xref: /openbmc/linux/kernel/workqueue.c (revision e2f1cf25)
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_assert(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_assert(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_assert(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 	dwork->cpu = cpu;
1462 	timer->expires = jiffies + delay;
1463 
1464 	if (unlikely(cpu != WORK_CPU_UNBOUND))
1465 		add_timer_on(timer, cpu);
1466 	else
1467 		add_timer(timer);
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 
1718 	/* prevent userland from meddling with cpumask of workqueue workers */
1719 	worker->task->flags |= PF_NO_SETAFFINITY;
1720 
1721 	/* successful, attach the worker to the pool */
1722 	worker_attach_to_pool(worker, pool);
1723 
1724 	/* start the newly created worker */
1725 	spin_lock_irq(&pool->lock);
1726 	worker->pool->nr_workers++;
1727 	worker_enter_idle(worker);
1728 	wake_up_process(worker->task);
1729 	spin_unlock_irq(&pool->lock);
1730 
1731 	return worker;
1732 
1733 fail:
1734 	if (id >= 0)
1735 		ida_simple_remove(&pool->worker_ida, id);
1736 	kfree(worker);
1737 	return NULL;
1738 }
1739 
1740 /**
1741  * destroy_worker - destroy a workqueue worker
1742  * @worker: worker to be destroyed
1743  *
1744  * Destroy @worker and adjust @pool stats accordingly.  The worker should
1745  * be idle.
1746  *
1747  * CONTEXT:
1748  * spin_lock_irq(pool->lock).
1749  */
1750 static void destroy_worker(struct worker *worker)
1751 {
1752 	struct worker_pool *pool = worker->pool;
1753 
1754 	lockdep_assert_held(&pool->lock);
1755 
1756 	/* sanity check frenzy */
1757 	if (WARN_ON(worker->current_work) ||
1758 	    WARN_ON(!list_empty(&worker->scheduled)) ||
1759 	    WARN_ON(!(worker->flags & WORKER_IDLE)))
1760 		return;
1761 
1762 	pool->nr_workers--;
1763 	pool->nr_idle--;
1764 
1765 	list_del_init(&worker->entry);
1766 	worker->flags |= WORKER_DIE;
1767 	wake_up_process(worker->task);
1768 }
1769 
1770 static void idle_worker_timeout(unsigned long __pool)
1771 {
1772 	struct worker_pool *pool = (void *)__pool;
1773 
1774 	spin_lock_irq(&pool->lock);
1775 
1776 	while (too_many_workers(pool)) {
1777 		struct worker *worker;
1778 		unsigned long expires;
1779 
1780 		/* idle_list is kept in LIFO order, check the last one */
1781 		worker = list_entry(pool->idle_list.prev, struct worker, entry);
1782 		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1783 
1784 		if (time_before(jiffies, expires)) {
1785 			mod_timer(&pool->idle_timer, expires);
1786 			break;
1787 		}
1788 
1789 		destroy_worker(worker);
1790 	}
1791 
1792 	spin_unlock_irq(&pool->lock);
1793 }
1794 
1795 static void send_mayday(struct work_struct *work)
1796 {
1797 	struct pool_workqueue *pwq = get_work_pwq(work);
1798 	struct workqueue_struct *wq = pwq->wq;
1799 
1800 	lockdep_assert_held(&wq_mayday_lock);
1801 
1802 	if (!wq->rescuer)
1803 		return;
1804 
1805 	/* mayday mayday mayday */
1806 	if (list_empty(&pwq->mayday_node)) {
1807 		/*
1808 		 * If @pwq is for an unbound wq, its base ref may be put at
1809 		 * any time due to an attribute change.  Pin @pwq until the
1810 		 * rescuer is done with it.
1811 		 */
1812 		get_pwq(pwq);
1813 		list_add_tail(&pwq->mayday_node, &wq->maydays);
1814 		wake_up_process(wq->rescuer->task);
1815 	}
1816 }
1817 
1818 static void pool_mayday_timeout(unsigned long __pool)
1819 {
1820 	struct worker_pool *pool = (void *)__pool;
1821 	struct work_struct *work;
1822 
1823 	spin_lock_irq(&pool->lock);
1824 	spin_lock(&wq_mayday_lock);		/* for wq->maydays */
1825 
1826 	if (need_to_create_worker(pool)) {
1827 		/*
1828 		 * We've been trying to create a new worker but
1829 		 * haven't been successful.  We might be hitting an
1830 		 * allocation deadlock.  Send distress signals to
1831 		 * rescuers.
1832 		 */
1833 		list_for_each_entry(work, &pool->worklist, entry)
1834 			send_mayday(work);
1835 	}
1836 
1837 	spin_unlock(&wq_mayday_lock);
1838 	spin_unlock_irq(&pool->lock);
1839 
1840 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1841 }
1842 
1843 /**
1844  * maybe_create_worker - create a new worker if necessary
1845  * @pool: pool to create a new worker for
1846  *
1847  * Create a new worker for @pool if necessary.  @pool is guaranteed to
1848  * have at least one idle worker on return from this function.  If
1849  * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1850  * sent to all rescuers with works scheduled on @pool to resolve
1851  * possible allocation deadlock.
1852  *
1853  * On return, need_to_create_worker() is guaranteed to be %false and
1854  * may_start_working() %true.
1855  *
1856  * LOCKING:
1857  * spin_lock_irq(pool->lock) which may be released and regrabbed
1858  * multiple times.  Does GFP_KERNEL allocations.  Called only from
1859  * manager.
1860  */
1861 static void maybe_create_worker(struct worker_pool *pool)
1862 __releases(&pool->lock)
1863 __acquires(&pool->lock)
1864 {
1865 restart:
1866 	spin_unlock_irq(&pool->lock);
1867 
1868 	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1869 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1870 
1871 	while (true) {
1872 		if (create_worker(pool) || !need_to_create_worker(pool))
1873 			break;
1874 
1875 		schedule_timeout_interruptible(CREATE_COOLDOWN);
1876 
1877 		if (!need_to_create_worker(pool))
1878 			break;
1879 	}
1880 
1881 	del_timer_sync(&pool->mayday_timer);
1882 	spin_lock_irq(&pool->lock);
1883 	/*
1884 	 * This is necessary even after a new worker was just successfully
1885 	 * created as @pool->lock was dropped and the new worker might have
1886 	 * already become busy.
1887 	 */
1888 	if (need_to_create_worker(pool))
1889 		goto restart;
1890 }
1891 
1892 /**
1893  * manage_workers - manage worker pool
1894  * @worker: self
1895  *
1896  * Assume the manager role and manage the worker pool @worker belongs
1897  * to.  At any given time, there can be only zero or one manager per
1898  * pool.  The exclusion is handled automatically by this function.
1899  *
1900  * The caller can safely start processing works on false return.  On
1901  * true return, it's guaranteed that need_to_create_worker() is false
1902  * and may_start_working() is true.
1903  *
1904  * CONTEXT:
1905  * spin_lock_irq(pool->lock) which may be released and regrabbed
1906  * multiple times.  Does GFP_KERNEL allocations.
1907  *
1908  * Return:
1909  * %false if the pool doesn't need management and the caller can safely
1910  * start processing works, %true if management function was performed and
1911  * the conditions that the caller verified before calling the function may
1912  * no longer be true.
1913  */
1914 static bool manage_workers(struct worker *worker)
1915 {
1916 	struct worker_pool *pool = worker->pool;
1917 
1918 	/*
1919 	 * Anyone who successfully grabs manager_arb wins the arbitration
1920 	 * and becomes the manager.  mutex_trylock() on pool->manager_arb
1921 	 * failure while holding pool->lock reliably indicates that someone
1922 	 * else is managing the pool and the worker which failed trylock
1923 	 * can proceed to executing work items.  This means that anyone
1924 	 * grabbing manager_arb is responsible for actually performing
1925 	 * manager duties.  If manager_arb is grabbed and released without
1926 	 * actual management, the pool may stall indefinitely.
1927 	 */
1928 	if (!mutex_trylock(&pool->manager_arb))
1929 		return false;
1930 	pool->manager = worker;
1931 
1932 	maybe_create_worker(pool);
1933 
1934 	pool->manager = NULL;
1935 	mutex_unlock(&pool->manager_arb);
1936 	return true;
1937 }
1938 
1939 /**
1940  * process_one_work - process single work
1941  * @worker: self
1942  * @work: work to process
1943  *
1944  * Process @work.  This function contains all the logics necessary to
1945  * process a single work including synchronization against and
1946  * interaction with other workers on the same cpu, queueing and
1947  * flushing.  As long as context requirement is met, any worker can
1948  * call this function to process a work.
1949  *
1950  * CONTEXT:
1951  * spin_lock_irq(pool->lock) which is released and regrabbed.
1952  */
1953 static void process_one_work(struct worker *worker, struct work_struct *work)
1954 __releases(&pool->lock)
1955 __acquires(&pool->lock)
1956 {
1957 	struct pool_workqueue *pwq = get_work_pwq(work);
1958 	struct worker_pool *pool = worker->pool;
1959 	bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
1960 	int work_color;
1961 	struct worker *collision;
1962 #ifdef CONFIG_LOCKDEP
1963 	/*
1964 	 * It is permissible to free the struct work_struct from
1965 	 * inside the function that is called from it, this we need to
1966 	 * take into account for lockdep too.  To avoid bogus "held
1967 	 * lock freed" warnings as well as problems when looking into
1968 	 * work->lockdep_map, make a copy and use that here.
1969 	 */
1970 	struct lockdep_map lockdep_map;
1971 
1972 	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
1973 #endif
1974 	/* ensure we're on the correct CPU */
1975 	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1976 		     raw_smp_processor_id() != pool->cpu);
1977 
1978 	/*
1979 	 * A single work shouldn't be executed concurrently by
1980 	 * multiple workers on a single cpu.  Check whether anyone is
1981 	 * already processing the work.  If so, defer the work to the
1982 	 * currently executing one.
1983 	 */
1984 	collision = find_worker_executing_work(pool, work);
1985 	if (unlikely(collision)) {
1986 		move_linked_works(work, &collision->scheduled, NULL);
1987 		return;
1988 	}
1989 
1990 	/* claim and dequeue */
1991 	debug_work_deactivate(work);
1992 	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
1993 	worker->current_work = work;
1994 	worker->current_func = work->func;
1995 	worker->current_pwq = pwq;
1996 	work_color = get_work_color(work);
1997 
1998 	list_del_init(&work->entry);
1999 
2000 	/*
2001 	 * CPU intensive works don't participate in concurrency management.
2002 	 * They're the scheduler's responsibility.  This takes @worker out
2003 	 * of concurrency management and the next code block will chain
2004 	 * execution of the pending work items.
2005 	 */
2006 	if (unlikely(cpu_intensive))
2007 		worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2008 
2009 	/*
2010 	 * Wake up another worker if necessary.  The condition is always
2011 	 * false for normal per-cpu workers since nr_running would always
2012 	 * be >= 1 at this point.  This is used to chain execution of the
2013 	 * pending work items for WORKER_NOT_RUNNING workers such as the
2014 	 * UNBOUND and CPU_INTENSIVE ones.
2015 	 */
2016 	if (need_more_worker(pool))
2017 		wake_up_worker(pool);
2018 
2019 	/*
2020 	 * Record the last pool and clear PENDING which should be the last
2021 	 * update to @work.  Also, do this inside @pool->lock so that
2022 	 * PENDING and queued state changes happen together while IRQ is
2023 	 * disabled.
2024 	 */
2025 	set_work_pool_and_clear_pending(work, pool->id);
2026 
2027 	spin_unlock_irq(&pool->lock);
2028 
2029 	lock_map_acquire_read(&pwq->wq->lockdep_map);
2030 	lock_map_acquire(&lockdep_map);
2031 	trace_workqueue_execute_start(work);
2032 	worker->current_func(work);
2033 	/*
2034 	 * While we must be careful to not use "work" after this, the trace
2035 	 * point will only record its address.
2036 	 */
2037 	trace_workqueue_execute_end(work);
2038 	lock_map_release(&lockdep_map);
2039 	lock_map_release(&pwq->wq->lockdep_map);
2040 
2041 	if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2042 		pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2043 		       "     last function: %pf\n",
2044 		       current->comm, preempt_count(), task_pid_nr(current),
2045 		       worker->current_func);
2046 		debug_show_held_locks(current);
2047 		dump_stack();
2048 	}
2049 
2050 	/*
2051 	 * The following prevents a kworker from hogging CPU on !PREEMPT
2052 	 * kernels, where a requeueing work item waiting for something to
2053 	 * happen could deadlock with stop_machine as such work item could
2054 	 * indefinitely requeue itself while all other CPUs are trapped in
2055 	 * stop_machine. At the same time, report a quiescent RCU state so
2056 	 * the same condition doesn't freeze RCU.
2057 	 */
2058 	cond_resched_rcu_qs();
2059 
2060 	spin_lock_irq(&pool->lock);
2061 
2062 	/* clear cpu intensive status */
2063 	if (unlikely(cpu_intensive))
2064 		worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2065 
2066 	/* we're done with it, release */
2067 	hash_del(&worker->hentry);
2068 	worker->current_work = NULL;
2069 	worker->current_func = NULL;
2070 	worker->current_pwq = NULL;
2071 	worker->desc_valid = false;
2072 	pwq_dec_nr_in_flight(pwq, work_color);
2073 }
2074 
2075 /**
2076  * process_scheduled_works - process scheduled works
2077  * @worker: self
2078  *
2079  * Process all scheduled works.  Please note that the scheduled list
2080  * may change while processing a work, so this function repeatedly
2081  * fetches a work from the top and executes it.
2082  *
2083  * CONTEXT:
2084  * spin_lock_irq(pool->lock) which may be released and regrabbed
2085  * multiple times.
2086  */
2087 static void process_scheduled_works(struct worker *worker)
2088 {
2089 	while (!list_empty(&worker->scheduled)) {
2090 		struct work_struct *work = list_first_entry(&worker->scheduled,
2091 						struct work_struct, entry);
2092 		process_one_work(worker, work);
2093 	}
2094 }
2095 
2096 /**
2097  * worker_thread - the worker thread function
2098  * @__worker: self
2099  *
2100  * The worker thread function.  All workers belong to a worker_pool -
2101  * either a per-cpu one or dynamic unbound one.  These workers process all
2102  * work items regardless of their specific target workqueue.  The only
2103  * exception is work items which belong to workqueues with a rescuer which
2104  * will be explained in rescuer_thread().
2105  *
2106  * Return: 0
2107  */
2108 static int worker_thread(void *__worker)
2109 {
2110 	struct worker *worker = __worker;
2111 	struct worker_pool *pool = worker->pool;
2112 
2113 	/* tell the scheduler that this is a workqueue worker */
2114 	worker->task->flags |= PF_WQ_WORKER;
2115 woke_up:
2116 	spin_lock_irq(&pool->lock);
2117 
2118 	/* am I supposed to die? */
2119 	if (unlikely(worker->flags & WORKER_DIE)) {
2120 		spin_unlock_irq(&pool->lock);
2121 		WARN_ON_ONCE(!list_empty(&worker->entry));
2122 		worker->task->flags &= ~PF_WQ_WORKER;
2123 
2124 		set_task_comm(worker->task, "kworker/dying");
2125 		ida_simple_remove(&pool->worker_ida, worker->id);
2126 		worker_detach_from_pool(worker, pool);
2127 		kfree(worker);
2128 		return 0;
2129 	}
2130 
2131 	worker_leave_idle(worker);
2132 recheck:
2133 	/* no more worker necessary? */
2134 	if (!need_more_worker(pool))
2135 		goto sleep;
2136 
2137 	/* do we need to manage? */
2138 	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2139 		goto recheck;
2140 
2141 	/*
2142 	 * ->scheduled list can only be filled while a worker is
2143 	 * preparing to process a work or actually processing it.
2144 	 * Make sure nobody diddled with it while I was sleeping.
2145 	 */
2146 	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2147 
2148 	/*
2149 	 * Finish PREP stage.  We're guaranteed to have at least one idle
2150 	 * worker or that someone else has already assumed the manager
2151 	 * role.  This is where @worker starts participating in concurrency
2152 	 * management if applicable and concurrency management is restored
2153 	 * after being rebound.  See rebind_workers() for details.
2154 	 */
2155 	worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2156 
2157 	do {
2158 		struct work_struct *work =
2159 			list_first_entry(&pool->worklist,
2160 					 struct work_struct, entry);
2161 
2162 		if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2163 			/* optimization path, not strictly necessary */
2164 			process_one_work(worker, work);
2165 			if (unlikely(!list_empty(&worker->scheduled)))
2166 				process_scheduled_works(worker);
2167 		} else {
2168 			move_linked_works(work, &worker->scheduled, NULL);
2169 			process_scheduled_works(worker);
2170 		}
2171 	} while (keep_working(pool));
2172 
2173 	worker_set_flags(worker, WORKER_PREP);
2174 sleep:
2175 	/*
2176 	 * pool->lock is held and there's no work to process and no need to
2177 	 * manage, sleep.  Workers are woken up only while holding
2178 	 * pool->lock or from local cpu, so setting the current state
2179 	 * before releasing pool->lock is enough to prevent losing any
2180 	 * event.
2181 	 */
2182 	worker_enter_idle(worker);
2183 	__set_current_state(TASK_INTERRUPTIBLE);
2184 	spin_unlock_irq(&pool->lock);
2185 	schedule();
2186 	goto woke_up;
2187 }
2188 
2189 /**
2190  * rescuer_thread - the rescuer thread function
2191  * @__rescuer: self
2192  *
2193  * Workqueue rescuer thread function.  There's one rescuer for each
2194  * workqueue which has WQ_MEM_RECLAIM set.
2195  *
2196  * Regular work processing on a pool may block trying to create a new
2197  * worker which uses GFP_KERNEL allocation which has slight chance of
2198  * developing into deadlock if some works currently on the same queue
2199  * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2200  * the problem rescuer solves.
2201  *
2202  * When such condition is possible, the pool summons rescuers of all
2203  * workqueues which have works queued on the pool and let them process
2204  * those works so that forward progress can be guaranteed.
2205  *
2206  * This should happen rarely.
2207  *
2208  * Return: 0
2209  */
2210 static int rescuer_thread(void *__rescuer)
2211 {
2212 	struct worker *rescuer = __rescuer;
2213 	struct workqueue_struct *wq = rescuer->rescue_wq;
2214 	struct list_head *scheduled = &rescuer->scheduled;
2215 	bool should_stop;
2216 
2217 	set_user_nice(current, RESCUER_NICE_LEVEL);
2218 
2219 	/*
2220 	 * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
2221 	 * doesn't participate in concurrency management.
2222 	 */
2223 	rescuer->task->flags |= PF_WQ_WORKER;
2224 repeat:
2225 	set_current_state(TASK_INTERRUPTIBLE);
2226 
2227 	/*
2228 	 * By the time the rescuer is requested to stop, the workqueue
2229 	 * shouldn't have any work pending, but @wq->maydays may still have
2230 	 * pwq(s) queued.  This can happen by non-rescuer workers consuming
2231 	 * all the work items before the rescuer got to them.  Go through
2232 	 * @wq->maydays processing before acting on should_stop so that the
2233 	 * list is always empty on exit.
2234 	 */
2235 	should_stop = kthread_should_stop();
2236 
2237 	/* see whether any pwq is asking for help */
2238 	spin_lock_irq(&wq_mayday_lock);
2239 
2240 	while (!list_empty(&wq->maydays)) {
2241 		struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2242 					struct pool_workqueue, mayday_node);
2243 		struct worker_pool *pool = pwq->pool;
2244 		struct work_struct *work, *n;
2245 
2246 		__set_current_state(TASK_RUNNING);
2247 		list_del_init(&pwq->mayday_node);
2248 
2249 		spin_unlock_irq(&wq_mayday_lock);
2250 
2251 		worker_attach_to_pool(rescuer, pool);
2252 
2253 		spin_lock_irq(&pool->lock);
2254 		rescuer->pool = pool;
2255 
2256 		/*
2257 		 * Slurp in all works issued via this workqueue and
2258 		 * process'em.
2259 		 */
2260 		WARN_ON_ONCE(!list_empty(scheduled));
2261 		list_for_each_entry_safe(work, n, &pool->worklist, entry)
2262 			if (get_work_pwq(work) == pwq)
2263 				move_linked_works(work, scheduled, &n);
2264 
2265 		if (!list_empty(scheduled)) {
2266 			process_scheduled_works(rescuer);
2267 
2268 			/*
2269 			 * The above execution of rescued work items could
2270 			 * have created more to rescue through
2271 			 * pwq_activate_first_delayed() or chained
2272 			 * queueing.  Let's put @pwq back on mayday list so
2273 			 * that such back-to-back work items, which may be
2274 			 * being used to relieve memory pressure, don't
2275 			 * incur MAYDAY_INTERVAL delay inbetween.
2276 			 */
2277 			if (need_to_create_worker(pool)) {
2278 				spin_lock(&wq_mayday_lock);
2279 				get_pwq(pwq);
2280 				list_move_tail(&pwq->mayday_node, &wq->maydays);
2281 				spin_unlock(&wq_mayday_lock);
2282 			}
2283 		}
2284 
2285 		/*
2286 		 * Put the reference grabbed by send_mayday().  @pool won't
2287 		 * go away while we're still attached to it.
2288 		 */
2289 		put_pwq(pwq);
2290 
2291 		/*
2292 		 * Leave this pool.  If need_more_worker() is %true, notify a
2293 		 * regular worker; otherwise, we end up with 0 concurrency
2294 		 * and stalling the execution.
2295 		 */
2296 		if (need_more_worker(pool))
2297 			wake_up_worker(pool);
2298 
2299 		rescuer->pool = NULL;
2300 		spin_unlock_irq(&pool->lock);
2301 
2302 		worker_detach_from_pool(rescuer, pool);
2303 
2304 		spin_lock_irq(&wq_mayday_lock);
2305 	}
2306 
2307 	spin_unlock_irq(&wq_mayday_lock);
2308 
2309 	if (should_stop) {
2310 		__set_current_state(TASK_RUNNING);
2311 		rescuer->task->flags &= ~PF_WQ_WORKER;
2312 		return 0;
2313 	}
2314 
2315 	/* rescuers should never participate in concurrency management */
2316 	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2317 	schedule();
2318 	goto repeat;
2319 }
2320 
2321 struct wq_barrier {
2322 	struct work_struct	work;
2323 	struct completion	done;
2324 	struct task_struct	*task;	/* purely informational */
2325 };
2326 
2327 static void wq_barrier_func(struct work_struct *work)
2328 {
2329 	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2330 	complete(&barr->done);
2331 }
2332 
2333 /**
2334  * insert_wq_barrier - insert a barrier work
2335  * @pwq: pwq to insert barrier into
2336  * @barr: wq_barrier to insert
2337  * @target: target work to attach @barr to
2338  * @worker: worker currently executing @target, NULL if @target is not executing
2339  *
2340  * @barr is linked to @target such that @barr is completed only after
2341  * @target finishes execution.  Please note that the ordering
2342  * guarantee is observed only with respect to @target and on the local
2343  * cpu.
2344  *
2345  * Currently, a queued barrier can't be canceled.  This is because
2346  * try_to_grab_pending() can't determine whether the work to be
2347  * grabbed is at the head of the queue and thus can't clear LINKED
2348  * flag of the previous work while there must be a valid next work
2349  * after a work with LINKED flag set.
2350  *
2351  * Note that when @worker is non-NULL, @target may be modified
2352  * underneath us, so we can't reliably determine pwq from @target.
2353  *
2354  * CONTEXT:
2355  * spin_lock_irq(pool->lock).
2356  */
2357 static void insert_wq_barrier(struct pool_workqueue *pwq,
2358 			      struct wq_barrier *barr,
2359 			      struct work_struct *target, struct worker *worker)
2360 {
2361 	struct list_head *head;
2362 	unsigned int linked = 0;
2363 
2364 	/*
2365 	 * debugobject calls are safe here even with pool->lock locked
2366 	 * as we know for sure that this will not trigger any of the
2367 	 * checks and call back into the fixup functions where we
2368 	 * might deadlock.
2369 	 */
2370 	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2371 	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2372 	init_completion(&barr->done);
2373 	barr->task = current;
2374 
2375 	/*
2376 	 * If @target is currently being executed, schedule the
2377 	 * barrier to the worker; otherwise, put it after @target.
2378 	 */
2379 	if (worker)
2380 		head = worker->scheduled.next;
2381 	else {
2382 		unsigned long *bits = work_data_bits(target);
2383 
2384 		head = target->entry.next;
2385 		/* there can already be other linked works, inherit and set */
2386 		linked = *bits & WORK_STRUCT_LINKED;
2387 		__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2388 	}
2389 
2390 	debug_work_activate(&barr->work);
2391 	insert_work(pwq, &barr->work, head,
2392 		    work_color_to_flags(WORK_NO_COLOR) | linked);
2393 }
2394 
2395 /**
2396  * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2397  * @wq: workqueue being flushed
2398  * @flush_color: new flush color, < 0 for no-op
2399  * @work_color: new work color, < 0 for no-op
2400  *
2401  * Prepare pwqs for workqueue flushing.
2402  *
2403  * If @flush_color is non-negative, flush_color on all pwqs should be
2404  * -1.  If no pwq has in-flight commands at the specified color, all
2405  * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
2406  * has in flight commands, its pwq->flush_color is set to
2407  * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2408  * wakeup logic is armed and %true is returned.
2409  *
2410  * The caller should have initialized @wq->first_flusher prior to
2411  * calling this function with non-negative @flush_color.  If
2412  * @flush_color is negative, no flush color update is done and %false
2413  * is returned.
2414  *
2415  * If @work_color is non-negative, all pwqs should have the same
2416  * work_color which is previous to @work_color and all will be
2417  * advanced to @work_color.
2418  *
2419  * CONTEXT:
2420  * mutex_lock(wq->mutex).
2421  *
2422  * Return:
2423  * %true if @flush_color >= 0 and there's something to flush.  %false
2424  * otherwise.
2425  */
2426 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2427 				      int flush_color, int work_color)
2428 {
2429 	bool wait = false;
2430 	struct pool_workqueue *pwq;
2431 
2432 	if (flush_color >= 0) {
2433 		WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2434 		atomic_set(&wq->nr_pwqs_to_flush, 1);
2435 	}
2436 
2437 	for_each_pwq(pwq, wq) {
2438 		struct worker_pool *pool = pwq->pool;
2439 
2440 		spin_lock_irq(&pool->lock);
2441 
2442 		if (flush_color >= 0) {
2443 			WARN_ON_ONCE(pwq->flush_color != -1);
2444 
2445 			if (pwq->nr_in_flight[flush_color]) {
2446 				pwq->flush_color = flush_color;
2447 				atomic_inc(&wq->nr_pwqs_to_flush);
2448 				wait = true;
2449 			}
2450 		}
2451 
2452 		if (work_color >= 0) {
2453 			WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2454 			pwq->work_color = work_color;
2455 		}
2456 
2457 		spin_unlock_irq(&pool->lock);
2458 	}
2459 
2460 	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2461 		complete(&wq->first_flusher->done);
2462 
2463 	return wait;
2464 }
2465 
2466 /**
2467  * flush_workqueue - ensure that any scheduled work has run to completion.
2468  * @wq: workqueue to flush
2469  *
2470  * This function sleeps until all work items which were queued on entry
2471  * have finished execution, but it is not livelocked by new incoming ones.
2472  */
2473 void flush_workqueue(struct workqueue_struct *wq)
2474 {
2475 	struct wq_flusher this_flusher = {
2476 		.list = LIST_HEAD_INIT(this_flusher.list),
2477 		.flush_color = -1,
2478 		.done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2479 	};
2480 	int next_color;
2481 
2482 	lock_map_acquire(&wq->lockdep_map);
2483 	lock_map_release(&wq->lockdep_map);
2484 
2485 	mutex_lock(&wq->mutex);
2486 
2487 	/*
2488 	 * Start-to-wait phase
2489 	 */
2490 	next_color = work_next_color(wq->work_color);
2491 
2492 	if (next_color != wq->flush_color) {
2493 		/*
2494 		 * Color space is not full.  The current work_color
2495 		 * becomes our flush_color and work_color is advanced
2496 		 * by one.
2497 		 */
2498 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2499 		this_flusher.flush_color = wq->work_color;
2500 		wq->work_color = next_color;
2501 
2502 		if (!wq->first_flusher) {
2503 			/* no flush in progress, become the first flusher */
2504 			WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2505 
2506 			wq->first_flusher = &this_flusher;
2507 
2508 			if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2509 						       wq->work_color)) {
2510 				/* nothing to flush, done */
2511 				wq->flush_color = next_color;
2512 				wq->first_flusher = NULL;
2513 				goto out_unlock;
2514 			}
2515 		} else {
2516 			/* wait in queue */
2517 			WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2518 			list_add_tail(&this_flusher.list, &wq->flusher_queue);
2519 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2520 		}
2521 	} else {
2522 		/*
2523 		 * Oops, color space is full, wait on overflow queue.
2524 		 * The next flush completion will assign us
2525 		 * flush_color and transfer to flusher_queue.
2526 		 */
2527 		list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2528 	}
2529 
2530 	mutex_unlock(&wq->mutex);
2531 
2532 	wait_for_completion(&this_flusher.done);
2533 
2534 	/*
2535 	 * Wake-up-and-cascade phase
2536 	 *
2537 	 * First flushers are responsible for cascading flushes and
2538 	 * handling overflow.  Non-first flushers can simply return.
2539 	 */
2540 	if (wq->first_flusher != &this_flusher)
2541 		return;
2542 
2543 	mutex_lock(&wq->mutex);
2544 
2545 	/* we might have raced, check again with mutex held */
2546 	if (wq->first_flusher != &this_flusher)
2547 		goto out_unlock;
2548 
2549 	wq->first_flusher = NULL;
2550 
2551 	WARN_ON_ONCE(!list_empty(&this_flusher.list));
2552 	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2553 
2554 	while (true) {
2555 		struct wq_flusher *next, *tmp;
2556 
2557 		/* complete all the flushers sharing the current flush color */
2558 		list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2559 			if (next->flush_color != wq->flush_color)
2560 				break;
2561 			list_del_init(&next->list);
2562 			complete(&next->done);
2563 		}
2564 
2565 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2566 			     wq->flush_color != work_next_color(wq->work_color));
2567 
2568 		/* this flush_color is finished, advance by one */
2569 		wq->flush_color = work_next_color(wq->flush_color);
2570 
2571 		/* one color has been freed, handle overflow queue */
2572 		if (!list_empty(&wq->flusher_overflow)) {
2573 			/*
2574 			 * Assign the same color to all overflowed
2575 			 * flushers, advance work_color and append to
2576 			 * flusher_queue.  This is the start-to-wait
2577 			 * phase for these overflowed flushers.
2578 			 */
2579 			list_for_each_entry(tmp, &wq->flusher_overflow, list)
2580 				tmp->flush_color = wq->work_color;
2581 
2582 			wq->work_color = work_next_color(wq->work_color);
2583 
2584 			list_splice_tail_init(&wq->flusher_overflow,
2585 					      &wq->flusher_queue);
2586 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2587 		}
2588 
2589 		if (list_empty(&wq->flusher_queue)) {
2590 			WARN_ON_ONCE(wq->flush_color != wq->work_color);
2591 			break;
2592 		}
2593 
2594 		/*
2595 		 * Need to flush more colors.  Make the next flusher
2596 		 * the new first flusher and arm pwqs.
2597 		 */
2598 		WARN_ON_ONCE(wq->flush_color == wq->work_color);
2599 		WARN_ON_ONCE(wq->flush_color != next->flush_color);
2600 
2601 		list_del_init(&next->list);
2602 		wq->first_flusher = next;
2603 
2604 		if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2605 			break;
2606 
2607 		/*
2608 		 * Meh... this color is already done, clear first
2609 		 * flusher and repeat cascading.
2610 		 */
2611 		wq->first_flusher = NULL;
2612 	}
2613 
2614 out_unlock:
2615 	mutex_unlock(&wq->mutex);
2616 }
2617 EXPORT_SYMBOL_GPL(flush_workqueue);
2618 
2619 /**
2620  * drain_workqueue - drain a workqueue
2621  * @wq: workqueue to drain
2622  *
2623  * Wait until the workqueue becomes empty.  While draining is in progress,
2624  * only chain queueing is allowed.  IOW, only currently pending or running
2625  * work items on @wq can queue further work items on it.  @wq is flushed
2626  * repeatedly until it becomes empty.  The number of flushing is determined
2627  * by the depth of chaining and should be relatively short.  Whine if it
2628  * takes too long.
2629  */
2630 void drain_workqueue(struct workqueue_struct *wq)
2631 {
2632 	unsigned int flush_cnt = 0;
2633 	struct pool_workqueue *pwq;
2634 
2635 	/*
2636 	 * __queue_work() needs to test whether there are drainers, is much
2637 	 * hotter than drain_workqueue() and already looks at @wq->flags.
2638 	 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2639 	 */
2640 	mutex_lock(&wq->mutex);
2641 	if (!wq->nr_drainers++)
2642 		wq->flags |= __WQ_DRAINING;
2643 	mutex_unlock(&wq->mutex);
2644 reflush:
2645 	flush_workqueue(wq);
2646 
2647 	mutex_lock(&wq->mutex);
2648 
2649 	for_each_pwq(pwq, wq) {
2650 		bool drained;
2651 
2652 		spin_lock_irq(&pwq->pool->lock);
2653 		drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2654 		spin_unlock_irq(&pwq->pool->lock);
2655 
2656 		if (drained)
2657 			continue;
2658 
2659 		if (++flush_cnt == 10 ||
2660 		    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2661 			pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2662 				wq->name, flush_cnt);
2663 
2664 		mutex_unlock(&wq->mutex);
2665 		goto reflush;
2666 	}
2667 
2668 	if (!--wq->nr_drainers)
2669 		wq->flags &= ~__WQ_DRAINING;
2670 	mutex_unlock(&wq->mutex);
2671 }
2672 EXPORT_SYMBOL_GPL(drain_workqueue);
2673 
2674 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2675 {
2676 	struct worker *worker = NULL;
2677 	struct worker_pool *pool;
2678 	struct pool_workqueue *pwq;
2679 
2680 	might_sleep();
2681 
2682 	local_irq_disable();
2683 	pool = get_work_pool(work);
2684 	if (!pool) {
2685 		local_irq_enable();
2686 		return false;
2687 	}
2688 
2689 	spin_lock(&pool->lock);
2690 	/* see the comment in try_to_grab_pending() with the same code */
2691 	pwq = get_work_pwq(work);
2692 	if (pwq) {
2693 		if (unlikely(pwq->pool != pool))
2694 			goto already_gone;
2695 	} else {
2696 		worker = find_worker_executing_work(pool, work);
2697 		if (!worker)
2698 			goto already_gone;
2699 		pwq = worker->current_pwq;
2700 	}
2701 
2702 	insert_wq_barrier(pwq, barr, work, worker);
2703 	spin_unlock_irq(&pool->lock);
2704 
2705 	/*
2706 	 * If @max_active is 1 or rescuer is in use, flushing another work
2707 	 * item on the same workqueue may lead to deadlock.  Make sure the
2708 	 * flusher is not running on the same workqueue by verifying write
2709 	 * access.
2710 	 */
2711 	if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2712 		lock_map_acquire(&pwq->wq->lockdep_map);
2713 	else
2714 		lock_map_acquire_read(&pwq->wq->lockdep_map);
2715 	lock_map_release(&pwq->wq->lockdep_map);
2716 
2717 	return true;
2718 already_gone:
2719 	spin_unlock_irq(&pool->lock);
2720 	return false;
2721 }
2722 
2723 /**
2724  * flush_work - wait for a work to finish executing the last queueing instance
2725  * @work: the work to flush
2726  *
2727  * Wait until @work has finished execution.  @work is guaranteed to be idle
2728  * on return if it hasn't been requeued since flush started.
2729  *
2730  * Return:
2731  * %true if flush_work() waited for the work to finish execution,
2732  * %false if it was already idle.
2733  */
2734 bool flush_work(struct work_struct *work)
2735 {
2736 	struct wq_barrier barr;
2737 
2738 	lock_map_acquire(&work->lockdep_map);
2739 	lock_map_release(&work->lockdep_map);
2740 
2741 	if (start_flush_work(work, &barr)) {
2742 		wait_for_completion(&barr.done);
2743 		destroy_work_on_stack(&barr.work);
2744 		return true;
2745 	} else {
2746 		return false;
2747 	}
2748 }
2749 EXPORT_SYMBOL_GPL(flush_work);
2750 
2751 struct cwt_wait {
2752 	wait_queue_t		wait;
2753 	struct work_struct	*work;
2754 };
2755 
2756 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2757 {
2758 	struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2759 
2760 	if (cwait->work != key)
2761 		return 0;
2762 	return autoremove_wake_function(wait, mode, sync, key);
2763 }
2764 
2765 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2766 {
2767 	static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2768 	unsigned long flags;
2769 	int ret;
2770 
2771 	do {
2772 		ret = try_to_grab_pending(work, is_dwork, &flags);
2773 		/*
2774 		 * If someone else is already canceling, wait for it to
2775 		 * finish.  flush_work() doesn't work for PREEMPT_NONE
2776 		 * because we may get scheduled between @work's completion
2777 		 * and the other canceling task resuming and clearing
2778 		 * CANCELING - flush_work() will return false immediately
2779 		 * as @work is no longer busy, try_to_grab_pending() will
2780 		 * return -ENOENT as @work is still being canceled and the
2781 		 * other canceling task won't be able to clear CANCELING as
2782 		 * we're hogging the CPU.
2783 		 *
2784 		 * Let's wait for completion using a waitqueue.  As this
2785 		 * may lead to the thundering herd problem, use a custom
2786 		 * wake function which matches @work along with exclusive
2787 		 * wait and wakeup.
2788 		 */
2789 		if (unlikely(ret == -ENOENT)) {
2790 			struct cwt_wait cwait;
2791 
2792 			init_wait(&cwait.wait);
2793 			cwait.wait.func = cwt_wakefn;
2794 			cwait.work = work;
2795 
2796 			prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2797 						  TASK_UNINTERRUPTIBLE);
2798 			if (work_is_canceling(work))
2799 				schedule();
2800 			finish_wait(&cancel_waitq, &cwait.wait);
2801 		}
2802 	} while (unlikely(ret < 0));
2803 
2804 	/* tell other tasks trying to grab @work to back off */
2805 	mark_work_canceling(work);
2806 	local_irq_restore(flags);
2807 
2808 	flush_work(work);
2809 	clear_work_data(work);
2810 
2811 	/*
2812 	 * Paired with prepare_to_wait() above so that either
2813 	 * waitqueue_active() is visible here or !work_is_canceling() is
2814 	 * visible there.
2815 	 */
2816 	smp_mb();
2817 	if (waitqueue_active(&cancel_waitq))
2818 		__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2819 
2820 	return ret;
2821 }
2822 
2823 /**
2824  * cancel_work_sync - cancel a work and wait for it to finish
2825  * @work: the work to cancel
2826  *
2827  * Cancel @work and wait for its execution to finish.  This function
2828  * can be used even if the work re-queues itself or migrates to
2829  * another workqueue.  On return from this function, @work is
2830  * guaranteed to be not pending or executing on any CPU.
2831  *
2832  * cancel_work_sync(&delayed_work->work) must not be used for
2833  * delayed_work's.  Use cancel_delayed_work_sync() instead.
2834  *
2835  * The caller must ensure that the workqueue on which @work was last
2836  * queued can't be destroyed before this function returns.
2837  *
2838  * Return:
2839  * %true if @work was pending, %false otherwise.
2840  */
2841 bool cancel_work_sync(struct work_struct *work)
2842 {
2843 	return __cancel_work_timer(work, false);
2844 }
2845 EXPORT_SYMBOL_GPL(cancel_work_sync);
2846 
2847 /**
2848  * flush_delayed_work - wait for a dwork to finish executing the last queueing
2849  * @dwork: the delayed work to flush
2850  *
2851  * Delayed timer is cancelled and the pending work is queued for
2852  * immediate execution.  Like flush_work(), this function only
2853  * considers the last queueing instance of @dwork.
2854  *
2855  * Return:
2856  * %true if flush_work() waited for the work to finish execution,
2857  * %false if it was already idle.
2858  */
2859 bool flush_delayed_work(struct delayed_work *dwork)
2860 {
2861 	local_irq_disable();
2862 	if (del_timer_sync(&dwork->timer))
2863 		__queue_work(dwork->cpu, dwork->wq, &dwork->work);
2864 	local_irq_enable();
2865 	return flush_work(&dwork->work);
2866 }
2867 EXPORT_SYMBOL(flush_delayed_work);
2868 
2869 /**
2870  * cancel_delayed_work - cancel a delayed work
2871  * @dwork: delayed_work to cancel
2872  *
2873  * Kill off a pending delayed_work.
2874  *
2875  * Return: %true if @dwork was pending and canceled; %false if it wasn't
2876  * pending.
2877  *
2878  * Note:
2879  * The work callback function may still be running on return, unless
2880  * it returns %true and the work doesn't re-arm itself.  Explicitly flush or
2881  * use cancel_delayed_work_sync() to wait on it.
2882  *
2883  * This function is safe to call from any context including IRQ handler.
2884  */
2885 bool cancel_delayed_work(struct delayed_work *dwork)
2886 {
2887 	unsigned long flags;
2888 	int ret;
2889 
2890 	do {
2891 		ret = try_to_grab_pending(&dwork->work, true, &flags);
2892 	} while (unlikely(ret == -EAGAIN));
2893 
2894 	if (unlikely(ret < 0))
2895 		return false;
2896 
2897 	set_work_pool_and_clear_pending(&dwork->work,
2898 					get_work_pool_id(&dwork->work));
2899 	local_irq_restore(flags);
2900 	return ret;
2901 }
2902 EXPORT_SYMBOL(cancel_delayed_work);
2903 
2904 /**
2905  * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2906  * @dwork: the delayed work cancel
2907  *
2908  * This is cancel_work_sync() for delayed works.
2909  *
2910  * Return:
2911  * %true if @dwork was pending, %false otherwise.
2912  */
2913 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2914 {
2915 	return __cancel_work_timer(&dwork->work, true);
2916 }
2917 EXPORT_SYMBOL(cancel_delayed_work_sync);
2918 
2919 /**
2920  * schedule_on_each_cpu - execute a function synchronously on each online CPU
2921  * @func: the function to call
2922  *
2923  * schedule_on_each_cpu() executes @func on each online CPU using the
2924  * system workqueue and blocks until all CPUs have completed.
2925  * schedule_on_each_cpu() is very slow.
2926  *
2927  * Return:
2928  * 0 on success, -errno on failure.
2929  */
2930 int schedule_on_each_cpu(work_func_t func)
2931 {
2932 	int cpu;
2933 	struct work_struct __percpu *works;
2934 
2935 	works = alloc_percpu(struct work_struct);
2936 	if (!works)
2937 		return -ENOMEM;
2938 
2939 	get_online_cpus();
2940 
2941 	for_each_online_cpu(cpu) {
2942 		struct work_struct *work = per_cpu_ptr(works, cpu);
2943 
2944 		INIT_WORK(work, func);
2945 		schedule_work_on(cpu, work);
2946 	}
2947 
2948 	for_each_online_cpu(cpu)
2949 		flush_work(per_cpu_ptr(works, cpu));
2950 
2951 	put_online_cpus();
2952 	free_percpu(works);
2953 	return 0;
2954 }
2955 
2956 /**
2957  * execute_in_process_context - reliably execute the routine with user context
2958  * @fn:		the function to execute
2959  * @ew:		guaranteed storage for the execute work structure (must
2960  *		be available when the work executes)
2961  *
2962  * Executes the function immediately if process context is available,
2963  * otherwise schedules the function for delayed execution.
2964  *
2965  * Return:	0 - function was executed
2966  *		1 - function was scheduled for execution
2967  */
2968 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2969 {
2970 	if (!in_interrupt()) {
2971 		fn(&ew->work);
2972 		return 0;
2973 	}
2974 
2975 	INIT_WORK(&ew->work, fn);
2976 	schedule_work(&ew->work);
2977 
2978 	return 1;
2979 }
2980 EXPORT_SYMBOL_GPL(execute_in_process_context);
2981 
2982 /**
2983  * free_workqueue_attrs - free a workqueue_attrs
2984  * @attrs: workqueue_attrs to free
2985  *
2986  * Undo alloc_workqueue_attrs().
2987  */
2988 void free_workqueue_attrs(struct workqueue_attrs *attrs)
2989 {
2990 	if (attrs) {
2991 		free_cpumask_var(attrs->cpumask);
2992 		kfree(attrs);
2993 	}
2994 }
2995 
2996 /**
2997  * alloc_workqueue_attrs - allocate a workqueue_attrs
2998  * @gfp_mask: allocation mask to use
2999  *
3000  * Allocate a new workqueue_attrs, initialize with default settings and
3001  * return it.
3002  *
3003  * Return: The allocated new workqueue_attr on success. %NULL on failure.
3004  */
3005 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3006 {
3007 	struct workqueue_attrs *attrs;
3008 
3009 	attrs = kzalloc(sizeof(*attrs), gfp_mask);
3010 	if (!attrs)
3011 		goto fail;
3012 	if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3013 		goto fail;
3014 
3015 	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3016 	return attrs;
3017 fail:
3018 	free_workqueue_attrs(attrs);
3019 	return NULL;
3020 }
3021 
3022 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3023 				 const struct workqueue_attrs *from)
3024 {
3025 	to->nice = from->nice;
3026 	cpumask_copy(to->cpumask, from->cpumask);
3027 	/*
3028 	 * Unlike hash and equality test, this function doesn't ignore
3029 	 * ->no_numa as it is used for both pool and wq attrs.  Instead,
3030 	 * get_unbound_pool() explicitly clears ->no_numa after copying.
3031 	 */
3032 	to->no_numa = from->no_numa;
3033 }
3034 
3035 /* hash value of the content of @attr */
3036 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3037 {
3038 	u32 hash = 0;
3039 
3040 	hash = jhash_1word(attrs->nice, hash);
3041 	hash = jhash(cpumask_bits(attrs->cpumask),
3042 		     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3043 	return hash;
3044 }
3045 
3046 /* content equality test */
3047 static bool wqattrs_equal(const struct workqueue_attrs *a,
3048 			  const struct workqueue_attrs *b)
3049 {
3050 	if (a->nice != b->nice)
3051 		return false;
3052 	if (!cpumask_equal(a->cpumask, b->cpumask))
3053 		return false;
3054 	return true;
3055 }
3056 
3057 /**
3058  * init_worker_pool - initialize a newly zalloc'd worker_pool
3059  * @pool: worker_pool to initialize
3060  *
3061  * Initialize a newly zalloc'd @pool.  It also allocates @pool->attrs.
3062  *
3063  * Return: 0 on success, -errno on failure.  Even on failure, all fields
3064  * inside @pool proper are initialized and put_unbound_pool() can be called
3065  * on @pool safely to release it.
3066  */
3067 static int init_worker_pool(struct worker_pool *pool)
3068 {
3069 	spin_lock_init(&pool->lock);
3070 	pool->id = -1;
3071 	pool->cpu = -1;
3072 	pool->node = NUMA_NO_NODE;
3073 	pool->flags |= POOL_DISASSOCIATED;
3074 	INIT_LIST_HEAD(&pool->worklist);
3075 	INIT_LIST_HEAD(&pool->idle_list);
3076 	hash_init(pool->busy_hash);
3077 
3078 	init_timer_deferrable(&pool->idle_timer);
3079 	pool->idle_timer.function = idle_worker_timeout;
3080 	pool->idle_timer.data = (unsigned long)pool;
3081 
3082 	setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3083 		    (unsigned long)pool);
3084 
3085 	mutex_init(&pool->manager_arb);
3086 	mutex_init(&pool->attach_mutex);
3087 	INIT_LIST_HEAD(&pool->workers);
3088 
3089 	ida_init(&pool->worker_ida);
3090 	INIT_HLIST_NODE(&pool->hash_node);
3091 	pool->refcnt = 1;
3092 
3093 	/* shouldn't fail above this point */
3094 	pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3095 	if (!pool->attrs)
3096 		return -ENOMEM;
3097 	return 0;
3098 }
3099 
3100 static void rcu_free_wq(struct rcu_head *rcu)
3101 {
3102 	struct workqueue_struct *wq =
3103 		container_of(rcu, struct workqueue_struct, rcu);
3104 
3105 	if (!(wq->flags & WQ_UNBOUND))
3106 		free_percpu(wq->cpu_pwqs);
3107 	else
3108 		free_workqueue_attrs(wq->unbound_attrs);
3109 
3110 	kfree(wq->rescuer);
3111 	kfree(wq);
3112 }
3113 
3114 static void rcu_free_pool(struct rcu_head *rcu)
3115 {
3116 	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3117 
3118 	ida_destroy(&pool->worker_ida);
3119 	free_workqueue_attrs(pool->attrs);
3120 	kfree(pool);
3121 }
3122 
3123 /**
3124  * put_unbound_pool - put a worker_pool
3125  * @pool: worker_pool to put
3126  *
3127  * Put @pool.  If its refcnt reaches zero, it gets destroyed in sched-RCU
3128  * safe manner.  get_unbound_pool() calls this function on its failure path
3129  * and this function should be able to release pools which went through,
3130  * successfully or not, init_worker_pool().
3131  *
3132  * Should be called with wq_pool_mutex held.
3133  */
3134 static void put_unbound_pool(struct worker_pool *pool)
3135 {
3136 	DECLARE_COMPLETION_ONSTACK(detach_completion);
3137 	struct worker *worker;
3138 
3139 	lockdep_assert_held(&wq_pool_mutex);
3140 
3141 	if (--pool->refcnt)
3142 		return;
3143 
3144 	/* sanity checks */
3145 	if (WARN_ON(!(pool->cpu < 0)) ||
3146 	    WARN_ON(!list_empty(&pool->worklist)))
3147 		return;
3148 
3149 	/* release id and unhash */
3150 	if (pool->id >= 0)
3151 		idr_remove(&worker_pool_idr, pool->id);
3152 	hash_del(&pool->hash_node);
3153 
3154 	/*
3155 	 * Become the manager and destroy all workers.  Grabbing
3156 	 * manager_arb prevents @pool's workers from blocking on
3157 	 * attach_mutex.
3158 	 */
3159 	mutex_lock(&pool->manager_arb);
3160 
3161 	spin_lock_irq(&pool->lock);
3162 	while ((worker = first_idle_worker(pool)))
3163 		destroy_worker(worker);
3164 	WARN_ON(pool->nr_workers || pool->nr_idle);
3165 	spin_unlock_irq(&pool->lock);
3166 
3167 	mutex_lock(&pool->attach_mutex);
3168 	if (!list_empty(&pool->workers))
3169 		pool->detach_completion = &detach_completion;
3170 	mutex_unlock(&pool->attach_mutex);
3171 
3172 	if (pool->detach_completion)
3173 		wait_for_completion(pool->detach_completion);
3174 
3175 	mutex_unlock(&pool->manager_arb);
3176 
3177 	/* shut down the timers */
3178 	del_timer_sync(&pool->idle_timer);
3179 	del_timer_sync(&pool->mayday_timer);
3180 
3181 	/* sched-RCU protected to allow dereferences from get_work_pool() */
3182 	call_rcu_sched(&pool->rcu, rcu_free_pool);
3183 }
3184 
3185 /**
3186  * get_unbound_pool - get a worker_pool with the specified attributes
3187  * @attrs: the attributes of the worker_pool to get
3188  *
3189  * Obtain a worker_pool which has the same attributes as @attrs, bump the
3190  * reference count and return it.  If there already is a matching
3191  * worker_pool, it will be used; otherwise, this function attempts to
3192  * create a new one.
3193  *
3194  * Should be called with wq_pool_mutex held.
3195  *
3196  * Return: On success, a worker_pool with the same attributes as @attrs.
3197  * On failure, %NULL.
3198  */
3199 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3200 {
3201 	u32 hash = wqattrs_hash(attrs);
3202 	struct worker_pool *pool;
3203 	int node;
3204 
3205 	lockdep_assert_held(&wq_pool_mutex);
3206 
3207 	/* do we already have a matching pool? */
3208 	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3209 		if (wqattrs_equal(pool->attrs, attrs)) {
3210 			pool->refcnt++;
3211 			return pool;
3212 		}
3213 	}
3214 
3215 	/* nope, create a new one */
3216 	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3217 	if (!pool || init_worker_pool(pool) < 0)
3218 		goto fail;
3219 
3220 	lockdep_set_subclass(&pool->lock, 1);	/* see put_pwq() */
3221 	copy_workqueue_attrs(pool->attrs, attrs);
3222 
3223 	/*
3224 	 * no_numa isn't a worker_pool attribute, always clear it.  See
3225 	 * 'struct workqueue_attrs' comments for detail.
3226 	 */
3227 	pool->attrs->no_numa = false;
3228 
3229 	/* if cpumask is contained inside a NUMA node, we belong to that node */
3230 	if (wq_numa_enabled) {
3231 		for_each_node(node) {
3232 			if (cpumask_subset(pool->attrs->cpumask,
3233 					   wq_numa_possible_cpumask[node])) {
3234 				pool->node = node;
3235 				break;
3236 			}
3237 		}
3238 	}
3239 
3240 	if (worker_pool_assign_id(pool) < 0)
3241 		goto fail;
3242 
3243 	/* create and start the initial worker */
3244 	if (!create_worker(pool))
3245 		goto fail;
3246 
3247 	/* install */
3248 	hash_add(unbound_pool_hash, &pool->hash_node, hash);
3249 
3250 	return pool;
3251 fail:
3252 	if (pool)
3253 		put_unbound_pool(pool);
3254 	return NULL;
3255 }
3256 
3257 static void rcu_free_pwq(struct rcu_head *rcu)
3258 {
3259 	kmem_cache_free(pwq_cache,
3260 			container_of(rcu, struct pool_workqueue, rcu));
3261 }
3262 
3263 /*
3264  * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3265  * and needs to be destroyed.
3266  */
3267 static void pwq_unbound_release_workfn(struct work_struct *work)
3268 {
3269 	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3270 						  unbound_release_work);
3271 	struct workqueue_struct *wq = pwq->wq;
3272 	struct worker_pool *pool = pwq->pool;
3273 	bool is_last;
3274 
3275 	if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3276 		return;
3277 
3278 	mutex_lock(&wq->mutex);
3279 	list_del_rcu(&pwq->pwqs_node);
3280 	is_last = list_empty(&wq->pwqs);
3281 	mutex_unlock(&wq->mutex);
3282 
3283 	mutex_lock(&wq_pool_mutex);
3284 	put_unbound_pool(pool);
3285 	mutex_unlock(&wq_pool_mutex);
3286 
3287 	call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3288 
3289 	/*
3290 	 * If we're the last pwq going away, @wq is already dead and no one
3291 	 * is gonna access it anymore.  Schedule RCU free.
3292 	 */
3293 	if (is_last)
3294 		call_rcu_sched(&wq->rcu, rcu_free_wq);
3295 }
3296 
3297 /**
3298  * pwq_adjust_max_active - update a pwq's max_active to the current setting
3299  * @pwq: target pool_workqueue
3300  *
3301  * If @pwq isn't freezing, set @pwq->max_active to the associated
3302  * workqueue's saved_max_active and activate delayed work items
3303  * accordingly.  If @pwq is freezing, clear @pwq->max_active to zero.
3304  */
3305 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3306 {
3307 	struct workqueue_struct *wq = pwq->wq;
3308 	bool freezable = wq->flags & WQ_FREEZABLE;
3309 
3310 	/* for @wq->saved_max_active */
3311 	lockdep_assert_held(&wq->mutex);
3312 
3313 	/* fast exit for non-freezable wqs */
3314 	if (!freezable && pwq->max_active == wq->saved_max_active)
3315 		return;
3316 
3317 	spin_lock_irq(&pwq->pool->lock);
3318 
3319 	/*
3320 	 * During [un]freezing, the caller is responsible for ensuring that
3321 	 * this function is called at least once after @workqueue_freezing
3322 	 * is updated and visible.
3323 	 */
3324 	if (!freezable || !workqueue_freezing) {
3325 		pwq->max_active = wq->saved_max_active;
3326 
3327 		while (!list_empty(&pwq->delayed_works) &&
3328 		       pwq->nr_active < pwq->max_active)
3329 			pwq_activate_first_delayed(pwq);
3330 
3331 		/*
3332 		 * Need to kick a worker after thawed or an unbound wq's
3333 		 * max_active is bumped.  It's a slow path.  Do it always.
3334 		 */
3335 		wake_up_worker(pwq->pool);
3336 	} else {
3337 		pwq->max_active = 0;
3338 	}
3339 
3340 	spin_unlock_irq(&pwq->pool->lock);
3341 }
3342 
3343 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3344 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3345 		     struct worker_pool *pool)
3346 {
3347 	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3348 
3349 	memset(pwq, 0, sizeof(*pwq));
3350 
3351 	pwq->pool = pool;
3352 	pwq->wq = wq;
3353 	pwq->flush_color = -1;
3354 	pwq->refcnt = 1;
3355 	INIT_LIST_HEAD(&pwq->delayed_works);
3356 	INIT_LIST_HEAD(&pwq->pwqs_node);
3357 	INIT_LIST_HEAD(&pwq->mayday_node);
3358 	INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3359 }
3360 
3361 /* sync @pwq with the current state of its associated wq and link it */
3362 static void link_pwq(struct pool_workqueue *pwq)
3363 {
3364 	struct workqueue_struct *wq = pwq->wq;
3365 
3366 	lockdep_assert_held(&wq->mutex);
3367 
3368 	/* may be called multiple times, ignore if already linked */
3369 	if (!list_empty(&pwq->pwqs_node))
3370 		return;
3371 
3372 	/* set the matching work_color */
3373 	pwq->work_color = wq->work_color;
3374 
3375 	/* sync max_active to the current setting */
3376 	pwq_adjust_max_active(pwq);
3377 
3378 	/* link in @pwq */
3379 	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3380 }
3381 
3382 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3383 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3384 					const struct workqueue_attrs *attrs)
3385 {
3386 	struct worker_pool *pool;
3387 	struct pool_workqueue *pwq;
3388 
3389 	lockdep_assert_held(&wq_pool_mutex);
3390 
3391 	pool = get_unbound_pool(attrs);
3392 	if (!pool)
3393 		return NULL;
3394 
3395 	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3396 	if (!pwq) {
3397 		put_unbound_pool(pool);
3398 		return NULL;
3399 	}
3400 
3401 	init_pwq(pwq, wq, pool);
3402 	return pwq;
3403 }
3404 
3405 /**
3406  * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3407  * @attrs: the wq_attrs of the default pwq of the target workqueue
3408  * @node: the target NUMA node
3409  * @cpu_going_down: if >= 0, the CPU to consider as offline
3410  * @cpumask: outarg, the resulting cpumask
3411  *
3412  * Calculate the cpumask a workqueue with @attrs should use on @node.  If
3413  * @cpu_going_down is >= 0, that cpu is considered offline during
3414  * calculation.  The result is stored in @cpumask.
3415  *
3416  * If NUMA affinity is not enabled, @attrs->cpumask is always used.  If
3417  * enabled and @node has online CPUs requested by @attrs, the returned
3418  * cpumask is the intersection of the possible CPUs of @node and
3419  * @attrs->cpumask.
3420  *
3421  * The caller is responsible for ensuring that the cpumask of @node stays
3422  * stable.
3423  *
3424  * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3425  * %false if equal.
3426  */
3427 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3428 				 int cpu_going_down, cpumask_t *cpumask)
3429 {
3430 	if (!wq_numa_enabled || attrs->no_numa)
3431 		goto use_dfl;
3432 
3433 	/* does @node have any online CPUs @attrs wants? */
3434 	cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3435 	if (cpu_going_down >= 0)
3436 		cpumask_clear_cpu(cpu_going_down, cpumask);
3437 
3438 	if (cpumask_empty(cpumask))
3439 		goto use_dfl;
3440 
3441 	/* yeap, return possible CPUs in @node that @attrs wants */
3442 	cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3443 	return !cpumask_equal(cpumask, attrs->cpumask);
3444 
3445 use_dfl:
3446 	cpumask_copy(cpumask, attrs->cpumask);
3447 	return false;
3448 }
3449 
3450 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3451 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3452 						   int node,
3453 						   struct pool_workqueue *pwq)
3454 {
3455 	struct pool_workqueue *old_pwq;
3456 
3457 	lockdep_assert_held(&wq_pool_mutex);
3458 	lockdep_assert_held(&wq->mutex);
3459 
3460 	/* link_pwq() can handle duplicate calls */
3461 	link_pwq(pwq);
3462 
3463 	old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3464 	rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3465 	return old_pwq;
3466 }
3467 
3468 /* context to store the prepared attrs & pwqs before applying */
3469 struct apply_wqattrs_ctx {
3470 	struct workqueue_struct	*wq;		/* target workqueue */
3471 	struct workqueue_attrs	*attrs;		/* attrs to apply */
3472 	struct list_head	list;		/* queued for batching commit */
3473 	struct pool_workqueue	*dfl_pwq;
3474 	struct pool_workqueue	*pwq_tbl[];
3475 };
3476 
3477 /* free the resources after success or abort */
3478 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3479 {
3480 	if (ctx) {
3481 		int node;
3482 
3483 		for_each_node(node)
3484 			put_pwq_unlocked(ctx->pwq_tbl[node]);
3485 		put_pwq_unlocked(ctx->dfl_pwq);
3486 
3487 		free_workqueue_attrs(ctx->attrs);
3488 
3489 		kfree(ctx);
3490 	}
3491 }
3492 
3493 /* allocate the attrs and pwqs for later installation */
3494 static struct apply_wqattrs_ctx *
3495 apply_wqattrs_prepare(struct workqueue_struct *wq,
3496 		      const struct workqueue_attrs *attrs)
3497 {
3498 	struct apply_wqattrs_ctx *ctx;
3499 	struct workqueue_attrs *new_attrs, *tmp_attrs;
3500 	int node;
3501 
3502 	lockdep_assert_held(&wq_pool_mutex);
3503 
3504 	ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3505 		      GFP_KERNEL);
3506 
3507 	new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3508 	tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3509 	if (!ctx || !new_attrs || !tmp_attrs)
3510 		goto out_free;
3511 
3512 	/*
3513 	 * Calculate the attrs of the default pwq.
3514 	 * If the user configured cpumask doesn't overlap with the
3515 	 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3516 	 */
3517 	copy_workqueue_attrs(new_attrs, attrs);
3518 	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3519 	if (unlikely(cpumask_empty(new_attrs->cpumask)))
3520 		cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3521 
3522 	/*
3523 	 * We may create multiple pwqs with differing cpumasks.  Make a
3524 	 * copy of @new_attrs which will be modified and used to obtain
3525 	 * pools.
3526 	 */
3527 	copy_workqueue_attrs(tmp_attrs, new_attrs);
3528 
3529 	/*
3530 	 * If something goes wrong during CPU up/down, we'll fall back to
3531 	 * the default pwq covering whole @attrs->cpumask.  Always create
3532 	 * it even if we don't use it immediately.
3533 	 */
3534 	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3535 	if (!ctx->dfl_pwq)
3536 		goto out_free;
3537 
3538 	for_each_node(node) {
3539 		if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3540 			ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3541 			if (!ctx->pwq_tbl[node])
3542 				goto out_free;
3543 		} else {
3544 			ctx->dfl_pwq->refcnt++;
3545 			ctx->pwq_tbl[node] = ctx->dfl_pwq;
3546 		}
3547 	}
3548 
3549 	/* save the user configured attrs and sanitize it. */
3550 	copy_workqueue_attrs(new_attrs, attrs);
3551 	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3552 	ctx->attrs = new_attrs;
3553 
3554 	ctx->wq = wq;
3555 	free_workqueue_attrs(tmp_attrs);
3556 	return ctx;
3557 
3558 out_free:
3559 	free_workqueue_attrs(tmp_attrs);
3560 	free_workqueue_attrs(new_attrs);
3561 	apply_wqattrs_cleanup(ctx);
3562 	return NULL;
3563 }
3564 
3565 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3566 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3567 {
3568 	int node;
3569 
3570 	/* all pwqs have been created successfully, let's install'em */
3571 	mutex_lock(&ctx->wq->mutex);
3572 
3573 	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3574 
3575 	/* save the previous pwq and install the new one */
3576 	for_each_node(node)
3577 		ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3578 							  ctx->pwq_tbl[node]);
3579 
3580 	/* @dfl_pwq might not have been used, ensure it's linked */
3581 	link_pwq(ctx->dfl_pwq);
3582 	swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3583 
3584 	mutex_unlock(&ctx->wq->mutex);
3585 }
3586 
3587 static void apply_wqattrs_lock(void)
3588 {
3589 	/* CPUs should stay stable across pwq creations and installations */
3590 	get_online_cpus();
3591 	mutex_lock(&wq_pool_mutex);
3592 }
3593 
3594 static void apply_wqattrs_unlock(void)
3595 {
3596 	mutex_unlock(&wq_pool_mutex);
3597 	put_online_cpus();
3598 }
3599 
3600 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3601 					const struct workqueue_attrs *attrs)
3602 {
3603 	struct apply_wqattrs_ctx *ctx;
3604 	int ret = -ENOMEM;
3605 
3606 	/* only unbound workqueues can change attributes */
3607 	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3608 		return -EINVAL;
3609 
3610 	/* creating multiple pwqs breaks ordering guarantee */
3611 	if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3612 		return -EINVAL;
3613 
3614 	ctx = apply_wqattrs_prepare(wq, attrs);
3615 
3616 	/* the ctx has been prepared successfully, let's commit it */
3617 	if (ctx) {
3618 		apply_wqattrs_commit(ctx);
3619 		ret = 0;
3620 	}
3621 
3622 	apply_wqattrs_cleanup(ctx);
3623 
3624 	return ret;
3625 }
3626 
3627 /**
3628  * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3629  * @wq: the target workqueue
3630  * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3631  *
3632  * Apply @attrs to an unbound workqueue @wq.  Unless disabled, on NUMA
3633  * machines, this function maps a separate pwq to each NUMA node with
3634  * possibles CPUs in @attrs->cpumask so that work items are affine to the
3635  * NUMA node it was issued on.  Older pwqs are released as in-flight work
3636  * items finish.  Note that a work item which repeatedly requeues itself
3637  * back-to-back will stay on its current pwq.
3638  *
3639  * Performs GFP_KERNEL allocations.
3640  *
3641  * Return: 0 on success and -errno on failure.
3642  */
3643 int apply_workqueue_attrs(struct workqueue_struct *wq,
3644 			  const struct workqueue_attrs *attrs)
3645 {
3646 	int ret;
3647 
3648 	apply_wqattrs_lock();
3649 	ret = apply_workqueue_attrs_locked(wq, attrs);
3650 	apply_wqattrs_unlock();
3651 
3652 	return ret;
3653 }
3654 
3655 /**
3656  * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3657  * @wq: the target workqueue
3658  * @cpu: the CPU coming up or going down
3659  * @online: whether @cpu is coming up or going down
3660  *
3661  * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3662  * %CPU_DOWN_FAILED.  @cpu is being hot[un]plugged, update NUMA affinity of
3663  * @wq accordingly.
3664  *
3665  * If NUMA affinity can't be adjusted due to memory allocation failure, it
3666  * falls back to @wq->dfl_pwq which may not be optimal but is always
3667  * correct.
3668  *
3669  * Note that when the last allowed CPU of a NUMA node goes offline for a
3670  * workqueue with a cpumask spanning multiple nodes, the workers which were
3671  * already executing the work items for the workqueue will lose their CPU
3672  * affinity and may execute on any CPU.  This is similar to how per-cpu
3673  * workqueues behave on CPU_DOWN.  If a workqueue user wants strict
3674  * affinity, it's the user's responsibility to flush the work item from
3675  * CPU_DOWN_PREPARE.
3676  */
3677 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3678 				   bool online)
3679 {
3680 	int node = cpu_to_node(cpu);
3681 	int cpu_off = online ? -1 : cpu;
3682 	struct pool_workqueue *old_pwq = NULL, *pwq;
3683 	struct workqueue_attrs *target_attrs;
3684 	cpumask_t *cpumask;
3685 
3686 	lockdep_assert_held(&wq_pool_mutex);
3687 
3688 	if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3689 	    wq->unbound_attrs->no_numa)
3690 		return;
3691 
3692 	/*
3693 	 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3694 	 * Let's use a preallocated one.  The following buf is protected by
3695 	 * CPU hotplug exclusion.
3696 	 */
3697 	target_attrs = wq_update_unbound_numa_attrs_buf;
3698 	cpumask = target_attrs->cpumask;
3699 
3700 	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3701 	pwq = unbound_pwq_by_node(wq, node);
3702 
3703 	/*
3704 	 * Let's determine what needs to be done.  If the target cpumask is
3705 	 * different from the default pwq's, we need to compare it to @pwq's
3706 	 * and create a new one if they don't match.  If the target cpumask
3707 	 * equals the default pwq's, the default pwq should be used.
3708 	 */
3709 	if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3710 		if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3711 			return;
3712 	} else {
3713 		goto use_dfl_pwq;
3714 	}
3715 
3716 	/* create a new pwq */
3717 	pwq = alloc_unbound_pwq(wq, target_attrs);
3718 	if (!pwq) {
3719 		pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3720 			wq->name);
3721 		goto use_dfl_pwq;
3722 	}
3723 
3724 	/* Install the new pwq. */
3725 	mutex_lock(&wq->mutex);
3726 	old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3727 	goto out_unlock;
3728 
3729 use_dfl_pwq:
3730 	mutex_lock(&wq->mutex);
3731 	spin_lock_irq(&wq->dfl_pwq->pool->lock);
3732 	get_pwq(wq->dfl_pwq);
3733 	spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3734 	old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3735 out_unlock:
3736 	mutex_unlock(&wq->mutex);
3737 	put_pwq_unlocked(old_pwq);
3738 }
3739 
3740 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3741 {
3742 	bool highpri = wq->flags & WQ_HIGHPRI;
3743 	int cpu, ret;
3744 
3745 	if (!(wq->flags & WQ_UNBOUND)) {
3746 		wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3747 		if (!wq->cpu_pwqs)
3748 			return -ENOMEM;
3749 
3750 		for_each_possible_cpu(cpu) {
3751 			struct pool_workqueue *pwq =
3752 				per_cpu_ptr(wq->cpu_pwqs, cpu);
3753 			struct worker_pool *cpu_pools =
3754 				per_cpu(cpu_worker_pools, cpu);
3755 
3756 			init_pwq(pwq, wq, &cpu_pools[highpri]);
3757 
3758 			mutex_lock(&wq->mutex);
3759 			link_pwq(pwq);
3760 			mutex_unlock(&wq->mutex);
3761 		}
3762 		return 0;
3763 	} else if (wq->flags & __WQ_ORDERED) {
3764 		ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3765 		/* there should only be single pwq for ordering guarantee */
3766 		WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3767 			      wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3768 		     "ordering guarantee broken for workqueue %s\n", wq->name);
3769 		return ret;
3770 	} else {
3771 		return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3772 	}
3773 }
3774 
3775 static int wq_clamp_max_active(int max_active, unsigned int flags,
3776 			       const char *name)
3777 {
3778 	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3779 
3780 	if (max_active < 1 || max_active > lim)
3781 		pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3782 			max_active, name, 1, lim);
3783 
3784 	return clamp_val(max_active, 1, lim);
3785 }
3786 
3787 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3788 					       unsigned int flags,
3789 					       int max_active,
3790 					       struct lock_class_key *key,
3791 					       const char *lock_name, ...)
3792 {
3793 	size_t tbl_size = 0;
3794 	va_list args;
3795 	struct workqueue_struct *wq;
3796 	struct pool_workqueue *pwq;
3797 
3798 	/* see the comment above the definition of WQ_POWER_EFFICIENT */
3799 	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3800 		flags |= WQ_UNBOUND;
3801 
3802 	/* allocate wq and format name */
3803 	if (flags & WQ_UNBOUND)
3804 		tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3805 
3806 	wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3807 	if (!wq)
3808 		return NULL;
3809 
3810 	if (flags & WQ_UNBOUND) {
3811 		wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3812 		if (!wq->unbound_attrs)
3813 			goto err_free_wq;
3814 	}
3815 
3816 	va_start(args, lock_name);
3817 	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3818 	va_end(args);
3819 
3820 	max_active = max_active ?: WQ_DFL_ACTIVE;
3821 	max_active = wq_clamp_max_active(max_active, flags, wq->name);
3822 
3823 	/* init wq */
3824 	wq->flags = flags;
3825 	wq->saved_max_active = max_active;
3826 	mutex_init(&wq->mutex);
3827 	atomic_set(&wq->nr_pwqs_to_flush, 0);
3828 	INIT_LIST_HEAD(&wq->pwqs);
3829 	INIT_LIST_HEAD(&wq->flusher_queue);
3830 	INIT_LIST_HEAD(&wq->flusher_overflow);
3831 	INIT_LIST_HEAD(&wq->maydays);
3832 
3833 	lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3834 	INIT_LIST_HEAD(&wq->list);
3835 
3836 	if (alloc_and_link_pwqs(wq) < 0)
3837 		goto err_free_wq;
3838 
3839 	/*
3840 	 * Workqueues which may be used during memory reclaim should
3841 	 * have a rescuer to guarantee forward progress.
3842 	 */
3843 	if (flags & WQ_MEM_RECLAIM) {
3844 		struct worker *rescuer;
3845 
3846 		rescuer = alloc_worker(NUMA_NO_NODE);
3847 		if (!rescuer)
3848 			goto err_destroy;
3849 
3850 		rescuer->rescue_wq = wq;
3851 		rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3852 					       wq->name);
3853 		if (IS_ERR(rescuer->task)) {
3854 			kfree(rescuer);
3855 			goto err_destroy;
3856 		}
3857 
3858 		wq->rescuer = rescuer;
3859 		rescuer->task->flags |= PF_NO_SETAFFINITY;
3860 		wake_up_process(rescuer->task);
3861 	}
3862 
3863 	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3864 		goto err_destroy;
3865 
3866 	/*
3867 	 * wq_pool_mutex protects global freeze state and workqueues list.
3868 	 * Grab it, adjust max_active and add the new @wq to workqueues
3869 	 * list.
3870 	 */
3871 	mutex_lock(&wq_pool_mutex);
3872 
3873 	mutex_lock(&wq->mutex);
3874 	for_each_pwq(pwq, wq)
3875 		pwq_adjust_max_active(pwq);
3876 	mutex_unlock(&wq->mutex);
3877 
3878 	list_add_tail_rcu(&wq->list, &workqueues);
3879 
3880 	mutex_unlock(&wq_pool_mutex);
3881 
3882 	return wq;
3883 
3884 err_free_wq:
3885 	free_workqueue_attrs(wq->unbound_attrs);
3886 	kfree(wq);
3887 	return NULL;
3888 err_destroy:
3889 	destroy_workqueue(wq);
3890 	return NULL;
3891 }
3892 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3893 
3894 /**
3895  * destroy_workqueue - safely terminate a workqueue
3896  * @wq: target workqueue
3897  *
3898  * Safely destroy a workqueue. All work currently pending will be done first.
3899  */
3900 void destroy_workqueue(struct workqueue_struct *wq)
3901 {
3902 	struct pool_workqueue *pwq;
3903 	int node;
3904 
3905 	/* drain it before proceeding with destruction */
3906 	drain_workqueue(wq);
3907 
3908 	/* sanity checks */
3909 	mutex_lock(&wq->mutex);
3910 	for_each_pwq(pwq, wq) {
3911 		int i;
3912 
3913 		for (i = 0; i < WORK_NR_COLORS; i++) {
3914 			if (WARN_ON(pwq->nr_in_flight[i])) {
3915 				mutex_unlock(&wq->mutex);
3916 				return;
3917 			}
3918 		}
3919 
3920 		if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
3921 		    WARN_ON(pwq->nr_active) ||
3922 		    WARN_ON(!list_empty(&pwq->delayed_works))) {
3923 			mutex_unlock(&wq->mutex);
3924 			return;
3925 		}
3926 	}
3927 	mutex_unlock(&wq->mutex);
3928 
3929 	/*
3930 	 * wq list is used to freeze wq, remove from list after
3931 	 * flushing is complete in case freeze races us.
3932 	 */
3933 	mutex_lock(&wq_pool_mutex);
3934 	list_del_rcu(&wq->list);
3935 	mutex_unlock(&wq_pool_mutex);
3936 
3937 	workqueue_sysfs_unregister(wq);
3938 
3939 	if (wq->rescuer)
3940 		kthread_stop(wq->rescuer->task);
3941 
3942 	if (!(wq->flags & WQ_UNBOUND)) {
3943 		/*
3944 		 * The base ref is never dropped on per-cpu pwqs.  Directly
3945 		 * schedule RCU free.
3946 		 */
3947 		call_rcu_sched(&wq->rcu, rcu_free_wq);
3948 	} else {
3949 		/*
3950 		 * We're the sole accessor of @wq at this point.  Directly
3951 		 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
3952 		 * @wq will be freed when the last pwq is released.
3953 		 */
3954 		for_each_node(node) {
3955 			pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3956 			RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
3957 			put_pwq_unlocked(pwq);
3958 		}
3959 
3960 		/*
3961 		 * Put dfl_pwq.  @wq may be freed any time after dfl_pwq is
3962 		 * put.  Don't access it afterwards.
3963 		 */
3964 		pwq = wq->dfl_pwq;
3965 		wq->dfl_pwq = NULL;
3966 		put_pwq_unlocked(pwq);
3967 	}
3968 }
3969 EXPORT_SYMBOL_GPL(destroy_workqueue);
3970 
3971 /**
3972  * workqueue_set_max_active - adjust max_active of a workqueue
3973  * @wq: target workqueue
3974  * @max_active: new max_active value.
3975  *
3976  * Set max_active of @wq to @max_active.
3977  *
3978  * CONTEXT:
3979  * Don't call from IRQ context.
3980  */
3981 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3982 {
3983 	struct pool_workqueue *pwq;
3984 
3985 	/* disallow meddling with max_active for ordered workqueues */
3986 	if (WARN_ON(wq->flags & __WQ_ORDERED))
3987 		return;
3988 
3989 	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3990 
3991 	mutex_lock(&wq->mutex);
3992 
3993 	wq->saved_max_active = max_active;
3994 
3995 	for_each_pwq(pwq, wq)
3996 		pwq_adjust_max_active(pwq);
3997 
3998 	mutex_unlock(&wq->mutex);
3999 }
4000 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4001 
4002 /**
4003  * current_is_workqueue_rescuer - is %current workqueue rescuer?
4004  *
4005  * Determine whether %current is a workqueue rescuer.  Can be used from
4006  * work functions to determine whether it's being run off the rescuer task.
4007  *
4008  * Return: %true if %current is a workqueue rescuer. %false otherwise.
4009  */
4010 bool current_is_workqueue_rescuer(void)
4011 {
4012 	struct worker *worker = current_wq_worker();
4013 
4014 	return worker && worker->rescue_wq;
4015 }
4016 
4017 /**
4018  * workqueue_congested - test whether a workqueue is congested
4019  * @cpu: CPU in question
4020  * @wq: target workqueue
4021  *
4022  * Test whether @wq's cpu workqueue for @cpu is congested.  There is
4023  * no synchronization around this function and the test result is
4024  * unreliable and only useful as advisory hints or for debugging.
4025  *
4026  * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4027  * Note that both per-cpu and unbound workqueues may be associated with
4028  * multiple pool_workqueues which have separate congested states.  A
4029  * workqueue being congested on one CPU doesn't mean the workqueue is also
4030  * contested on other CPUs / NUMA nodes.
4031  *
4032  * Return:
4033  * %true if congested, %false otherwise.
4034  */
4035 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4036 {
4037 	struct pool_workqueue *pwq;
4038 	bool ret;
4039 
4040 	rcu_read_lock_sched();
4041 
4042 	if (cpu == WORK_CPU_UNBOUND)
4043 		cpu = smp_processor_id();
4044 
4045 	if (!(wq->flags & WQ_UNBOUND))
4046 		pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4047 	else
4048 		pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4049 
4050 	ret = !list_empty(&pwq->delayed_works);
4051 	rcu_read_unlock_sched();
4052 
4053 	return ret;
4054 }
4055 EXPORT_SYMBOL_GPL(workqueue_congested);
4056 
4057 /**
4058  * work_busy - test whether a work is currently pending or running
4059  * @work: the work to be tested
4060  *
4061  * Test whether @work is currently pending or running.  There is no
4062  * synchronization around this function and the test result is
4063  * unreliable and only useful as advisory hints or for debugging.
4064  *
4065  * Return:
4066  * OR'd bitmask of WORK_BUSY_* bits.
4067  */
4068 unsigned int work_busy(struct work_struct *work)
4069 {
4070 	struct worker_pool *pool;
4071 	unsigned long flags;
4072 	unsigned int ret = 0;
4073 
4074 	if (work_pending(work))
4075 		ret |= WORK_BUSY_PENDING;
4076 
4077 	local_irq_save(flags);
4078 	pool = get_work_pool(work);
4079 	if (pool) {
4080 		spin_lock(&pool->lock);
4081 		if (find_worker_executing_work(pool, work))
4082 			ret |= WORK_BUSY_RUNNING;
4083 		spin_unlock(&pool->lock);
4084 	}
4085 	local_irq_restore(flags);
4086 
4087 	return ret;
4088 }
4089 EXPORT_SYMBOL_GPL(work_busy);
4090 
4091 /**
4092  * set_worker_desc - set description for the current work item
4093  * @fmt: printf-style format string
4094  * @...: arguments for the format string
4095  *
4096  * This function can be called by a running work function to describe what
4097  * the work item is about.  If the worker task gets dumped, this
4098  * information will be printed out together to help debugging.  The
4099  * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4100  */
4101 void set_worker_desc(const char *fmt, ...)
4102 {
4103 	struct worker *worker = current_wq_worker();
4104 	va_list args;
4105 
4106 	if (worker) {
4107 		va_start(args, fmt);
4108 		vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4109 		va_end(args);
4110 		worker->desc_valid = true;
4111 	}
4112 }
4113 
4114 /**
4115  * print_worker_info - print out worker information and description
4116  * @log_lvl: the log level to use when printing
4117  * @task: target task
4118  *
4119  * If @task is a worker and currently executing a work item, print out the
4120  * name of the workqueue being serviced and worker description set with
4121  * set_worker_desc() by the currently executing work item.
4122  *
4123  * This function can be safely called on any task as long as the
4124  * task_struct itself is accessible.  While safe, this function isn't
4125  * synchronized and may print out mixups or garbages of limited length.
4126  */
4127 void print_worker_info(const char *log_lvl, struct task_struct *task)
4128 {
4129 	work_func_t *fn = NULL;
4130 	char name[WQ_NAME_LEN] = { };
4131 	char desc[WORKER_DESC_LEN] = { };
4132 	struct pool_workqueue *pwq = NULL;
4133 	struct workqueue_struct *wq = NULL;
4134 	bool desc_valid = false;
4135 	struct worker *worker;
4136 
4137 	if (!(task->flags & PF_WQ_WORKER))
4138 		return;
4139 
4140 	/*
4141 	 * This function is called without any synchronization and @task
4142 	 * could be in any state.  Be careful with dereferences.
4143 	 */
4144 	worker = probe_kthread_data(task);
4145 
4146 	/*
4147 	 * Carefully copy the associated workqueue's workfn and name.  Keep
4148 	 * the original last '\0' in case the original contains garbage.
4149 	 */
4150 	probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4151 	probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4152 	probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4153 	probe_kernel_read(name, wq->name, sizeof(name) - 1);
4154 
4155 	/* copy worker description */
4156 	probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4157 	if (desc_valid)
4158 		probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4159 
4160 	if (fn || name[0] || desc[0]) {
4161 		printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4162 		if (desc[0])
4163 			pr_cont(" (%s)", desc);
4164 		pr_cont("\n");
4165 	}
4166 }
4167 
4168 static void pr_cont_pool_info(struct worker_pool *pool)
4169 {
4170 	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4171 	if (pool->node != NUMA_NO_NODE)
4172 		pr_cont(" node=%d", pool->node);
4173 	pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4174 }
4175 
4176 static void pr_cont_work(bool comma, struct work_struct *work)
4177 {
4178 	if (work->func == wq_barrier_func) {
4179 		struct wq_barrier *barr;
4180 
4181 		barr = container_of(work, struct wq_barrier, work);
4182 
4183 		pr_cont("%s BAR(%d)", comma ? "," : "",
4184 			task_pid_nr(barr->task));
4185 	} else {
4186 		pr_cont("%s %pf", comma ? "," : "", work->func);
4187 	}
4188 }
4189 
4190 static void show_pwq(struct pool_workqueue *pwq)
4191 {
4192 	struct worker_pool *pool = pwq->pool;
4193 	struct work_struct *work;
4194 	struct worker *worker;
4195 	bool has_in_flight = false, has_pending = false;
4196 	int bkt;
4197 
4198 	pr_info("  pwq %d:", pool->id);
4199 	pr_cont_pool_info(pool);
4200 
4201 	pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4202 		!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4203 
4204 	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4205 		if (worker->current_pwq == pwq) {
4206 			has_in_flight = true;
4207 			break;
4208 		}
4209 	}
4210 	if (has_in_flight) {
4211 		bool comma = false;
4212 
4213 		pr_info("    in-flight:");
4214 		hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4215 			if (worker->current_pwq != pwq)
4216 				continue;
4217 
4218 			pr_cont("%s %d%s:%pf", comma ? "," : "",
4219 				task_pid_nr(worker->task),
4220 				worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4221 				worker->current_func);
4222 			list_for_each_entry(work, &worker->scheduled, entry)
4223 				pr_cont_work(false, work);
4224 			comma = true;
4225 		}
4226 		pr_cont("\n");
4227 	}
4228 
4229 	list_for_each_entry(work, &pool->worklist, entry) {
4230 		if (get_work_pwq(work) == pwq) {
4231 			has_pending = true;
4232 			break;
4233 		}
4234 	}
4235 	if (has_pending) {
4236 		bool comma = false;
4237 
4238 		pr_info("    pending:");
4239 		list_for_each_entry(work, &pool->worklist, entry) {
4240 			if (get_work_pwq(work) != pwq)
4241 				continue;
4242 
4243 			pr_cont_work(comma, work);
4244 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4245 		}
4246 		pr_cont("\n");
4247 	}
4248 
4249 	if (!list_empty(&pwq->delayed_works)) {
4250 		bool comma = false;
4251 
4252 		pr_info("    delayed:");
4253 		list_for_each_entry(work, &pwq->delayed_works, entry) {
4254 			pr_cont_work(comma, work);
4255 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4256 		}
4257 		pr_cont("\n");
4258 	}
4259 }
4260 
4261 /**
4262  * show_workqueue_state - dump workqueue state
4263  *
4264  * Called from a sysrq handler and prints out all busy workqueues and
4265  * pools.
4266  */
4267 void show_workqueue_state(void)
4268 {
4269 	struct workqueue_struct *wq;
4270 	struct worker_pool *pool;
4271 	unsigned long flags;
4272 	int pi;
4273 
4274 	rcu_read_lock_sched();
4275 
4276 	pr_info("Showing busy workqueues and worker pools:\n");
4277 
4278 	list_for_each_entry_rcu(wq, &workqueues, list) {
4279 		struct pool_workqueue *pwq;
4280 		bool idle = true;
4281 
4282 		for_each_pwq(pwq, wq) {
4283 			if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4284 				idle = false;
4285 				break;
4286 			}
4287 		}
4288 		if (idle)
4289 			continue;
4290 
4291 		pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4292 
4293 		for_each_pwq(pwq, wq) {
4294 			spin_lock_irqsave(&pwq->pool->lock, flags);
4295 			if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4296 				show_pwq(pwq);
4297 			spin_unlock_irqrestore(&pwq->pool->lock, flags);
4298 		}
4299 	}
4300 
4301 	for_each_pool(pool, pi) {
4302 		struct worker *worker;
4303 		bool first = true;
4304 
4305 		spin_lock_irqsave(&pool->lock, flags);
4306 		if (pool->nr_workers == pool->nr_idle)
4307 			goto next_pool;
4308 
4309 		pr_info("pool %d:", pool->id);
4310 		pr_cont_pool_info(pool);
4311 		pr_cont(" workers=%d", pool->nr_workers);
4312 		if (pool->manager)
4313 			pr_cont(" manager: %d",
4314 				task_pid_nr(pool->manager->task));
4315 		list_for_each_entry(worker, &pool->idle_list, entry) {
4316 			pr_cont(" %s%d", first ? "idle: " : "",
4317 				task_pid_nr(worker->task));
4318 			first = false;
4319 		}
4320 		pr_cont("\n");
4321 	next_pool:
4322 		spin_unlock_irqrestore(&pool->lock, flags);
4323 	}
4324 
4325 	rcu_read_unlock_sched();
4326 }
4327 
4328 /*
4329  * CPU hotplug.
4330  *
4331  * There are two challenges in supporting CPU hotplug.  Firstly, there
4332  * are a lot of assumptions on strong associations among work, pwq and
4333  * pool which make migrating pending and scheduled works very
4334  * difficult to implement without impacting hot paths.  Secondly,
4335  * worker pools serve mix of short, long and very long running works making
4336  * blocked draining impractical.
4337  *
4338  * This is solved by allowing the pools to be disassociated from the CPU
4339  * running as an unbound one and allowing it to be reattached later if the
4340  * cpu comes back online.
4341  */
4342 
4343 static void wq_unbind_fn(struct work_struct *work)
4344 {
4345 	int cpu = smp_processor_id();
4346 	struct worker_pool *pool;
4347 	struct worker *worker;
4348 
4349 	for_each_cpu_worker_pool(pool, cpu) {
4350 		mutex_lock(&pool->attach_mutex);
4351 		spin_lock_irq(&pool->lock);
4352 
4353 		/*
4354 		 * We've blocked all attach/detach operations. Make all workers
4355 		 * unbound and set DISASSOCIATED.  Before this, all workers
4356 		 * except for the ones which are still executing works from
4357 		 * before the last CPU down must be on the cpu.  After
4358 		 * this, they may become diasporas.
4359 		 */
4360 		for_each_pool_worker(worker, pool)
4361 			worker->flags |= WORKER_UNBOUND;
4362 
4363 		pool->flags |= POOL_DISASSOCIATED;
4364 
4365 		spin_unlock_irq(&pool->lock);
4366 		mutex_unlock(&pool->attach_mutex);
4367 
4368 		/*
4369 		 * Call schedule() so that we cross rq->lock and thus can
4370 		 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4371 		 * This is necessary as scheduler callbacks may be invoked
4372 		 * from other cpus.
4373 		 */
4374 		schedule();
4375 
4376 		/*
4377 		 * Sched callbacks are disabled now.  Zap nr_running.
4378 		 * After this, nr_running stays zero and need_more_worker()
4379 		 * and keep_working() are always true as long as the
4380 		 * worklist is not empty.  This pool now behaves as an
4381 		 * unbound (in terms of concurrency management) pool which
4382 		 * are served by workers tied to the pool.
4383 		 */
4384 		atomic_set(&pool->nr_running, 0);
4385 
4386 		/*
4387 		 * With concurrency management just turned off, a busy
4388 		 * worker blocking could lead to lengthy stalls.  Kick off
4389 		 * unbound chain execution of currently pending work items.
4390 		 */
4391 		spin_lock_irq(&pool->lock);
4392 		wake_up_worker(pool);
4393 		spin_unlock_irq(&pool->lock);
4394 	}
4395 }
4396 
4397 /**
4398  * rebind_workers - rebind all workers of a pool to the associated CPU
4399  * @pool: pool of interest
4400  *
4401  * @pool->cpu is coming online.  Rebind all workers to the CPU.
4402  */
4403 static void rebind_workers(struct worker_pool *pool)
4404 {
4405 	struct worker *worker;
4406 
4407 	lockdep_assert_held(&pool->attach_mutex);
4408 
4409 	/*
4410 	 * Restore CPU affinity of all workers.  As all idle workers should
4411 	 * be on the run-queue of the associated CPU before any local
4412 	 * wake-ups for concurrency management happen, restore CPU affinity
4413 	 * of all workers first and then clear UNBOUND.  As we're called
4414 	 * from CPU_ONLINE, the following shouldn't fail.
4415 	 */
4416 	for_each_pool_worker(worker, pool)
4417 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4418 						  pool->attrs->cpumask) < 0);
4419 
4420 	spin_lock_irq(&pool->lock);
4421 	pool->flags &= ~POOL_DISASSOCIATED;
4422 
4423 	for_each_pool_worker(worker, pool) {
4424 		unsigned int worker_flags = worker->flags;
4425 
4426 		/*
4427 		 * A bound idle worker should actually be on the runqueue
4428 		 * of the associated CPU for local wake-ups targeting it to
4429 		 * work.  Kick all idle workers so that they migrate to the
4430 		 * associated CPU.  Doing this in the same loop as
4431 		 * replacing UNBOUND with REBOUND is safe as no worker will
4432 		 * be bound before @pool->lock is released.
4433 		 */
4434 		if (worker_flags & WORKER_IDLE)
4435 			wake_up_process(worker->task);
4436 
4437 		/*
4438 		 * We want to clear UNBOUND but can't directly call
4439 		 * worker_clr_flags() or adjust nr_running.  Atomically
4440 		 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4441 		 * @worker will clear REBOUND using worker_clr_flags() when
4442 		 * it initiates the next execution cycle thus restoring
4443 		 * concurrency management.  Note that when or whether
4444 		 * @worker clears REBOUND doesn't affect correctness.
4445 		 *
4446 		 * ACCESS_ONCE() is necessary because @worker->flags may be
4447 		 * tested without holding any lock in
4448 		 * wq_worker_waking_up().  Without it, NOT_RUNNING test may
4449 		 * fail incorrectly leading to premature concurrency
4450 		 * management operations.
4451 		 */
4452 		WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4453 		worker_flags |= WORKER_REBOUND;
4454 		worker_flags &= ~WORKER_UNBOUND;
4455 		ACCESS_ONCE(worker->flags) = worker_flags;
4456 	}
4457 
4458 	spin_unlock_irq(&pool->lock);
4459 }
4460 
4461 /**
4462  * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4463  * @pool: unbound pool of interest
4464  * @cpu: the CPU which is coming up
4465  *
4466  * An unbound pool may end up with a cpumask which doesn't have any online
4467  * CPUs.  When a worker of such pool get scheduled, the scheduler resets
4468  * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
4469  * online CPU before, cpus_allowed of all its workers should be restored.
4470  */
4471 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4472 {
4473 	static cpumask_t cpumask;
4474 	struct worker *worker;
4475 
4476 	lockdep_assert_held(&pool->attach_mutex);
4477 
4478 	/* is @cpu allowed for @pool? */
4479 	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4480 		return;
4481 
4482 	/* is @cpu the only online CPU? */
4483 	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4484 	if (cpumask_weight(&cpumask) != 1)
4485 		return;
4486 
4487 	/* as we're called from CPU_ONLINE, the following shouldn't fail */
4488 	for_each_pool_worker(worker, pool)
4489 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4490 						  pool->attrs->cpumask) < 0);
4491 }
4492 
4493 /*
4494  * Workqueues should be brought up before normal priority CPU notifiers.
4495  * This will be registered high priority CPU notifier.
4496  */
4497 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4498 					       unsigned long action,
4499 					       void *hcpu)
4500 {
4501 	int cpu = (unsigned long)hcpu;
4502 	struct worker_pool *pool;
4503 	struct workqueue_struct *wq;
4504 	int pi;
4505 
4506 	switch (action & ~CPU_TASKS_FROZEN) {
4507 	case CPU_UP_PREPARE:
4508 		for_each_cpu_worker_pool(pool, cpu) {
4509 			if (pool->nr_workers)
4510 				continue;
4511 			if (!create_worker(pool))
4512 				return NOTIFY_BAD;
4513 		}
4514 		break;
4515 
4516 	case CPU_DOWN_FAILED:
4517 	case CPU_ONLINE:
4518 		mutex_lock(&wq_pool_mutex);
4519 
4520 		for_each_pool(pool, pi) {
4521 			mutex_lock(&pool->attach_mutex);
4522 
4523 			if (pool->cpu == cpu)
4524 				rebind_workers(pool);
4525 			else if (pool->cpu < 0)
4526 				restore_unbound_workers_cpumask(pool, cpu);
4527 
4528 			mutex_unlock(&pool->attach_mutex);
4529 		}
4530 
4531 		/* update NUMA affinity of unbound workqueues */
4532 		list_for_each_entry(wq, &workqueues, list)
4533 			wq_update_unbound_numa(wq, cpu, true);
4534 
4535 		mutex_unlock(&wq_pool_mutex);
4536 		break;
4537 	}
4538 	return NOTIFY_OK;
4539 }
4540 
4541 /*
4542  * Workqueues should be brought down after normal priority CPU notifiers.
4543  * This will be registered as low priority CPU notifier.
4544  */
4545 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4546 						 unsigned long action,
4547 						 void *hcpu)
4548 {
4549 	int cpu = (unsigned long)hcpu;
4550 	struct work_struct unbind_work;
4551 	struct workqueue_struct *wq;
4552 
4553 	switch (action & ~CPU_TASKS_FROZEN) {
4554 	case CPU_DOWN_PREPARE:
4555 		/* unbinding per-cpu workers should happen on the local CPU */
4556 		INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4557 		queue_work_on(cpu, system_highpri_wq, &unbind_work);
4558 
4559 		/* update NUMA affinity of unbound workqueues */
4560 		mutex_lock(&wq_pool_mutex);
4561 		list_for_each_entry(wq, &workqueues, list)
4562 			wq_update_unbound_numa(wq, cpu, false);
4563 		mutex_unlock(&wq_pool_mutex);
4564 
4565 		/* wait for per-cpu unbinding to finish */
4566 		flush_work(&unbind_work);
4567 		destroy_work_on_stack(&unbind_work);
4568 		break;
4569 	}
4570 	return NOTIFY_OK;
4571 }
4572 
4573 #ifdef CONFIG_SMP
4574 
4575 struct work_for_cpu {
4576 	struct work_struct work;
4577 	long (*fn)(void *);
4578 	void *arg;
4579 	long ret;
4580 };
4581 
4582 static void work_for_cpu_fn(struct work_struct *work)
4583 {
4584 	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4585 
4586 	wfc->ret = wfc->fn(wfc->arg);
4587 }
4588 
4589 /**
4590  * work_on_cpu - run a function in user context on a particular cpu
4591  * @cpu: the cpu to run on
4592  * @fn: the function to run
4593  * @arg: the function arg
4594  *
4595  * It is up to the caller to ensure that the cpu doesn't go offline.
4596  * The caller must not hold any locks which would prevent @fn from completing.
4597  *
4598  * Return: The value @fn returns.
4599  */
4600 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4601 {
4602 	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4603 
4604 	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4605 	schedule_work_on(cpu, &wfc.work);
4606 	flush_work(&wfc.work);
4607 	destroy_work_on_stack(&wfc.work);
4608 	return wfc.ret;
4609 }
4610 EXPORT_SYMBOL_GPL(work_on_cpu);
4611 #endif /* CONFIG_SMP */
4612 
4613 #ifdef CONFIG_FREEZER
4614 
4615 /**
4616  * freeze_workqueues_begin - begin freezing workqueues
4617  *
4618  * Start freezing workqueues.  After this function returns, all freezable
4619  * workqueues will queue new works to their delayed_works list instead of
4620  * pool->worklist.
4621  *
4622  * CONTEXT:
4623  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4624  */
4625 void freeze_workqueues_begin(void)
4626 {
4627 	struct workqueue_struct *wq;
4628 	struct pool_workqueue *pwq;
4629 
4630 	mutex_lock(&wq_pool_mutex);
4631 
4632 	WARN_ON_ONCE(workqueue_freezing);
4633 	workqueue_freezing = true;
4634 
4635 	list_for_each_entry(wq, &workqueues, list) {
4636 		mutex_lock(&wq->mutex);
4637 		for_each_pwq(pwq, wq)
4638 			pwq_adjust_max_active(pwq);
4639 		mutex_unlock(&wq->mutex);
4640 	}
4641 
4642 	mutex_unlock(&wq_pool_mutex);
4643 }
4644 
4645 /**
4646  * freeze_workqueues_busy - are freezable workqueues still busy?
4647  *
4648  * Check whether freezing is complete.  This function must be called
4649  * between freeze_workqueues_begin() and thaw_workqueues().
4650  *
4651  * CONTEXT:
4652  * Grabs and releases wq_pool_mutex.
4653  *
4654  * Return:
4655  * %true if some freezable workqueues are still busy.  %false if freezing
4656  * is complete.
4657  */
4658 bool freeze_workqueues_busy(void)
4659 {
4660 	bool busy = false;
4661 	struct workqueue_struct *wq;
4662 	struct pool_workqueue *pwq;
4663 
4664 	mutex_lock(&wq_pool_mutex);
4665 
4666 	WARN_ON_ONCE(!workqueue_freezing);
4667 
4668 	list_for_each_entry(wq, &workqueues, list) {
4669 		if (!(wq->flags & WQ_FREEZABLE))
4670 			continue;
4671 		/*
4672 		 * nr_active is monotonically decreasing.  It's safe
4673 		 * to peek without lock.
4674 		 */
4675 		rcu_read_lock_sched();
4676 		for_each_pwq(pwq, wq) {
4677 			WARN_ON_ONCE(pwq->nr_active < 0);
4678 			if (pwq->nr_active) {
4679 				busy = true;
4680 				rcu_read_unlock_sched();
4681 				goto out_unlock;
4682 			}
4683 		}
4684 		rcu_read_unlock_sched();
4685 	}
4686 out_unlock:
4687 	mutex_unlock(&wq_pool_mutex);
4688 	return busy;
4689 }
4690 
4691 /**
4692  * thaw_workqueues - thaw workqueues
4693  *
4694  * Thaw workqueues.  Normal queueing is restored and all collected
4695  * frozen works are transferred to their respective pool worklists.
4696  *
4697  * CONTEXT:
4698  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4699  */
4700 void thaw_workqueues(void)
4701 {
4702 	struct workqueue_struct *wq;
4703 	struct pool_workqueue *pwq;
4704 
4705 	mutex_lock(&wq_pool_mutex);
4706 
4707 	if (!workqueue_freezing)
4708 		goto out_unlock;
4709 
4710 	workqueue_freezing = false;
4711 
4712 	/* restore max_active and repopulate worklist */
4713 	list_for_each_entry(wq, &workqueues, list) {
4714 		mutex_lock(&wq->mutex);
4715 		for_each_pwq(pwq, wq)
4716 			pwq_adjust_max_active(pwq);
4717 		mutex_unlock(&wq->mutex);
4718 	}
4719 
4720 out_unlock:
4721 	mutex_unlock(&wq_pool_mutex);
4722 }
4723 #endif /* CONFIG_FREEZER */
4724 
4725 static int workqueue_apply_unbound_cpumask(void)
4726 {
4727 	LIST_HEAD(ctxs);
4728 	int ret = 0;
4729 	struct workqueue_struct *wq;
4730 	struct apply_wqattrs_ctx *ctx, *n;
4731 
4732 	lockdep_assert_held(&wq_pool_mutex);
4733 
4734 	list_for_each_entry(wq, &workqueues, list) {
4735 		if (!(wq->flags & WQ_UNBOUND))
4736 			continue;
4737 		/* creating multiple pwqs breaks ordering guarantee */
4738 		if (wq->flags & __WQ_ORDERED)
4739 			continue;
4740 
4741 		ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4742 		if (!ctx) {
4743 			ret = -ENOMEM;
4744 			break;
4745 		}
4746 
4747 		list_add_tail(&ctx->list, &ctxs);
4748 	}
4749 
4750 	list_for_each_entry_safe(ctx, n, &ctxs, list) {
4751 		if (!ret)
4752 			apply_wqattrs_commit(ctx);
4753 		apply_wqattrs_cleanup(ctx);
4754 	}
4755 
4756 	return ret;
4757 }
4758 
4759 /**
4760  *  workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4761  *  @cpumask: the cpumask to set
4762  *
4763  *  The low-level workqueues cpumask is a global cpumask that limits
4764  *  the affinity of all unbound workqueues.  This function check the @cpumask
4765  *  and apply it to all unbound workqueues and updates all pwqs of them.
4766  *
4767  *  Retun:	0	- Success
4768  *  		-EINVAL	- Invalid @cpumask
4769  *  		-ENOMEM	- Failed to allocate memory for attrs or pwqs.
4770  */
4771 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4772 {
4773 	int ret = -EINVAL;
4774 	cpumask_var_t saved_cpumask;
4775 
4776 	if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4777 		return -ENOMEM;
4778 
4779 	cpumask_and(cpumask, cpumask, cpu_possible_mask);
4780 	if (!cpumask_empty(cpumask)) {
4781 		apply_wqattrs_lock();
4782 
4783 		/* save the old wq_unbound_cpumask. */
4784 		cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4785 
4786 		/* update wq_unbound_cpumask at first and apply it to wqs. */
4787 		cpumask_copy(wq_unbound_cpumask, cpumask);
4788 		ret = workqueue_apply_unbound_cpumask();
4789 
4790 		/* restore the wq_unbound_cpumask when failed. */
4791 		if (ret < 0)
4792 			cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4793 
4794 		apply_wqattrs_unlock();
4795 	}
4796 
4797 	free_cpumask_var(saved_cpumask);
4798 	return ret;
4799 }
4800 
4801 #ifdef CONFIG_SYSFS
4802 /*
4803  * Workqueues with WQ_SYSFS flag set is visible to userland via
4804  * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
4805  * following attributes.
4806  *
4807  *  per_cpu	RO bool	: whether the workqueue is per-cpu or unbound
4808  *  max_active	RW int	: maximum number of in-flight work items
4809  *
4810  * Unbound workqueues have the following extra attributes.
4811  *
4812  *  id		RO int	: the associated pool ID
4813  *  nice	RW int	: nice value of the workers
4814  *  cpumask	RW mask	: bitmask of allowed CPUs for the workers
4815  */
4816 struct wq_device {
4817 	struct workqueue_struct		*wq;
4818 	struct device			dev;
4819 };
4820 
4821 static struct workqueue_struct *dev_to_wq(struct device *dev)
4822 {
4823 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4824 
4825 	return wq_dev->wq;
4826 }
4827 
4828 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4829 			    char *buf)
4830 {
4831 	struct workqueue_struct *wq = dev_to_wq(dev);
4832 
4833 	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4834 }
4835 static DEVICE_ATTR_RO(per_cpu);
4836 
4837 static ssize_t max_active_show(struct device *dev,
4838 			       struct device_attribute *attr, char *buf)
4839 {
4840 	struct workqueue_struct *wq = dev_to_wq(dev);
4841 
4842 	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4843 }
4844 
4845 static ssize_t max_active_store(struct device *dev,
4846 				struct device_attribute *attr, const char *buf,
4847 				size_t count)
4848 {
4849 	struct workqueue_struct *wq = dev_to_wq(dev);
4850 	int val;
4851 
4852 	if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4853 		return -EINVAL;
4854 
4855 	workqueue_set_max_active(wq, val);
4856 	return count;
4857 }
4858 static DEVICE_ATTR_RW(max_active);
4859 
4860 static struct attribute *wq_sysfs_attrs[] = {
4861 	&dev_attr_per_cpu.attr,
4862 	&dev_attr_max_active.attr,
4863 	NULL,
4864 };
4865 ATTRIBUTE_GROUPS(wq_sysfs);
4866 
4867 static ssize_t wq_pool_ids_show(struct device *dev,
4868 				struct device_attribute *attr, char *buf)
4869 {
4870 	struct workqueue_struct *wq = dev_to_wq(dev);
4871 	const char *delim = "";
4872 	int node, written = 0;
4873 
4874 	rcu_read_lock_sched();
4875 	for_each_node(node) {
4876 		written += scnprintf(buf + written, PAGE_SIZE - written,
4877 				     "%s%d:%d", delim, node,
4878 				     unbound_pwq_by_node(wq, node)->pool->id);
4879 		delim = " ";
4880 	}
4881 	written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4882 	rcu_read_unlock_sched();
4883 
4884 	return written;
4885 }
4886 
4887 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
4888 			    char *buf)
4889 {
4890 	struct workqueue_struct *wq = dev_to_wq(dev);
4891 	int written;
4892 
4893 	mutex_lock(&wq->mutex);
4894 	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
4895 	mutex_unlock(&wq->mutex);
4896 
4897 	return written;
4898 }
4899 
4900 /* prepare workqueue_attrs for sysfs store operations */
4901 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
4902 {
4903 	struct workqueue_attrs *attrs;
4904 
4905 	lockdep_assert_held(&wq_pool_mutex);
4906 
4907 	attrs = alloc_workqueue_attrs(GFP_KERNEL);
4908 	if (!attrs)
4909 		return NULL;
4910 
4911 	copy_workqueue_attrs(attrs, wq->unbound_attrs);
4912 	return attrs;
4913 }
4914 
4915 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
4916 			     const char *buf, size_t count)
4917 {
4918 	struct workqueue_struct *wq = dev_to_wq(dev);
4919 	struct workqueue_attrs *attrs;
4920 	int ret = -ENOMEM;
4921 
4922 	apply_wqattrs_lock();
4923 
4924 	attrs = wq_sysfs_prep_attrs(wq);
4925 	if (!attrs)
4926 		goto out_unlock;
4927 
4928 	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
4929 	    attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
4930 		ret = apply_workqueue_attrs_locked(wq, attrs);
4931 	else
4932 		ret = -EINVAL;
4933 
4934 out_unlock:
4935 	apply_wqattrs_unlock();
4936 	free_workqueue_attrs(attrs);
4937 	return ret ?: count;
4938 }
4939 
4940 static ssize_t wq_cpumask_show(struct device *dev,
4941 			       struct device_attribute *attr, char *buf)
4942 {
4943 	struct workqueue_struct *wq = dev_to_wq(dev);
4944 	int written;
4945 
4946 	mutex_lock(&wq->mutex);
4947 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
4948 			    cpumask_pr_args(wq->unbound_attrs->cpumask));
4949 	mutex_unlock(&wq->mutex);
4950 	return written;
4951 }
4952 
4953 static ssize_t wq_cpumask_store(struct device *dev,
4954 				struct device_attribute *attr,
4955 				const char *buf, size_t count)
4956 {
4957 	struct workqueue_struct *wq = dev_to_wq(dev);
4958 	struct workqueue_attrs *attrs;
4959 	int ret = -ENOMEM;
4960 
4961 	apply_wqattrs_lock();
4962 
4963 	attrs = wq_sysfs_prep_attrs(wq);
4964 	if (!attrs)
4965 		goto out_unlock;
4966 
4967 	ret = cpumask_parse(buf, attrs->cpumask);
4968 	if (!ret)
4969 		ret = apply_workqueue_attrs_locked(wq, attrs);
4970 
4971 out_unlock:
4972 	apply_wqattrs_unlock();
4973 	free_workqueue_attrs(attrs);
4974 	return ret ?: count;
4975 }
4976 
4977 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
4978 			    char *buf)
4979 {
4980 	struct workqueue_struct *wq = dev_to_wq(dev);
4981 	int written;
4982 
4983 	mutex_lock(&wq->mutex);
4984 	written = scnprintf(buf, PAGE_SIZE, "%d\n",
4985 			    !wq->unbound_attrs->no_numa);
4986 	mutex_unlock(&wq->mutex);
4987 
4988 	return written;
4989 }
4990 
4991 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
4992 			     const char *buf, size_t count)
4993 {
4994 	struct workqueue_struct *wq = dev_to_wq(dev);
4995 	struct workqueue_attrs *attrs;
4996 	int v, ret = -ENOMEM;
4997 
4998 	apply_wqattrs_lock();
4999 
5000 	attrs = wq_sysfs_prep_attrs(wq);
5001 	if (!attrs)
5002 		goto out_unlock;
5003 
5004 	ret = -EINVAL;
5005 	if (sscanf(buf, "%d", &v) == 1) {
5006 		attrs->no_numa = !v;
5007 		ret = apply_workqueue_attrs_locked(wq, attrs);
5008 	}
5009 
5010 out_unlock:
5011 	apply_wqattrs_unlock();
5012 	free_workqueue_attrs(attrs);
5013 	return ret ?: count;
5014 }
5015 
5016 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5017 	__ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5018 	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5019 	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5020 	__ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5021 	__ATTR_NULL,
5022 };
5023 
5024 static struct bus_type wq_subsys = {
5025 	.name				= "workqueue",
5026 	.dev_groups			= wq_sysfs_groups,
5027 };
5028 
5029 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5030 		struct device_attribute *attr, char *buf)
5031 {
5032 	int written;
5033 
5034 	mutex_lock(&wq_pool_mutex);
5035 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5036 			    cpumask_pr_args(wq_unbound_cpumask));
5037 	mutex_unlock(&wq_pool_mutex);
5038 
5039 	return written;
5040 }
5041 
5042 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5043 		struct device_attribute *attr, const char *buf, size_t count)
5044 {
5045 	cpumask_var_t cpumask;
5046 	int ret;
5047 
5048 	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5049 		return -ENOMEM;
5050 
5051 	ret = cpumask_parse(buf, cpumask);
5052 	if (!ret)
5053 		ret = workqueue_set_unbound_cpumask(cpumask);
5054 
5055 	free_cpumask_var(cpumask);
5056 	return ret ? ret : count;
5057 }
5058 
5059 static struct device_attribute wq_sysfs_cpumask_attr =
5060 	__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5061 	       wq_unbound_cpumask_store);
5062 
5063 static int __init wq_sysfs_init(void)
5064 {
5065 	int err;
5066 
5067 	err = subsys_virtual_register(&wq_subsys, NULL);
5068 	if (err)
5069 		return err;
5070 
5071 	return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5072 }
5073 core_initcall(wq_sysfs_init);
5074 
5075 static void wq_device_release(struct device *dev)
5076 {
5077 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5078 
5079 	kfree(wq_dev);
5080 }
5081 
5082 /**
5083  * workqueue_sysfs_register - make a workqueue visible in sysfs
5084  * @wq: the workqueue to register
5085  *
5086  * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5087  * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5088  * which is the preferred method.
5089  *
5090  * Workqueue user should use this function directly iff it wants to apply
5091  * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5092  * apply_workqueue_attrs() may race against userland updating the
5093  * attributes.
5094  *
5095  * Return: 0 on success, -errno on failure.
5096  */
5097 int workqueue_sysfs_register(struct workqueue_struct *wq)
5098 {
5099 	struct wq_device *wq_dev;
5100 	int ret;
5101 
5102 	/*
5103 	 * Adjusting max_active or creating new pwqs by applying
5104 	 * attributes breaks ordering guarantee.  Disallow exposing ordered
5105 	 * workqueues.
5106 	 */
5107 	if (WARN_ON(wq->flags & __WQ_ORDERED))
5108 		return -EINVAL;
5109 
5110 	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5111 	if (!wq_dev)
5112 		return -ENOMEM;
5113 
5114 	wq_dev->wq = wq;
5115 	wq_dev->dev.bus = &wq_subsys;
5116 	wq_dev->dev.init_name = wq->name;
5117 	wq_dev->dev.release = wq_device_release;
5118 
5119 	/*
5120 	 * unbound_attrs are created separately.  Suppress uevent until
5121 	 * everything is ready.
5122 	 */
5123 	dev_set_uevent_suppress(&wq_dev->dev, true);
5124 
5125 	ret = device_register(&wq_dev->dev);
5126 	if (ret) {
5127 		kfree(wq_dev);
5128 		wq->wq_dev = NULL;
5129 		return ret;
5130 	}
5131 
5132 	if (wq->flags & WQ_UNBOUND) {
5133 		struct device_attribute *attr;
5134 
5135 		for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5136 			ret = device_create_file(&wq_dev->dev, attr);
5137 			if (ret) {
5138 				device_unregister(&wq_dev->dev);
5139 				wq->wq_dev = NULL;
5140 				return ret;
5141 			}
5142 		}
5143 	}
5144 
5145 	dev_set_uevent_suppress(&wq_dev->dev, false);
5146 	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5147 	return 0;
5148 }
5149 
5150 /**
5151  * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5152  * @wq: the workqueue to unregister
5153  *
5154  * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5155  */
5156 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5157 {
5158 	struct wq_device *wq_dev = wq->wq_dev;
5159 
5160 	if (!wq->wq_dev)
5161 		return;
5162 
5163 	wq->wq_dev = NULL;
5164 	device_unregister(&wq_dev->dev);
5165 }
5166 #else	/* CONFIG_SYSFS */
5167 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)	{ }
5168 #endif	/* CONFIG_SYSFS */
5169 
5170 static void __init wq_numa_init(void)
5171 {
5172 	cpumask_var_t *tbl;
5173 	int node, cpu;
5174 
5175 	if (num_possible_nodes() <= 1)
5176 		return;
5177 
5178 	if (wq_disable_numa) {
5179 		pr_info("workqueue: NUMA affinity support disabled\n");
5180 		return;
5181 	}
5182 
5183 	wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5184 	BUG_ON(!wq_update_unbound_numa_attrs_buf);
5185 
5186 	/*
5187 	 * We want masks of possible CPUs of each node which isn't readily
5188 	 * available.  Build one from cpu_to_node() which should have been
5189 	 * fully initialized by now.
5190 	 */
5191 	tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5192 	BUG_ON(!tbl);
5193 
5194 	for_each_node(node)
5195 		BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5196 				node_online(node) ? node : NUMA_NO_NODE));
5197 
5198 	for_each_possible_cpu(cpu) {
5199 		node = cpu_to_node(cpu);
5200 		if (WARN_ON(node == NUMA_NO_NODE)) {
5201 			pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5202 			/* happens iff arch is bonkers, let's just proceed */
5203 			return;
5204 		}
5205 		cpumask_set_cpu(cpu, tbl[node]);
5206 	}
5207 
5208 	wq_numa_possible_cpumask = tbl;
5209 	wq_numa_enabled = true;
5210 }
5211 
5212 static int __init init_workqueues(void)
5213 {
5214 	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5215 	int i, cpu;
5216 
5217 	WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5218 
5219 	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5220 	cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5221 
5222 	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5223 
5224 	cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5225 	hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5226 
5227 	wq_numa_init();
5228 
5229 	/* initialize CPU pools */
5230 	for_each_possible_cpu(cpu) {
5231 		struct worker_pool *pool;
5232 
5233 		i = 0;
5234 		for_each_cpu_worker_pool(pool, cpu) {
5235 			BUG_ON(init_worker_pool(pool));
5236 			pool->cpu = cpu;
5237 			cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5238 			pool->attrs->nice = std_nice[i++];
5239 			pool->node = cpu_to_node(cpu);
5240 
5241 			/* alloc pool ID */
5242 			mutex_lock(&wq_pool_mutex);
5243 			BUG_ON(worker_pool_assign_id(pool));
5244 			mutex_unlock(&wq_pool_mutex);
5245 		}
5246 	}
5247 
5248 	/* create the initial worker */
5249 	for_each_online_cpu(cpu) {
5250 		struct worker_pool *pool;
5251 
5252 		for_each_cpu_worker_pool(pool, cpu) {
5253 			pool->flags &= ~POOL_DISASSOCIATED;
5254 			BUG_ON(!create_worker(pool));
5255 		}
5256 	}
5257 
5258 	/* create default unbound and ordered wq attrs */
5259 	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5260 		struct workqueue_attrs *attrs;
5261 
5262 		BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5263 		attrs->nice = std_nice[i];
5264 		unbound_std_wq_attrs[i] = attrs;
5265 
5266 		/*
5267 		 * An ordered wq should have only one pwq as ordering is
5268 		 * guaranteed by max_active which is enforced by pwqs.
5269 		 * Turn off NUMA so that dfl_pwq is used for all nodes.
5270 		 */
5271 		BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5272 		attrs->nice = std_nice[i];
5273 		attrs->no_numa = true;
5274 		ordered_wq_attrs[i] = attrs;
5275 	}
5276 
5277 	system_wq = alloc_workqueue("events", 0, 0);
5278 	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5279 	system_long_wq = alloc_workqueue("events_long", 0, 0);
5280 	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5281 					    WQ_UNBOUND_MAX_ACTIVE);
5282 	system_freezable_wq = alloc_workqueue("events_freezable",
5283 					      WQ_FREEZABLE, 0);
5284 	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5285 					      WQ_POWER_EFFICIENT, 0);
5286 	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5287 					      WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5288 					      0);
5289 	BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5290 	       !system_unbound_wq || !system_freezable_wq ||
5291 	       !system_power_efficient_wq ||
5292 	       !system_freezable_power_efficient_wq);
5293 	return 0;
5294 }
5295 early_initcall(init_workqueues);
5296