xref: /linux/kernel/workqueue.c (revision b85d45947951d23cb22d90caecf4c1eb81342c96)
1 /*
2  * kernel/workqueue.c - generic async execution with shared worker pool
3  *
4  * Copyright (C) 2002		Ingo Molnar
5  *
6  *   Derived from the taskqueue/keventd code by:
7  *     David Woodhouse <dwmw2@infradead.org>
8  *     Andrew Morton
9  *     Kai Petzke <wpp@marie.physik.tu-berlin.de>
10  *     Theodore Ts'o <tytso@mit.edu>
11  *
12  * Made to use alloc_percpu by Christoph Lameter.
13  *
14  * Copyright (C) 2010		SUSE Linux Products GmbH
15  * Copyright (C) 2010		Tejun Heo <tj@kernel.org>
16  *
17  * This is the generic async execution mechanism.  Work items as are
18  * executed in process context.  The worker pool is shared and
19  * automatically managed.  There are two worker pools for each CPU (one for
20  * normal work items and the other for high priority ones) and some extra
21  * pools for workqueues which are not bound to any specific CPU - the
22  * number of these backing pools is dynamic.
23  *
24  * Please read Documentation/workqueue.txt for details.
25  */
26 
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 
52 #include "workqueue_internal.h"
53 
54 enum {
55 	/*
56 	 * worker_pool flags
57 	 *
58 	 * A bound pool is either associated or disassociated with its CPU.
59 	 * While associated (!DISASSOCIATED), all workers are bound to the
60 	 * CPU and none has %WORKER_UNBOUND set and concurrency management
61 	 * is in effect.
62 	 *
63 	 * While DISASSOCIATED, the cpu may be offline and all workers have
64 	 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 	 * be executing on any CPU.  The pool behaves as an unbound one.
66 	 *
67 	 * Note that DISASSOCIATED should be flipped only while holding
68 	 * attach_mutex to avoid changing binding state while
69 	 * worker_attach_to_pool() is in progress.
70 	 */
71 	POOL_DISASSOCIATED	= 1 << 2,	/* cpu can't serve workers */
72 
73 	/* worker flags */
74 	WORKER_DIE		= 1 << 1,	/* die die die */
75 	WORKER_IDLE		= 1 << 2,	/* is idle */
76 	WORKER_PREP		= 1 << 3,	/* preparing to run works */
77 	WORKER_CPU_INTENSIVE	= 1 << 6,	/* cpu intensive */
78 	WORKER_UNBOUND		= 1 << 7,	/* worker is unbound */
79 	WORKER_REBOUND		= 1 << 8,	/* worker was rebound */
80 
81 	WORKER_NOT_RUNNING	= WORKER_PREP | WORKER_CPU_INTENSIVE |
82 				  WORKER_UNBOUND | WORKER_REBOUND,
83 
84 	NR_STD_WORKER_POOLS	= 2,		/* # standard pools per cpu */
85 
86 	UNBOUND_POOL_HASH_ORDER	= 6,		/* hashed by pool->attrs */
87 	BUSY_WORKER_HASH_ORDER	= 6,		/* 64 pointers */
88 
89 	MAX_IDLE_WORKERS_RATIO	= 4,		/* 1/4 of busy can be idle */
90 	IDLE_WORKER_TIMEOUT	= 300 * HZ,	/* keep idle ones for 5 mins */
91 
92 	MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
93 						/* call for help after 10ms
94 						   (min two ticks) */
95 	MAYDAY_INTERVAL		= HZ / 10,	/* and then every 100ms */
96 	CREATE_COOLDOWN		= HZ,		/* time to breath after fail */
97 
98 	/*
99 	 * Rescue workers are used only on emergencies and shared by
100 	 * all cpus.  Give MIN_NICE.
101 	 */
102 	RESCUER_NICE_LEVEL	= MIN_NICE,
103 	HIGHPRI_NICE_LEVEL	= MIN_NICE,
104 
105 	WQ_NAME_LEN		= 24,
106 };
107 
108 /*
109  * Structure fields follow one of the following exclusion rules.
110  *
111  * I: Modifiable by initialization/destruction paths and read-only for
112  *    everyone else.
113  *
114  * P: Preemption protected.  Disabling preemption is enough and should
115  *    only be modified and accessed from the local cpu.
116  *
117  * L: pool->lock protected.  Access with pool->lock held.
118  *
119  * X: During normal operation, modification requires pool->lock and should
120  *    be done only from local cpu.  Either disabling preemption on local
121  *    cpu or grabbing pool->lock is enough for read access.  If
122  *    POOL_DISASSOCIATED is set, it's identical to L.
123  *
124  * A: pool->attach_mutex protected.
125  *
126  * PL: wq_pool_mutex protected.
127  *
128  * PR: wq_pool_mutex protected for writes.  Sched-RCU protected for reads.
129  *
130  * PW: wq_pool_mutex and wq->mutex protected for writes.  Either for reads.
131  *
132  * PWR: wq_pool_mutex and wq->mutex protected for writes.  Either or
133  *      sched-RCU for reads.
134  *
135  * WQ: wq->mutex protected.
136  *
137  * WR: wq->mutex protected for writes.  Sched-RCU protected for reads.
138  *
139  * MD: wq_mayday_lock protected.
140  */
141 
142 /* struct worker is defined in workqueue_internal.h */
143 
144 struct worker_pool {
145 	spinlock_t		lock;		/* the pool lock */
146 	int			cpu;		/* I: the associated cpu */
147 	int			node;		/* I: the associated node ID */
148 	int			id;		/* I: pool ID */
149 	unsigned int		flags;		/* X: flags */
150 
151 	struct list_head	worklist;	/* L: list of pending works */
152 	int			nr_workers;	/* L: total number of workers */
153 
154 	/* nr_idle includes the ones off idle_list for rebinding */
155 	int			nr_idle;	/* L: currently idle ones */
156 
157 	struct list_head	idle_list;	/* X: list of idle workers */
158 	struct timer_list	idle_timer;	/* L: worker idle timeout */
159 	struct timer_list	mayday_timer;	/* L: SOS timer for workers */
160 
161 	/* a workers is either on busy_hash or idle_list, or the manager */
162 	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
163 						/* L: hash of busy workers */
164 
165 	/* see manage_workers() for details on the two manager mutexes */
166 	struct mutex		manager_arb;	/* manager arbitration */
167 	struct worker		*manager;	/* L: purely informational */
168 	struct mutex		attach_mutex;	/* attach/detach exclusion */
169 	struct list_head	workers;	/* A: attached workers */
170 	struct completion	*detach_completion; /* all workers detached */
171 
172 	struct ida		worker_ida;	/* worker IDs for task name */
173 
174 	struct workqueue_attrs	*attrs;		/* I: worker attributes */
175 	struct hlist_node	hash_node;	/* PL: unbound_pool_hash node */
176 	int			refcnt;		/* PL: refcnt for unbound pools */
177 
178 	/*
179 	 * The current concurrency level.  As it's likely to be accessed
180 	 * from other CPUs during try_to_wake_up(), put it in a separate
181 	 * cacheline.
182 	 */
183 	atomic_t		nr_running ____cacheline_aligned_in_smp;
184 
185 	/*
186 	 * Destruction of pool is sched-RCU protected to allow dereferences
187 	 * from get_work_pool().
188 	 */
189 	struct rcu_head		rcu;
190 } ____cacheline_aligned_in_smp;
191 
192 /*
193  * The per-pool workqueue.  While queued, the lower WORK_STRUCT_FLAG_BITS
194  * of work_struct->data are used for flags and the remaining high bits
195  * point to the pwq; thus, pwqs need to be aligned at two's power of the
196  * number of flag bits.
197  */
198 struct pool_workqueue {
199 	struct worker_pool	*pool;		/* I: the associated pool */
200 	struct workqueue_struct *wq;		/* I: the owning workqueue */
201 	int			work_color;	/* L: current color */
202 	int			flush_color;	/* L: flushing color */
203 	int			refcnt;		/* L: reference count */
204 	int			nr_in_flight[WORK_NR_COLORS];
205 						/* L: nr of in_flight works */
206 	int			nr_active;	/* L: nr of active works */
207 	int			max_active;	/* L: max active works */
208 	struct list_head	delayed_works;	/* L: delayed works */
209 	struct list_head	pwqs_node;	/* WR: node on wq->pwqs */
210 	struct list_head	mayday_node;	/* MD: node on wq->maydays */
211 
212 	/*
213 	 * Release of unbound pwq is punted to system_wq.  See put_pwq()
214 	 * and pwq_unbound_release_workfn() for details.  pool_workqueue
215 	 * itself is also sched-RCU protected so that the first pwq can be
216 	 * determined without grabbing wq->mutex.
217 	 */
218 	struct work_struct	unbound_release_work;
219 	struct rcu_head		rcu;
220 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
221 
222 /*
223  * Structure used to wait for workqueue flush.
224  */
225 struct wq_flusher {
226 	struct list_head	list;		/* WQ: list of flushers */
227 	int			flush_color;	/* WQ: flush color waiting for */
228 	struct completion	done;		/* flush completion */
229 };
230 
231 struct wq_device;
232 
233 /*
234  * The externally visible workqueue.  It relays the issued work items to
235  * the appropriate worker_pool through its pool_workqueues.
236  */
237 struct workqueue_struct {
238 	struct list_head	pwqs;		/* WR: all pwqs of this wq */
239 	struct list_head	list;		/* PR: list of all workqueues */
240 
241 	struct mutex		mutex;		/* protects this wq */
242 	int			work_color;	/* WQ: current work color */
243 	int			flush_color;	/* WQ: current flush color */
244 	atomic_t		nr_pwqs_to_flush; /* flush in progress */
245 	struct wq_flusher	*first_flusher;	/* WQ: first flusher */
246 	struct list_head	flusher_queue;	/* WQ: flush waiters */
247 	struct list_head	flusher_overflow; /* WQ: flush overflow list */
248 
249 	struct list_head	maydays;	/* MD: pwqs requesting rescue */
250 	struct worker		*rescuer;	/* I: rescue worker */
251 
252 	int			nr_drainers;	/* WQ: drain in progress */
253 	int			saved_max_active; /* WQ: saved pwq max_active */
254 
255 	struct workqueue_attrs	*unbound_attrs;	/* PW: only for unbound wqs */
256 	struct pool_workqueue	*dfl_pwq;	/* PW: only for unbound wqs */
257 
258 #ifdef CONFIG_SYSFS
259 	struct wq_device	*wq_dev;	/* I: for sysfs interface */
260 #endif
261 #ifdef CONFIG_LOCKDEP
262 	struct lockdep_map	lockdep_map;
263 #endif
264 	char			name[WQ_NAME_LEN]; /* I: workqueue name */
265 
266 	/*
267 	 * Destruction of workqueue_struct is sched-RCU protected to allow
268 	 * walking the workqueues list without grabbing wq_pool_mutex.
269 	 * This is used to dump all workqueues from sysrq.
270 	 */
271 	struct rcu_head		rcu;
272 
273 	/* hot fields used during command issue, aligned to cacheline */
274 	unsigned int		flags ____cacheline_aligned; /* WQ: WQ_* flags */
275 	struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
276 	struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
277 };
278 
279 static struct kmem_cache *pwq_cache;
280 
281 static cpumask_var_t *wq_numa_possible_cpumask;
282 					/* possible CPUs of each node */
283 
284 static bool wq_disable_numa;
285 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
286 
287 /* see the comment above the definition of WQ_POWER_EFFICIENT */
288 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
289 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
290 
291 static bool wq_numa_enabled;		/* unbound NUMA affinity enabled */
292 
293 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
294 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
295 
296 static DEFINE_MUTEX(wq_pool_mutex);	/* protects pools and workqueues list */
297 static DEFINE_SPINLOCK(wq_mayday_lock);	/* protects wq->maydays list */
298 
299 static LIST_HEAD(workqueues);		/* PR: list of all workqueues */
300 static bool workqueue_freezing;		/* PL: have wqs started freezing? */
301 
302 static cpumask_var_t wq_unbound_cpumask; /* PL: low level cpumask for all unbound wqs */
303 
304 /* the per-cpu worker pools */
305 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
306 				     cpu_worker_pools);
307 
308 static DEFINE_IDR(worker_pool_idr);	/* PR: idr of all pools */
309 
310 /* PL: hash of all unbound pools keyed by pool->attrs */
311 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
312 
313 /* I: attributes used when instantiating standard unbound pools on demand */
314 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
315 
316 /* I: attributes used when instantiating ordered pools on demand */
317 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
318 
319 struct workqueue_struct *system_wq __read_mostly;
320 EXPORT_SYMBOL(system_wq);
321 struct workqueue_struct *system_highpri_wq __read_mostly;
322 EXPORT_SYMBOL_GPL(system_highpri_wq);
323 struct workqueue_struct *system_long_wq __read_mostly;
324 EXPORT_SYMBOL_GPL(system_long_wq);
325 struct workqueue_struct *system_unbound_wq __read_mostly;
326 EXPORT_SYMBOL_GPL(system_unbound_wq);
327 struct workqueue_struct *system_freezable_wq __read_mostly;
328 EXPORT_SYMBOL_GPL(system_freezable_wq);
329 struct workqueue_struct *system_power_efficient_wq __read_mostly;
330 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
331 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
332 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
333 
334 static int worker_thread(void *__worker);
335 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
336 
337 #define CREATE_TRACE_POINTS
338 #include <trace/events/workqueue.h>
339 
340 #define assert_rcu_or_pool_mutex()					\
341 	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() &&			\
342 			 !lockdep_is_held(&wq_pool_mutex),		\
343 			 "sched RCU or wq_pool_mutex should be held")
344 
345 #define assert_rcu_or_wq_mutex(wq)					\
346 	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() &&			\
347 			 !lockdep_is_held(&wq->mutex),			\
348 			 "sched RCU or wq->mutex should be held")
349 
350 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq)			\
351 	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() &&			\
352 			 !lockdep_is_held(&wq->mutex) &&		\
353 			 !lockdep_is_held(&wq_pool_mutex),		\
354 			 "sched RCU, wq->mutex or wq_pool_mutex should be held")
355 
356 #define for_each_cpu_worker_pool(pool, cpu)				\
357 	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0];		\
358 	     (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
359 	     (pool)++)
360 
361 /**
362  * for_each_pool - iterate through all worker_pools in the system
363  * @pool: iteration cursor
364  * @pi: integer used for iteration
365  *
366  * This must be called either with wq_pool_mutex held or sched RCU read
367  * locked.  If the pool needs to be used beyond the locking in effect, the
368  * caller is responsible for guaranteeing that the pool stays online.
369  *
370  * The if/else clause exists only for the lockdep assertion and can be
371  * ignored.
372  */
373 #define for_each_pool(pool, pi)						\
374 	idr_for_each_entry(&worker_pool_idr, pool, pi)			\
375 		if (({ assert_rcu_or_pool_mutex(); false; })) { }	\
376 		else
377 
378 /**
379  * for_each_pool_worker - iterate through all workers of a worker_pool
380  * @worker: iteration cursor
381  * @pool: worker_pool to iterate workers of
382  *
383  * This must be called with @pool->attach_mutex.
384  *
385  * The if/else clause exists only for the lockdep assertion and can be
386  * ignored.
387  */
388 #define for_each_pool_worker(worker, pool)				\
389 	list_for_each_entry((worker), &(pool)->workers, node)		\
390 		if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
391 		else
392 
393 /**
394  * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
395  * @pwq: iteration cursor
396  * @wq: the target workqueue
397  *
398  * This must be called either with wq->mutex held or sched RCU read locked.
399  * If the pwq needs to be used beyond the locking in effect, the caller is
400  * responsible for guaranteeing that the pwq stays online.
401  *
402  * The if/else clause exists only for the lockdep assertion and can be
403  * ignored.
404  */
405 #define for_each_pwq(pwq, wq)						\
406 	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node)		\
407 		if (({ assert_rcu_or_wq_mutex(wq); false; })) { }	\
408 		else
409 
410 #ifdef CONFIG_DEBUG_OBJECTS_WORK
411 
412 static struct debug_obj_descr work_debug_descr;
413 
414 static void *work_debug_hint(void *addr)
415 {
416 	return ((struct work_struct *) addr)->func;
417 }
418 
419 /*
420  * fixup_init is called when:
421  * - an active object is initialized
422  */
423 static int work_fixup_init(void *addr, enum debug_obj_state state)
424 {
425 	struct work_struct *work = addr;
426 
427 	switch (state) {
428 	case ODEBUG_STATE_ACTIVE:
429 		cancel_work_sync(work);
430 		debug_object_init(work, &work_debug_descr);
431 		return 1;
432 	default:
433 		return 0;
434 	}
435 }
436 
437 /*
438  * fixup_activate is called when:
439  * - an active object is activated
440  * - an unknown object is activated (might be a statically initialized object)
441  */
442 static int work_fixup_activate(void *addr, enum debug_obj_state state)
443 {
444 	struct work_struct *work = addr;
445 
446 	switch (state) {
447 
448 	case ODEBUG_STATE_NOTAVAILABLE:
449 		/*
450 		 * This is not really a fixup. The work struct was
451 		 * statically initialized. We just make sure that it
452 		 * is tracked in the object tracker.
453 		 */
454 		if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
455 			debug_object_init(work, &work_debug_descr);
456 			debug_object_activate(work, &work_debug_descr);
457 			return 0;
458 		}
459 		WARN_ON_ONCE(1);
460 		return 0;
461 
462 	case ODEBUG_STATE_ACTIVE:
463 		WARN_ON(1);
464 
465 	default:
466 		return 0;
467 	}
468 }
469 
470 /*
471  * fixup_free is called when:
472  * - an active object is freed
473  */
474 static int work_fixup_free(void *addr, enum debug_obj_state state)
475 {
476 	struct work_struct *work = addr;
477 
478 	switch (state) {
479 	case ODEBUG_STATE_ACTIVE:
480 		cancel_work_sync(work);
481 		debug_object_free(work, &work_debug_descr);
482 		return 1;
483 	default:
484 		return 0;
485 	}
486 }
487 
488 static struct debug_obj_descr work_debug_descr = {
489 	.name		= "work_struct",
490 	.debug_hint	= work_debug_hint,
491 	.fixup_init	= work_fixup_init,
492 	.fixup_activate	= work_fixup_activate,
493 	.fixup_free	= work_fixup_free,
494 };
495 
496 static inline void debug_work_activate(struct work_struct *work)
497 {
498 	debug_object_activate(work, &work_debug_descr);
499 }
500 
501 static inline void debug_work_deactivate(struct work_struct *work)
502 {
503 	debug_object_deactivate(work, &work_debug_descr);
504 }
505 
506 void __init_work(struct work_struct *work, int onstack)
507 {
508 	if (onstack)
509 		debug_object_init_on_stack(work, &work_debug_descr);
510 	else
511 		debug_object_init(work, &work_debug_descr);
512 }
513 EXPORT_SYMBOL_GPL(__init_work);
514 
515 void destroy_work_on_stack(struct work_struct *work)
516 {
517 	debug_object_free(work, &work_debug_descr);
518 }
519 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
520 
521 void destroy_delayed_work_on_stack(struct delayed_work *work)
522 {
523 	destroy_timer_on_stack(&work->timer);
524 	debug_object_free(&work->work, &work_debug_descr);
525 }
526 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
527 
528 #else
529 static inline void debug_work_activate(struct work_struct *work) { }
530 static inline void debug_work_deactivate(struct work_struct *work) { }
531 #endif
532 
533 /**
534  * worker_pool_assign_id - allocate ID and assing it to @pool
535  * @pool: the pool pointer of interest
536  *
537  * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
538  * successfully, -errno on failure.
539  */
540 static int worker_pool_assign_id(struct worker_pool *pool)
541 {
542 	int ret;
543 
544 	lockdep_assert_held(&wq_pool_mutex);
545 
546 	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
547 			GFP_KERNEL);
548 	if (ret >= 0) {
549 		pool->id = ret;
550 		return 0;
551 	}
552 	return ret;
553 }
554 
555 /**
556  * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
557  * @wq: the target workqueue
558  * @node: the node ID
559  *
560  * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
561  * read locked.
562  * If the pwq needs to be used beyond the locking in effect, the caller is
563  * responsible for guaranteeing that the pwq stays online.
564  *
565  * Return: The unbound pool_workqueue for @node.
566  */
567 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
568 						  int node)
569 {
570 	assert_rcu_or_wq_mutex_or_pool_mutex(wq);
571 	return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
572 }
573 
574 static unsigned int work_color_to_flags(int color)
575 {
576 	return color << WORK_STRUCT_COLOR_SHIFT;
577 }
578 
579 static int get_work_color(struct work_struct *work)
580 {
581 	return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
582 		((1 << WORK_STRUCT_COLOR_BITS) - 1);
583 }
584 
585 static int work_next_color(int color)
586 {
587 	return (color + 1) % WORK_NR_COLORS;
588 }
589 
590 /*
591  * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
592  * contain the pointer to the queued pwq.  Once execution starts, the flag
593  * is cleared and the high bits contain OFFQ flags and pool ID.
594  *
595  * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
596  * and clear_work_data() can be used to set the pwq, pool or clear
597  * work->data.  These functions should only be called while the work is
598  * owned - ie. while the PENDING bit is set.
599  *
600  * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
601  * corresponding to a work.  Pool is available once the work has been
602  * queued anywhere after initialization until it is sync canceled.  pwq is
603  * available only while the work item is queued.
604  *
605  * %WORK_OFFQ_CANCELING is used to mark a work item which is being
606  * canceled.  While being canceled, a work item may have its PENDING set
607  * but stay off timer and worklist for arbitrarily long and nobody should
608  * try to steal the PENDING bit.
609  */
610 static inline void set_work_data(struct work_struct *work, unsigned long data,
611 				 unsigned long flags)
612 {
613 	WARN_ON_ONCE(!work_pending(work));
614 	atomic_long_set(&work->data, data | flags | work_static(work));
615 }
616 
617 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
618 			 unsigned long extra_flags)
619 {
620 	set_work_data(work, (unsigned long)pwq,
621 		      WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
622 }
623 
624 static void set_work_pool_and_keep_pending(struct work_struct *work,
625 					   int pool_id)
626 {
627 	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
628 		      WORK_STRUCT_PENDING);
629 }
630 
631 static void set_work_pool_and_clear_pending(struct work_struct *work,
632 					    int pool_id)
633 {
634 	/*
635 	 * The following wmb is paired with the implied mb in
636 	 * test_and_set_bit(PENDING) and ensures all updates to @work made
637 	 * here are visible to and precede any updates by the next PENDING
638 	 * owner.
639 	 */
640 	smp_wmb();
641 	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
642 }
643 
644 static void clear_work_data(struct work_struct *work)
645 {
646 	smp_wmb();	/* see set_work_pool_and_clear_pending() */
647 	set_work_data(work, WORK_STRUCT_NO_POOL, 0);
648 }
649 
650 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
651 {
652 	unsigned long data = atomic_long_read(&work->data);
653 
654 	if (data & WORK_STRUCT_PWQ)
655 		return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
656 	else
657 		return NULL;
658 }
659 
660 /**
661  * get_work_pool - return the worker_pool a given work was associated with
662  * @work: the work item of interest
663  *
664  * Pools are created and destroyed under wq_pool_mutex, and allows read
665  * access under sched-RCU read lock.  As such, this function should be
666  * called under wq_pool_mutex or with preemption disabled.
667  *
668  * All fields of the returned pool are accessible as long as the above
669  * mentioned locking is in effect.  If the returned pool needs to be used
670  * beyond the critical section, the caller is responsible for ensuring the
671  * returned pool is and stays online.
672  *
673  * Return: The worker_pool @work was last associated with.  %NULL if none.
674  */
675 static struct worker_pool *get_work_pool(struct work_struct *work)
676 {
677 	unsigned long data = atomic_long_read(&work->data);
678 	int pool_id;
679 
680 	assert_rcu_or_pool_mutex();
681 
682 	if (data & WORK_STRUCT_PWQ)
683 		return ((struct pool_workqueue *)
684 			(data & WORK_STRUCT_WQ_DATA_MASK))->pool;
685 
686 	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
687 	if (pool_id == WORK_OFFQ_POOL_NONE)
688 		return NULL;
689 
690 	return idr_find(&worker_pool_idr, pool_id);
691 }
692 
693 /**
694  * get_work_pool_id - return the worker pool ID a given work is associated with
695  * @work: the work item of interest
696  *
697  * Return: The worker_pool ID @work was last associated with.
698  * %WORK_OFFQ_POOL_NONE if none.
699  */
700 static int get_work_pool_id(struct work_struct *work)
701 {
702 	unsigned long data = atomic_long_read(&work->data);
703 
704 	if (data & WORK_STRUCT_PWQ)
705 		return ((struct pool_workqueue *)
706 			(data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
707 
708 	return data >> WORK_OFFQ_POOL_SHIFT;
709 }
710 
711 static void mark_work_canceling(struct work_struct *work)
712 {
713 	unsigned long pool_id = get_work_pool_id(work);
714 
715 	pool_id <<= WORK_OFFQ_POOL_SHIFT;
716 	set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
717 }
718 
719 static bool work_is_canceling(struct work_struct *work)
720 {
721 	unsigned long data = atomic_long_read(&work->data);
722 
723 	return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
724 }
725 
726 /*
727  * Policy functions.  These define the policies on how the global worker
728  * pools are managed.  Unless noted otherwise, these functions assume that
729  * they're being called with pool->lock held.
730  */
731 
732 static bool __need_more_worker(struct worker_pool *pool)
733 {
734 	return !atomic_read(&pool->nr_running);
735 }
736 
737 /*
738  * Need to wake up a worker?  Called from anything but currently
739  * running workers.
740  *
741  * Note that, because unbound workers never contribute to nr_running, this
742  * function will always return %true for unbound pools as long as the
743  * worklist isn't empty.
744  */
745 static bool need_more_worker(struct worker_pool *pool)
746 {
747 	return !list_empty(&pool->worklist) && __need_more_worker(pool);
748 }
749 
750 /* Can I start working?  Called from busy but !running workers. */
751 static bool may_start_working(struct worker_pool *pool)
752 {
753 	return pool->nr_idle;
754 }
755 
756 /* Do I need to keep working?  Called from currently running workers. */
757 static bool keep_working(struct worker_pool *pool)
758 {
759 	return !list_empty(&pool->worklist) &&
760 		atomic_read(&pool->nr_running) <= 1;
761 }
762 
763 /* Do we need a new worker?  Called from manager. */
764 static bool need_to_create_worker(struct worker_pool *pool)
765 {
766 	return need_more_worker(pool) && !may_start_working(pool);
767 }
768 
769 /* Do we have too many workers and should some go away? */
770 static bool too_many_workers(struct worker_pool *pool)
771 {
772 	bool managing = mutex_is_locked(&pool->manager_arb);
773 	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
774 	int nr_busy = pool->nr_workers - nr_idle;
775 
776 	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
777 }
778 
779 /*
780  * Wake up functions.
781  */
782 
783 /* Return the first idle worker.  Safe with preemption disabled */
784 static struct worker *first_idle_worker(struct worker_pool *pool)
785 {
786 	if (unlikely(list_empty(&pool->idle_list)))
787 		return NULL;
788 
789 	return list_first_entry(&pool->idle_list, struct worker, entry);
790 }
791 
792 /**
793  * wake_up_worker - wake up an idle worker
794  * @pool: worker pool to wake worker from
795  *
796  * Wake up the first idle worker of @pool.
797  *
798  * CONTEXT:
799  * spin_lock_irq(pool->lock).
800  */
801 static void wake_up_worker(struct worker_pool *pool)
802 {
803 	struct worker *worker = first_idle_worker(pool);
804 
805 	if (likely(worker))
806 		wake_up_process(worker->task);
807 }
808 
809 /**
810  * wq_worker_waking_up - a worker is waking up
811  * @task: task waking up
812  * @cpu: CPU @task is waking up to
813  *
814  * This function is called during try_to_wake_up() when a worker is
815  * being awoken.
816  *
817  * CONTEXT:
818  * spin_lock_irq(rq->lock)
819  */
820 void wq_worker_waking_up(struct task_struct *task, int cpu)
821 {
822 	struct worker *worker = kthread_data(task);
823 
824 	if (!(worker->flags & WORKER_NOT_RUNNING)) {
825 		WARN_ON_ONCE(worker->pool->cpu != cpu);
826 		atomic_inc(&worker->pool->nr_running);
827 	}
828 }
829 
830 /**
831  * wq_worker_sleeping - a worker is going to sleep
832  * @task: task going to sleep
833  * @cpu: CPU in question, must be the current CPU number
834  *
835  * This function is called during schedule() when a busy worker is
836  * going to sleep.  Worker on the same cpu can be woken up by
837  * returning pointer to its task.
838  *
839  * CONTEXT:
840  * spin_lock_irq(rq->lock)
841  *
842  * Return:
843  * Worker task on @cpu to wake up, %NULL if none.
844  */
845 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
846 {
847 	struct worker *worker = kthread_data(task), *to_wakeup = NULL;
848 	struct worker_pool *pool;
849 
850 	/*
851 	 * Rescuers, which may not have all the fields set up like normal
852 	 * workers, also reach here, let's not access anything before
853 	 * checking NOT_RUNNING.
854 	 */
855 	if (worker->flags & WORKER_NOT_RUNNING)
856 		return NULL;
857 
858 	pool = worker->pool;
859 
860 	/* this can only happen on the local cpu */
861 	if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu))
862 		return NULL;
863 
864 	/*
865 	 * The counterpart of the following dec_and_test, implied mb,
866 	 * worklist not empty test sequence is in insert_work().
867 	 * Please read comment there.
868 	 *
869 	 * NOT_RUNNING is clear.  This means that we're bound to and
870 	 * running on the local cpu w/ rq lock held and preemption
871 	 * disabled, which in turn means that none else could be
872 	 * manipulating idle_list, so dereferencing idle_list without pool
873 	 * lock is safe.
874 	 */
875 	if (atomic_dec_and_test(&pool->nr_running) &&
876 	    !list_empty(&pool->worklist))
877 		to_wakeup = first_idle_worker(pool);
878 	return to_wakeup ? to_wakeup->task : NULL;
879 }
880 
881 /**
882  * worker_set_flags - set worker flags and adjust nr_running accordingly
883  * @worker: self
884  * @flags: flags to set
885  *
886  * Set @flags in @worker->flags and adjust nr_running accordingly.
887  *
888  * CONTEXT:
889  * spin_lock_irq(pool->lock)
890  */
891 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
892 {
893 	struct worker_pool *pool = worker->pool;
894 
895 	WARN_ON_ONCE(worker->task != current);
896 
897 	/* If transitioning into NOT_RUNNING, adjust nr_running. */
898 	if ((flags & WORKER_NOT_RUNNING) &&
899 	    !(worker->flags & WORKER_NOT_RUNNING)) {
900 		atomic_dec(&pool->nr_running);
901 	}
902 
903 	worker->flags |= flags;
904 }
905 
906 /**
907  * worker_clr_flags - clear worker flags and adjust nr_running accordingly
908  * @worker: self
909  * @flags: flags to clear
910  *
911  * Clear @flags in @worker->flags and adjust nr_running accordingly.
912  *
913  * CONTEXT:
914  * spin_lock_irq(pool->lock)
915  */
916 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
917 {
918 	struct worker_pool *pool = worker->pool;
919 	unsigned int oflags = worker->flags;
920 
921 	WARN_ON_ONCE(worker->task != current);
922 
923 	worker->flags &= ~flags;
924 
925 	/*
926 	 * If transitioning out of NOT_RUNNING, increment nr_running.  Note
927 	 * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
928 	 * of multiple flags, not a single flag.
929 	 */
930 	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
931 		if (!(worker->flags & WORKER_NOT_RUNNING))
932 			atomic_inc(&pool->nr_running);
933 }
934 
935 /**
936  * find_worker_executing_work - find worker which is executing a work
937  * @pool: pool of interest
938  * @work: work to find worker for
939  *
940  * Find a worker which is executing @work on @pool by searching
941  * @pool->busy_hash which is keyed by the address of @work.  For a worker
942  * to match, its current execution should match the address of @work and
943  * its work function.  This is to avoid unwanted dependency between
944  * unrelated work executions through a work item being recycled while still
945  * being executed.
946  *
947  * This is a bit tricky.  A work item may be freed once its execution
948  * starts and nothing prevents the freed area from being recycled for
949  * another work item.  If the same work item address ends up being reused
950  * before the original execution finishes, workqueue will identify the
951  * recycled work item as currently executing and make it wait until the
952  * current execution finishes, introducing an unwanted dependency.
953  *
954  * This function checks the work item address and work function to avoid
955  * false positives.  Note that this isn't complete as one may construct a
956  * work function which can introduce dependency onto itself through a
957  * recycled work item.  Well, if somebody wants to shoot oneself in the
958  * foot that badly, there's only so much we can do, and if such deadlock
959  * actually occurs, it should be easy to locate the culprit work function.
960  *
961  * CONTEXT:
962  * spin_lock_irq(pool->lock).
963  *
964  * Return:
965  * Pointer to worker which is executing @work if found, %NULL
966  * otherwise.
967  */
968 static struct worker *find_worker_executing_work(struct worker_pool *pool,
969 						 struct work_struct *work)
970 {
971 	struct worker *worker;
972 
973 	hash_for_each_possible(pool->busy_hash, worker, hentry,
974 			       (unsigned long)work)
975 		if (worker->current_work == work &&
976 		    worker->current_func == work->func)
977 			return worker;
978 
979 	return NULL;
980 }
981 
982 /**
983  * move_linked_works - move linked works to a list
984  * @work: start of series of works to be scheduled
985  * @head: target list to append @work to
986  * @nextp: out parameter for nested worklist walking
987  *
988  * Schedule linked works starting from @work to @head.  Work series to
989  * be scheduled starts at @work and includes any consecutive work with
990  * WORK_STRUCT_LINKED set in its predecessor.
991  *
992  * If @nextp is not NULL, it's updated to point to the next work of
993  * the last scheduled work.  This allows move_linked_works() to be
994  * nested inside outer list_for_each_entry_safe().
995  *
996  * CONTEXT:
997  * spin_lock_irq(pool->lock).
998  */
999 static void move_linked_works(struct work_struct *work, struct list_head *head,
1000 			      struct work_struct **nextp)
1001 {
1002 	struct work_struct *n;
1003 
1004 	/*
1005 	 * Linked worklist will always end before the end of the list,
1006 	 * use NULL for list head.
1007 	 */
1008 	list_for_each_entry_safe_from(work, n, NULL, entry) {
1009 		list_move_tail(&work->entry, head);
1010 		if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1011 			break;
1012 	}
1013 
1014 	/*
1015 	 * If we're already inside safe list traversal and have moved
1016 	 * multiple works to the scheduled queue, the next position
1017 	 * needs to be updated.
1018 	 */
1019 	if (nextp)
1020 		*nextp = n;
1021 }
1022 
1023 /**
1024  * get_pwq - get an extra reference on the specified pool_workqueue
1025  * @pwq: pool_workqueue to get
1026  *
1027  * Obtain an extra reference on @pwq.  The caller should guarantee that
1028  * @pwq has positive refcnt and be holding the matching pool->lock.
1029  */
1030 static void get_pwq(struct pool_workqueue *pwq)
1031 {
1032 	lockdep_assert_held(&pwq->pool->lock);
1033 	WARN_ON_ONCE(pwq->refcnt <= 0);
1034 	pwq->refcnt++;
1035 }
1036 
1037 /**
1038  * put_pwq - put a pool_workqueue reference
1039  * @pwq: pool_workqueue to put
1040  *
1041  * Drop a reference of @pwq.  If its refcnt reaches zero, schedule its
1042  * destruction.  The caller should be holding the matching pool->lock.
1043  */
1044 static void put_pwq(struct pool_workqueue *pwq)
1045 {
1046 	lockdep_assert_held(&pwq->pool->lock);
1047 	if (likely(--pwq->refcnt))
1048 		return;
1049 	if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1050 		return;
1051 	/*
1052 	 * @pwq can't be released under pool->lock, bounce to
1053 	 * pwq_unbound_release_workfn().  This never recurses on the same
1054 	 * pool->lock as this path is taken only for unbound workqueues and
1055 	 * the release work item is scheduled on a per-cpu workqueue.  To
1056 	 * avoid lockdep warning, unbound pool->locks are given lockdep
1057 	 * subclass of 1 in get_unbound_pool().
1058 	 */
1059 	schedule_work(&pwq->unbound_release_work);
1060 }
1061 
1062 /**
1063  * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1064  * @pwq: pool_workqueue to put (can be %NULL)
1065  *
1066  * put_pwq() with locking.  This function also allows %NULL @pwq.
1067  */
1068 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1069 {
1070 	if (pwq) {
1071 		/*
1072 		 * As both pwqs and pools are sched-RCU protected, the
1073 		 * following lock operations are safe.
1074 		 */
1075 		spin_lock_irq(&pwq->pool->lock);
1076 		put_pwq(pwq);
1077 		spin_unlock_irq(&pwq->pool->lock);
1078 	}
1079 }
1080 
1081 static void pwq_activate_delayed_work(struct work_struct *work)
1082 {
1083 	struct pool_workqueue *pwq = get_work_pwq(work);
1084 
1085 	trace_workqueue_activate_work(work);
1086 	move_linked_works(work, &pwq->pool->worklist, NULL);
1087 	__clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1088 	pwq->nr_active++;
1089 }
1090 
1091 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1092 {
1093 	struct work_struct *work = list_first_entry(&pwq->delayed_works,
1094 						    struct work_struct, entry);
1095 
1096 	pwq_activate_delayed_work(work);
1097 }
1098 
1099 /**
1100  * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1101  * @pwq: pwq of interest
1102  * @color: color of work which left the queue
1103  *
1104  * A work either has completed or is removed from pending queue,
1105  * decrement nr_in_flight of its pwq and handle workqueue flushing.
1106  *
1107  * CONTEXT:
1108  * spin_lock_irq(pool->lock).
1109  */
1110 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1111 {
1112 	/* uncolored work items don't participate in flushing or nr_active */
1113 	if (color == WORK_NO_COLOR)
1114 		goto out_put;
1115 
1116 	pwq->nr_in_flight[color]--;
1117 
1118 	pwq->nr_active--;
1119 	if (!list_empty(&pwq->delayed_works)) {
1120 		/* one down, submit a delayed one */
1121 		if (pwq->nr_active < pwq->max_active)
1122 			pwq_activate_first_delayed(pwq);
1123 	}
1124 
1125 	/* is flush in progress and are we at the flushing tip? */
1126 	if (likely(pwq->flush_color != color))
1127 		goto out_put;
1128 
1129 	/* are there still in-flight works? */
1130 	if (pwq->nr_in_flight[color])
1131 		goto out_put;
1132 
1133 	/* this pwq is done, clear flush_color */
1134 	pwq->flush_color = -1;
1135 
1136 	/*
1137 	 * If this was the last pwq, wake up the first flusher.  It
1138 	 * will handle the rest.
1139 	 */
1140 	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1141 		complete(&pwq->wq->first_flusher->done);
1142 out_put:
1143 	put_pwq(pwq);
1144 }
1145 
1146 /**
1147  * try_to_grab_pending - steal work item from worklist and disable irq
1148  * @work: work item to steal
1149  * @is_dwork: @work is a delayed_work
1150  * @flags: place to store irq state
1151  *
1152  * Try to grab PENDING bit of @work.  This function can handle @work in any
1153  * stable state - idle, on timer or on worklist.
1154  *
1155  * Return:
1156  *  1		if @work was pending and we successfully stole PENDING
1157  *  0		if @work was idle and we claimed PENDING
1158  *  -EAGAIN	if PENDING couldn't be grabbed at the moment, safe to busy-retry
1159  *  -ENOENT	if someone else is canceling @work, this state may persist
1160  *		for arbitrarily long
1161  *
1162  * Note:
1163  * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1164  * interrupted while holding PENDING and @work off queue, irq must be
1165  * disabled on entry.  This, combined with delayed_work->timer being
1166  * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1167  *
1168  * On successful return, >= 0, irq is disabled and the caller is
1169  * responsible for releasing it using local_irq_restore(*@flags).
1170  *
1171  * This function is safe to call from any context including IRQ handler.
1172  */
1173 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1174 			       unsigned long *flags)
1175 {
1176 	struct worker_pool *pool;
1177 	struct pool_workqueue *pwq;
1178 
1179 	local_irq_save(*flags);
1180 
1181 	/* try to steal the timer if it exists */
1182 	if (is_dwork) {
1183 		struct delayed_work *dwork = to_delayed_work(work);
1184 
1185 		/*
1186 		 * dwork->timer is irqsafe.  If del_timer() fails, it's
1187 		 * guaranteed that the timer is not queued anywhere and not
1188 		 * running on the local CPU.
1189 		 */
1190 		if (likely(del_timer(&dwork->timer)))
1191 			return 1;
1192 	}
1193 
1194 	/* try to claim PENDING the normal way */
1195 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1196 		return 0;
1197 
1198 	/*
1199 	 * The queueing is in progress, or it is already queued. Try to
1200 	 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1201 	 */
1202 	pool = get_work_pool(work);
1203 	if (!pool)
1204 		goto fail;
1205 
1206 	spin_lock(&pool->lock);
1207 	/*
1208 	 * work->data is guaranteed to point to pwq only while the work
1209 	 * item is queued on pwq->wq, and both updating work->data to point
1210 	 * to pwq on queueing and to pool on dequeueing are done under
1211 	 * pwq->pool->lock.  This in turn guarantees that, if work->data
1212 	 * points to pwq which is associated with a locked pool, the work
1213 	 * item is currently queued on that pool.
1214 	 */
1215 	pwq = get_work_pwq(work);
1216 	if (pwq && pwq->pool == pool) {
1217 		debug_work_deactivate(work);
1218 
1219 		/*
1220 		 * A delayed work item cannot be grabbed directly because
1221 		 * it might have linked NO_COLOR work items which, if left
1222 		 * on the delayed_list, will confuse pwq->nr_active
1223 		 * management later on and cause stall.  Make sure the work
1224 		 * item is activated before grabbing.
1225 		 */
1226 		if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1227 			pwq_activate_delayed_work(work);
1228 
1229 		list_del_init(&work->entry);
1230 		pwq_dec_nr_in_flight(pwq, get_work_color(work));
1231 
1232 		/* work->data points to pwq iff queued, point to pool */
1233 		set_work_pool_and_keep_pending(work, pool->id);
1234 
1235 		spin_unlock(&pool->lock);
1236 		return 1;
1237 	}
1238 	spin_unlock(&pool->lock);
1239 fail:
1240 	local_irq_restore(*flags);
1241 	if (work_is_canceling(work))
1242 		return -ENOENT;
1243 	cpu_relax();
1244 	return -EAGAIN;
1245 }
1246 
1247 /**
1248  * insert_work - insert a work into a pool
1249  * @pwq: pwq @work belongs to
1250  * @work: work to insert
1251  * @head: insertion point
1252  * @extra_flags: extra WORK_STRUCT_* flags to set
1253  *
1254  * Insert @work which belongs to @pwq after @head.  @extra_flags is or'd to
1255  * work_struct flags.
1256  *
1257  * CONTEXT:
1258  * spin_lock_irq(pool->lock).
1259  */
1260 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1261 			struct list_head *head, unsigned int extra_flags)
1262 {
1263 	struct worker_pool *pool = pwq->pool;
1264 
1265 	/* we own @work, set data and link */
1266 	set_work_pwq(work, pwq, extra_flags);
1267 	list_add_tail(&work->entry, head);
1268 	get_pwq(pwq);
1269 
1270 	/*
1271 	 * Ensure either wq_worker_sleeping() sees the above
1272 	 * list_add_tail() or we see zero nr_running to avoid workers lying
1273 	 * around lazily while there are works to be processed.
1274 	 */
1275 	smp_mb();
1276 
1277 	if (__need_more_worker(pool))
1278 		wake_up_worker(pool);
1279 }
1280 
1281 /*
1282  * Test whether @work is being queued from another work executing on the
1283  * same workqueue.
1284  */
1285 static bool is_chained_work(struct workqueue_struct *wq)
1286 {
1287 	struct worker *worker;
1288 
1289 	worker = current_wq_worker();
1290 	/*
1291 	 * Return %true iff I'm a worker execuing a work item on @wq.  If
1292 	 * I'm @worker, it's safe to dereference it without locking.
1293 	 */
1294 	return worker && worker->current_pwq->wq == wq;
1295 }
1296 
1297 static void __queue_work(int cpu, struct workqueue_struct *wq,
1298 			 struct work_struct *work)
1299 {
1300 	struct pool_workqueue *pwq;
1301 	struct worker_pool *last_pool;
1302 	struct list_head *worklist;
1303 	unsigned int work_flags;
1304 	unsigned int req_cpu = cpu;
1305 
1306 	/*
1307 	 * While a work item is PENDING && off queue, a task trying to
1308 	 * steal the PENDING will busy-loop waiting for it to either get
1309 	 * queued or lose PENDING.  Grabbing PENDING and queueing should
1310 	 * happen with IRQ disabled.
1311 	 */
1312 	WARN_ON_ONCE(!irqs_disabled());
1313 
1314 	debug_work_activate(work);
1315 
1316 	/* if draining, only works from the same workqueue are allowed */
1317 	if (unlikely(wq->flags & __WQ_DRAINING) &&
1318 	    WARN_ON_ONCE(!is_chained_work(wq)))
1319 		return;
1320 retry:
1321 	if (req_cpu == WORK_CPU_UNBOUND)
1322 		cpu = raw_smp_processor_id();
1323 
1324 	/* pwq which will be used unless @work is executing elsewhere */
1325 	if (!(wq->flags & WQ_UNBOUND))
1326 		pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1327 	else
1328 		pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1329 
1330 	/*
1331 	 * If @work was previously on a different pool, it might still be
1332 	 * running there, in which case the work needs to be queued on that
1333 	 * pool to guarantee non-reentrancy.
1334 	 */
1335 	last_pool = get_work_pool(work);
1336 	if (last_pool && last_pool != pwq->pool) {
1337 		struct worker *worker;
1338 
1339 		spin_lock(&last_pool->lock);
1340 
1341 		worker = find_worker_executing_work(last_pool, work);
1342 
1343 		if (worker && worker->current_pwq->wq == wq) {
1344 			pwq = worker->current_pwq;
1345 		} else {
1346 			/* meh... not running there, queue here */
1347 			spin_unlock(&last_pool->lock);
1348 			spin_lock(&pwq->pool->lock);
1349 		}
1350 	} else {
1351 		spin_lock(&pwq->pool->lock);
1352 	}
1353 
1354 	/*
1355 	 * pwq is determined and locked.  For unbound pools, we could have
1356 	 * raced with pwq release and it could already be dead.  If its
1357 	 * refcnt is zero, repeat pwq selection.  Note that pwqs never die
1358 	 * without another pwq replacing it in the numa_pwq_tbl or while
1359 	 * work items are executing on it, so the retrying is guaranteed to
1360 	 * make forward-progress.
1361 	 */
1362 	if (unlikely(!pwq->refcnt)) {
1363 		if (wq->flags & WQ_UNBOUND) {
1364 			spin_unlock(&pwq->pool->lock);
1365 			cpu_relax();
1366 			goto retry;
1367 		}
1368 		/* oops */
1369 		WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1370 			  wq->name, cpu);
1371 	}
1372 
1373 	/* pwq determined, queue */
1374 	trace_workqueue_queue_work(req_cpu, pwq, work);
1375 
1376 	if (WARN_ON(!list_empty(&work->entry))) {
1377 		spin_unlock(&pwq->pool->lock);
1378 		return;
1379 	}
1380 
1381 	pwq->nr_in_flight[pwq->work_color]++;
1382 	work_flags = work_color_to_flags(pwq->work_color);
1383 
1384 	if (likely(pwq->nr_active < pwq->max_active)) {
1385 		trace_workqueue_activate_work(work);
1386 		pwq->nr_active++;
1387 		worklist = &pwq->pool->worklist;
1388 	} else {
1389 		work_flags |= WORK_STRUCT_DELAYED;
1390 		worklist = &pwq->delayed_works;
1391 	}
1392 
1393 	insert_work(pwq, work, worklist, work_flags);
1394 
1395 	spin_unlock(&pwq->pool->lock);
1396 }
1397 
1398 /**
1399  * queue_work_on - queue work on specific cpu
1400  * @cpu: CPU number to execute work on
1401  * @wq: workqueue to use
1402  * @work: work to queue
1403  *
1404  * We queue the work to a specific CPU, the caller must ensure it
1405  * can't go away.
1406  *
1407  * Return: %false if @work was already on a queue, %true otherwise.
1408  */
1409 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1410 		   struct work_struct *work)
1411 {
1412 	bool ret = false;
1413 	unsigned long flags;
1414 
1415 	local_irq_save(flags);
1416 
1417 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1418 		__queue_work(cpu, wq, work);
1419 		ret = true;
1420 	}
1421 
1422 	local_irq_restore(flags);
1423 	return ret;
1424 }
1425 EXPORT_SYMBOL(queue_work_on);
1426 
1427 void delayed_work_timer_fn(unsigned long __data)
1428 {
1429 	struct delayed_work *dwork = (struct delayed_work *)__data;
1430 
1431 	/* should have been called from irqsafe timer with irq already off */
1432 	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
1433 }
1434 EXPORT_SYMBOL(delayed_work_timer_fn);
1435 
1436 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1437 				struct delayed_work *dwork, unsigned long delay)
1438 {
1439 	struct timer_list *timer = &dwork->timer;
1440 	struct work_struct *work = &dwork->work;
1441 
1442 	WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1443 		     timer->data != (unsigned long)dwork);
1444 	WARN_ON_ONCE(timer_pending(timer));
1445 	WARN_ON_ONCE(!list_empty(&work->entry));
1446 
1447 	/*
1448 	 * If @delay is 0, queue @dwork->work immediately.  This is for
1449 	 * both optimization and correctness.  The earliest @timer can
1450 	 * expire is on the closest next tick and delayed_work users depend
1451 	 * on that there's no such delay when @delay is 0.
1452 	 */
1453 	if (!delay) {
1454 		__queue_work(cpu, wq, &dwork->work);
1455 		return;
1456 	}
1457 
1458 	timer_stats_timer_set_start_info(&dwork->timer);
1459 
1460 	dwork->wq = wq;
1461 	/* timer isn't guaranteed to run in this cpu, record earlier */
1462 	if (cpu == WORK_CPU_UNBOUND)
1463 		cpu = raw_smp_processor_id();
1464 	dwork->cpu = cpu;
1465 	timer->expires = jiffies + delay;
1466 
1467 	add_timer_on(timer, cpu);
1468 }
1469 
1470 /**
1471  * queue_delayed_work_on - queue work on specific CPU after delay
1472  * @cpu: CPU number to execute work on
1473  * @wq: workqueue to use
1474  * @dwork: work to queue
1475  * @delay: number of jiffies to wait before queueing
1476  *
1477  * Return: %false if @work was already on a queue, %true otherwise.  If
1478  * @delay is zero and @dwork is idle, it will be scheduled for immediate
1479  * execution.
1480  */
1481 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1482 			   struct delayed_work *dwork, unsigned long delay)
1483 {
1484 	struct work_struct *work = &dwork->work;
1485 	bool ret = false;
1486 	unsigned long flags;
1487 
1488 	/* read the comment in __queue_work() */
1489 	local_irq_save(flags);
1490 
1491 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1492 		__queue_delayed_work(cpu, wq, dwork, delay);
1493 		ret = true;
1494 	}
1495 
1496 	local_irq_restore(flags);
1497 	return ret;
1498 }
1499 EXPORT_SYMBOL(queue_delayed_work_on);
1500 
1501 /**
1502  * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1503  * @cpu: CPU number to execute work on
1504  * @wq: workqueue to use
1505  * @dwork: work to queue
1506  * @delay: number of jiffies to wait before queueing
1507  *
1508  * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1509  * modify @dwork's timer so that it expires after @delay.  If @delay is
1510  * zero, @work is guaranteed to be scheduled immediately regardless of its
1511  * current state.
1512  *
1513  * Return: %false if @dwork was idle and queued, %true if @dwork was
1514  * pending and its timer was modified.
1515  *
1516  * This function is safe to call from any context including IRQ handler.
1517  * See try_to_grab_pending() for details.
1518  */
1519 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1520 			 struct delayed_work *dwork, unsigned long delay)
1521 {
1522 	unsigned long flags;
1523 	int ret;
1524 
1525 	do {
1526 		ret = try_to_grab_pending(&dwork->work, true, &flags);
1527 	} while (unlikely(ret == -EAGAIN));
1528 
1529 	if (likely(ret >= 0)) {
1530 		__queue_delayed_work(cpu, wq, dwork, delay);
1531 		local_irq_restore(flags);
1532 	}
1533 
1534 	/* -ENOENT from try_to_grab_pending() becomes %true */
1535 	return ret;
1536 }
1537 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1538 
1539 /**
1540  * worker_enter_idle - enter idle state
1541  * @worker: worker which is entering idle state
1542  *
1543  * @worker is entering idle state.  Update stats and idle timer if
1544  * necessary.
1545  *
1546  * LOCKING:
1547  * spin_lock_irq(pool->lock).
1548  */
1549 static void worker_enter_idle(struct worker *worker)
1550 {
1551 	struct worker_pool *pool = worker->pool;
1552 
1553 	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1554 	    WARN_ON_ONCE(!list_empty(&worker->entry) &&
1555 			 (worker->hentry.next || worker->hentry.pprev)))
1556 		return;
1557 
1558 	/* can't use worker_set_flags(), also called from create_worker() */
1559 	worker->flags |= WORKER_IDLE;
1560 	pool->nr_idle++;
1561 	worker->last_active = jiffies;
1562 
1563 	/* idle_list is LIFO */
1564 	list_add(&worker->entry, &pool->idle_list);
1565 
1566 	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1567 		mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1568 
1569 	/*
1570 	 * Sanity check nr_running.  Because wq_unbind_fn() releases
1571 	 * pool->lock between setting %WORKER_UNBOUND and zapping
1572 	 * nr_running, the warning may trigger spuriously.  Check iff
1573 	 * unbind is not in progress.
1574 	 */
1575 	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1576 		     pool->nr_workers == pool->nr_idle &&
1577 		     atomic_read(&pool->nr_running));
1578 }
1579 
1580 /**
1581  * worker_leave_idle - leave idle state
1582  * @worker: worker which is leaving idle state
1583  *
1584  * @worker is leaving idle state.  Update stats.
1585  *
1586  * LOCKING:
1587  * spin_lock_irq(pool->lock).
1588  */
1589 static void worker_leave_idle(struct worker *worker)
1590 {
1591 	struct worker_pool *pool = worker->pool;
1592 
1593 	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1594 		return;
1595 	worker_clr_flags(worker, WORKER_IDLE);
1596 	pool->nr_idle--;
1597 	list_del_init(&worker->entry);
1598 }
1599 
1600 static struct worker *alloc_worker(int node)
1601 {
1602 	struct worker *worker;
1603 
1604 	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1605 	if (worker) {
1606 		INIT_LIST_HEAD(&worker->entry);
1607 		INIT_LIST_HEAD(&worker->scheduled);
1608 		INIT_LIST_HEAD(&worker->node);
1609 		/* on creation a worker is in !idle && prep state */
1610 		worker->flags = WORKER_PREP;
1611 	}
1612 	return worker;
1613 }
1614 
1615 /**
1616  * worker_attach_to_pool() - attach a worker to a pool
1617  * @worker: worker to be attached
1618  * @pool: the target pool
1619  *
1620  * Attach @worker to @pool.  Once attached, the %WORKER_UNBOUND flag and
1621  * cpu-binding of @worker are kept coordinated with the pool across
1622  * cpu-[un]hotplugs.
1623  */
1624 static void worker_attach_to_pool(struct worker *worker,
1625 				   struct worker_pool *pool)
1626 {
1627 	mutex_lock(&pool->attach_mutex);
1628 
1629 	/*
1630 	 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1631 	 * online CPUs.  It'll be re-applied when any of the CPUs come up.
1632 	 */
1633 	set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1634 
1635 	/*
1636 	 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1637 	 * stable across this function.  See the comments above the
1638 	 * flag definition for details.
1639 	 */
1640 	if (pool->flags & POOL_DISASSOCIATED)
1641 		worker->flags |= WORKER_UNBOUND;
1642 
1643 	list_add_tail(&worker->node, &pool->workers);
1644 
1645 	mutex_unlock(&pool->attach_mutex);
1646 }
1647 
1648 /**
1649  * worker_detach_from_pool() - detach a worker from its pool
1650  * @worker: worker which is attached to its pool
1651  * @pool: the pool @worker is attached to
1652  *
1653  * Undo the attaching which had been done in worker_attach_to_pool().  The
1654  * caller worker shouldn't access to the pool after detached except it has
1655  * other reference to the pool.
1656  */
1657 static void worker_detach_from_pool(struct worker *worker,
1658 				    struct worker_pool *pool)
1659 {
1660 	struct completion *detach_completion = NULL;
1661 
1662 	mutex_lock(&pool->attach_mutex);
1663 	list_del(&worker->node);
1664 	if (list_empty(&pool->workers))
1665 		detach_completion = pool->detach_completion;
1666 	mutex_unlock(&pool->attach_mutex);
1667 
1668 	/* clear leftover flags without pool->lock after it is detached */
1669 	worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1670 
1671 	if (detach_completion)
1672 		complete(detach_completion);
1673 }
1674 
1675 /**
1676  * create_worker - create a new workqueue worker
1677  * @pool: pool the new worker will belong to
1678  *
1679  * Create and start a new worker which is attached to @pool.
1680  *
1681  * CONTEXT:
1682  * Might sleep.  Does GFP_KERNEL allocations.
1683  *
1684  * Return:
1685  * Pointer to the newly created worker.
1686  */
1687 static struct worker *create_worker(struct worker_pool *pool)
1688 {
1689 	struct worker *worker = NULL;
1690 	int id = -1;
1691 	char id_buf[16];
1692 
1693 	/* ID is needed to determine kthread name */
1694 	id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1695 	if (id < 0)
1696 		goto fail;
1697 
1698 	worker = alloc_worker(pool->node);
1699 	if (!worker)
1700 		goto fail;
1701 
1702 	worker->pool = pool;
1703 	worker->id = id;
1704 
1705 	if (pool->cpu >= 0)
1706 		snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1707 			 pool->attrs->nice < 0  ? "H" : "");
1708 	else
1709 		snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1710 
1711 	worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1712 					      "kworker/%s", id_buf);
1713 	if (IS_ERR(worker->task))
1714 		goto fail;
1715 
1716 	set_user_nice(worker->task, pool->attrs->nice);
1717 	kthread_bind_mask(worker->task, pool->attrs->cpumask);
1718 
1719 	/* successful, attach the worker to the pool */
1720 	worker_attach_to_pool(worker, pool);
1721 
1722 	/* start the newly created worker */
1723 	spin_lock_irq(&pool->lock);
1724 	worker->pool->nr_workers++;
1725 	worker_enter_idle(worker);
1726 	wake_up_process(worker->task);
1727 	spin_unlock_irq(&pool->lock);
1728 
1729 	return worker;
1730 
1731 fail:
1732 	if (id >= 0)
1733 		ida_simple_remove(&pool->worker_ida, id);
1734 	kfree(worker);
1735 	return NULL;
1736 }
1737 
1738 /**
1739  * destroy_worker - destroy a workqueue worker
1740  * @worker: worker to be destroyed
1741  *
1742  * Destroy @worker and adjust @pool stats accordingly.  The worker should
1743  * be idle.
1744  *
1745  * CONTEXT:
1746  * spin_lock_irq(pool->lock).
1747  */
1748 static void destroy_worker(struct worker *worker)
1749 {
1750 	struct worker_pool *pool = worker->pool;
1751 
1752 	lockdep_assert_held(&pool->lock);
1753 
1754 	/* sanity check frenzy */
1755 	if (WARN_ON(worker->current_work) ||
1756 	    WARN_ON(!list_empty(&worker->scheduled)) ||
1757 	    WARN_ON(!(worker->flags & WORKER_IDLE)))
1758 		return;
1759 
1760 	pool->nr_workers--;
1761 	pool->nr_idle--;
1762 
1763 	list_del_init(&worker->entry);
1764 	worker->flags |= WORKER_DIE;
1765 	wake_up_process(worker->task);
1766 }
1767 
1768 static void idle_worker_timeout(unsigned long __pool)
1769 {
1770 	struct worker_pool *pool = (void *)__pool;
1771 
1772 	spin_lock_irq(&pool->lock);
1773 
1774 	while (too_many_workers(pool)) {
1775 		struct worker *worker;
1776 		unsigned long expires;
1777 
1778 		/* idle_list is kept in LIFO order, check the last one */
1779 		worker = list_entry(pool->idle_list.prev, struct worker, entry);
1780 		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1781 
1782 		if (time_before(jiffies, expires)) {
1783 			mod_timer(&pool->idle_timer, expires);
1784 			break;
1785 		}
1786 
1787 		destroy_worker(worker);
1788 	}
1789 
1790 	spin_unlock_irq(&pool->lock);
1791 }
1792 
1793 static void send_mayday(struct work_struct *work)
1794 {
1795 	struct pool_workqueue *pwq = get_work_pwq(work);
1796 	struct workqueue_struct *wq = pwq->wq;
1797 
1798 	lockdep_assert_held(&wq_mayday_lock);
1799 
1800 	if (!wq->rescuer)
1801 		return;
1802 
1803 	/* mayday mayday mayday */
1804 	if (list_empty(&pwq->mayday_node)) {
1805 		/*
1806 		 * If @pwq is for an unbound wq, its base ref may be put at
1807 		 * any time due to an attribute change.  Pin @pwq until the
1808 		 * rescuer is done with it.
1809 		 */
1810 		get_pwq(pwq);
1811 		list_add_tail(&pwq->mayday_node, &wq->maydays);
1812 		wake_up_process(wq->rescuer->task);
1813 	}
1814 }
1815 
1816 static void pool_mayday_timeout(unsigned long __pool)
1817 {
1818 	struct worker_pool *pool = (void *)__pool;
1819 	struct work_struct *work;
1820 
1821 	spin_lock_irq(&pool->lock);
1822 	spin_lock(&wq_mayday_lock);		/* for wq->maydays */
1823 
1824 	if (need_to_create_worker(pool)) {
1825 		/*
1826 		 * We've been trying to create a new worker but
1827 		 * haven't been successful.  We might be hitting an
1828 		 * allocation deadlock.  Send distress signals to
1829 		 * rescuers.
1830 		 */
1831 		list_for_each_entry(work, &pool->worklist, entry)
1832 			send_mayday(work);
1833 	}
1834 
1835 	spin_unlock(&wq_mayday_lock);
1836 	spin_unlock_irq(&pool->lock);
1837 
1838 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1839 }
1840 
1841 /**
1842  * maybe_create_worker - create a new worker if necessary
1843  * @pool: pool to create a new worker for
1844  *
1845  * Create a new worker for @pool if necessary.  @pool is guaranteed to
1846  * have at least one idle worker on return from this function.  If
1847  * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1848  * sent to all rescuers with works scheduled on @pool to resolve
1849  * possible allocation deadlock.
1850  *
1851  * On return, need_to_create_worker() is guaranteed to be %false and
1852  * may_start_working() %true.
1853  *
1854  * LOCKING:
1855  * spin_lock_irq(pool->lock) which may be released and regrabbed
1856  * multiple times.  Does GFP_KERNEL allocations.  Called only from
1857  * manager.
1858  */
1859 static void maybe_create_worker(struct worker_pool *pool)
1860 __releases(&pool->lock)
1861 __acquires(&pool->lock)
1862 {
1863 restart:
1864 	spin_unlock_irq(&pool->lock);
1865 
1866 	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1867 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1868 
1869 	while (true) {
1870 		if (create_worker(pool) || !need_to_create_worker(pool))
1871 			break;
1872 
1873 		schedule_timeout_interruptible(CREATE_COOLDOWN);
1874 
1875 		if (!need_to_create_worker(pool))
1876 			break;
1877 	}
1878 
1879 	del_timer_sync(&pool->mayday_timer);
1880 	spin_lock_irq(&pool->lock);
1881 	/*
1882 	 * This is necessary even after a new worker was just successfully
1883 	 * created as @pool->lock was dropped and the new worker might have
1884 	 * already become busy.
1885 	 */
1886 	if (need_to_create_worker(pool))
1887 		goto restart;
1888 }
1889 
1890 /**
1891  * manage_workers - manage worker pool
1892  * @worker: self
1893  *
1894  * Assume the manager role and manage the worker pool @worker belongs
1895  * to.  At any given time, there can be only zero or one manager per
1896  * pool.  The exclusion is handled automatically by this function.
1897  *
1898  * The caller can safely start processing works on false return.  On
1899  * true return, it's guaranteed that need_to_create_worker() is false
1900  * and may_start_working() is true.
1901  *
1902  * CONTEXT:
1903  * spin_lock_irq(pool->lock) which may be released and regrabbed
1904  * multiple times.  Does GFP_KERNEL allocations.
1905  *
1906  * Return:
1907  * %false if the pool doesn't need management and the caller can safely
1908  * start processing works, %true if management function was performed and
1909  * the conditions that the caller verified before calling the function may
1910  * no longer be true.
1911  */
1912 static bool manage_workers(struct worker *worker)
1913 {
1914 	struct worker_pool *pool = worker->pool;
1915 
1916 	/*
1917 	 * Anyone who successfully grabs manager_arb wins the arbitration
1918 	 * and becomes the manager.  mutex_trylock() on pool->manager_arb
1919 	 * failure while holding pool->lock reliably indicates that someone
1920 	 * else is managing the pool and the worker which failed trylock
1921 	 * can proceed to executing work items.  This means that anyone
1922 	 * grabbing manager_arb is responsible for actually performing
1923 	 * manager duties.  If manager_arb is grabbed and released without
1924 	 * actual management, the pool may stall indefinitely.
1925 	 */
1926 	if (!mutex_trylock(&pool->manager_arb))
1927 		return false;
1928 	pool->manager = worker;
1929 
1930 	maybe_create_worker(pool);
1931 
1932 	pool->manager = NULL;
1933 	mutex_unlock(&pool->manager_arb);
1934 	return true;
1935 }
1936 
1937 /**
1938  * process_one_work - process single work
1939  * @worker: self
1940  * @work: work to process
1941  *
1942  * Process @work.  This function contains all the logics necessary to
1943  * process a single work including synchronization against and
1944  * interaction with other workers on the same cpu, queueing and
1945  * flushing.  As long as context requirement is met, any worker can
1946  * call this function to process a work.
1947  *
1948  * CONTEXT:
1949  * spin_lock_irq(pool->lock) which is released and regrabbed.
1950  */
1951 static void process_one_work(struct worker *worker, struct work_struct *work)
1952 __releases(&pool->lock)
1953 __acquires(&pool->lock)
1954 {
1955 	struct pool_workqueue *pwq = get_work_pwq(work);
1956 	struct worker_pool *pool = worker->pool;
1957 	bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
1958 	int work_color;
1959 	struct worker *collision;
1960 #ifdef CONFIG_LOCKDEP
1961 	/*
1962 	 * It is permissible to free the struct work_struct from
1963 	 * inside the function that is called from it, this we need to
1964 	 * take into account for lockdep too.  To avoid bogus "held
1965 	 * lock freed" warnings as well as problems when looking into
1966 	 * work->lockdep_map, make a copy and use that here.
1967 	 */
1968 	struct lockdep_map lockdep_map;
1969 
1970 	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
1971 #endif
1972 	/* ensure we're on the correct CPU */
1973 	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1974 		     raw_smp_processor_id() != pool->cpu);
1975 
1976 	/*
1977 	 * A single work shouldn't be executed concurrently by
1978 	 * multiple workers on a single cpu.  Check whether anyone is
1979 	 * already processing the work.  If so, defer the work to the
1980 	 * currently executing one.
1981 	 */
1982 	collision = find_worker_executing_work(pool, work);
1983 	if (unlikely(collision)) {
1984 		move_linked_works(work, &collision->scheduled, NULL);
1985 		return;
1986 	}
1987 
1988 	/* claim and dequeue */
1989 	debug_work_deactivate(work);
1990 	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
1991 	worker->current_work = work;
1992 	worker->current_func = work->func;
1993 	worker->current_pwq = pwq;
1994 	work_color = get_work_color(work);
1995 
1996 	list_del_init(&work->entry);
1997 
1998 	/*
1999 	 * CPU intensive works don't participate in concurrency management.
2000 	 * They're the scheduler's responsibility.  This takes @worker out
2001 	 * of concurrency management and the next code block will chain
2002 	 * execution of the pending work items.
2003 	 */
2004 	if (unlikely(cpu_intensive))
2005 		worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2006 
2007 	/*
2008 	 * Wake up another worker if necessary.  The condition is always
2009 	 * false for normal per-cpu workers since nr_running would always
2010 	 * be >= 1 at this point.  This is used to chain execution of the
2011 	 * pending work items for WORKER_NOT_RUNNING workers such as the
2012 	 * UNBOUND and CPU_INTENSIVE ones.
2013 	 */
2014 	if (need_more_worker(pool))
2015 		wake_up_worker(pool);
2016 
2017 	/*
2018 	 * Record the last pool and clear PENDING which should be the last
2019 	 * update to @work.  Also, do this inside @pool->lock so that
2020 	 * PENDING and queued state changes happen together while IRQ is
2021 	 * disabled.
2022 	 */
2023 	set_work_pool_and_clear_pending(work, pool->id);
2024 
2025 	spin_unlock_irq(&pool->lock);
2026 
2027 	lock_map_acquire_read(&pwq->wq->lockdep_map);
2028 	lock_map_acquire(&lockdep_map);
2029 	trace_workqueue_execute_start(work);
2030 	worker->current_func(work);
2031 	/*
2032 	 * While we must be careful to not use "work" after this, the trace
2033 	 * point will only record its address.
2034 	 */
2035 	trace_workqueue_execute_end(work);
2036 	lock_map_release(&lockdep_map);
2037 	lock_map_release(&pwq->wq->lockdep_map);
2038 
2039 	if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2040 		pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2041 		       "     last function: %pf\n",
2042 		       current->comm, preempt_count(), task_pid_nr(current),
2043 		       worker->current_func);
2044 		debug_show_held_locks(current);
2045 		dump_stack();
2046 	}
2047 
2048 	/*
2049 	 * The following prevents a kworker from hogging CPU on !PREEMPT
2050 	 * kernels, where a requeueing work item waiting for something to
2051 	 * happen could deadlock with stop_machine as such work item could
2052 	 * indefinitely requeue itself while all other CPUs are trapped in
2053 	 * stop_machine. At the same time, report a quiescent RCU state so
2054 	 * the same condition doesn't freeze RCU.
2055 	 */
2056 	cond_resched_rcu_qs();
2057 
2058 	spin_lock_irq(&pool->lock);
2059 
2060 	/* clear cpu intensive status */
2061 	if (unlikely(cpu_intensive))
2062 		worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2063 
2064 	/* we're done with it, release */
2065 	hash_del(&worker->hentry);
2066 	worker->current_work = NULL;
2067 	worker->current_func = NULL;
2068 	worker->current_pwq = NULL;
2069 	worker->desc_valid = false;
2070 	pwq_dec_nr_in_flight(pwq, work_color);
2071 }
2072 
2073 /**
2074  * process_scheduled_works - process scheduled works
2075  * @worker: self
2076  *
2077  * Process all scheduled works.  Please note that the scheduled list
2078  * may change while processing a work, so this function repeatedly
2079  * fetches a work from the top and executes it.
2080  *
2081  * CONTEXT:
2082  * spin_lock_irq(pool->lock) which may be released and regrabbed
2083  * multiple times.
2084  */
2085 static void process_scheduled_works(struct worker *worker)
2086 {
2087 	while (!list_empty(&worker->scheduled)) {
2088 		struct work_struct *work = list_first_entry(&worker->scheduled,
2089 						struct work_struct, entry);
2090 		process_one_work(worker, work);
2091 	}
2092 }
2093 
2094 /**
2095  * worker_thread - the worker thread function
2096  * @__worker: self
2097  *
2098  * The worker thread function.  All workers belong to a worker_pool -
2099  * either a per-cpu one or dynamic unbound one.  These workers process all
2100  * work items regardless of their specific target workqueue.  The only
2101  * exception is work items which belong to workqueues with a rescuer which
2102  * will be explained in rescuer_thread().
2103  *
2104  * Return: 0
2105  */
2106 static int worker_thread(void *__worker)
2107 {
2108 	struct worker *worker = __worker;
2109 	struct worker_pool *pool = worker->pool;
2110 
2111 	/* tell the scheduler that this is a workqueue worker */
2112 	worker->task->flags |= PF_WQ_WORKER;
2113 woke_up:
2114 	spin_lock_irq(&pool->lock);
2115 
2116 	/* am I supposed to die? */
2117 	if (unlikely(worker->flags & WORKER_DIE)) {
2118 		spin_unlock_irq(&pool->lock);
2119 		WARN_ON_ONCE(!list_empty(&worker->entry));
2120 		worker->task->flags &= ~PF_WQ_WORKER;
2121 
2122 		set_task_comm(worker->task, "kworker/dying");
2123 		ida_simple_remove(&pool->worker_ida, worker->id);
2124 		worker_detach_from_pool(worker, pool);
2125 		kfree(worker);
2126 		return 0;
2127 	}
2128 
2129 	worker_leave_idle(worker);
2130 recheck:
2131 	/* no more worker necessary? */
2132 	if (!need_more_worker(pool))
2133 		goto sleep;
2134 
2135 	/* do we need to manage? */
2136 	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2137 		goto recheck;
2138 
2139 	/*
2140 	 * ->scheduled list can only be filled while a worker is
2141 	 * preparing to process a work or actually processing it.
2142 	 * Make sure nobody diddled with it while I was sleeping.
2143 	 */
2144 	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2145 
2146 	/*
2147 	 * Finish PREP stage.  We're guaranteed to have at least one idle
2148 	 * worker or that someone else has already assumed the manager
2149 	 * role.  This is where @worker starts participating in concurrency
2150 	 * management if applicable and concurrency management is restored
2151 	 * after being rebound.  See rebind_workers() for details.
2152 	 */
2153 	worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2154 
2155 	do {
2156 		struct work_struct *work =
2157 			list_first_entry(&pool->worklist,
2158 					 struct work_struct, entry);
2159 
2160 		if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2161 			/* optimization path, not strictly necessary */
2162 			process_one_work(worker, work);
2163 			if (unlikely(!list_empty(&worker->scheduled)))
2164 				process_scheduled_works(worker);
2165 		} else {
2166 			move_linked_works(work, &worker->scheduled, NULL);
2167 			process_scheduled_works(worker);
2168 		}
2169 	} while (keep_working(pool));
2170 
2171 	worker_set_flags(worker, WORKER_PREP);
2172 sleep:
2173 	/*
2174 	 * pool->lock is held and there's no work to process and no need to
2175 	 * manage, sleep.  Workers are woken up only while holding
2176 	 * pool->lock or from local cpu, so setting the current state
2177 	 * before releasing pool->lock is enough to prevent losing any
2178 	 * event.
2179 	 */
2180 	worker_enter_idle(worker);
2181 	__set_current_state(TASK_INTERRUPTIBLE);
2182 	spin_unlock_irq(&pool->lock);
2183 	schedule();
2184 	goto woke_up;
2185 }
2186 
2187 /**
2188  * rescuer_thread - the rescuer thread function
2189  * @__rescuer: self
2190  *
2191  * Workqueue rescuer thread function.  There's one rescuer for each
2192  * workqueue which has WQ_MEM_RECLAIM set.
2193  *
2194  * Regular work processing on a pool may block trying to create a new
2195  * worker which uses GFP_KERNEL allocation which has slight chance of
2196  * developing into deadlock if some works currently on the same queue
2197  * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2198  * the problem rescuer solves.
2199  *
2200  * When such condition is possible, the pool summons rescuers of all
2201  * workqueues which have works queued on the pool and let them process
2202  * those works so that forward progress can be guaranteed.
2203  *
2204  * This should happen rarely.
2205  *
2206  * Return: 0
2207  */
2208 static int rescuer_thread(void *__rescuer)
2209 {
2210 	struct worker *rescuer = __rescuer;
2211 	struct workqueue_struct *wq = rescuer->rescue_wq;
2212 	struct list_head *scheduled = &rescuer->scheduled;
2213 	bool should_stop;
2214 
2215 	set_user_nice(current, RESCUER_NICE_LEVEL);
2216 
2217 	/*
2218 	 * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
2219 	 * doesn't participate in concurrency management.
2220 	 */
2221 	rescuer->task->flags |= PF_WQ_WORKER;
2222 repeat:
2223 	set_current_state(TASK_INTERRUPTIBLE);
2224 
2225 	/*
2226 	 * By the time the rescuer is requested to stop, the workqueue
2227 	 * shouldn't have any work pending, but @wq->maydays may still have
2228 	 * pwq(s) queued.  This can happen by non-rescuer workers consuming
2229 	 * all the work items before the rescuer got to them.  Go through
2230 	 * @wq->maydays processing before acting on should_stop so that the
2231 	 * list is always empty on exit.
2232 	 */
2233 	should_stop = kthread_should_stop();
2234 
2235 	/* see whether any pwq is asking for help */
2236 	spin_lock_irq(&wq_mayday_lock);
2237 
2238 	while (!list_empty(&wq->maydays)) {
2239 		struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2240 					struct pool_workqueue, mayday_node);
2241 		struct worker_pool *pool = pwq->pool;
2242 		struct work_struct *work, *n;
2243 
2244 		__set_current_state(TASK_RUNNING);
2245 		list_del_init(&pwq->mayday_node);
2246 
2247 		spin_unlock_irq(&wq_mayday_lock);
2248 
2249 		worker_attach_to_pool(rescuer, pool);
2250 
2251 		spin_lock_irq(&pool->lock);
2252 		rescuer->pool = pool;
2253 
2254 		/*
2255 		 * Slurp in all works issued via this workqueue and
2256 		 * process'em.
2257 		 */
2258 		WARN_ON_ONCE(!list_empty(scheduled));
2259 		list_for_each_entry_safe(work, n, &pool->worklist, entry)
2260 			if (get_work_pwq(work) == pwq)
2261 				move_linked_works(work, scheduled, &n);
2262 
2263 		if (!list_empty(scheduled)) {
2264 			process_scheduled_works(rescuer);
2265 
2266 			/*
2267 			 * The above execution of rescued work items could
2268 			 * have created more to rescue through
2269 			 * pwq_activate_first_delayed() or chained
2270 			 * queueing.  Let's put @pwq back on mayday list so
2271 			 * that such back-to-back work items, which may be
2272 			 * being used to relieve memory pressure, don't
2273 			 * incur MAYDAY_INTERVAL delay inbetween.
2274 			 */
2275 			if (need_to_create_worker(pool)) {
2276 				spin_lock(&wq_mayday_lock);
2277 				get_pwq(pwq);
2278 				list_move_tail(&pwq->mayday_node, &wq->maydays);
2279 				spin_unlock(&wq_mayday_lock);
2280 			}
2281 		}
2282 
2283 		/*
2284 		 * Put the reference grabbed by send_mayday().  @pool won't
2285 		 * go away while we're still attached to it.
2286 		 */
2287 		put_pwq(pwq);
2288 
2289 		/*
2290 		 * Leave this pool.  If need_more_worker() is %true, notify a
2291 		 * regular worker; otherwise, we end up with 0 concurrency
2292 		 * and stalling the execution.
2293 		 */
2294 		if (need_more_worker(pool))
2295 			wake_up_worker(pool);
2296 
2297 		rescuer->pool = NULL;
2298 		spin_unlock_irq(&pool->lock);
2299 
2300 		worker_detach_from_pool(rescuer, pool);
2301 
2302 		spin_lock_irq(&wq_mayday_lock);
2303 	}
2304 
2305 	spin_unlock_irq(&wq_mayday_lock);
2306 
2307 	if (should_stop) {
2308 		__set_current_state(TASK_RUNNING);
2309 		rescuer->task->flags &= ~PF_WQ_WORKER;
2310 		return 0;
2311 	}
2312 
2313 	/* rescuers should never participate in concurrency management */
2314 	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2315 	schedule();
2316 	goto repeat;
2317 }
2318 
2319 struct wq_barrier {
2320 	struct work_struct	work;
2321 	struct completion	done;
2322 	struct task_struct	*task;	/* purely informational */
2323 };
2324 
2325 static void wq_barrier_func(struct work_struct *work)
2326 {
2327 	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2328 	complete(&barr->done);
2329 }
2330 
2331 /**
2332  * insert_wq_barrier - insert a barrier work
2333  * @pwq: pwq to insert barrier into
2334  * @barr: wq_barrier to insert
2335  * @target: target work to attach @barr to
2336  * @worker: worker currently executing @target, NULL if @target is not executing
2337  *
2338  * @barr is linked to @target such that @barr is completed only after
2339  * @target finishes execution.  Please note that the ordering
2340  * guarantee is observed only with respect to @target and on the local
2341  * cpu.
2342  *
2343  * Currently, a queued barrier can't be canceled.  This is because
2344  * try_to_grab_pending() can't determine whether the work to be
2345  * grabbed is at the head of the queue and thus can't clear LINKED
2346  * flag of the previous work while there must be a valid next work
2347  * after a work with LINKED flag set.
2348  *
2349  * Note that when @worker is non-NULL, @target may be modified
2350  * underneath us, so we can't reliably determine pwq from @target.
2351  *
2352  * CONTEXT:
2353  * spin_lock_irq(pool->lock).
2354  */
2355 static void insert_wq_barrier(struct pool_workqueue *pwq,
2356 			      struct wq_barrier *barr,
2357 			      struct work_struct *target, struct worker *worker)
2358 {
2359 	struct list_head *head;
2360 	unsigned int linked = 0;
2361 
2362 	/*
2363 	 * debugobject calls are safe here even with pool->lock locked
2364 	 * as we know for sure that this will not trigger any of the
2365 	 * checks and call back into the fixup functions where we
2366 	 * might deadlock.
2367 	 */
2368 	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2369 	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2370 	init_completion(&barr->done);
2371 	barr->task = current;
2372 
2373 	/*
2374 	 * If @target is currently being executed, schedule the
2375 	 * barrier to the worker; otherwise, put it after @target.
2376 	 */
2377 	if (worker)
2378 		head = worker->scheduled.next;
2379 	else {
2380 		unsigned long *bits = work_data_bits(target);
2381 
2382 		head = target->entry.next;
2383 		/* there can already be other linked works, inherit and set */
2384 		linked = *bits & WORK_STRUCT_LINKED;
2385 		__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2386 	}
2387 
2388 	debug_work_activate(&barr->work);
2389 	insert_work(pwq, &barr->work, head,
2390 		    work_color_to_flags(WORK_NO_COLOR) | linked);
2391 }
2392 
2393 /**
2394  * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2395  * @wq: workqueue being flushed
2396  * @flush_color: new flush color, < 0 for no-op
2397  * @work_color: new work color, < 0 for no-op
2398  *
2399  * Prepare pwqs for workqueue flushing.
2400  *
2401  * If @flush_color is non-negative, flush_color on all pwqs should be
2402  * -1.  If no pwq has in-flight commands at the specified color, all
2403  * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
2404  * has in flight commands, its pwq->flush_color is set to
2405  * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2406  * wakeup logic is armed and %true is returned.
2407  *
2408  * The caller should have initialized @wq->first_flusher prior to
2409  * calling this function with non-negative @flush_color.  If
2410  * @flush_color is negative, no flush color update is done and %false
2411  * is returned.
2412  *
2413  * If @work_color is non-negative, all pwqs should have the same
2414  * work_color which is previous to @work_color and all will be
2415  * advanced to @work_color.
2416  *
2417  * CONTEXT:
2418  * mutex_lock(wq->mutex).
2419  *
2420  * Return:
2421  * %true if @flush_color >= 0 and there's something to flush.  %false
2422  * otherwise.
2423  */
2424 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2425 				      int flush_color, int work_color)
2426 {
2427 	bool wait = false;
2428 	struct pool_workqueue *pwq;
2429 
2430 	if (flush_color >= 0) {
2431 		WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2432 		atomic_set(&wq->nr_pwqs_to_flush, 1);
2433 	}
2434 
2435 	for_each_pwq(pwq, wq) {
2436 		struct worker_pool *pool = pwq->pool;
2437 
2438 		spin_lock_irq(&pool->lock);
2439 
2440 		if (flush_color >= 0) {
2441 			WARN_ON_ONCE(pwq->flush_color != -1);
2442 
2443 			if (pwq->nr_in_flight[flush_color]) {
2444 				pwq->flush_color = flush_color;
2445 				atomic_inc(&wq->nr_pwqs_to_flush);
2446 				wait = true;
2447 			}
2448 		}
2449 
2450 		if (work_color >= 0) {
2451 			WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2452 			pwq->work_color = work_color;
2453 		}
2454 
2455 		spin_unlock_irq(&pool->lock);
2456 	}
2457 
2458 	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2459 		complete(&wq->first_flusher->done);
2460 
2461 	return wait;
2462 }
2463 
2464 /**
2465  * flush_workqueue - ensure that any scheduled work has run to completion.
2466  * @wq: workqueue to flush
2467  *
2468  * This function sleeps until all work items which were queued on entry
2469  * have finished execution, but it is not livelocked by new incoming ones.
2470  */
2471 void flush_workqueue(struct workqueue_struct *wq)
2472 {
2473 	struct wq_flusher this_flusher = {
2474 		.list = LIST_HEAD_INIT(this_flusher.list),
2475 		.flush_color = -1,
2476 		.done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2477 	};
2478 	int next_color;
2479 
2480 	lock_map_acquire(&wq->lockdep_map);
2481 	lock_map_release(&wq->lockdep_map);
2482 
2483 	mutex_lock(&wq->mutex);
2484 
2485 	/*
2486 	 * Start-to-wait phase
2487 	 */
2488 	next_color = work_next_color(wq->work_color);
2489 
2490 	if (next_color != wq->flush_color) {
2491 		/*
2492 		 * Color space is not full.  The current work_color
2493 		 * becomes our flush_color and work_color is advanced
2494 		 * by one.
2495 		 */
2496 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2497 		this_flusher.flush_color = wq->work_color;
2498 		wq->work_color = next_color;
2499 
2500 		if (!wq->first_flusher) {
2501 			/* no flush in progress, become the first flusher */
2502 			WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2503 
2504 			wq->first_flusher = &this_flusher;
2505 
2506 			if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2507 						       wq->work_color)) {
2508 				/* nothing to flush, done */
2509 				wq->flush_color = next_color;
2510 				wq->first_flusher = NULL;
2511 				goto out_unlock;
2512 			}
2513 		} else {
2514 			/* wait in queue */
2515 			WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2516 			list_add_tail(&this_flusher.list, &wq->flusher_queue);
2517 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2518 		}
2519 	} else {
2520 		/*
2521 		 * Oops, color space is full, wait on overflow queue.
2522 		 * The next flush completion will assign us
2523 		 * flush_color and transfer to flusher_queue.
2524 		 */
2525 		list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2526 	}
2527 
2528 	mutex_unlock(&wq->mutex);
2529 
2530 	wait_for_completion(&this_flusher.done);
2531 
2532 	/*
2533 	 * Wake-up-and-cascade phase
2534 	 *
2535 	 * First flushers are responsible for cascading flushes and
2536 	 * handling overflow.  Non-first flushers can simply return.
2537 	 */
2538 	if (wq->first_flusher != &this_flusher)
2539 		return;
2540 
2541 	mutex_lock(&wq->mutex);
2542 
2543 	/* we might have raced, check again with mutex held */
2544 	if (wq->first_flusher != &this_flusher)
2545 		goto out_unlock;
2546 
2547 	wq->first_flusher = NULL;
2548 
2549 	WARN_ON_ONCE(!list_empty(&this_flusher.list));
2550 	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2551 
2552 	while (true) {
2553 		struct wq_flusher *next, *tmp;
2554 
2555 		/* complete all the flushers sharing the current flush color */
2556 		list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2557 			if (next->flush_color != wq->flush_color)
2558 				break;
2559 			list_del_init(&next->list);
2560 			complete(&next->done);
2561 		}
2562 
2563 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2564 			     wq->flush_color != work_next_color(wq->work_color));
2565 
2566 		/* this flush_color is finished, advance by one */
2567 		wq->flush_color = work_next_color(wq->flush_color);
2568 
2569 		/* one color has been freed, handle overflow queue */
2570 		if (!list_empty(&wq->flusher_overflow)) {
2571 			/*
2572 			 * Assign the same color to all overflowed
2573 			 * flushers, advance work_color and append to
2574 			 * flusher_queue.  This is the start-to-wait
2575 			 * phase for these overflowed flushers.
2576 			 */
2577 			list_for_each_entry(tmp, &wq->flusher_overflow, list)
2578 				tmp->flush_color = wq->work_color;
2579 
2580 			wq->work_color = work_next_color(wq->work_color);
2581 
2582 			list_splice_tail_init(&wq->flusher_overflow,
2583 					      &wq->flusher_queue);
2584 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2585 		}
2586 
2587 		if (list_empty(&wq->flusher_queue)) {
2588 			WARN_ON_ONCE(wq->flush_color != wq->work_color);
2589 			break;
2590 		}
2591 
2592 		/*
2593 		 * Need to flush more colors.  Make the next flusher
2594 		 * the new first flusher and arm pwqs.
2595 		 */
2596 		WARN_ON_ONCE(wq->flush_color == wq->work_color);
2597 		WARN_ON_ONCE(wq->flush_color != next->flush_color);
2598 
2599 		list_del_init(&next->list);
2600 		wq->first_flusher = next;
2601 
2602 		if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2603 			break;
2604 
2605 		/*
2606 		 * Meh... this color is already done, clear first
2607 		 * flusher and repeat cascading.
2608 		 */
2609 		wq->first_flusher = NULL;
2610 	}
2611 
2612 out_unlock:
2613 	mutex_unlock(&wq->mutex);
2614 }
2615 EXPORT_SYMBOL(flush_workqueue);
2616 
2617 /**
2618  * drain_workqueue - drain a workqueue
2619  * @wq: workqueue to drain
2620  *
2621  * Wait until the workqueue becomes empty.  While draining is in progress,
2622  * only chain queueing is allowed.  IOW, only currently pending or running
2623  * work items on @wq can queue further work items on it.  @wq is flushed
2624  * repeatedly until it becomes empty.  The number of flushing is determined
2625  * by the depth of chaining and should be relatively short.  Whine if it
2626  * takes too long.
2627  */
2628 void drain_workqueue(struct workqueue_struct *wq)
2629 {
2630 	unsigned int flush_cnt = 0;
2631 	struct pool_workqueue *pwq;
2632 
2633 	/*
2634 	 * __queue_work() needs to test whether there are drainers, is much
2635 	 * hotter than drain_workqueue() and already looks at @wq->flags.
2636 	 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2637 	 */
2638 	mutex_lock(&wq->mutex);
2639 	if (!wq->nr_drainers++)
2640 		wq->flags |= __WQ_DRAINING;
2641 	mutex_unlock(&wq->mutex);
2642 reflush:
2643 	flush_workqueue(wq);
2644 
2645 	mutex_lock(&wq->mutex);
2646 
2647 	for_each_pwq(pwq, wq) {
2648 		bool drained;
2649 
2650 		spin_lock_irq(&pwq->pool->lock);
2651 		drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2652 		spin_unlock_irq(&pwq->pool->lock);
2653 
2654 		if (drained)
2655 			continue;
2656 
2657 		if (++flush_cnt == 10 ||
2658 		    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2659 			pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2660 				wq->name, flush_cnt);
2661 
2662 		mutex_unlock(&wq->mutex);
2663 		goto reflush;
2664 	}
2665 
2666 	if (!--wq->nr_drainers)
2667 		wq->flags &= ~__WQ_DRAINING;
2668 	mutex_unlock(&wq->mutex);
2669 }
2670 EXPORT_SYMBOL_GPL(drain_workqueue);
2671 
2672 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2673 {
2674 	struct worker *worker = NULL;
2675 	struct worker_pool *pool;
2676 	struct pool_workqueue *pwq;
2677 
2678 	might_sleep();
2679 
2680 	local_irq_disable();
2681 	pool = get_work_pool(work);
2682 	if (!pool) {
2683 		local_irq_enable();
2684 		return false;
2685 	}
2686 
2687 	spin_lock(&pool->lock);
2688 	/* see the comment in try_to_grab_pending() with the same code */
2689 	pwq = get_work_pwq(work);
2690 	if (pwq) {
2691 		if (unlikely(pwq->pool != pool))
2692 			goto already_gone;
2693 	} else {
2694 		worker = find_worker_executing_work(pool, work);
2695 		if (!worker)
2696 			goto already_gone;
2697 		pwq = worker->current_pwq;
2698 	}
2699 
2700 	insert_wq_barrier(pwq, barr, work, worker);
2701 	spin_unlock_irq(&pool->lock);
2702 
2703 	/*
2704 	 * If @max_active is 1 or rescuer is in use, flushing another work
2705 	 * item on the same workqueue may lead to deadlock.  Make sure the
2706 	 * flusher is not running on the same workqueue by verifying write
2707 	 * access.
2708 	 */
2709 	if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2710 		lock_map_acquire(&pwq->wq->lockdep_map);
2711 	else
2712 		lock_map_acquire_read(&pwq->wq->lockdep_map);
2713 	lock_map_release(&pwq->wq->lockdep_map);
2714 
2715 	return true;
2716 already_gone:
2717 	spin_unlock_irq(&pool->lock);
2718 	return false;
2719 }
2720 
2721 /**
2722  * flush_work - wait for a work to finish executing the last queueing instance
2723  * @work: the work to flush
2724  *
2725  * Wait until @work has finished execution.  @work is guaranteed to be idle
2726  * on return if it hasn't been requeued since flush started.
2727  *
2728  * Return:
2729  * %true if flush_work() waited for the work to finish execution,
2730  * %false if it was already idle.
2731  */
2732 bool flush_work(struct work_struct *work)
2733 {
2734 	struct wq_barrier barr;
2735 
2736 	lock_map_acquire(&work->lockdep_map);
2737 	lock_map_release(&work->lockdep_map);
2738 
2739 	if (start_flush_work(work, &barr)) {
2740 		wait_for_completion(&barr.done);
2741 		destroy_work_on_stack(&barr.work);
2742 		return true;
2743 	} else {
2744 		return false;
2745 	}
2746 }
2747 EXPORT_SYMBOL_GPL(flush_work);
2748 
2749 struct cwt_wait {
2750 	wait_queue_t		wait;
2751 	struct work_struct	*work;
2752 };
2753 
2754 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2755 {
2756 	struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2757 
2758 	if (cwait->work != key)
2759 		return 0;
2760 	return autoremove_wake_function(wait, mode, sync, key);
2761 }
2762 
2763 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2764 {
2765 	static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2766 	unsigned long flags;
2767 	int ret;
2768 
2769 	do {
2770 		ret = try_to_grab_pending(work, is_dwork, &flags);
2771 		/*
2772 		 * If someone else is already canceling, wait for it to
2773 		 * finish.  flush_work() doesn't work for PREEMPT_NONE
2774 		 * because we may get scheduled between @work's completion
2775 		 * and the other canceling task resuming and clearing
2776 		 * CANCELING - flush_work() will return false immediately
2777 		 * as @work is no longer busy, try_to_grab_pending() will
2778 		 * return -ENOENT as @work is still being canceled and the
2779 		 * other canceling task won't be able to clear CANCELING as
2780 		 * we're hogging the CPU.
2781 		 *
2782 		 * Let's wait for completion using a waitqueue.  As this
2783 		 * may lead to the thundering herd problem, use a custom
2784 		 * wake function which matches @work along with exclusive
2785 		 * wait and wakeup.
2786 		 */
2787 		if (unlikely(ret == -ENOENT)) {
2788 			struct cwt_wait cwait;
2789 
2790 			init_wait(&cwait.wait);
2791 			cwait.wait.func = cwt_wakefn;
2792 			cwait.work = work;
2793 
2794 			prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2795 						  TASK_UNINTERRUPTIBLE);
2796 			if (work_is_canceling(work))
2797 				schedule();
2798 			finish_wait(&cancel_waitq, &cwait.wait);
2799 		}
2800 	} while (unlikely(ret < 0));
2801 
2802 	/* tell other tasks trying to grab @work to back off */
2803 	mark_work_canceling(work);
2804 	local_irq_restore(flags);
2805 
2806 	flush_work(work);
2807 	clear_work_data(work);
2808 
2809 	/*
2810 	 * Paired with prepare_to_wait() above so that either
2811 	 * waitqueue_active() is visible here or !work_is_canceling() is
2812 	 * visible there.
2813 	 */
2814 	smp_mb();
2815 	if (waitqueue_active(&cancel_waitq))
2816 		__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2817 
2818 	return ret;
2819 }
2820 
2821 /**
2822  * cancel_work_sync - cancel a work and wait for it to finish
2823  * @work: the work to cancel
2824  *
2825  * Cancel @work and wait for its execution to finish.  This function
2826  * can be used even if the work re-queues itself or migrates to
2827  * another workqueue.  On return from this function, @work is
2828  * guaranteed to be not pending or executing on any CPU.
2829  *
2830  * cancel_work_sync(&delayed_work->work) must not be used for
2831  * delayed_work's.  Use cancel_delayed_work_sync() instead.
2832  *
2833  * The caller must ensure that the workqueue on which @work was last
2834  * queued can't be destroyed before this function returns.
2835  *
2836  * Return:
2837  * %true if @work was pending, %false otherwise.
2838  */
2839 bool cancel_work_sync(struct work_struct *work)
2840 {
2841 	return __cancel_work_timer(work, false);
2842 }
2843 EXPORT_SYMBOL_GPL(cancel_work_sync);
2844 
2845 /**
2846  * flush_delayed_work - wait for a dwork to finish executing the last queueing
2847  * @dwork: the delayed work to flush
2848  *
2849  * Delayed timer is cancelled and the pending work is queued for
2850  * immediate execution.  Like flush_work(), this function only
2851  * considers the last queueing instance of @dwork.
2852  *
2853  * Return:
2854  * %true if flush_work() waited for the work to finish execution,
2855  * %false if it was already idle.
2856  */
2857 bool flush_delayed_work(struct delayed_work *dwork)
2858 {
2859 	local_irq_disable();
2860 	if (del_timer_sync(&dwork->timer))
2861 		__queue_work(dwork->cpu, dwork->wq, &dwork->work);
2862 	local_irq_enable();
2863 	return flush_work(&dwork->work);
2864 }
2865 EXPORT_SYMBOL(flush_delayed_work);
2866 
2867 /**
2868  * cancel_delayed_work - cancel a delayed work
2869  * @dwork: delayed_work to cancel
2870  *
2871  * Kill off a pending delayed_work.
2872  *
2873  * Return: %true if @dwork was pending and canceled; %false if it wasn't
2874  * pending.
2875  *
2876  * Note:
2877  * The work callback function may still be running on return, unless
2878  * it returns %true and the work doesn't re-arm itself.  Explicitly flush or
2879  * use cancel_delayed_work_sync() to wait on it.
2880  *
2881  * This function is safe to call from any context including IRQ handler.
2882  */
2883 bool cancel_delayed_work(struct delayed_work *dwork)
2884 {
2885 	unsigned long flags;
2886 	int ret;
2887 
2888 	do {
2889 		ret = try_to_grab_pending(&dwork->work, true, &flags);
2890 	} while (unlikely(ret == -EAGAIN));
2891 
2892 	if (unlikely(ret < 0))
2893 		return false;
2894 
2895 	set_work_pool_and_clear_pending(&dwork->work,
2896 					get_work_pool_id(&dwork->work));
2897 	local_irq_restore(flags);
2898 	return ret;
2899 }
2900 EXPORT_SYMBOL(cancel_delayed_work);
2901 
2902 /**
2903  * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2904  * @dwork: the delayed work cancel
2905  *
2906  * This is cancel_work_sync() for delayed works.
2907  *
2908  * Return:
2909  * %true if @dwork was pending, %false otherwise.
2910  */
2911 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2912 {
2913 	return __cancel_work_timer(&dwork->work, true);
2914 }
2915 EXPORT_SYMBOL(cancel_delayed_work_sync);
2916 
2917 /**
2918  * schedule_on_each_cpu - execute a function synchronously on each online CPU
2919  * @func: the function to call
2920  *
2921  * schedule_on_each_cpu() executes @func on each online CPU using the
2922  * system workqueue and blocks until all CPUs have completed.
2923  * schedule_on_each_cpu() is very slow.
2924  *
2925  * Return:
2926  * 0 on success, -errno on failure.
2927  */
2928 int schedule_on_each_cpu(work_func_t func)
2929 {
2930 	int cpu;
2931 	struct work_struct __percpu *works;
2932 
2933 	works = alloc_percpu(struct work_struct);
2934 	if (!works)
2935 		return -ENOMEM;
2936 
2937 	get_online_cpus();
2938 
2939 	for_each_online_cpu(cpu) {
2940 		struct work_struct *work = per_cpu_ptr(works, cpu);
2941 
2942 		INIT_WORK(work, func);
2943 		schedule_work_on(cpu, work);
2944 	}
2945 
2946 	for_each_online_cpu(cpu)
2947 		flush_work(per_cpu_ptr(works, cpu));
2948 
2949 	put_online_cpus();
2950 	free_percpu(works);
2951 	return 0;
2952 }
2953 
2954 /**
2955  * execute_in_process_context - reliably execute the routine with user context
2956  * @fn:		the function to execute
2957  * @ew:		guaranteed storage for the execute work structure (must
2958  *		be available when the work executes)
2959  *
2960  * Executes the function immediately if process context is available,
2961  * otherwise schedules the function for delayed execution.
2962  *
2963  * Return:	0 - function was executed
2964  *		1 - function was scheduled for execution
2965  */
2966 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2967 {
2968 	if (!in_interrupt()) {
2969 		fn(&ew->work);
2970 		return 0;
2971 	}
2972 
2973 	INIT_WORK(&ew->work, fn);
2974 	schedule_work(&ew->work);
2975 
2976 	return 1;
2977 }
2978 EXPORT_SYMBOL_GPL(execute_in_process_context);
2979 
2980 /**
2981  * free_workqueue_attrs - free a workqueue_attrs
2982  * @attrs: workqueue_attrs to free
2983  *
2984  * Undo alloc_workqueue_attrs().
2985  */
2986 void free_workqueue_attrs(struct workqueue_attrs *attrs)
2987 {
2988 	if (attrs) {
2989 		free_cpumask_var(attrs->cpumask);
2990 		kfree(attrs);
2991 	}
2992 }
2993 
2994 /**
2995  * alloc_workqueue_attrs - allocate a workqueue_attrs
2996  * @gfp_mask: allocation mask to use
2997  *
2998  * Allocate a new workqueue_attrs, initialize with default settings and
2999  * return it.
3000  *
3001  * Return: The allocated new workqueue_attr on success. %NULL on failure.
3002  */
3003 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3004 {
3005 	struct workqueue_attrs *attrs;
3006 
3007 	attrs = kzalloc(sizeof(*attrs), gfp_mask);
3008 	if (!attrs)
3009 		goto fail;
3010 	if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3011 		goto fail;
3012 
3013 	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3014 	return attrs;
3015 fail:
3016 	free_workqueue_attrs(attrs);
3017 	return NULL;
3018 }
3019 
3020 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3021 				 const struct workqueue_attrs *from)
3022 {
3023 	to->nice = from->nice;
3024 	cpumask_copy(to->cpumask, from->cpumask);
3025 	/*
3026 	 * Unlike hash and equality test, this function doesn't ignore
3027 	 * ->no_numa as it is used for both pool and wq attrs.  Instead,
3028 	 * get_unbound_pool() explicitly clears ->no_numa after copying.
3029 	 */
3030 	to->no_numa = from->no_numa;
3031 }
3032 
3033 /* hash value of the content of @attr */
3034 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3035 {
3036 	u32 hash = 0;
3037 
3038 	hash = jhash_1word(attrs->nice, hash);
3039 	hash = jhash(cpumask_bits(attrs->cpumask),
3040 		     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3041 	return hash;
3042 }
3043 
3044 /* content equality test */
3045 static bool wqattrs_equal(const struct workqueue_attrs *a,
3046 			  const struct workqueue_attrs *b)
3047 {
3048 	if (a->nice != b->nice)
3049 		return false;
3050 	if (!cpumask_equal(a->cpumask, b->cpumask))
3051 		return false;
3052 	return true;
3053 }
3054 
3055 /**
3056  * init_worker_pool - initialize a newly zalloc'd worker_pool
3057  * @pool: worker_pool to initialize
3058  *
3059  * Initialize a newly zalloc'd @pool.  It also allocates @pool->attrs.
3060  *
3061  * Return: 0 on success, -errno on failure.  Even on failure, all fields
3062  * inside @pool proper are initialized and put_unbound_pool() can be called
3063  * on @pool safely to release it.
3064  */
3065 static int init_worker_pool(struct worker_pool *pool)
3066 {
3067 	spin_lock_init(&pool->lock);
3068 	pool->id = -1;
3069 	pool->cpu = -1;
3070 	pool->node = NUMA_NO_NODE;
3071 	pool->flags |= POOL_DISASSOCIATED;
3072 	INIT_LIST_HEAD(&pool->worklist);
3073 	INIT_LIST_HEAD(&pool->idle_list);
3074 	hash_init(pool->busy_hash);
3075 
3076 	init_timer_deferrable(&pool->idle_timer);
3077 	pool->idle_timer.function = idle_worker_timeout;
3078 	pool->idle_timer.data = (unsigned long)pool;
3079 
3080 	setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3081 		    (unsigned long)pool);
3082 
3083 	mutex_init(&pool->manager_arb);
3084 	mutex_init(&pool->attach_mutex);
3085 	INIT_LIST_HEAD(&pool->workers);
3086 
3087 	ida_init(&pool->worker_ida);
3088 	INIT_HLIST_NODE(&pool->hash_node);
3089 	pool->refcnt = 1;
3090 
3091 	/* shouldn't fail above this point */
3092 	pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3093 	if (!pool->attrs)
3094 		return -ENOMEM;
3095 	return 0;
3096 }
3097 
3098 static void rcu_free_wq(struct rcu_head *rcu)
3099 {
3100 	struct workqueue_struct *wq =
3101 		container_of(rcu, struct workqueue_struct, rcu);
3102 
3103 	if (!(wq->flags & WQ_UNBOUND))
3104 		free_percpu(wq->cpu_pwqs);
3105 	else
3106 		free_workqueue_attrs(wq->unbound_attrs);
3107 
3108 	kfree(wq->rescuer);
3109 	kfree(wq);
3110 }
3111 
3112 static void rcu_free_pool(struct rcu_head *rcu)
3113 {
3114 	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3115 
3116 	ida_destroy(&pool->worker_ida);
3117 	free_workqueue_attrs(pool->attrs);
3118 	kfree(pool);
3119 }
3120 
3121 /**
3122  * put_unbound_pool - put a worker_pool
3123  * @pool: worker_pool to put
3124  *
3125  * Put @pool.  If its refcnt reaches zero, it gets destroyed in sched-RCU
3126  * safe manner.  get_unbound_pool() calls this function on its failure path
3127  * and this function should be able to release pools which went through,
3128  * successfully or not, init_worker_pool().
3129  *
3130  * Should be called with wq_pool_mutex held.
3131  */
3132 static void put_unbound_pool(struct worker_pool *pool)
3133 {
3134 	DECLARE_COMPLETION_ONSTACK(detach_completion);
3135 	struct worker *worker;
3136 
3137 	lockdep_assert_held(&wq_pool_mutex);
3138 
3139 	if (--pool->refcnt)
3140 		return;
3141 
3142 	/* sanity checks */
3143 	if (WARN_ON(!(pool->cpu < 0)) ||
3144 	    WARN_ON(!list_empty(&pool->worklist)))
3145 		return;
3146 
3147 	/* release id and unhash */
3148 	if (pool->id >= 0)
3149 		idr_remove(&worker_pool_idr, pool->id);
3150 	hash_del(&pool->hash_node);
3151 
3152 	/*
3153 	 * Become the manager and destroy all workers.  Grabbing
3154 	 * manager_arb prevents @pool's workers from blocking on
3155 	 * attach_mutex.
3156 	 */
3157 	mutex_lock(&pool->manager_arb);
3158 
3159 	spin_lock_irq(&pool->lock);
3160 	while ((worker = first_idle_worker(pool)))
3161 		destroy_worker(worker);
3162 	WARN_ON(pool->nr_workers || pool->nr_idle);
3163 	spin_unlock_irq(&pool->lock);
3164 
3165 	mutex_lock(&pool->attach_mutex);
3166 	if (!list_empty(&pool->workers))
3167 		pool->detach_completion = &detach_completion;
3168 	mutex_unlock(&pool->attach_mutex);
3169 
3170 	if (pool->detach_completion)
3171 		wait_for_completion(pool->detach_completion);
3172 
3173 	mutex_unlock(&pool->manager_arb);
3174 
3175 	/* shut down the timers */
3176 	del_timer_sync(&pool->idle_timer);
3177 	del_timer_sync(&pool->mayday_timer);
3178 
3179 	/* sched-RCU protected to allow dereferences from get_work_pool() */
3180 	call_rcu_sched(&pool->rcu, rcu_free_pool);
3181 }
3182 
3183 /**
3184  * get_unbound_pool - get a worker_pool with the specified attributes
3185  * @attrs: the attributes of the worker_pool to get
3186  *
3187  * Obtain a worker_pool which has the same attributes as @attrs, bump the
3188  * reference count and return it.  If there already is a matching
3189  * worker_pool, it will be used; otherwise, this function attempts to
3190  * create a new one.
3191  *
3192  * Should be called with wq_pool_mutex held.
3193  *
3194  * Return: On success, a worker_pool with the same attributes as @attrs.
3195  * On failure, %NULL.
3196  */
3197 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3198 {
3199 	u32 hash = wqattrs_hash(attrs);
3200 	struct worker_pool *pool;
3201 	int node;
3202 
3203 	lockdep_assert_held(&wq_pool_mutex);
3204 
3205 	/* do we already have a matching pool? */
3206 	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3207 		if (wqattrs_equal(pool->attrs, attrs)) {
3208 			pool->refcnt++;
3209 			return pool;
3210 		}
3211 	}
3212 
3213 	/* nope, create a new one */
3214 	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3215 	if (!pool || init_worker_pool(pool) < 0)
3216 		goto fail;
3217 
3218 	lockdep_set_subclass(&pool->lock, 1);	/* see put_pwq() */
3219 	copy_workqueue_attrs(pool->attrs, attrs);
3220 
3221 	/*
3222 	 * no_numa isn't a worker_pool attribute, always clear it.  See
3223 	 * 'struct workqueue_attrs' comments for detail.
3224 	 */
3225 	pool->attrs->no_numa = false;
3226 
3227 	/* if cpumask is contained inside a NUMA node, we belong to that node */
3228 	if (wq_numa_enabled) {
3229 		for_each_node(node) {
3230 			if (cpumask_subset(pool->attrs->cpumask,
3231 					   wq_numa_possible_cpumask[node])) {
3232 				pool->node = node;
3233 				break;
3234 			}
3235 		}
3236 	}
3237 
3238 	if (worker_pool_assign_id(pool) < 0)
3239 		goto fail;
3240 
3241 	/* create and start the initial worker */
3242 	if (!create_worker(pool))
3243 		goto fail;
3244 
3245 	/* install */
3246 	hash_add(unbound_pool_hash, &pool->hash_node, hash);
3247 
3248 	return pool;
3249 fail:
3250 	if (pool)
3251 		put_unbound_pool(pool);
3252 	return NULL;
3253 }
3254 
3255 static void rcu_free_pwq(struct rcu_head *rcu)
3256 {
3257 	kmem_cache_free(pwq_cache,
3258 			container_of(rcu, struct pool_workqueue, rcu));
3259 }
3260 
3261 /*
3262  * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3263  * and needs to be destroyed.
3264  */
3265 static void pwq_unbound_release_workfn(struct work_struct *work)
3266 {
3267 	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3268 						  unbound_release_work);
3269 	struct workqueue_struct *wq = pwq->wq;
3270 	struct worker_pool *pool = pwq->pool;
3271 	bool is_last;
3272 
3273 	if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3274 		return;
3275 
3276 	mutex_lock(&wq->mutex);
3277 	list_del_rcu(&pwq->pwqs_node);
3278 	is_last = list_empty(&wq->pwqs);
3279 	mutex_unlock(&wq->mutex);
3280 
3281 	mutex_lock(&wq_pool_mutex);
3282 	put_unbound_pool(pool);
3283 	mutex_unlock(&wq_pool_mutex);
3284 
3285 	call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3286 
3287 	/*
3288 	 * If we're the last pwq going away, @wq is already dead and no one
3289 	 * is gonna access it anymore.  Schedule RCU free.
3290 	 */
3291 	if (is_last)
3292 		call_rcu_sched(&wq->rcu, rcu_free_wq);
3293 }
3294 
3295 /**
3296  * pwq_adjust_max_active - update a pwq's max_active to the current setting
3297  * @pwq: target pool_workqueue
3298  *
3299  * If @pwq isn't freezing, set @pwq->max_active to the associated
3300  * workqueue's saved_max_active and activate delayed work items
3301  * accordingly.  If @pwq is freezing, clear @pwq->max_active to zero.
3302  */
3303 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3304 {
3305 	struct workqueue_struct *wq = pwq->wq;
3306 	bool freezable = wq->flags & WQ_FREEZABLE;
3307 
3308 	/* for @wq->saved_max_active */
3309 	lockdep_assert_held(&wq->mutex);
3310 
3311 	/* fast exit for non-freezable wqs */
3312 	if (!freezable && pwq->max_active == wq->saved_max_active)
3313 		return;
3314 
3315 	spin_lock_irq(&pwq->pool->lock);
3316 
3317 	/*
3318 	 * During [un]freezing, the caller is responsible for ensuring that
3319 	 * this function is called at least once after @workqueue_freezing
3320 	 * is updated and visible.
3321 	 */
3322 	if (!freezable || !workqueue_freezing) {
3323 		pwq->max_active = wq->saved_max_active;
3324 
3325 		while (!list_empty(&pwq->delayed_works) &&
3326 		       pwq->nr_active < pwq->max_active)
3327 			pwq_activate_first_delayed(pwq);
3328 
3329 		/*
3330 		 * Need to kick a worker after thawed or an unbound wq's
3331 		 * max_active is bumped.  It's a slow path.  Do it always.
3332 		 */
3333 		wake_up_worker(pwq->pool);
3334 	} else {
3335 		pwq->max_active = 0;
3336 	}
3337 
3338 	spin_unlock_irq(&pwq->pool->lock);
3339 }
3340 
3341 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3342 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3343 		     struct worker_pool *pool)
3344 {
3345 	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3346 
3347 	memset(pwq, 0, sizeof(*pwq));
3348 
3349 	pwq->pool = pool;
3350 	pwq->wq = wq;
3351 	pwq->flush_color = -1;
3352 	pwq->refcnt = 1;
3353 	INIT_LIST_HEAD(&pwq->delayed_works);
3354 	INIT_LIST_HEAD(&pwq->pwqs_node);
3355 	INIT_LIST_HEAD(&pwq->mayday_node);
3356 	INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3357 }
3358 
3359 /* sync @pwq with the current state of its associated wq and link it */
3360 static void link_pwq(struct pool_workqueue *pwq)
3361 {
3362 	struct workqueue_struct *wq = pwq->wq;
3363 
3364 	lockdep_assert_held(&wq->mutex);
3365 
3366 	/* may be called multiple times, ignore if already linked */
3367 	if (!list_empty(&pwq->pwqs_node))
3368 		return;
3369 
3370 	/* set the matching work_color */
3371 	pwq->work_color = wq->work_color;
3372 
3373 	/* sync max_active to the current setting */
3374 	pwq_adjust_max_active(pwq);
3375 
3376 	/* link in @pwq */
3377 	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3378 }
3379 
3380 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3381 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3382 					const struct workqueue_attrs *attrs)
3383 {
3384 	struct worker_pool *pool;
3385 	struct pool_workqueue *pwq;
3386 
3387 	lockdep_assert_held(&wq_pool_mutex);
3388 
3389 	pool = get_unbound_pool(attrs);
3390 	if (!pool)
3391 		return NULL;
3392 
3393 	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3394 	if (!pwq) {
3395 		put_unbound_pool(pool);
3396 		return NULL;
3397 	}
3398 
3399 	init_pwq(pwq, wq, pool);
3400 	return pwq;
3401 }
3402 
3403 /**
3404  * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3405  * @attrs: the wq_attrs of the default pwq of the target workqueue
3406  * @node: the target NUMA node
3407  * @cpu_going_down: if >= 0, the CPU to consider as offline
3408  * @cpumask: outarg, the resulting cpumask
3409  *
3410  * Calculate the cpumask a workqueue with @attrs should use on @node.  If
3411  * @cpu_going_down is >= 0, that cpu is considered offline during
3412  * calculation.  The result is stored in @cpumask.
3413  *
3414  * If NUMA affinity is not enabled, @attrs->cpumask is always used.  If
3415  * enabled and @node has online CPUs requested by @attrs, the returned
3416  * cpumask is the intersection of the possible CPUs of @node and
3417  * @attrs->cpumask.
3418  *
3419  * The caller is responsible for ensuring that the cpumask of @node stays
3420  * stable.
3421  *
3422  * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3423  * %false if equal.
3424  */
3425 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3426 				 int cpu_going_down, cpumask_t *cpumask)
3427 {
3428 	if (!wq_numa_enabled || attrs->no_numa)
3429 		goto use_dfl;
3430 
3431 	/* does @node have any online CPUs @attrs wants? */
3432 	cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3433 	if (cpu_going_down >= 0)
3434 		cpumask_clear_cpu(cpu_going_down, cpumask);
3435 
3436 	if (cpumask_empty(cpumask))
3437 		goto use_dfl;
3438 
3439 	/* yeap, return possible CPUs in @node that @attrs wants */
3440 	cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3441 	return !cpumask_equal(cpumask, attrs->cpumask);
3442 
3443 use_dfl:
3444 	cpumask_copy(cpumask, attrs->cpumask);
3445 	return false;
3446 }
3447 
3448 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3449 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3450 						   int node,
3451 						   struct pool_workqueue *pwq)
3452 {
3453 	struct pool_workqueue *old_pwq;
3454 
3455 	lockdep_assert_held(&wq_pool_mutex);
3456 	lockdep_assert_held(&wq->mutex);
3457 
3458 	/* link_pwq() can handle duplicate calls */
3459 	link_pwq(pwq);
3460 
3461 	old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3462 	rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3463 	return old_pwq;
3464 }
3465 
3466 /* context to store the prepared attrs & pwqs before applying */
3467 struct apply_wqattrs_ctx {
3468 	struct workqueue_struct	*wq;		/* target workqueue */
3469 	struct workqueue_attrs	*attrs;		/* attrs to apply */
3470 	struct list_head	list;		/* queued for batching commit */
3471 	struct pool_workqueue	*dfl_pwq;
3472 	struct pool_workqueue	*pwq_tbl[];
3473 };
3474 
3475 /* free the resources after success or abort */
3476 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3477 {
3478 	if (ctx) {
3479 		int node;
3480 
3481 		for_each_node(node)
3482 			put_pwq_unlocked(ctx->pwq_tbl[node]);
3483 		put_pwq_unlocked(ctx->dfl_pwq);
3484 
3485 		free_workqueue_attrs(ctx->attrs);
3486 
3487 		kfree(ctx);
3488 	}
3489 }
3490 
3491 /* allocate the attrs and pwqs for later installation */
3492 static struct apply_wqattrs_ctx *
3493 apply_wqattrs_prepare(struct workqueue_struct *wq,
3494 		      const struct workqueue_attrs *attrs)
3495 {
3496 	struct apply_wqattrs_ctx *ctx;
3497 	struct workqueue_attrs *new_attrs, *tmp_attrs;
3498 	int node;
3499 
3500 	lockdep_assert_held(&wq_pool_mutex);
3501 
3502 	ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3503 		      GFP_KERNEL);
3504 
3505 	new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3506 	tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3507 	if (!ctx || !new_attrs || !tmp_attrs)
3508 		goto out_free;
3509 
3510 	/*
3511 	 * Calculate the attrs of the default pwq.
3512 	 * If the user configured cpumask doesn't overlap with the
3513 	 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3514 	 */
3515 	copy_workqueue_attrs(new_attrs, attrs);
3516 	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3517 	if (unlikely(cpumask_empty(new_attrs->cpumask)))
3518 		cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3519 
3520 	/*
3521 	 * We may create multiple pwqs with differing cpumasks.  Make a
3522 	 * copy of @new_attrs which will be modified and used to obtain
3523 	 * pools.
3524 	 */
3525 	copy_workqueue_attrs(tmp_attrs, new_attrs);
3526 
3527 	/*
3528 	 * If something goes wrong during CPU up/down, we'll fall back to
3529 	 * the default pwq covering whole @attrs->cpumask.  Always create
3530 	 * it even if we don't use it immediately.
3531 	 */
3532 	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3533 	if (!ctx->dfl_pwq)
3534 		goto out_free;
3535 
3536 	for_each_node(node) {
3537 		if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3538 			ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3539 			if (!ctx->pwq_tbl[node])
3540 				goto out_free;
3541 		} else {
3542 			ctx->dfl_pwq->refcnt++;
3543 			ctx->pwq_tbl[node] = ctx->dfl_pwq;
3544 		}
3545 	}
3546 
3547 	/* save the user configured attrs and sanitize it. */
3548 	copy_workqueue_attrs(new_attrs, attrs);
3549 	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3550 	ctx->attrs = new_attrs;
3551 
3552 	ctx->wq = wq;
3553 	free_workqueue_attrs(tmp_attrs);
3554 	return ctx;
3555 
3556 out_free:
3557 	free_workqueue_attrs(tmp_attrs);
3558 	free_workqueue_attrs(new_attrs);
3559 	apply_wqattrs_cleanup(ctx);
3560 	return NULL;
3561 }
3562 
3563 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3564 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3565 {
3566 	int node;
3567 
3568 	/* all pwqs have been created successfully, let's install'em */
3569 	mutex_lock(&ctx->wq->mutex);
3570 
3571 	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3572 
3573 	/* save the previous pwq and install the new one */
3574 	for_each_node(node)
3575 		ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3576 							  ctx->pwq_tbl[node]);
3577 
3578 	/* @dfl_pwq might not have been used, ensure it's linked */
3579 	link_pwq(ctx->dfl_pwq);
3580 	swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3581 
3582 	mutex_unlock(&ctx->wq->mutex);
3583 }
3584 
3585 static void apply_wqattrs_lock(void)
3586 {
3587 	/* CPUs should stay stable across pwq creations and installations */
3588 	get_online_cpus();
3589 	mutex_lock(&wq_pool_mutex);
3590 }
3591 
3592 static void apply_wqattrs_unlock(void)
3593 {
3594 	mutex_unlock(&wq_pool_mutex);
3595 	put_online_cpus();
3596 }
3597 
3598 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3599 					const struct workqueue_attrs *attrs)
3600 {
3601 	struct apply_wqattrs_ctx *ctx;
3602 	int ret = -ENOMEM;
3603 
3604 	/* only unbound workqueues can change attributes */
3605 	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3606 		return -EINVAL;
3607 
3608 	/* creating multiple pwqs breaks ordering guarantee */
3609 	if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3610 		return -EINVAL;
3611 
3612 	ctx = apply_wqattrs_prepare(wq, attrs);
3613 
3614 	/* the ctx has been prepared successfully, let's commit it */
3615 	if (ctx) {
3616 		apply_wqattrs_commit(ctx);
3617 		ret = 0;
3618 	}
3619 
3620 	apply_wqattrs_cleanup(ctx);
3621 
3622 	return ret;
3623 }
3624 
3625 /**
3626  * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3627  * @wq: the target workqueue
3628  * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3629  *
3630  * Apply @attrs to an unbound workqueue @wq.  Unless disabled, on NUMA
3631  * machines, this function maps a separate pwq to each NUMA node with
3632  * possibles CPUs in @attrs->cpumask so that work items are affine to the
3633  * NUMA node it was issued on.  Older pwqs are released as in-flight work
3634  * items finish.  Note that a work item which repeatedly requeues itself
3635  * back-to-back will stay on its current pwq.
3636  *
3637  * Performs GFP_KERNEL allocations.
3638  *
3639  * Return: 0 on success and -errno on failure.
3640  */
3641 int apply_workqueue_attrs(struct workqueue_struct *wq,
3642 			  const struct workqueue_attrs *attrs)
3643 {
3644 	int ret;
3645 
3646 	apply_wqattrs_lock();
3647 	ret = apply_workqueue_attrs_locked(wq, attrs);
3648 	apply_wqattrs_unlock();
3649 
3650 	return ret;
3651 }
3652 
3653 /**
3654  * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3655  * @wq: the target workqueue
3656  * @cpu: the CPU coming up or going down
3657  * @online: whether @cpu is coming up or going down
3658  *
3659  * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3660  * %CPU_DOWN_FAILED.  @cpu is being hot[un]plugged, update NUMA affinity of
3661  * @wq accordingly.
3662  *
3663  * If NUMA affinity can't be adjusted due to memory allocation failure, it
3664  * falls back to @wq->dfl_pwq which may not be optimal but is always
3665  * correct.
3666  *
3667  * Note that when the last allowed CPU of a NUMA node goes offline for a
3668  * workqueue with a cpumask spanning multiple nodes, the workers which were
3669  * already executing the work items for the workqueue will lose their CPU
3670  * affinity and may execute on any CPU.  This is similar to how per-cpu
3671  * workqueues behave on CPU_DOWN.  If a workqueue user wants strict
3672  * affinity, it's the user's responsibility to flush the work item from
3673  * CPU_DOWN_PREPARE.
3674  */
3675 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3676 				   bool online)
3677 {
3678 	int node = cpu_to_node(cpu);
3679 	int cpu_off = online ? -1 : cpu;
3680 	struct pool_workqueue *old_pwq = NULL, *pwq;
3681 	struct workqueue_attrs *target_attrs;
3682 	cpumask_t *cpumask;
3683 
3684 	lockdep_assert_held(&wq_pool_mutex);
3685 
3686 	if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3687 	    wq->unbound_attrs->no_numa)
3688 		return;
3689 
3690 	/*
3691 	 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3692 	 * Let's use a preallocated one.  The following buf is protected by
3693 	 * CPU hotplug exclusion.
3694 	 */
3695 	target_attrs = wq_update_unbound_numa_attrs_buf;
3696 	cpumask = target_attrs->cpumask;
3697 
3698 	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3699 	pwq = unbound_pwq_by_node(wq, node);
3700 
3701 	/*
3702 	 * Let's determine what needs to be done.  If the target cpumask is
3703 	 * different from the default pwq's, we need to compare it to @pwq's
3704 	 * and create a new one if they don't match.  If the target cpumask
3705 	 * equals the default pwq's, the default pwq should be used.
3706 	 */
3707 	if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3708 		if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3709 			return;
3710 	} else {
3711 		goto use_dfl_pwq;
3712 	}
3713 
3714 	/* create a new pwq */
3715 	pwq = alloc_unbound_pwq(wq, target_attrs);
3716 	if (!pwq) {
3717 		pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3718 			wq->name);
3719 		goto use_dfl_pwq;
3720 	}
3721 
3722 	/* Install the new pwq. */
3723 	mutex_lock(&wq->mutex);
3724 	old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3725 	goto out_unlock;
3726 
3727 use_dfl_pwq:
3728 	mutex_lock(&wq->mutex);
3729 	spin_lock_irq(&wq->dfl_pwq->pool->lock);
3730 	get_pwq(wq->dfl_pwq);
3731 	spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3732 	old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3733 out_unlock:
3734 	mutex_unlock(&wq->mutex);
3735 	put_pwq_unlocked(old_pwq);
3736 }
3737 
3738 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3739 {
3740 	bool highpri = wq->flags & WQ_HIGHPRI;
3741 	int cpu, ret;
3742 
3743 	if (!(wq->flags & WQ_UNBOUND)) {
3744 		wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3745 		if (!wq->cpu_pwqs)
3746 			return -ENOMEM;
3747 
3748 		for_each_possible_cpu(cpu) {
3749 			struct pool_workqueue *pwq =
3750 				per_cpu_ptr(wq->cpu_pwqs, cpu);
3751 			struct worker_pool *cpu_pools =
3752 				per_cpu(cpu_worker_pools, cpu);
3753 
3754 			init_pwq(pwq, wq, &cpu_pools[highpri]);
3755 
3756 			mutex_lock(&wq->mutex);
3757 			link_pwq(pwq);
3758 			mutex_unlock(&wq->mutex);
3759 		}
3760 		return 0;
3761 	} else if (wq->flags & __WQ_ORDERED) {
3762 		ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3763 		/* there should only be single pwq for ordering guarantee */
3764 		WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3765 			      wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3766 		     "ordering guarantee broken for workqueue %s\n", wq->name);
3767 		return ret;
3768 	} else {
3769 		return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3770 	}
3771 }
3772 
3773 static int wq_clamp_max_active(int max_active, unsigned int flags,
3774 			       const char *name)
3775 {
3776 	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3777 
3778 	if (max_active < 1 || max_active > lim)
3779 		pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3780 			max_active, name, 1, lim);
3781 
3782 	return clamp_val(max_active, 1, lim);
3783 }
3784 
3785 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3786 					       unsigned int flags,
3787 					       int max_active,
3788 					       struct lock_class_key *key,
3789 					       const char *lock_name, ...)
3790 {
3791 	size_t tbl_size = 0;
3792 	va_list args;
3793 	struct workqueue_struct *wq;
3794 	struct pool_workqueue *pwq;
3795 
3796 	/* see the comment above the definition of WQ_POWER_EFFICIENT */
3797 	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3798 		flags |= WQ_UNBOUND;
3799 
3800 	/* allocate wq and format name */
3801 	if (flags & WQ_UNBOUND)
3802 		tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3803 
3804 	wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3805 	if (!wq)
3806 		return NULL;
3807 
3808 	if (flags & WQ_UNBOUND) {
3809 		wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3810 		if (!wq->unbound_attrs)
3811 			goto err_free_wq;
3812 	}
3813 
3814 	va_start(args, lock_name);
3815 	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3816 	va_end(args);
3817 
3818 	max_active = max_active ?: WQ_DFL_ACTIVE;
3819 	max_active = wq_clamp_max_active(max_active, flags, wq->name);
3820 
3821 	/* init wq */
3822 	wq->flags = flags;
3823 	wq->saved_max_active = max_active;
3824 	mutex_init(&wq->mutex);
3825 	atomic_set(&wq->nr_pwqs_to_flush, 0);
3826 	INIT_LIST_HEAD(&wq->pwqs);
3827 	INIT_LIST_HEAD(&wq->flusher_queue);
3828 	INIT_LIST_HEAD(&wq->flusher_overflow);
3829 	INIT_LIST_HEAD(&wq->maydays);
3830 
3831 	lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3832 	INIT_LIST_HEAD(&wq->list);
3833 
3834 	if (alloc_and_link_pwqs(wq) < 0)
3835 		goto err_free_wq;
3836 
3837 	/*
3838 	 * Workqueues which may be used during memory reclaim should
3839 	 * have a rescuer to guarantee forward progress.
3840 	 */
3841 	if (flags & WQ_MEM_RECLAIM) {
3842 		struct worker *rescuer;
3843 
3844 		rescuer = alloc_worker(NUMA_NO_NODE);
3845 		if (!rescuer)
3846 			goto err_destroy;
3847 
3848 		rescuer->rescue_wq = wq;
3849 		rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3850 					       wq->name);
3851 		if (IS_ERR(rescuer->task)) {
3852 			kfree(rescuer);
3853 			goto err_destroy;
3854 		}
3855 
3856 		wq->rescuer = rescuer;
3857 		kthread_bind_mask(rescuer->task, cpu_possible_mask);
3858 		wake_up_process(rescuer->task);
3859 	}
3860 
3861 	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3862 		goto err_destroy;
3863 
3864 	/*
3865 	 * wq_pool_mutex protects global freeze state and workqueues list.
3866 	 * Grab it, adjust max_active and add the new @wq to workqueues
3867 	 * list.
3868 	 */
3869 	mutex_lock(&wq_pool_mutex);
3870 
3871 	mutex_lock(&wq->mutex);
3872 	for_each_pwq(pwq, wq)
3873 		pwq_adjust_max_active(pwq);
3874 	mutex_unlock(&wq->mutex);
3875 
3876 	list_add_tail_rcu(&wq->list, &workqueues);
3877 
3878 	mutex_unlock(&wq_pool_mutex);
3879 
3880 	return wq;
3881 
3882 err_free_wq:
3883 	free_workqueue_attrs(wq->unbound_attrs);
3884 	kfree(wq);
3885 	return NULL;
3886 err_destroy:
3887 	destroy_workqueue(wq);
3888 	return NULL;
3889 }
3890 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3891 
3892 /**
3893  * destroy_workqueue - safely terminate a workqueue
3894  * @wq: target workqueue
3895  *
3896  * Safely destroy a workqueue. All work currently pending will be done first.
3897  */
3898 void destroy_workqueue(struct workqueue_struct *wq)
3899 {
3900 	struct pool_workqueue *pwq;
3901 	int node;
3902 
3903 	/* drain it before proceeding with destruction */
3904 	drain_workqueue(wq);
3905 
3906 	/* sanity checks */
3907 	mutex_lock(&wq->mutex);
3908 	for_each_pwq(pwq, wq) {
3909 		int i;
3910 
3911 		for (i = 0; i < WORK_NR_COLORS; i++) {
3912 			if (WARN_ON(pwq->nr_in_flight[i])) {
3913 				mutex_unlock(&wq->mutex);
3914 				return;
3915 			}
3916 		}
3917 
3918 		if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
3919 		    WARN_ON(pwq->nr_active) ||
3920 		    WARN_ON(!list_empty(&pwq->delayed_works))) {
3921 			mutex_unlock(&wq->mutex);
3922 			return;
3923 		}
3924 	}
3925 	mutex_unlock(&wq->mutex);
3926 
3927 	/*
3928 	 * wq list is used to freeze wq, remove from list after
3929 	 * flushing is complete in case freeze races us.
3930 	 */
3931 	mutex_lock(&wq_pool_mutex);
3932 	list_del_rcu(&wq->list);
3933 	mutex_unlock(&wq_pool_mutex);
3934 
3935 	workqueue_sysfs_unregister(wq);
3936 
3937 	if (wq->rescuer)
3938 		kthread_stop(wq->rescuer->task);
3939 
3940 	if (!(wq->flags & WQ_UNBOUND)) {
3941 		/*
3942 		 * The base ref is never dropped on per-cpu pwqs.  Directly
3943 		 * schedule RCU free.
3944 		 */
3945 		call_rcu_sched(&wq->rcu, rcu_free_wq);
3946 	} else {
3947 		/*
3948 		 * We're the sole accessor of @wq at this point.  Directly
3949 		 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
3950 		 * @wq will be freed when the last pwq is released.
3951 		 */
3952 		for_each_node(node) {
3953 			pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3954 			RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
3955 			put_pwq_unlocked(pwq);
3956 		}
3957 
3958 		/*
3959 		 * Put dfl_pwq.  @wq may be freed any time after dfl_pwq is
3960 		 * put.  Don't access it afterwards.
3961 		 */
3962 		pwq = wq->dfl_pwq;
3963 		wq->dfl_pwq = NULL;
3964 		put_pwq_unlocked(pwq);
3965 	}
3966 }
3967 EXPORT_SYMBOL_GPL(destroy_workqueue);
3968 
3969 /**
3970  * workqueue_set_max_active - adjust max_active of a workqueue
3971  * @wq: target workqueue
3972  * @max_active: new max_active value.
3973  *
3974  * Set max_active of @wq to @max_active.
3975  *
3976  * CONTEXT:
3977  * Don't call from IRQ context.
3978  */
3979 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3980 {
3981 	struct pool_workqueue *pwq;
3982 
3983 	/* disallow meddling with max_active for ordered workqueues */
3984 	if (WARN_ON(wq->flags & __WQ_ORDERED))
3985 		return;
3986 
3987 	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3988 
3989 	mutex_lock(&wq->mutex);
3990 
3991 	wq->saved_max_active = max_active;
3992 
3993 	for_each_pwq(pwq, wq)
3994 		pwq_adjust_max_active(pwq);
3995 
3996 	mutex_unlock(&wq->mutex);
3997 }
3998 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3999 
4000 /**
4001  * current_is_workqueue_rescuer - is %current workqueue rescuer?
4002  *
4003  * Determine whether %current is a workqueue rescuer.  Can be used from
4004  * work functions to determine whether it's being run off the rescuer task.
4005  *
4006  * Return: %true if %current is a workqueue rescuer. %false otherwise.
4007  */
4008 bool current_is_workqueue_rescuer(void)
4009 {
4010 	struct worker *worker = current_wq_worker();
4011 
4012 	return worker && worker->rescue_wq;
4013 }
4014 
4015 /**
4016  * workqueue_congested - test whether a workqueue is congested
4017  * @cpu: CPU in question
4018  * @wq: target workqueue
4019  *
4020  * Test whether @wq's cpu workqueue for @cpu is congested.  There is
4021  * no synchronization around this function and the test result is
4022  * unreliable and only useful as advisory hints or for debugging.
4023  *
4024  * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4025  * Note that both per-cpu and unbound workqueues may be associated with
4026  * multiple pool_workqueues which have separate congested states.  A
4027  * workqueue being congested on one CPU doesn't mean the workqueue is also
4028  * contested on other CPUs / NUMA nodes.
4029  *
4030  * Return:
4031  * %true if congested, %false otherwise.
4032  */
4033 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4034 {
4035 	struct pool_workqueue *pwq;
4036 	bool ret;
4037 
4038 	rcu_read_lock_sched();
4039 
4040 	if (cpu == WORK_CPU_UNBOUND)
4041 		cpu = smp_processor_id();
4042 
4043 	if (!(wq->flags & WQ_UNBOUND))
4044 		pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4045 	else
4046 		pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4047 
4048 	ret = !list_empty(&pwq->delayed_works);
4049 	rcu_read_unlock_sched();
4050 
4051 	return ret;
4052 }
4053 EXPORT_SYMBOL_GPL(workqueue_congested);
4054 
4055 /**
4056  * work_busy - test whether a work is currently pending or running
4057  * @work: the work to be tested
4058  *
4059  * Test whether @work is currently pending or running.  There is no
4060  * synchronization around this function and the test result is
4061  * unreliable and only useful as advisory hints or for debugging.
4062  *
4063  * Return:
4064  * OR'd bitmask of WORK_BUSY_* bits.
4065  */
4066 unsigned int work_busy(struct work_struct *work)
4067 {
4068 	struct worker_pool *pool;
4069 	unsigned long flags;
4070 	unsigned int ret = 0;
4071 
4072 	if (work_pending(work))
4073 		ret |= WORK_BUSY_PENDING;
4074 
4075 	local_irq_save(flags);
4076 	pool = get_work_pool(work);
4077 	if (pool) {
4078 		spin_lock(&pool->lock);
4079 		if (find_worker_executing_work(pool, work))
4080 			ret |= WORK_BUSY_RUNNING;
4081 		spin_unlock(&pool->lock);
4082 	}
4083 	local_irq_restore(flags);
4084 
4085 	return ret;
4086 }
4087 EXPORT_SYMBOL_GPL(work_busy);
4088 
4089 /**
4090  * set_worker_desc - set description for the current work item
4091  * @fmt: printf-style format string
4092  * @...: arguments for the format string
4093  *
4094  * This function can be called by a running work function to describe what
4095  * the work item is about.  If the worker task gets dumped, this
4096  * information will be printed out together to help debugging.  The
4097  * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4098  */
4099 void set_worker_desc(const char *fmt, ...)
4100 {
4101 	struct worker *worker = current_wq_worker();
4102 	va_list args;
4103 
4104 	if (worker) {
4105 		va_start(args, fmt);
4106 		vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4107 		va_end(args);
4108 		worker->desc_valid = true;
4109 	}
4110 }
4111 
4112 /**
4113  * print_worker_info - print out worker information and description
4114  * @log_lvl: the log level to use when printing
4115  * @task: target task
4116  *
4117  * If @task is a worker and currently executing a work item, print out the
4118  * name of the workqueue being serviced and worker description set with
4119  * set_worker_desc() by the currently executing work item.
4120  *
4121  * This function can be safely called on any task as long as the
4122  * task_struct itself is accessible.  While safe, this function isn't
4123  * synchronized and may print out mixups or garbages of limited length.
4124  */
4125 void print_worker_info(const char *log_lvl, struct task_struct *task)
4126 {
4127 	work_func_t *fn = NULL;
4128 	char name[WQ_NAME_LEN] = { };
4129 	char desc[WORKER_DESC_LEN] = { };
4130 	struct pool_workqueue *pwq = NULL;
4131 	struct workqueue_struct *wq = NULL;
4132 	bool desc_valid = false;
4133 	struct worker *worker;
4134 
4135 	if (!(task->flags & PF_WQ_WORKER))
4136 		return;
4137 
4138 	/*
4139 	 * This function is called without any synchronization and @task
4140 	 * could be in any state.  Be careful with dereferences.
4141 	 */
4142 	worker = probe_kthread_data(task);
4143 
4144 	/*
4145 	 * Carefully copy the associated workqueue's workfn and name.  Keep
4146 	 * the original last '\0' in case the original contains garbage.
4147 	 */
4148 	probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4149 	probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4150 	probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4151 	probe_kernel_read(name, wq->name, sizeof(name) - 1);
4152 
4153 	/* copy worker description */
4154 	probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4155 	if (desc_valid)
4156 		probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4157 
4158 	if (fn || name[0] || desc[0]) {
4159 		printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4160 		if (desc[0])
4161 			pr_cont(" (%s)", desc);
4162 		pr_cont("\n");
4163 	}
4164 }
4165 
4166 static void pr_cont_pool_info(struct worker_pool *pool)
4167 {
4168 	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4169 	if (pool->node != NUMA_NO_NODE)
4170 		pr_cont(" node=%d", pool->node);
4171 	pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4172 }
4173 
4174 static void pr_cont_work(bool comma, struct work_struct *work)
4175 {
4176 	if (work->func == wq_barrier_func) {
4177 		struct wq_barrier *barr;
4178 
4179 		barr = container_of(work, struct wq_barrier, work);
4180 
4181 		pr_cont("%s BAR(%d)", comma ? "," : "",
4182 			task_pid_nr(barr->task));
4183 	} else {
4184 		pr_cont("%s %pf", comma ? "," : "", work->func);
4185 	}
4186 }
4187 
4188 static void show_pwq(struct pool_workqueue *pwq)
4189 {
4190 	struct worker_pool *pool = pwq->pool;
4191 	struct work_struct *work;
4192 	struct worker *worker;
4193 	bool has_in_flight = false, has_pending = false;
4194 	int bkt;
4195 
4196 	pr_info("  pwq %d:", pool->id);
4197 	pr_cont_pool_info(pool);
4198 
4199 	pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4200 		!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4201 
4202 	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4203 		if (worker->current_pwq == pwq) {
4204 			has_in_flight = true;
4205 			break;
4206 		}
4207 	}
4208 	if (has_in_flight) {
4209 		bool comma = false;
4210 
4211 		pr_info("    in-flight:");
4212 		hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4213 			if (worker->current_pwq != pwq)
4214 				continue;
4215 
4216 			pr_cont("%s %d%s:%pf", comma ? "," : "",
4217 				task_pid_nr(worker->task),
4218 				worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4219 				worker->current_func);
4220 			list_for_each_entry(work, &worker->scheduled, entry)
4221 				pr_cont_work(false, work);
4222 			comma = true;
4223 		}
4224 		pr_cont("\n");
4225 	}
4226 
4227 	list_for_each_entry(work, &pool->worklist, entry) {
4228 		if (get_work_pwq(work) == pwq) {
4229 			has_pending = true;
4230 			break;
4231 		}
4232 	}
4233 	if (has_pending) {
4234 		bool comma = false;
4235 
4236 		pr_info("    pending:");
4237 		list_for_each_entry(work, &pool->worklist, entry) {
4238 			if (get_work_pwq(work) != pwq)
4239 				continue;
4240 
4241 			pr_cont_work(comma, work);
4242 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4243 		}
4244 		pr_cont("\n");
4245 	}
4246 
4247 	if (!list_empty(&pwq->delayed_works)) {
4248 		bool comma = false;
4249 
4250 		pr_info("    delayed:");
4251 		list_for_each_entry(work, &pwq->delayed_works, entry) {
4252 			pr_cont_work(comma, work);
4253 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4254 		}
4255 		pr_cont("\n");
4256 	}
4257 }
4258 
4259 /**
4260  * show_workqueue_state - dump workqueue state
4261  *
4262  * Called from a sysrq handler and prints out all busy workqueues and
4263  * pools.
4264  */
4265 void show_workqueue_state(void)
4266 {
4267 	struct workqueue_struct *wq;
4268 	struct worker_pool *pool;
4269 	unsigned long flags;
4270 	int pi;
4271 
4272 	rcu_read_lock_sched();
4273 
4274 	pr_info("Showing busy workqueues and worker pools:\n");
4275 
4276 	list_for_each_entry_rcu(wq, &workqueues, list) {
4277 		struct pool_workqueue *pwq;
4278 		bool idle = true;
4279 
4280 		for_each_pwq(pwq, wq) {
4281 			if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4282 				idle = false;
4283 				break;
4284 			}
4285 		}
4286 		if (idle)
4287 			continue;
4288 
4289 		pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4290 
4291 		for_each_pwq(pwq, wq) {
4292 			spin_lock_irqsave(&pwq->pool->lock, flags);
4293 			if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4294 				show_pwq(pwq);
4295 			spin_unlock_irqrestore(&pwq->pool->lock, flags);
4296 		}
4297 	}
4298 
4299 	for_each_pool(pool, pi) {
4300 		struct worker *worker;
4301 		bool first = true;
4302 
4303 		spin_lock_irqsave(&pool->lock, flags);
4304 		if (pool->nr_workers == pool->nr_idle)
4305 			goto next_pool;
4306 
4307 		pr_info("pool %d:", pool->id);
4308 		pr_cont_pool_info(pool);
4309 		pr_cont(" workers=%d", pool->nr_workers);
4310 		if (pool->manager)
4311 			pr_cont(" manager: %d",
4312 				task_pid_nr(pool->manager->task));
4313 		list_for_each_entry(worker, &pool->idle_list, entry) {
4314 			pr_cont(" %s%d", first ? "idle: " : "",
4315 				task_pid_nr(worker->task));
4316 			first = false;
4317 		}
4318 		pr_cont("\n");
4319 	next_pool:
4320 		spin_unlock_irqrestore(&pool->lock, flags);
4321 	}
4322 
4323 	rcu_read_unlock_sched();
4324 }
4325 
4326 /*
4327  * CPU hotplug.
4328  *
4329  * There are two challenges in supporting CPU hotplug.  Firstly, there
4330  * are a lot of assumptions on strong associations among work, pwq and
4331  * pool which make migrating pending and scheduled works very
4332  * difficult to implement without impacting hot paths.  Secondly,
4333  * worker pools serve mix of short, long and very long running works making
4334  * blocked draining impractical.
4335  *
4336  * This is solved by allowing the pools to be disassociated from the CPU
4337  * running as an unbound one and allowing it to be reattached later if the
4338  * cpu comes back online.
4339  */
4340 
4341 static void wq_unbind_fn(struct work_struct *work)
4342 {
4343 	int cpu = smp_processor_id();
4344 	struct worker_pool *pool;
4345 	struct worker *worker;
4346 
4347 	for_each_cpu_worker_pool(pool, cpu) {
4348 		mutex_lock(&pool->attach_mutex);
4349 		spin_lock_irq(&pool->lock);
4350 
4351 		/*
4352 		 * We've blocked all attach/detach operations. Make all workers
4353 		 * unbound and set DISASSOCIATED.  Before this, all workers
4354 		 * except for the ones which are still executing works from
4355 		 * before the last CPU down must be on the cpu.  After
4356 		 * this, they may become diasporas.
4357 		 */
4358 		for_each_pool_worker(worker, pool)
4359 			worker->flags |= WORKER_UNBOUND;
4360 
4361 		pool->flags |= POOL_DISASSOCIATED;
4362 
4363 		spin_unlock_irq(&pool->lock);
4364 		mutex_unlock(&pool->attach_mutex);
4365 
4366 		/*
4367 		 * Call schedule() so that we cross rq->lock and thus can
4368 		 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4369 		 * This is necessary as scheduler callbacks may be invoked
4370 		 * from other cpus.
4371 		 */
4372 		schedule();
4373 
4374 		/*
4375 		 * Sched callbacks are disabled now.  Zap nr_running.
4376 		 * After this, nr_running stays zero and need_more_worker()
4377 		 * and keep_working() are always true as long as the
4378 		 * worklist is not empty.  This pool now behaves as an
4379 		 * unbound (in terms of concurrency management) pool which
4380 		 * are served by workers tied to the pool.
4381 		 */
4382 		atomic_set(&pool->nr_running, 0);
4383 
4384 		/*
4385 		 * With concurrency management just turned off, a busy
4386 		 * worker blocking could lead to lengthy stalls.  Kick off
4387 		 * unbound chain execution of currently pending work items.
4388 		 */
4389 		spin_lock_irq(&pool->lock);
4390 		wake_up_worker(pool);
4391 		spin_unlock_irq(&pool->lock);
4392 	}
4393 }
4394 
4395 /**
4396  * rebind_workers - rebind all workers of a pool to the associated CPU
4397  * @pool: pool of interest
4398  *
4399  * @pool->cpu is coming online.  Rebind all workers to the CPU.
4400  */
4401 static void rebind_workers(struct worker_pool *pool)
4402 {
4403 	struct worker *worker;
4404 
4405 	lockdep_assert_held(&pool->attach_mutex);
4406 
4407 	/*
4408 	 * Restore CPU affinity of all workers.  As all idle workers should
4409 	 * be on the run-queue of the associated CPU before any local
4410 	 * wake-ups for concurrency management happen, restore CPU affinity
4411 	 * of all workers first and then clear UNBOUND.  As we're called
4412 	 * from CPU_ONLINE, the following shouldn't fail.
4413 	 */
4414 	for_each_pool_worker(worker, pool)
4415 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4416 						  pool->attrs->cpumask) < 0);
4417 
4418 	spin_lock_irq(&pool->lock);
4419 	pool->flags &= ~POOL_DISASSOCIATED;
4420 
4421 	for_each_pool_worker(worker, pool) {
4422 		unsigned int worker_flags = worker->flags;
4423 
4424 		/*
4425 		 * A bound idle worker should actually be on the runqueue
4426 		 * of the associated CPU for local wake-ups targeting it to
4427 		 * work.  Kick all idle workers so that they migrate to the
4428 		 * associated CPU.  Doing this in the same loop as
4429 		 * replacing UNBOUND with REBOUND is safe as no worker will
4430 		 * be bound before @pool->lock is released.
4431 		 */
4432 		if (worker_flags & WORKER_IDLE)
4433 			wake_up_process(worker->task);
4434 
4435 		/*
4436 		 * We want to clear UNBOUND but can't directly call
4437 		 * worker_clr_flags() or adjust nr_running.  Atomically
4438 		 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4439 		 * @worker will clear REBOUND using worker_clr_flags() when
4440 		 * it initiates the next execution cycle thus restoring
4441 		 * concurrency management.  Note that when or whether
4442 		 * @worker clears REBOUND doesn't affect correctness.
4443 		 *
4444 		 * ACCESS_ONCE() is necessary because @worker->flags may be
4445 		 * tested without holding any lock in
4446 		 * wq_worker_waking_up().  Without it, NOT_RUNNING test may
4447 		 * fail incorrectly leading to premature concurrency
4448 		 * management operations.
4449 		 */
4450 		WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4451 		worker_flags |= WORKER_REBOUND;
4452 		worker_flags &= ~WORKER_UNBOUND;
4453 		ACCESS_ONCE(worker->flags) = worker_flags;
4454 	}
4455 
4456 	spin_unlock_irq(&pool->lock);
4457 }
4458 
4459 /**
4460  * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4461  * @pool: unbound pool of interest
4462  * @cpu: the CPU which is coming up
4463  *
4464  * An unbound pool may end up with a cpumask which doesn't have any online
4465  * CPUs.  When a worker of such pool get scheduled, the scheduler resets
4466  * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
4467  * online CPU before, cpus_allowed of all its workers should be restored.
4468  */
4469 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4470 {
4471 	static cpumask_t cpumask;
4472 	struct worker *worker;
4473 
4474 	lockdep_assert_held(&pool->attach_mutex);
4475 
4476 	/* is @cpu allowed for @pool? */
4477 	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4478 		return;
4479 
4480 	/* is @cpu the only online CPU? */
4481 	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4482 	if (cpumask_weight(&cpumask) != 1)
4483 		return;
4484 
4485 	/* as we're called from CPU_ONLINE, the following shouldn't fail */
4486 	for_each_pool_worker(worker, pool)
4487 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4488 						  pool->attrs->cpumask) < 0);
4489 }
4490 
4491 /*
4492  * Workqueues should be brought up before normal priority CPU notifiers.
4493  * This will be registered high priority CPU notifier.
4494  */
4495 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4496 					       unsigned long action,
4497 					       void *hcpu)
4498 {
4499 	int cpu = (unsigned long)hcpu;
4500 	struct worker_pool *pool;
4501 	struct workqueue_struct *wq;
4502 	int pi;
4503 
4504 	switch (action & ~CPU_TASKS_FROZEN) {
4505 	case CPU_UP_PREPARE:
4506 		for_each_cpu_worker_pool(pool, cpu) {
4507 			if (pool->nr_workers)
4508 				continue;
4509 			if (!create_worker(pool))
4510 				return NOTIFY_BAD;
4511 		}
4512 		break;
4513 
4514 	case CPU_DOWN_FAILED:
4515 	case CPU_ONLINE:
4516 		mutex_lock(&wq_pool_mutex);
4517 
4518 		for_each_pool(pool, pi) {
4519 			mutex_lock(&pool->attach_mutex);
4520 
4521 			if (pool->cpu == cpu)
4522 				rebind_workers(pool);
4523 			else if (pool->cpu < 0)
4524 				restore_unbound_workers_cpumask(pool, cpu);
4525 
4526 			mutex_unlock(&pool->attach_mutex);
4527 		}
4528 
4529 		/* update NUMA affinity of unbound workqueues */
4530 		list_for_each_entry(wq, &workqueues, list)
4531 			wq_update_unbound_numa(wq, cpu, true);
4532 
4533 		mutex_unlock(&wq_pool_mutex);
4534 		break;
4535 	}
4536 	return NOTIFY_OK;
4537 }
4538 
4539 /*
4540  * Workqueues should be brought down after normal priority CPU notifiers.
4541  * This will be registered as low priority CPU notifier.
4542  */
4543 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4544 						 unsigned long action,
4545 						 void *hcpu)
4546 {
4547 	int cpu = (unsigned long)hcpu;
4548 	struct work_struct unbind_work;
4549 	struct workqueue_struct *wq;
4550 
4551 	switch (action & ~CPU_TASKS_FROZEN) {
4552 	case CPU_DOWN_PREPARE:
4553 		/* unbinding per-cpu workers should happen on the local CPU */
4554 		INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4555 		queue_work_on(cpu, system_highpri_wq, &unbind_work);
4556 
4557 		/* update NUMA affinity of unbound workqueues */
4558 		mutex_lock(&wq_pool_mutex);
4559 		list_for_each_entry(wq, &workqueues, list)
4560 			wq_update_unbound_numa(wq, cpu, false);
4561 		mutex_unlock(&wq_pool_mutex);
4562 
4563 		/* wait for per-cpu unbinding to finish */
4564 		flush_work(&unbind_work);
4565 		destroy_work_on_stack(&unbind_work);
4566 		break;
4567 	}
4568 	return NOTIFY_OK;
4569 }
4570 
4571 #ifdef CONFIG_SMP
4572 
4573 struct work_for_cpu {
4574 	struct work_struct work;
4575 	long (*fn)(void *);
4576 	void *arg;
4577 	long ret;
4578 };
4579 
4580 static void work_for_cpu_fn(struct work_struct *work)
4581 {
4582 	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4583 
4584 	wfc->ret = wfc->fn(wfc->arg);
4585 }
4586 
4587 /**
4588  * work_on_cpu - run a function in user context on a particular cpu
4589  * @cpu: the cpu to run on
4590  * @fn: the function to run
4591  * @arg: the function arg
4592  *
4593  * It is up to the caller to ensure that the cpu doesn't go offline.
4594  * The caller must not hold any locks which would prevent @fn from completing.
4595  *
4596  * Return: The value @fn returns.
4597  */
4598 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4599 {
4600 	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4601 
4602 	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4603 	schedule_work_on(cpu, &wfc.work);
4604 	flush_work(&wfc.work);
4605 	destroy_work_on_stack(&wfc.work);
4606 	return wfc.ret;
4607 }
4608 EXPORT_SYMBOL_GPL(work_on_cpu);
4609 #endif /* CONFIG_SMP */
4610 
4611 #ifdef CONFIG_FREEZER
4612 
4613 /**
4614  * freeze_workqueues_begin - begin freezing workqueues
4615  *
4616  * Start freezing workqueues.  After this function returns, all freezable
4617  * workqueues will queue new works to their delayed_works list instead of
4618  * pool->worklist.
4619  *
4620  * CONTEXT:
4621  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4622  */
4623 void freeze_workqueues_begin(void)
4624 {
4625 	struct workqueue_struct *wq;
4626 	struct pool_workqueue *pwq;
4627 
4628 	mutex_lock(&wq_pool_mutex);
4629 
4630 	WARN_ON_ONCE(workqueue_freezing);
4631 	workqueue_freezing = true;
4632 
4633 	list_for_each_entry(wq, &workqueues, list) {
4634 		mutex_lock(&wq->mutex);
4635 		for_each_pwq(pwq, wq)
4636 			pwq_adjust_max_active(pwq);
4637 		mutex_unlock(&wq->mutex);
4638 	}
4639 
4640 	mutex_unlock(&wq_pool_mutex);
4641 }
4642 
4643 /**
4644  * freeze_workqueues_busy - are freezable workqueues still busy?
4645  *
4646  * Check whether freezing is complete.  This function must be called
4647  * between freeze_workqueues_begin() and thaw_workqueues().
4648  *
4649  * CONTEXT:
4650  * Grabs and releases wq_pool_mutex.
4651  *
4652  * Return:
4653  * %true if some freezable workqueues are still busy.  %false if freezing
4654  * is complete.
4655  */
4656 bool freeze_workqueues_busy(void)
4657 {
4658 	bool busy = false;
4659 	struct workqueue_struct *wq;
4660 	struct pool_workqueue *pwq;
4661 
4662 	mutex_lock(&wq_pool_mutex);
4663 
4664 	WARN_ON_ONCE(!workqueue_freezing);
4665 
4666 	list_for_each_entry(wq, &workqueues, list) {
4667 		if (!(wq->flags & WQ_FREEZABLE))
4668 			continue;
4669 		/*
4670 		 * nr_active is monotonically decreasing.  It's safe
4671 		 * to peek without lock.
4672 		 */
4673 		rcu_read_lock_sched();
4674 		for_each_pwq(pwq, wq) {
4675 			WARN_ON_ONCE(pwq->nr_active < 0);
4676 			if (pwq->nr_active) {
4677 				busy = true;
4678 				rcu_read_unlock_sched();
4679 				goto out_unlock;
4680 			}
4681 		}
4682 		rcu_read_unlock_sched();
4683 	}
4684 out_unlock:
4685 	mutex_unlock(&wq_pool_mutex);
4686 	return busy;
4687 }
4688 
4689 /**
4690  * thaw_workqueues - thaw workqueues
4691  *
4692  * Thaw workqueues.  Normal queueing is restored and all collected
4693  * frozen works are transferred to their respective pool worklists.
4694  *
4695  * CONTEXT:
4696  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4697  */
4698 void thaw_workqueues(void)
4699 {
4700 	struct workqueue_struct *wq;
4701 	struct pool_workqueue *pwq;
4702 
4703 	mutex_lock(&wq_pool_mutex);
4704 
4705 	if (!workqueue_freezing)
4706 		goto out_unlock;
4707 
4708 	workqueue_freezing = false;
4709 
4710 	/* restore max_active and repopulate worklist */
4711 	list_for_each_entry(wq, &workqueues, list) {
4712 		mutex_lock(&wq->mutex);
4713 		for_each_pwq(pwq, wq)
4714 			pwq_adjust_max_active(pwq);
4715 		mutex_unlock(&wq->mutex);
4716 	}
4717 
4718 out_unlock:
4719 	mutex_unlock(&wq_pool_mutex);
4720 }
4721 #endif /* CONFIG_FREEZER */
4722 
4723 static int workqueue_apply_unbound_cpumask(void)
4724 {
4725 	LIST_HEAD(ctxs);
4726 	int ret = 0;
4727 	struct workqueue_struct *wq;
4728 	struct apply_wqattrs_ctx *ctx, *n;
4729 
4730 	lockdep_assert_held(&wq_pool_mutex);
4731 
4732 	list_for_each_entry(wq, &workqueues, list) {
4733 		if (!(wq->flags & WQ_UNBOUND))
4734 			continue;
4735 		/* creating multiple pwqs breaks ordering guarantee */
4736 		if (wq->flags & __WQ_ORDERED)
4737 			continue;
4738 
4739 		ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4740 		if (!ctx) {
4741 			ret = -ENOMEM;
4742 			break;
4743 		}
4744 
4745 		list_add_tail(&ctx->list, &ctxs);
4746 	}
4747 
4748 	list_for_each_entry_safe(ctx, n, &ctxs, list) {
4749 		if (!ret)
4750 			apply_wqattrs_commit(ctx);
4751 		apply_wqattrs_cleanup(ctx);
4752 	}
4753 
4754 	return ret;
4755 }
4756 
4757 /**
4758  *  workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4759  *  @cpumask: the cpumask to set
4760  *
4761  *  The low-level workqueues cpumask is a global cpumask that limits
4762  *  the affinity of all unbound workqueues.  This function check the @cpumask
4763  *  and apply it to all unbound workqueues and updates all pwqs of them.
4764  *
4765  *  Retun:	0	- Success
4766  *  		-EINVAL	- Invalid @cpumask
4767  *  		-ENOMEM	- Failed to allocate memory for attrs or pwqs.
4768  */
4769 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4770 {
4771 	int ret = -EINVAL;
4772 	cpumask_var_t saved_cpumask;
4773 
4774 	if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4775 		return -ENOMEM;
4776 
4777 	cpumask_and(cpumask, cpumask, cpu_possible_mask);
4778 	if (!cpumask_empty(cpumask)) {
4779 		apply_wqattrs_lock();
4780 
4781 		/* save the old wq_unbound_cpumask. */
4782 		cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4783 
4784 		/* update wq_unbound_cpumask at first and apply it to wqs. */
4785 		cpumask_copy(wq_unbound_cpumask, cpumask);
4786 		ret = workqueue_apply_unbound_cpumask();
4787 
4788 		/* restore the wq_unbound_cpumask when failed. */
4789 		if (ret < 0)
4790 			cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4791 
4792 		apply_wqattrs_unlock();
4793 	}
4794 
4795 	free_cpumask_var(saved_cpumask);
4796 	return ret;
4797 }
4798 
4799 #ifdef CONFIG_SYSFS
4800 /*
4801  * Workqueues with WQ_SYSFS flag set is visible to userland via
4802  * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
4803  * following attributes.
4804  *
4805  *  per_cpu	RO bool	: whether the workqueue is per-cpu or unbound
4806  *  max_active	RW int	: maximum number of in-flight work items
4807  *
4808  * Unbound workqueues have the following extra attributes.
4809  *
4810  *  id		RO int	: the associated pool ID
4811  *  nice	RW int	: nice value of the workers
4812  *  cpumask	RW mask	: bitmask of allowed CPUs for the workers
4813  */
4814 struct wq_device {
4815 	struct workqueue_struct		*wq;
4816 	struct device			dev;
4817 };
4818 
4819 static struct workqueue_struct *dev_to_wq(struct device *dev)
4820 {
4821 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4822 
4823 	return wq_dev->wq;
4824 }
4825 
4826 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4827 			    char *buf)
4828 {
4829 	struct workqueue_struct *wq = dev_to_wq(dev);
4830 
4831 	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4832 }
4833 static DEVICE_ATTR_RO(per_cpu);
4834 
4835 static ssize_t max_active_show(struct device *dev,
4836 			       struct device_attribute *attr, char *buf)
4837 {
4838 	struct workqueue_struct *wq = dev_to_wq(dev);
4839 
4840 	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4841 }
4842 
4843 static ssize_t max_active_store(struct device *dev,
4844 				struct device_attribute *attr, const char *buf,
4845 				size_t count)
4846 {
4847 	struct workqueue_struct *wq = dev_to_wq(dev);
4848 	int val;
4849 
4850 	if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4851 		return -EINVAL;
4852 
4853 	workqueue_set_max_active(wq, val);
4854 	return count;
4855 }
4856 static DEVICE_ATTR_RW(max_active);
4857 
4858 static struct attribute *wq_sysfs_attrs[] = {
4859 	&dev_attr_per_cpu.attr,
4860 	&dev_attr_max_active.attr,
4861 	NULL,
4862 };
4863 ATTRIBUTE_GROUPS(wq_sysfs);
4864 
4865 static ssize_t wq_pool_ids_show(struct device *dev,
4866 				struct device_attribute *attr, char *buf)
4867 {
4868 	struct workqueue_struct *wq = dev_to_wq(dev);
4869 	const char *delim = "";
4870 	int node, written = 0;
4871 
4872 	rcu_read_lock_sched();
4873 	for_each_node(node) {
4874 		written += scnprintf(buf + written, PAGE_SIZE - written,
4875 				     "%s%d:%d", delim, node,
4876 				     unbound_pwq_by_node(wq, node)->pool->id);
4877 		delim = " ";
4878 	}
4879 	written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4880 	rcu_read_unlock_sched();
4881 
4882 	return written;
4883 }
4884 
4885 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
4886 			    char *buf)
4887 {
4888 	struct workqueue_struct *wq = dev_to_wq(dev);
4889 	int written;
4890 
4891 	mutex_lock(&wq->mutex);
4892 	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
4893 	mutex_unlock(&wq->mutex);
4894 
4895 	return written;
4896 }
4897 
4898 /* prepare workqueue_attrs for sysfs store operations */
4899 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
4900 {
4901 	struct workqueue_attrs *attrs;
4902 
4903 	lockdep_assert_held(&wq_pool_mutex);
4904 
4905 	attrs = alloc_workqueue_attrs(GFP_KERNEL);
4906 	if (!attrs)
4907 		return NULL;
4908 
4909 	copy_workqueue_attrs(attrs, wq->unbound_attrs);
4910 	return attrs;
4911 }
4912 
4913 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
4914 			     const char *buf, size_t count)
4915 {
4916 	struct workqueue_struct *wq = dev_to_wq(dev);
4917 	struct workqueue_attrs *attrs;
4918 	int ret = -ENOMEM;
4919 
4920 	apply_wqattrs_lock();
4921 
4922 	attrs = wq_sysfs_prep_attrs(wq);
4923 	if (!attrs)
4924 		goto out_unlock;
4925 
4926 	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
4927 	    attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
4928 		ret = apply_workqueue_attrs_locked(wq, attrs);
4929 	else
4930 		ret = -EINVAL;
4931 
4932 out_unlock:
4933 	apply_wqattrs_unlock();
4934 	free_workqueue_attrs(attrs);
4935 	return ret ?: count;
4936 }
4937 
4938 static ssize_t wq_cpumask_show(struct device *dev,
4939 			       struct device_attribute *attr, char *buf)
4940 {
4941 	struct workqueue_struct *wq = dev_to_wq(dev);
4942 	int written;
4943 
4944 	mutex_lock(&wq->mutex);
4945 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
4946 			    cpumask_pr_args(wq->unbound_attrs->cpumask));
4947 	mutex_unlock(&wq->mutex);
4948 	return written;
4949 }
4950 
4951 static ssize_t wq_cpumask_store(struct device *dev,
4952 				struct device_attribute *attr,
4953 				const char *buf, size_t count)
4954 {
4955 	struct workqueue_struct *wq = dev_to_wq(dev);
4956 	struct workqueue_attrs *attrs;
4957 	int ret = -ENOMEM;
4958 
4959 	apply_wqattrs_lock();
4960 
4961 	attrs = wq_sysfs_prep_attrs(wq);
4962 	if (!attrs)
4963 		goto out_unlock;
4964 
4965 	ret = cpumask_parse(buf, attrs->cpumask);
4966 	if (!ret)
4967 		ret = apply_workqueue_attrs_locked(wq, attrs);
4968 
4969 out_unlock:
4970 	apply_wqattrs_unlock();
4971 	free_workqueue_attrs(attrs);
4972 	return ret ?: count;
4973 }
4974 
4975 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
4976 			    char *buf)
4977 {
4978 	struct workqueue_struct *wq = dev_to_wq(dev);
4979 	int written;
4980 
4981 	mutex_lock(&wq->mutex);
4982 	written = scnprintf(buf, PAGE_SIZE, "%d\n",
4983 			    !wq->unbound_attrs->no_numa);
4984 	mutex_unlock(&wq->mutex);
4985 
4986 	return written;
4987 }
4988 
4989 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
4990 			     const char *buf, size_t count)
4991 {
4992 	struct workqueue_struct *wq = dev_to_wq(dev);
4993 	struct workqueue_attrs *attrs;
4994 	int v, ret = -ENOMEM;
4995 
4996 	apply_wqattrs_lock();
4997 
4998 	attrs = wq_sysfs_prep_attrs(wq);
4999 	if (!attrs)
5000 		goto out_unlock;
5001 
5002 	ret = -EINVAL;
5003 	if (sscanf(buf, "%d", &v) == 1) {
5004 		attrs->no_numa = !v;
5005 		ret = apply_workqueue_attrs_locked(wq, attrs);
5006 	}
5007 
5008 out_unlock:
5009 	apply_wqattrs_unlock();
5010 	free_workqueue_attrs(attrs);
5011 	return ret ?: count;
5012 }
5013 
5014 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5015 	__ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5016 	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5017 	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5018 	__ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5019 	__ATTR_NULL,
5020 };
5021 
5022 static struct bus_type wq_subsys = {
5023 	.name				= "workqueue",
5024 	.dev_groups			= wq_sysfs_groups,
5025 };
5026 
5027 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5028 		struct device_attribute *attr, char *buf)
5029 {
5030 	int written;
5031 
5032 	mutex_lock(&wq_pool_mutex);
5033 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5034 			    cpumask_pr_args(wq_unbound_cpumask));
5035 	mutex_unlock(&wq_pool_mutex);
5036 
5037 	return written;
5038 }
5039 
5040 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5041 		struct device_attribute *attr, const char *buf, size_t count)
5042 {
5043 	cpumask_var_t cpumask;
5044 	int ret;
5045 
5046 	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5047 		return -ENOMEM;
5048 
5049 	ret = cpumask_parse(buf, cpumask);
5050 	if (!ret)
5051 		ret = workqueue_set_unbound_cpumask(cpumask);
5052 
5053 	free_cpumask_var(cpumask);
5054 	return ret ? ret : count;
5055 }
5056 
5057 static struct device_attribute wq_sysfs_cpumask_attr =
5058 	__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5059 	       wq_unbound_cpumask_store);
5060 
5061 static int __init wq_sysfs_init(void)
5062 {
5063 	int err;
5064 
5065 	err = subsys_virtual_register(&wq_subsys, NULL);
5066 	if (err)
5067 		return err;
5068 
5069 	return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5070 }
5071 core_initcall(wq_sysfs_init);
5072 
5073 static void wq_device_release(struct device *dev)
5074 {
5075 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5076 
5077 	kfree(wq_dev);
5078 }
5079 
5080 /**
5081  * workqueue_sysfs_register - make a workqueue visible in sysfs
5082  * @wq: the workqueue to register
5083  *
5084  * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5085  * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5086  * which is the preferred method.
5087  *
5088  * Workqueue user should use this function directly iff it wants to apply
5089  * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5090  * apply_workqueue_attrs() may race against userland updating the
5091  * attributes.
5092  *
5093  * Return: 0 on success, -errno on failure.
5094  */
5095 int workqueue_sysfs_register(struct workqueue_struct *wq)
5096 {
5097 	struct wq_device *wq_dev;
5098 	int ret;
5099 
5100 	/*
5101 	 * Adjusting max_active or creating new pwqs by applying
5102 	 * attributes breaks ordering guarantee.  Disallow exposing ordered
5103 	 * workqueues.
5104 	 */
5105 	if (WARN_ON(wq->flags & __WQ_ORDERED))
5106 		return -EINVAL;
5107 
5108 	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5109 	if (!wq_dev)
5110 		return -ENOMEM;
5111 
5112 	wq_dev->wq = wq;
5113 	wq_dev->dev.bus = &wq_subsys;
5114 	wq_dev->dev.init_name = wq->name;
5115 	wq_dev->dev.release = wq_device_release;
5116 
5117 	/*
5118 	 * unbound_attrs are created separately.  Suppress uevent until
5119 	 * everything is ready.
5120 	 */
5121 	dev_set_uevent_suppress(&wq_dev->dev, true);
5122 
5123 	ret = device_register(&wq_dev->dev);
5124 	if (ret) {
5125 		kfree(wq_dev);
5126 		wq->wq_dev = NULL;
5127 		return ret;
5128 	}
5129 
5130 	if (wq->flags & WQ_UNBOUND) {
5131 		struct device_attribute *attr;
5132 
5133 		for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5134 			ret = device_create_file(&wq_dev->dev, attr);
5135 			if (ret) {
5136 				device_unregister(&wq_dev->dev);
5137 				wq->wq_dev = NULL;
5138 				return ret;
5139 			}
5140 		}
5141 	}
5142 
5143 	dev_set_uevent_suppress(&wq_dev->dev, false);
5144 	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5145 	return 0;
5146 }
5147 
5148 /**
5149  * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5150  * @wq: the workqueue to unregister
5151  *
5152  * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5153  */
5154 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5155 {
5156 	struct wq_device *wq_dev = wq->wq_dev;
5157 
5158 	if (!wq->wq_dev)
5159 		return;
5160 
5161 	wq->wq_dev = NULL;
5162 	device_unregister(&wq_dev->dev);
5163 }
5164 #else	/* CONFIG_SYSFS */
5165 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)	{ }
5166 #endif	/* CONFIG_SYSFS */
5167 
5168 static void __init wq_numa_init(void)
5169 {
5170 	cpumask_var_t *tbl;
5171 	int node, cpu;
5172 
5173 	if (num_possible_nodes() <= 1)
5174 		return;
5175 
5176 	if (wq_disable_numa) {
5177 		pr_info("workqueue: NUMA affinity support disabled\n");
5178 		return;
5179 	}
5180 
5181 	wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5182 	BUG_ON(!wq_update_unbound_numa_attrs_buf);
5183 
5184 	/*
5185 	 * We want masks of possible CPUs of each node which isn't readily
5186 	 * available.  Build one from cpu_to_node() which should have been
5187 	 * fully initialized by now.
5188 	 */
5189 	tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5190 	BUG_ON(!tbl);
5191 
5192 	for_each_node(node)
5193 		BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5194 				node_online(node) ? node : NUMA_NO_NODE));
5195 
5196 	for_each_possible_cpu(cpu) {
5197 		node = cpu_to_node(cpu);
5198 		if (WARN_ON(node == NUMA_NO_NODE)) {
5199 			pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5200 			/* happens iff arch is bonkers, let's just proceed */
5201 			return;
5202 		}
5203 		cpumask_set_cpu(cpu, tbl[node]);
5204 	}
5205 
5206 	wq_numa_possible_cpumask = tbl;
5207 	wq_numa_enabled = true;
5208 }
5209 
5210 static int __init init_workqueues(void)
5211 {
5212 	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5213 	int i, cpu;
5214 
5215 	WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5216 
5217 	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5218 	cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5219 
5220 	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5221 
5222 	cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5223 	hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5224 
5225 	wq_numa_init();
5226 
5227 	/* initialize CPU pools */
5228 	for_each_possible_cpu(cpu) {
5229 		struct worker_pool *pool;
5230 
5231 		i = 0;
5232 		for_each_cpu_worker_pool(pool, cpu) {
5233 			BUG_ON(init_worker_pool(pool));
5234 			pool->cpu = cpu;
5235 			cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5236 			pool->attrs->nice = std_nice[i++];
5237 			pool->node = cpu_to_node(cpu);
5238 
5239 			/* alloc pool ID */
5240 			mutex_lock(&wq_pool_mutex);
5241 			BUG_ON(worker_pool_assign_id(pool));
5242 			mutex_unlock(&wq_pool_mutex);
5243 		}
5244 	}
5245 
5246 	/* create the initial worker */
5247 	for_each_online_cpu(cpu) {
5248 		struct worker_pool *pool;
5249 
5250 		for_each_cpu_worker_pool(pool, cpu) {
5251 			pool->flags &= ~POOL_DISASSOCIATED;
5252 			BUG_ON(!create_worker(pool));
5253 		}
5254 	}
5255 
5256 	/* create default unbound and ordered wq attrs */
5257 	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5258 		struct workqueue_attrs *attrs;
5259 
5260 		BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5261 		attrs->nice = std_nice[i];
5262 		unbound_std_wq_attrs[i] = attrs;
5263 
5264 		/*
5265 		 * An ordered wq should have only one pwq as ordering is
5266 		 * guaranteed by max_active which is enforced by pwqs.
5267 		 * Turn off NUMA so that dfl_pwq is used for all nodes.
5268 		 */
5269 		BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5270 		attrs->nice = std_nice[i];
5271 		attrs->no_numa = true;
5272 		ordered_wq_attrs[i] = attrs;
5273 	}
5274 
5275 	system_wq = alloc_workqueue("events", 0, 0);
5276 	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5277 	system_long_wq = alloc_workqueue("events_long", 0, 0);
5278 	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5279 					    WQ_UNBOUND_MAX_ACTIVE);
5280 	system_freezable_wq = alloc_workqueue("events_freezable",
5281 					      WQ_FREEZABLE, 0);
5282 	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5283 					      WQ_POWER_EFFICIENT, 0);
5284 	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5285 					      WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5286 					      0);
5287 	BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5288 	       !system_unbound_wq || !system_freezable_wq ||
5289 	       !system_power_efficient_wq ||
5290 	       !system_freezable_power_efficient_wq);
5291 	return 0;
5292 }
5293 early_initcall(init_workqueues);
5294