215 	PWQ_STAT_MAYDAY,	/* maydays to rescuer */
216 	PWQ_STAT_RESCUED,	/* linked work items executed by rescuer */
217 
218 	PWQ_NR_STATS,
219 };
220 
221 /*
222  * The per-pool workqueue.  While queued, the lower WORK_STRUCT_FLAG_BITS
223  * of work_struct->data are used for flags and the remaining high bits
224  * point to the pwq; thus, pwqs need to be aligned at two's power of the
225  * number of flag bits.
226  */
227 struct pool_workqueue {
228 	struct worker_pool	*pool;		/* I: the associated pool */
229 	struct workqueue_struct *wq;		/* I: the owning workqueue */
230 	int			work_color;	/* L: current color */
231 	int			flush_color;	/* L: flushing color */
232 	int			refcnt;		/* L: reference count */
233 	int			nr_in_flight[WORK_NR_COLORS];
234 						/* L: nr of in_flight works */
235 
236 	/*
237 	 * nr_active management and WORK_STRUCT_INACTIVE:
238 	 *
239 	 * When pwq->nr_active >= max_active, new work item is queued to
240 	 * pwq->inactive_works instead of pool->worklist and marked with
241 	 * WORK_STRUCT_INACTIVE.
242 	 *
243 	 * All work items marked with WORK_STRUCT_INACTIVE do not participate
244 	 * in pwq->nr_active and all work items in pwq->inactive_works are
245 	 * marked with WORK_STRUCT_INACTIVE.  But not all WORK_STRUCT_INACTIVE
246 	 * work items are in pwq->inactive_works.  Some of them are ready to
247 	 * run in pool->worklist or worker->scheduled.  Those work itmes are
248 	 * only struct wq_barrier which is used for flush_work() and should
249 	 * not participate in pwq->nr_active.  For non-barrier work item, it
250 	 * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
251 	 */
252 	int			nr_active;	/* L: nr of active works */
253 	int			max_active;	/* L: max active works */
254 	struct list_head	inactive_works;	/* L: inactive works */
255 	struct list_head	pwqs_node;	/* WR: node on wq->pwqs */
256 	struct list_head	mayday_node;	/* MD: node on wq->maydays */
257 
258 	u64			stats[PWQ_NR_STATS];
259 
260 	/*
261 	 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
262 	 * and pwq_release_workfn() for details. pool_workqueue itself is also
263 	 * RCU protected so that the first pwq can be determined without
264 	 * grabbing wq->mutex.
265 	 */
266 	struct kthread_work	release_work;
267 	struct rcu_head		rcu;
268 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
269 
270 /*
271  * Structure used to wait for workqueue flush.
272  */
273 struct wq_flusher {
274 	struct list_head	list;		/* WQ: list of flushers */
275 	int			flush_color;	/* WQ: flush color waiting for */
276 	struct completion	done;		/* flush completion */
277 };
278 
279 struct wq_device;
280 
281 /*
282  * The externally visible workqueue.  It relays the issued work items to
283  * the appropriate worker_pool through its pool_workqueues.
284  */
285 struct workqueue_struct {
286 	struct list_head	pwqs;		/* WR: all pwqs of this wq */
287 	struct list_head	list;		/* PR: list of all workqueues */
288 
289 	struct mutex		mutex;		/* protects this wq */
290 	int			work_color;	/* WQ: current work color */
291 	int			flush_color;	/* WQ: current flush color */
292 	atomic_t		nr_pwqs_to_flush; /* flush in progress */
293 	struct wq_flusher	*first_flusher;	/* WQ: first flusher */
294 	struct list_head	flusher_queue;	/* WQ: flush waiters */
295 	struct list_head	flusher_overflow; /* WQ: flush overflow list */
296 
297 	struct list_head	maydays;	/* MD: pwqs requesting rescue */
298 	struct worker		*rescuer;	/* MD: rescue worker */
299 
300 	int			nr_drainers;	/* WQ: drain in progress */
301 	int			saved_max_active; /* WQ: saved pwq max_active */
302 
303 	struct workqueue_attrs	*unbound_attrs;	/* PW: only for unbound wqs */
304 	struct pool_workqueue	*dfl_pwq;	/* PW: only for unbound wqs */
305 
306 #ifdef CONFIG_SYSFS
307 	struct wq_device	*wq_dev;	/* I: for sysfs interface */
308 #endif
309 #ifdef CONFIG_LOCKDEP
310 	char			*lock_name;
311 	struct lock_class_key	key;
312 	struct lockdep_map	lockdep_map;
313 #endif
314 	char			name[WQ_NAME_LEN]; /* I: workqueue name */
315 
316 	/*
317 	 * Destruction of workqueue_struct is RCU protected to allow walking
318 	 * the workqueues list without grabbing wq_pool_mutex.
319 	 * This is used to dump all workqueues from sysrq.
320 	 */
321 	struct rcu_head		rcu;
322 
323 	/* hot fields used during command issue, aligned to cacheline */
324 	unsigned int		flags ____cacheline_aligned; /* WQ: WQ_* flags */
325 	struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
326 };
327 
328 static struct kmem_cache *pwq_cache;
329 
330 /*
331  * Each pod type describes how CPUs should be grouped for unbound workqueues.
332  * See the comment above workqueue_attrs->affn_scope.
333  */
334 struct wq_pod_type {
335 	int			nr_pods;	/* number of pods */
336 	cpumask_var_t		*pod_cpus;	/* pod -> cpus */
337 	int			*pod_node;	/* pod -> node */
338 	int			*cpu_pod;	/* cpu -> pod */
339 };
340 
341 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
342 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_DFL;
343 
344 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
345 	[WQ_AFFN_CPU]			= "cpu",
346 	[WQ_AFFN_SMT]			= "smt",
347 	[WQ_AFFN_CACHE]			= "cache",
348 	[WQ_AFFN_NUMA]			= "numa",
349 	[WQ_AFFN_SYSTEM]		= "system",
350 };
351 
352 /*
353  * Per-cpu work items which run for longer than the following threshold are
354  * automatically considered CPU intensive and excluded from concurrency
355  * management to prevent them from noticeably delaying other per-cpu work items.
356  * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
357  * The actual value is initialized in wq_cpu_intensive_thresh_init().
358  */
359 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
360 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
361 
362 /* see the comment above the definition of WQ_POWER_EFFICIENT */
363 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
364 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
365 
366 static bool wq_online;			/* can kworkers be created yet? */
367 
368 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
369 static struct workqueue_attrs *wq_update_pod_attrs_buf;
370 
371 static DEFINE_MUTEX(wq_pool_mutex);	/* protects pools and workqueues list */
372 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
373 static DEFINE_RAW_SPINLOCK(wq_mayday_lock);	/* protects wq->maydays list */
374 /* wait for manager to go away */
375 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
376 
377 static LIST_HEAD(workqueues);		/* PR: list of all workqueues */
378 static bool workqueue_freezing;		/* PL: have wqs started freezing? */
379 
380 /* PL&A: allowable cpus for unbound wqs and work items */
381 static cpumask_var_t wq_unbound_cpumask;
382 
383 /* for further constrain wq_unbound_cpumask by cmdline parameter*/
384 static struct cpumask wq_cmdline_cpumask __initdata;
385 
386 /* CPU where unbound work was last round robin scheduled from this CPU */
387 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
388 
389 /*
390  * Local execution of unbound work items is no longer guaranteed.  The
391  * following always forces round-robin CPU selection on unbound work items
392  * to uncover usages which depend on it.
393  */
394 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
395 static bool wq_debug_force_rr_cpu = true;
396 #else
397 static bool wq_debug_force_rr_cpu = false;
398 #endif
399 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
400 
401 /* the per-cpu worker pools */
402 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
403 
404 static DEFINE_IDR(worker_pool_idr);	/* PR: idr of all pools */
405 
406 /* PL: hash of all unbound pools keyed by pool->attrs */
407 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
408 
409 /* I: attributes used when instantiating standard unbound pools on demand */
410 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
411 
412 /* I: attributes used when instantiating ordered pools on demand */
413 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
414 
415 /*
416  * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
417  * process context while holding a pool lock. Bounce to a dedicated kthread
418  * worker to avoid A-A deadlocks.
419  */
420 static struct kthread_worker *pwq_release_worker;
421 
422 struct workqueue_struct *system_wq __read_mostly;
423 EXPORT_SYMBOL(system_wq);
424 struct workqueue_struct *system_highpri_wq __read_mostly;
425 EXPORT_SYMBOL_GPL(system_highpri_wq);
426 struct workqueue_struct *system_long_wq __read_mostly;
427 EXPORT_SYMBOL_GPL(system_long_wq);
428 struct workqueue_struct *system_unbound_wq __read_mostly;
429 EXPORT_SYMBOL_GPL(system_unbound_wq);
430 struct workqueue_struct *system_freezable_wq __read_mostly;
431 EXPORT_SYMBOL_GPL(system_freezable_wq);
432 struct workqueue_struct *system_power_efficient_wq __read_mostly;
433 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
434 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
435 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
436 
437 static int worker_thread(void *__worker);
438 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
439 static void show_pwq(struct pool_workqueue *pwq);
440 static void show_one_worker_pool(struct worker_pool *pool);
441 
442 #define CREATE_TRACE_POINTS
443 #include <trace/events/workqueue.h>
444 
445 #define assert_rcu_or_pool_mutex()					\
446 	RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&			\
447 			 !lockdep_is_held(&wq_pool_mutex),		\
448 			 "RCU or wq_pool_mutex should be held")
449 
450 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq)			\
451 	RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&			\
452 			 !lockdep_is_held(&wq->mutex) &&		\
453 			 !lockdep_is_held(&wq_pool_mutex),		\
454 			 "RCU, wq->mutex or wq_pool_mutex should be held")
455 
456 #define for_each_cpu_worker_pool(pool, cpu)				\
457 	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0];		\
458 	     (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
459 	     (pool)++)
460 
461 /**
462  * for_each_pool - iterate through all worker_pools in the system
463  * @pool: iteration cursor
464  * @pi: integer used for iteration
465  *
466  * This must be called either with wq_pool_mutex held or RCU read
467  * locked.  If the pool needs to be used beyond the locking in effect, the
468  * caller is responsible for guaranteeing that the pool stays online.
469  *
470  * The if/else clause exists only for the lockdep assertion and can be
471  * ignored.
472  */
473 #define for_each_pool(pool, pi)						\
474 	idr_for_each_entry(&worker_pool_idr, pool, pi)			\
475 		if (({ assert_rcu_or_pool_mutex(); false; })) { }	\
476 		else
477 
478 /**
479  * for_each_pool_worker - iterate through all workers of a worker_pool
480  * @worker: iteration cursor
481  * @pool: worker_pool to iterate workers of
482  *
483  * This must be called with wq_pool_attach_mutex.
484  *
485  * The if/else clause exists only for the lockdep assertion and can be
486  * ignored.
487  */
488 #define for_each_pool_worker(worker, pool)				\
489 	list_for_each_entry((worker), &(pool)->workers, node)		\
490 		if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
491 		else
492 
493 /**
494  * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
495  * @pwq: iteration cursor
496  * @wq: the target workqueue
497  *
498  * This must be called either with wq->mutex held or RCU read locked.
499  * If the pwq needs to be used beyond the locking in effect, the caller is
500  * responsible for guaranteeing that the pwq stays online.
501  *
502  * The if/else clause exists only for the lockdep assertion and can be
503  * ignored.
504  */
505 #define for_each_pwq(pwq, wq)						\
506 	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node,		\
507 				 lockdep_is_held(&(wq->mutex)))
508 
509 #ifdef CONFIG_DEBUG_OBJECTS_WORK
510 
511 static const struct debug_obj_descr work_debug_descr;
512 
513 static void *work_debug_hint(void *addr)
514 {
515 	return ((struct work_struct *) addr)->func;
516 }
517 
518 static bool work_is_static_object(void *addr)
519 {
520 	struct work_struct *work = addr;
521 
522 	return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
523 }
524 
525 /*
526  * fixup_init is called when:
527  * - an active object is initialized
528  */
529 static bool work_fixup_init(void *addr, enum debug_obj_state state)
530 {
531 	struct work_struct *work = addr;
532 
533 	switch (state) {
534 	case ODEBUG_STATE_ACTIVE:
535 		cancel_work_sync(work);
536 		debug_object_init(work, &work_debug_descr);
537 		return true;
538 	default:
539 		return false;
540 	}
541 }
542 
543 /*
544  * fixup_free is called when:
545  * - an active object is freed
546  */
547 static bool work_fixup_free(void *addr, enum debug_obj_state state)
548 {
549 	struct work_struct *work = addr;
550 
551 	switch (state) {
552 	case ODEBUG_STATE_ACTIVE:
553 		cancel_work_sync(work);
554 		debug_object_free(work, &work_debug_descr);
555 		return true;
556 	default:
557 		return false;
558 	}
559 }
560 
561 static const struct debug_obj_descr work_debug_descr = {
562 	.name		= "work_struct",
563 	.debug_hint	= work_debug_hint,
564 	.is_static_object = work_is_static_object,
565 	.fixup_init	= work_fixup_init,
566 	.fixup_free	= work_fixup_free,
567 };
568 
569 static inline void debug_work_activate(struct work_struct *work)
570 {
571 	debug_object_activate(work, &work_debug_descr);
572 }
573 
574 static inline void debug_work_deactivate(struct work_struct *work)
575 {
576 	debug_object_deactivate(work, &work_debug_descr);
577 }
578 
579 void __init_work(struct work_struct *work, int onstack)
580 {
581 	if (onstack)
582 		debug_object_init_on_stack(work, &work_debug_descr);
583 	else
584 		debug_object_init(work, &work_debug_descr);
585 }
586 EXPORT_SYMBOL_GPL(__init_work);
587 
588 void destroy_work_on_stack(struct work_struct *work)
589 {
590 	debug_object_free(work, &work_debug_descr);
591 }
592 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
593 
594 void destroy_delayed_work_on_stack(struct delayed_work *work)
595 {
596 	destroy_timer_on_stack(&work->timer);
597 	debug_object_free(&work->work, &work_debug_descr);
598 }
599 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
600 
601 #else
602 static inline void debug_work_activate(struct work_struct *work) { }
603 static inline void debug_work_deactivate(struct work_struct *work) { }
604 #endif
605 
606 /**
607  * worker_pool_assign_id - allocate ID and assign it to @pool
608  * @pool: the pool pointer of interest
609  *
610  * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
611  * successfully, -errno on failure.
612  */
613 static int worker_pool_assign_id(struct worker_pool *pool)
614 {
615 	int ret;
616 
617 	lockdep_assert_held(&wq_pool_mutex);
618 
619 	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
620 			GFP_KERNEL);
621 	if (ret >= 0) {
622 		pool->id = ret;
623 		return 0;
624 	}
625 	return ret;
626 }
627 
628 static unsigned int work_color_to_flags(int color)
629 {
630 	return color << WORK_STRUCT_COLOR_SHIFT;
631 }
632 
633 static int get_work_color(unsigned long work_data)
634 {
635 	return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
636 		((1 << WORK_STRUCT_COLOR_BITS) - 1);
637 }
638 
639 static int work_next_color(int color)
640 {
641 	return (color + 1) % WORK_NR_COLORS;
642 }
643 
644 /*
645  * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
646  * contain the pointer to the queued pwq.  Once execution starts, the flag
647  * is cleared and the high bits contain OFFQ flags and pool ID.
648  *
649  * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
650  * and clear_work_data() can be used to set the pwq, pool or clear
651  * work->data.  These functions should only be called while the work is
652  * owned - ie. while the PENDING bit is set.
653  *
654  * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
655  * corresponding to a work.  Pool is available once the work has been
656  * queued anywhere after initialization until it is sync canceled.  pwq is
657  * available only while the work item is queued.
658  *
659  * %WORK_OFFQ_CANCELING is used to mark a work item which is being
660  * canceled.  While being canceled, a work item may have its PENDING set
661  * but stay off timer and worklist for arbitrarily long and nobody should
662  * try to steal the PENDING bit.
663  */
664 static inline void set_work_data(struct work_struct *work, unsigned long data,
665 				 unsigned long flags)
666 {
667 	WARN_ON_ONCE(!work_pending(work));
668 	atomic_long_set(&work->data, data | flags | work_static(work));
669 }
670 
671 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
672 			 unsigned long extra_flags)
673 {
674 	set_work_data(work, (unsigned long)pwq,
675 		      WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
676 }
677 
678 static void set_work_pool_and_keep_pending(struct work_struct *work,
679 					   int pool_id)
680 {
681 	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
682 		      WORK_STRUCT_PENDING);
683 }
684 
685 static void set_work_pool_and_clear_pending(struct work_struct *work,
686 					    int pool_id)
687 {
688 	/*
689 	 * The following wmb is paired with the implied mb in
690 	 * test_and_set_bit(PENDING) and ensures all updates to @work made
691 	 * here are visible to and precede any updates by the next PENDING
692 	 * owner.
693 	 */
694 	smp_wmb();
695 	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
696 	/*
697 	 * The following mb guarantees that previous clear of a PENDING bit
698 	 * will not be reordered with any speculative LOADS or STORES from
699 	 * work->current_func, which is executed afterwards.  This possible
700 	 * reordering can lead to a missed execution on attempt to queue
701 	 * the same @work.  E.g. consider this case:
702 	 *
703 	 *   CPU#0                         CPU#1
704 	 *   ----------------------------  --------------------------------
705 	 *
706 	 * 1  STORE event_indicated
707 	 * 2  queue_work_on() {
708 	 * 3    test_and_set_bit(PENDING)
709 	 * 4 }                             set_..._and_clear_pending() {
710 	 * 5                                 set_work_data() # clear bit
711 	 * 6                                 smp_mb()
712 	 * 7                               work->current_func() {
713 	 * 8				      LOAD event_indicated
714 	 *				   }
715 	 *
716 	 * Without an explicit full barrier speculative LOAD on line 8 can
717 	 * be executed before CPU#0 does STORE on line 1.  If that happens,
718 	 * CPU#0 observes the PENDING bit is still set and new execution of
719 	 * a @work is not queued in a hope, that CPU#1 will eventually
720 	 * finish the queued @work.  Meanwhile CPU#1 does not see
721 	 * event_indicated is set, because speculative LOAD was executed
722 	 * before actual STORE.
723 	 */
724 	smp_mb();
725 }
726 
727 static void clear_work_data(struct work_struct *work)
728 {
729 	smp_wmb();	/* see set_work_pool_and_clear_pending() */
730 	set_work_data(work, WORK_STRUCT_NO_POOL, 0);
731 }
732 
733 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
734 {
735 	return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
736 }
737 
738 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
739 {
740 	unsigned long data = atomic_long_read(&work->data);
741 
742 	if (data & WORK_STRUCT_PWQ)
743 		return work_struct_pwq(data);
744 	else
745 		return NULL;
746 }
747 
748 /**
749  * get_work_pool - return the worker_pool a given work was associated with
750  * @work: the work item of interest
751  *
752  * Pools are created and destroyed under wq_pool_mutex, and allows read
753  * access under RCU read lock.  As such, this function should be
754  * called under wq_pool_mutex or inside of a rcu_read_lock() region.
755  *
756  * All fields of the returned pool are accessible as long as the above
757  * mentioned locking is in effect.  If the returned pool needs to be used
758  * beyond the critical section, the caller is responsible for ensuring the
759  * returned pool is and stays online.
760  *
761  * Return: The worker_pool @work was last associated with.  %NULL if none.
762  */
763 static struct worker_pool *get_work_pool(struct work_struct *work)
764 {
765 	unsigned long data = atomic_long_read(&work->data);
766 	int pool_id;
767 
768 	assert_rcu_or_pool_mutex();
769 
770 	if (data & WORK_STRUCT_PWQ)
771 		return work_struct_pwq(data)->pool;
772 
773 	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
774 	if (pool_id == WORK_OFFQ_POOL_NONE)
775 		return NULL;
776 
777 	return idr_find(&worker_pool_idr, pool_id);
778 }
779 
780 /**
781  * get_work_pool_id - return the worker pool ID a given work is associated with
782  * @work: the work item of interest
783  *
784  * Return: The worker_pool ID @work was last associated with.
785  * %WORK_OFFQ_POOL_NONE if none.
786  */
787 static int get_work_pool_id(struct work_struct *work)
788 {
789 	unsigned long data = atomic_long_read(&work->data);
790 
791 	if (data & WORK_STRUCT_PWQ)
792 		return work_struct_pwq(data)->pool->id;
793 
794 	return data >> WORK_OFFQ_POOL_SHIFT;
795 }
796 
797 static void mark_work_canceling(struct work_struct *work)
798 {
799 	unsigned long pool_id = get_work_pool_id(work);
800 
801 	pool_id <<= WORK_OFFQ_POOL_SHIFT;
802 	set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
803 }
804 
805 static bool work_is_canceling(struct work_struct *work)
806 {
807 	unsigned long data = atomic_long_read(&work->data);
808 
809 	return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
810 }
811 
812 /*
813  * Policy functions.  These define the policies on how the global worker
814  * pools are managed.  Unless noted otherwise, these functions assume that
815  * they're being called with pool->lock held.
816  */
817 
818 /*
819  * Need to wake up a worker?  Called from anything but currently
820  * running workers.
821  *
822  * Note that, because unbound workers never contribute to nr_running, this
823  * function will always return %true for unbound pools as long as the
824  * worklist isn't empty.
825  */
826 static bool need_more_worker(struct worker_pool *pool)
827 {
828 	return !list_empty(&pool->worklist) && !pool->nr_running;
829 }
830 
831 /* Can I start working?  Called from busy but !running workers. */
832 static bool may_start_working(struct worker_pool *pool)
833 {
834 	return pool->nr_idle;
835 }
836 
837 /* Do I need to keep working?  Called from currently running workers. */
838 static bool keep_working(struct worker_pool *pool)
839 {
840 	return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
841 }
842 
843 /* Do we need a new worker?  Called from manager. */
844 static bool need_to_create_worker(struct worker_pool *pool)
845 {
846 	return need_more_worker(pool) && !may_start_working(pool);
847 }
848 
849 /* Do we have too many workers and should some go away? */
850 static bool too_many_workers(struct worker_pool *pool)
851 {
852 	bool managing = pool->flags & POOL_MANAGER_ACTIVE;
853 	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
854 	int nr_busy = pool->nr_workers - nr_idle;
855 
856 	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
857 }
858 
859 /**
860  * worker_set_flags - set worker flags and adjust nr_running accordingly
861  * @worker: self
862  * @flags: flags to set
863  *
864  * Set @flags in @worker->flags and adjust nr_running accordingly.
865  */
866 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
867 {
868 	struct worker_pool *pool = worker->pool;
869 
870 	lockdep_assert_held(&pool->lock);
871 
872 	/* If transitioning into NOT_RUNNING, adjust nr_running. */
873 	if ((flags & WORKER_NOT_RUNNING) &&
874 	    !(worker->flags & WORKER_NOT_RUNNING)) {
875 		pool->nr_running--;
876 	}
877 
878 	worker->flags |= flags;
879 }
880 
881 /**
882  * worker_clr_flags - clear worker flags and adjust nr_running accordingly
883  * @worker: self
884  * @flags: flags to clear
885  *
886  * Clear @flags in @worker->flags and adjust nr_running accordingly.
887  */
888 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
889 {
890 	struct worker_pool *pool = worker->pool;
891 	unsigned int oflags = worker->flags;
892 
893 	lockdep_assert_held(&pool->lock);
894 
895 	worker->flags &= ~flags;
896 
897 	/*
898 	 * If transitioning out of NOT_RUNNING, increment nr_running.  Note
899 	 * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
900 	 * of multiple flags, not a single flag.
901 	 */
902 	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
903 		if (!(worker->flags & WORKER_NOT_RUNNING))
904 			pool->nr_running++;
905 }
906 
907 /* Return the first idle worker.  Called with pool->lock held. */
908 static struct worker *first_idle_worker(struct worker_pool *pool)
909 {
910 	if (unlikely(list_empty(&pool->idle_list)))
911 		return NULL;
912 
913 	return list_first_entry(&pool->idle_list, struct worker, entry);
914 }
915 
916 /**
917  * worker_enter_idle - enter idle state
918  * @worker: worker which is entering idle state
919  *
920  * @worker is entering idle state.  Update stats and idle timer if
921  * necessary.
922  *
923  * LOCKING:
924  * raw_spin_lock_irq(pool->lock).
925  */
926 static void worker_enter_idle(struct worker *worker)
927 {
928 	struct worker_pool *pool = worker->pool;
929 
930 	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
931 	    WARN_ON_ONCE(!list_empty(&worker->entry) &&
932 			 (worker->hentry.next || worker->hentry.pprev)))
933 		return;
934 
935 	/* can't use worker_set_flags(), also called from create_worker() */
936 	worker->flags |= WORKER_IDLE;
937 	pool->nr_idle++;
938 	worker->last_active = jiffies;
939 
940 	/* idle_list is LIFO */
941 	list_add(&worker->entry, &pool->idle_list);
942 
943 	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
944 		mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
945 
946 	/* Sanity check nr_running. */
947 	WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
948 }
949 
950 /**
951  * worker_leave_idle - leave idle state
952  * @worker: worker which is leaving idle state
953  *
954  * @worker is leaving idle state.  Update stats.
955  *
956  * LOCKING:
957  * raw_spin_lock_irq(pool->lock).
958  */
959 static void worker_leave_idle(struct worker *worker)
960 {
961 	struct worker_pool *pool = worker->pool;
962 
963 	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
964 		return;
965 	worker_clr_flags(worker, WORKER_IDLE);
966 	pool->nr_idle--;
967 	list_del_init(&worker->entry);
968 }
969 
970 /**
971  * find_worker_executing_work - find worker which is executing a work
972  * @pool: pool of interest
973  * @work: work to find worker for
974  *
975  * Find a worker which is executing @work on @pool by searching
976  * @pool->busy_hash which is keyed by the address of @work.  For a worker
977  * to match, its current execution should match the address of @work and
978  * its work function.  This is to avoid unwanted dependency between
979  * unrelated work executions through a work item being recycled while still
980  * being executed.
981  *
982  * This is a bit tricky.  A work item may be freed once its execution
983  * starts and nothing prevents the freed area from being recycled for
984  * another work item.  If the same work item address ends up being reused
985  * before the original execution finishes, workqueue will identify the
986  * recycled work item as currently executing and make it wait until the
987  * current execution finishes, introducing an unwanted dependency.
988  *
989  * This function checks the work item address and work function to avoid
990  * false positives.  Note that this isn't complete as one may construct a
991  * work function which can introduce dependency onto itself through a
992  * recycled work item.  Well, if somebody wants to shoot oneself in the
993  * foot that badly, there's only so much we can do, and if such deadlock
994  * actually occurs, it should be easy to locate the culprit work function.
995  *
996  * CONTEXT:
997  * raw_spin_lock_irq(pool->lock).
998  *
999  * Return:
1000  * Pointer to worker which is executing @work if found, %NULL
1001  * otherwise.
1002  */
1003 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1004 						 struct work_struct *work)
1005 {
1006 	struct worker *worker;
1007 
1008 	hash_for_each_possible(pool->busy_hash, worker, hentry,
1009 			       (unsigned long)work)
1010 		if (worker->current_work == work &&
1011 		    worker->current_func == work->func)
1012 			return worker;
1013 
1014 	return NULL;
1015 }
1016 
1017 /**
1018  * move_linked_works - move linked works to a list
1019  * @work: start of series of works to be scheduled
1020  * @head: target list to append @work to
1021  * @nextp: out parameter for nested worklist walking
1022  *
1023  * Schedule linked works starting from @work to @head. Work series to be
1024  * scheduled starts at @work and includes any consecutive work with
1025  * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1026  * @nextp.
1027  *
1028  * CONTEXT:
1029  * raw_spin_lock_irq(pool->lock).
1030  */
1031 static void move_linked_works(struct work_struct *work, struct list_head *head,
1032 			      struct work_struct **nextp)
1033 {
1034 	struct work_struct *n;
1035 
1036 	/*
1037 	 * Linked worklist will always end before the end of the list,
1038 	 * use NULL for list head.
1039 	 */
1040 	list_for_each_entry_safe_from(work, n, NULL, entry) {
1041 		list_move_tail(&work->entry, head);
1042 		if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1043 			break;
1044 	}
1045 
1046 	/*
1047 	 * If we're already inside safe list traversal and have moved
1048 	 * multiple works to the scheduled queue, the next position
1049 	 * needs to be updated.
1050 	 */
1051 	if (nextp)
1052 		*nextp = n;
1053 }
1054 
1055 /**
1056  * assign_work - assign a work item and its linked work items to a worker
1057  * @work: work to assign
1058  * @worker: worker to assign to
1059  * @nextp: out parameter for nested worklist walking
1060  *
1061  * Assign @work and its linked work items to @worker. If @work is already being
1062  * executed by another worker in the same pool, it'll be punted there.
1063  *
1064  * If @nextp is not NULL, it's updated to point to the next work of the last
1065  * scheduled work. This allows assign_work() to be nested inside
1066  * list_for_each_entry_safe().
1067  *
1068  * Returns %true if @work was successfully assigned to @worker. %false if @work
1069  * was punted to another worker already executing it.
1070  */
1071 static bool assign_work(struct work_struct *work, struct worker *worker,
1072 			struct work_struct **nextp)
1073 {
1074 	struct worker_pool *pool = worker->pool;
1075 	struct worker *collision;
1076 
1077 	lockdep_assert_held(&pool->lock);
1078 
1079 	/*
1080 	 * A single work shouldn't be executed concurrently by multiple workers.
1081 	 * __queue_work() ensures that @work doesn't jump to a different pool
1082 	 * while still running in the previous pool. Here, we should ensure that
1083 	 * @work is not executed concurrently by multiple workers from the same
1084 	 * pool. Check whether anyone is already processing the work. If so,
1085 	 * defer the work to the currently executing one.
1086 	 */
1087 	collision = find_worker_executing_work(pool, work);
1088 	if (unlikely(collision)) {
1089 		move_linked_works(work, &collision->scheduled, nextp);
1090 		return false;
1091 	}
1092 
1093 	move_linked_works(work, &worker->scheduled, nextp);
1094 	return true;
1095 }
1096 
1097 /**
1098  * kick_pool - wake up an idle worker if necessary
1099  * @pool: pool to kick
1100  *
1101  * @pool may have pending work items. Wake up worker if necessary. Returns
1102  * whether a worker was woken up.
1103  */
1104 static bool kick_pool(struct worker_pool *pool)
1105 {
1106 	struct worker *worker = first_idle_worker(pool);
1107 	struct task_struct *p;
1108 
1109 	lockdep_assert_held(&pool->lock);
1110 
1111 	if (!need_more_worker(pool) || !worker)
1112 		return false;
1113 
1114 	p = worker->task;
1115 
1116 #ifdef CONFIG_SMP
1117 	/*
1118 	 * Idle @worker is about to execute @work and waking up provides an
1119 	 * opportunity to migrate @worker at a lower cost by setting the task's
1120 	 * wake_cpu field. Let's see if we want to move @worker to improve
1121 	 * execution locality.
1122 	 *
1123 	 * We're waking the worker that went idle the latest and there's some
1124 	 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1125 	 * so, setting the wake_cpu won't do anything. As this is a best-effort
1126 	 * optimization and the race window is narrow, let's leave as-is for
1127 	 * now. If this becomes pronounced, we can skip over workers which are
1128 	 * still on cpu when picking an idle worker.
1129 	 *
1130 	 * If @pool has non-strict affinity, @worker might have ended up outside
1131 	 * its affinity scope. Repatriate.
1132 	 */
1133 	if (!pool->attrs->affn_strict &&
1134 	    !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1135 		struct work_struct *work = list_first_entry(&pool->worklist,
1136 						struct work_struct, entry);
1137 		p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask);
1138 		get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1139 	}
1140 #endif
1141 	wake_up_process(p);
1142 	return true;
1143 }
1144 
1145 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1146 
1147 /*
1148  * Concurrency-managed per-cpu work items that hog CPU for longer than
1149  * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1150  * which prevents them from stalling other concurrency-managed work items. If a
1151  * work function keeps triggering this mechanism, it's likely that the work item
1152  * should be using an unbound workqueue instead.
1153  *
1154  * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1155  * and report them so that they can be examined and converted to use unbound
1156  * workqueues as appropriate. To avoid flooding the console, each violating work
1157  * function is tracked and reported with exponential backoff.
1158  */
1159 #define WCI_MAX_ENTS 128
1160 
1161 struct wci_ent {
1162 	work_func_t		func;
1163 	atomic64_t		cnt;
1164 	struct hlist_node	hash_node;
1165 };
1166 
1167 static struct wci_ent wci_ents[WCI_MAX_ENTS];
1168 static int wci_nr_ents;
1169 static DEFINE_RAW_SPINLOCK(wci_lock);
1170 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1171 
1172 static struct wci_ent *wci_find_ent(work_func_t func)
1173 {
1174 	struct wci_ent *ent;
1175 
1176 	hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1177 				   (unsigned long)func) {
1178 		if (ent->func == func)
1179 			return ent;
1180 	}
1181 	return NULL;
1182 }
1183 
1184 static void wq_cpu_intensive_report(work_func_t func)
1185 {
1186 	struct wci_ent *ent;
1187 
1188 restart:
1189 	ent = wci_find_ent(func);
1190 	if (ent) {
1191 		u64 cnt;
1192 
1193 		/*
1194 		 * Start reporting from the fourth time and back off
1195 		 * exponentially.
1196 		 */
1197 		cnt = atomic64_inc_return_relaxed(&ent->cnt);
1198 		if (cnt >= 4 && is_power_of_2(cnt))
1199 			printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1200 					ent->func, wq_cpu_intensive_thresh_us,
1201 					atomic64_read(&ent->cnt));
1202 		return;
1203 	}
1204 
1205 	/*
1206 	 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1207 	 * is exhausted, something went really wrong and we probably made enough
1208 	 * noise already.
1209 	 */
1210 	if (wci_nr_ents >= WCI_MAX_ENTS)
1211 		return;
1212 
1213 	raw_spin_lock(&wci_lock);
1214 
1215 	if (wci_nr_ents >= WCI_MAX_ENTS) {
1216 		raw_spin_unlock(&wci_lock);
1217 		return;
1218 	}
1219 
1220 	if (wci_find_ent(func)) {
1221 		raw_spin_unlock(&wci_lock);
1222 		goto restart;
1223 	}
1224 
1225 	ent = &wci_ents[wci_nr_ents++];
1226 	ent->func = func;
1227 	atomic64_set(&ent->cnt, 1);
1228 	hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1229 
1230 	raw_spin_unlock(&wci_lock);
1231 }
1232 
1233 #else	/* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1234 static void wq_cpu_intensive_report(work_func_t func) {}
1235 #endif	/* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1236 
1237 /**
1238  * wq_worker_running - a worker is running again
1239  * @task: task waking up
1240  *
1241  * This function is called when a worker returns from schedule()
1242  */
1243 void wq_worker_running(struct task_struct *task)
1244 {
1245 	struct worker *worker = kthread_data(task);
1246 
1247 	if (!READ_ONCE(worker->sleeping))
1248 		return;
1249 
1250 	/*
1251 	 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1252 	 * and the nr_running increment below, we may ruin the nr_running reset
1253 	 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1254 	 * pool. Protect against such race.
1255 	 */
1256 	preempt_disable();
1257 	if (!(worker->flags & WORKER_NOT_RUNNING))
1258 		worker->pool->nr_running++;
1259 	preempt_enable();
1260 
1261 	/*
1262 	 * CPU intensive auto-detection cares about how long a work item hogged
1263 	 * CPU without sleeping. Reset the starting timestamp on wakeup.
1264 	 */
1265 	worker->current_at = worker->task->se.sum_exec_runtime;
1266 
1267 	WRITE_ONCE(worker->sleeping, 0);
1268 }
1269 
1270 /**
1271  * wq_worker_sleeping - a worker is going to sleep
1272  * @task: task going to sleep
1273  *
1274  * This function is called from schedule() when a busy worker is
1275  * going to sleep.
1276  */
1277 void wq_worker_sleeping(struct task_struct *task)
1278 {
1279 	struct worker *worker = kthread_data(task);
1280 	struct worker_pool *pool;
1281 
1282 	/*
1283 	 * Rescuers, which may not have all the fields set up like normal
1284 	 * workers, also reach here, let's not access anything before
1285 	 * checking NOT_RUNNING.
1286 	 */
1287 	if (worker->flags & WORKER_NOT_RUNNING)
1288 		return;
1289 
1290 	pool = worker->pool;
1291 
1292 	/* Return if preempted before wq_worker_running() was reached */
1293 	if (READ_ONCE(worker->sleeping))
1294 		return;
1295 
1296 	WRITE_ONCE(worker->sleeping, 1);
1297 	raw_spin_lock_irq(&pool->lock);
1298 
1299 	/*
1300 	 * Recheck in case unbind_workers() preempted us. We don't
1301 	 * want to decrement nr_running after the worker is unbound
1302 	 * and nr_running has been reset.
1303 	 */
1304 	if (worker->flags & WORKER_NOT_RUNNING) {
1305 		raw_spin_unlock_irq(&pool->lock);
1306 		return;
1307 	}
1308 
1309 	pool->nr_running--;
1310 	if (kick_pool(pool))
1311 		worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1312 
1313 	raw_spin_unlock_irq(&pool->lock);
1314 }
1315 
1316 /**
1317  * wq_worker_tick - a scheduler tick occurred while a kworker is running
1318  * @task: task currently running
1319  *
1320  * Called from scheduler_tick(). We're in the IRQ context and the current
1321  * worker's fields which follow the 'K' locking rule can be accessed safely.
1322  */
1323 void wq_worker_tick(struct task_struct *task)
1324 {
1325 	struct worker *worker = kthread_data(task);
1326 	struct pool_workqueue *pwq = worker->current_pwq;
1327 	struct worker_pool *pool = worker->pool;
1328 
1329 	if (!pwq)
1330 		return;
1331 
1332 	pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1333 
1334 	if (!wq_cpu_intensive_thresh_us)
1335 		return;
1336 
1337 	/*
1338 	 * If the current worker is concurrency managed and hogged the CPU for
1339 	 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1340 	 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1341 	 *
1342 	 * Set @worker->sleeping means that @worker is in the process of
1343 	 * switching out voluntarily and won't be contributing to
1344 	 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1345 	 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1346 	 * double decrements. The task is releasing the CPU anyway. Let's skip.
1347 	 * We probably want to make this prettier in the future.
1348 	 */
1349 	if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1350 	    worker->task->se.sum_exec_runtime - worker->current_at <
1351 	    wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1352 		return;
1353 
1354 	raw_spin_lock(&pool->lock);
1355 
1356 	worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1357 	wq_cpu_intensive_report(worker->current_func);
1358 	pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1359 
1360 	if (kick_pool(pool))
1361 		pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1362 
1363 	raw_spin_unlock(&pool->lock);
1364 }
1365 
1366 /**
1367  * wq_worker_last_func - retrieve worker's last work function
1368  * @task: Task to retrieve last work function of.
1369  *
1370  * Determine the last function a worker executed. This is called from
1371  * the scheduler to get a worker's last known identity.
1372  *
1373  * CONTEXT:
1374  * raw_spin_lock_irq(rq->lock)
1375  *
1376  * This function is called during schedule() when a kworker is going
1377  * to sleep. It's used by psi to identify aggregation workers during
1378  * dequeuing, to allow periodic aggregation to shut-off when that
1379  * worker is the last task in the system or cgroup to go to sleep.
1380  *
1381  * As this function doesn't involve any workqueue-related locking, it
1382  * only returns stable values when called from inside the scheduler's
1383  * queuing and dequeuing paths, when @task, which must be a kworker,
1384  * is guaranteed to not be processing any works.
1385  *
1386  * Return:
1387  * The last work function %current executed as a worker, NULL if it
1388  * hasn't executed any work yet.
1389  */
1390 work_func_t wq_worker_last_func(struct task_struct *task)
1391 {
1392 	struct worker *worker = kthread_data(task);
1393 
1394 	return worker->last_func;
1395 }
1396 
1397 /**
1398  * get_pwq - get an extra reference on the specified pool_workqueue
1399  * @pwq: pool_workqueue to get
1400  *
1401  * Obtain an extra reference on @pwq.  The caller should guarantee that
1402  * @pwq has positive refcnt and be holding the matching pool->lock.
1403  */
1404 static void get_pwq(struct pool_workqueue *pwq)
1405 {
1406 	lockdep_assert_held(&pwq->pool->lock);
1407 	WARN_ON_ONCE(pwq->refcnt <= 0);
1408 	pwq->refcnt++;
1409 }
1410 
1411 /**
1412  * put_pwq - put a pool_workqueue reference
1413  * @pwq: pool_workqueue to put
1414  *
1415  * Drop a reference of @pwq.  If its refcnt reaches zero, schedule its
1416  * destruction.  The caller should be holding the matching pool->lock.
1417  */
1418 static void put_pwq(struct pool_workqueue *pwq)
1419 {
1420 	lockdep_assert_held(&pwq->pool->lock);
1421 	if (likely(--pwq->refcnt))
1422 		return;
1423 	/*
1424 	 * @pwq can't be released under pool->lock, bounce to a dedicated
1425 	 * kthread_worker to avoid A-A deadlocks.
1426 	 */
1427 	kthread_queue_work(pwq_release_worker, &pwq->release_work);
1428 }
1429 
1430 /**
1431  * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1432  * @pwq: pool_workqueue to put (can be %NULL)
1433  *
1434  * put_pwq() with locking.  This function also allows %NULL @pwq.
1435  */
1436 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1437 {
1438 	if (pwq) {
1439 		/*
1440 		 * As both pwqs and pools are RCU protected, the
1441 		 * following lock operations are safe.
1442 		 */
1443 		raw_spin_lock_irq(&pwq->pool->lock);
1444 		put_pwq(pwq);
1445 		raw_spin_unlock_irq(&pwq->pool->lock);
1446 	}
1447 }
1448 
1449 static void pwq_activate_inactive_work(struct work_struct *work)
1450 {
1451 	struct pool_workqueue *pwq = get_work_pwq(work);
1452 
1453 	trace_workqueue_activate_work(work);
1454 	if (list_empty(&pwq->pool->worklist))
1455 		pwq->pool->watchdog_ts = jiffies;
1456 	move_linked_works(work, &pwq->pool->worklist, NULL);
1457 	__clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work));
1458 	pwq->nr_active++;
1459 }
1460 
1461 static void pwq_activate_first_inactive(struct pool_workqueue *pwq)
1462 {
1463 	struct work_struct *work = list_first_entry(&pwq->inactive_works,
1464 						    struct work_struct, entry);
1465 
1466 	pwq_activate_inactive_work(work);
1467 }
1468 
1469 /**
1470  * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1471  * @pwq: pwq of interest
1472  * @work_data: work_data of work which left the queue
1473  *
1474  * A work either has completed or is removed from pending queue,
1475  * decrement nr_in_flight of its pwq and handle workqueue flushing.
1476  *
1477  * CONTEXT:
1478  * raw_spin_lock_irq(pool->lock).
1479  */
1480 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1481 {
1482 	int color = get_work_color(work_data);
1483 
1484 	if (!(work_data & WORK_STRUCT_INACTIVE)) {
1485 		pwq->nr_active--;
1486 		if (!list_empty(&pwq->inactive_works)) {
1487 			/* one down, submit an inactive one */
1488 			if (pwq->nr_active < pwq->max_active)
1489 				pwq_activate_first_inactive(pwq);
1490 		}
1491 	}
1492 
1493 	pwq->nr_in_flight[color]--;
1494 
1495 	/* is flush in progress and are we at the flushing tip? */
1496 	if (likely(pwq->flush_color != color))
1497 		goto out_put;
1498 
1499 	/* are there still in-flight works? */
1500 	if (pwq->nr_in_flight[color])
1501 		goto out_put;
1502 
1503 	/* this pwq is done, clear flush_color */
1504 	pwq->flush_color = -1;
1505 
1506 	/*
1507 	 * If this was the last pwq, wake up the first flusher.  It
1508 	 * will handle the rest.
1509 	 */
1510 	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1511 		complete(&pwq->wq->first_flusher->done);
1512 out_put:
1513 	put_pwq(pwq);
1514 }
1515 
1516 /**
1517  * try_to_grab_pending - steal work item from worklist and disable irq
1518  * @work: work item to steal
1519  * @is_dwork: @work is a delayed_work
1520  * @flags: place to store irq state
1521  *
1522  * Try to grab PENDING bit of @work.  This function can handle @work in any
1523  * stable state - idle, on timer or on worklist.
1524  *
1525  * Return:
1526  *
1527  *  ========	================================================================
1528  *  1		if @work was pending and we successfully stole PENDING
1529  *  0		if @work was idle and we claimed PENDING
1530  *  -EAGAIN	if PENDING couldn't be grabbed at the moment, safe to busy-retry
1531  *  -ENOENT	if someone else is canceling @work, this state may persist
1532  *		for arbitrarily long
1533  *  ========	================================================================
1534  *
1535  * Note:
1536  * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1537  * interrupted while holding PENDING and @work off queue, irq must be
1538  * disabled on entry.  This, combined with delayed_work->timer being
1539  * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1540  *
1541  * On successful return, >= 0, irq is disabled and the caller is
1542  * responsible for releasing it using local_irq_restore(*@flags).
1543  *
1544  * This function is safe to call from any context including IRQ handler.
1545  */
1546 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1547 			       unsigned long *flags)
1548 {
1549 	struct worker_pool *pool;
1550 	struct pool_workqueue *pwq;
1551 
1552 	local_irq_save(*flags);
1553 
1554 	/* try to steal the timer if it exists */
1555 	if (is_dwork) {
1556 		struct delayed_work *dwork = to_delayed_work(work);
1557 
1558 		/*
1559 		 * dwork->timer is irqsafe.  If del_timer() fails, it's
1560 		 * guaranteed that the timer is not queued anywhere and not
1561 		 * running on the local CPU.
1562 		 */
1563 		if (likely(del_timer(&dwork->timer)))
1564 			return 1;
1565 	}
1566 
1567 	/* try to claim PENDING the normal way */
1568 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1569 		return 0;
1570 
1571 	rcu_read_lock();
1572 	/*
1573 	 * The queueing is in progress, or it is already queued. Try to
1574 	 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1575 	 */
1576 	pool = get_work_pool(work);
1577 	if (!pool)
1578 		goto fail;
1579 
1580 	raw_spin_lock(&pool->lock);
1581 	/*
1582 	 * work->data is guaranteed to point to pwq only while the work
1583 	 * item is queued on pwq->wq, and both updating work->data to point
1584 	 * to pwq on queueing and to pool on dequeueing are done under
1585 	 * pwq->pool->lock.  This in turn guarantees that, if work->data
1586 	 * points to pwq which is associated with a locked pool, the work
1587 	 * item is currently queued on that pool.
1588 	 */
1589 	pwq = get_work_pwq(work);
1590 	if (pwq && pwq->pool == pool) {
1591 		debug_work_deactivate(work);
1592 
1593 		/*
1594 		 * A cancelable inactive work item must be in the
1595 		 * pwq->inactive_works since a queued barrier can't be
1596 		 * canceled (see the comments in insert_wq_barrier()).
1597 		 *
1598 		 * An inactive work item cannot be grabbed directly because
1599 		 * it might have linked barrier work items which, if left
1600 		 * on the inactive_works list, will confuse pwq->nr_active
1601 		 * management later on and cause stall.  Make sure the work
1602 		 * item is activated before grabbing.
1603 		 */
1604 		if (*work_data_bits(work) & WORK_STRUCT_INACTIVE)
1605 			pwq_activate_inactive_work(work);
1606 
1607 		list_del_init(&work->entry);
1608 		pwq_dec_nr_in_flight(pwq, *work_data_bits(work));
1609 
1610 		/* work->data points to pwq iff queued, point to pool */
1611 		set_work_pool_and_keep_pending(work, pool->id);
1612 
1613 		raw_spin_unlock(&pool->lock);
1614 		rcu_read_unlock();
1615 		return 1;
1616 	}
1617 	raw_spin_unlock(&pool->lock);
1618 fail:
1619 	rcu_read_unlock();
1620 	local_irq_restore(*flags);
1621 	if (work_is_canceling(work))
1622 		return -ENOENT;
1623 	cpu_relax();
1624 	return -EAGAIN;
1625 }
1626 
1627 /**
1628  * insert_work - insert a work into a pool
1629  * @pwq: pwq @work belongs to
1630  * @work: work to insert
1631  * @head: insertion point
1632  * @extra_flags: extra WORK_STRUCT_* flags to set
1633  *
1634  * Insert @work which belongs to @pwq after @head.  @extra_flags is or'd to
1635  * work_struct flags.
1636  *
1637  * CONTEXT:
1638  * raw_spin_lock_irq(pool->lock).
1639  */
1640 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1641 			struct list_head *head, unsigned int extra_flags)
1642 {
1643 	debug_work_activate(work);
1644 
1645 	/* record the work call stack in order to print it in KASAN reports */
1646 	kasan_record_aux_stack_noalloc(work);
1647 
1648 	/* we own @work, set data and link */
1649 	set_work_pwq(work, pwq, extra_flags);
1650 	list_add_tail(&work->entry, head);
1651 	get_pwq(pwq);
1652 }
1653 
1654 /*
1655  * Test whether @work is being queued from another work executing on the
1656  * same workqueue.
1657  */
1658 static bool is_chained_work(struct workqueue_struct *wq)
1659 {
1660 	struct worker *worker;
1661 
1662 	worker = current_wq_worker();
1663 	/*
1664 	 * Return %true iff I'm a worker executing a work item on @wq.  If
1665 	 * I'm @worker, it's safe to dereference it without locking.
1666 	 */
1667 	return worker && worker->current_pwq->wq == wq;
1668 }
1669 
1670 /*
1671  * When queueing an unbound work item to a wq, prefer local CPU if allowed
1672  * by wq_unbound_cpumask.  Otherwise, round robin among the allowed ones to
1673  * avoid perturbing sensitive tasks.
1674  */
1675 static int wq_select_unbound_cpu(int cpu)
1676 {
1677 	int new_cpu;
1678 
1679 	if (likely(!wq_debug_force_rr_cpu)) {
1680 		if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1681 			return cpu;
1682 	} else {
1683 		pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
1684 	}
1685 
1686 	if (cpumask_empty(wq_unbound_cpumask))
1687 		return cpu;
1688 
1689 	new_cpu = __this_cpu_read(wq_rr_cpu_last);
1690 	new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1691 	if (unlikely(new_cpu >= nr_cpu_ids)) {
1692 		new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1693 		if (unlikely(new_cpu >= nr_cpu_ids))
1694 			return cpu;
1695 	}
1696 	__this_cpu_write(wq_rr_cpu_last, new_cpu);
1697 
1698 	return new_cpu;
1699 }
1700 
1701 static void __queue_work(int cpu, struct workqueue_struct *wq,
1702 			 struct work_struct *work)
1703 {
1704 	struct pool_workqueue *pwq;
1705 	struct worker_pool *last_pool, *pool;
1706 	unsigned int work_flags;
1707 	unsigned int req_cpu = cpu;
1708 
1709 	/*
1710 	 * While a work item is PENDING && off queue, a task trying to
1711 	 * steal the PENDING will busy-loop waiting for it to either get
1712 	 * queued or lose PENDING.  Grabbing PENDING and queueing should
1713 	 * happen with IRQ disabled.
1714 	 */
1715 	lockdep_assert_irqs_disabled();
1716 
1717 
1718 	/*
1719 	 * For a draining wq, only works from the same workqueue are
1720 	 * allowed. The __WQ_DESTROYING helps to spot the issue that
1721 	 * queues a new work item to a wq after destroy_workqueue(wq).
1722 	 */
1723 	if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
1724 		     WARN_ON_ONCE(!is_chained_work(wq))))
1725 		return;
1726 	rcu_read_lock();
1727 retry:
1728 	/* pwq which will be used unless @work is executing elsewhere */
1729 	if (req_cpu == WORK_CPU_UNBOUND) {
1730 		if (wq->flags & WQ_UNBOUND)
1731 			cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1732 		else
1733 			cpu = raw_smp_processor_id();
1734 	}
1735 
1736 	pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
1737 	pool = pwq->pool;
1738 
1739 	/*
1740 	 * If @work was previously on a different pool, it might still be
1741 	 * running there, in which case the work needs to be queued on that
1742 	 * pool to guarantee non-reentrancy.
1743 	 */
1744 	last_pool = get_work_pool(work);
1745 	if (last_pool && last_pool != pool) {
1746 		struct worker *worker;
1747 
1748 		raw_spin_lock(&last_pool->lock);
1749 
1750 		worker = find_worker_executing_work(last_pool, work);
1751 
1752 		if (worker && worker->current_pwq->wq == wq) {
1753 			pwq = worker->current_pwq;
1754 			pool = pwq->pool;
1755 			WARN_ON_ONCE(pool != last_pool);
1756 		} else {
1757 			/* meh... not running there, queue here */
1758 			raw_spin_unlock(&last_pool->lock);
1759 			raw_spin_lock(&pool->lock);
1760 		}
1761 	} else {
1762 		raw_spin_lock(&pool->lock);
1763 	}
1764 
1765 	/*
1766 	 * pwq is determined and locked. For unbound pools, we could have raced
1767 	 * with pwq release and it could already be dead. If its refcnt is zero,
1768 	 * repeat pwq selection. Note that unbound pwqs never die without
1769 	 * another pwq replacing it in cpu_pwq or while work items are executing
1770 	 * on it, so the retrying is guaranteed to make forward-progress.
1771 	 */
1772 	if (unlikely(!pwq->refcnt)) {
1773 		if (wq->flags & WQ_UNBOUND) {
1774 			raw_spin_unlock(&pool->lock);
1775 			cpu_relax();
1776 			goto retry;
1777 		}
1778 		/* oops */
1779 		WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1780 			  wq->name, cpu);
1781 	}
1782 
1783 	/* pwq determined, queue */
1784 	trace_workqueue_queue_work(req_cpu, pwq, work);
1785 
1786 	if (WARN_ON(!list_empty(&work->entry)))
1787 		goto out;
1788 
1789 	pwq->nr_in_flight[pwq->work_color]++;
1790 	work_flags = work_color_to_flags(pwq->work_color);
1791 
1792 	if (likely(pwq->nr_active < pwq->max_active)) {
1793 		if (list_empty(&pool->worklist))
1794 			pool->watchdog_ts = jiffies;
1795 
1796 		trace_workqueue_activate_work(work);
1797 		pwq->nr_active++;
1798 		insert_work(pwq, work, &pool->worklist, work_flags);
1799 		kick_pool(pool);
1800 	} else {
1801 		work_flags |= WORK_STRUCT_INACTIVE;
1802 		insert_work(pwq, work, &pwq->inactive_works, work_flags);
1803 	}
1804 
1805 out:
1806 	raw_spin_unlock(&pool->lock);
1807 	rcu_read_unlock();
1808 }
1809 
1810 /**
1811  * queue_work_on - queue work on specific cpu
1812  * @cpu: CPU number to execute work on
1813  * @wq: workqueue to use
1814  * @work: work to queue
1815  *
1816  * We queue the work to a specific CPU, the caller must ensure it
1817  * can't go away.  Callers that fail to ensure that the specified
1818  * CPU cannot go away will execute on a randomly chosen CPU.
1819  * But note well that callers specifying a CPU that never has been
1820  * online will get a splat.
1821  *
1822  * Return: %false if @work was already on a queue, %true otherwise.
1823  */
1824 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1825 		   struct work_struct *work)
1826 {
1827 	bool ret = false;
1828 	unsigned long flags;
1829 
1830 	local_irq_save(flags);
1831 
1832 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1833 		__queue_work(cpu, wq, work);
1834 		ret = true;
1835 	}
1836 
1837 	local_irq_restore(flags);
1838 	return ret;
1839 }
1840 EXPORT_SYMBOL(queue_work_on);
1841 
1842 /**
1843  * select_numa_node_cpu - Select a CPU based on NUMA node
1844  * @node: NUMA node ID that we want to select a CPU from
1845  *
1846  * This function will attempt to find a "random" cpu available on a given
1847  * node. If there are no CPUs available on the given node it will return
1848  * WORK_CPU_UNBOUND indicating that we should just schedule to any
1849  * available CPU if we need to schedule this work.
1850  */
1851 static int select_numa_node_cpu(int node)
1852 {
1853 	int cpu;
1854 
1855 	/* Delay binding to CPU if node is not valid or online */
1856 	if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1857 		return WORK_CPU_UNBOUND;
1858 
1859 	/* Use local node/cpu if we are already there */
1860 	cpu = raw_smp_processor_id();
1861 	if (node == cpu_to_node(cpu))
1862 		return cpu;
1863 
1864 	/* Use "random" otherwise know as "first" online CPU of node */
1865 	cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1866 
1867 	/* If CPU is valid return that, otherwise just defer */
1868 	return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1869 }
1870 
1871 /**
1872  * queue_work_node - queue work on a "random" cpu for a given NUMA node
1873  * @node: NUMA node that we are targeting the work for
1874  * @wq: workqueue to use
1875  * @work: work to queue
1876  *
1877  * We queue the work to a "random" CPU within a given NUMA node. The basic
1878  * idea here is to provide a way to somehow associate work with a given
1879  * NUMA node.
1880  *
1881  * This function will only make a best effort attempt at getting this onto
1882  * the right NUMA node. If no node is requested or the requested node is
1883  * offline then we just fall back to standard queue_work behavior.
1884  *
1885  * Currently the "random" CPU ends up being the first available CPU in the
1886  * intersection of cpu_online_mask and the cpumask of the node, unless we
1887  * are running on the node. In that case we just use the current CPU.
1888  *
1889  * Return: %false if @work was already on a queue, %true otherwise.
1890  */
1891 bool queue_work_node(int node, struct workqueue_struct *wq,
1892 		     struct work_struct *work)
1893 {
1894 	unsigned long flags;
1895 	bool ret = false;
1896 
1897 	/*
1898 	 * This current implementation is specific to unbound workqueues.
1899 	 * Specifically we only return the first available CPU for a given
1900 	 * node instead of cycling through individual CPUs within the node.
1901 	 *
1902 	 * If this is used with a per-cpu workqueue then the logic in
1903 	 * workqueue_select_cpu_near would need to be updated to allow for
1904 	 * some round robin type logic.
1905 	 */
1906 	WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1907 
1908 	local_irq_save(flags);
1909 
1910 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1911 		int cpu = select_numa_node_cpu(node);
1912 
1913 		__queue_work(cpu, wq, work);
1914 		ret = true;
1915 	}
1916 
1917 	local_irq_restore(flags);
1918 	return ret;
1919 }
1920 EXPORT_SYMBOL_GPL(queue_work_node);
1921 
1922 void delayed_work_timer_fn(struct timer_list *t)
1923 {
1924 	struct delayed_work *dwork = from_timer(dwork, t, timer);
1925 
1926 	/* should have been called from irqsafe timer with irq already off */
1927 	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
1928 }
1929 EXPORT_SYMBOL(delayed_work_timer_fn);
1930 
1931 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1932 				struct delayed_work *dwork, unsigned long delay)
1933 {
1934 	struct timer_list *timer = &dwork->timer;
1935 	struct work_struct *work = &dwork->work;
1936 
1937 	WARN_ON_ONCE(!wq);
1938 	WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1939 	WARN_ON_ONCE(timer_pending(timer));
1940 	WARN_ON_ONCE(!list_empty(&work->entry));
1941 
1942 	/*
1943 	 * If @delay is 0, queue @dwork->work immediately.  This is for
1944 	 * both optimization and correctness.  The earliest @timer can
1945 	 * expire is on the closest next tick and delayed_work users depend
1946 	 * on that there's no such delay when @delay is 0.
1947 	 */
1948 	if (!delay) {
1949 		__queue_work(cpu, wq, &dwork->work);
1950 		return;
1951 	}
1952 
1953 	dwork->wq = wq;
1954 	dwork->cpu = cpu;
1955 	timer->expires = jiffies + delay;
1956 
1957 	if (unlikely(cpu != WORK_CPU_UNBOUND))
1958 		add_timer_on(timer, cpu);
1959 	else
1960 		add_timer(timer);
1961 }
1962 
1963 /**
1964  * queue_delayed_work_on - queue work on specific CPU after delay
1965  * @cpu: CPU number to execute work on
1966  * @wq: workqueue to use
1967  * @dwork: work to queue
1968  * @delay: number of jiffies to wait before queueing
1969  *
1970  * Return: %false if @work was already on a queue, %true otherwise.  If
1971  * @delay is zero and @dwork is idle, it will be scheduled for immediate
1972  * execution.
1973  */
1974 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1975 			   struct delayed_work *dwork, unsigned long delay)
1976 {
1977 	struct work_struct *work = &dwork->work;
1978 	bool ret = false;
1979 	unsigned long flags;
1980 
1981 	/* read the comment in __queue_work() */
1982 	local_irq_save(flags);
1983 
1984 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1985 		__queue_delayed_work(cpu, wq, dwork, delay);
1986 		ret = true;
1987 	}
1988 
1989 	local_irq_restore(flags);
1990 	return ret;
1991 }
1992 EXPORT_SYMBOL(queue_delayed_work_on);
1993 
1994 /**
1995  * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1996  * @cpu: CPU number to execute work on
1997  * @wq: workqueue to use
1998  * @dwork: work to queue
1999  * @delay: number of jiffies to wait before queueing
2000  *
2001  * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2002  * modify @dwork's timer so that it expires after @delay.  If @delay is
2003  * zero, @work is guaranteed to be scheduled immediately regardless of its
2004  * current state.
2005  *
2006  * Return: %false if @dwork was idle and queued, %true if @dwork was
2007  * pending and its timer was modified.
2008  *
2009  * This function is safe to call from any context including IRQ handler.
2010  * See try_to_grab_pending() for details.
2011  */
2012 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2013 			 struct delayed_work *dwork, unsigned long delay)
2014 {
2015 	unsigned long flags;
2016 	int ret;
2017 
2018 	do {
2019 		ret = try_to_grab_pending(&dwork->work, true, &flags);
2020 	} while (unlikely(ret == -EAGAIN));
2021 
2022 	if (likely(ret >= 0)) {
2023 		__queue_delayed_work(cpu, wq, dwork, delay);
2024 		local_irq_restore(flags);
2025 	}
2026 
2027 	/* -ENOENT from try_to_grab_pending() becomes %true */
2028 	return ret;
2029 }
2030 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2031 
2032 static void rcu_work_rcufn(struct rcu_head *rcu)
2033 {
2034 	struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2035 
2036 	/* read the comment in __queue_work() */
2037 	local_irq_disable();
2038 	__queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2039 	local_irq_enable();
2040 }
2041 
2042 /**
2043  * queue_rcu_work - queue work after a RCU grace period
2044  * @wq: workqueue to use
2045  * @rwork: work to queue
2046  *
2047  * Return: %false if @rwork was already pending, %true otherwise.  Note
2048  * that a full RCU grace period is guaranteed only after a %true return.
2049  * While @rwork is guaranteed to be executed after a %false return, the
2050  * execution may happen before a full RCU grace period has passed.
2051  */
2052 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2053 {
2054 	struct work_struct *work = &rwork->work;
2055 
2056 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2057 		rwork->wq = wq;
2058 		call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2059 		return true;
2060 	}
2061 
2062 	return false;
2063 }
2064 EXPORT_SYMBOL(queue_rcu_work);
2065 
2066 static struct worker *alloc_worker(int node)
2067 {
2068 	struct worker *worker;
2069 
2070 	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2071 	if (worker) {
2072 		INIT_LIST_HEAD(&worker->entry);
2073 		INIT_LIST_HEAD(&worker->scheduled);
2074 		INIT_LIST_HEAD(&worker->node);
2075 		/* on creation a worker is in !idle && prep state */
2076 		worker->flags = WORKER_PREP;
2077 	}
2078 	return worker;
2079 }
2080 
2081 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2082 {
2083 	if (pool->cpu < 0 && pool->attrs->affn_strict)
2084 		return pool->attrs->__pod_cpumask;
2085 	else
2086 		return pool->attrs->cpumask;
2087 }
2088 
2089 /**
2090  * worker_attach_to_pool() - attach a worker to a pool
2091  * @worker: worker to be attached
2092  * @pool: the target pool
2093  *
2094  * Attach @worker to @pool.  Once attached, the %WORKER_UNBOUND flag and
2095  * cpu-binding of @worker are kept coordinated with the pool across
2096  * cpu-[un]hotplugs.
2097  */
2098 static void worker_attach_to_pool(struct worker *worker,
2099 				   struct worker_pool *pool)
2100 {
2101 	mutex_lock(&wq_pool_attach_mutex);
2102 
2103 	/*
2104 	 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
2105 	 * stable across this function.  See the comments above the flag
2106 	 * definition for details.
2107 	 */
2108 	if (pool->flags & POOL_DISASSOCIATED)
2109 		worker->flags |= WORKER_UNBOUND;
2110 	else
2111 		kthread_set_per_cpu(worker->task, pool->cpu);
2112 
2113 	if (worker->rescue_wq)
2114 		set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2115 
2116 	list_add_tail(&worker->node, &pool->workers);
2117 	worker->pool = pool;
2118 
2119 	mutex_unlock(&wq_pool_attach_mutex);
2120 }
2121 
2122 /**
2123  * worker_detach_from_pool() - detach a worker from its pool
2124  * @worker: worker which is attached to its pool
2125  *
2126  * Undo the attaching which had been done in worker_attach_to_pool().  The
2127  * caller worker shouldn't access to the pool after detached except it has
2128  * other reference to the pool.
2129  */
2130 static void worker_detach_from_pool(struct worker *worker)
2131 {
2132 	struct worker_pool *pool = worker->pool;
2133 	struct completion *detach_completion = NULL;
2134 
2135 	mutex_lock(&wq_pool_attach_mutex);
2136 
2137 	kthread_set_per_cpu(worker->task, -1);
2138 	list_del(&worker->node);
2139 	worker->pool = NULL;
2140 
2141 	if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
2142 		detach_completion = pool->detach_completion;
2143 	mutex_unlock(&wq_pool_attach_mutex);
2144 
2145 	/* clear leftover flags without pool->lock after it is detached */
2146 	worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2147 
2148 	if (detach_completion)
2149 		complete(detach_completion);
2150 }
2151 
2152 /**
2153  * create_worker - create a new workqueue worker
2154  * @pool: pool the new worker will belong to
2155  *
2156  * Create and start a new worker which is attached to @pool.
2157  *
2158  * CONTEXT:
2159  * Might sleep.  Does GFP_KERNEL allocations.
2160  *
2161  * Return:
2162  * Pointer to the newly created worker.
2163  */
2164 static struct worker *create_worker(struct worker_pool *pool)
2165 {
2166 	struct worker *worker;
2167 	int id;
2168 	char id_buf[16];
2169 
2170 	/* ID is needed to determine kthread name */
2171 	id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2172 	if (id < 0) {
2173 		pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2174 			    ERR_PTR(id));
2175 		return NULL;
2176 	}
2177 
2178 	worker = alloc_worker(pool->node);
2179 	if (!worker) {
2180 		pr_err_once("workqueue: Failed to allocate a worker\n");
2181 		goto fail;
2182 	}
2183 
2184 	worker->id = id;
2185 
2186 	if (pool->cpu >= 0)
2187 		snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
2188 			 pool->attrs->nice < 0  ? "H" : "");
2189 	else
2190 		snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2191 
2192 	worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
2193 					      "kworker/%s", id_buf);
2194 	if (IS_ERR(worker->task)) {
2195 		if (PTR_ERR(worker->task) == -EINTR) {
2196 			pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2197 			       id_buf);
2198 		} else {
2199 			pr_err_once("workqueue: Failed to create a worker thread: %pe",
2200 				    worker->task);
2201 		}
2202 		goto fail;
2203 	}
2204 
2205 	set_user_nice(worker->task, pool->attrs->nice);
2206 	kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2207 
2208 	/* successful, attach the worker to the pool */
2209 	worker_attach_to_pool(worker, pool);
2210 
2211 	/* start the newly created worker */
2212 	raw_spin_lock_irq(&pool->lock);
2213 
2214 	worker->pool->nr_workers++;
2215 	worker_enter_idle(worker);
2216 	kick_pool(pool);
2217 
2218 	/*
2219 	 * @worker is waiting on a completion in kthread() and will trigger hung
2220 	 * check if not woken up soon. As kick_pool() might not have waken it
2221 	 * up, wake it up explicitly once more.
2222 	 */
2223 	wake_up_process(worker->task);
2224 
2225 	raw_spin_unlock_irq(&pool->lock);
2226 
2227 	return worker;
2228 
2229 fail:
2230 	ida_free(&pool->worker_ida, id);
2231 	kfree(worker);
2232 	return NULL;
2233 }
2234 
2235 static void unbind_worker(struct worker *worker)
2236 {
2237 	lockdep_assert_held(&wq_pool_attach_mutex);
2238 
2239 	kthread_set_per_cpu(worker->task, -1);
2240 	if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2241 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2242 	else
2243 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2244 }
2245 
2246 static void wake_dying_workers(struct list_head *cull_list)
2247 {
2248 	struct worker *worker, *tmp;
2249 
2250 	list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2251 		list_del_init(&worker->entry);
2252 		unbind_worker(worker);
2253 		/*
2254 		 * If the worker was somehow already running, then it had to be
2255 		 * in pool->idle_list when set_worker_dying() happened or we
2256 		 * wouldn't have gotten here.
2257 		 *
2258 		 * Thus, the worker must either have observed the WORKER_DIE
2259 		 * flag, or have set its state to TASK_IDLE. Either way, the
2260 		 * below will be observed by the worker and is safe to do
2261 		 * outside of pool->lock.
2262 		 */
2263 		wake_up_process(worker->task);
2264 	}
2265 }
2266 
2267 /**
2268  * set_worker_dying - Tag a worker for destruction
2269  * @worker: worker to be destroyed
2270  * @list: transfer worker away from its pool->idle_list and into list
2271  *
2272  * Tag @worker for destruction and adjust @pool stats accordingly.  The worker
2273  * should be idle.
2274  *
2275  * CONTEXT:
2276  * raw_spin_lock_irq(pool->lock).
2277  */
2278 static void set_worker_dying(struct worker *worker, struct list_head *list)
2279 {
2280 	struct worker_pool *pool = worker->pool;
2281 
2282 	lockdep_assert_held(&pool->lock);
2283 	lockdep_assert_held(&wq_pool_attach_mutex);
2284 
2285 	/* sanity check frenzy */
2286 	if (WARN_ON(worker->current_work) ||
2287 	    WARN_ON(!list_empty(&worker->scheduled)) ||
2288 	    WARN_ON(!(worker->flags & WORKER_IDLE)))
2289 		return;
2290 
2291 	pool->nr_workers--;
2292 	pool->nr_idle--;
2293 
2294 	worker->flags |= WORKER_DIE;
2295 
2296 	list_move(&worker->entry, list);
2297 	list_move(&worker->node, &pool->dying_workers);
2298 }
2299 
2300 /**
2301  * idle_worker_timeout - check if some idle workers can now be deleted.
2302  * @t: The pool's idle_timer that just expired
2303  *
2304  * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2305  * worker_leave_idle(), as a worker flicking between idle and active while its
2306  * pool is at the too_many_workers() tipping point would cause too much timer
2307  * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2308  * it expire and re-evaluate things from there.
2309  */
2310 static void idle_worker_timeout(struct timer_list *t)
2311 {
2312 	struct worker_pool *pool = from_timer(pool, t, idle_timer);
2313 	bool do_cull = false;
2314 
2315 	if (work_pending(&pool->idle_cull_work))
2316 		return;
2317 
2318 	raw_spin_lock_irq(&pool->lock);
2319 
2320 	if (too_many_workers(pool)) {
2321 		struct worker *worker;
2322 		unsigned long expires;
2323 
2324 		/* idle_list is kept in LIFO order, check the last one */
2325 		worker = list_entry(pool->idle_list.prev, struct worker, entry);
2326 		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2327 		do_cull = !time_before(jiffies, expires);
2328 
2329 		if (!do_cull)
2330 			mod_timer(&pool->idle_timer, expires);
2331 	}
2332 	raw_spin_unlock_irq(&pool->lock);
2333 
2334 	if (do_cull)
2335 		queue_work(system_unbound_wq, &pool->idle_cull_work);
2336 }
2337 
2338 /**
2339  * idle_cull_fn - cull workers that have been idle for too long.
2340  * @work: the pool's work for handling these idle workers
2341  *
2342  * This goes through a pool's idle workers and gets rid of those that have been
2343  * idle for at least IDLE_WORKER_TIMEOUT seconds.
2344  *
2345  * We don't want to disturb isolated CPUs because of a pcpu kworker being
2346  * culled, so this also resets worker affinity. This requires a sleepable
2347  * context, hence the split between timer callback and work item.
2348  */
2349 static void idle_cull_fn(struct work_struct *work)
2350 {
2351 	struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2352 	LIST_HEAD(cull_list);
2353 
2354 	/*
2355 	 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2356 	 * cannot proceed beyong worker_detach_from_pool() in its self-destruct
2357 	 * path. This is required as a previously-preempted worker could run after
2358 	 * set_worker_dying() has happened but before wake_dying_workers() did.
2359 	 */
2360 	mutex_lock(&wq_pool_attach_mutex);
2361 	raw_spin_lock_irq(&pool->lock);
2362 
2363 	while (too_many_workers(pool)) {
2364 		struct worker *worker;
2365 		unsigned long expires;
2366 
2367 		worker = list_entry(pool->idle_list.prev, struct worker, entry);
2368 		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2369 
2370 		if (time_before(jiffies, expires)) {
2371 			mod_timer(&pool->idle_timer, expires);
2372 			break;
2373 		}
2374 
2375 		set_worker_dying(worker, &cull_list);
2376 	}
2377 
2378 	raw_spin_unlock_irq(&pool->lock);
2379 	wake_dying_workers(&cull_list);
2380 	mutex_unlock(&wq_pool_attach_mutex);
2381 }
2382 
2383 static void send_mayday(struct work_struct *work)
2384 {
2385 	struct pool_workqueue *pwq = get_work_pwq(work);
2386 	struct workqueue_struct *wq = pwq->wq;
2387 
2388 	lockdep_assert_held(&wq_mayday_lock);
2389 
2390 	if (!wq->rescuer)
2391 		return;
2392 
2393 	/* mayday mayday mayday */
2394 	if (list_empty(&pwq->mayday_node)) {
2395 		/*
2396 		 * If @pwq is for an unbound wq, its base ref may be put at
2397 		 * any time due to an attribute change.  Pin @pwq until the
2398 		 * rescuer is done with it.
2399 		 */
2400 		get_pwq(pwq);
2401 		list_add_tail(&pwq->mayday_node, &wq->maydays);
2402 		wake_up_process(wq->rescuer->task);
2403 		pwq->stats[PWQ_STAT_MAYDAY]++;
2404 	}
2405 }
2406 
2407 static void pool_mayday_timeout(struct timer_list *t)
2408 {
2409 	struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2410 	struct work_struct *work;
2411 
2412 	raw_spin_lock_irq(&pool->lock);
2413 	raw_spin_lock(&wq_mayday_lock);		/* for wq->maydays */
2414 
2415 	if (need_to_create_worker(pool)) {
2416 		/*
2417 		 * We've been trying to create a new worker but
2418 		 * haven't been successful.  We might be hitting an
2419 		 * allocation deadlock.  Send distress signals to
2420 		 * rescuers.
2421 		 */
2422 		list_for_each_entry(work, &pool->worklist, entry)
2423 			send_mayday(work);
2424 	}
2425 
2426 	raw_spin_unlock(&wq_mayday_lock);
2427 	raw_spin_unlock_irq(&pool->lock);
2428 
2429 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2430 }
2431 
2432 /**
2433  * maybe_create_worker - create a new worker if necessary
2434  * @pool: pool to create a new worker for
2435  *
2436  * Create a new worker for @pool if necessary.  @pool is guaranteed to
2437  * have at least one idle worker on return from this function.  If
2438  * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2439  * sent to all rescuers with works scheduled on @pool to resolve
2440  * possible allocation deadlock.
2441  *
2442  * On return, need_to_create_worker() is guaranteed to be %false and
2443  * may_start_working() %true.
2444  *
2445  * LOCKING:
2446  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2447  * multiple times.  Does GFP_KERNEL allocations.  Called only from
2448  * manager.
2449  */
2450 static void maybe_create_worker(struct worker_pool *pool)
2451 __releases(&pool->lock)
2452 __acquires(&pool->lock)
2453 {
2454 restart:
2455 	raw_spin_unlock_irq(&pool->lock);
2456 
2457 	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2458 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2459 
2460 	while (true) {
2461 		if (create_worker(pool) || !need_to_create_worker(pool))
2462 			break;
2463 
2464 		schedule_timeout_interruptible(CREATE_COOLDOWN);
2465 
2466 		if (!need_to_create_worker(pool))
2467 			break;
2468 	}
2469 
2470 	del_timer_sync(&pool->mayday_timer);
2471 	raw_spin_lock_irq(&pool->lock);
2472 	/*
2473 	 * This is necessary even after a new worker was just successfully
2474 	 * created as @pool->lock was dropped and the new worker might have
2475 	 * already become busy.
2476 	 */
2477 	if (need_to_create_worker(pool))
2478 		goto restart;
2479 }
2480 
2481 /**
2482  * manage_workers - manage worker pool
2483  * @worker: self
2484  *
2485  * Assume the manager role and manage the worker pool @worker belongs
2486  * to.  At any given time, there can be only zero or one manager per
2487  * pool.  The exclusion is handled automatically by this function.
2488  *
2489  * The caller can safely start processing works on false return.  On
2490  * true return, it's guaranteed that need_to_create_worker() is false
2491  * and may_start_working() is true.
2492  *
2493  * CONTEXT:
2494  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2495  * multiple times.  Does GFP_KERNEL allocations.
2496  *
2497  * Return:
2498  * %false if the pool doesn't need management and the caller can safely
2499  * start processing works, %true if management function was performed and
2500  * the conditions that the caller verified before calling the function may
2501  * no longer be true.
2502  */
2503 static bool manage_workers(struct worker *worker)
2504 {
2505 	struct worker_pool *pool = worker->pool;
2506 
2507 	if (pool->flags & POOL_MANAGER_ACTIVE)
2508 		return false;
2509 
2510 	pool->flags |= POOL_MANAGER_ACTIVE;
2511 	pool->manager = worker;
2512 
2513 	maybe_create_worker(pool);
2514 
2515 	pool->manager = NULL;
2516 	pool->flags &= ~POOL_MANAGER_ACTIVE;
2517 	rcuwait_wake_up(&manager_wait);
2518 	return true;
2519 }
2520 
2521 /**
2522  * process_one_work - process single work
2523  * @worker: self
2524  * @work: work to process
2525  *
2526  * Process @work.  This function contains all the logics necessary to
2527  * process a single work including synchronization against and
2528  * interaction with other workers on the same cpu, queueing and
2529  * flushing.  As long as context requirement is met, any worker can
2530  * call this function to process a work.
2531  *
2532  * CONTEXT:
2533  * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2534  */
2535 static void process_one_work(struct worker *worker, struct work_struct *work)
2536 __releases(&pool->lock)
2537 __acquires(&pool->lock)
2538 {
2539 	struct pool_workqueue *pwq = get_work_pwq(work);
2540 	struct worker_pool *pool = worker->pool;
2541 	unsigned long work_data;
2542 #ifdef CONFIG_LOCKDEP
2543 	/*
2544 	 * It is permissible to free the struct work_struct from
2545 	 * inside the function that is called from it, this we need to
2546 	 * take into account for lockdep too.  To avoid bogus "held
2547 	 * lock freed" warnings as well as problems when looking into
2548 	 * work->lockdep_map, make a copy and use that here.
2549 	 */
2550 	struct lockdep_map lockdep_map;
2551 
2552 	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2553 #endif
2554 	/* ensure we're on the correct CPU */
2555 	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2556 		     raw_smp_processor_id() != pool->cpu);
2557 
2558 	/* claim and dequeue */
2559 	debug_work_deactivate(work);
2560 	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2561 	worker->current_work = work;
2562 	worker->current_func = work->func;
2563 	worker->current_pwq = pwq;
2564 	worker->current_at = worker->task->se.sum_exec_runtime;
2565 	work_data = *work_data_bits(work);
2566 	worker->current_color = get_work_color(work_data);
2567 
2568 	/*
2569 	 * Record wq name for cmdline and debug reporting, may get
2570 	 * overridden through set_worker_desc().
2571 	 */
2572 	strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2573 
2574 	list_del_init(&work->entry);
2575 
2576 	/*
2577 	 * CPU intensive works don't participate in concurrency management.
2578 	 * They're the scheduler's responsibility.  This takes @worker out
2579 	 * of concurrency management and the next code block will chain
2580 	 * execution of the pending work items.
2581 	 */
2582 	if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
2583 		worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2584 
2585 	/*
2586 	 * Kick @pool if necessary. It's always noop for per-cpu worker pools
2587 	 * since nr_running would always be >= 1 at this point. This is used to
2588 	 * chain execution of the pending work items for WORKER_NOT_RUNNING
2589 	 * workers such as the UNBOUND and CPU_INTENSIVE ones.
2590 	 */
2591 	kick_pool(pool);
2592 
2593 	/*
2594 	 * Record the last pool and clear PENDING which should be the last
2595 	 * update to @work.  Also, do this inside @pool->lock so that
2596 	 * PENDING and queued state changes happen together while IRQ is
2597 	 * disabled.
2598 	 */
2599 	set_work_pool_and_clear_pending(work, pool->id);
2600 
2601 	raw_spin_unlock_irq(&pool->lock);
2602 
2603 	lock_map_acquire(&pwq->wq->lockdep_map);
2604 	lock_map_acquire(&lockdep_map);
2605 	/*
2606 	 * Strictly speaking we should mark the invariant state without holding
2607 	 * any locks, that is, before these two lock_map_acquire()'s.
2608 	 *
2609 	 * However, that would result in:
2610 	 *
2611 	 *   A(W1)
2612 	 *   WFC(C)
2613 	 *		A(W1)
2614 	 *		C(C)
2615 	 *
2616 	 * Which would create W1->C->W1 dependencies, even though there is no
2617 	 * actual deadlock possible. There are two solutions, using a
2618 	 * read-recursive acquire on the work(queue) 'locks', but this will then
2619 	 * hit the lockdep limitation on recursive locks, or simply discard
2620 	 * these locks.
2621 	 *
2622 	 * AFAICT there is no possible deadlock scenario between the
2623 	 * flush_work() and complete() primitives (except for single-threaded
2624 	 * workqueues), so hiding them isn't a problem.
2625 	 */
2626 	lockdep_invariant_state(true);
2627 	pwq->stats[PWQ_STAT_STARTED]++;
2628 	trace_workqueue_execute_start(work);
2629 	worker->current_func(work);
2630 	/*
2631 	 * While we must be careful to not use "work" after this, the trace
2632 	 * point will only record its address.
2633 	 */
2634 	trace_workqueue_execute_end(work, worker->current_func);
2635 	pwq->stats[PWQ_STAT_COMPLETED]++;
2636 	lock_map_release(&lockdep_map);
2637 	lock_map_release(&pwq->wq->lockdep_map);
2638 
2639 	if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2640 		pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2641 		       "     last function: %ps\n",
2642 		       current->comm, preempt_count(), task_pid_nr(current),
2643 		       worker->current_func);
2644 		debug_show_held_locks(current);
2645 		dump_stack();
2646 	}
2647 
2648 	/*
2649 	 * The following prevents a kworker from hogging CPU on !PREEMPTION
2650 	 * kernels, where a requeueing work item waiting for something to
2651 	 * happen could deadlock with stop_machine as such work item could
2652 	 * indefinitely requeue itself while all other CPUs are trapped in
2653 	 * stop_machine. At the same time, report a quiescent RCU state so
2654 	 * the same condition doesn't freeze RCU.
2655 	 */
2656 	cond_resched();
2657 
2658 	raw_spin_lock_irq(&pool->lock);
2659 
2660 	/*
2661 	 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
2662 	 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
2663 	 * wq_cpu_intensive_thresh_us. Clear it.
2664 	 */
2665 	worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2666 
2667 	/* tag the worker for identification in schedule() */
2668 	worker->last_func = worker->current_func;
2669 
2670 	/* we're done with it, release */
2671 	hash_del(&worker->hentry);
2672 	worker->current_work = NULL;
2673 	worker->current_func = NULL;
2674 	worker->current_pwq = NULL;
2675 	worker->current_color = INT_MAX;
2676 	pwq_dec_nr_in_flight(pwq, work_data);
2677 }
2678 
2679 /**
2680  * process_scheduled_works - process scheduled works
2681  * @worker: self
2682  *
2683  * Process all scheduled works.  Please note that the scheduled list
2684  * may change while processing a work, so this function repeatedly
2685  * fetches a work from the top and executes it.
2686  *
2687  * CONTEXT:
2688  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2689  * multiple times.
2690  */
2691 static void process_scheduled_works(struct worker *worker)
2692 {
2693 	struct work_struct *work;
2694 	bool first = true;
2695 
2696 	while ((work = list_first_entry_or_null(&worker->scheduled,
2697 						struct work_struct, entry))) {
2698 		if (first) {
2699 			worker->pool->watchdog_ts = jiffies;
2700 			first = false;
2701 		}
2702 		process_one_work(worker, work);
2703 	}
2704 }
2705 
2706 static void set_pf_worker(bool val)
2707 {
2708 	mutex_lock(&wq_pool_attach_mutex);
2709 	if (val)
2710 		current->flags |= PF_WQ_WORKER;
2711 	else
2712 		current->flags &= ~PF_WQ_WORKER;
2713 	mutex_unlock(&wq_pool_attach_mutex);
2714 }
2715 
2716 /**
2717  * worker_thread - the worker thread function
2718  * @__worker: self
2719  *
2720  * The worker thread function.  All workers belong to a worker_pool -
2721  * either a per-cpu one or dynamic unbound one.  These workers process all
2722  * work items regardless of their specific target workqueue.  The only
2723  * exception is work items which belong to workqueues with a rescuer which
2724  * will be explained in rescuer_thread().
2725  *
2726  * Return: 0
2727  */
2728 static int worker_thread(void *__worker)
2729 {
2730 	struct worker *worker = __worker;
2731 	struct worker_pool *pool = worker->pool;
2732 
2733 	/* tell the scheduler that this is a workqueue worker */
2734 	set_pf_worker(true);
2735 woke_up:
2736 	raw_spin_lock_irq(&pool->lock);
2737 
2738 	/* am I supposed to die? */
2739 	if (unlikely(worker->flags & WORKER_DIE)) {
2740 		raw_spin_unlock_irq(&pool->lock);
2741 		set_pf_worker(false);
2742 
2743 		set_task_comm(worker->task, "kworker/dying");
2744 		ida_free(&pool->worker_ida, worker->id);
2745 		worker_detach_from_pool(worker);
2746 		WARN_ON_ONCE(!list_empty(&worker->entry));
2747 		kfree(worker);
2748 		return 0;
2749 	}
2750 
2751 	worker_leave_idle(worker);
2752 recheck:
2753 	/* no more worker necessary? */
2754 	if (!need_more_worker(pool))
2755 		goto sleep;
2756 
2757 	/* do we need to manage? */
2758 	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2759 		goto recheck;
2760 
2761 	/*
2762 	 * ->scheduled list can only be filled while a worker is
2763 	 * preparing to process a work or actually processing it.
2764 	 * Make sure nobody diddled with it while I was sleeping.
2765 	 */
2766 	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2767 
2768 	/*
2769 	 * Finish PREP stage.  We're guaranteed to have at least one idle
2770 	 * worker or that someone else has already assumed the manager
2771 	 * role.  This is where @worker starts participating in concurrency
2772 	 * management if applicable and concurrency management is restored
2773 	 * after being rebound.  See rebind_workers() for details.
2774 	 */
2775 	worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2776 
2777 	do {
2778 		struct work_struct *work =
2779 			list_first_entry(&pool->worklist,
2780 					 struct work_struct, entry);
2781 
2782 		if (assign_work(work, worker, NULL))
2783 			process_scheduled_works(worker);
2784 	} while (keep_working(pool));
2785 
2786 	worker_set_flags(worker, WORKER_PREP);
2787 sleep:
2788 	/*
2789 	 * pool->lock is held and there's no work to process and no need to
2790 	 * manage, sleep.  Workers are woken up only while holding
2791 	 * pool->lock or from local cpu, so setting the current state
2792 	 * before releasing pool->lock is enough to prevent losing any
2793 	 * event.
2794 	 */
2795 	worker_enter_idle(worker);
2796 	__set_current_state(TASK_IDLE);
2797 	raw_spin_unlock_irq(&pool->lock);
2798 	schedule();
2799 	goto woke_up;
2800 }
2801 
2802 /**
2803  * rescuer_thread - the rescuer thread function
2804  * @__rescuer: self
2805  *
2806  * Workqueue rescuer thread function.  There's one rescuer for each
2807  * workqueue which has WQ_MEM_RECLAIM set.
2808  *
2809  * Regular work processing on a pool may block trying to create a new
2810  * worker which uses GFP_KERNEL allocation which has slight chance of
2811  * developing into deadlock if some works currently on the same queue
2812  * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2813  * the problem rescuer solves.
2814  *
2815  * When such condition is possible, the pool summons rescuers of all
2816  * workqueues which have works queued on the pool and let them process
2817  * those works so that forward progress can be guaranteed.
2818  *
2819  * This should happen rarely.
2820  *
2821  * Return: 0
2822  */
2823 static int rescuer_thread(void *__rescuer)
2824 {
2825 	struct worker *rescuer = __rescuer;
2826 	struct workqueue_struct *wq = rescuer->rescue_wq;
2827 	bool should_stop;
2828 
2829 	set_user_nice(current, RESCUER_NICE_LEVEL);
2830 
2831 	/*
2832 	 * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
2833 	 * doesn't participate in concurrency management.
2834 	 */
2835 	set_pf_worker(true);
2836 repeat:
2837 	set_current_state(TASK_IDLE);
2838 
2839 	/*
2840 	 * By the time the rescuer is requested to stop, the workqueue
2841 	 * shouldn't have any work pending, but @wq->maydays may still have
2842 	 * pwq(s) queued.  This can happen by non-rescuer workers consuming
2843 	 * all the work items before the rescuer got to them.  Go through
2844 	 * @wq->maydays processing before acting on should_stop so that the
2845 	 * list is always empty on exit.
2846 	 */
2847 	should_stop = kthread_should_stop();
2848 
2849 	/* see whether any pwq is asking for help */
2850 	raw_spin_lock_irq(&wq_mayday_lock);
2851 
2852 	while (!list_empty(&wq->maydays)) {
2853 		struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2854 					struct pool_workqueue, mayday_node);
2855 		struct worker_pool *pool = pwq->pool;
2856 		struct work_struct *work, *n;
2857 
2858 		__set_current_state(TASK_RUNNING);
2859 		list_del_init(&pwq->mayday_node);
2860 
2861 		raw_spin_unlock_irq(&wq_mayday_lock);
2862 
2863 		worker_attach_to_pool(rescuer, pool);
2864 
2865 		raw_spin_lock_irq(&pool->lock);
2866 
2867 		/*
2868 		 * Slurp in all works issued via this workqueue and
2869 		 * process'em.
2870 		 */
2871 		WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2872 		list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2873 			if (get_work_pwq(work) == pwq &&
2874 			    assign_work(work, rescuer, &n))
2875 				pwq->stats[PWQ_STAT_RESCUED]++;
2876 		}
2877 
2878 		if (!list_empty(&rescuer->scheduled)) {
2879 			process_scheduled_works(rescuer);
2880 
2881 			/*
2882 			 * The above execution of rescued work items could
2883 			 * have created more to rescue through
2884 			 * pwq_activate_first_inactive() or chained
2885 			 * queueing.  Let's put @pwq back on mayday list so
2886 			 * that such back-to-back work items, which may be
2887 			 * being used to relieve memory pressure, don't
2888 			 * incur MAYDAY_INTERVAL delay inbetween.
2889 			 */
2890 			if (pwq->nr_active && need_to_create_worker(pool)) {
2891 				raw_spin_lock(&wq_mayday_lock);
2892 				/*
2893 				 * Queue iff we aren't racing destruction
2894 				 * and somebody else hasn't queued it already.
2895 				 */
2896 				if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2897 					get_pwq(pwq);
2898 					list_add_tail(&pwq->mayday_node, &wq->maydays);
2899 				}
2900 				raw_spin_unlock(&wq_mayday_lock);
2901 			}
2902 		}
2903 
2904 		/*
2905 		 * Put the reference grabbed by send_mayday().  @pool won't
2906 		 * go away while we're still attached to it.
2907 		 */
2908 		put_pwq(pwq);
2909 
2910 		/*
2911 		 * Leave this pool. Notify regular workers; otherwise, we end up
2912 		 * with 0 concurrency and stalling the execution.
2913 		 */
2914 		kick_pool(pool);
2915 
2916 		raw_spin_unlock_irq(&pool->lock);
2917 
2918 		worker_detach_from_pool(rescuer);
2919 
2920 		raw_spin_lock_irq(&wq_mayday_lock);
2921 	}
2922 
2923 	raw_spin_unlock_irq(&wq_mayday_lock);
2924 
2925 	if (should_stop) {
2926 		__set_current_state(TASK_RUNNING);
2927 		set_pf_worker(false);
2928 		return 0;
2929 	}
2930 
2931 	/* rescuers should never participate in concurrency management */
2932 	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2933 	schedule();
2934 	goto repeat;
2935 }
2936 
2937 /**
2938  * check_flush_dependency - check for flush dependency sanity
2939  * @target_wq: workqueue being flushed
2940  * @target_work: work item being flushed (NULL for workqueue flushes)
2941  *
2942  * %current is trying to flush the whole @target_wq or @target_work on it.
2943  * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2944  * reclaiming memory or running on a workqueue which doesn't have
2945  * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2946  * a deadlock.
2947  */
2948 static void check_flush_dependency(struct workqueue_struct *target_wq,
2949 				   struct work_struct *target_work)
2950 {
2951 	work_func_t target_func = target_work ? target_work->func : NULL;
2952 	struct worker *worker;
2953 
2954 	if (target_wq->flags & WQ_MEM_RECLAIM)
2955 		return;
2956 
2957 	worker = current_wq_worker();
2958 
2959 	WARN_ONCE(current->flags & PF_MEMALLOC,
2960 		  "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2961 		  current->pid, current->comm, target_wq->name, target_func);
2962 	WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2963 			      (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2964 		  "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2965 		  worker->current_pwq->wq->name, worker->current_func,
2966 		  target_wq->name, target_func);
2967 }
2968 
2969 struct wq_barrier {
2970 	struct work_struct	work;
2971 	struct completion	done;
2972 	struct task_struct	*task;	/* purely informational */
2973 };
2974 
2975 static void wq_barrier_func(struct work_struct *work)
2976 {
2977 	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2978 	complete(&barr->done);
2979 }
2980 
2981 /**
2982  * insert_wq_barrier - insert a barrier work
2983  * @pwq: pwq to insert barrier into
2984  * @barr: wq_barrier to insert
2985  * @target: target work to attach @barr to
2986  * @worker: worker currently executing @target, NULL if @target is not executing
2987  *
2988  * @barr is linked to @target such that @barr is completed only after
2989  * @target finishes execution.  Please note that the ordering
2990  * guarantee is observed only with respect to @target and on the local
2991  * cpu.
2992  *
2993  * Currently, a queued barrier can't be canceled.  This is because
2994  * try_to_grab_pending() can't determine whether the work to be
2995  * grabbed is at the head of the queue and thus can't clear LINKED
2996  * flag of the previous work while there must be a valid next work
2997  * after a work with LINKED flag set.
2998  *
2999  * Note that when @worker is non-NULL, @target may be modified
3000  * underneath us, so we can't reliably determine pwq from @target.
3001  *
3002  * CONTEXT:
3003  * raw_spin_lock_irq(pool->lock).
3004  */
3005 static void insert_wq_barrier(struct pool_workqueue *pwq,
3006 			      struct wq_barrier *barr,
3007 			      struct work_struct *target, struct worker *worker)
3008 {
3009 	unsigned int work_flags = 0;
3010 	unsigned int work_color;
3011 	struct list_head *head;
3012 
3013 	/*
3014 	 * debugobject calls are safe here even with pool->lock locked
3015 	 * as we know for sure that this will not trigger any of the
3016 	 * checks and call back into the fixup functions where we
3017 	 * might deadlock.
3018 	 */
3019 	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
3020 	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3021 
3022 	init_completion_map(&barr->done, &target->lockdep_map);
3023 
3024 	barr->task = current;
3025 
3026 	/* The barrier work item does not participate in pwq->nr_active. */
3027 	work_flags |= WORK_STRUCT_INACTIVE;
3028 
3029 	/*
3030 	 * If @target is currently being executed, schedule the
3031 	 * barrier to the worker; otherwise, put it after @target.
3032 	 */
3033 	if (worker) {
3034 		head = worker->scheduled.next;
3035 		work_color = worker->current_color;
3036 	} else {
3037 		unsigned long *bits = work_data_bits(target);
3038 
3039 		head = target->entry.next;
3040 		/* there can already be other linked works, inherit and set */
3041 		work_flags |= *bits & WORK_STRUCT_LINKED;
3042 		work_color = get_work_color(*bits);
3043 		__set_bit(WORK_STRUCT_LINKED_BIT, bits);
3044 	}
3045 
3046 	pwq->nr_in_flight[work_color]++;
3047 	work_flags |= work_color_to_flags(work_color);
3048 
3049 	insert_work(pwq, &barr->work, head, work_flags);
3050 }
3051 
3052 /**
3053  * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3054  * @wq: workqueue being flushed
3055  * @flush_color: new flush color, < 0 for no-op
3056  * @work_color: new work color, < 0 for no-op
3057  *
3058  * Prepare pwqs for workqueue flushing.
3059  *
3060  * If @flush_color is non-negative, flush_color on all pwqs should be
3061  * -1.  If no pwq has in-flight commands at the specified color, all
3062  * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
3063  * has in flight commands, its pwq->flush_color is set to
3064  * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3065  * wakeup logic is armed and %true is returned.
3066  *
3067  * The caller should have initialized @wq->first_flusher prior to
3068  * calling this function with non-negative @flush_color.  If
3069  * @flush_color is negative, no flush color update is done and %false
3070  * is returned.
3071  *
3072  * If @work_color is non-negative, all pwqs should have the same
3073  * work_color which is previous to @work_color and all will be
3074  * advanced to @work_color.
3075  *
3076  * CONTEXT:
3077  * mutex_lock(wq->mutex).
3078  *
3079  * Return:
3080  * %true if @flush_color >= 0 and there's something to flush.  %false
3081  * otherwise.
3082  */
3083 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3084 				      int flush_color, int work_color)
3085 {
3086 	bool wait = false;
3087 	struct pool_workqueue *pwq;
3088 
3089 	if (flush_color >= 0) {
3090 		WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3091 		atomic_set(&wq->nr_pwqs_to_flush, 1);
3092 	}
3093 
3094 	for_each_pwq(pwq, wq) {
3095 		struct worker_pool *pool = pwq->pool;
3096 
3097 		raw_spin_lock_irq(&pool->lock);
3098 
3099 		if (flush_color >= 0) {
3100 			WARN_ON_ONCE(pwq->flush_color != -1);
3101 
3102 			if (pwq->nr_in_flight[flush_color]) {
3103 				pwq->flush_color = flush_color;
3104 				atomic_inc(&wq->nr_pwqs_to_flush);
3105 				wait = true;
3106 			}
3107 		}
3108 
3109 		if (work_color >= 0) {
3110 			WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3111 			pwq->work_color = work_color;
3112 		}
3113 
3114 		raw_spin_unlock_irq(&pool->lock);
3115 	}
3116 
3117 	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3118 		complete(&wq->first_flusher->done);
3119 
3120 	return wait;
3121 }
3122 
3123 /**
3124  * __flush_workqueue - ensure that any scheduled work has run to completion.
3125  * @wq: workqueue to flush
3126  *
3127  * This function sleeps until all work items which were queued on entry
3128  * have finished execution, but it is not livelocked by new incoming ones.
3129  */
3130 void __flush_workqueue(struct workqueue_struct *wq)
3131 {
3132 	struct wq_flusher this_flusher = {
3133 		.list = LIST_HEAD_INIT(this_flusher.list),
3134 		.flush_color = -1,
3135 		.done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3136 	};
3137 	int next_color;
3138 
3139 	if (WARN_ON(!wq_online))
3140 		return;
3141 
3142 	lock_map_acquire(&wq->lockdep_map);
3143 	lock_map_release(&wq->lockdep_map);
3144 
3145 	mutex_lock(&wq->mutex);
3146 
3147 	/*
3148 	 * Start-to-wait phase
3149 	 */
3150 	next_color = work_next_color(wq->work_color);
3151 
3152 	if (next_color != wq->flush_color) {
3153 		/*
3154 		 * Color space is not full.  The current work_color
3155 		 * becomes our flush_color and work_color is advanced
3156 		 * by one.
3157 		 */
3158 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3159 		this_flusher.flush_color = wq->work_color;
3160 		wq->work_color = next_color;
3161 
3162 		if (!wq->first_flusher) {
3163 			/* no flush in progress, become the first flusher */
3164 			WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3165 
3166 			wq->first_flusher = &this_flusher;
3167 
3168 			if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3169 						       wq->work_color)) {
3170 				/* nothing to flush, done */
3171 				wq->flush_color = next_color;
3172 				wq->first_flusher = NULL;
3173 				goto out_unlock;
3174 			}
3175 		} else {
3176 			/* wait in queue */
3177 			WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3178 			list_add_tail(&this_flusher.list, &wq->flusher_queue);
3179 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3180 		}
3181 	} else {
3182 		/*
3183 		 * Oops, color space is full, wait on overflow queue.
3184 		 * The next flush completion will assign us
3185 		 * flush_color and transfer to flusher_queue.
3186 		 */
3187 		list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3188 	}
3189 
3190 	check_flush_dependency(wq, NULL);
3191 
3192 	mutex_unlock(&wq->mutex);
3193 
3194 	wait_for_completion(&this_flusher.done);
3195 
3196 	/*
3197 	 * Wake-up-and-cascade phase
3198 	 *
3199 	 * First flushers are responsible for cascading flushes and
3200 	 * handling overflow.  Non-first flushers can simply return.
3201 	 */
3202 	if (READ_ONCE(wq->first_flusher) != &this_flusher)
3203 		return;
3204 
3205 	mutex_lock(&wq->mutex);
3206 
3207 	/* we might have raced, check again with mutex held */
3208 	if (wq->first_flusher != &this_flusher)
3209 		goto out_unlock;
3210 
3211 	WRITE_ONCE(wq->first_flusher, NULL);
3212 
3213 	WARN_ON_ONCE(!list_empty(&this_flusher.list));
3214 	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3215 
3216 	while (true) {
3217 		struct wq_flusher *next, *tmp;
3218 
3219 		/* complete all the flushers sharing the current flush color */
3220 		list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3221 			if (next->flush_color != wq->flush_color)
3222 				break;
3223 			list_del_init(&next->list);
3224 			complete(&next->done);
3225 		}
3226 
3227 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
3228 			     wq->flush_color != work_next_color(wq->work_color));
3229 
3230 		/* this flush_color is finished, advance by one */
3231 		wq->flush_color = work_next_color(wq->flush_color);
3232 
3233 		/* one color has been freed, handle overflow queue */
3234 		if (!list_empty(&wq->flusher_overflow)) {
3235 			/*
3236 			 * Assign the same color to all overflowed
3237 			 * flushers, advance work_color and append to
3238 			 * flusher_queue.  This is the start-to-wait
3239 			 * phase for these overflowed flushers.
3240 			 */
3241 			list_for_each_entry(tmp, &wq->flusher_overflow, list)
3242 				tmp->flush_color = wq->work_color;
3243 
3244 			wq->work_color = work_next_color(wq->work_color);
3245 
3246 			list_splice_tail_init(&wq->flusher_overflow,
3247 					      &wq->flusher_queue);
3248 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3249 		}
3250 
3251 		if (list_empty(&wq->flusher_queue)) {
3252 			WARN_ON_ONCE(wq->flush_color != wq->work_color);
3253 			break;
3254 		}
3255 
3256 		/*
3257 		 * Need to flush more colors.  Make the next flusher
3258 		 * the new first flusher and arm pwqs.
3259 		 */
3260 		WARN_ON_ONCE(wq->flush_color == wq->work_color);
3261 		WARN_ON_ONCE(wq->flush_color != next->flush_color);
3262 
3263 		list_del_init(&next->list);
3264 		wq->first_flusher = next;
3265 
3266 		if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
3267 			break;
3268 
3269 		/*
3270 		 * Meh... this color is already done, clear first
3271 		 * flusher and repeat cascading.
3272 		 */
3273 		wq->first_flusher = NULL;
3274 	}
3275 
3276 out_unlock:
3277 	mutex_unlock(&wq->mutex);
3278 }
3279 EXPORT_SYMBOL(__flush_workqueue);
3280 
3281 /**
3282  * drain_workqueue - drain a workqueue
3283  * @wq: workqueue to drain
3284  *
3285  * Wait until the workqueue becomes empty.  While draining is in progress,
3286  * only chain queueing is allowed.  IOW, only currently pending or running
3287  * work items on @wq can queue further work items on it.  @wq is flushed
3288  * repeatedly until it becomes empty.  The number of flushing is determined
3289  * by the depth of chaining and should be relatively short.  Whine if it
3290  * takes too long.
3291  */
3292 void drain_workqueue(struct workqueue_struct *wq)
3293 {
3294 	unsigned int flush_cnt = 0;
3295 	struct pool_workqueue *pwq;
3296 
3297 	/*
3298 	 * __queue_work() needs to test whether there are drainers, is much
3299 	 * hotter than drain_workqueue() and already looks at @wq->flags.
3300 	 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
3301 	 */
3302 	mutex_lock(&wq->mutex);
3303 	if (!wq->nr_drainers++)
3304 		wq->flags |= __WQ_DRAINING;
3305 	mutex_unlock(&wq->mutex);
3306 reflush:
3307 	__flush_workqueue(wq);
3308 
3309 	mutex_lock(&wq->mutex);
3310 
3311 	for_each_pwq(pwq, wq) {
3312 		bool drained;
3313 
3314 		raw_spin_lock_irq(&pwq->pool->lock);
3315 		drained = !pwq->nr_active && list_empty(&pwq->inactive_works);
3316 		raw_spin_unlock_irq(&pwq->pool->lock);
3317 
3318 		if (drained)
3319 			continue;
3320 
3321 		if (++flush_cnt == 10 ||
3322 		    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3323 			pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
3324 				wq->name, __func__, flush_cnt);
3325 
3326 		mutex_unlock(&wq->mutex);
3327 		goto reflush;
3328 	}
3329 
3330 	if (!--wq->nr_drainers)
3331 		wq->flags &= ~__WQ_DRAINING;
3332 	mutex_unlock(&wq->mutex);
3333 }
3334 EXPORT_SYMBOL_GPL(drain_workqueue);
3335 
3336 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
3337 			     bool from_cancel)
3338 {
3339 	struct worker *worker = NULL;
3340 	struct worker_pool *pool;
3341 	struct pool_workqueue *pwq;
3342 
3343 	might_sleep();
3344 
3345 	rcu_read_lock();
3346 	pool = get_work_pool(work);
3347 	if (!pool) {
3348 		rcu_read_unlock();
3349 		return false;
3350 	}
3351 
3352 	raw_spin_lock_irq(&pool->lock);
3353 	/* see the comment in try_to_grab_pending() with the same code */
3354 	pwq = get_work_pwq(work);
3355 	if (pwq) {
3356 		if (unlikely(pwq->pool != pool))
3357 			goto already_gone;
3358 	} else {
3359 		worker = find_worker_executing_work(pool, work);
3360 		if (!worker)
3361 			goto already_gone;
3362 		pwq = worker->current_pwq;
3363 	}
3364 
3365 	check_flush_dependency(pwq->wq, work);
3366 
3367 	insert_wq_barrier(pwq, barr, work, worker);
3368 	raw_spin_unlock_irq(&pool->lock);
3369 
3370 	/*
3371 	 * Force a lock recursion deadlock when using flush_work() inside a
3372 	 * single-threaded or rescuer equipped workqueue.
3373 	 *
3374 	 * For single threaded workqueues the deadlock happens when the work
3375 	 * is after the work issuing the flush_work(). For rescuer equipped
3376 	 * workqueues the deadlock happens when the rescuer stalls, blocking
3377 	 * forward progress.
3378 	 */
3379 	if (!from_cancel &&
3380 	    (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3381 		lock_map_acquire(&pwq->wq->lockdep_map);
3382 		lock_map_release(&pwq->wq->lockdep_map);
3383 	}
3384 	rcu_read_unlock();
3385 	return true;
3386 already_gone:
3387 	raw_spin_unlock_irq(&pool->lock);
3388 	rcu_read_unlock();
3389 	return false;
3390 }
3391 
3392 static bool __flush_work(struct work_struct *work, bool from_cancel)
3393 {
3394 	struct wq_barrier barr;
3395 
3396 	if (WARN_ON(!wq_online))
3397 		return false;
3398 
3399 	if (WARN_ON(!work->func))
3400 		return false;
3401 
3402 	lock_map_acquire(&work->lockdep_map);
3403 	lock_map_release(&work->lockdep_map);
3404 
3405 	if (start_flush_work(work, &barr, from_cancel)) {
3406 		wait_for_completion(&barr.done);
3407 		destroy_work_on_stack(&barr.work);
3408 		return true;
3409 	} else {
3410 		return false;
3411 	}
3412 }
3413 
3414 /**
3415  * flush_work - wait for a work to finish executing the last queueing instance
3416  * @work: the work to flush
3417  *
3418  * Wait until @work has finished execution.  @work is guaranteed to be idle
3419  * on return if it hasn't been requeued since flush started.
3420  *
3421  * Return:
3422  * %true if flush_work() waited for the work to finish execution,
3423  * %false if it was already idle.
3424  */
3425 bool flush_work(struct work_struct *work)
3426 {
3427 	return __flush_work(work, false);
3428 }
3429 EXPORT_SYMBOL_GPL(flush_work);
3430 
3431 struct cwt_wait {
3432 	wait_queue_entry_t		wait;
3433 	struct work_struct	*work;
3434 };
3435 
3436 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3437 {
3438 	struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3439 
3440 	if (cwait->work != key)
3441 		return 0;
3442 	return autoremove_wake_function(wait, mode, sync, key);
3443 }
3444 
3445 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3446 {
3447 	static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3448 	unsigned long flags;
3449 	int ret;
3450 
3451 	do {
3452 		ret = try_to_grab_pending(work, is_dwork, &flags);
3453 		/*
3454 		 * If someone else is already canceling, wait for it to
3455 		 * finish.  flush_work() doesn't work for PREEMPT_NONE
3456 		 * because we may get scheduled between @work's completion
3457 		 * and the other canceling task resuming and clearing
3458 		 * CANCELING - flush_work() will return false immediately
3459 		 * as @work is no longer busy, try_to_grab_pending() will
3460 		 * return -ENOENT as @work is still being canceled and the
3461 		 * other canceling task won't be able to clear CANCELING as
3462 		 * we're hogging the CPU.
3463 		 *
3464 		 * Let's wait for completion using a waitqueue.  As this
3465 		 * may lead to the thundering herd problem, use a custom
3466 		 * wake function which matches @work along with exclusive
3467 		 * wait and wakeup.
3468 		 */
3469 		if (unlikely(ret == -ENOENT)) {
3470 			struct cwt_wait cwait;
3471 
3472 			init_wait(&cwait.wait);
3473 			cwait.wait.func = cwt_wakefn;
3474 			cwait.work = work;
3475 
3476 			prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3477 						  TASK_UNINTERRUPTIBLE);
3478 			if (work_is_canceling(work))
3479 				schedule();
3480 			finish_wait(&cancel_waitq, &cwait.wait);
3481 		}
3482 	} while (unlikely(ret < 0));
3483 
3484 	/* tell other tasks trying to grab @work to back off */
3485 	mark_work_canceling(work);
3486 	local_irq_restore(flags);
3487 
3488 	/*
3489 	 * This allows canceling during early boot.  We know that @work
3490 	 * isn't executing.
3491 	 */
3492 	if (wq_online)
3493 		__flush_work(work, true);
3494 
3495 	clear_work_data(work);
3496 
3497 	/*
3498 	 * Paired with prepare_to_wait() above so that either
3499 	 * waitqueue_active() is visible here or !work_is_canceling() is
3500 	 * visible there.
3501 	 */
3502 	smp_mb();
3503 	if (waitqueue_active(&cancel_waitq))
3504 		__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3505 
3506 	return ret;
3507 }
3508 
3509 /**
3510  * cancel_work_sync - cancel a work and wait for it to finish
3511  * @work: the work to cancel
3512  *
3513  * Cancel @work and wait for its execution to finish.  This function
3514  * can be used even if the work re-queues itself or migrates to
3515  * another workqueue.  On return from this function, @work is
3516  * guaranteed to be not pending or executing on any CPU.
3517  *
3518  * cancel_work_sync(&delayed_work->work) must not be used for
3519  * delayed_work's.  Use cancel_delayed_work_sync() instead.
3520  *
3521  * The caller must ensure that the workqueue on which @work was last
3522  * queued can't be destroyed before this function returns.
3523  *
3524  * Return:
3525  * %true if @work was pending, %false otherwise.
3526  */
3527 bool cancel_work_sync(struct work_struct *work)
3528 {
3529 	return __cancel_work_timer(work, false);
3530 }
3531 EXPORT_SYMBOL_GPL(cancel_work_sync);
3532 
3533 /**
3534  * flush_delayed_work - wait for a dwork to finish executing the last queueing
3535  * @dwork: the delayed work to flush
3536  *
3537  * Delayed timer is cancelled and the pending work is queued for
3538  * immediate execution.  Like flush_work(), this function only
3539  * considers the last queueing instance of @dwork.
3540  *
3541  * Return:
3542  * %true if flush_work() waited for the work to finish execution,
3543  * %false if it was already idle.
3544  */
3545 bool flush_delayed_work(struct delayed_work *dwork)
3546 {
3547 	local_irq_disable();
3548 	if (del_timer_sync(&dwork->timer))
3549 		__queue_work(dwork->cpu, dwork->wq, &dwork->work);
3550 	local_irq_enable();
3551 	return flush_work(&dwork->work);
3552 }
3553 EXPORT_SYMBOL(flush_delayed_work);
3554 
3555 /**
3556  * flush_rcu_work - wait for a rwork to finish executing the last queueing
3557  * @rwork: the rcu work to flush
3558  *
3559  * Return:
3560  * %true if flush_rcu_work() waited for the work to finish execution,
3561  * %false if it was already idle.
3562  */
3563 bool flush_rcu_work(struct rcu_work *rwork)
3564 {
3565 	if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3566 		rcu_barrier();
3567 		flush_work(&rwork->work);
3568 		return true;
3569 	} else {
3570 		return flush_work(&rwork->work);
3571 	}
3572 }
3573 EXPORT_SYMBOL(flush_rcu_work);
3574 
3575 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3576 {
3577 	unsigned long flags;
3578 	int ret;
3579 
3580 	do {
3581 		ret = try_to_grab_pending(work, is_dwork, &flags);
3582 	} while (unlikely(ret == -EAGAIN));
3583 
3584 	if (unlikely(ret < 0))
3585 		return false;
3586 
3587 	set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3588 	local_irq_restore(flags);
3589 	return ret;
3590 }
3591 
3592 /*
3593  * See cancel_delayed_work()
3594  */
3595 bool cancel_work(struct work_struct *work)
3596 {
3597 	return __cancel_work(work, false);
3598 }
3599 EXPORT_SYMBOL(cancel_work);
3600 
3601 /**
3602  * cancel_delayed_work - cancel a delayed work
3603  * @dwork: delayed_work to cancel
3604  *
3605  * Kill off a pending delayed_work.
3606  *
3607  * Return: %true if @dwork was pending and canceled; %false if it wasn't
3608  * pending.
3609  *
3610  * Note:
3611  * The work callback function may still be running on return, unless
3612  * it returns %true and the work doesn't re-arm itself.  Explicitly flush or
3613  * use cancel_delayed_work_sync() to wait on it.
3614  *
3615  * This function is safe to call from any context including IRQ handler.
3616  */
3617 bool cancel_delayed_work(struct delayed_work *dwork)
3618 {
3619 	return __cancel_work(&dwork->work, true);
3620 }
3621 EXPORT_SYMBOL(cancel_delayed_work);
3622 
3623 /**
3624  * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3625  * @dwork: the delayed work cancel
3626  *
3627  * This is cancel_work_sync() for delayed works.
3628  *
3629  * Return:
3630  * %true if @dwork was pending, %false otherwise.
3631  */
3632 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3633 {
3634 	return __cancel_work_timer(&dwork->work, true);
3635 }
3636 EXPORT_SYMBOL(cancel_delayed_work_sync);
3637 
3638 /**
3639  * schedule_on_each_cpu - execute a function synchronously on each online CPU
3640  * @func: the function to call
3641  *
3642  * schedule_on_each_cpu() executes @func on each online CPU using the
3643  * system workqueue and blocks until all CPUs have completed.
3644  * schedule_on_each_cpu() is very slow.
3645  *
3646  * Return:
3647  * 0 on success, -errno on failure.
3648  */
3649 int schedule_on_each_cpu(work_func_t func)
3650 {
3651 	int cpu;
3652 	struct work_struct __percpu *works;
3653 
3654 	works = alloc_percpu(struct work_struct);
3655 	if (!works)
3656 		return -ENOMEM;
3657 
3658 	cpus_read_lock();
3659 
3660 	for_each_online_cpu(cpu) {
3661 		struct work_struct *work = per_cpu_ptr(works, cpu);
3662 
3663 		INIT_WORK(work, func);
3664 		schedule_work_on(cpu, work);
3665 	}
3666 
3667 	for_each_online_cpu(cpu)
3668 		flush_work(per_cpu_ptr(works, cpu));
3669 
3670 	cpus_read_unlock();
3671 	free_percpu(works);
3672 	return 0;
3673 }
3674 
3675 /**
3676  * execute_in_process_context - reliably execute the routine with user context
3677  * @fn:		the function to execute
3678  * @ew:		guaranteed storage for the execute work structure (must
3679  *		be available when the work executes)
3680  *
3681  * Executes the function immediately if process context is available,
3682  * otherwise schedules the function for delayed execution.
3683  *
3684  * Return:	0 - function was executed
3685  *		1 - function was scheduled for execution
3686  */
3687 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3688 {
3689 	if (!in_interrupt()) {
3690 		fn(&ew->work);
3691 		return 0;
3692 	}
3693 
3694 	INIT_WORK(&ew->work, fn);
3695 	schedule_work(&ew->work);
3696 
3697 	return 1;
3698 }
3699 EXPORT_SYMBOL_GPL(execute_in_process_context);
3700 
3701 /**
3702  * free_workqueue_attrs - free a workqueue_attrs
3703  * @attrs: workqueue_attrs to free
3704  *
3705  * Undo alloc_workqueue_attrs().
3706  */
3707 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3708 {
3709 	if (attrs) {
3710 		free_cpumask_var(attrs->cpumask);
3711 		free_cpumask_var(attrs->__pod_cpumask);
3712 		kfree(attrs);
3713 	}
3714 }
3715 
3716 /**
3717  * alloc_workqueue_attrs - allocate a workqueue_attrs
3718  *
3719  * Allocate a new workqueue_attrs, initialize with default settings and
3720  * return it.
3721  *
3722  * Return: The allocated new workqueue_attr on success. %NULL on failure.
3723  */
3724 struct workqueue_attrs *alloc_workqueue_attrs(void)
3725 {
3726 	struct workqueue_attrs *attrs;
3727 
3728 	attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3729 	if (!attrs)
3730 		goto fail;
3731 	if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3732 		goto fail;
3733 	if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
3734 		goto fail;
3735 
3736 	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3737 	attrs->affn_scope = wq_affn_dfl;
3738 	return attrs;
3739 fail:
3740 	free_workqueue_attrs(attrs);
3741 	return NULL;
3742 }
3743 
3744 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3745 				 const struct workqueue_attrs *from)
3746 {
3747 	to->nice = from->nice;
3748 	cpumask_copy(to->cpumask, from->cpumask);
3749 	cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
3750 	to->affn_strict = from->affn_strict;
3751 
3752 	/*
3753 	 * Unlike hash and equality test, copying shouldn't ignore wq-only
3754 	 * fields as copying is used for both pool and wq attrs. Instead,
3755 	 * get_unbound_pool() explicitly clears the fields.
3756 	 */
3757 	to->affn_scope = from->affn_scope;
3758 	to->ordered = from->ordered;
3759 }
3760 
3761 /*
3762  * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
3763  * comments in 'struct workqueue_attrs' definition.
3764  */
3765 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
3766 {
3767 	attrs->affn_scope = WQ_AFFN_NR_TYPES;
3768 	attrs->ordered = false;
3769 }
3770 
3771 /* hash value of the content of @attr */
3772 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3773 {
3774 	u32 hash = 0;
3775 
3776 	hash = jhash_1word(attrs->nice, hash);
3777 	hash = jhash(cpumask_bits(attrs->cpumask),
3778 		     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3779 	hash = jhash(cpumask_bits(attrs->__pod_cpumask),
3780 		     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3781 	hash = jhash_1word(attrs->affn_strict, hash);
3782 	return hash;
3783 }
3784 
3785 /* content equality test */
3786 static bool wqattrs_equal(const struct workqueue_attrs *a,
3787 			  const struct workqueue_attrs *b)
3788 {
3789 	if (a->nice != b->nice)
3790 		return false;
3791 	if (!cpumask_equal(a->cpumask, b->cpumask))
3792 		return false;
3793 	if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
3794 		return false;
3795 	if (a->affn_strict != b->affn_strict)
3796 		return false;
3797 	return true;
3798 }
3799 
3800 /* Update @attrs with actually available CPUs */
3801 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
3802 				      const cpumask_t *unbound_cpumask)
3803 {
3804 	/*
3805 	 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
3806 	 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
3807 	 * @unbound_cpumask.
3808 	 */
3809 	cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
3810 	if (unlikely(cpumask_empty(attrs->cpumask)))
3811 		cpumask_copy(attrs->cpumask, unbound_cpumask);
3812 }
3813 
3814 /* find wq_pod_type to use for @attrs */
3815 static const struct wq_pod_type *
3816 wqattrs_pod_type(const struct workqueue_attrs *attrs)
3817 {
3818 	struct wq_pod_type *pt = &wq_pod_types[attrs->affn_scope];
3819 
3820 	if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
3821 	    likely(pt->nr_pods))
3822 		return pt;
3823 
3824 	/*
3825 	 * Before workqueue_init_topology(), only SYSTEM is available which is
3826 	 * initialized in workqueue_init_early().
3827 	 */
3828 	pt = &wq_pod_types[WQ_AFFN_SYSTEM];
3829 	BUG_ON(!pt->nr_pods);
3830 	return pt;
3831 }
3832 
3833 /**
3834  * init_worker_pool - initialize a newly zalloc'd worker_pool
3835  * @pool: worker_pool to initialize
3836  *
3837  * Initialize a newly zalloc'd @pool.  It also allocates @pool->attrs.
3838  *
3839  * Return: 0 on success, -errno on failure.  Even on failure, all fields
3840  * inside @pool proper are initialized and put_unbound_pool() can be called
3841  * on @pool safely to release it.
3842  */
3843 static int init_worker_pool(struct worker_pool *pool)
3844 {
3845 	raw_spin_lock_init(&pool->lock);
3846 	pool->id = -1;
3847 	pool->cpu = -1;
3848 	pool->node = NUMA_NO_NODE;
3849 	pool->flags |= POOL_DISASSOCIATED;
3850 	pool->watchdog_ts = jiffies;
3851 	INIT_LIST_HEAD(&pool->worklist);
3852 	INIT_LIST_HEAD(&pool->idle_list);
3853 	hash_init(pool->busy_hash);
3854 
3855 	timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3856 	INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
3857 
3858 	timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3859 
3860 	INIT_LIST_HEAD(&pool->workers);
3861 	INIT_LIST_HEAD(&pool->dying_workers);
3862 
3863 	ida_init(&pool->worker_ida);
3864 	INIT_HLIST_NODE(&pool->hash_node);
3865 	pool->refcnt = 1;
3866 
3867 	/* shouldn't fail above this point */
3868 	pool->attrs = alloc_workqueue_attrs();
3869 	if (!pool->attrs)
3870 		return -ENOMEM;
3871 
3872 	wqattrs_clear_for_pool(pool->attrs);
3873 
3874 	return 0;
3875 }
3876 
3877 #ifdef CONFIG_LOCKDEP
3878 static void wq_init_lockdep(struct workqueue_struct *wq)
3879 {
3880 	char *lock_name;
3881 
3882 	lockdep_register_key(&wq->key);
3883 	lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3884 	if (!lock_name)
3885 		lock_name = wq->name;
3886 
3887 	wq->lock_name = lock_name;
3888 	lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3889 }
3890 
3891 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3892 {
3893 	lockdep_unregister_key(&wq->key);
3894 }
3895 
3896 static void wq_free_lockdep(struct workqueue_struct *wq)
3897 {
3898 	if (wq->lock_name != wq->name)
3899 		kfree(wq->lock_name);
3900 }
3901 #else
3902 static void wq_init_lockdep(struct workqueue_struct *wq)
3903 {
3904 }
3905 
3906 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3907 {
3908 }
3909 
3910 static void wq_free_lockdep(struct workqueue_struct *wq)
3911 {
3912 }
3913 #endif
3914 
3915 static void rcu_free_wq(struct rcu_head *rcu)
3916 {
3917 	struct workqueue_struct *wq =
3918 		container_of(rcu, struct workqueue_struct, rcu);
3919 
3920 	wq_free_lockdep(wq);
3921 	free_percpu(wq->cpu_pwq);
3922 	free_workqueue_attrs(wq->unbound_attrs);
3923 	kfree(wq);
3924 }
3925 
3926 static void rcu_free_pool(struct rcu_head *rcu)
3927 {
3928 	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3929 
3930 	ida_destroy(&pool->worker_ida);
3931 	free_workqueue_attrs(pool->attrs);
3932 	kfree(pool);
3933 }
3934 
3935 /**
3936  * put_unbound_pool - put a worker_pool
3937  * @pool: worker_pool to put
3938  *
3939  * Put @pool.  If its refcnt reaches zero, it gets destroyed in RCU
3940  * safe manner.  get_unbound_pool() calls this function on its failure path
3941  * and this function should be able to release pools which went through,
3942  * successfully or not, init_worker_pool().
3943  *
3944  * Should be called with wq_pool_mutex held.
3945  */
3946 static void put_unbound_pool(struct worker_pool *pool)
3947 {
3948 	DECLARE_COMPLETION_ONSTACK(detach_completion);
3949 	struct worker *worker;
3950 	LIST_HEAD(cull_list);
3951 
3952 	lockdep_assert_held(&wq_pool_mutex);
3953 
3954 	if (--pool->refcnt)
3955 		return;
3956 
3957 	/* sanity checks */
3958 	if (WARN_ON(!(pool->cpu < 0)) ||
3959 	    WARN_ON(!list_empty(&pool->worklist)))
3960 		return;
3961 
3962 	/* release id and unhash */
3963 	if (pool->id >= 0)
3964 		idr_remove(&worker_pool_idr, pool->id);
3965 	hash_del(&pool->hash_node);
3966 
3967 	/*
3968 	 * Become the manager and destroy all workers.  This prevents
3969 	 * @pool's workers from blocking on attach_mutex.  We're the last
3970 	 * manager and @pool gets freed with the flag set.
3971 	 *
3972 	 * Having a concurrent manager is quite unlikely to happen as we can
3973 	 * only get here with
3974 	 *   pwq->refcnt == pool->refcnt == 0
3975 	 * which implies no work queued to the pool, which implies no worker can
3976 	 * become the manager. However a worker could have taken the role of
3977 	 * manager before the refcnts dropped to 0, since maybe_create_worker()
3978 	 * drops pool->lock
3979 	 */
3980 	while (true) {
3981 		rcuwait_wait_event(&manager_wait,
3982 				   !(pool->flags & POOL_MANAGER_ACTIVE),
3983 				   TASK_UNINTERRUPTIBLE);
3984 
3985 		mutex_lock(&wq_pool_attach_mutex);
3986 		raw_spin_lock_irq(&pool->lock);
3987 		if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
3988 			pool->flags |= POOL_MANAGER_ACTIVE;
3989 			break;
3990 		}
3991 		raw_spin_unlock_irq(&pool->lock);
3992 		mutex_unlock(&wq_pool_attach_mutex);
3993 	}
3994 
3995 	while ((worker = first_idle_worker(pool)))
3996 		set_worker_dying(worker, &cull_list);
3997 	WARN_ON(pool->nr_workers || pool->nr_idle);
3998 	raw_spin_unlock_irq(&pool->lock);
3999 
4000 	wake_dying_workers(&cull_list);
4001 
4002 	if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
4003 		pool->detach_completion = &detach_completion;
4004 	mutex_unlock(&wq_pool_attach_mutex);
4005 
4006 	if (pool->detach_completion)
4007 		wait_for_completion(pool->detach_completion);
4008 
4009 	/* shut down the timers */
4010 	del_timer_sync(&pool->idle_timer);
4011 	cancel_work_sync(&pool->idle_cull_work);
4012 	del_timer_sync(&pool->mayday_timer);
4013 
4014 	/* RCU protected to allow dereferences from get_work_pool() */
4015 	call_rcu(&pool->rcu, rcu_free_pool);
4016 }
4017 
4018 /**
4019  * get_unbound_pool - get a worker_pool with the specified attributes
4020  * @attrs: the attributes of the worker_pool to get
4021  *
4022  * Obtain a worker_pool which has the same attributes as @attrs, bump the
4023  * reference count and return it.  If there already is a matching
4024  * worker_pool, it will be used; otherwise, this function attempts to
4025  * create a new one.
4026  *
4027  * Should be called with wq_pool_mutex held.
4028  *
4029  * Return: On success, a worker_pool with the same attributes as @attrs.
4030  * On failure, %NULL.
4031  */
4032 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4033 {
4034 	struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4035 	u32 hash = wqattrs_hash(attrs);
4036 	struct worker_pool *pool;
4037 	int pod, node = NUMA_NO_NODE;
4038 
4039 	lockdep_assert_held(&wq_pool_mutex);
4040 
4041 	/* do we already have a matching pool? */
4042 	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4043 		if (wqattrs_equal(pool->attrs, attrs)) {
4044 			pool->refcnt++;
4045 			return pool;
4046 		}
4047 	}
4048 
4049 	/* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4050 	for (pod = 0; pod < pt->nr_pods; pod++) {
4051 		if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4052 			node = pt->pod_node[pod];
4053 			break;
4054 		}
4055 	}
4056 
4057 	/* nope, create a new one */
4058 	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4059 	if (!pool || init_worker_pool(pool) < 0)
4060 		goto fail;
4061 
4062 	pool->node = node;
4063 	copy_workqueue_attrs(pool->attrs, attrs);
4064 	wqattrs_clear_for_pool(pool->attrs);
4065 
4066 	if (worker_pool_assign_id(pool) < 0)
4067 		goto fail;
4068 
4069 	/* create and start the initial worker */
4070 	if (wq_online && !create_worker(pool))
4071 		goto fail;
4072 
4073 	/* install */
4074 	hash_add(unbound_pool_hash, &pool->hash_node, hash);
4075 
4076 	return pool;
4077 fail:
4078 	if (pool)
4079 		put_unbound_pool(pool);
4080 	return NULL;
4081 }
4082 
4083 static void rcu_free_pwq(struct rcu_head *rcu)
4084 {
4085 	kmem_cache_free(pwq_cache,
4086 			container_of(rcu, struct pool_workqueue, rcu));
4087 }
4088 
4089 /*
4090  * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
4091  * refcnt and needs to be destroyed.
4092  */
4093 static void pwq_release_workfn(struct kthread_work *work)
4094 {
4095 	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
4096 						  release_work);
4097 	struct workqueue_struct *wq = pwq->wq;
4098 	struct worker_pool *pool = pwq->pool;
4099 	bool is_last = false;
4100 
4101 	/*
4102 	 * When @pwq is not linked, it doesn't hold any reference to the
4103 	 * @wq, and @wq is invalid to access.
4104 	 */
4105 	if (!list_empty(&pwq->pwqs_node)) {
4106 		mutex_lock(&wq->mutex);
4107 		list_del_rcu(&pwq->pwqs_node);
4108 		is_last = list_empty(&wq->pwqs);
4109 		mutex_unlock(&wq->mutex);
4110 	}
4111 
4112 	if (wq->flags & WQ_UNBOUND) {
4113 		mutex_lock(&wq_pool_mutex);
4114 		put_unbound_pool(pool);
4115 		mutex_unlock(&wq_pool_mutex);
4116 	}
4117 
4118 	call_rcu(&pwq->rcu, rcu_free_pwq);
4119 
4120 	/*
4121 	 * If we're the last pwq going away, @wq is already dead and no one
4122 	 * is gonna access it anymore.  Schedule RCU free.
4123 	 */
4124 	if (is_last) {
4125 		wq_unregister_lockdep(wq);
4126 		call_rcu(&wq->rcu, rcu_free_wq);
4127 	}
4128 }
4129 
4130 /**
4131  * pwq_adjust_max_active - update a pwq's max_active to the current setting
4132  * @pwq: target pool_workqueue
4133  *
4134  * If @pwq isn't freezing, set @pwq->max_active to the associated
4135  * workqueue's saved_max_active and activate inactive work items
4136  * accordingly.  If @pwq is freezing, clear @pwq->max_active to zero.
4137  */
4138 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
4139 {
4140 	struct workqueue_struct *wq = pwq->wq;
4141 	bool freezable = wq->flags & WQ_FREEZABLE;
4142 	unsigned long flags;
4143 
4144 	/* for @wq->saved_max_active */
4145 	lockdep_assert_held(&wq->mutex);
4146 
4147 	/* fast exit for non-freezable wqs */
4148 	if (!freezable && pwq->max_active == wq->saved_max_active)
4149 		return;
4150 
4151 	/* this function can be called during early boot w/ irq disabled */
4152 	raw_spin_lock_irqsave(&pwq->pool->lock, flags);
4153 
4154 	/*
4155 	 * During [un]freezing, the caller is responsible for ensuring that
4156 	 * this function is called at least once after @workqueue_freezing
4157 	 * is updated and visible.
4158 	 */
4159 	if (!freezable || !workqueue_freezing) {
4160 		pwq->max_active = wq->saved_max_active;
4161 
4162 		while (!list_empty(&pwq->inactive_works) &&
4163 		       pwq->nr_active < pwq->max_active)
4164 			pwq_activate_first_inactive(pwq);
4165 
4166 		kick_pool(pwq->pool);
4167 	} else {
4168 		pwq->max_active = 0;
4169 	}
4170 
4171 	raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
4172 }
4173 
4174 /* initialize newly allocated @pwq which is associated with @wq and @pool */
4175 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
4176 		     struct worker_pool *pool)
4177 {
4178 	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
4179 
4180 	memset(pwq, 0, sizeof(*pwq));
4181 
4182 	pwq->pool = pool;
4183 	pwq->wq = wq;
4184 	pwq->flush_color = -1;
4185 	pwq->refcnt = 1;
4186 	INIT_LIST_HEAD(&pwq->inactive_works);
4187 	INIT_LIST_HEAD(&pwq->pwqs_node);
4188 	INIT_LIST_HEAD(&pwq->mayday_node);
4189 	kthread_init_work(&pwq->release_work, pwq_release_workfn);
4190 }
4191 
4192 /* sync @pwq with the current state of its associated wq and link it */
4193 static void link_pwq(struct pool_workqueue *pwq)
4194 {
4195 	struct workqueue_struct *wq = pwq->wq;
4196 
4197 	lockdep_assert_held(&wq->mutex);
4198 
4199 	/* may be called multiple times, ignore if already linked */
4200 	if (!list_empty(&pwq->pwqs_node))
4201 		return;
4202 
4203 	/* set the matching work_color */
4204 	pwq->work_color = wq->work_color;
4205 
4206 	/* sync max_active to the current setting */
4207 	pwq_adjust_max_active(pwq);
4208 
4209 	/* link in @pwq */
4210 	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
4211 }
4212 
4213 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
4214 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
4215 					const struct workqueue_attrs *attrs)
4216 {
4217 	struct worker_pool *pool;
4218 	struct pool_workqueue *pwq;
4219 
4220 	lockdep_assert_held(&wq_pool_mutex);
4221 
4222 	pool = get_unbound_pool(attrs);
4223 	if (!pool)
4224 		return NULL;
4225 
4226 	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
4227 	if (!pwq) {
4228 		put_unbound_pool(pool);
4229 		return NULL;
4230 	}
4231 
4232 	init_pwq(pwq, wq, pool);
4233 	return pwq;
4234 }
4235 
4236 /**
4237  * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
4238  * @attrs: the wq_attrs of the default pwq of the target workqueue
4239  * @cpu: the target CPU
4240  * @cpu_going_down: if >= 0, the CPU to consider as offline
4241  *
4242  * Calculate the cpumask a workqueue with @attrs should use on @pod. If
4243  * @cpu_going_down is >= 0, that cpu is considered offline during calculation.
4244  * The result is stored in @attrs->__pod_cpumask.
4245  *
4246  * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
4247  * and @pod has online CPUs requested by @attrs, the returned cpumask is the
4248  * intersection of the possible CPUs of @pod and @attrs->cpumask.
4249  *
4250  * The caller is responsible for ensuring that the cpumask of @pod stays stable.
4251  */
4252 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
4253 				int cpu_going_down)
4254 {
4255 	const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
4256 	int pod = pt->cpu_pod[cpu];
4257 
4258 	/* does @pod have any online CPUs @attrs wants? */
4259 	cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
4260 	cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
4261 	if (cpu_going_down >= 0)
4262 		cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
4263 
4264 	if (cpumask_empty(attrs->__pod_cpumask)) {
4265 		cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
4266 		return;
4267 	}
4268 
4269 	/* yeap, return possible CPUs in @pod that @attrs wants */
4270 	cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
4271 
4272 	if (cpumask_empty(attrs->__pod_cpumask))
4273 		pr_warn_once("WARNING: workqueue cpumask: online intersect > "
4274 				"possible intersect\n");
4275 }
4276 
4277 /* install @pwq into @wq's cpu_pwq and return the old pwq */
4278 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
4279 					int cpu, struct pool_workqueue *pwq)
4280 {
4281 	struct pool_workqueue *old_pwq;
4282 
4283 	lockdep_assert_held(&wq_pool_mutex);
4284 	lockdep_assert_held(&wq->mutex);
4285 
4286 	/* link_pwq() can handle duplicate calls */
4287 	link_pwq(pwq);
4288 
4289 	old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
4290 	rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq);
4291 	return old_pwq;
4292 }
4293 
4294 /* context to store the prepared attrs & pwqs before applying */
4295 struct apply_wqattrs_ctx {
4296 	struct workqueue_struct	*wq;		/* target workqueue */
4297 	struct workqueue_attrs	*attrs;		/* attrs to apply */
4298 	struct list_head	list;		/* queued for batching commit */
4299 	struct pool_workqueue	*dfl_pwq;
4300 	struct pool_workqueue	*pwq_tbl[];
4301 };
4302 
4303 /* free the resources after success or abort */
4304 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
4305 {
4306 	if (ctx) {
4307 		int cpu;
4308 
4309 		for_each_possible_cpu(cpu)
4310 			put_pwq_unlocked(ctx->pwq_tbl[cpu]);
4311 		put_pwq_unlocked(ctx->dfl_pwq);
4312 
4313 		free_workqueue_attrs(ctx->attrs);
4314 
4315 		kfree(ctx);
4316 	}
4317 }
4318 
4319 /* allocate the attrs and pwqs for later installation */
4320 static struct apply_wqattrs_ctx *
4321 apply_wqattrs_prepare(struct workqueue_struct *wq,
4322 		      const struct workqueue_attrs *attrs,
4323 		      const cpumask_var_t unbound_cpumask)
4324 {
4325 	struct apply_wqattrs_ctx *ctx;
4326 	struct workqueue_attrs *new_attrs;
4327 	int cpu;
4328 
4329 	lockdep_assert_held(&wq_pool_mutex);
4330 
4331 	if (WARN_ON(attrs->affn_scope < 0 ||
4332 		    attrs->affn_scope >= WQ_AFFN_NR_TYPES))
4333 		return ERR_PTR(-EINVAL);
4334 
4335 	ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
4336 
4337 	new_attrs = alloc_workqueue_attrs();
4338 	if (!ctx || !new_attrs)
4339 		goto out_free;
4340 
4341 	/*
4342 	 * If something goes wrong during CPU up/down, we'll fall back to
4343 	 * the default pwq covering whole @attrs->cpumask.  Always create
4344 	 * it even if we don't use it immediately.
4345 	 */
4346 	copy_workqueue_attrs(new_attrs, attrs);
4347 	wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
4348 	cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4349 	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
4350 	if (!ctx->dfl_pwq)
4351 		goto out_free;
4352 
4353 	for_each_possible_cpu(cpu) {
4354 		if (new_attrs->ordered) {
4355 			ctx->dfl_pwq->refcnt++;
4356 			ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
4357 		} else {
4358 			wq_calc_pod_cpumask(new_attrs, cpu, -1);
4359 			ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
4360 			if (!ctx->pwq_tbl[cpu])
4361 				goto out_free;
4362 		}
4363 	}
4364 
4365 	/* save the user configured attrs and sanitize it. */
4366 	copy_workqueue_attrs(new_attrs, attrs);
4367 	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
4368 	cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4369 	ctx->attrs = new_attrs;
4370 
4371 	ctx->wq = wq;
4372 	return ctx;
4373 
4374 out_free:
4375 	free_workqueue_attrs(new_attrs);
4376 	apply_wqattrs_cleanup(ctx);
4377 	return ERR_PTR(-ENOMEM);
4378 }
4379 
4380 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
4381 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4382 {
4383 	int cpu;
4384 
4385 	/* all pwqs have been created successfully, let's install'em */
4386 	mutex_lock(&ctx->wq->mutex);
4387 
4388 	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4389 
4390 	/* save the previous pwq and install the new one */
4391 	for_each_possible_cpu(cpu)
4392 		ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
4393 							ctx->pwq_tbl[cpu]);
4394 
4395 	/* @dfl_pwq might not have been used, ensure it's linked */
4396 	link_pwq(ctx->dfl_pwq);
4397 	swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
4398 
4399 	mutex_unlock(&ctx->wq->mutex);
4400 }
4401 
4402 static void apply_wqattrs_lock(void)
4403 {
4404 	/* CPUs should stay stable across pwq creations and installations */
4405 	cpus_read_lock();
4406 	mutex_lock(&wq_pool_mutex);
4407 }
4408 
4409 static void apply_wqattrs_unlock(void)
4410 {
4411 	mutex_unlock(&wq_pool_mutex);
4412 	cpus_read_unlock();
4413 }
4414 
4415 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4416 					const struct workqueue_attrs *attrs)
4417 {
4418 	struct apply_wqattrs_ctx *ctx;
4419 
4420 	/* only unbound workqueues can change attributes */
4421 	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4422 		return -EINVAL;
4423 
4424 	/* creating multiple pwqs breaks ordering guarantee */
4425 	if (!list_empty(&wq->pwqs)) {
4426 		if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4427 			return -EINVAL;
4428 
4429 		wq->flags &= ~__WQ_ORDERED;
4430 	}
4431 
4432 	ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
4433 	if (IS_ERR(ctx))
4434 		return PTR_ERR(ctx);
4435 
4436 	/* the ctx has been prepared successfully, let's commit it */
4437 	apply_wqattrs_commit(ctx);
4438 	apply_wqattrs_cleanup(ctx);
4439 
4440 	return 0;
4441 }
4442 
4443 /**
4444  * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4445  * @wq: the target workqueue
4446  * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4447  *
4448  * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
4449  * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
4450  * work items are affine to the pod it was issued on. Older pwqs are released as
4451  * in-flight work items finish. Note that a work item which repeatedly requeues
4452  * itself back-to-back will stay on its current pwq.
4453  *
4454  * Performs GFP_KERNEL allocations.
4455  *
4456  * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
4457  *
4458  * Return: 0 on success and -errno on failure.
4459  */
4460 int apply_workqueue_attrs(struct workqueue_struct *wq,
4461 			  const struct workqueue_attrs *attrs)
4462 {
4463 	int ret;
4464 
4465 	lockdep_assert_cpus_held();
4466 
4467 	mutex_lock(&wq_pool_mutex);
4468 	ret = apply_workqueue_attrs_locked(wq, attrs);
4469 	mutex_unlock(&wq_pool_mutex);
4470 
4471 	return ret;
4472 }
4473 
4474 /**
4475  * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
4476  * @wq: the target workqueue
4477  * @cpu: the CPU to update pool association for
4478  * @hotplug_cpu: the CPU coming up or going down
4479  * @online: whether @cpu is coming up or going down
4480  *
4481  * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4482  * %CPU_DOWN_FAILED.  @cpu is being hot[un]plugged, update pod affinity of
4483  * @wq accordingly.
4484  *
4485  *
4486  * If pod affinity can't be adjusted due to memory allocation failure, it falls
4487  * back to @wq->dfl_pwq which may not be optimal but is always correct.
4488  *
4489  * Note that when the last allowed CPU of a pod goes offline for a workqueue
4490  * with a cpumask spanning multiple pods, the workers which were already
4491  * executing the work items for the workqueue will lose their CPU affinity and
4492  * may execute on any CPU. This is similar to how per-cpu workqueues behave on
4493  * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
4494  * responsibility to flush the work item from CPU_DOWN_PREPARE.
4495  */
4496 static void wq_update_pod(struct workqueue_struct *wq, int cpu,
4497 			  int hotplug_cpu, bool online)
4498 {
4499 	int off_cpu = online ? -1 : hotplug_cpu;
4500 	struct pool_workqueue *old_pwq = NULL, *pwq;
4501 	struct workqueue_attrs *target_attrs;
4502 
4503 	lockdep_assert_held(&wq_pool_mutex);
4504 
4505 	if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
4506 		return;
4507 
4508 	/*
4509 	 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4510 	 * Let's use a preallocated one.  The following buf is protected by
4511 	 * CPU hotplug exclusion.
4512 	 */
4513 	target_attrs = wq_update_pod_attrs_buf;
4514 
4515 	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4516 	wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
4517 
4518 	/* nothing to do if the target cpumask matches the current pwq */
4519 	wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
4520 	pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu),
4521 					lockdep_is_held(&wq_pool_mutex));
4522 	if (wqattrs_equal(target_attrs, pwq->pool->attrs))
4523 		return;
4524 
4525 	/* create a new pwq */
4526 	pwq = alloc_unbound_pwq(wq, target_attrs);
4527 	if (!pwq) {
4528 		pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
4529 			wq->name);
4530 		goto use_dfl_pwq;
4531 	}
4532 
4533 	/* Install the new pwq. */
4534 	mutex_lock(&wq->mutex);
4535 	old_pwq = install_unbound_pwq(wq, cpu, pwq);
4536 	goto out_unlock;
4537 
4538 use_dfl_pwq:
4539 	mutex_lock(&wq->mutex);
4540 	raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
4541 	get_pwq(wq->dfl_pwq);
4542 	raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4543 	old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq);
4544 out_unlock:
4545 	mutex_unlock(&wq->mutex);
4546 	put_pwq_unlocked(old_pwq);
4547 }
4548 
4549 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4550 {
4551 	bool highpri = wq->flags & WQ_HIGHPRI;
4552 	int cpu, ret;
4553 
4554 	wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
4555 	if (!wq->cpu_pwq)
4556 		goto enomem;
4557 
4558 	if (!(wq->flags & WQ_UNBOUND)) {
4559 		for_each_possible_cpu(cpu) {
4560 			struct pool_workqueue **pwq_p =
4561 				per_cpu_ptr(wq->cpu_pwq, cpu);
4562 			struct worker_pool *pool =
4563 				&(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]);
4564 
4565 			*pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
4566 						       pool->node);
4567 			if (!*pwq_p)
4568 				goto enomem;
4569 
4570 			init_pwq(*pwq_p, wq, pool);
4571 
4572 			mutex_lock(&wq->mutex);
4573 			link_pwq(*pwq_p);
4574 			mutex_unlock(&wq->mutex);
4575 		}
4576 		return 0;
4577 	}
4578 
4579 	cpus_read_lock();
4580 	if (wq->flags & __WQ_ORDERED) {
4581 		ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4582 		/* there should only be single pwq for ordering guarantee */
4583 		WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4584 			      wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4585 		     "ordering guarantee broken for workqueue %s\n", wq->name);
4586 	} else {
4587 		ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4588 	}
4589 	cpus_read_unlock();
4590 
4591 	return ret;
4592 
4593 enomem:
4594 	if (wq->cpu_pwq) {
4595 		for_each_possible_cpu(cpu)
4596 			kfree(*per_cpu_ptr(wq->cpu_pwq, cpu));
4597 		free_percpu(wq->cpu_pwq);
4598 		wq->cpu_pwq = NULL;
4599 	}
4600 	return -ENOMEM;
4601 }
4602 
4603 static int wq_clamp_max_active(int max_active, unsigned int flags,
4604 			       const char *name)
4605 {
4606 	if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
4607 		pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4608 			max_active, name, 1, WQ_MAX_ACTIVE);
4609 
4610 	return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
4611 }
4612 
4613 /*
4614  * Workqueues which may be used during memory reclaim should have a rescuer
4615  * to guarantee forward progress.
4616  */
4617 static int init_rescuer(struct workqueue_struct *wq)
4618 {
4619 	struct worker *rescuer;
4620 	int ret;
4621 
4622 	if (!(wq->flags & WQ_MEM_RECLAIM))
4623 		return 0;
4624 
4625 	rescuer = alloc_worker(NUMA_NO_NODE);
4626 	if (!rescuer) {
4627 		pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
4628 		       wq->name);
4629 		return -ENOMEM;
4630 	}
4631 
4632 	rescuer->rescue_wq = wq;
4633 	rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4634 	if (IS_ERR(rescuer->task)) {
4635 		ret = PTR_ERR(rescuer->task);
4636 		pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
4637 		       wq->name, ERR_PTR(ret));
4638 		kfree(rescuer);
4639 		return ret;
4640 	}
4641 
4642 	wq->rescuer = rescuer;
4643 	kthread_bind_mask(rescuer->task, cpu_possible_mask);
4644 	wake_up_process(rescuer->task);
4645 
4646 	return 0;
4647 }
4648 
4649 __printf(1, 4)
4650 struct workqueue_struct *alloc_workqueue(const char *fmt,
4651 					 unsigned int flags,
4652 					 int max_active, ...)
4653 {
4654 	va_list args;
4655 	struct workqueue_struct *wq;
4656 	struct pool_workqueue *pwq;
4657 
4658 	/*
4659 	 * Unbound && max_active == 1 used to imply ordered, which is no longer
4660 	 * the case on many machines due to per-pod pools. While
4661 	 * alloc_ordered_workqueue() is the right way to create an ordered
4662 	 * workqueue, keep the previous behavior to avoid subtle breakages.
4663 	 */
4664 	if ((flags & WQ_UNBOUND) && max_active == 1)
4665 		flags |= __WQ_ORDERED;
4666 
4667 	/* see the comment above the definition of WQ_POWER_EFFICIENT */
4668 	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4669 		flags |= WQ_UNBOUND;
4670 
4671 	/* allocate wq and format name */
4672 	wq = kzalloc(sizeof(*wq), GFP_KERNEL);
4673 	if (!wq)
4674 		return NULL;
4675 
4676 	if (flags & WQ_UNBOUND) {
4677 		wq->unbound_attrs = alloc_workqueue_attrs();
4678 		if (!wq->unbound_attrs)
4679 			goto err_free_wq;
4680 	}
4681 
4682 	va_start(args, max_active);
4683 	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4684 	va_end(args);
4685 
4686 	max_active = max_active ?: WQ_DFL_ACTIVE;
4687 	max_active = wq_clamp_max_active(max_active, flags, wq->name);
4688 
4689 	/* init wq */
4690 	wq->flags = flags;
4691 	wq->saved_max_active = max_active;
4692 	mutex_init(&wq->mutex);
4693 	atomic_set(&wq->nr_pwqs_to_flush, 0);
4694 	INIT_LIST_HEAD(&wq->pwqs);
4695 	INIT_LIST_HEAD(&wq->flusher_queue);
4696 	INIT_LIST_HEAD(&wq->flusher_overflow);
4697 	INIT_LIST_HEAD(&wq->maydays);
4698 
4699 	wq_init_lockdep(wq);
4700 	INIT_LIST_HEAD(&wq->list);
4701 
4702 	if (alloc_and_link_pwqs(wq) < 0)
4703 		goto err_unreg_lockdep;
4704 
4705 	if (wq_online && init_rescuer(wq) < 0)
4706 		goto err_destroy;
4707 
4708 	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4709 		goto err_destroy;
4710 
4711 	/*
4712 	 * wq_pool_mutex protects global freeze state and workqueues list.
4713 	 * Grab it, adjust max_active and add the new @wq to workqueues
4714 	 * list.
4715 	 */
4716 	mutex_lock(&wq_pool_mutex);
4717 
4718 	mutex_lock(&wq->mutex);
4719 	for_each_pwq(pwq, wq)
4720 		pwq_adjust_max_active(pwq);
4721 	mutex_unlock(&wq->mutex);
4722 
4723 	list_add_tail_rcu(&wq->list, &workqueues);
4724 
4725 	mutex_unlock(&wq_pool_mutex);
4726 
4727 	return wq;
4728 
4729 err_unreg_lockdep:
4730 	wq_unregister_lockdep(wq);
4731 	wq_free_lockdep(wq);
4732 err_free_wq:
4733 	free_workqueue_attrs(wq->unbound_attrs);
4734 	kfree(wq);
4735 	return NULL;
4736 err_destroy:
4737 	destroy_workqueue(wq);
4738 	return NULL;
4739 }
4740 EXPORT_SYMBOL_GPL(alloc_workqueue);
4741 
4742 static bool pwq_busy(struct pool_workqueue *pwq)
4743 {
4744 	int i;
4745 
4746 	for (i = 0; i < WORK_NR_COLORS; i++)
4747 		if (pwq->nr_in_flight[i])
4748 			return true;
4749 
4750 	if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4751 		return true;
4752 	if (pwq->nr_active || !list_empty(&pwq->inactive_works))
4753 		return true;
4754 
4755 	return false;
4756 }
4757 
4758 /**
4759  * destroy_workqueue - safely terminate a workqueue
4760  * @wq: target workqueue
4761  *
4762  * Safely destroy a workqueue. All work currently pending will be done first.
4763  */
4764 void destroy_workqueue(struct workqueue_struct *wq)
4765 {
4766 	struct pool_workqueue *pwq;
4767 	int cpu;
4768 
4769 	/*
4770 	 * Remove it from sysfs first so that sanity check failure doesn't
4771 	 * lead to sysfs name conflicts.
4772 	 */
4773 	workqueue_sysfs_unregister(wq);
4774 
4775 	/* mark the workqueue destruction is in progress */
4776 	mutex_lock(&wq->mutex);
4777 	wq->flags |= __WQ_DESTROYING;
4778 	mutex_unlock(&wq->mutex);
4779 
4780 	/* drain it before proceeding with destruction */
4781 	drain_workqueue(wq);
4782 
4783 	/* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4784 	if (wq->rescuer) {
4785 		struct worker *rescuer = wq->rescuer;
4786 
4787 		/* this prevents new queueing */
4788 		raw_spin_lock_irq(&wq_mayday_lock);
4789 		wq->rescuer = NULL;
4790 		raw_spin_unlock_irq(&wq_mayday_lock);
4791 
4792 		/* rescuer will empty maydays list before exiting */
4793 		kthread_stop(rescuer->task);
4794 		kfree(rescuer);
4795 	}
4796 
4797 	/*
4798 	 * Sanity checks - grab all the locks so that we wait for all
4799 	 * in-flight operations which may do put_pwq().
4800 	 */
4801 	mutex_lock(&wq_pool_mutex);
4802 	mutex_lock(&wq->mutex);
4803 	for_each_pwq(pwq, wq) {
4804 		raw_spin_lock_irq(&pwq->pool->lock);
4805 		if (WARN_ON(pwq_busy(pwq))) {
4806 			pr_warn("%s: %s has the following busy pwq\n",
4807 				__func__, wq->name);
4808 			show_pwq(pwq);
4809 			raw_spin_unlock_irq(&pwq->pool->lock);
4810 			mutex_unlock(&wq->mutex);
4811 			mutex_unlock(&wq_pool_mutex);
4812 			show_one_workqueue(wq);
4813 			return;
4814 		}
4815 		raw_spin_unlock_irq(&pwq->pool->lock);
4816 	}
4817 	mutex_unlock(&wq->mutex);
4818 
4819 	/*
4820 	 * wq list is used to freeze wq, remove from list after
4821 	 * flushing is complete in case freeze races us.
4822 	 */
4823 	list_del_rcu(&wq->list);
4824 	mutex_unlock(&wq_pool_mutex);
4825 
4826 	/*
4827 	 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
4828 	 * to put the base refs. @wq will be auto-destroyed from the last
4829 	 * pwq_put. RCU read lock prevents @wq from going away from under us.
4830 	 */
4831 	rcu_read_lock();
4832 
4833 	for_each_possible_cpu(cpu) {
4834 		pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
4835 		RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL);
4836 		put_pwq_unlocked(pwq);
4837 	}
4838 
4839 	put_pwq_unlocked(wq->dfl_pwq);
4840 	wq->dfl_pwq = NULL;
4841 
4842 	rcu_read_unlock();
4843 }
4844 EXPORT_SYMBOL_GPL(destroy_workqueue);
4845 
4846 /**
4847  * workqueue_set_max_active - adjust max_active of a workqueue
4848  * @wq: target workqueue
4849  * @max_active: new max_active value.
4850  *
4851  * Set max_active of @wq to @max_active.
4852  *
4853  * CONTEXT:
4854  * Don't call from IRQ context.
4855  */
4856 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4857 {
4858 	struct pool_workqueue *pwq;
4859 
4860 	/* disallow meddling with max_active for ordered workqueues */
4861 	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4862 		return;
4863 
4864 	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4865 
4866 	mutex_lock(&wq->mutex);
4867 
4868 	wq->flags &= ~__WQ_ORDERED;
4869 	wq->saved_max_active = max_active;
4870 
4871 	for_each_pwq(pwq, wq)
4872 		pwq_adjust_max_active(pwq);
4873 
4874 	mutex_unlock(&wq->mutex);
4875 }
4876 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4877 
4878 /**
4879  * current_work - retrieve %current task's work struct
4880  *
4881  * Determine if %current task is a workqueue worker and what it's working on.
4882  * Useful to find out the context that the %current task is running in.
4883  *
4884  * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4885  */
4886 struct work_struct *current_work(void)
4887 {
4888 	struct worker *worker = current_wq_worker();
4889 
4890 	return worker ? worker->current_work : NULL;
4891 }
4892 EXPORT_SYMBOL(current_work);
4893 
4894 /**
4895  * current_is_workqueue_rescuer - is %current workqueue rescuer?
4896  *
4897  * Determine whether %current is a workqueue rescuer.  Can be used from
4898  * work functions to determine whether it's being run off the rescuer task.
4899  *
4900  * Return: %true if %current is a workqueue rescuer. %false otherwise.
4901  */
4902 bool current_is_workqueue_rescuer(void)
4903 {
4904 	struct worker *worker = current_wq_worker();
4905 
4906 	return worker && worker->rescue_wq;
4907 }
4908 
4909 /**
4910  * workqueue_congested - test whether a workqueue is congested
4911  * @cpu: CPU in question
4912  * @wq: target workqueue
4913  *
4914  * Test whether @wq's cpu workqueue for @cpu is congested.  There is
4915  * no synchronization around this function and the test result is
4916  * unreliable and only useful as advisory hints or for debugging.
4917  *
4918  * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4919  *
4920  * With the exception of ordered workqueues, all workqueues have per-cpu
4921  * pool_workqueues, each with its own congested state. A workqueue being
4922  * congested on one CPU doesn't mean that the workqueue is contested on any
4923  * other CPUs.
4924  *
4925  * Return:
4926  * %true if congested, %false otherwise.
4927  */
4928 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4929 {
4930 	struct pool_workqueue *pwq;
4931 	bool ret;
4932 
4933 	rcu_read_lock();
4934 	preempt_disable();
4935 
4936 	if (cpu == WORK_CPU_UNBOUND)
4937 		cpu = smp_processor_id();
4938 
4939 	pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
4940 	ret = !list_empty(&pwq->inactive_works);
4941 
4942 	preempt_enable();
4943 	rcu_read_unlock();
4944 
4945 	return ret;
4946 }
4947 EXPORT_SYMBOL_GPL(workqueue_congested);
4948 
4949 /**
4950  * work_busy - test whether a work is currently pending or running
4951  * @work: the work to be tested
4952  *
4953  * Test whether @work is currently pending or running.  There is no
4954  * synchronization around this function and the test result is
4955  * unreliable and only useful as advisory hints or for debugging.
4956  *
4957  * Return:
4958  * OR'd bitmask of WORK_BUSY_* bits.
4959  */
4960 unsigned int work_busy(struct work_struct *work)
4961 {
4962 	struct worker_pool *pool;
4963 	unsigned long flags;
4964 	unsigned int ret = 0;
4965 
4966 	if (work_pending(work))
4967 		ret |= WORK_BUSY_PENDING;
4968 
4969 	rcu_read_lock();
4970 	pool = get_work_pool(work);
4971 	if (pool) {
4972 		raw_spin_lock_irqsave(&pool->lock, flags);
4973 		if (find_worker_executing_work(pool, work))
4974 			ret |= WORK_BUSY_RUNNING;
4975 		raw_spin_unlock_irqrestore(&pool->lock, flags);
4976 	}
4977 	rcu_read_unlock();
4978 
4979 	return ret;
4980 }
4981 EXPORT_SYMBOL_GPL(work_busy);
4982 
4983 /**
4984  * set_worker_desc - set description for the current work item
4985  * @fmt: printf-style format string
4986  * @...: arguments for the format string
4987  *
4988  * This function can be called by a running work function to describe what
4989  * the work item is about.  If the worker task gets dumped, this
4990  * information will be printed out together to help debugging.  The
4991  * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4992  */
4993 void set_worker_desc(const char *fmt, ...)
4994 {
4995 	struct worker *worker = current_wq_worker();
4996 	va_list args;
4997 
4998 	if (worker) {
4999 		va_start(args, fmt);
5000 		vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
5001 		va_end(args);
5002 	}
5003 }
5004 EXPORT_SYMBOL_GPL(set_worker_desc);
5005 
5006 /**
5007  * print_worker_info - print out worker information and description
5008  * @log_lvl: the log level to use when printing
5009  * @task: target task
5010  *
5011  * If @task is a worker and currently executing a work item, print out the
5012  * name of the workqueue being serviced and worker description set with
5013  * set_worker_desc() by the currently executing work item.
5014  *
5015  * This function can be safely called on any task as long as the
5016  * task_struct itself is accessible.  While safe, this function isn't
5017  * synchronized and may print out mixups or garbages of limited length.
5018  */
5019 void print_worker_info(const char *log_lvl, struct task_struct *task)
5020 {
5021 	work_func_t *fn = NULL;
5022 	char name[WQ_NAME_LEN] = { };
5023 	char desc[WORKER_DESC_LEN] = { };
5024 	struct pool_workqueue *pwq = NULL;
5025 	struct workqueue_struct *wq = NULL;
5026 	struct worker *worker;
5027 
5028 	if (!(task->flags & PF_WQ_WORKER))
5029 		return;
5030 
5031 	/*
5032 	 * This function is called without any synchronization and @task
5033 	 * could be in any state.  Be careful with dereferences.
5034 	 */
5035 	worker = kthread_probe_data(task);
5036 
5037 	/*
5038 	 * Carefully copy the associated workqueue's workfn, name and desc.
5039 	 * Keep the original last '\0' in case the original is garbage.
5040 	 */
5041 	copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
5042 	copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
5043 	copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
5044 	copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
5045 	copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
5046 
5047 	if (fn || name[0] || desc[0]) {
5048 		printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
5049 		if (strcmp(name, desc))
5050 			pr_cont(" (%s)", desc);
5051 		pr_cont("\n");
5052 	}
5053 }
5054 
5055 static void pr_cont_pool_info(struct worker_pool *pool)
5056 {
5057 	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
5058 	if (pool->node != NUMA_NO_NODE)
5059 		pr_cont(" node=%d", pool->node);
5060 	pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
5061 }
5062 
5063 struct pr_cont_work_struct {
5064 	bool comma;
5065 	work_func_t func;
5066 	long ctr;
5067 };
5068 
5069 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
5070 {
5071 	if (!pcwsp->ctr)
5072 		goto out_record;
5073 	if (func == pcwsp->func) {
5074 		pcwsp->ctr++;
5075 		return;
5076 	}
5077 	if (pcwsp->ctr == 1)
5078 		pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
5079 	else
5080 		pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
5081 	pcwsp->ctr = 0;
5082 out_record:
5083 	if ((long)func == -1L)
5084 		return;
5085 	pcwsp->comma = comma;
5086 	pcwsp->func = func;
5087 	pcwsp->ctr = 1;
5088 }
5089 
5090 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
5091 {
5092 	if (work->func == wq_barrier_func) {
5093 		struct wq_barrier *barr;
5094 
5095 		barr = container_of(work, struct wq_barrier, work);
5096 
5097 		pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5098 		pr_cont("%s BAR(%d)", comma ? "," : "",
5099 			task_pid_nr(barr->task));
5100 	} else {
5101 		if (!comma)
5102 			pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5103 		pr_cont_work_flush(comma, work->func, pcwsp);
5104 	}
5105 }
5106 
5107 static void show_pwq(struct pool_workqueue *pwq)
5108 {
5109 	struct pr_cont_work_struct pcws = { .ctr = 0, };
5110 	struct worker_pool *pool = pwq->pool;
5111 	struct work_struct *work;
5112 	struct worker *worker;
5113 	bool has_in_flight = false, has_pending = false;
5114 	int bkt;
5115 
5116 	pr_info("  pwq %d:", pool->id);
5117 	pr_cont_pool_info(pool);
5118 
5119 	pr_cont(" active=%d/%d refcnt=%d%s\n",
5120 		pwq->nr_active, pwq->max_active, pwq->refcnt,
5121 		!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
5122 
5123 	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5124 		if (worker->current_pwq == pwq) {
5125 			has_in_flight = true;
5126 			break;
5127 		}
5128 	}
5129 	if (has_in_flight) {
5130 		bool comma = false;
5131 
5132 		pr_info("    in-flight:");
5133 		hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5134 			if (worker->current_pwq != pwq)
5135 				continue;
5136 
5137 			pr_cont("%s %d%s:%ps", comma ? "," : "",
5138 				task_pid_nr(worker->task),
5139 				worker->rescue_wq ? "(RESCUER)" : "",
5140 				worker->current_func);
5141 			list_for_each_entry(work, &worker->scheduled, entry)
5142 				pr_cont_work(false, work, &pcws);
5143 			pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5144 			comma = true;
5145 		}
5146 		pr_cont("\n");
5147 	}
5148 
5149 	list_for_each_entry(work, &pool->worklist, entry) {
5150 		if (get_work_pwq(work) == pwq) {
5151 			has_pending = true;
5152 			break;
5153 		}
5154 	}
5155 	if (has_pending) {
5156 		bool comma = false;
5157 
5158 		pr_info("    pending:");
5159 		list_for_each_entry(work, &pool->worklist, entry) {
5160 			if (get_work_pwq(work) != pwq)
5161 				continue;
5162 
5163 			pr_cont_work(comma, work, &pcws);
5164 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5165 		}
5166 		pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5167 		pr_cont("\n");
5168 	}
5169 
5170 	if (!list_empty(&pwq->inactive_works)) {
5171 		bool comma = false;
5172 
5173 		pr_info("    inactive:");
5174 		list_for_each_entry(work, &pwq->inactive_works, entry) {
5175 			pr_cont_work(comma, work, &pcws);
5176 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5177 		}
5178 		pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5179 		pr_cont("\n");
5180 	}
5181 }
5182 
5183 /**
5184  * show_one_workqueue - dump state of specified workqueue
5185  * @wq: workqueue whose state will be printed
5186  */
5187 void show_one_workqueue(struct workqueue_struct *wq)
5188 {
5189 	struct pool_workqueue *pwq;
5190 	bool idle = true;
5191 	unsigned long flags;
5192 
5193 	for_each_pwq(pwq, wq) {
5194 		if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
5195 			idle = false;
5196 			break;
5197 		}
5198 	}
5199 	if (idle) /* Nothing to print for idle workqueue */
5200 		return;
5201 
5202 	pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
5203 
5204 	for_each_pwq(pwq, wq) {
5205 		raw_spin_lock_irqsave(&pwq->pool->lock, flags);
5206 		if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
5207 			/*
5208 			 * Defer printing to avoid deadlocks in console
5209 			 * drivers that queue work while holding locks
5210 			 * also taken in their write paths.
5211 			 */
5212 			printk_deferred_enter();
5213 			show_pwq(pwq);
5214 			printk_deferred_exit();
5215 		}
5216 		raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
5217 		/*
5218 		 * We could be printing a lot from atomic context, e.g.
5219 		 * sysrq-t -> show_all_workqueues(). Avoid triggering
5220 		 * hard lockup.
5221 		 */
5222 		touch_nmi_watchdog();
5223 	}
5224 
5225 }
5226 
5227 /**
5228  * show_one_worker_pool - dump state of specified worker pool
5229  * @pool: worker pool whose state will be printed
5230  */
5231 static void show_one_worker_pool(struct worker_pool *pool)
5232 {
5233 	struct worker *worker;
5234 	bool first = true;
5235 	unsigned long flags;
5236 	unsigned long hung = 0;
5237 
5238 	raw_spin_lock_irqsave(&pool->lock, flags);
5239 	if (pool->nr_workers == pool->nr_idle)
5240 		goto next_pool;
5241 
5242 	/* How long the first pending work is waiting for a worker. */
5243 	if (!list_empty(&pool->worklist))
5244 		hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
5245 
5246 	/*
5247 	 * Defer printing to avoid deadlocks in console drivers that
5248 	 * queue work while holding locks also taken in their write
5249 	 * paths.
5250 	 */
5251 	printk_deferred_enter();
5252 	pr_info("pool %d:", pool->id);
5253 	pr_cont_pool_info(pool);
5254 	pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
5255 	if (pool->manager)
5256 		pr_cont(" manager: %d",
5257 			task_pid_nr(pool->manager->task));
5258 	list_for_each_entry(worker, &pool->idle_list, entry) {
5259 		pr_cont(" %s%d", first ? "idle: " : "",
5260 			task_pid_nr(worker->task));
5261 		first = false;
5262 	}
5263 	pr_cont("\n");
5264 	printk_deferred_exit();
5265 next_pool:
5266 	raw_spin_unlock_irqrestore(&pool->lock, flags);
5267 	/*
5268 	 * We could be printing a lot from atomic context, e.g.
5269 	 * sysrq-t -> show_all_workqueues(). Avoid triggering
5270 	 * hard lockup.
5271 	 */
5272 	touch_nmi_watchdog();
5273 
5274 }
5275 
5276 /**
5277  * show_all_workqueues - dump workqueue state
5278  *
5279  * Called from a sysrq handler and prints out all busy workqueues and pools.
5280  */
5281 void show_all_workqueues(void)
5282 {
5283 	struct workqueue_struct *wq;
5284 	struct worker_pool *pool;
5285 	int pi;
5286 
5287 	rcu_read_lock();
5288 
5289 	pr_info("Showing busy workqueues and worker pools:\n");
5290 
5291 	list_for_each_entry_rcu(wq, &workqueues, list)
5292 		show_one_workqueue(wq);
5293 
5294 	for_each_pool(pool, pi)
5295 		show_one_worker_pool(pool);
5296 
5297 	rcu_read_unlock();
5298 }
5299 
5300 /**
5301  * show_freezable_workqueues - dump freezable workqueue state
5302  *
5303  * Called from try_to_freeze_tasks() and prints out all freezable workqueues
5304  * still busy.
5305  */
5306 void show_freezable_workqueues(void)
5307 {
5308 	struct workqueue_struct *wq;
5309 
5310 	rcu_read_lock();
5311 
5312 	pr_info("Showing freezable workqueues that are still busy:\n");
5313 
5314 	list_for_each_entry_rcu(wq, &workqueues, list) {
5315 		if (!(wq->flags & WQ_FREEZABLE))
5316 			continue;
5317 		show_one_workqueue(wq);
5318 	}
5319 
5320 	rcu_read_unlock();
5321 }
5322 
5323 /* used to show worker information through /proc/PID/{comm,stat,status} */
5324 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
5325 {
5326 	int off;
5327 
5328 	/* always show the actual comm */
5329 	off = strscpy(buf, task->comm, size);
5330 	if (off < 0)
5331 		return;
5332 
5333 	/* stabilize PF_WQ_WORKER and worker pool association */
5334 	mutex_lock(&wq_pool_attach_mutex);
5335 
5336 	if (task->flags & PF_WQ_WORKER) {
5337 		struct worker *worker = kthread_data(task);
5338 		struct worker_pool *pool = worker->pool;
5339 
5340 		if (pool) {
5341 			raw_spin_lock_irq(&pool->lock);
5342 			/*
5343 			 * ->desc tracks information (wq name or
5344 			 * set_worker_desc()) for the latest execution.  If
5345 			 * current, prepend '+', otherwise '-'.
5346 			 */
5347 			if (worker->desc[0] != '\0') {
5348 				if (worker->current_work)
5349 					scnprintf(buf + off, size - off, "+%s",
5350 						  worker->desc);
5351 				else
5352 					scnprintf(buf + off, size - off, "-%s",
5353 						  worker->desc);
5354 			}
5355 			raw_spin_unlock_irq(&pool->lock);
5356 		}
5357 	}
5358 
5359 	mutex_unlock(&wq_pool_attach_mutex);
5360 }
5361 
5362 #ifdef CONFIG_SMP
5363 
5364 /*
5365  * CPU hotplug.
5366  *
5367  * There are two challenges in supporting CPU hotplug.  Firstly, there
5368  * are a lot of assumptions on strong associations among work, pwq and
5369  * pool which make migrating pending and scheduled works very
5370  * difficult to implement without impacting hot paths.  Secondly,
5371  * worker pools serve mix of short, long and very long running works making
5372  * blocked draining impractical.
5373  *
5374  * This is solved by allowing the pools to be disassociated from the CPU
5375  * running as an unbound one and allowing it to be reattached later if the
5376  * cpu comes back online.
5377  */
5378 
5379 static void unbind_workers(int cpu)
5380 {
5381 	struct worker_pool *pool;
5382 	struct worker *worker;
5383 
5384 	for_each_cpu_worker_pool(pool, cpu) {
5385 		mutex_lock(&wq_pool_attach_mutex);
5386 		raw_spin_lock_irq(&pool->lock);
5387 
5388 		/*
5389 		 * We've blocked all attach/detach operations. Make all workers
5390 		 * unbound and set DISASSOCIATED.  Before this, all workers
5391 		 * must be on the cpu.  After this, they may become diasporas.
5392 		 * And the preemption disabled section in their sched callbacks
5393 		 * are guaranteed to see WORKER_UNBOUND since the code here
5394 		 * is on the same cpu.
5395 		 */
5396 		for_each_pool_worker(worker, pool)
5397 			worker->flags |= WORKER_UNBOUND;
5398 
5399 		pool->flags |= POOL_DISASSOCIATED;
5400 
5401 		/*
5402 		 * The handling of nr_running in sched callbacks are disabled
5403 		 * now.  Zap nr_running.  After this, nr_running stays zero and
5404 		 * need_more_worker() and keep_working() are always true as
5405 		 * long as the worklist is not empty.  This pool now behaves as
5406 		 * an unbound (in terms of concurrency management) pool which
5407 		 * are served by workers tied to the pool.
5408 		 */
5409 		pool->nr_running = 0;
5410 
5411 		/*
5412 		 * With concurrency management just turned off, a busy
5413 		 * worker blocking could lead to lengthy stalls.  Kick off
5414 		 * unbound chain execution of currently pending work items.
5415 		 */
5416 		kick_pool(pool);
5417 
5418 		raw_spin_unlock_irq(&pool->lock);
5419 
5420 		for_each_pool_worker(worker, pool)
5421 			unbind_worker(worker);
5422 
5423 		mutex_unlock(&wq_pool_attach_mutex);
5424 	}
5425 }
5426 
5427 /**
5428  * rebind_workers - rebind all workers of a pool to the associated CPU
5429  * @pool: pool of interest
5430  *
5431  * @pool->cpu is coming online.  Rebind all workers to the CPU.
5432  */
5433 static void rebind_workers(struct worker_pool *pool)
5434 {
5435 	struct worker *worker;
5436 
5437 	lockdep_assert_held(&wq_pool_attach_mutex);
5438 
5439 	/*
5440 	 * Restore CPU affinity of all workers.  As all idle workers should
5441 	 * be on the run-queue of the associated CPU before any local
5442 	 * wake-ups for concurrency management happen, restore CPU affinity
5443 	 * of all workers first and then clear UNBOUND.  As we're called
5444 	 * from CPU_ONLINE, the following shouldn't fail.
5445 	 */
5446 	for_each_pool_worker(worker, pool) {
5447 		kthread_set_per_cpu(worker->task, pool->cpu);
5448 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
5449 						  pool_allowed_cpus(pool)) < 0);
5450 	}
5451 
5452 	raw_spin_lock_irq(&pool->lock);
5453 
5454 	pool->flags &= ~POOL_DISASSOCIATED;
5455 
5456 	for_each_pool_worker(worker, pool) {
5457 		unsigned int worker_flags = worker->flags;
5458 
5459 		/*
5460 		 * We want to clear UNBOUND but can't directly call
5461 		 * worker_clr_flags() or adjust nr_running.  Atomically
5462 		 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5463 		 * @worker will clear REBOUND using worker_clr_flags() when
5464 		 * it initiates the next execution cycle thus restoring
5465 		 * concurrency management.  Note that when or whether
5466 		 * @worker clears REBOUND doesn't affect correctness.
5467 		 *
5468 		 * WRITE_ONCE() is necessary because @worker->flags may be
5469 		 * tested without holding any lock in
5470 		 * wq_worker_running().  Without it, NOT_RUNNING test may
5471 		 * fail incorrectly leading to premature concurrency
5472 		 * management operations.
5473 		 */
5474 		WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
5475 		worker_flags |= WORKER_REBOUND;
5476 		worker_flags &= ~WORKER_UNBOUND;
5477 		WRITE_ONCE(worker->flags, worker_flags);
5478 	}
5479 
5480 	raw_spin_unlock_irq(&pool->lock);
5481 }
5482 
5483 /**
5484  * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5485  * @pool: unbound pool of interest
5486  * @cpu: the CPU which is coming up
5487  *
5488  * An unbound pool may end up with a cpumask which doesn't have any online
5489  * CPUs.  When a worker of such pool get scheduled, the scheduler resets
5490  * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
5491  * online CPU before, cpus_allowed of all its workers should be restored.
5492  */
5493 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5494 {
5495 	static cpumask_t cpumask;
5496 	struct worker *worker;
5497 
5498 	lockdep_assert_held(&wq_pool_attach_mutex);
5499 
5500 	/* is @cpu allowed for @pool? */
5501 	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5502 		return;
5503 
5504 	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5505 
5506 	/* as we're called from CPU_ONLINE, the following shouldn't fail */
5507 	for_each_pool_worker(worker, pool)
5508 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5509 }
5510 
5511 int workqueue_prepare_cpu(unsigned int cpu)
5512 {
5513 	struct worker_pool *pool;
5514 
5515 	for_each_cpu_worker_pool(pool, cpu) {
5516 		if (pool->nr_workers)
5517 			continue;
5518 		if (!create_worker(pool))
5519 			return -ENOMEM;
5520 	}
5521 	return 0;
5522 }
5523 
5524 int workqueue_online_cpu(unsigned int cpu)
5525 {
5526 	struct worker_pool *pool;
5527 	struct workqueue_struct *wq;
5528 	int pi;
5529 
5530 	mutex_lock(&wq_pool_mutex);
5531 
5532 	for_each_pool(pool, pi) {
5533 		mutex_lock(&wq_pool_attach_mutex);
5534 
5535 		if (pool->cpu == cpu)
5536 			rebind_workers(pool);
5537 		else if (pool->cpu < 0)
5538 			restore_unbound_workers_cpumask(pool, cpu);
5539 
5540 		mutex_unlock(&wq_pool_attach_mutex);
5541 	}
5542 
5543 	/* update pod affinity of unbound workqueues */
5544 	list_for_each_entry(wq, &workqueues, list) {
5545 		struct workqueue_attrs *attrs = wq->unbound_attrs;
5546 
5547 		if (attrs) {
5548 			const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5549 			int tcpu;
5550 
5551 			for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
5552 				wq_update_pod(wq, tcpu, cpu, true);
5553 		}
5554 	}
5555 
5556 	mutex_unlock(&wq_pool_mutex);
5557 	return 0;
5558 }
5559 
5560 int workqueue_offline_cpu(unsigned int cpu)
5561 {
5562 	struct workqueue_struct *wq;
5563 
5564 	/* unbinding per-cpu workers should happen on the local CPU */
5565 	if (WARN_ON(cpu != smp_processor_id()))
5566 		return -1;
5567 
5568 	unbind_workers(cpu);
5569 
5570 	/* update pod affinity of unbound workqueues */
5571 	mutex_lock(&wq_pool_mutex);
5572 	list_for_each_entry(wq, &workqueues, list) {
5573 		struct workqueue_attrs *attrs = wq->unbound_attrs;
5574 
5575 		if (attrs) {
5576 			const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5577 			int tcpu;
5578 
5579 			for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
5580 				wq_update_pod(wq, tcpu, cpu, false);
5581 		}
5582 	}
5583 	mutex_unlock(&wq_pool_mutex);
5584 
5585 	return 0;
5586 }
5587 
5588 struct work_for_cpu {
5589 	struct work_struct work;
5590 	long (*fn)(void *);
5591 	void *arg;
5592 	long ret;
5593 };
5594 
5595 static void work_for_cpu_fn(struct work_struct *work)
5596 {
5597 	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5598 
5599 	wfc->ret = wfc->fn(wfc->arg);
5600 }
5601 
5602 /**
5603  * work_on_cpu - run a function in thread context on a particular cpu
5604  * @cpu: the cpu to run on
5605  * @fn: the function to run
5606  * @arg: the function arg
5607  *
5608  * It is up to the caller to ensure that the cpu doesn't go offline.
5609  * The caller must not hold any locks which would prevent @fn from completing.
5610  *
5611  * Return: The value @fn returns.
5612  */
5613 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5614 {
5615 	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5616 
5617 	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5618 	schedule_work_on(cpu, &wfc.work);
5619 	flush_work(&wfc.work);
5620 	destroy_work_on_stack(&wfc.work);
5621 	return wfc.ret;
5622 }
5623 EXPORT_SYMBOL_GPL(work_on_cpu);
5624 
5625 /**
5626  * work_on_cpu_safe - run a function in thread context on a particular cpu
5627  * @cpu: the cpu to run on
5628  * @fn:  the function to run
5629  * @arg: the function argument
5630  *
5631  * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5632  * any locks which would prevent @fn from completing.
5633  *
5634  * Return: The value @fn returns.
5635  */
5636 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5637 {
5638 	long ret = -ENODEV;
5639 
5640 	cpus_read_lock();
5641 	if (cpu_online(cpu))
5642 		ret = work_on_cpu(cpu, fn, arg);
5643 	cpus_read_unlock();
5644 	return ret;
5645 }
5646 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5647 #endif /* CONFIG_SMP */
5648 
5649 #ifdef CONFIG_FREEZER
5650 
5651 /**
5652  * freeze_workqueues_begin - begin freezing workqueues
5653  *
5654  * Start freezing workqueues.  After this function returns, all freezable
5655  * workqueues will queue new works to their inactive_works list instead of
5656  * pool->worklist.
5657  *
5658  * CONTEXT:
5659  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5660  */
5661 void freeze_workqueues_begin(void)
5662 {
5663 	struct workqueue_struct *wq;
5664 	struct pool_workqueue *pwq;
5665 
5666 	mutex_lock(&wq_pool_mutex);
5667 
5668 	WARN_ON_ONCE(workqueue_freezing);
5669 	workqueue_freezing = true;
5670 
5671 	list_for_each_entry(wq, &workqueues, list) {
5672 		mutex_lock(&wq->mutex);
5673 		for_each_pwq(pwq, wq)
5674 			pwq_adjust_max_active(pwq);
5675 		mutex_unlock(&wq->mutex);
5676 	}
5677 
5678 	mutex_unlock(&wq_pool_mutex);
5679 }
5680 
5681 /**
5682  * freeze_workqueues_busy - are freezable workqueues still busy?
5683  *
5684  * Check whether freezing is complete.  This function must be called
5685  * between freeze_workqueues_begin() and thaw_workqueues().
5686  *
5687  * CONTEXT:
5688  * Grabs and releases wq_pool_mutex.
5689  *
5690  * Return:
5691  * %true if some freezable workqueues are still busy.  %false if freezing
5692  * is complete.
5693  */
5694 bool freeze_workqueues_busy(void)
5695 {
5696 	bool busy = false;
5697 	struct workqueue_struct *wq;
5698 	struct pool_workqueue *pwq;
5699 
5700 	mutex_lock(&wq_pool_mutex);
5701 
5702 	WARN_ON_ONCE(!workqueue_freezing);
5703 
5704 	list_for_each_entry(wq, &workqueues, list) {
5705 		if (!(wq->flags & WQ_FREEZABLE))
5706 			continue;
5707 		/*
5708 		 * nr_active is monotonically decreasing.  It's safe
5709 		 * to peek without lock.
5710 		 */
5711 		rcu_read_lock();
5712 		for_each_pwq(pwq, wq) {
5713 			WARN_ON_ONCE(pwq->nr_active < 0);
5714 			if (pwq->nr_active) {
5715 				busy = true;
5716 				rcu_read_unlock();
5717 				goto out_unlock;
5718 			}
5719 		}
5720 		rcu_read_unlock();
5721 	}
5722 out_unlock:
5723 	mutex_unlock(&wq_pool_mutex);
5724 	return busy;
5725 }
5726 
5727 /**
5728  * thaw_workqueues - thaw workqueues
5729  *
5730  * Thaw workqueues.  Normal queueing is restored and all collected
5731  * frozen works are transferred to their respective pool worklists.
5732  *
5733  * CONTEXT:
5734  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5735  */
5736 void thaw_workqueues(void)
5737 {
5738 	struct workqueue_struct *wq;
5739 	struct pool_workqueue *pwq;
5740 
5741 	mutex_lock(&wq_pool_mutex);
5742 
5743 	if (!workqueue_freezing)
5744 		goto out_unlock;
5745 
5746 	workqueue_freezing = false;
5747 
5748 	/* restore max_active and repopulate worklist */
5749 	list_for_each_entry(wq, &workqueues, list) {
5750 		mutex_lock(&wq->mutex);
5751 		for_each_pwq(pwq, wq)
5752 			pwq_adjust_max_active(pwq);
5753 		mutex_unlock(&wq->mutex);
5754 	}
5755 
5756 out_unlock:
5757 	mutex_unlock(&wq_pool_mutex);
5758 }
5759 #endif /* CONFIG_FREEZER */
5760 
5761 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
5762 {
5763 	LIST_HEAD(ctxs);
5764 	int ret = 0;
5765 	struct workqueue_struct *wq;
5766 	struct apply_wqattrs_ctx *ctx, *n;
5767 
5768 	lockdep_assert_held(&wq_pool_mutex);
5769 
5770 	list_for_each_entry(wq, &workqueues, list) {
5771 		if (!(wq->flags & WQ_UNBOUND))
5772 			continue;
5773 		/* creating multiple pwqs breaks ordering guarantee */
5774 		if (wq->flags & __WQ_ORDERED)
5775 			continue;
5776 
5777 		ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
5778 		if (IS_ERR(ctx)) {
5779 			ret = PTR_ERR(ctx);
5780 			break;
5781 		}
5782 
5783 		list_add_tail(&ctx->list, &ctxs);
5784 	}
5785 
5786 	list_for_each_entry_safe(ctx, n, &ctxs, list) {
5787 		if (!ret)
5788 			apply_wqattrs_commit(ctx);
5789 		apply_wqattrs_cleanup(ctx);
5790 	}
5791 
5792 	if (!ret) {
5793 		mutex_lock(&wq_pool_attach_mutex);
5794 		cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
5795 		mutex_unlock(&wq_pool_attach_mutex);
5796 	}
5797 	return ret;
5798 }
5799 
5800 /**
5801  *  workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5802  *  @cpumask: the cpumask to set
5803  *
5804  *  The low-level workqueues cpumask is a global cpumask that limits
5805  *  the affinity of all unbound workqueues.  This function check the @cpumask
5806  *  and apply it to all unbound workqueues and updates all pwqs of them.
5807  *
5808  *  Return:	0	- Success
5809  *  		-EINVAL	- Invalid @cpumask
5810  *  		-ENOMEM	- Failed to allocate memory for attrs or pwqs.
5811  */
5812 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5813 {
5814 	int ret = -EINVAL;
5815 
5816 	/*
5817 	 * Not excluding isolated cpus on purpose.
5818 	 * If the user wishes to include them, we allow that.
5819 	 */
5820 	cpumask_and(cpumask, cpumask, cpu_possible_mask);
5821 	if (!cpumask_empty(cpumask)) {
5822 		apply_wqattrs_lock();
5823 		if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
5824 			ret = 0;
5825 			goto out_unlock;
5826 		}
5827 
5828 		ret = workqueue_apply_unbound_cpumask(cpumask);
5829 
5830 out_unlock:
5831 		apply_wqattrs_unlock();
5832 	}
5833 
5834 	return ret;
5835 }
5836 
5837 static int parse_affn_scope(const char *val)
5838 {
5839 	int i;
5840 
5841 	for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
5842 		if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
5843 			return i;
5844 	}
5845 	return -EINVAL;
5846 }
5847 
5848 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
5849 {
5850 	int affn;
5851 
5852 	affn = parse_affn_scope(val);
5853 	if (affn < 0)
5854 		return affn;
5855 
5856 	wq_affn_dfl = affn;
5857 	return 0;
5858 }
5859 
5860 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
5861 {
5862 	return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
5863 }
5864 
5865 static const struct kernel_param_ops wq_affn_dfl_ops = {
5866 	.set	= wq_affn_dfl_set,
5867 	.get	= wq_affn_dfl_get,
5868 };
5869 
5870 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
5871 
5872 #ifdef CONFIG_SYSFS
5873 /*
5874  * Workqueues with WQ_SYSFS flag set is visible to userland via
5875  * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
5876  * following attributes.
5877  *
5878  *  per_cpu		RO bool	: whether the workqueue is per-cpu or unbound
5879  *  max_active		RW int	: maximum number of in-flight work items
5880  *
5881  * Unbound workqueues have the following extra attributes.
5882  *
5883  *  nice		RW int	: nice value of the workers
5884  *  cpumask		RW mask	: bitmask of allowed CPUs for the workers
5885  *  affinity_scope	RW str  : worker CPU affinity scope (cache, numa, none)
5886  *  affinity_strict	RW bool : worker CPU affinity is strict
5887  */
5888 struct wq_device {
5889 	struct workqueue_struct		*wq;
5890 	struct device			dev;
5891 };
5892 
5893 static struct workqueue_struct *dev_to_wq(struct device *dev)
5894 {
5895 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5896 
5897 	return wq_dev->wq;
5898 }
5899 
5900 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5901 			    char *buf)
5902 {
5903 	struct workqueue_struct *wq = dev_to_wq(dev);
5904 
5905 	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5906 }
5907 static DEVICE_ATTR_RO(per_cpu);
5908 
5909 static ssize_t max_active_show(struct device *dev,
5910 			       struct device_attribute *attr, char *buf)
5911 {
5912 	struct workqueue_struct *wq = dev_to_wq(dev);
5913 
5914 	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5915 }
5916 
5917 static ssize_t max_active_store(struct device *dev,
5918 				struct device_attribute *attr, const char *buf,
5919 				size_t count)
5920 {
5921 	struct workqueue_struct *wq = dev_to_wq(dev);
5922 	int val;
5923 
5924 	if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5925 		return -EINVAL;
5926 
5927 	workqueue_set_max_active(wq, val);
5928 	return count;
5929 }
5930 static DEVICE_ATTR_RW(max_active);
5931 
5932 static struct attribute *wq_sysfs_attrs[] = {
5933 	&dev_attr_per_cpu.attr,
5934 	&dev_attr_max_active.attr,
5935 	NULL,
5936 };
5937 ATTRIBUTE_GROUPS(wq_sysfs);
5938 
5939 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5940 			    char *buf)
5941 {
5942 	struct workqueue_struct *wq = dev_to_wq(dev);
5943 	int written;
5944 
5945 	mutex_lock(&wq->mutex);
5946 	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5947 	mutex_unlock(&wq->mutex);
5948 
5949 	return written;
5950 }
5951 
5952 /* prepare workqueue_attrs for sysfs store operations */
5953 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5954 {
5955 	struct workqueue_attrs *attrs;
5956 
5957 	lockdep_assert_held(&wq_pool_mutex);
5958 
5959 	attrs = alloc_workqueue_attrs();
5960 	if (!attrs)
5961 		return NULL;
5962 
5963 	copy_workqueue_attrs(attrs, wq->unbound_attrs);
5964 	return attrs;
5965 }
5966 
5967 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5968 			     const char *buf, size_t count)
5969 {
5970 	struct workqueue_struct *wq = dev_to_wq(dev);
5971 	struct workqueue_attrs *attrs;
5972 	int ret = -ENOMEM;
5973 
5974 	apply_wqattrs_lock();
5975 
5976 	attrs = wq_sysfs_prep_attrs(wq);
5977 	if (!attrs)
5978 		goto out_unlock;
5979 
5980 	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5981 	    attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5982 		ret = apply_workqueue_attrs_locked(wq, attrs);
5983 	else
5984 		ret = -EINVAL;
5985 
5986 out_unlock:
5987 	apply_wqattrs_unlock();
5988 	free_workqueue_attrs(attrs);
5989 	return ret ?: count;
5990 }
5991 
5992 static ssize_t wq_cpumask_show(struct device *dev,
5993 			       struct device_attribute *attr, char *buf)
5994 {
5995 	struct workqueue_struct *wq = dev_to_wq(dev);
5996 	int written;
5997 
5998 	mutex_lock(&wq->mutex);
5999 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6000 			    cpumask_pr_args(wq->unbound_attrs->cpumask));
6001 	mutex_unlock(&wq->mutex);
6002 	return written;
6003 }
6004 
6005 static ssize_t wq_cpumask_store(struct device *dev,
6006 				struct device_attribute *attr,
6007 				const char *buf, size_t count)
6008 {
6009 	struct workqueue_struct *wq = dev_to_wq(dev);
6010 	struct workqueue_attrs *attrs;
6011 	int ret = -ENOMEM;
6012 
6013 	apply_wqattrs_lock();
6014 
6015 	attrs = wq_sysfs_prep_attrs(wq);
6016 	if (!attrs)
6017 		goto out_unlock;
6018 
6019 	ret = cpumask_parse(buf, attrs->cpumask);
6020 	if (!ret)
6021 		ret = apply_workqueue_attrs_locked(wq, attrs);
6022 
6023 out_unlock:
6024 	apply_wqattrs_unlock();
6025 	free_workqueue_attrs(attrs);
6026 	return ret ?: count;
6027 }
6028 
6029 static ssize_t wq_affn_scope_show(struct device *dev,
6030 				  struct device_attribute *attr, char *buf)
6031 {
6032 	struct workqueue_struct *wq = dev_to_wq(dev);
6033 	int written;
6034 
6035 	mutex_lock(&wq->mutex);
6036 	written = scnprintf(buf, PAGE_SIZE, "%s\n",
6037 			    wq_affn_names[wq->unbound_attrs->affn_scope]);
6038 	mutex_unlock(&wq->mutex);
6039 
6040 	return written;
6041 }
6042 
6043 static ssize_t wq_affn_scope_store(struct device *dev,
6044 				   struct device_attribute *attr,
6045 				   const char *buf, size_t count)
6046 {
6047 	struct workqueue_struct *wq = dev_to_wq(dev);
6048 	struct workqueue_attrs *attrs;
6049 	int affn, ret = -ENOMEM;
6050 
6051 	affn = parse_affn_scope(buf);
6052 	if (affn < 0)
6053 		return affn;
6054 
6055 	apply_wqattrs_lock();
6056 	attrs = wq_sysfs_prep_attrs(wq);
6057 	if (attrs) {
6058 		attrs->affn_scope = affn;
6059 		ret = apply_workqueue_attrs_locked(wq, attrs);
6060 	}
6061 	apply_wqattrs_unlock();
6062 	free_workqueue_attrs(attrs);
6063 	return ret ?: count;
6064 }
6065 
6066 static ssize_t wq_affinity_strict_show(struct device *dev,
6067 				       struct device_attribute *attr, char *buf)
6068 {
6069 	struct workqueue_struct *wq = dev_to_wq(dev);
6070 
6071 	return scnprintf(buf, PAGE_SIZE, "%d\n",
6072 			 wq->unbound_attrs->affn_strict);
6073 }
6074 
6075 static ssize_t wq_affinity_strict_store(struct device *dev,
6076 					struct device_attribute *attr,
6077 					const char *buf, size_t count)
6078 {
6079 	struct workqueue_struct *wq = dev_to_wq(dev);
6080 	struct workqueue_attrs *attrs;
6081 	int v, ret = -ENOMEM;
6082 
6083 	if (sscanf(buf, "%d", &v) != 1)
6084 		return -EINVAL;
6085 
6086 	apply_wqattrs_lock();
6087 	attrs = wq_sysfs_prep_attrs(wq);
6088 	if (attrs) {
6089 		attrs->affn_strict = (bool)v;
6090 		ret = apply_workqueue_attrs_locked(wq, attrs);
6091 	}
6092 	apply_wqattrs_unlock();
6093 	free_workqueue_attrs(attrs);
6094 	return ret ?: count;
6095 }
6096 
6097 static struct device_attribute wq_sysfs_unbound_attrs[] = {
6098 	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
6099 	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
6100 	__ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
6101 	__ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
6102 	__ATTR_NULL,
6103 };
6104 
6105 static struct bus_type wq_subsys = {
6106 	.name				= "workqueue",
6107 	.dev_groups			= wq_sysfs_groups,
6108 };
6109 
6110 static ssize_t wq_unbound_cpumask_show(struct device *dev,
6111 		struct device_attribute *attr, char *buf)
6112 {
6113 	int written;
6114 
6115 	mutex_lock(&wq_pool_mutex);
6116 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6117 			    cpumask_pr_args(wq_unbound_cpumask));
6118 	mutex_unlock(&wq_pool_mutex);
6119 
6120 	return written;
6121 }
6122 
6123 static ssize_t wq_unbound_cpumask_store(struct device *dev,
6124 		struct device_attribute *attr, const char *buf, size_t count)
6125 {
6126 	cpumask_var_t cpumask;
6127 	int ret;
6128 
6129 	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6130 		return -ENOMEM;
6131 
6132 	ret = cpumask_parse(buf, cpumask);
6133 	if (!ret)
6134 		ret = workqueue_set_unbound_cpumask(cpumask);
6135 
6136 	free_cpumask_var(cpumask);
6137 	return ret ? ret : count;
6138 }
6139 
6140 static struct device_attribute wq_sysfs_cpumask_attr =
6141 	__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
6142 	       wq_unbound_cpumask_store);
6143 
6144 static int __init wq_sysfs_init(void)
6145 {
6146 	struct device *dev_root;
6147 	int err;
6148 
6149 	err = subsys_virtual_register(&wq_subsys, NULL);
6150 	if (err)
6151 		return err;
6152 
6153 	dev_root = bus_get_dev_root(&wq_subsys);
6154 	if (dev_root) {
6155 		err = device_create_file(dev_root, &wq_sysfs_cpumask_attr);
6156 		put_device(dev_root);
6157 	}
6158 	return err;
6159 }
6160 core_initcall(wq_sysfs_init);
6161 
6162 static void wq_device_release(struct device *dev)
6163 {
6164 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6165 
6166 	kfree(wq_dev);
6167 }
6168 
6169 /**
6170  * workqueue_sysfs_register - make a workqueue visible in sysfs
6171  * @wq: the workqueue to register
6172  *
6173  * Expose @wq in sysfs under /sys/bus/workqueue/devices.
6174  * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
6175  * which is the preferred method.
6176  *
6177  * Workqueue user should use this function directly iff it wants to apply
6178  * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
6179  * apply_workqueue_attrs() may race against userland updating the
6180  * attributes.
6181  *
6182  * Return: 0 on success, -errno on failure.
6183  */
6184 int workqueue_sysfs_register(struct workqueue_struct *wq)
6185 {
6186 	struct wq_device *wq_dev;
6187 	int ret;
6188 
6189 	/*
6190 	 * Adjusting max_active or creating new pwqs by applying
6191 	 * attributes breaks ordering guarantee.  Disallow exposing ordered
6192 	 * workqueues.
6193 	 */
6194 	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
6195 		return -EINVAL;
6196 
6197 	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
6198 	if (!wq_dev)
6199 		return -ENOMEM;
6200 
6201 	wq_dev->wq = wq;
6202 	wq_dev->dev.bus = &wq_subsys;
6203 	wq_dev->dev.release = wq_device_release;
6204 	dev_set_name(&wq_dev->dev, "%s", wq->name);
6205 
6206 	/*
6207 	 * unbound_attrs are created separately.  Suppress uevent until
6208 	 * everything is ready.
6209 	 */
6210 	dev_set_uevent_suppress(&wq_dev->dev, true);
6211 
6212 	ret = device_register(&wq_dev->dev);
6213 	if (ret) {
6214 		put_device(&wq_dev->dev);
6215 		wq->wq_dev = NULL;
6216 		return ret;
6217 	}
6218 
6219 	if (wq->flags & WQ_UNBOUND) {
6220 		struct device_attribute *attr;
6221 
6222 		for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
6223 			ret = device_create_file(&wq_dev->dev, attr);
6224 			if (ret) {
6225 				device_unregister(&wq_dev->dev);
6226 				wq->wq_dev = NULL;
6227 				return ret;
6228 			}
6229 		}
6230 	}
6231 
6232 	dev_set_uevent_suppress(&wq_dev->dev, false);
6233 	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
6234 	return 0;
6235 }
6236 
6237 /**
6238  * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
6239  * @wq: the workqueue to unregister
6240  *
6241  * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
6242  */
6243 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
6244 {
6245 	struct wq_device *wq_dev = wq->wq_dev;
6246 
6247 	if (!wq->wq_dev)
6248 		return;
6249 
6250 	wq->wq_dev = NULL;
6251 	device_unregister(&wq_dev->dev);
6252 }
6253 #else	/* CONFIG_SYSFS */
6254 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)	{ }
6255 #endif	/* CONFIG_SYSFS */
6256 
6257 /*
6258  * Workqueue watchdog.
6259  *
6260  * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
6261  * flush dependency, a concurrency managed work item which stays RUNNING
6262  * indefinitely.  Workqueue stalls can be very difficult to debug as the
6263  * usual warning mechanisms don't trigger and internal workqueue state is
6264  * largely opaque.
6265  *
6266  * Workqueue watchdog monitors all worker pools periodically and dumps
6267  * state if some pools failed to make forward progress for a while where
6268  * forward progress is defined as the first item on ->worklist changing.
6269  *
6270  * This mechanism is controlled through the kernel parameter
6271  * "workqueue.watchdog_thresh" which can be updated at runtime through the
6272  * corresponding sysfs parameter file.
6273  */
6274 #ifdef CONFIG_WQ_WATCHDOG
6275 
6276 static unsigned long wq_watchdog_thresh = 30;
6277 static struct timer_list wq_watchdog_timer;
6278 
6279 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
6280 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
6281 
6282 /*
6283  * Show workers that might prevent the processing of pending work items.
6284  * The only candidates are CPU-bound workers in the running state.
6285  * Pending work items should be handled by another idle worker
6286  * in all other situations.
6287  */
6288 static void show_cpu_pool_hog(struct worker_pool *pool)
6289 {
6290 	struct worker *worker;
6291 	unsigned long flags;
6292 	int bkt;
6293 
6294 	raw_spin_lock_irqsave(&pool->lock, flags);
6295 
6296 	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6297 		if (task_is_running(worker->task)) {
6298 			/*
6299 			 * Defer printing to avoid deadlocks in console
6300 			 * drivers that queue work while holding locks
6301 			 * also taken in their write paths.
6302 			 */
6303 			printk_deferred_enter();
6304 
6305 			pr_info("pool %d:\n", pool->id);
6306 			sched_show_task(worker->task);
6307 
6308 			printk_deferred_exit();
6309 		}
6310 	}
6311 
6312 	raw_spin_unlock_irqrestore(&pool->lock, flags);
6313 }
6314 
6315 static void show_cpu_pools_hogs(void)
6316 {
6317 	struct worker_pool *pool;
6318 	int pi;
6319 
6320 	pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
6321 
6322 	rcu_read_lock();
6323 
6324 	for_each_pool(pool, pi) {
6325 		if (pool->cpu_stall)
6326 			show_cpu_pool_hog(pool);
6327 
6328 	}
6329 
6330 	rcu_read_unlock();
6331 }
6332 
6333 static void wq_watchdog_reset_touched(void)
6334 {
6335 	int cpu;
6336 
6337 	wq_watchdog_touched = jiffies;
6338 	for_each_possible_cpu(cpu)
6339 		per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
6340 }
6341 
6342 static void wq_watchdog_timer_fn(struct timer_list *unused)
6343 {
6344 	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
6345 	bool lockup_detected = false;
6346 	bool cpu_pool_stall = false;
6347 	unsigned long now = jiffies;
6348 	struct worker_pool *pool;
6349 	int pi;
6350 
6351 	if (!thresh)
6352 		return;
6353 
6354 	rcu_read_lock();
6355 
6356 	for_each_pool(pool, pi) {
6357 		unsigned long pool_ts, touched, ts;
6358 
6359 		pool->cpu_stall = false;
6360 		if (list_empty(&pool->worklist))
6361 			continue;
6362 
6363 		/*
6364 		 * If a virtual machine is stopped by the host it can look to
6365 		 * the watchdog like a stall.
6366 		 */
6367 		kvm_check_and_clear_guest_paused();
6368 
6369 		/* get the latest of pool and touched timestamps */
6370 		if (pool->cpu >= 0)
6371 			touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
6372 		else
6373 			touched = READ_ONCE(wq_watchdog_touched);
6374 		pool_ts = READ_ONCE(pool->watchdog_ts);
6375 
6376 		if (time_after(pool_ts, touched))
6377 			ts = pool_ts;
6378 		else
6379 			ts = touched;
6380 
6381 		/* did we stall? */
6382 		if (time_after(now, ts + thresh)) {
6383 			lockup_detected = true;
6384 			if (pool->cpu >= 0) {
6385 				pool->cpu_stall = true;
6386 				cpu_pool_stall = true;
6387 			}
6388 			pr_emerg("BUG: workqueue lockup - pool");
6389 			pr_cont_pool_info(pool);
6390 			pr_cont(" stuck for %us!\n",
6391 				jiffies_to_msecs(now - pool_ts) / 1000);
6392 		}
6393 
6394 
6395 	}
6396 
6397 	rcu_read_unlock();
6398 
6399 	if (lockup_detected)
6400 		show_all_workqueues();
6401 
6402 	if (cpu_pool_stall)
6403 		show_cpu_pools_hogs();
6404 
6405 	wq_watchdog_reset_touched();
6406 	mod_timer(&wq_watchdog_timer, jiffies + thresh);
6407 }
6408 
6409 notrace void wq_watchdog_touch(int cpu)
6410 {
6411 	if (cpu >= 0)
6412 		per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
6413 
6414 	wq_watchdog_touched = jiffies;
6415 }
6416 
6417 static void wq_watchdog_set_thresh(unsigned long thresh)
6418 {
6419 	wq_watchdog_thresh = 0;
6420 	del_timer_sync(&wq_watchdog_timer);
6421 
6422 	if (thresh) {
6423 		wq_watchdog_thresh = thresh;
6424 		wq_watchdog_reset_touched();
6425 		mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
6426 	}
6427 }
6428 
6429 static int wq_watchdog_param_set_thresh(const char *val,
6430 					const struct kernel_param *kp)
6431 {
6432 	unsigned long thresh;
6433 	int ret;
6434 
6435 	ret = kstrtoul(val, 0, &thresh);
6436 	if (ret)
6437 		return ret;
6438 
6439 	if (system_wq)
6440 		wq_watchdog_set_thresh(thresh);
6441 	else
6442 		wq_watchdog_thresh = thresh;
6443 
6444 	return 0;
6445 }
6446 
6447 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
6448 	.set	= wq_watchdog_param_set_thresh,
6449 	.get	= param_get_ulong,
6450 };
6451 
6452 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
6453 		0644);
6454 
6455 static void wq_watchdog_init(void)
6456 {
6457 	timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
6458 	wq_watchdog_set_thresh(wq_watchdog_thresh);
6459 }
6460 
6461 #else	/* CONFIG_WQ_WATCHDOG */
6462 
6463 static inline void wq_watchdog_init(void) { }
6464 
6465 #endif	/* CONFIG_WQ_WATCHDOG */
6466 
6467 /**
6468  * workqueue_init_early - early init for workqueue subsystem
6469  *
6470  * This is the first step of three-staged workqueue subsystem initialization and
6471  * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
6472  * up. It sets up all the data structures and system workqueues and allows early
6473  * boot code to create workqueues and queue/cancel work items. Actual work item
6474  * execution starts only after kthreads can be created and scheduled right
6475  * before early initcalls.
6476  */
6477 void __init workqueue_init_early(void)
6478 {
6479 	struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
6480 	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
6481 	int i, cpu;
6482 
6483 	BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
6484 
6485 	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
6486 	cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ));
6487 	cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN));
6488 
6489 	if (!cpumask_empty(&wq_cmdline_cpumask))
6490 		cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, &wq_cmdline_cpumask);
6491 
6492 	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
6493 
6494 	wq_update_pod_attrs_buf = alloc_workqueue_attrs();
6495 	BUG_ON(!wq_update_pod_attrs_buf);
6496 
6497 	/* initialize WQ_AFFN_SYSTEM pods */
6498 	pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
6499 	pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
6500 	pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
6501 	BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
6502 
6503 	BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
6504 
6505 	wq_update_pod_attrs_buf = alloc_workqueue_attrs();
6506 	BUG_ON(!wq_update_pod_attrs_buf);
6507 
6508 	pt->nr_pods = 1;
6509 	cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
6510 	pt->pod_node[0] = NUMA_NO_NODE;
6511 	pt->cpu_pod[0] = 0;
6512 
6513 	/* initialize CPU pools */
6514 	for_each_possible_cpu(cpu) {
6515 		struct worker_pool *pool;
6516 
6517 		i = 0;
6518 		for_each_cpu_worker_pool(pool, cpu) {
6519 			BUG_ON(init_worker_pool(pool));
6520 			pool->cpu = cpu;
6521 			cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
6522 			cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
6523 			pool->attrs->nice = std_nice[i++];
6524 			pool->attrs->affn_strict = true;
6525 			pool->node = cpu_to_node(cpu);
6526 
6527 			/* alloc pool ID */
6528 			mutex_lock(&wq_pool_mutex);
6529 			BUG_ON(worker_pool_assign_id(pool));
6530 			mutex_unlock(&wq_pool_mutex);
6531 		}
6532 	}
6533 
6534 	/* create default unbound and ordered wq attrs */
6535 	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
6536 		struct workqueue_attrs *attrs;
6537 
6538 		BUG_ON(!(attrs = alloc_workqueue_attrs()));
6539 		attrs->nice = std_nice[i];
6540 		unbound_std_wq_attrs[i] = attrs;
6541 
6542 		/*
6543 		 * An ordered wq should have only one pwq as ordering is
6544 		 * guaranteed by max_active which is enforced by pwqs.
6545 		 */
6546 		BUG_ON(!(attrs = alloc_workqueue_attrs()));
6547 		attrs->nice = std_nice[i];
6548 		attrs->ordered = true;
6549 		ordered_wq_attrs[i] = attrs;
6550 	}
6551 
6552 	system_wq = alloc_workqueue("events", 0, 0);
6553 	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
6554 	system_long_wq = alloc_workqueue("events_long", 0, 0);
6555 	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
6556 					    WQ_MAX_ACTIVE);
6557 	system_freezable_wq = alloc_workqueue("events_freezable",
6558 					      WQ_FREEZABLE, 0);
6559 	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
6560 					      WQ_POWER_EFFICIENT, 0);
6561 	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
6562 					      WQ_FREEZABLE | WQ_POWER_EFFICIENT,
6563 					      0);
6564 	BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
6565 	       !system_unbound_wq || !system_freezable_wq ||
6566 	       !system_power_efficient_wq ||
6567 	       !system_freezable_power_efficient_wq);
6568 }
6569 
6570 static void __init wq_cpu_intensive_thresh_init(void)
6571 {
6572 	unsigned long thresh;
6573 	unsigned long bogo;
6574 
6575 	/* if the user set it to a specific value, keep it */
6576 	if (wq_cpu_intensive_thresh_us != ULONG_MAX)
6577 		return;
6578 
6579 	pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
6580 	BUG_ON(IS_ERR(pwq_release_worker));
6581 
6582 	/*
6583 	 * The default of 10ms is derived from the fact that most modern (as of
6584 	 * 2023) processors can do a lot in 10ms and that it's just below what
6585 	 * most consider human-perceivable. However, the kernel also runs on a
6586 	 * lot slower CPUs including microcontrollers where the threshold is way
6587 	 * too low.
6588 	 *
6589 	 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
6590 	 * This is by no means accurate but it doesn't have to be. The mechanism
6591 	 * is still useful even when the threshold is fully scaled up. Also, as
6592 	 * the reports would usually be applicable to everyone, some machines
6593 	 * operating on longer thresholds won't significantly diminish their
6594 	 * usefulness.
6595 	 */
6596 	thresh = 10 * USEC_PER_MSEC;
6597 
6598 	/* see init/calibrate.c for lpj -> BogoMIPS calculation */
6599 	bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
6600 	if (bogo < 4000)
6601 		thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
6602 
6603 	pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
6604 		 loops_per_jiffy, bogo, thresh);
6605 
6606 	wq_cpu_intensive_thresh_us = thresh;
6607 }
6608 
6609 /**
6610  * workqueue_init - bring workqueue subsystem fully online
6611  *
6612  * This is the second step of three-staged workqueue subsystem initialization
6613  * and invoked as soon as kthreads can be created and scheduled. Workqueues have
6614  * been created and work items queued on them, but there are no kworkers
6615  * executing the work items yet. Populate the worker pools with the initial
6616  * workers and enable future kworker creations.
6617  */
6618 void __init workqueue_init(void)
6619 {
6620 	struct workqueue_struct *wq;
6621 	struct worker_pool *pool;
6622 	int cpu, bkt;
6623 
6624 	wq_cpu_intensive_thresh_init();
6625 
6626 	mutex_lock(&wq_pool_mutex);
6627 
6628 	/*
6629 	 * Per-cpu pools created earlier could be missing node hint. Fix them
6630 	 * up. Also, create a rescuer for workqueues that requested it.
6631 	 */
6632 	for_each_possible_cpu(cpu) {
6633 		for_each_cpu_worker_pool(pool, cpu) {
6634 			pool->node = cpu_to_node(cpu);
6635 		}
6636 	}
6637 
6638 	list_for_each_entry(wq, &workqueues, list) {
6639 		WARN(init_rescuer(wq),
6640 		     "workqueue: failed to create early rescuer for %s",
6641 		     wq->name);
6642 	}
6643 
6644 	mutex_unlock(&wq_pool_mutex);
6645 
6646 	/* create the initial workers */
6647 	for_each_online_cpu(cpu) {
6648 		for_each_cpu_worker_pool(pool, cpu) {
6649 			pool->flags &= ~POOL_DISASSOCIATED;
6650 			BUG_ON(!create_worker(pool));
6651 		}
6652 	}
6653 
6654 	hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6655 		BUG_ON(!create_worker(pool));
6656 
6657 	wq_online = true;
6658 	wq_watchdog_init();
6659 }
6660 
6661 /*
6662  * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
6663  * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
6664  * and consecutive pod ID. The rest of @pt is initialized accordingly.
6665  */
6666 static void __init init_pod_type(struct wq_pod_type *pt,
6667 				 bool (*cpus_share_pod)(int, int))
6668 {
6669 	int cur, pre, cpu, pod;
6670 
6671 	pt->nr_pods = 0;
6672 
6673 	/* init @pt->cpu_pod[] according to @cpus_share_pod() */
6674 	pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
6675 	BUG_ON(!pt->cpu_pod);
6676 
6677 	for_each_possible_cpu(cur) {
6678 		for_each_possible_cpu(pre) {
6679 			if (pre >= cur) {
6680 				pt->cpu_pod[cur] = pt->nr_pods++;
6681 				break;
6682 			}
6683 			if (cpus_share_pod(cur, pre)) {
6684 				pt->cpu_pod[cur] = pt->cpu_pod[pre];
6685 				break;
6686 			}
6687 		}
6688 	}
6689 
6690 	/* init the rest to match @pt->cpu_pod[] */
6691 	pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
6692 	pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
6693 	BUG_ON(!pt->pod_cpus || !pt->pod_node);
6694 
6695 	for (pod = 0; pod < pt->nr_pods; pod++)
6696 		BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
6697 
6698 	for_each_possible_cpu(cpu) {
6699 		cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
6700 		pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
6701 	}
6702 }
6703 
6704 static bool __init cpus_dont_share(int cpu0, int cpu1)
6705 {
6706 	return false;
6707 }
6708 
6709 static bool __init cpus_share_smt(int cpu0, int cpu1)
6710 {
6711 #ifdef CONFIG_SCHED_SMT
6712 	return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
6713 #else
6714 	return false;
6715 #endif
6716 }
6717 
6718 static bool __init cpus_share_numa(int cpu0, int cpu1)
6719 {
6720 	return cpu_to_node(cpu0) == cpu_to_node(cpu1);
6721 }
6722 
6723 /**
6724  * workqueue_init_topology - initialize CPU pods for unbound workqueues
6725  *
6726  * This is the third step of there-staged workqueue subsystem initialization and
6727  * invoked after SMP and topology information are fully initialized. It
6728  * initializes the unbound CPU pods accordingly.
6729  */
6730 void __init workqueue_init_topology(void)
6731 {
6732 	struct workqueue_struct *wq;
6733 	int cpu;
6734 
6735 	init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
6736 	init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
6737 	init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
6738 	init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
6739 
6740 	mutex_lock(&wq_pool_mutex);
6741 
6742 	/*
6743 	 * Workqueues allocated earlier would have all CPUs sharing the default
6744 	 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
6745 	 * combinations to apply per-pod sharing.
6746 	 */
6747 	list_for_each_entry(wq, &workqueues, list) {
6748 		for_each_online_cpu(cpu) {
6749 			wq_update_pod(wq, cpu, cpu, true);
6750 		}
6751 	}
6752 
6753 	mutex_unlock(&wq_pool_mutex);
6754 }
6755 
6756 void __warn_flushing_systemwide_wq(void)
6757 {
6758 	pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
6759 	dump_stack();
6760 }
6761 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
6762 
6763 static int __init workqueue_unbound_cpus_setup(char *str)
6764 {
6765 	if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
6766 		cpumask_clear(&wq_cmdline_cpumask);
6767 		pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
6768 	}
6769 
6770 	return 1;
6771 }
6772 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);
6773