xref: /openbmc/linux/kernel/workqueue.c (revision f35e839a)
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 is one worker pool for each CPU and
20  * one extra for works which are better served by workers which are
21  * not bound to any specific CPU.
22  *
23  * Please read Documentation/workqueue.txt for details.
24  */
25 
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/jhash.h>
45 #include <linux/hashtable.h>
46 #include <linux/rculist.h>
47 #include <linux/nodemask.h>
48 #include <linux/moduleparam.h>
49 #include <linux/uaccess.h>
50 
51 #include "workqueue_internal.h"
52 
53 enum {
54 	/*
55 	 * worker_pool flags
56 	 *
57 	 * A bound pool is either associated or disassociated with its CPU.
58 	 * While associated (!DISASSOCIATED), all workers are bound to the
59 	 * CPU and none has %WORKER_UNBOUND set and concurrency management
60 	 * is in effect.
61 	 *
62 	 * While DISASSOCIATED, the cpu may be offline and all workers have
63 	 * %WORKER_UNBOUND set and concurrency management disabled, and may
64 	 * be executing on any CPU.  The pool behaves as an unbound one.
65 	 *
66 	 * Note that DISASSOCIATED should be flipped only while holding
67 	 * manager_mutex to avoid changing binding state while
68 	 * create_worker() is in progress.
69 	 */
70 	POOL_MANAGE_WORKERS	= 1 << 0,	/* need to manage workers */
71 	POOL_DISASSOCIATED	= 1 << 2,	/* cpu can't serve workers */
72 	POOL_FREEZING		= 1 << 3,	/* freeze in progress */
73 
74 	/* worker flags */
75 	WORKER_STARTED		= 1 << 0,	/* started */
76 	WORKER_DIE		= 1 << 1,	/* die die die */
77 	WORKER_IDLE		= 1 << 2,	/* is idle */
78 	WORKER_PREP		= 1 << 3,	/* preparing to run works */
79 	WORKER_CPU_INTENSIVE	= 1 << 6,	/* cpu intensive */
80 	WORKER_UNBOUND		= 1 << 7,	/* worker is unbound */
81 	WORKER_REBOUND		= 1 << 8,	/* worker was rebound */
82 
83 	WORKER_NOT_RUNNING	= WORKER_PREP | WORKER_CPU_INTENSIVE |
84 				  WORKER_UNBOUND | WORKER_REBOUND,
85 
86 	NR_STD_WORKER_POOLS	= 2,		/* # standard pools per cpu */
87 
88 	UNBOUND_POOL_HASH_ORDER	= 6,		/* hashed by pool->attrs */
89 	BUSY_WORKER_HASH_ORDER	= 6,		/* 64 pointers */
90 
91 	MAX_IDLE_WORKERS_RATIO	= 4,		/* 1/4 of busy can be idle */
92 	IDLE_WORKER_TIMEOUT	= 300 * HZ,	/* keep idle ones for 5 mins */
93 
94 	MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
95 						/* call for help after 10ms
96 						   (min two ticks) */
97 	MAYDAY_INTERVAL		= HZ / 10,	/* and then every 100ms */
98 	CREATE_COOLDOWN		= HZ,		/* time to breath after fail */
99 
100 	/*
101 	 * Rescue workers are used only on emergencies and shared by
102 	 * all cpus.  Give -20.
103 	 */
104 	RESCUER_NICE_LEVEL	= -20,
105 	HIGHPRI_NICE_LEVEL	= -20,
106 
107 	WQ_NAME_LEN		= 24,
108 };
109 
110 /*
111  * Structure fields follow one of the following exclusion rules.
112  *
113  * I: Modifiable by initialization/destruction paths and read-only for
114  *    everyone else.
115  *
116  * P: Preemption protected.  Disabling preemption is enough and should
117  *    only be modified and accessed from the local cpu.
118  *
119  * L: pool->lock protected.  Access with pool->lock held.
120  *
121  * X: During normal operation, modification requires pool->lock and should
122  *    be done only from local cpu.  Either disabling preemption on local
123  *    cpu or grabbing pool->lock is enough for read access.  If
124  *    POOL_DISASSOCIATED is set, it's identical to L.
125  *
126  * MG: pool->manager_mutex and pool->lock protected.  Writes require both
127  *     locks.  Reads can happen under either lock.
128  *
129  * PL: wq_pool_mutex protected.
130  *
131  * PR: wq_pool_mutex protected for writes.  Sched-RCU protected for reads.
132  *
133  * WQ: wq->mutex protected.
134  *
135  * WR: wq->mutex protected for writes.  Sched-RCU protected for reads.
136  *
137  * MD: wq_mayday_lock protected.
138  */
139 
140 /* struct worker is defined in workqueue_internal.h */
141 
142 struct worker_pool {
143 	spinlock_t		lock;		/* the pool lock */
144 	int			cpu;		/* I: the associated cpu */
145 	int			node;		/* I: the associated node ID */
146 	int			id;		/* I: pool ID */
147 	unsigned int		flags;		/* X: flags */
148 
149 	struct list_head	worklist;	/* L: list of pending works */
150 	int			nr_workers;	/* L: total number of workers */
151 
152 	/* nr_idle includes the ones off idle_list for rebinding */
153 	int			nr_idle;	/* L: currently idle ones */
154 
155 	struct list_head	idle_list;	/* X: list of idle workers */
156 	struct timer_list	idle_timer;	/* L: worker idle timeout */
157 	struct timer_list	mayday_timer;	/* L: SOS timer for workers */
158 
159 	/* a workers is either on busy_hash or idle_list, or the manager */
160 	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
161 						/* L: hash of busy workers */
162 
163 	/* see manage_workers() for details on the two manager mutexes */
164 	struct mutex		manager_arb;	/* manager arbitration */
165 	struct mutex		manager_mutex;	/* manager exclusion */
166 	struct idr		worker_idr;	/* MG: worker IDs and iteration */
167 
168 	struct workqueue_attrs	*attrs;		/* I: worker attributes */
169 	struct hlist_node	hash_node;	/* PL: unbound_pool_hash node */
170 	int			refcnt;		/* PL: refcnt for unbound pools */
171 
172 	/*
173 	 * The current concurrency level.  As it's likely to be accessed
174 	 * from other CPUs during try_to_wake_up(), put it in a separate
175 	 * cacheline.
176 	 */
177 	atomic_t		nr_running ____cacheline_aligned_in_smp;
178 
179 	/*
180 	 * Destruction of pool is sched-RCU protected to allow dereferences
181 	 * from get_work_pool().
182 	 */
183 	struct rcu_head		rcu;
184 } ____cacheline_aligned_in_smp;
185 
186 /*
187  * The per-pool workqueue.  While queued, the lower WORK_STRUCT_FLAG_BITS
188  * of work_struct->data are used for flags and the remaining high bits
189  * point to the pwq; thus, pwqs need to be aligned at two's power of the
190  * number of flag bits.
191  */
192 struct pool_workqueue {
193 	struct worker_pool	*pool;		/* I: the associated pool */
194 	struct workqueue_struct *wq;		/* I: the owning workqueue */
195 	int			work_color;	/* L: current color */
196 	int			flush_color;	/* L: flushing color */
197 	int			refcnt;		/* L: reference count */
198 	int			nr_in_flight[WORK_NR_COLORS];
199 						/* L: nr of in_flight works */
200 	int			nr_active;	/* L: nr of active works */
201 	int			max_active;	/* L: max active works */
202 	struct list_head	delayed_works;	/* L: delayed works */
203 	struct list_head	pwqs_node;	/* WR: node on wq->pwqs */
204 	struct list_head	mayday_node;	/* MD: node on wq->maydays */
205 
206 	/*
207 	 * Release of unbound pwq is punted to system_wq.  See put_pwq()
208 	 * and pwq_unbound_release_workfn() for details.  pool_workqueue
209 	 * itself is also sched-RCU protected so that the first pwq can be
210 	 * determined without grabbing wq->mutex.
211 	 */
212 	struct work_struct	unbound_release_work;
213 	struct rcu_head		rcu;
214 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
215 
216 /*
217  * Structure used to wait for workqueue flush.
218  */
219 struct wq_flusher {
220 	struct list_head	list;		/* WQ: list of flushers */
221 	int			flush_color;	/* WQ: flush color waiting for */
222 	struct completion	done;		/* flush completion */
223 };
224 
225 struct wq_device;
226 
227 /*
228  * The externally visible workqueue.  It relays the issued work items to
229  * the appropriate worker_pool through its pool_workqueues.
230  */
231 struct workqueue_struct {
232 	struct list_head	pwqs;		/* WR: all pwqs of this wq */
233 	struct list_head	list;		/* PL: list of all workqueues */
234 
235 	struct mutex		mutex;		/* protects this wq */
236 	int			work_color;	/* WQ: current work color */
237 	int			flush_color;	/* WQ: current flush color */
238 	atomic_t		nr_pwqs_to_flush; /* flush in progress */
239 	struct wq_flusher	*first_flusher;	/* WQ: first flusher */
240 	struct list_head	flusher_queue;	/* WQ: flush waiters */
241 	struct list_head	flusher_overflow; /* WQ: flush overflow list */
242 
243 	struct list_head	maydays;	/* MD: pwqs requesting rescue */
244 	struct worker		*rescuer;	/* I: rescue worker */
245 
246 	int			nr_drainers;	/* WQ: drain in progress */
247 	int			saved_max_active; /* WQ: saved pwq max_active */
248 
249 	struct workqueue_attrs	*unbound_attrs;	/* WQ: only for unbound wqs */
250 	struct pool_workqueue	*dfl_pwq;	/* WQ: only for unbound wqs */
251 
252 #ifdef CONFIG_SYSFS
253 	struct wq_device	*wq_dev;	/* I: for sysfs interface */
254 #endif
255 #ifdef CONFIG_LOCKDEP
256 	struct lockdep_map	lockdep_map;
257 #endif
258 	char			name[WQ_NAME_LEN]; /* I: workqueue name */
259 
260 	/* hot fields used during command issue, aligned to cacheline */
261 	unsigned int		flags ____cacheline_aligned; /* WQ: WQ_* flags */
262 	struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
263 	struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
264 };
265 
266 static struct kmem_cache *pwq_cache;
267 
268 static int wq_numa_tbl_len;		/* highest possible NUMA node id + 1 */
269 static cpumask_var_t *wq_numa_possible_cpumask;
270 					/* possible CPUs of each node */
271 
272 static bool wq_disable_numa;
273 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
274 
275 static bool wq_numa_enabled;		/* unbound NUMA affinity enabled */
276 
277 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
278 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
279 
280 static DEFINE_MUTEX(wq_pool_mutex);	/* protects pools and workqueues list */
281 static DEFINE_SPINLOCK(wq_mayday_lock);	/* protects wq->maydays list */
282 
283 static LIST_HEAD(workqueues);		/* PL: list of all workqueues */
284 static bool workqueue_freezing;		/* PL: have wqs started freezing? */
285 
286 /* the per-cpu worker pools */
287 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
288 				     cpu_worker_pools);
289 
290 static DEFINE_IDR(worker_pool_idr);	/* PR: idr of all pools */
291 
292 /* PL: hash of all unbound pools keyed by pool->attrs */
293 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
294 
295 /* I: attributes used when instantiating standard unbound pools on demand */
296 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
297 
298 struct workqueue_struct *system_wq __read_mostly;
299 EXPORT_SYMBOL_GPL(system_wq);
300 struct workqueue_struct *system_highpri_wq __read_mostly;
301 EXPORT_SYMBOL_GPL(system_highpri_wq);
302 struct workqueue_struct *system_long_wq __read_mostly;
303 EXPORT_SYMBOL_GPL(system_long_wq);
304 struct workqueue_struct *system_unbound_wq __read_mostly;
305 EXPORT_SYMBOL_GPL(system_unbound_wq);
306 struct workqueue_struct *system_freezable_wq __read_mostly;
307 EXPORT_SYMBOL_GPL(system_freezable_wq);
308 
309 static int worker_thread(void *__worker);
310 static void copy_workqueue_attrs(struct workqueue_attrs *to,
311 				 const struct workqueue_attrs *from);
312 
313 #define CREATE_TRACE_POINTS
314 #include <trace/events/workqueue.h>
315 
316 #define assert_rcu_or_pool_mutex()					\
317 	rcu_lockdep_assert(rcu_read_lock_sched_held() ||		\
318 			   lockdep_is_held(&wq_pool_mutex),		\
319 			   "sched RCU or wq_pool_mutex should be held")
320 
321 #define assert_rcu_or_wq_mutex(wq)					\
322 	rcu_lockdep_assert(rcu_read_lock_sched_held() ||		\
323 			   lockdep_is_held(&wq->mutex),			\
324 			   "sched RCU or wq->mutex should be held")
325 
326 #ifdef CONFIG_LOCKDEP
327 #define assert_manager_or_pool_lock(pool)				\
328 	WARN_ONCE(debug_locks &&					\
329 		  !lockdep_is_held(&(pool)->manager_mutex) &&		\
330 		  !lockdep_is_held(&(pool)->lock),			\
331 		  "pool->manager_mutex or ->lock should be held")
332 #else
333 #define assert_manager_or_pool_lock(pool)	do { } while (0)
334 #endif
335 
336 #define for_each_cpu_worker_pool(pool, cpu)				\
337 	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0];		\
338 	     (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
339 	     (pool)++)
340 
341 /**
342  * for_each_pool - iterate through all worker_pools in the system
343  * @pool: iteration cursor
344  * @pi: integer used for iteration
345  *
346  * This must be called either with wq_pool_mutex held or sched RCU read
347  * locked.  If the pool needs to be used beyond the locking in effect, the
348  * caller is responsible for guaranteeing that the pool stays online.
349  *
350  * The if/else clause exists only for the lockdep assertion and can be
351  * ignored.
352  */
353 #define for_each_pool(pool, pi)						\
354 	idr_for_each_entry(&worker_pool_idr, pool, pi)			\
355 		if (({ assert_rcu_or_pool_mutex(); false; })) { }	\
356 		else
357 
358 /**
359  * for_each_pool_worker - iterate through all workers of a worker_pool
360  * @worker: iteration cursor
361  * @wi: integer used for iteration
362  * @pool: worker_pool to iterate workers of
363  *
364  * This must be called with either @pool->manager_mutex or ->lock held.
365  *
366  * The if/else clause exists only for the lockdep assertion and can be
367  * ignored.
368  */
369 #define for_each_pool_worker(worker, wi, pool)				\
370 	idr_for_each_entry(&(pool)->worker_idr, (worker), (wi))		\
371 		if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
372 		else
373 
374 /**
375  * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
376  * @pwq: iteration cursor
377  * @wq: the target workqueue
378  *
379  * This must be called either with wq->mutex held or sched RCU read locked.
380  * If the pwq needs to be used beyond the locking in effect, the caller is
381  * responsible for guaranteeing that the pwq stays online.
382  *
383  * The if/else clause exists only for the lockdep assertion and can be
384  * ignored.
385  */
386 #define for_each_pwq(pwq, wq)						\
387 	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node)		\
388 		if (({ assert_rcu_or_wq_mutex(wq); false; })) { }	\
389 		else
390 
391 #ifdef CONFIG_DEBUG_OBJECTS_WORK
392 
393 static struct debug_obj_descr work_debug_descr;
394 
395 static void *work_debug_hint(void *addr)
396 {
397 	return ((struct work_struct *) addr)->func;
398 }
399 
400 /*
401  * fixup_init is called when:
402  * - an active object is initialized
403  */
404 static int work_fixup_init(void *addr, enum debug_obj_state state)
405 {
406 	struct work_struct *work = addr;
407 
408 	switch (state) {
409 	case ODEBUG_STATE_ACTIVE:
410 		cancel_work_sync(work);
411 		debug_object_init(work, &work_debug_descr);
412 		return 1;
413 	default:
414 		return 0;
415 	}
416 }
417 
418 /*
419  * fixup_activate is called when:
420  * - an active object is activated
421  * - an unknown object is activated (might be a statically initialized object)
422  */
423 static int work_fixup_activate(void *addr, enum debug_obj_state state)
424 {
425 	struct work_struct *work = addr;
426 
427 	switch (state) {
428 
429 	case ODEBUG_STATE_NOTAVAILABLE:
430 		/*
431 		 * This is not really a fixup. The work struct was
432 		 * statically initialized. We just make sure that it
433 		 * is tracked in the object tracker.
434 		 */
435 		if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
436 			debug_object_init(work, &work_debug_descr);
437 			debug_object_activate(work, &work_debug_descr);
438 			return 0;
439 		}
440 		WARN_ON_ONCE(1);
441 		return 0;
442 
443 	case ODEBUG_STATE_ACTIVE:
444 		WARN_ON(1);
445 
446 	default:
447 		return 0;
448 	}
449 }
450 
451 /*
452  * fixup_free is called when:
453  * - an active object is freed
454  */
455 static int work_fixup_free(void *addr, enum debug_obj_state state)
456 {
457 	struct work_struct *work = addr;
458 
459 	switch (state) {
460 	case ODEBUG_STATE_ACTIVE:
461 		cancel_work_sync(work);
462 		debug_object_free(work, &work_debug_descr);
463 		return 1;
464 	default:
465 		return 0;
466 	}
467 }
468 
469 static struct debug_obj_descr work_debug_descr = {
470 	.name		= "work_struct",
471 	.debug_hint	= work_debug_hint,
472 	.fixup_init	= work_fixup_init,
473 	.fixup_activate	= work_fixup_activate,
474 	.fixup_free	= work_fixup_free,
475 };
476 
477 static inline void debug_work_activate(struct work_struct *work)
478 {
479 	debug_object_activate(work, &work_debug_descr);
480 }
481 
482 static inline void debug_work_deactivate(struct work_struct *work)
483 {
484 	debug_object_deactivate(work, &work_debug_descr);
485 }
486 
487 void __init_work(struct work_struct *work, int onstack)
488 {
489 	if (onstack)
490 		debug_object_init_on_stack(work, &work_debug_descr);
491 	else
492 		debug_object_init(work, &work_debug_descr);
493 }
494 EXPORT_SYMBOL_GPL(__init_work);
495 
496 void destroy_work_on_stack(struct work_struct *work)
497 {
498 	debug_object_free(work, &work_debug_descr);
499 }
500 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
501 
502 #else
503 static inline void debug_work_activate(struct work_struct *work) { }
504 static inline void debug_work_deactivate(struct work_struct *work) { }
505 #endif
506 
507 /* allocate ID and assign it to @pool */
508 static int worker_pool_assign_id(struct worker_pool *pool)
509 {
510 	int ret;
511 
512 	lockdep_assert_held(&wq_pool_mutex);
513 
514 	ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
515 	if (ret >= 0) {
516 		pool->id = ret;
517 		return 0;
518 	}
519 	return ret;
520 }
521 
522 /**
523  * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
524  * @wq: the target workqueue
525  * @node: the node ID
526  *
527  * This must be called either with pwq_lock held or sched RCU read locked.
528  * If the pwq needs to be used beyond the locking in effect, the caller is
529  * responsible for guaranteeing that the pwq stays online.
530  */
531 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
532 						  int node)
533 {
534 	assert_rcu_or_wq_mutex(wq);
535 	return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
536 }
537 
538 static unsigned int work_color_to_flags(int color)
539 {
540 	return color << WORK_STRUCT_COLOR_SHIFT;
541 }
542 
543 static int get_work_color(struct work_struct *work)
544 {
545 	return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
546 		((1 << WORK_STRUCT_COLOR_BITS) - 1);
547 }
548 
549 static int work_next_color(int color)
550 {
551 	return (color + 1) % WORK_NR_COLORS;
552 }
553 
554 /*
555  * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
556  * contain the pointer to the queued pwq.  Once execution starts, the flag
557  * is cleared and the high bits contain OFFQ flags and pool ID.
558  *
559  * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
560  * and clear_work_data() can be used to set the pwq, pool or clear
561  * work->data.  These functions should only be called while the work is
562  * owned - ie. while the PENDING bit is set.
563  *
564  * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
565  * corresponding to a work.  Pool is available once the work has been
566  * queued anywhere after initialization until it is sync canceled.  pwq is
567  * available only while the work item is queued.
568  *
569  * %WORK_OFFQ_CANCELING is used to mark a work item which is being
570  * canceled.  While being canceled, a work item may have its PENDING set
571  * but stay off timer and worklist for arbitrarily long and nobody should
572  * try to steal the PENDING bit.
573  */
574 static inline void set_work_data(struct work_struct *work, unsigned long data,
575 				 unsigned long flags)
576 {
577 	WARN_ON_ONCE(!work_pending(work));
578 	atomic_long_set(&work->data, data | flags | work_static(work));
579 }
580 
581 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
582 			 unsigned long extra_flags)
583 {
584 	set_work_data(work, (unsigned long)pwq,
585 		      WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
586 }
587 
588 static void set_work_pool_and_keep_pending(struct work_struct *work,
589 					   int pool_id)
590 {
591 	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
592 		      WORK_STRUCT_PENDING);
593 }
594 
595 static void set_work_pool_and_clear_pending(struct work_struct *work,
596 					    int pool_id)
597 {
598 	/*
599 	 * The following wmb is paired with the implied mb in
600 	 * test_and_set_bit(PENDING) and ensures all updates to @work made
601 	 * here are visible to and precede any updates by the next PENDING
602 	 * owner.
603 	 */
604 	smp_wmb();
605 	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
606 }
607 
608 static void clear_work_data(struct work_struct *work)
609 {
610 	smp_wmb();	/* see set_work_pool_and_clear_pending() */
611 	set_work_data(work, WORK_STRUCT_NO_POOL, 0);
612 }
613 
614 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
615 {
616 	unsigned long data = atomic_long_read(&work->data);
617 
618 	if (data & WORK_STRUCT_PWQ)
619 		return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
620 	else
621 		return NULL;
622 }
623 
624 /**
625  * get_work_pool - return the worker_pool a given work was associated with
626  * @work: the work item of interest
627  *
628  * Return the worker_pool @work was last associated with.  %NULL if none.
629  *
630  * Pools are created and destroyed under wq_pool_mutex, and allows read
631  * access under sched-RCU read lock.  As such, this function should be
632  * called under wq_pool_mutex or with preemption disabled.
633  *
634  * All fields of the returned pool are accessible as long as the above
635  * mentioned locking is in effect.  If the returned pool needs to be used
636  * beyond the critical section, the caller is responsible for ensuring the
637  * returned pool is and stays online.
638  */
639 static struct worker_pool *get_work_pool(struct work_struct *work)
640 {
641 	unsigned long data = atomic_long_read(&work->data);
642 	int pool_id;
643 
644 	assert_rcu_or_pool_mutex();
645 
646 	if (data & WORK_STRUCT_PWQ)
647 		return ((struct pool_workqueue *)
648 			(data & WORK_STRUCT_WQ_DATA_MASK))->pool;
649 
650 	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
651 	if (pool_id == WORK_OFFQ_POOL_NONE)
652 		return NULL;
653 
654 	return idr_find(&worker_pool_idr, pool_id);
655 }
656 
657 /**
658  * get_work_pool_id - return the worker pool ID a given work is associated with
659  * @work: the work item of interest
660  *
661  * Return the worker_pool ID @work was last associated with.
662  * %WORK_OFFQ_POOL_NONE if none.
663  */
664 static int get_work_pool_id(struct work_struct *work)
665 {
666 	unsigned long data = atomic_long_read(&work->data);
667 
668 	if (data & WORK_STRUCT_PWQ)
669 		return ((struct pool_workqueue *)
670 			(data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
671 
672 	return data >> WORK_OFFQ_POOL_SHIFT;
673 }
674 
675 static void mark_work_canceling(struct work_struct *work)
676 {
677 	unsigned long pool_id = get_work_pool_id(work);
678 
679 	pool_id <<= WORK_OFFQ_POOL_SHIFT;
680 	set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
681 }
682 
683 static bool work_is_canceling(struct work_struct *work)
684 {
685 	unsigned long data = atomic_long_read(&work->data);
686 
687 	return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
688 }
689 
690 /*
691  * Policy functions.  These define the policies on how the global worker
692  * pools are managed.  Unless noted otherwise, these functions assume that
693  * they're being called with pool->lock held.
694  */
695 
696 static bool __need_more_worker(struct worker_pool *pool)
697 {
698 	return !atomic_read(&pool->nr_running);
699 }
700 
701 /*
702  * Need to wake up a worker?  Called from anything but currently
703  * running workers.
704  *
705  * Note that, because unbound workers never contribute to nr_running, this
706  * function will always return %true for unbound pools as long as the
707  * worklist isn't empty.
708  */
709 static bool need_more_worker(struct worker_pool *pool)
710 {
711 	return !list_empty(&pool->worklist) && __need_more_worker(pool);
712 }
713 
714 /* Can I start working?  Called from busy but !running workers. */
715 static bool may_start_working(struct worker_pool *pool)
716 {
717 	return pool->nr_idle;
718 }
719 
720 /* Do I need to keep working?  Called from currently running workers. */
721 static bool keep_working(struct worker_pool *pool)
722 {
723 	return !list_empty(&pool->worklist) &&
724 		atomic_read(&pool->nr_running) <= 1;
725 }
726 
727 /* Do we need a new worker?  Called from manager. */
728 static bool need_to_create_worker(struct worker_pool *pool)
729 {
730 	return need_more_worker(pool) && !may_start_working(pool);
731 }
732 
733 /* Do I need to be the manager? */
734 static bool need_to_manage_workers(struct worker_pool *pool)
735 {
736 	return need_to_create_worker(pool) ||
737 		(pool->flags & POOL_MANAGE_WORKERS);
738 }
739 
740 /* Do we have too many workers and should some go away? */
741 static bool too_many_workers(struct worker_pool *pool)
742 {
743 	bool managing = mutex_is_locked(&pool->manager_arb);
744 	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
745 	int nr_busy = pool->nr_workers - nr_idle;
746 
747 	/*
748 	 * nr_idle and idle_list may disagree if idle rebinding is in
749 	 * progress.  Never return %true if idle_list is empty.
750 	 */
751 	if (list_empty(&pool->idle_list))
752 		return false;
753 
754 	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
755 }
756 
757 /*
758  * Wake up functions.
759  */
760 
761 /* Return the first worker.  Safe with preemption disabled */
762 static struct worker *first_worker(struct worker_pool *pool)
763 {
764 	if (unlikely(list_empty(&pool->idle_list)))
765 		return NULL;
766 
767 	return list_first_entry(&pool->idle_list, struct worker, entry);
768 }
769 
770 /**
771  * wake_up_worker - wake up an idle worker
772  * @pool: worker pool to wake worker from
773  *
774  * Wake up the first idle worker of @pool.
775  *
776  * CONTEXT:
777  * spin_lock_irq(pool->lock).
778  */
779 static void wake_up_worker(struct worker_pool *pool)
780 {
781 	struct worker *worker = first_worker(pool);
782 
783 	if (likely(worker))
784 		wake_up_process(worker->task);
785 }
786 
787 /**
788  * wq_worker_waking_up - a worker is waking up
789  * @task: task waking up
790  * @cpu: CPU @task is waking up to
791  *
792  * This function is called during try_to_wake_up() when a worker is
793  * being awoken.
794  *
795  * CONTEXT:
796  * spin_lock_irq(rq->lock)
797  */
798 void wq_worker_waking_up(struct task_struct *task, int cpu)
799 {
800 	struct worker *worker = kthread_data(task);
801 
802 	if (!(worker->flags & WORKER_NOT_RUNNING)) {
803 		WARN_ON_ONCE(worker->pool->cpu != cpu);
804 		atomic_inc(&worker->pool->nr_running);
805 	}
806 }
807 
808 /**
809  * wq_worker_sleeping - a worker is going to sleep
810  * @task: task going to sleep
811  * @cpu: CPU in question, must be the current CPU number
812  *
813  * This function is called during schedule() when a busy worker is
814  * going to sleep.  Worker on the same cpu can be woken up by
815  * returning pointer to its task.
816  *
817  * CONTEXT:
818  * spin_lock_irq(rq->lock)
819  *
820  * RETURNS:
821  * Worker task on @cpu to wake up, %NULL if none.
822  */
823 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
824 {
825 	struct worker *worker = kthread_data(task), *to_wakeup = NULL;
826 	struct worker_pool *pool;
827 
828 	/*
829 	 * Rescuers, which may not have all the fields set up like normal
830 	 * workers, also reach here, let's not access anything before
831 	 * checking NOT_RUNNING.
832 	 */
833 	if (worker->flags & WORKER_NOT_RUNNING)
834 		return NULL;
835 
836 	pool = worker->pool;
837 
838 	/* this can only happen on the local cpu */
839 	if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
840 		return NULL;
841 
842 	/*
843 	 * The counterpart of the following dec_and_test, implied mb,
844 	 * worklist not empty test sequence is in insert_work().
845 	 * Please read comment there.
846 	 *
847 	 * NOT_RUNNING is clear.  This means that we're bound to and
848 	 * running on the local cpu w/ rq lock held and preemption
849 	 * disabled, which in turn means that none else could be
850 	 * manipulating idle_list, so dereferencing idle_list without pool
851 	 * lock is safe.
852 	 */
853 	if (atomic_dec_and_test(&pool->nr_running) &&
854 	    !list_empty(&pool->worklist))
855 		to_wakeup = first_worker(pool);
856 	return to_wakeup ? to_wakeup->task : NULL;
857 }
858 
859 /**
860  * worker_set_flags - set worker flags and adjust nr_running accordingly
861  * @worker: self
862  * @flags: flags to set
863  * @wakeup: wakeup an idle worker if necessary
864  *
865  * Set @flags in @worker->flags and adjust nr_running accordingly.  If
866  * nr_running becomes zero and @wakeup is %true, an idle worker is
867  * woken up.
868  *
869  * CONTEXT:
870  * spin_lock_irq(pool->lock)
871  */
872 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
873 				    bool wakeup)
874 {
875 	struct worker_pool *pool = worker->pool;
876 
877 	WARN_ON_ONCE(worker->task != current);
878 
879 	/*
880 	 * If transitioning into NOT_RUNNING, adjust nr_running and
881 	 * wake up an idle worker as necessary if requested by
882 	 * @wakeup.
883 	 */
884 	if ((flags & WORKER_NOT_RUNNING) &&
885 	    !(worker->flags & WORKER_NOT_RUNNING)) {
886 		if (wakeup) {
887 			if (atomic_dec_and_test(&pool->nr_running) &&
888 			    !list_empty(&pool->worklist))
889 				wake_up_worker(pool);
890 		} else
891 			atomic_dec(&pool->nr_running);
892 	}
893 
894 	worker->flags |= flags;
895 }
896 
897 /**
898  * worker_clr_flags - clear worker flags and adjust nr_running accordingly
899  * @worker: self
900  * @flags: flags to clear
901  *
902  * Clear @flags in @worker->flags and adjust nr_running accordingly.
903  *
904  * CONTEXT:
905  * spin_lock_irq(pool->lock)
906  */
907 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
908 {
909 	struct worker_pool *pool = worker->pool;
910 	unsigned int oflags = worker->flags;
911 
912 	WARN_ON_ONCE(worker->task != current);
913 
914 	worker->flags &= ~flags;
915 
916 	/*
917 	 * If transitioning out of NOT_RUNNING, increment nr_running.  Note
918 	 * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
919 	 * of multiple flags, not a single flag.
920 	 */
921 	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
922 		if (!(worker->flags & WORKER_NOT_RUNNING))
923 			atomic_inc(&pool->nr_running);
924 }
925 
926 /**
927  * find_worker_executing_work - find worker which is executing a work
928  * @pool: pool of interest
929  * @work: work to find worker for
930  *
931  * Find a worker which is executing @work on @pool by searching
932  * @pool->busy_hash which is keyed by the address of @work.  For a worker
933  * to match, its current execution should match the address of @work and
934  * its work function.  This is to avoid unwanted dependency between
935  * unrelated work executions through a work item being recycled while still
936  * being executed.
937  *
938  * This is a bit tricky.  A work item may be freed once its execution
939  * starts and nothing prevents the freed area from being recycled for
940  * another work item.  If the same work item address ends up being reused
941  * before the original execution finishes, workqueue will identify the
942  * recycled work item as currently executing and make it wait until the
943  * current execution finishes, introducing an unwanted dependency.
944  *
945  * This function checks the work item address and work function to avoid
946  * false positives.  Note that this isn't complete as one may construct a
947  * work function which can introduce dependency onto itself through a
948  * recycled work item.  Well, if somebody wants to shoot oneself in the
949  * foot that badly, there's only so much we can do, and if such deadlock
950  * actually occurs, it should be easy to locate the culprit work function.
951  *
952  * CONTEXT:
953  * spin_lock_irq(pool->lock).
954  *
955  * RETURNS:
956  * Pointer to worker which is executing @work if found, NULL
957  * otherwise.
958  */
959 static struct worker *find_worker_executing_work(struct worker_pool *pool,
960 						 struct work_struct *work)
961 {
962 	struct worker *worker;
963 
964 	hash_for_each_possible(pool->busy_hash, worker, hentry,
965 			       (unsigned long)work)
966 		if (worker->current_work == work &&
967 		    worker->current_func == work->func)
968 			return worker;
969 
970 	return NULL;
971 }
972 
973 /**
974  * move_linked_works - move linked works to a list
975  * @work: start of series of works to be scheduled
976  * @head: target list to append @work to
977  * @nextp: out paramter for nested worklist walking
978  *
979  * Schedule linked works starting from @work to @head.  Work series to
980  * be scheduled starts at @work and includes any consecutive work with
981  * WORK_STRUCT_LINKED set in its predecessor.
982  *
983  * If @nextp is not NULL, it's updated to point to the next work of
984  * the last scheduled work.  This allows move_linked_works() to be
985  * nested inside outer list_for_each_entry_safe().
986  *
987  * CONTEXT:
988  * spin_lock_irq(pool->lock).
989  */
990 static void move_linked_works(struct work_struct *work, struct list_head *head,
991 			      struct work_struct **nextp)
992 {
993 	struct work_struct *n;
994 
995 	/*
996 	 * Linked worklist will always end before the end of the list,
997 	 * use NULL for list head.
998 	 */
999 	list_for_each_entry_safe_from(work, n, NULL, entry) {
1000 		list_move_tail(&work->entry, head);
1001 		if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1002 			break;
1003 	}
1004 
1005 	/*
1006 	 * If we're already inside safe list traversal and have moved
1007 	 * multiple works to the scheduled queue, the next position
1008 	 * needs to be updated.
1009 	 */
1010 	if (nextp)
1011 		*nextp = n;
1012 }
1013 
1014 /**
1015  * get_pwq - get an extra reference on the specified pool_workqueue
1016  * @pwq: pool_workqueue to get
1017  *
1018  * Obtain an extra reference on @pwq.  The caller should guarantee that
1019  * @pwq has positive refcnt and be holding the matching pool->lock.
1020  */
1021 static void get_pwq(struct pool_workqueue *pwq)
1022 {
1023 	lockdep_assert_held(&pwq->pool->lock);
1024 	WARN_ON_ONCE(pwq->refcnt <= 0);
1025 	pwq->refcnt++;
1026 }
1027 
1028 /**
1029  * put_pwq - put a pool_workqueue reference
1030  * @pwq: pool_workqueue to put
1031  *
1032  * Drop a reference of @pwq.  If its refcnt reaches zero, schedule its
1033  * destruction.  The caller should be holding the matching pool->lock.
1034  */
1035 static void put_pwq(struct pool_workqueue *pwq)
1036 {
1037 	lockdep_assert_held(&pwq->pool->lock);
1038 	if (likely(--pwq->refcnt))
1039 		return;
1040 	if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1041 		return;
1042 	/*
1043 	 * @pwq can't be released under pool->lock, bounce to
1044 	 * pwq_unbound_release_workfn().  This never recurses on the same
1045 	 * pool->lock as this path is taken only for unbound workqueues and
1046 	 * the release work item is scheduled on a per-cpu workqueue.  To
1047 	 * avoid lockdep warning, unbound pool->locks are given lockdep
1048 	 * subclass of 1 in get_unbound_pool().
1049 	 */
1050 	schedule_work(&pwq->unbound_release_work);
1051 }
1052 
1053 /**
1054  * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1055  * @pwq: pool_workqueue to put (can be %NULL)
1056  *
1057  * put_pwq() with locking.  This function also allows %NULL @pwq.
1058  */
1059 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1060 {
1061 	if (pwq) {
1062 		/*
1063 		 * As both pwqs and pools are sched-RCU protected, the
1064 		 * following lock operations are safe.
1065 		 */
1066 		spin_lock_irq(&pwq->pool->lock);
1067 		put_pwq(pwq);
1068 		spin_unlock_irq(&pwq->pool->lock);
1069 	}
1070 }
1071 
1072 static void pwq_activate_delayed_work(struct work_struct *work)
1073 {
1074 	struct pool_workqueue *pwq = get_work_pwq(work);
1075 
1076 	trace_workqueue_activate_work(work);
1077 	move_linked_works(work, &pwq->pool->worklist, NULL);
1078 	__clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1079 	pwq->nr_active++;
1080 }
1081 
1082 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1083 {
1084 	struct work_struct *work = list_first_entry(&pwq->delayed_works,
1085 						    struct work_struct, entry);
1086 
1087 	pwq_activate_delayed_work(work);
1088 }
1089 
1090 /**
1091  * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1092  * @pwq: pwq of interest
1093  * @color: color of work which left the queue
1094  *
1095  * A work either has completed or is removed from pending queue,
1096  * decrement nr_in_flight of its pwq and handle workqueue flushing.
1097  *
1098  * CONTEXT:
1099  * spin_lock_irq(pool->lock).
1100  */
1101 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1102 {
1103 	/* uncolored work items don't participate in flushing or nr_active */
1104 	if (color == WORK_NO_COLOR)
1105 		goto out_put;
1106 
1107 	pwq->nr_in_flight[color]--;
1108 
1109 	pwq->nr_active--;
1110 	if (!list_empty(&pwq->delayed_works)) {
1111 		/* one down, submit a delayed one */
1112 		if (pwq->nr_active < pwq->max_active)
1113 			pwq_activate_first_delayed(pwq);
1114 	}
1115 
1116 	/* is flush in progress and are we at the flushing tip? */
1117 	if (likely(pwq->flush_color != color))
1118 		goto out_put;
1119 
1120 	/* are there still in-flight works? */
1121 	if (pwq->nr_in_flight[color])
1122 		goto out_put;
1123 
1124 	/* this pwq is done, clear flush_color */
1125 	pwq->flush_color = -1;
1126 
1127 	/*
1128 	 * If this was the last pwq, wake up the first flusher.  It
1129 	 * will handle the rest.
1130 	 */
1131 	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1132 		complete(&pwq->wq->first_flusher->done);
1133 out_put:
1134 	put_pwq(pwq);
1135 }
1136 
1137 /**
1138  * try_to_grab_pending - steal work item from worklist and disable irq
1139  * @work: work item to steal
1140  * @is_dwork: @work is a delayed_work
1141  * @flags: place to store irq state
1142  *
1143  * Try to grab PENDING bit of @work.  This function can handle @work in any
1144  * stable state - idle, on timer or on worklist.  Return values are
1145  *
1146  *  1		if @work was pending and we successfully stole PENDING
1147  *  0		if @work was idle and we claimed PENDING
1148  *  -EAGAIN	if PENDING couldn't be grabbed at the moment, safe to busy-retry
1149  *  -ENOENT	if someone else is canceling @work, this state may persist
1150  *		for arbitrarily long
1151  *
1152  * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1153  * interrupted while holding PENDING and @work off queue, irq must be
1154  * disabled on entry.  This, combined with delayed_work->timer being
1155  * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1156  *
1157  * On successful return, >= 0, irq is disabled and the caller is
1158  * responsible for releasing it using local_irq_restore(*@flags).
1159  *
1160  * This function is safe to call from any context including IRQ handler.
1161  */
1162 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1163 			       unsigned long *flags)
1164 {
1165 	struct worker_pool *pool;
1166 	struct pool_workqueue *pwq;
1167 
1168 	local_irq_save(*flags);
1169 
1170 	/* try to steal the timer if it exists */
1171 	if (is_dwork) {
1172 		struct delayed_work *dwork = to_delayed_work(work);
1173 
1174 		/*
1175 		 * dwork->timer is irqsafe.  If del_timer() fails, it's
1176 		 * guaranteed that the timer is not queued anywhere and not
1177 		 * running on the local CPU.
1178 		 */
1179 		if (likely(del_timer(&dwork->timer)))
1180 			return 1;
1181 	}
1182 
1183 	/* try to claim PENDING the normal way */
1184 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1185 		return 0;
1186 
1187 	/*
1188 	 * The queueing is in progress, or it is already queued. Try to
1189 	 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1190 	 */
1191 	pool = get_work_pool(work);
1192 	if (!pool)
1193 		goto fail;
1194 
1195 	spin_lock(&pool->lock);
1196 	/*
1197 	 * work->data is guaranteed to point to pwq only while the work
1198 	 * item is queued on pwq->wq, and both updating work->data to point
1199 	 * to pwq on queueing and to pool on dequeueing are done under
1200 	 * pwq->pool->lock.  This in turn guarantees that, if work->data
1201 	 * points to pwq which is associated with a locked pool, the work
1202 	 * item is currently queued on that pool.
1203 	 */
1204 	pwq = get_work_pwq(work);
1205 	if (pwq && pwq->pool == pool) {
1206 		debug_work_deactivate(work);
1207 
1208 		/*
1209 		 * A delayed work item cannot be grabbed directly because
1210 		 * it might have linked NO_COLOR work items which, if left
1211 		 * on the delayed_list, will confuse pwq->nr_active
1212 		 * management later on and cause stall.  Make sure the work
1213 		 * item is activated before grabbing.
1214 		 */
1215 		if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1216 			pwq_activate_delayed_work(work);
1217 
1218 		list_del_init(&work->entry);
1219 		pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1220 
1221 		/* work->data points to pwq iff queued, point to pool */
1222 		set_work_pool_and_keep_pending(work, pool->id);
1223 
1224 		spin_unlock(&pool->lock);
1225 		return 1;
1226 	}
1227 	spin_unlock(&pool->lock);
1228 fail:
1229 	local_irq_restore(*flags);
1230 	if (work_is_canceling(work))
1231 		return -ENOENT;
1232 	cpu_relax();
1233 	return -EAGAIN;
1234 }
1235 
1236 /**
1237  * insert_work - insert a work into a pool
1238  * @pwq: pwq @work belongs to
1239  * @work: work to insert
1240  * @head: insertion point
1241  * @extra_flags: extra WORK_STRUCT_* flags to set
1242  *
1243  * Insert @work which belongs to @pwq after @head.  @extra_flags is or'd to
1244  * work_struct flags.
1245  *
1246  * CONTEXT:
1247  * spin_lock_irq(pool->lock).
1248  */
1249 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1250 			struct list_head *head, unsigned int extra_flags)
1251 {
1252 	struct worker_pool *pool = pwq->pool;
1253 
1254 	/* we own @work, set data and link */
1255 	set_work_pwq(work, pwq, extra_flags);
1256 	list_add_tail(&work->entry, head);
1257 	get_pwq(pwq);
1258 
1259 	/*
1260 	 * Ensure either wq_worker_sleeping() sees the above
1261 	 * list_add_tail() or we see zero nr_running to avoid workers lying
1262 	 * around lazily while there are works to be processed.
1263 	 */
1264 	smp_mb();
1265 
1266 	if (__need_more_worker(pool))
1267 		wake_up_worker(pool);
1268 }
1269 
1270 /*
1271  * Test whether @work is being queued from another work executing on the
1272  * same workqueue.
1273  */
1274 static bool is_chained_work(struct workqueue_struct *wq)
1275 {
1276 	struct worker *worker;
1277 
1278 	worker = current_wq_worker();
1279 	/*
1280 	 * Return %true iff I'm a worker execuing a work item on @wq.  If
1281 	 * I'm @worker, it's safe to dereference it without locking.
1282 	 */
1283 	return worker && worker->current_pwq->wq == wq;
1284 }
1285 
1286 static void __queue_work(int cpu, struct workqueue_struct *wq,
1287 			 struct work_struct *work)
1288 {
1289 	struct pool_workqueue *pwq;
1290 	struct worker_pool *last_pool;
1291 	struct list_head *worklist;
1292 	unsigned int work_flags;
1293 	unsigned int req_cpu = cpu;
1294 
1295 	/*
1296 	 * While a work item is PENDING && off queue, a task trying to
1297 	 * steal the PENDING will busy-loop waiting for it to either get
1298 	 * queued or lose PENDING.  Grabbing PENDING and queueing should
1299 	 * happen with IRQ disabled.
1300 	 */
1301 	WARN_ON_ONCE(!irqs_disabled());
1302 
1303 	debug_work_activate(work);
1304 
1305 	/* if dying, only works from the same workqueue are allowed */
1306 	if (unlikely(wq->flags & __WQ_DRAINING) &&
1307 	    WARN_ON_ONCE(!is_chained_work(wq)))
1308 		return;
1309 retry:
1310 	if (req_cpu == WORK_CPU_UNBOUND)
1311 		cpu = raw_smp_processor_id();
1312 
1313 	/* pwq which will be used unless @work is executing elsewhere */
1314 	if (!(wq->flags & WQ_UNBOUND))
1315 		pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1316 	else
1317 		pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1318 
1319 	/*
1320 	 * If @work was previously on a different pool, it might still be
1321 	 * running there, in which case the work needs to be queued on that
1322 	 * pool to guarantee non-reentrancy.
1323 	 */
1324 	last_pool = get_work_pool(work);
1325 	if (last_pool && last_pool != pwq->pool) {
1326 		struct worker *worker;
1327 
1328 		spin_lock(&last_pool->lock);
1329 
1330 		worker = find_worker_executing_work(last_pool, work);
1331 
1332 		if (worker && worker->current_pwq->wq == wq) {
1333 			pwq = worker->current_pwq;
1334 		} else {
1335 			/* meh... not running there, queue here */
1336 			spin_unlock(&last_pool->lock);
1337 			spin_lock(&pwq->pool->lock);
1338 		}
1339 	} else {
1340 		spin_lock(&pwq->pool->lock);
1341 	}
1342 
1343 	/*
1344 	 * pwq is determined and locked.  For unbound pools, we could have
1345 	 * raced with pwq release and it could already be dead.  If its
1346 	 * refcnt is zero, repeat pwq selection.  Note that pwqs never die
1347 	 * without another pwq replacing it in the numa_pwq_tbl or while
1348 	 * work items are executing on it, so the retrying is guaranteed to
1349 	 * make forward-progress.
1350 	 */
1351 	if (unlikely(!pwq->refcnt)) {
1352 		if (wq->flags & WQ_UNBOUND) {
1353 			spin_unlock(&pwq->pool->lock);
1354 			cpu_relax();
1355 			goto retry;
1356 		}
1357 		/* oops */
1358 		WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1359 			  wq->name, cpu);
1360 	}
1361 
1362 	/* pwq determined, queue */
1363 	trace_workqueue_queue_work(req_cpu, pwq, work);
1364 
1365 	if (WARN_ON(!list_empty(&work->entry))) {
1366 		spin_unlock(&pwq->pool->lock);
1367 		return;
1368 	}
1369 
1370 	pwq->nr_in_flight[pwq->work_color]++;
1371 	work_flags = work_color_to_flags(pwq->work_color);
1372 
1373 	if (likely(pwq->nr_active < pwq->max_active)) {
1374 		trace_workqueue_activate_work(work);
1375 		pwq->nr_active++;
1376 		worklist = &pwq->pool->worklist;
1377 	} else {
1378 		work_flags |= WORK_STRUCT_DELAYED;
1379 		worklist = &pwq->delayed_works;
1380 	}
1381 
1382 	insert_work(pwq, work, worklist, work_flags);
1383 
1384 	spin_unlock(&pwq->pool->lock);
1385 }
1386 
1387 /**
1388  * queue_work_on - queue work on specific cpu
1389  * @cpu: CPU number to execute work on
1390  * @wq: workqueue to use
1391  * @work: work to queue
1392  *
1393  * Returns %false if @work was already on a queue, %true otherwise.
1394  *
1395  * We queue the work to a specific CPU, the caller must ensure it
1396  * can't go away.
1397  */
1398 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1399 		   struct work_struct *work)
1400 {
1401 	bool ret = false;
1402 	unsigned long flags;
1403 
1404 	local_irq_save(flags);
1405 
1406 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1407 		__queue_work(cpu, wq, work);
1408 		ret = true;
1409 	}
1410 
1411 	local_irq_restore(flags);
1412 	return ret;
1413 }
1414 EXPORT_SYMBOL_GPL(queue_work_on);
1415 
1416 void delayed_work_timer_fn(unsigned long __data)
1417 {
1418 	struct delayed_work *dwork = (struct delayed_work *)__data;
1419 
1420 	/* should have been called from irqsafe timer with irq already off */
1421 	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
1422 }
1423 EXPORT_SYMBOL(delayed_work_timer_fn);
1424 
1425 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1426 				struct delayed_work *dwork, unsigned long delay)
1427 {
1428 	struct timer_list *timer = &dwork->timer;
1429 	struct work_struct *work = &dwork->work;
1430 
1431 	WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1432 		     timer->data != (unsigned long)dwork);
1433 	WARN_ON_ONCE(timer_pending(timer));
1434 	WARN_ON_ONCE(!list_empty(&work->entry));
1435 
1436 	/*
1437 	 * If @delay is 0, queue @dwork->work immediately.  This is for
1438 	 * both optimization and correctness.  The earliest @timer can
1439 	 * expire is on the closest next tick and delayed_work users depend
1440 	 * on that there's no such delay when @delay is 0.
1441 	 */
1442 	if (!delay) {
1443 		__queue_work(cpu, wq, &dwork->work);
1444 		return;
1445 	}
1446 
1447 	timer_stats_timer_set_start_info(&dwork->timer);
1448 
1449 	dwork->wq = wq;
1450 	dwork->cpu = cpu;
1451 	timer->expires = jiffies + delay;
1452 
1453 	if (unlikely(cpu != WORK_CPU_UNBOUND))
1454 		add_timer_on(timer, cpu);
1455 	else
1456 		add_timer(timer);
1457 }
1458 
1459 /**
1460  * queue_delayed_work_on - queue work on specific CPU after delay
1461  * @cpu: CPU number to execute work on
1462  * @wq: workqueue to use
1463  * @dwork: work to queue
1464  * @delay: number of jiffies to wait before queueing
1465  *
1466  * Returns %false if @work was already on a queue, %true otherwise.  If
1467  * @delay is zero and @dwork is idle, it will be scheduled for immediate
1468  * execution.
1469  */
1470 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1471 			   struct delayed_work *dwork, unsigned long delay)
1472 {
1473 	struct work_struct *work = &dwork->work;
1474 	bool ret = false;
1475 	unsigned long flags;
1476 
1477 	/* read the comment in __queue_work() */
1478 	local_irq_save(flags);
1479 
1480 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1481 		__queue_delayed_work(cpu, wq, dwork, delay);
1482 		ret = true;
1483 	}
1484 
1485 	local_irq_restore(flags);
1486 	return ret;
1487 }
1488 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1489 
1490 /**
1491  * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1492  * @cpu: CPU number to execute work on
1493  * @wq: workqueue to use
1494  * @dwork: work to queue
1495  * @delay: number of jiffies to wait before queueing
1496  *
1497  * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1498  * modify @dwork's timer so that it expires after @delay.  If @delay is
1499  * zero, @work is guaranteed to be scheduled immediately regardless of its
1500  * current state.
1501  *
1502  * Returns %false if @dwork was idle and queued, %true if @dwork was
1503  * pending and its timer was modified.
1504  *
1505  * This function is safe to call from any context including IRQ handler.
1506  * See try_to_grab_pending() for details.
1507  */
1508 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1509 			 struct delayed_work *dwork, unsigned long delay)
1510 {
1511 	unsigned long flags;
1512 	int ret;
1513 
1514 	do {
1515 		ret = try_to_grab_pending(&dwork->work, true, &flags);
1516 	} while (unlikely(ret == -EAGAIN));
1517 
1518 	if (likely(ret >= 0)) {
1519 		__queue_delayed_work(cpu, wq, dwork, delay);
1520 		local_irq_restore(flags);
1521 	}
1522 
1523 	/* -ENOENT from try_to_grab_pending() becomes %true */
1524 	return ret;
1525 }
1526 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1527 
1528 /**
1529  * worker_enter_idle - enter idle state
1530  * @worker: worker which is entering idle state
1531  *
1532  * @worker is entering idle state.  Update stats and idle timer if
1533  * necessary.
1534  *
1535  * LOCKING:
1536  * spin_lock_irq(pool->lock).
1537  */
1538 static void worker_enter_idle(struct worker *worker)
1539 {
1540 	struct worker_pool *pool = worker->pool;
1541 
1542 	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1543 	    WARN_ON_ONCE(!list_empty(&worker->entry) &&
1544 			 (worker->hentry.next || worker->hentry.pprev)))
1545 		return;
1546 
1547 	/* can't use worker_set_flags(), also called from start_worker() */
1548 	worker->flags |= WORKER_IDLE;
1549 	pool->nr_idle++;
1550 	worker->last_active = jiffies;
1551 
1552 	/* idle_list is LIFO */
1553 	list_add(&worker->entry, &pool->idle_list);
1554 
1555 	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1556 		mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1557 
1558 	/*
1559 	 * Sanity check nr_running.  Because wq_unbind_fn() releases
1560 	 * pool->lock between setting %WORKER_UNBOUND and zapping
1561 	 * nr_running, the warning may trigger spuriously.  Check iff
1562 	 * unbind is not in progress.
1563 	 */
1564 	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1565 		     pool->nr_workers == pool->nr_idle &&
1566 		     atomic_read(&pool->nr_running));
1567 }
1568 
1569 /**
1570  * worker_leave_idle - leave idle state
1571  * @worker: worker which is leaving idle state
1572  *
1573  * @worker is leaving idle state.  Update stats.
1574  *
1575  * LOCKING:
1576  * spin_lock_irq(pool->lock).
1577  */
1578 static void worker_leave_idle(struct worker *worker)
1579 {
1580 	struct worker_pool *pool = worker->pool;
1581 
1582 	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1583 		return;
1584 	worker_clr_flags(worker, WORKER_IDLE);
1585 	pool->nr_idle--;
1586 	list_del_init(&worker->entry);
1587 }
1588 
1589 /**
1590  * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1591  * @pool: target worker_pool
1592  *
1593  * Bind %current to the cpu of @pool if it is associated and lock @pool.
1594  *
1595  * Works which are scheduled while the cpu is online must at least be
1596  * scheduled to a worker which is bound to the cpu so that if they are
1597  * flushed from cpu callbacks while cpu is going down, they are
1598  * guaranteed to execute on the cpu.
1599  *
1600  * This function is to be used by unbound workers and rescuers to bind
1601  * themselves to the target cpu and may race with cpu going down or
1602  * coming online.  kthread_bind() can't be used because it may put the
1603  * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1604  * verbatim as it's best effort and blocking and pool may be
1605  * [dis]associated in the meantime.
1606  *
1607  * This function tries set_cpus_allowed() and locks pool and verifies the
1608  * binding against %POOL_DISASSOCIATED which is set during
1609  * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1610  * enters idle state or fetches works without dropping lock, it can
1611  * guarantee the scheduling requirement described in the first paragraph.
1612  *
1613  * CONTEXT:
1614  * Might sleep.  Called without any lock but returns with pool->lock
1615  * held.
1616  *
1617  * RETURNS:
1618  * %true if the associated pool is online (@worker is successfully
1619  * bound), %false if offline.
1620  */
1621 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1622 __acquires(&pool->lock)
1623 {
1624 	while (true) {
1625 		/*
1626 		 * The following call may fail, succeed or succeed
1627 		 * without actually migrating the task to the cpu if
1628 		 * it races with cpu hotunplug operation.  Verify
1629 		 * against POOL_DISASSOCIATED.
1630 		 */
1631 		if (!(pool->flags & POOL_DISASSOCIATED))
1632 			set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1633 
1634 		spin_lock_irq(&pool->lock);
1635 		if (pool->flags & POOL_DISASSOCIATED)
1636 			return false;
1637 		if (task_cpu(current) == pool->cpu &&
1638 		    cpumask_equal(&current->cpus_allowed, pool->attrs->cpumask))
1639 			return true;
1640 		spin_unlock_irq(&pool->lock);
1641 
1642 		/*
1643 		 * We've raced with CPU hot[un]plug.  Give it a breather
1644 		 * and retry migration.  cond_resched() is required here;
1645 		 * otherwise, we might deadlock against cpu_stop trying to
1646 		 * bring down the CPU on non-preemptive kernel.
1647 		 */
1648 		cpu_relax();
1649 		cond_resched();
1650 	}
1651 }
1652 
1653 static struct worker *alloc_worker(void)
1654 {
1655 	struct worker *worker;
1656 
1657 	worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1658 	if (worker) {
1659 		INIT_LIST_HEAD(&worker->entry);
1660 		INIT_LIST_HEAD(&worker->scheduled);
1661 		/* on creation a worker is in !idle && prep state */
1662 		worker->flags = WORKER_PREP;
1663 	}
1664 	return worker;
1665 }
1666 
1667 /**
1668  * create_worker - create a new workqueue worker
1669  * @pool: pool the new worker will belong to
1670  *
1671  * Create a new worker which is bound to @pool.  The returned worker
1672  * can be started by calling start_worker() or destroyed using
1673  * destroy_worker().
1674  *
1675  * CONTEXT:
1676  * Might sleep.  Does GFP_KERNEL allocations.
1677  *
1678  * RETURNS:
1679  * Pointer to the newly created worker.
1680  */
1681 static struct worker *create_worker(struct worker_pool *pool)
1682 {
1683 	struct worker *worker = NULL;
1684 	int id = -1;
1685 	char id_buf[16];
1686 
1687 	lockdep_assert_held(&pool->manager_mutex);
1688 
1689 	/*
1690 	 * ID is needed to determine kthread name.  Allocate ID first
1691 	 * without installing the pointer.
1692 	 */
1693 	idr_preload(GFP_KERNEL);
1694 	spin_lock_irq(&pool->lock);
1695 
1696 	id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1697 
1698 	spin_unlock_irq(&pool->lock);
1699 	idr_preload_end();
1700 	if (id < 0)
1701 		goto fail;
1702 
1703 	worker = alloc_worker();
1704 	if (!worker)
1705 		goto fail;
1706 
1707 	worker->pool = pool;
1708 	worker->id = id;
1709 
1710 	if (pool->cpu >= 0)
1711 		snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1712 			 pool->attrs->nice < 0  ? "H" : "");
1713 	else
1714 		snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1715 
1716 	worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1717 					      "kworker/%s", id_buf);
1718 	if (IS_ERR(worker->task))
1719 		goto fail;
1720 
1721 	/*
1722 	 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1723 	 * online CPUs.  It'll be re-applied when any of the CPUs come up.
1724 	 */
1725 	set_user_nice(worker->task, pool->attrs->nice);
1726 	set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1727 
1728 	/* prevent userland from meddling with cpumask of workqueue workers */
1729 	worker->task->flags |= PF_NO_SETAFFINITY;
1730 
1731 	/*
1732 	 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1733 	 * remains stable across this function.  See the comments above the
1734 	 * flag definition for details.
1735 	 */
1736 	if (pool->flags & POOL_DISASSOCIATED)
1737 		worker->flags |= WORKER_UNBOUND;
1738 
1739 	/* successful, commit the pointer to idr */
1740 	spin_lock_irq(&pool->lock);
1741 	idr_replace(&pool->worker_idr, worker, worker->id);
1742 	spin_unlock_irq(&pool->lock);
1743 
1744 	return worker;
1745 
1746 fail:
1747 	if (id >= 0) {
1748 		spin_lock_irq(&pool->lock);
1749 		idr_remove(&pool->worker_idr, id);
1750 		spin_unlock_irq(&pool->lock);
1751 	}
1752 	kfree(worker);
1753 	return NULL;
1754 }
1755 
1756 /**
1757  * start_worker - start a newly created worker
1758  * @worker: worker to start
1759  *
1760  * Make the pool aware of @worker and start it.
1761  *
1762  * CONTEXT:
1763  * spin_lock_irq(pool->lock).
1764  */
1765 static void start_worker(struct worker *worker)
1766 {
1767 	worker->flags |= WORKER_STARTED;
1768 	worker->pool->nr_workers++;
1769 	worker_enter_idle(worker);
1770 	wake_up_process(worker->task);
1771 }
1772 
1773 /**
1774  * create_and_start_worker - create and start a worker for a pool
1775  * @pool: the target pool
1776  *
1777  * Grab the managership of @pool and create and start a new worker for it.
1778  */
1779 static int create_and_start_worker(struct worker_pool *pool)
1780 {
1781 	struct worker *worker;
1782 
1783 	mutex_lock(&pool->manager_mutex);
1784 
1785 	worker = create_worker(pool);
1786 	if (worker) {
1787 		spin_lock_irq(&pool->lock);
1788 		start_worker(worker);
1789 		spin_unlock_irq(&pool->lock);
1790 	}
1791 
1792 	mutex_unlock(&pool->manager_mutex);
1793 
1794 	return worker ? 0 : -ENOMEM;
1795 }
1796 
1797 /**
1798  * destroy_worker - destroy a workqueue worker
1799  * @worker: worker to be destroyed
1800  *
1801  * Destroy @worker and adjust @pool stats accordingly.
1802  *
1803  * CONTEXT:
1804  * spin_lock_irq(pool->lock) which is released and regrabbed.
1805  */
1806 static void destroy_worker(struct worker *worker)
1807 {
1808 	struct worker_pool *pool = worker->pool;
1809 
1810 	lockdep_assert_held(&pool->manager_mutex);
1811 	lockdep_assert_held(&pool->lock);
1812 
1813 	/* sanity check frenzy */
1814 	if (WARN_ON(worker->current_work) ||
1815 	    WARN_ON(!list_empty(&worker->scheduled)))
1816 		return;
1817 
1818 	if (worker->flags & WORKER_STARTED)
1819 		pool->nr_workers--;
1820 	if (worker->flags & WORKER_IDLE)
1821 		pool->nr_idle--;
1822 
1823 	list_del_init(&worker->entry);
1824 	worker->flags |= WORKER_DIE;
1825 
1826 	idr_remove(&pool->worker_idr, worker->id);
1827 
1828 	spin_unlock_irq(&pool->lock);
1829 
1830 	kthread_stop(worker->task);
1831 	kfree(worker);
1832 
1833 	spin_lock_irq(&pool->lock);
1834 }
1835 
1836 static void idle_worker_timeout(unsigned long __pool)
1837 {
1838 	struct worker_pool *pool = (void *)__pool;
1839 
1840 	spin_lock_irq(&pool->lock);
1841 
1842 	if (too_many_workers(pool)) {
1843 		struct worker *worker;
1844 		unsigned long expires;
1845 
1846 		/* idle_list is kept in LIFO order, check the last one */
1847 		worker = list_entry(pool->idle_list.prev, struct worker, entry);
1848 		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1849 
1850 		if (time_before(jiffies, expires))
1851 			mod_timer(&pool->idle_timer, expires);
1852 		else {
1853 			/* it's been idle for too long, wake up manager */
1854 			pool->flags |= POOL_MANAGE_WORKERS;
1855 			wake_up_worker(pool);
1856 		}
1857 	}
1858 
1859 	spin_unlock_irq(&pool->lock);
1860 }
1861 
1862 static void send_mayday(struct work_struct *work)
1863 {
1864 	struct pool_workqueue *pwq = get_work_pwq(work);
1865 	struct workqueue_struct *wq = pwq->wq;
1866 
1867 	lockdep_assert_held(&wq_mayday_lock);
1868 
1869 	if (!wq->rescuer)
1870 		return;
1871 
1872 	/* mayday mayday mayday */
1873 	if (list_empty(&pwq->mayday_node)) {
1874 		list_add_tail(&pwq->mayday_node, &wq->maydays);
1875 		wake_up_process(wq->rescuer->task);
1876 	}
1877 }
1878 
1879 static void pool_mayday_timeout(unsigned long __pool)
1880 {
1881 	struct worker_pool *pool = (void *)__pool;
1882 	struct work_struct *work;
1883 
1884 	spin_lock_irq(&wq_mayday_lock);		/* for wq->maydays */
1885 	spin_lock(&pool->lock);
1886 
1887 	if (need_to_create_worker(pool)) {
1888 		/*
1889 		 * We've been trying to create a new worker but
1890 		 * haven't been successful.  We might be hitting an
1891 		 * allocation deadlock.  Send distress signals to
1892 		 * rescuers.
1893 		 */
1894 		list_for_each_entry(work, &pool->worklist, entry)
1895 			send_mayday(work);
1896 	}
1897 
1898 	spin_unlock(&pool->lock);
1899 	spin_unlock_irq(&wq_mayday_lock);
1900 
1901 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1902 }
1903 
1904 /**
1905  * maybe_create_worker - create a new worker if necessary
1906  * @pool: pool to create a new worker for
1907  *
1908  * Create a new worker for @pool if necessary.  @pool is guaranteed to
1909  * have at least one idle worker on return from this function.  If
1910  * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1911  * sent to all rescuers with works scheduled on @pool to resolve
1912  * possible allocation deadlock.
1913  *
1914  * On return, need_to_create_worker() is guaranteed to be %false and
1915  * may_start_working() %true.
1916  *
1917  * LOCKING:
1918  * spin_lock_irq(pool->lock) which may be released and regrabbed
1919  * multiple times.  Does GFP_KERNEL allocations.  Called only from
1920  * manager.
1921  *
1922  * RETURNS:
1923  * %false if no action was taken and pool->lock stayed locked, %true
1924  * otherwise.
1925  */
1926 static bool maybe_create_worker(struct worker_pool *pool)
1927 __releases(&pool->lock)
1928 __acquires(&pool->lock)
1929 {
1930 	if (!need_to_create_worker(pool))
1931 		return false;
1932 restart:
1933 	spin_unlock_irq(&pool->lock);
1934 
1935 	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1936 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1937 
1938 	while (true) {
1939 		struct worker *worker;
1940 
1941 		worker = create_worker(pool);
1942 		if (worker) {
1943 			del_timer_sync(&pool->mayday_timer);
1944 			spin_lock_irq(&pool->lock);
1945 			start_worker(worker);
1946 			if (WARN_ON_ONCE(need_to_create_worker(pool)))
1947 				goto restart;
1948 			return true;
1949 		}
1950 
1951 		if (!need_to_create_worker(pool))
1952 			break;
1953 
1954 		__set_current_state(TASK_INTERRUPTIBLE);
1955 		schedule_timeout(CREATE_COOLDOWN);
1956 
1957 		if (!need_to_create_worker(pool))
1958 			break;
1959 	}
1960 
1961 	del_timer_sync(&pool->mayday_timer);
1962 	spin_lock_irq(&pool->lock);
1963 	if (need_to_create_worker(pool))
1964 		goto restart;
1965 	return true;
1966 }
1967 
1968 /**
1969  * maybe_destroy_worker - destroy workers which have been idle for a while
1970  * @pool: pool to destroy workers for
1971  *
1972  * Destroy @pool workers which have been idle for longer than
1973  * IDLE_WORKER_TIMEOUT.
1974  *
1975  * LOCKING:
1976  * spin_lock_irq(pool->lock) which may be released and regrabbed
1977  * multiple times.  Called only from manager.
1978  *
1979  * RETURNS:
1980  * %false if no action was taken and pool->lock stayed locked, %true
1981  * otherwise.
1982  */
1983 static bool maybe_destroy_workers(struct worker_pool *pool)
1984 {
1985 	bool ret = false;
1986 
1987 	while (too_many_workers(pool)) {
1988 		struct worker *worker;
1989 		unsigned long expires;
1990 
1991 		worker = list_entry(pool->idle_list.prev, struct worker, entry);
1992 		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1993 
1994 		if (time_before(jiffies, expires)) {
1995 			mod_timer(&pool->idle_timer, expires);
1996 			break;
1997 		}
1998 
1999 		destroy_worker(worker);
2000 		ret = true;
2001 	}
2002 
2003 	return ret;
2004 }
2005 
2006 /**
2007  * manage_workers - manage worker pool
2008  * @worker: self
2009  *
2010  * Assume the manager role and manage the worker pool @worker belongs
2011  * to.  At any given time, there can be only zero or one manager per
2012  * pool.  The exclusion is handled automatically by this function.
2013  *
2014  * The caller can safely start processing works on false return.  On
2015  * true return, it's guaranteed that need_to_create_worker() is false
2016  * and may_start_working() is true.
2017  *
2018  * CONTEXT:
2019  * spin_lock_irq(pool->lock) which may be released and regrabbed
2020  * multiple times.  Does GFP_KERNEL allocations.
2021  *
2022  * RETURNS:
2023  * spin_lock_irq(pool->lock) which may be released and regrabbed
2024  * multiple times.  Does GFP_KERNEL allocations.
2025  */
2026 static bool manage_workers(struct worker *worker)
2027 {
2028 	struct worker_pool *pool = worker->pool;
2029 	bool ret = false;
2030 
2031 	/*
2032 	 * Managership is governed by two mutexes - manager_arb and
2033 	 * manager_mutex.  manager_arb handles arbitration of manager role.
2034 	 * Anyone who successfully grabs manager_arb wins the arbitration
2035 	 * and becomes the manager.  mutex_trylock() on pool->manager_arb
2036 	 * failure while holding pool->lock reliably indicates that someone
2037 	 * else is managing the pool and the worker which failed trylock
2038 	 * can proceed to executing work items.  This means that anyone
2039 	 * grabbing manager_arb is responsible for actually performing
2040 	 * manager duties.  If manager_arb is grabbed and released without
2041 	 * actual management, the pool may stall indefinitely.
2042 	 *
2043 	 * manager_mutex is used for exclusion of actual management
2044 	 * operations.  The holder of manager_mutex can be sure that none
2045 	 * of management operations, including creation and destruction of
2046 	 * workers, won't take place until the mutex is released.  Because
2047 	 * manager_mutex doesn't interfere with manager role arbitration,
2048 	 * it is guaranteed that the pool's management, while may be
2049 	 * delayed, won't be disturbed by someone else grabbing
2050 	 * manager_mutex.
2051 	 */
2052 	if (!mutex_trylock(&pool->manager_arb))
2053 		return ret;
2054 
2055 	/*
2056 	 * With manager arbitration won, manager_mutex would be free in
2057 	 * most cases.  trylock first without dropping @pool->lock.
2058 	 */
2059 	if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2060 		spin_unlock_irq(&pool->lock);
2061 		mutex_lock(&pool->manager_mutex);
2062 		ret = true;
2063 	}
2064 
2065 	pool->flags &= ~POOL_MANAGE_WORKERS;
2066 
2067 	/*
2068 	 * Destroy and then create so that may_start_working() is true
2069 	 * on return.
2070 	 */
2071 	ret |= maybe_destroy_workers(pool);
2072 	ret |= maybe_create_worker(pool);
2073 
2074 	mutex_unlock(&pool->manager_mutex);
2075 	mutex_unlock(&pool->manager_arb);
2076 	return ret;
2077 }
2078 
2079 /**
2080  * process_one_work - process single work
2081  * @worker: self
2082  * @work: work to process
2083  *
2084  * Process @work.  This function contains all the logics necessary to
2085  * process a single work including synchronization against and
2086  * interaction with other workers on the same cpu, queueing and
2087  * flushing.  As long as context requirement is met, any worker can
2088  * call this function to process a work.
2089  *
2090  * CONTEXT:
2091  * spin_lock_irq(pool->lock) which is released and regrabbed.
2092  */
2093 static void process_one_work(struct worker *worker, struct work_struct *work)
2094 __releases(&pool->lock)
2095 __acquires(&pool->lock)
2096 {
2097 	struct pool_workqueue *pwq = get_work_pwq(work);
2098 	struct worker_pool *pool = worker->pool;
2099 	bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2100 	int work_color;
2101 	struct worker *collision;
2102 #ifdef CONFIG_LOCKDEP
2103 	/*
2104 	 * It is permissible to free the struct work_struct from
2105 	 * inside the function that is called from it, this we need to
2106 	 * take into account for lockdep too.  To avoid bogus "held
2107 	 * lock freed" warnings as well as problems when looking into
2108 	 * work->lockdep_map, make a copy and use that here.
2109 	 */
2110 	struct lockdep_map lockdep_map;
2111 
2112 	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2113 #endif
2114 	/*
2115 	 * Ensure we're on the correct CPU.  DISASSOCIATED test is
2116 	 * necessary to avoid spurious warnings from rescuers servicing the
2117 	 * unbound or a disassociated pool.
2118 	 */
2119 	WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2120 		     !(pool->flags & POOL_DISASSOCIATED) &&
2121 		     raw_smp_processor_id() != pool->cpu);
2122 
2123 	/*
2124 	 * A single work shouldn't be executed concurrently by
2125 	 * multiple workers on a single cpu.  Check whether anyone is
2126 	 * already processing the work.  If so, defer the work to the
2127 	 * currently executing one.
2128 	 */
2129 	collision = find_worker_executing_work(pool, work);
2130 	if (unlikely(collision)) {
2131 		move_linked_works(work, &collision->scheduled, NULL);
2132 		return;
2133 	}
2134 
2135 	/* claim and dequeue */
2136 	debug_work_deactivate(work);
2137 	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2138 	worker->current_work = work;
2139 	worker->current_func = work->func;
2140 	worker->current_pwq = pwq;
2141 	work_color = get_work_color(work);
2142 
2143 	list_del_init(&work->entry);
2144 
2145 	/*
2146 	 * CPU intensive works don't participate in concurrency
2147 	 * management.  They're the scheduler's responsibility.
2148 	 */
2149 	if (unlikely(cpu_intensive))
2150 		worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2151 
2152 	/*
2153 	 * Unbound pool isn't concurrency managed and work items should be
2154 	 * executed ASAP.  Wake up another worker if necessary.
2155 	 */
2156 	if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2157 		wake_up_worker(pool);
2158 
2159 	/*
2160 	 * Record the last pool and clear PENDING which should be the last
2161 	 * update to @work.  Also, do this inside @pool->lock so that
2162 	 * PENDING and queued state changes happen together while IRQ is
2163 	 * disabled.
2164 	 */
2165 	set_work_pool_and_clear_pending(work, pool->id);
2166 
2167 	spin_unlock_irq(&pool->lock);
2168 
2169 	lock_map_acquire_read(&pwq->wq->lockdep_map);
2170 	lock_map_acquire(&lockdep_map);
2171 	trace_workqueue_execute_start(work);
2172 	worker->current_func(work);
2173 	/*
2174 	 * While we must be careful to not use "work" after this, the trace
2175 	 * point will only record its address.
2176 	 */
2177 	trace_workqueue_execute_end(work);
2178 	lock_map_release(&lockdep_map);
2179 	lock_map_release(&pwq->wq->lockdep_map);
2180 
2181 	if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2182 		pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2183 		       "     last function: %pf\n",
2184 		       current->comm, preempt_count(), task_pid_nr(current),
2185 		       worker->current_func);
2186 		debug_show_held_locks(current);
2187 		dump_stack();
2188 	}
2189 
2190 	spin_lock_irq(&pool->lock);
2191 
2192 	/* clear cpu intensive status */
2193 	if (unlikely(cpu_intensive))
2194 		worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2195 
2196 	/* we're done with it, release */
2197 	hash_del(&worker->hentry);
2198 	worker->current_work = NULL;
2199 	worker->current_func = NULL;
2200 	worker->current_pwq = NULL;
2201 	worker->desc_valid = false;
2202 	pwq_dec_nr_in_flight(pwq, work_color);
2203 }
2204 
2205 /**
2206  * process_scheduled_works - process scheduled works
2207  * @worker: self
2208  *
2209  * Process all scheduled works.  Please note that the scheduled list
2210  * may change while processing a work, so this function repeatedly
2211  * fetches a work from the top and executes it.
2212  *
2213  * CONTEXT:
2214  * spin_lock_irq(pool->lock) which may be released and regrabbed
2215  * multiple times.
2216  */
2217 static void process_scheduled_works(struct worker *worker)
2218 {
2219 	while (!list_empty(&worker->scheduled)) {
2220 		struct work_struct *work = list_first_entry(&worker->scheduled,
2221 						struct work_struct, entry);
2222 		process_one_work(worker, work);
2223 	}
2224 }
2225 
2226 /**
2227  * worker_thread - the worker thread function
2228  * @__worker: self
2229  *
2230  * The worker thread function.  All workers belong to a worker_pool -
2231  * either a per-cpu one or dynamic unbound one.  These workers process all
2232  * work items regardless of their specific target workqueue.  The only
2233  * exception is work items which belong to workqueues with a rescuer which
2234  * will be explained in rescuer_thread().
2235  */
2236 static int worker_thread(void *__worker)
2237 {
2238 	struct worker *worker = __worker;
2239 	struct worker_pool *pool = worker->pool;
2240 
2241 	/* tell the scheduler that this is a workqueue worker */
2242 	worker->task->flags |= PF_WQ_WORKER;
2243 woke_up:
2244 	spin_lock_irq(&pool->lock);
2245 
2246 	/* am I supposed to die? */
2247 	if (unlikely(worker->flags & WORKER_DIE)) {
2248 		spin_unlock_irq(&pool->lock);
2249 		WARN_ON_ONCE(!list_empty(&worker->entry));
2250 		worker->task->flags &= ~PF_WQ_WORKER;
2251 		return 0;
2252 	}
2253 
2254 	worker_leave_idle(worker);
2255 recheck:
2256 	/* no more worker necessary? */
2257 	if (!need_more_worker(pool))
2258 		goto sleep;
2259 
2260 	/* do we need to manage? */
2261 	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2262 		goto recheck;
2263 
2264 	/*
2265 	 * ->scheduled list can only be filled while a worker is
2266 	 * preparing to process a work or actually processing it.
2267 	 * Make sure nobody diddled with it while I was sleeping.
2268 	 */
2269 	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2270 
2271 	/*
2272 	 * Finish PREP stage.  We're guaranteed to have at least one idle
2273 	 * worker or that someone else has already assumed the manager
2274 	 * role.  This is where @worker starts participating in concurrency
2275 	 * management if applicable and concurrency management is restored
2276 	 * after being rebound.  See rebind_workers() for details.
2277 	 */
2278 	worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2279 
2280 	do {
2281 		struct work_struct *work =
2282 			list_first_entry(&pool->worklist,
2283 					 struct work_struct, entry);
2284 
2285 		if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2286 			/* optimization path, not strictly necessary */
2287 			process_one_work(worker, work);
2288 			if (unlikely(!list_empty(&worker->scheduled)))
2289 				process_scheduled_works(worker);
2290 		} else {
2291 			move_linked_works(work, &worker->scheduled, NULL);
2292 			process_scheduled_works(worker);
2293 		}
2294 	} while (keep_working(pool));
2295 
2296 	worker_set_flags(worker, WORKER_PREP, false);
2297 sleep:
2298 	if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2299 		goto recheck;
2300 
2301 	/*
2302 	 * pool->lock is held and there's no work to process and no need to
2303 	 * manage, sleep.  Workers are woken up only while holding
2304 	 * pool->lock or from local cpu, so setting the current state
2305 	 * before releasing pool->lock is enough to prevent losing any
2306 	 * event.
2307 	 */
2308 	worker_enter_idle(worker);
2309 	__set_current_state(TASK_INTERRUPTIBLE);
2310 	spin_unlock_irq(&pool->lock);
2311 	schedule();
2312 	goto woke_up;
2313 }
2314 
2315 /**
2316  * rescuer_thread - the rescuer thread function
2317  * @__rescuer: self
2318  *
2319  * Workqueue rescuer thread function.  There's one rescuer for each
2320  * workqueue which has WQ_MEM_RECLAIM set.
2321  *
2322  * Regular work processing on a pool may block trying to create a new
2323  * worker which uses GFP_KERNEL allocation which has slight chance of
2324  * developing into deadlock if some works currently on the same queue
2325  * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2326  * the problem rescuer solves.
2327  *
2328  * When such condition is possible, the pool summons rescuers of all
2329  * workqueues which have works queued on the pool and let them process
2330  * those works so that forward progress can be guaranteed.
2331  *
2332  * This should happen rarely.
2333  */
2334 static int rescuer_thread(void *__rescuer)
2335 {
2336 	struct worker *rescuer = __rescuer;
2337 	struct workqueue_struct *wq = rescuer->rescue_wq;
2338 	struct list_head *scheduled = &rescuer->scheduled;
2339 
2340 	set_user_nice(current, RESCUER_NICE_LEVEL);
2341 
2342 	/*
2343 	 * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
2344 	 * doesn't participate in concurrency management.
2345 	 */
2346 	rescuer->task->flags |= PF_WQ_WORKER;
2347 repeat:
2348 	set_current_state(TASK_INTERRUPTIBLE);
2349 
2350 	if (kthread_should_stop()) {
2351 		__set_current_state(TASK_RUNNING);
2352 		rescuer->task->flags &= ~PF_WQ_WORKER;
2353 		return 0;
2354 	}
2355 
2356 	/* see whether any pwq is asking for help */
2357 	spin_lock_irq(&wq_mayday_lock);
2358 
2359 	while (!list_empty(&wq->maydays)) {
2360 		struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2361 					struct pool_workqueue, mayday_node);
2362 		struct worker_pool *pool = pwq->pool;
2363 		struct work_struct *work, *n;
2364 
2365 		__set_current_state(TASK_RUNNING);
2366 		list_del_init(&pwq->mayday_node);
2367 
2368 		spin_unlock_irq(&wq_mayday_lock);
2369 
2370 		/* migrate to the target cpu if possible */
2371 		worker_maybe_bind_and_lock(pool);
2372 		rescuer->pool = pool;
2373 
2374 		/*
2375 		 * Slurp in all works issued via this workqueue and
2376 		 * process'em.
2377 		 */
2378 		WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2379 		list_for_each_entry_safe(work, n, &pool->worklist, entry)
2380 			if (get_work_pwq(work) == pwq)
2381 				move_linked_works(work, scheduled, &n);
2382 
2383 		process_scheduled_works(rescuer);
2384 
2385 		/*
2386 		 * Leave this pool.  If keep_working() is %true, notify a
2387 		 * regular worker; otherwise, we end up with 0 concurrency
2388 		 * and stalling the execution.
2389 		 */
2390 		if (keep_working(pool))
2391 			wake_up_worker(pool);
2392 
2393 		rescuer->pool = NULL;
2394 		spin_unlock(&pool->lock);
2395 		spin_lock(&wq_mayday_lock);
2396 	}
2397 
2398 	spin_unlock_irq(&wq_mayday_lock);
2399 
2400 	/* rescuers should never participate in concurrency management */
2401 	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2402 	schedule();
2403 	goto repeat;
2404 }
2405 
2406 struct wq_barrier {
2407 	struct work_struct	work;
2408 	struct completion	done;
2409 };
2410 
2411 static void wq_barrier_func(struct work_struct *work)
2412 {
2413 	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2414 	complete(&barr->done);
2415 }
2416 
2417 /**
2418  * insert_wq_barrier - insert a barrier work
2419  * @pwq: pwq to insert barrier into
2420  * @barr: wq_barrier to insert
2421  * @target: target work to attach @barr to
2422  * @worker: worker currently executing @target, NULL if @target is not executing
2423  *
2424  * @barr is linked to @target such that @barr is completed only after
2425  * @target finishes execution.  Please note that the ordering
2426  * guarantee is observed only with respect to @target and on the local
2427  * cpu.
2428  *
2429  * Currently, a queued barrier can't be canceled.  This is because
2430  * try_to_grab_pending() can't determine whether the work to be
2431  * grabbed is at the head of the queue and thus can't clear LINKED
2432  * flag of the previous work while there must be a valid next work
2433  * after a work with LINKED flag set.
2434  *
2435  * Note that when @worker is non-NULL, @target may be modified
2436  * underneath us, so we can't reliably determine pwq from @target.
2437  *
2438  * CONTEXT:
2439  * spin_lock_irq(pool->lock).
2440  */
2441 static void insert_wq_barrier(struct pool_workqueue *pwq,
2442 			      struct wq_barrier *barr,
2443 			      struct work_struct *target, struct worker *worker)
2444 {
2445 	struct list_head *head;
2446 	unsigned int linked = 0;
2447 
2448 	/*
2449 	 * debugobject calls are safe here even with pool->lock locked
2450 	 * as we know for sure that this will not trigger any of the
2451 	 * checks and call back into the fixup functions where we
2452 	 * might deadlock.
2453 	 */
2454 	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2455 	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2456 	init_completion(&barr->done);
2457 
2458 	/*
2459 	 * If @target is currently being executed, schedule the
2460 	 * barrier to the worker; otherwise, put it after @target.
2461 	 */
2462 	if (worker)
2463 		head = worker->scheduled.next;
2464 	else {
2465 		unsigned long *bits = work_data_bits(target);
2466 
2467 		head = target->entry.next;
2468 		/* there can already be other linked works, inherit and set */
2469 		linked = *bits & WORK_STRUCT_LINKED;
2470 		__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2471 	}
2472 
2473 	debug_work_activate(&barr->work);
2474 	insert_work(pwq, &barr->work, head,
2475 		    work_color_to_flags(WORK_NO_COLOR) | linked);
2476 }
2477 
2478 /**
2479  * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2480  * @wq: workqueue being flushed
2481  * @flush_color: new flush color, < 0 for no-op
2482  * @work_color: new work color, < 0 for no-op
2483  *
2484  * Prepare pwqs for workqueue flushing.
2485  *
2486  * If @flush_color is non-negative, flush_color on all pwqs should be
2487  * -1.  If no pwq has in-flight commands at the specified color, all
2488  * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
2489  * has in flight commands, its pwq->flush_color is set to
2490  * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2491  * wakeup logic is armed and %true is returned.
2492  *
2493  * The caller should have initialized @wq->first_flusher prior to
2494  * calling this function with non-negative @flush_color.  If
2495  * @flush_color is negative, no flush color update is done and %false
2496  * is returned.
2497  *
2498  * If @work_color is non-negative, all pwqs should have the same
2499  * work_color which is previous to @work_color and all will be
2500  * advanced to @work_color.
2501  *
2502  * CONTEXT:
2503  * mutex_lock(wq->mutex).
2504  *
2505  * RETURNS:
2506  * %true if @flush_color >= 0 and there's something to flush.  %false
2507  * otherwise.
2508  */
2509 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2510 				      int flush_color, int work_color)
2511 {
2512 	bool wait = false;
2513 	struct pool_workqueue *pwq;
2514 
2515 	if (flush_color >= 0) {
2516 		WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2517 		atomic_set(&wq->nr_pwqs_to_flush, 1);
2518 	}
2519 
2520 	for_each_pwq(pwq, wq) {
2521 		struct worker_pool *pool = pwq->pool;
2522 
2523 		spin_lock_irq(&pool->lock);
2524 
2525 		if (flush_color >= 0) {
2526 			WARN_ON_ONCE(pwq->flush_color != -1);
2527 
2528 			if (pwq->nr_in_flight[flush_color]) {
2529 				pwq->flush_color = flush_color;
2530 				atomic_inc(&wq->nr_pwqs_to_flush);
2531 				wait = true;
2532 			}
2533 		}
2534 
2535 		if (work_color >= 0) {
2536 			WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2537 			pwq->work_color = work_color;
2538 		}
2539 
2540 		spin_unlock_irq(&pool->lock);
2541 	}
2542 
2543 	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2544 		complete(&wq->first_flusher->done);
2545 
2546 	return wait;
2547 }
2548 
2549 /**
2550  * flush_workqueue - ensure that any scheduled work has run to completion.
2551  * @wq: workqueue to flush
2552  *
2553  * This function sleeps until all work items which were queued on entry
2554  * have finished execution, but it is not livelocked by new incoming ones.
2555  */
2556 void flush_workqueue(struct workqueue_struct *wq)
2557 {
2558 	struct wq_flusher this_flusher = {
2559 		.list = LIST_HEAD_INIT(this_flusher.list),
2560 		.flush_color = -1,
2561 		.done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2562 	};
2563 	int next_color;
2564 
2565 	lock_map_acquire(&wq->lockdep_map);
2566 	lock_map_release(&wq->lockdep_map);
2567 
2568 	mutex_lock(&wq->mutex);
2569 
2570 	/*
2571 	 * Start-to-wait phase
2572 	 */
2573 	next_color = work_next_color(wq->work_color);
2574 
2575 	if (next_color != wq->flush_color) {
2576 		/*
2577 		 * Color space is not full.  The current work_color
2578 		 * becomes our flush_color and work_color is advanced
2579 		 * by one.
2580 		 */
2581 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2582 		this_flusher.flush_color = wq->work_color;
2583 		wq->work_color = next_color;
2584 
2585 		if (!wq->first_flusher) {
2586 			/* no flush in progress, become the first flusher */
2587 			WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2588 
2589 			wq->first_flusher = &this_flusher;
2590 
2591 			if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2592 						       wq->work_color)) {
2593 				/* nothing to flush, done */
2594 				wq->flush_color = next_color;
2595 				wq->first_flusher = NULL;
2596 				goto out_unlock;
2597 			}
2598 		} else {
2599 			/* wait in queue */
2600 			WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2601 			list_add_tail(&this_flusher.list, &wq->flusher_queue);
2602 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2603 		}
2604 	} else {
2605 		/*
2606 		 * Oops, color space is full, wait on overflow queue.
2607 		 * The next flush completion will assign us
2608 		 * flush_color and transfer to flusher_queue.
2609 		 */
2610 		list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2611 	}
2612 
2613 	mutex_unlock(&wq->mutex);
2614 
2615 	wait_for_completion(&this_flusher.done);
2616 
2617 	/*
2618 	 * Wake-up-and-cascade phase
2619 	 *
2620 	 * First flushers are responsible for cascading flushes and
2621 	 * handling overflow.  Non-first flushers can simply return.
2622 	 */
2623 	if (wq->first_flusher != &this_flusher)
2624 		return;
2625 
2626 	mutex_lock(&wq->mutex);
2627 
2628 	/* we might have raced, check again with mutex held */
2629 	if (wq->first_flusher != &this_flusher)
2630 		goto out_unlock;
2631 
2632 	wq->first_flusher = NULL;
2633 
2634 	WARN_ON_ONCE(!list_empty(&this_flusher.list));
2635 	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2636 
2637 	while (true) {
2638 		struct wq_flusher *next, *tmp;
2639 
2640 		/* complete all the flushers sharing the current flush color */
2641 		list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2642 			if (next->flush_color != wq->flush_color)
2643 				break;
2644 			list_del_init(&next->list);
2645 			complete(&next->done);
2646 		}
2647 
2648 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2649 			     wq->flush_color != work_next_color(wq->work_color));
2650 
2651 		/* this flush_color is finished, advance by one */
2652 		wq->flush_color = work_next_color(wq->flush_color);
2653 
2654 		/* one color has been freed, handle overflow queue */
2655 		if (!list_empty(&wq->flusher_overflow)) {
2656 			/*
2657 			 * Assign the same color to all overflowed
2658 			 * flushers, advance work_color and append to
2659 			 * flusher_queue.  This is the start-to-wait
2660 			 * phase for these overflowed flushers.
2661 			 */
2662 			list_for_each_entry(tmp, &wq->flusher_overflow, list)
2663 				tmp->flush_color = wq->work_color;
2664 
2665 			wq->work_color = work_next_color(wq->work_color);
2666 
2667 			list_splice_tail_init(&wq->flusher_overflow,
2668 					      &wq->flusher_queue);
2669 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2670 		}
2671 
2672 		if (list_empty(&wq->flusher_queue)) {
2673 			WARN_ON_ONCE(wq->flush_color != wq->work_color);
2674 			break;
2675 		}
2676 
2677 		/*
2678 		 * Need to flush more colors.  Make the next flusher
2679 		 * the new first flusher and arm pwqs.
2680 		 */
2681 		WARN_ON_ONCE(wq->flush_color == wq->work_color);
2682 		WARN_ON_ONCE(wq->flush_color != next->flush_color);
2683 
2684 		list_del_init(&next->list);
2685 		wq->first_flusher = next;
2686 
2687 		if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2688 			break;
2689 
2690 		/*
2691 		 * Meh... this color is already done, clear first
2692 		 * flusher and repeat cascading.
2693 		 */
2694 		wq->first_flusher = NULL;
2695 	}
2696 
2697 out_unlock:
2698 	mutex_unlock(&wq->mutex);
2699 }
2700 EXPORT_SYMBOL_GPL(flush_workqueue);
2701 
2702 /**
2703  * drain_workqueue - drain a workqueue
2704  * @wq: workqueue to drain
2705  *
2706  * Wait until the workqueue becomes empty.  While draining is in progress,
2707  * only chain queueing is allowed.  IOW, only currently pending or running
2708  * work items on @wq can queue further work items on it.  @wq is flushed
2709  * repeatedly until it becomes empty.  The number of flushing is detemined
2710  * by the depth of chaining and should be relatively short.  Whine if it
2711  * takes too long.
2712  */
2713 void drain_workqueue(struct workqueue_struct *wq)
2714 {
2715 	unsigned int flush_cnt = 0;
2716 	struct pool_workqueue *pwq;
2717 
2718 	/*
2719 	 * __queue_work() needs to test whether there are drainers, is much
2720 	 * hotter than drain_workqueue() and already looks at @wq->flags.
2721 	 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2722 	 */
2723 	mutex_lock(&wq->mutex);
2724 	if (!wq->nr_drainers++)
2725 		wq->flags |= __WQ_DRAINING;
2726 	mutex_unlock(&wq->mutex);
2727 reflush:
2728 	flush_workqueue(wq);
2729 
2730 	mutex_lock(&wq->mutex);
2731 
2732 	for_each_pwq(pwq, wq) {
2733 		bool drained;
2734 
2735 		spin_lock_irq(&pwq->pool->lock);
2736 		drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2737 		spin_unlock_irq(&pwq->pool->lock);
2738 
2739 		if (drained)
2740 			continue;
2741 
2742 		if (++flush_cnt == 10 ||
2743 		    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2744 			pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2745 				wq->name, flush_cnt);
2746 
2747 		mutex_unlock(&wq->mutex);
2748 		goto reflush;
2749 	}
2750 
2751 	if (!--wq->nr_drainers)
2752 		wq->flags &= ~__WQ_DRAINING;
2753 	mutex_unlock(&wq->mutex);
2754 }
2755 EXPORT_SYMBOL_GPL(drain_workqueue);
2756 
2757 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2758 {
2759 	struct worker *worker = NULL;
2760 	struct worker_pool *pool;
2761 	struct pool_workqueue *pwq;
2762 
2763 	might_sleep();
2764 
2765 	local_irq_disable();
2766 	pool = get_work_pool(work);
2767 	if (!pool) {
2768 		local_irq_enable();
2769 		return false;
2770 	}
2771 
2772 	spin_lock(&pool->lock);
2773 	/* see the comment in try_to_grab_pending() with the same code */
2774 	pwq = get_work_pwq(work);
2775 	if (pwq) {
2776 		if (unlikely(pwq->pool != pool))
2777 			goto already_gone;
2778 	} else {
2779 		worker = find_worker_executing_work(pool, work);
2780 		if (!worker)
2781 			goto already_gone;
2782 		pwq = worker->current_pwq;
2783 	}
2784 
2785 	insert_wq_barrier(pwq, barr, work, worker);
2786 	spin_unlock_irq(&pool->lock);
2787 
2788 	/*
2789 	 * If @max_active is 1 or rescuer is in use, flushing another work
2790 	 * item on the same workqueue may lead to deadlock.  Make sure the
2791 	 * flusher is not running on the same workqueue by verifying write
2792 	 * access.
2793 	 */
2794 	if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2795 		lock_map_acquire(&pwq->wq->lockdep_map);
2796 	else
2797 		lock_map_acquire_read(&pwq->wq->lockdep_map);
2798 	lock_map_release(&pwq->wq->lockdep_map);
2799 
2800 	return true;
2801 already_gone:
2802 	spin_unlock_irq(&pool->lock);
2803 	return false;
2804 }
2805 
2806 /**
2807  * flush_work - wait for a work to finish executing the last queueing instance
2808  * @work: the work to flush
2809  *
2810  * Wait until @work has finished execution.  @work is guaranteed to be idle
2811  * on return if it hasn't been requeued since flush started.
2812  *
2813  * RETURNS:
2814  * %true if flush_work() waited for the work to finish execution,
2815  * %false if it was already idle.
2816  */
2817 bool flush_work(struct work_struct *work)
2818 {
2819 	struct wq_barrier barr;
2820 
2821 	lock_map_acquire(&work->lockdep_map);
2822 	lock_map_release(&work->lockdep_map);
2823 
2824 	if (start_flush_work(work, &barr)) {
2825 		wait_for_completion(&barr.done);
2826 		destroy_work_on_stack(&barr.work);
2827 		return true;
2828 	} else {
2829 		return false;
2830 	}
2831 }
2832 EXPORT_SYMBOL_GPL(flush_work);
2833 
2834 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2835 {
2836 	unsigned long flags;
2837 	int ret;
2838 
2839 	do {
2840 		ret = try_to_grab_pending(work, is_dwork, &flags);
2841 		/*
2842 		 * If someone else is canceling, wait for the same event it
2843 		 * would be waiting for before retrying.
2844 		 */
2845 		if (unlikely(ret == -ENOENT))
2846 			flush_work(work);
2847 	} while (unlikely(ret < 0));
2848 
2849 	/* tell other tasks trying to grab @work to back off */
2850 	mark_work_canceling(work);
2851 	local_irq_restore(flags);
2852 
2853 	flush_work(work);
2854 	clear_work_data(work);
2855 	return ret;
2856 }
2857 
2858 /**
2859  * cancel_work_sync - cancel a work and wait for it to finish
2860  * @work: the work to cancel
2861  *
2862  * Cancel @work and wait for its execution to finish.  This function
2863  * can be used even if the work re-queues itself or migrates to
2864  * another workqueue.  On return from this function, @work is
2865  * guaranteed to be not pending or executing on any CPU.
2866  *
2867  * cancel_work_sync(&delayed_work->work) must not be used for
2868  * delayed_work's.  Use cancel_delayed_work_sync() instead.
2869  *
2870  * The caller must ensure that the workqueue on which @work was last
2871  * queued can't be destroyed before this function returns.
2872  *
2873  * RETURNS:
2874  * %true if @work was pending, %false otherwise.
2875  */
2876 bool cancel_work_sync(struct work_struct *work)
2877 {
2878 	return __cancel_work_timer(work, false);
2879 }
2880 EXPORT_SYMBOL_GPL(cancel_work_sync);
2881 
2882 /**
2883  * flush_delayed_work - wait for a dwork to finish executing the last queueing
2884  * @dwork: the delayed work to flush
2885  *
2886  * Delayed timer is cancelled and the pending work is queued for
2887  * immediate execution.  Like flush_work(), this function only
2888  * considers the last queueing instance of @dwork.
2889  *
2890  * RETURNS:
2891  * %true if flush_work() waited for the work to finish execution,
2892  * %false if it was already idle.
2893  */
2894 bool flush_delayed_work(struct delayed_work *dwork)
2895 {
2896 	local_irq_disable();
2897 	if (del_timer_sync(&dwork->timer))
2898 		__queue_work(dwork->cpu, dwork->wq, &dwork->work);
2899 	local_irq_enable();
2900 	return flush_work(&dwork->work);
2901 }
2902 EXPORT_SYMBOL(flush_delayed_work);
2903 
2904 /**
2905  * cancel_delayed_work - cancel a delayed work
2906  * @dwork: delayed_work to cancel
2907  *
2908  * Kill off a pending delayed_work.  Returns %true if @dwork was pending
2909  * and canceled; %false if wasn't pending.  Note that the work callback
2910  * function may still be running on return, unless it returns %true and the
2911  * work doesn't re-arm itself.  Explicitly flush or use
2912  * cancel_delayed_work_sync() to wait on it.
2913  *
2914  * This function is safe to call from any context including IRQ handler.
2915  */
2916 bool cancel_delayed_work(struct delayed_work *dwork)
2917 {
2918 	unsigned long flags;
2919 	int ret;
2920 
2921 	do {
2922 		ret = try_to_grab_pending(&dwork->work, true, &flags);
2923 	} while (unlikely(ret == -EAGAIN));
2924 
2925 	if (unlikely(ret < 0))
2926 		return false;
2927 
2928 	set_work_pool_and_clear_pending(&dwork->work,
2929 					get_work_pool_id(&dwork->work));
2930 	local_irq_restore(flags);
2931 	return ret;
2932 }
2933 EXPORT_SYMBOL(cancel_delayed_work);
2934 
2935 /**
2936  * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2937  * @dwork: the delayed work cancel
2938  *
2939  * This is cancel_work_sync() for delayed works.
2940  *
2941  * RETURNS:
2942  * %true if @dwork was pending, %false otherwise.
2943  */
2944 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2945 {
2946 	return __cancel_work_timer(&dwork->work, true);
2947 }
2948 EXPORT_SYMBOL(cancel_delayed_work_sync);
2949 
2950 /**
2951  * schedule_on_each_cpu - execute a function synchronously on each online CPU
2952  * @func: the function to call
2953  *
2954  * schedule_on_each_cpu() executes @func on each online CPU using the
2955  * system workqueue and blocks until all CPUs have completed.
2956  * schedule_on_each_cpu() is very slow.
2957  *
2958  * RETURNS:
2959  * 0 on success, -errno on failure.
2960  */
2961 int schedule_on_each_cpu(work_func_t func)
2962 {
2963 	int cpu;
2964 	struct work_struct __percpu *works;
2965 
2966 	works = alloc_percpu(struct work_struct);
2967 	if (!works)
2968 		return -ENOMEM;
2969 
2970 	get_online_cpus();
2971 
2972 	for_each_online_cpu(cpu) {
2973 		struct work_struct *work = per_cpu_ptr(works, cpu);
2974 
2975 		INIT_WORK(work, func);
2976 		schedule_work_on(cpu, work);
2977 	}
2978 
2979 	for_each_online_cpu(cpu)
2980 		flush_work(per_cpu_ptr(works, cpu));
2981 
2982 	put_online_cpus();
2983 	free_percpu(works);
2984 	return 0;
2985 }
2986 
2987 /**
2988  * flush_scheduled_work - ensure that any scheduled work has run to completion.
2989  *
2990  * Forces execution of the kernel-global workqueue and blocks until its
2991  * completion.
2992  *
2993  * Think twice before calling this function!  It's very easy to get into
2994  * trouble if you don't take great care.  Either of the following situations
2995  * will lead to deadlock:
2996  *
2997  *	One of the work items currently on the workqueue needs to acquire
2998  *	a lock held by your code or its caller.
2999  *
3000  *	Your code is running in the context of a work routine.
3001  *
3002  * They will be detected by lockdep when they occur, but the first might not
3003  * occur very often.  It depends on what work items are on the workqueue and
3004  * what locks they need, which you have no control over.
3005  *
3006  * In most situations flushing the entire workqueue is overkill; you merely
3007  * need to know that a particular work item isn't queued and isn't running.
3008  * In such cases you should use cancel_delayed_work_sync() or
3009  * cancel_work_sync() instead.
3010  */
3011 void flush_scheduled_work(void)
3012 {
3013 	flush_workqueue(system_wq);
3014 }
3015 EXPORT_SYMBOL(flush_scheduled_work);
3016 
3017 /**
3018  * execute_in_process_context - reliably execute the routine with user context
3019  * @fn:		the function to execute
3020  * @ew:		guaranteed storage for the execute work structure (must
3021  *		be available when the work executes)
3022  *
3023  * Executes the function immediately if process context is available,
3024  * otherwise schedules the function for delayed execution.
3025  *
3026  * Returns:	0 - function was executed
3027  *		1 - function was scheduled for execution
3028  */
3029 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3030 {
3031 	if (!in_interrupt()) {
3032 		fn(&ew->work);
3033 		return 0;
3034 	}
3035 
3036 	INIT_WORK(&ew->work, fn);
3037 	schedule_work(&ew->work);
3038 
3039 	return 1;
3040 }
3041 EXPORT_SYMBOL_GPL(execute_in_process_context);
3042 
3043 #ifdef CONFIG_SYSFS
3044 /*
3045  * Workqueues with WQ_SYSFS flag set is visible to userland via
3046  * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
3047  * following attributes.
3048  *
3049  *  per_cpu	RO bool	: whether the workqueue is per-cpu or unbound
3050  *  max_active	RW int	: maximum number of in-flight work items
3051  *
3052  * Unbound workqueues have the following extra attributes.
3053  *
3054  *  id		RO int	: the associated pool ID
3055  *  nice	RW int	: nice value of the workers
3056  *  cpumask	RW mask	: bitmask of allowed CPUs for the workers
3057  */
3058 struct wq_device {
3059 	struct workqueue_struct		*wq;
3060 	struct device			dev;
3061 };
3062 
3063 static struct workqueue_struct *dev_to_wq(struct device *dev)
3064 {
3065 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3066 
3067 	return wq_dev->wq;
3068 }
3069 
3070 static ssize_t wq_per_cpu_show(struct device *dev,
3071 			       struct device_attribute *attr, char *buf)
3072 {
3073 	struct workqueue_struct *wq = dev_to_wq(dev);
3074 
3075 	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3076 }
3077 
3078 static ssize_t wq_max_active_show(struct device *dev,
3079 				  struct device_attribute *attr, char *buf)
3080 {
3081 	struct workqueue_struct *wq = dev_to_wq(dev);
3082 
3083 	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3084 }
3085 
3086 static ssize_t wq_max_active_store(struct device *dev,
3087 				   struct device_attribute *attr,
3088 				   const char *buf, size_t count)
3089 {
3090 	struct workqueue_struct *wq = dev_to_wq(dev);
3091 	int val;
3092 
3093 	if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3094 		return -EINVAL;
3095 
3096 	workqueue_set_max_active(wq, val);
3097 	return count;
3098 }
3099 
3100 static struct device_attribute wq_sysfs_attrs[] = {
3101 	__ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3102 	__ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3103 	__ATTR_NULL,
3104 };
3105 
3106 static ssize_t wq_pool_ids_show(struct device *dev,
3107 				struct device_attribute *attr, char *buf)
3108 {
3109 	struct workqueue_struct *wq = dev_to_wq(dev);
3110 	const char *delim = "";
3111 	int node, written = 0;
3112 
3113 	rcu_read_lock_sched();
3114 	for_each_node(node) {
3115 		written += scnprintf(buf + written, PAGE_SIZE - written,
3116 				     "%s%d:%d", delim, node,
3117 				     unbound_pwq_by_node(wq, node)->pool->id);
3118 		delim = " ";
3119 	}
3120 	written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3121 	rcu_read_unlock_sched();
3122 
3123 	return written;
3124 }
3125 
3126 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3127 			    char *buf)
3128 {
3129 	struct workqueue_struct *wq = dev_to_wq(dev);
3130 	int written;
3131 
3132 	mutex_lock(&wq->mutex);
3133 	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3134 	mutex_unlock(&wq->mutex);
3135 
3136 	return written;
3137 }
3138 
3139 /* prepare workqueue_attrs for sysfs store operations */
3140 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3141 {
3142 	struct workqueue_attrs *attrs;
3143 
3144 	attrs = alloc_workqueue_attrs(GFP_KERNEL);
3145 	if (!attrs)
3146 		return NULL;
3147 
3148 	mutex_lock(&wq->mutex);
3149 	copy_workqueue_attrs(attrs, wq->unbound_attrs);
3150 	mutex_unlock(&wq->mutex);
3151 	return attrs;
3152 }
3153 
3154 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3155 			     const char *buf, size_t count)
3156 {
3157 	struct workqueue_struct *wq = dev_to_wq(dev);
3158 	struct workqueue_attrs *attrs;
3159 	int ret;
3160 
3161 	attrs = wq_sysfs_prep_attrs(wq);
3162 	if (!attrs)
3163 		return -ENOMEM;
3164 
3165 	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3166 	    attrs->nice >= -20 && attrs->nice <= 19)
3167 		ret = apply_workqueue_attrs(wq, attrs);
3168 	else
3169 		ret = -EINVAL;
3170 
3171 	free_workqueue_attrs(attrs);
3172 	return ret ?: count;
3173 }
3174 
3175 static ssize_t wq_cpumask_show(struct device *dev,
3176 			       struct device_attribute *attr, char *buf)
3177 {
3178 	struct workqueue_struct *wq = dev_to_wq(dev);
3179 	int written;
3180 
3181 	mutex_lock(&wq->mutex);
3182 	written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3183 	mutex_unlock(&wq->mutex);
3184 
3185 	written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3186 	return written;
3187 }
3188 
3189 static ssize_t wq_cpumask_store(struct device *dev,
3190 				struct device_attribute *attr,
3191 				const char *buf, size_t count)
3192 {
3193 	struct workqueue_struct *wq = dev_to_wq(dev);
3194 	struct workqueue_attrs *attrs;
3195 	int ret;
3196 
3197 	attrs = wq_sysfs_prep_attrs(wq);
3198 	if (!attrs)
3199 		return -ENOMEM;
3200 
3201 	ret = cpumask_parse(buf, attrs->cpumask);
3202 	if (!ret)
3203 		ret = apply_workqueue_attrs(wq, attrs);
3204 
3205 	free_workqueue_attrs(attrs);
3206 	return ret ?: count;
3207 }
3208 
3209 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3210 			    char *buf)
3211 {
3212 	struct workqueue_struct *wq = dev_to_wq(dev);
3213 	int written;
3214 
3215 	mutex_lock(&wq->mutex);
3216 	written = scnprintf(buf, PAGE_SIZE, "%d\n",
3217 			    !wq->unbound_attrs->no_numa);
3218 	mutex_unlock(&wq->mutex);
3219 
3220 	return written;
3221 }
3222 
3223 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3224 			     const char *buf, size_t count)
3225 {
3226 	struct workqueue_struct *wq = dev_to_wq(dev);
3227 	struct workqueue_attrs *attrs;
3228 	int v, ret;
3229 
3230 	attrs = wq_sysfs_prep_attrs(wq);
3231 	if (!attrs)
3232 		return -ENOMEM;
3233 
3234 	ret = -EINVAL;
3235 	if (sscanf(buf, "%d", &v) == 1) {
3236 		attrs->no_numa = !v;
3237 		ret = apply_workqueue_attrs(wq, attrs);
3238 	}
3239 
3240 	free_workqueue_attrs(attrs);
3241 	return ret ?: count;
3242 }
3243 
3244 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3245 	__ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3246 	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3247 	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3248 	__ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3249 	__ATTR_NULL,
3250 };
3251 
3252 static struct bus_type wq_subsys = {
3253 	.name				= "workqueue",
3254 	.dev_attrs			= wq_sysfs_attrs,
3255 };
3256 
3257 static int __init wq_sysfs_init(void)
3258 {
3259 	return subsys_virtual_register(&wq_subsys, NULL);
3260 }
3261 core_initcall(wq_sysfs_init);
3262 
3263 static void wq_device_release(struct device *dev)
3264 {
3265 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3266 
3267 	kfree(wq_dev);
3268 }
3269 
3270 /**
3271  * workqueue_sysfs_register - make a workqueue visible in sysfs
3272  * @wq: the workqueue to register
3273  *
3274  * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3275  * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3276  * which is the preferred method.
3277  *
3278  * Workqueue user should use this function directly iff it wants to apply
3279  * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3280  * apply_workqueue_attrs() may race against userland updating the
3281  * attributes.
3282  *
3283  * Returns 0 on success, -errno on failure.
3284  */
3285 int workqueue_sysfs_register(struct workqueue_struct *wq)
3286 {
3287 	struct wq_device *wq_dev;
3288 	int ret;
3289 
3290 	/*
3291 	 * Adjusting max_active or creating new pwqs by applyting
3292 	 * attributes breaks ordering guarantee.  Disallow exposing ordered
3293 	 * workqueues.
3294 	 */
3295 	if (WARN_ON(wq->flags & __WQ_ORDERED))
3296 		return -EINVAL;
3297 
3298 	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3299 	if (!wq_dev)
3300 		return -ENOMEM;
3301 
3302 	wq_dev->wq = wq;
3303 	wq_dev->dev.bus = &wq_subsys;
3304 	wq_dev->dev.init_name = wq->name;
3305 	wq_dev->dev.release = wq_device_release;
3306 
3307 	/*
3308 	 * unbound_attrs are created separately.  Suppress uevent until
3309 	 * everything is ready.
3310 	 */
3311 	dev_set_uevent_suppress(&wq_dev->dev, true);
3312 
3313 	ret = device_register(&wq_dev->dev);
3314 	if (ret) {
3315 		kfree(wq_dev);
3316 		wq->wq_dev = NULL;
3317 		return ret;
3318 	}
3319 
3320 	if (wq->flags & WQ_UNBOUND) {
3321 		struct device_attribute *attr;
3322 
3323 		for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3324 			ret = device_create_file(&wq_dev->dev, attr);
3325 			if (ret) {
3326 				device_unregister(&wq_dev->dev);
3327 				wq->wq_dev = NULL;
3328 				return ret;
3329 			}
3330 		}
3331 	}
3332 
3333 	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3334 	return 0;
3335 }
3336 
3337 /**
3338  * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3339  * @wq: the workqueue to unregister
3340  *
3341  * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3342  */
3343 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3344 {
3345 	struct wq_device *wq_dev = wq->wq_dev;
3346 
3347 	if (!wq->wq_dev)
3348 		return;
3349 
3350 	wq->wq_dev = NULL;
3351 	device_unregister(&wq_dev->dev);
3352 }
3353 #else	/* CONFIG_SYSFS */
3354 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)	{ }
3355 #endif	/* CONFIG_SYSFS */
3356 
3357 /**
3358  * free_workqueue_attrs - free a workqueue_attrs
3359  * @attrs: workqueue_attrs to free
3360  *
3361  * Undo alloc_workqueue_attrs().
3362  */
3363 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3364 {
3365 	if (attrs) {
3366 		free_cpumask_var(attrs->cpumask);
3367 		kfree(attrs);
3368 	}
3369 }
3370 
3371 /**
3372  * alloc_workqueue_attrs - allocate a workqueue_attrs
3373  * @gfp_mask: allocation mask to use
3374  *
3375  * Allocate a new workqueue_attrs, initialize with default settings and
3376  * return it.  Returns NULL on failure.
3377  */
3378 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3379 {
3380 	struct workqueue_attrs *attrs;
3381 
3382 	attrs = kzalloc(sizeof(*attrs), gfp_mask);
3383 	if (!attrs)
3384 		goto fail;
3385 	if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3386 		goto fail;
3387 
3388 	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3389 	return attrs;
3390 fail:
3391 	free_workqueue_attrs(attrs);
3392 	return NULL;
3393 }
3394 
3395 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3396 				 const struct workqueue_attrs *from)
3397 {
3398 	to->nice = from->nice;
3399 	cpumask_copy(to->cpumask, from->cpumask);
3400 }
3401 
3402 /* hash value of the content of @attr */
3403 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3404 {
3405 	u32 hash = 0;
3406 
3407 	hash = jhash_1word(attrs->nice, hash);
3408 	hash = jhash(cpumask_bits(attrs->cpumask),
3409 		     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3410 	return hash;
3411 }
3412 
3413 /* content equality test */
3414 static bool wqattrs_equal(const struct workqueue_attrs *a,
3415 			  const struct workqueue_attrs *b)
3416 {
3417 	if (a->nice != b->nice)
3418 		return false;
3419 	if (!cpumask_equal(a->cpumask, b->cpumask))
3420 		return false;
3421 	return true;
3422 }
3423 
3424 /**
3425  * init_worker_pool - initialize a newly zalloc'd worker_pool
3426  * @pool: worker_pool to initialize
3427  *
3428  * Initiailize a newly zalloc'd @pool.  It also allocates @pool->attrs.
3429  * Returns 0 on success, -errno on failure.  Even on failure, all fields
3430  * inside @pool proper are initialized and put_unbound_pool() can be called
3431  * on @pool safely to release it.
3432  */
3433 static int init_worker_pool(struct worker_pool *pool)
3434 {
3435 	spin_lock_init(&pool->lock);
3436 	pool->id = -1;
3437 	pool->cpu = -1;
3438 	pool->node = NUMA_NO_NODE;
3439 	pool->flags |= POOL_DISASSOCIATED;
3440 	INIT_LIST_HEAD(&pool->worklist);
3441 	INIT_LIST_HEAD(&pool->idle_list);
3442 	hash_init(pool->busy_hash);
3443 
3444 	init_timer_deferrable(&pool->idle_timer);
3445 	pool->idle_timer.function = idle_worker_timeout;
3446 	pool->idle_timer.data = (unsigned long)pool;
3447 
3448 	setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3449 		    (unsigned long)pool);
3450 
3451 	mutex_init(&pool->manager_arb);
3452 	mutex_init(&pool->manager_mutex);
3453 	idr_init(&pool->worker_idr);
3454 
3455 	INIT_HLIST_NODE(&pool->hash_node);
3456 	pool->refcnt = 1;
3457 
3458 	/* shouldn't fail above this point */
3459 	pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3460 	if (!pool->attrs)
3461 		return -ENOMEM;
3462 	return 0;
3463 }
3464 
3465 static void rcu_free_pool(struct rcu_head *rcu)
3466 {
3467 	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3468 
3469 	idr_destroy(&pool->worker_idr);
3470 	free_workqueue_attrs(pool->attrs);
3471 	kfree(pool);
3472 }
3473 
3474 /**
3475  * put_unbound_pool - put a worker_pool
3476  * @pool: worker_pool to put
3477  *
3478  * Put @pool.  If its refcnt reaches zero, it gets destroyed in sched-RCU
3479  * safe manner.  get_unbound_pool() calls this function on its failure path
3480  * and this function should be able to release pools which went through,
3481  * successfully or not, init_worker_pool().
3482  *
3483  * Should be called with wq_pool_mutex held.
3484  */
3485 static void put_unbound_pool(struct worker_pool *pool)
3486 {
3487 	struct worker *worker;
3488 
3489 	lockdep_assert_held(&wq_pool_mutex);
3490 
3491 	if (--pool->refcnt)
3492 		return;
3493 
3494 	/* sanity checks */
3495 	if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3496 	    WARN_ON(!list_empty(&pool->worklist)))
3497 		return;
3498 
3499 	/* release id and unhash */
3500 	if (pool->id >= 0)
3501 		idr_remove(&worker_pool_idr, pool->id);
3502 	hash_del(&pool->hash_node);
3503 
3504 	/*
3505 	 * Become the manager and destroy all workers.  Grabbing
3506 	 * manager_arb prevents @pool's workers from blocking on
3507 	 * manager_mutex.
3508 	 */
3509 	mutex_lock(&pool->manager_arb);
3510 	mutex_lock(&pool->manager_mutex);
3511 	spin_lock_irq(&pool->lock);
3512 
3513 	while ((worker = first_worker(pool)))
3514 		destroy_worker(worker);
3515 	WARN_ON(pool->nr_workers || pool->nr_idle);
3516 
3517 	spin_unlock_irq(&pool->lock);
3518 	mutex_unlock(&pool->manager_mutex);
3519 	mutex_unlock(&pool->manager_arb);
3520 
3521 	/* shut down the timers */
3522 	del_timer_sync(&pool->idle_timer);
3523 	del_timer_sync(&pool->mayday_timer);
3524 
3525 	/* sched-RCU protected to allow dereferences from get_work_pool() */
3526 	call_rcu_sched(&pool->rcu, rcu_free_pool);
3527 }
3528 
3529 /**
3530  * get_unbound_pool - get a worker_pool with the specified attributes
3531  * @attrs: the attributes of the worker_pool to get
3532  *
3533  * Obtain a worker_pool which has the same attributes as @attrs, bump the
3534  * reference count and return it.  If there already is a matching
3535  * worker_pool, it will be used; otherwise, this function attempts to
3536  * create a new one.  On failure, returns NULL.
3537  *
3538  * Should be called with wq_pool_mutex held.
3539  */
3540 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3541 {
3542 	u32 hash = wqattrs_hash(attrs);
3543 	struct worker_pool *pool;
3544 	int node;
3545 
3546 	lockdep_assert_held(&wq_pool_mutex);
3547 
3548 	/* do we already have a matching pool? */
3549 	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3550 		if (wqattrs_equal(pool->attrs, attrs)) {
3551 			pool->refcnt++;
3552 			goto out_unlock;
3553 		}
3554 	}
3555 
3556 	/* nope, create a new one */
3557 	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3558 	if (!pool || init_worker_pool(pool) < 0)
3559 		goto fail;
3560 
3561 	if (workqueue_freezing)
3562 		pool->flags |= POOL_FREEZING;
3563 
3564 	lockdep_set_subclass(&pool->lock, 1);	/* see put_pwq() */
3565 	copy_workqueue_attrs(pool->attrs, attrs);
3566 
3567 	/* if cpumask is contained inside a NUMA node, we belong to that node */
3568 	if (wq_numa_enabled) {
3569 		for_each_node(node) {
3570 			if (cpumask_subset(pool->attrs->cpumask,
3571 					   wq_numa_possible_cpumask[node])) {
3572 				pool->node = node;
3573 				break;
3574 			}
3575 		}
3576 	}
3577 
3578 	if (worker_pool_assign_id(pool) < 0)
3579 		goto fail;
3580 
3581 	/* create and start the initial worker */
3582 	if (create_and_start_worker(pool) < 0)
3583 		goto fail;
3584 
3585 	/* install */
3586 	hash_add(unbound_pool_hash, &pool->hash_node, hash);
3587 out_unlock:
3588 	return pool;
3589 fail:
3590 	if (pool)
3591 		put_unbound_pool(pool);
3592 	return NULL;
3593 }
3594 
3595 static void rcu_free_pwq(struct rcu_head *rcu)
3596 {
3597 	kmem_cache_free(pwq_cache,
3598 			container_of(rcu, struct pool_workqueue, rcu));
3599 }
3600 
3601 /*
3602  * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3603  * and needs to be destroyed.
3604  */
3605 static void pwq_unbound_release_workfn(struct work_struct *work)
3606 {
3607 	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3608 						  unbound_release_work);
3609 	struct workqueue_struct *wq = pwq->wq;
3610 	struct worker_pool *pool = pwq->pool;
3611 	bool is_last;
3612 
3613 	if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3614 		return;
3615 
3616 	/*
3617 	 * Unlink @pwq.  Synchronization against wq->mutex isn't strictly
3618 	 * necessary on release but do it anyway.  It's easier to verify
3619 	 * and consistent with the linking path.
3620 	 */
3621 	mutex_lock(&wq->mutex);
3622 	list_del_rcu(&pwq->pwqs_node);
3623 	is_last = list_empty(&wq->pwqs);
3624 	mutex_unlock(&wq->mutex);
3625 
3626 	mutex_lock(&wq_pool_mutex);
3627 	put_unbound_pool(pool);
3628 	mutex_unlock(&wq_pool_mutex);
3629 
3630 	call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3631 
3632 	/*
3633 	 * If we're the last pwq going away, @wq is already dead and no one
3634 	 * is gonna access it anymore.  Free it.
3635 	 */
3636 	if (is_last) {
3637 		free_workqueue_attrs(wq->unbound_attrs);
3638 		kfree(wq);
3639 	}
3640 }
3641 
3642 /**
3643  * pwq_adjust_max_active - update a pwq's max_active to the current setting
3644  * @pwq: target pool_workqueue
3645  *
3646  * If @pwq isn't freezing, set @pwq->max_active to the associated
3647  * workqueue's saved_max_active and activate delayed work items
3648  * accordingly.  If @pwq is freezing, clear @pwq->max_active to zero.
3649  */
3650 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3651 {
3652 	struct workqueue_struct *wq = pwq->wq;
3653 	bool freezable = wq->flags & WQ_FREEZABLE;
3654 
3655 	/* for @wq->saved_max_active */
3656 	lockdep_assert_held(&wq->mutex);
3657 
3658 	/* fast exit for non-freezable wqs */
3659 	if (!freezable && pwq->max_active == wq->saved_max_active)
3660 		return;
3661 
3662 	spin_lock_irq(&pwq->pool->lock);
3663 
3664 	if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3665 		pwq->max_active = wq->saved_max_active;
3666 
3667 		while (!list_empty(&pwq->delayed_works) &&
3668 		       pwq->nr_active < pwq->max_active)
3669 			pwq_activate_first_delayed(pwq);
3670 
3671 		/*
3672 		 * Need to kick a worker after thawed or an unbound wq's
3673 		 * max_active is bumped.  It's a slow path.  Do it always.
3674 		 */
3675 		wake_up_worker(pwq->pool);
3676 	} else {
3677 		pwq->max_active = 0;
3678 	}
3679 
3680 	spin_unlock_irq(&pwq->pool->lock);
3681 }
3682 
3683 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3684 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3685 		     struct worker_pool *pool)
3686 {
3687 	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3688 
3689 	memset(pwq, 0, sizeof(*pwq));
3690 
3691 	pwq->pool = pool;
3692 	pwq->wq = wq;
3693 	pwq->flush_color = -1;
3694 	pwq->refcnt = 1;
3695 	INIT_LIST_HEAD(&pwq->delayed_works);
3696 	INIT_LIST_HEAD(&pwq->pwqs_node);
3697 	INIT_LIST_HEAD(&pwq->mayday_node);
3698 	INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3699 }
3700 
3701 /* sync @pwq with the current state of its associated wq and link it */
3702 static void link_pwq(struct pool_workqueue *pwq)
3703 {
3704 	struct workqueue_struct *wq = pwq->wq;
3705 
3706 	lockdep_assert_held(&wq->mutex);
3707 
3708 	/* may be called multiple times, ignore if already linked */
3709 	if (!list_empty(&pwq->pwqs_node))
3710 		return;
3711 
3712 	/*
3713 	 * Set the matching work_color.  This is synchronized with
3714 	 * wq->mutex to avoid confusing flush_workqueue().
3715 	 */
3716 	pwq->work_color = wq->work_color;
3717 
3718 	/* sync max_active to the current setting */
3719 	pwq_adjust_max_active(pwq);
3720 
3721 	/* link in @pwq */
3722 	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3723 }
3724 
3725 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3726 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3727 					const struct workqueue_attrs *attrs)
3728 {
3729 	struct worker_pool *pool;
3730 	struct pool_workqueue *pwq;
3731 
3732 	lockdep_assert_held(&wq_pool_mutex);
3733 
3734 	pool = get_unbound_pool(attrs);
3735 	if (!pool)
3736 		return NULL;
3737 
3738 	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3739 	if (!pwq) {
3740 		put_unbound_pool(pool);
3741 		return NULL;
3742 	}
3743 
3744 	init_pwq(pwq, wq, pool);
3745 	return pwq;
3746 }
3747 
3748 /* undo alloc_unbound_pwq(), used only in the error path */
3749 static void free_unbound_pwq(struct pool_workqueue *pwq)
3750 {
3751 	lockdep_assert_held(&wq_pool_mutex);
3752 
3753 	if (pwq) {
3754 		put_unbound_pool(pwq->pool);
3755 		kmem_cache_free(pwq_cache, pwq);
3756 	}
3757 }
3758 
3759 /**
3760  * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3761  * @attrs: the wq_attrs of interest
3762  * @node: the target NUMA node
3763  * @cpu_going_down: if >= 0, the CPU to consider as offline
3764  * @cpumask: outarg, the resulting cpumask
3765  *
3766  * Calculate the cpumask a workqueue with @attrs should use on @node.  If
3767  * @cpu_going_down is >= 0, that cpu is considered offline during
3768  * calculation.  The result is stored in @cpumask.  This function returns
3769  * %true if the resulting @cpumask is different from @attrs->cpumask,
3770  * %false if equal.
3771  *
3772  * If NUMA affinity is not enabled, @attrs->cpumask is always used.  If
3773  * enabled and @node has online CPUs requested by @attrs, the returned
3774  * cpumask is the intersection of the possible CPUs of @node and
3775  * @attrs->cpumask.
3776  *
3777  * The caller is responsible for ensuring that the cpumask of @node stays
3778  * stable.
3779  */
3780 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3781 				 int cpu_going_down, cpumask_t *cpumask)
3782 {
3783 	if (!wq_numa_enabled || attrs->no_numa)
3784 		goto use_dfl;
3785 
3786 	/* does @node have any online CPUs @attrs wants? */
3787 	cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3788 	if (cpu_going_down >= 0)
3789 		cpumask_clear_cpu(cpu_going_down, cpumask);
3790 
3791 	if (cpumask_empty(cpumask))
3792 		goto use_dfl;
3793 
3794 	/* yeap, return possible CPUs in @node that @attrs wants */
3795 	cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3796 	return !cpumask_equal(cpumask, attrs->cpumask);
3797 
3798 use_dfl:
3799 	cpumask_copy(cpumask, attrs->cpumask);
3800 	return false;
3801 }
3802 
3803 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3804 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3805 						   int node,
3806 						   struct pool_workqueue *pwq)
3807 {
3808 	struct pool_workqueue *old_pwq;
3809 
3810 	lockdep_assert_held(&wq->mutex);
3811 
3812 	/* link_pwq() can handle duplicate calls */
3813 	link_pwq(pwq);
3814 
3815 	old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3816 	rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3817 	return old_pwq;
3818 }
3819 
3820 /**
3821  * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3822  * @wq: the target workqueue
3823  * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3824  *
3825  * Apply @attrs to an unbound workqueue @wq.  Unless disabled, on NUMA
3826  * machines, this function maps a separate pwq to each NUMA node with
3827  * possibles CPUs in @attrs->cpumask so that work items are affine to the
3828  * NUMA node it was issued on.  Older pwqs are released as in-flight work
3829  * items finish.  Note that a work item which repeatedly requeues itself
3830  * back-to-back will stay on its current pwq.
3831  *
3832  * Performs GFP_KERNEL allocations.  Returns 0 on success and -errno on
3833  * failure.
3834  */
3835 int apply_workqueue_attrs(struct workqueue_struct *wq,
3836 			  const struct workqueue_attrs *attrs)
3837 {
3838 	struct workqueue_attrs *new_attrs, *tmp_attrs;
3839 	struct pool_workqueue **pwq_tbl, *dfl_pwq;
3840 	int node, ret;
3841 
3842 	/* only unbound workqueues can change attributes */
3843 	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3844 		return -EINVAL;
3845 
3846 	/* creating multiple pwqs breaks ordering guarantee */
3847 	if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3848 		return -EINVAL;
3849 
3850 	pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3851 	new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3852 	tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3853 	if (!pwq_tbl || !new_attrs || !tmp_attrs)
3854 		goto enomem;
3855 
3856 	/* make a copy of @attrs and sanitize it */
3857 	copy_workqueue_attrs(new_attrs, attrs);
3858 	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3859 
3860 	/*
3861 	 * We may create multiple pwqs with differing cpumasks.  Make a
3862 	 * copy of @new_attrs which will be modified and used to obtain
3863 	 * pools.
3864 	 */
3865 	copy_workqueue_attrs(tmp_attrs, new_attrs);
3866 
3867 	/*
3868 	 * CPUs should stay stable across pwq creations and installations.
3869 	 * Pin CPUs, determine the target cpumask for each node and create
3870 	 * pwqs accordingly.
3871 	 */
3872 	get_online_cpus();
3873 
3874 	mutex_lock(&wq_pool_mutex);
3875 
3876 	/*
3877 	 * If something goes wrong during CPU up/down, we'll fall back to
3878 	 * the default pwq covering whole @attrs->cpumask.  Always create
3879 	 * it even if we don't use it immediately.
3880 	 */
3881 	dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3882 	if (!dfl_pwq)
3883 		goto enomem_pwq;
3884 
3885 	for_each_node(node) {
3886 		if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3887 			pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3888 			if (!pwq_tbl[node])
3889 				goto enomem_pwq;
3890 		} else {
3891 			dfl_pwq->refcnt++;
3892 			pwq_tbl[node] = dfl_pwq;
3893 		}
3894 	}
3895 
3896 	mutex_unlock(&wq_pool_mutex);
3897 
3898 	/* all pwqs have been created successfully, let's install'em */
3899 	mutex_lock(&wq->mutex);
3900 
3901 	copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3902 
3903 	/* save the previous pwq and install the new one */
3904 	for_each_node(node)
3905 		pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3906 
3907 	/* @dfl_pwq might not have been used, ensure it's linked */
3908 	link_pwq(dfl_pwq);
3909 	swap(wq->dfl_pwq, dfl_pwq);
3910 
3911 	mutex_unlock(&wq->mutex);
3912 
3913 	/* put the old pwqs */
3914 	for_each_node(node)
3915 		put_pwq_unlocked(pwq_tbl[node]);
3916 	put_pwq_unlocked(dfl_pwq);
3917 
3918 	put_online_cpus();
3919 	ret = 0;
3920 	/* fall through */
3921 out_free:
3922 	free_workqueue_attrs(tmp_attrs);
3923 	free_workqueue_attrs(new_attrs);
3924 	kfree(pwq_tbl);
3925 	return ret;
3926 
3927 enomem_pwq:
3928 	free_unbound_pwq(dfl_pwq);
3929 	for_each_node(node)
3930 		if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3931 			free_unbound_pwq(pwq_tbl[node]);
3932 	mutex_unlock(&wq_pool_mutex);
3933 	put_online_cpus();
3934 enomem:
3935 	ret = -ENOMEM;
3936 	goto out_free;
3937 }
3938 
3939 /**
3940  * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3941  * @wq: the target workqueue
3942  * @cpu: the CPU coming up or going down
3943  * @online: whether @cpu is coming up or going down
3944  *
3945  * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3946  * %CPU_DOWN_FAILED.  @cpu is being hot[un]plugged, update NUMA affinity of
3947  * @wq accordingly.
3948  *
3949  * If NUMA affinity can't be adjusted due to memory allocation failure, it
3950  * falls back to @wq->dfl_pwq which may not be optimal but is always
3951  * correct.
3952  *
3953  * Note that when the last allowed CPU of a NUMA node goes offline for a
3954  * workqueue with a cpumask spanning multiple nodes, the workers which were
3955  * already executing the work items for the workqueue will lose their CPU
3956  * affinity and may execute on any CPU.  This is similar to how per-cpu
3957  * workqueues behave on CPU_DOWN.  If a workqueue user wants strict
3958  * affinity, it's the user's responsibility to flush the work item from
3959  * CPU_DOWN_PREPARE.
3960  */
3961 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3962 				   bool online)
3963 {
3964 	int node = cpu_to_node(cpu);
3965 	int cpu_off = online ? -1 : cpu;
3966 	struct pool_workqueue *old_pwq = NULL, *pwq;
3967 	struct workqueue_attrs *target_attrs;
3968 	cpumask_t *cpumask;
3969 
3970 	lockdep_assert_held(&wq_pool_mutex);
3971 
3972 	if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
3973 		return;
3974 
3975 	/*
3976 	 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3977 	 * Let's use a preallocated one.  The following buf is protected by
3978 	 * CPU hotplug exclusion.
3979 	 */
3980 	target_attrs = wq_update_unbound_numa_attrs_buf;
3981 	cpumask = target_attrs->cpumask;
3982 
3983 	mutex_lock(&wq->mutex);
3984 	if (wq->unbound_attrs->no_numa)
3985 		goto out_unlock;
3986 
3987 	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3988 	pwq = unbound_pwq_by_node(wq, node);
3989 
3990 	/*
3991 	 * Let's determine what needs to be done.  If the target cpumask is
3992 	 * different from wq's, we need to compare it to @pwq's and create
3993 	 * a new one if they don't match.  If the target cpumask equals
3994 	 * wq's, the default pwq should be used.  If @pwq is already the
3995 	 * default one, nothing to do; otherwise, install the default one.
3996 	 */
3997 	if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
3998 		if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3999 			goto out_unlock;
4000 	} else {
4001 		if (pwq == wq->dfl_pwq)
4002 			goto out_unlock;
4003 		else
4004 			goto use_dfl_pwq;
4005 	}
4006 
4007 	mutex_unlock(&wq->mutex);
4008 
4009 	/* create a new pwq */
4010 	pwq = alloc_unbound_pwq(wq, target_attrs);
4011 	if (!pwq) {
4012 		pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4013 			   wq->name);
4014 		goto out_unlock;
4015 	}
4016 
4017 	/*
4018 	 * Install the new pwq.  As this function is called only from CPU
4019 	 * hotplug callbacks and applying a new attrs is wrapped with
4020 	 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4021 	 * inbetween.
4022 	 */
4023 	mutex_lock(&wq->mutex);
4024 	old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4025 	goto out_unlock;
4026 
4027 use_dfl_pwq:
4028 	spin_lock_irq(&wq->dfl_pwq->pool->lock);
4029 	get_pwq(wq->dfl_pwq);
4030 	spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4031 	old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4032 out_unlock:
4033 	mutex_unlock(&wq->mutex);
4034 	put_pwq_unlocked(old_pwq);
4035 }
4036 
4037 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4038 {
4039 	bool highpri = wq->flags & WQ_HIGHPRI;
4040 	int cpu;
4041 
4042 	if (!(wq->flags & WQ_UNBOUND)) {
4043 		wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4044 		if (!wq->cpu_pwqs)
4045 			return -ENOMEM;
4046 
4047 		for_each_possible_cpu(cpu) {
4048 			struct pool_workqueue *pwq =
4049 				per_cpu_ptr(wq->cpu_pwqs, cpu);
4050 			struct worker_pool *cpu_pools =
4051 				per_cpu(cpu_worker_pools, cpu);
4052 
4053 			init_pwq(pwq, wq, &cpu_pools[highpri]);
4054 
4055 			mutex_lock(&wq->mutex);
4056 			link_pwq(pwq);
4057 			mutex_unlock(&wq->mutex);
4058 		}
4059 		return 0;
4060 	} else {
4061 		return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4062 	}
4063 }
4064 
4065 static int wq_clamp_max_active(int max_active, unsigned int flags,
4066 			       const char *name)
4067 {
4068 	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4069 
4070 	if (max_active < 1 || max_active > lim)
4071 		pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4072 			max_active, name, 1, lim);
4073 
4074 	return clamp_val(max_active, 1, lim);
4075 }
4076 
4077 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4078 					       unsigned int flags,
4079 					       int max_active,
4080 					       struct lock_class_key *key,
4081 					       const char *lock_name, ...)
4082 {
4083 	size_t tbl_size = 0;
4084 	va_list args;
4085 	struct workqueue_struct *wq;
4086 	struct pool_workqueue *pwq;
4087 
4088 	/* allocate wq and format name */
4089 	if (flags & WQ_UNBOUND)
4090 		tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4091 
4092 	wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4093 	if (!wq)
4094 		return NULL;
4095 
4096 	if (flags & WQ_UNBOUND) {
4097 		wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4098 		if (!wq->unbound_attrs)
4099 			goto err_free_wq;
4100 	}
4101 
4102 	va_start(args, lock_name);
4103 	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4104 	va_end(args);
4105 
4106 	max_active = max_active ?: WQ_DFL_ACTIVE;
4107 	max_active = wq_clamp_max_active(max_active, flags, wq->name);
4108 
4109 	/* init wq */
4110 	wq->flags = flags;
4111 	wq->saved_max_active = max_active;
4112 	mutex_init(&wq->mutex);
4113 	atomic_set(&wq->nr_pwqs_to_flush, 0);
4114 	INIT_LIST_HEAD(&wq->pwqs);
4115 	INIT_LIST_HEAD(&wq->flusher_queue);
4116 	INIT_LIST_HEAD(&wq->flusher_overflow);
4117 	INIT_LIST_HEAD(&wq->maydays);
4118 
4119 	lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4120 	INIT_LIST_HEAD(&wq->list);
4121 
4122 	if (alloc_and_link_pwqs(wq) < 0)
4123 		goto err_free_wq;
4124 
4125 	/*
4126 	 * Workqueues which may be used during memory reclaim should
4127 	 * have a rescuer to guarantee forward progress.
4128 	 */
4129 	if (flags & WQ_MEM_RECLAIM) {
4130 		struct worker *rescuer;
4131 
4132 		rescuer = alloc_worker();
4133 		if (!rescuer)
4134 			goto err_destroy;
4135 
4136 		rescuer->rescue_wq = wq;
4137 		rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4138 					       wq->name);
4139 		if (IS_ERR(rescuer->task)) {
4140 			kfree(rescuer);
4141 			goto err_destroy;
4142 		}
4143 
4144 		wq->rescuer = rescuer;
4145 		rescuer->task->flags |= PF_NO_SETAFFINITY;
4146 		wake_up_process(rescuer->task);
4147 	}
4148 
4149 	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4150 		goto err_destroy;
4151 
4152 	/*
4153 	 * wq_pool_mutex protects global freeze state and workqueues list.
4154 	 * Grab it, adjust max_active and add the new @wq to workqueues
4155 	 * list.
4156 	 */
4157 	mutex_lock(&wq_pool_mutex);
4158 
4159 	mutex_lock(&wq->mutex);
4160 	for_each_pwq(pwq, wq)
4161 		pwq_adjust_max_active(pwq);
4162 	mutex_unlock(&wq->mutex);
4163 
4164 	list_add(&wq->list, &workqueues);
4165 
4166 	mutex_unlock(&wq_pool_mutex);
4167 
4168 	return wq;
4169 
4170 err_free_wq:
4171 	free_workqueue_attrs(wq->unbound_attrs);
4172 	kfree(wq);
4173 	return NULL;
4174 err_destroy:
4175 	destroy_workqueue(wq);
4176 	return NULL;
4177 }
4178 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4179 
4180 /**
4181  * destroy_workqueue - safely terminate a workqueue
4182  * @wq: target workqueue
4183  *
4184  * Safely destroy a workqueue. All work currently pending will be done first.
4185  */
4186 void destroy_workqueue(struct workqueue_struct *wq)
4187 {
4188 	struct pool_workqueue *pwq;
4189 	int node;
4190 
4191 	/* drain it before proceeding with destruction */
4192 	drain_workqueue(wq);
4193 
4194 	/* sanity checks */
4195 	mutex_lock(&wq->mutex);
4196 	for_each_pwq(pwq, wq) {
4197 		int i;
4198 
4199 		for (i = 0; i < WORK_NR_COLORS; i++) {
4200 			if (WARN_ON(pwq->nr_in_flight[i])) {
4201 				mutex_unlock(&wq->mutex);
4202 				return;
4203 			}
4204 		}
4205 
4206 		if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4207 		    WARN_ON(pwq->nr_active) ||
4208 		    WARN_ON(!list_empty(&pwq->delayed_works))) {
4209 			mutex_unlock(&wq->mutex);
4210 			return;
4211 		}
4212 	}
4213 	mutex_unlock(&wq->mutex);
4214 
4215 	/*
4216 	 * wq list is used to freeze wq, remove from list after
4217 	 * flushing is complete in case freeze races us.
4218 	 */
4219 	mutex_lock(&wq_pool_mutex);
4220 	list_del_init(&wq->list);
4221 	mutex_unlock(&wq_pool_mutex);
4222 
4223 	workqueue_sysfs_unregister(wq);
4224 
4225 	if (wq->rescuer) {
4226 		kthread_stop(wq->rescuer->task);
4227 		kfree(wq->rescuer);
4228 		wq->rescuer = NULL;
4229 	}
4230 
4231 	if (!(wq->flags & WQ_UNBOUND)) {
4232 		/*
4233 		 * The base ref is never dropped on per-cpu pwqs.  Directly
4234 		 * free the pwqs and wq.
4235 		 */
4236 		free_percpu(wq->cpu_pwqs);
4237 		kfree(wq);
4238 	} else {
4239 		/*
4240 		 * We're the sole accessor of @wq at this point.  Directly
4241 		 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4242 		 * @wq will be freed when the last pwq is released.
4243 		 */
4244 		for_each_node(node) {
4245 			pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4246 			RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4247 			put_pwq_unlocked(pwq);
4248 		}
4249 
4250 		/*
4251 		 * Put dfl_pwq.  @wq may be freed any time after dfl_pwq is
4252 		 * put.  Don't access it afterwards.
4253 		 */
4254 		pwq = wq->dfl_pwq;
4255 		wq->dfl_pwq = NULL;
4256 		put_pwq_unlocked(pwq);
4257 	}
4258 }
4259 EXPORT_SYMBOL_GPL(destroy_workqueue);
4260 
4261 /**
4262  * workqueue_set_max_active - adjust max_active of a workqueue
4263  * @wq: target workqueue
4264  * @max_active: new max_active value.
4265  *
4266  * Set max_active of @wq to @max_active.
4267  *
4268  * CONTEXT:
4269  * Don't call from IRQ context.
4270  */
4271 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4272 {
4273 	struct pool_workqueue *pwq;
4274 
4275 	/* disallow meddling with max_active for ordered workqueues */
4276 	if (WARN_ON(wq->flags & __WQ_ORDERED))
4277 		return;
4278 
4279 	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4280 
4281 	mutex_lock(&wq->mutex);
4282 
4283 	wq->saved_max_active = max_active;
4284 
4285 	for_each_pwq(pwq, wq)
4286 		pwq_adjust_max_active(pwq);
4287 
4288 	mutex_unlock(&wq->mutex);
4289 }
4290 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4291 
4292 /**
4293  * current_is_workqueue_rescuer - is %current workqueue rescuer?
4294  *
4295  * Determine whether %current is a workqueue rescuer.  Can be used from
4296  * work functions to determine whether it's being run off the rescuer task.
4297  */
4298 bool current_is_workqueue_rescuer(void)
4299 {
4300 	struct worker *worker = current_wq_worker();
4301 
4302 	return worker && worker->rescue_wq;
4303 }
4304 
4305 /**
4306  * workqueue_congested - test whether a workqueue is congested
4307  * @cpu: CPU in question
4308  * @wq: target workqueue
4309  *
4310  * Test whether @wq's cpu workqueue for @cpu is congested.  There is
4311  * no synchronization around this function and the test result is
4312  * unreliable and only useful as advisory hints or for debugging.
4313  *
4314  * RETURNS:
4315  * %true if congested, %false otherwise.
4316  */
4317 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4318 {
4319 	struct pool_workqueue *pwq;
4320 	bool ret;
4321 
4322 	rcu_read_lock_sched();
4323 
4324 	if (!(wq->flags & WQ_UNBOUND))
4325 		pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4326 	else
4327 		pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4328 
4329 	ret = !list_empty(&pwq->delayed_works);
4330 	rcu_read_unlock_sched();
4331 
4332 	return ret;
4333 }
4334 EXPORT_SYMBOL_GPL(workqueue_congested);
4335 
4336 /**
4337  * work_busy - test whether a work is currently pending or running
4338  * @work: the work to be tested
4339  *
4340  * Test whether @work is currently pending or running.  There is no
4341  * synchronization around this function and the test result is
4342  * unreliable and only useful as advisory hints or for debugging.
4343  *
4344  * RETURNS:
4345  * OR'd bitmask of WORK_BUSY_* bits.
4346  */
4347 unsigned int work_busy(struct work_struct *work)
4348 {
4349 	struct worker_pool *pool;
4350 	unsigned long flags;
4351 	unsigned int ret = 0;
4352 
4353 	if (work_pending(work))
4354 		ret |= WORK_BUSY_PENDING;
4355 
4356 	local_irq_save(flags);
4357 	pool = get_work_pool(work);
4358 	if (pool) {
4359 		spin_lock(&pool->lock);
4360 		if (find_worker_executing_work(pool, work))
4361 			ret |= WORK_BUSY_RUNNING;
4362 		spin_unlock(&pool->lock);
4363 	}
4364 	local_irq_restore(flags);
4365 
4366 	return ret;
4367 }
4368 EXPORT_SYMBOL_GPL(work_busy);
4369 
4370 /**
4371  * set_worker_desc - set description for the current work item
4372  * @fmt: printf-style format string
4373  * @...: arguments for the format string
4374  *
4375  * This function can be called by a running work function to describe what
4376  * the work item is about.  If the worker task gets dumped, this
4377  * information will be printed out together to help debugging.  The
4378  * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4379  */
4380 void set_worker_desc(const char *fmt, ...)
4381 {
4382 	struct worker *worker = current_wq_worker();
4383 	va_list args;
4384 
4385 	if (worker) {
4386 		va_start(args, fmt);
4387 		vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4388 		va_end(args);
4389 		worker->desc_valid = true;
4390 	}
4391 }
4392 
4393 /**
4394  * print_worker_info - print out worker information and description
4395  * @log_lvl: the log level to use when printing
4396  * @task: target task
4397  *
4398  * If @task is a worker and currently executing a work item, print out the
4399  * name of the workqueue being serviced and worker description set with
4400  * set_worker_desc() by the currently executing work item.
4401  *
4402  * This function can be safely called on any task as long as the
4403  * task_struct itself is accessible.  While safe, this function isn't
4404  * synchronized and may print out mixups or garbages of limited length.
4405  */
4406 void print_worker_info(const char *log_lvl, struct task_struct *task)
4407 {
4408 	work_func_t *fn = NULL;
4409 	char name[WQ_NAME_LEN] = { };
4410 	char desc[WORKER_DESC_LEN] = { };
4411 	struct pool_workqueue *pwq = NULL;
4412 	struct workqueue_struct *wq = NULL;
4413 	bool desc_valid = false;
4414 	struct worker *worker;
4415 
4416 	if (!(task->flags & PF_WQ_WORKER))
4417 		return;
4418 
4419 	/*
4420 	 * This function is called without any synchronization and @task
4421 	 * could be in any state.  Be careful with dereferences.
4422 	 */
4423 	worker = probe_kthread_data(task);
4424 
4425 	/*
4426 	 * Carefully copy the associated workqueue's workfn and name.  Keep
4427 	 * the original last '\0' in case the original contains garbage.
4428 	 */
4429 	probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4430 	probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4431 	probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4432 	probe_kernel_read(name, wq->name, sizeof(name) - 1);
4433 
4434 	/* copy worker description */
4435 	probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4436 	if (desc_valid)
4437 		probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4438 
4439 	if (fn || name[0] || desc[0]) {
4440 		printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4441 		if (desc[0])
4442 			pr_cont(" (%s)", desc);
4443 		pr_cont("\n");
4444 	}
4445 }
4446 
4447 /*
4448  * CPU hotplug.
4449  *
4450  * There are two challenges in supporting CPU hotplug.  Firstly, there
4451  * are a lot of assumptions on strong associations among work, pwq and
4452  * pool which make migrating pending and scheduled works very
4453  * difficult to implement without impacting hot paths.  Secondly,
4454  * worker pools serve mix of short, long and very long running works making
4455  * blocked draining impractical.
4456  *
4457  * This is solved by allowing the pools to be disassociated from the CPU
4458  * running as an unbound one and allowing it to be reattached later if the
4459  * cpu comes back online.
4460  */
4461 
4462 static void wq_unbind_fn(struct work_struct *work)
4463 {
4464 	int cpu = smp_processor_id();
4465 	struct worker_pool *pool;
4466 	struct worker *worker;
4467 	int wi;
4468 
4469 	for_each_cpu_worker_pool(pool, cpu) {
4470 		WARN_ON_ONCE(cpu != smp_processor_id());
4471 
4472 		mutex_lock(&pool->manager_mutex);
4473 		spin_lock_irq(&pool->lock);
4474 
4475 		/*
4476 		 * We've blocked all manager operations.  Make all workers
4477 		 * unbound and set DISASSOCIATED.  Before this, all workers
4478 		 * except for the ones which are still executing works from
4479 		 * before the last CPU down must be on the cpu.  After
4480 		 * this, they may become diasporas.
4481 		 */
4482 		for_each_pool_worker(worker, wi, pool)
4483 			worker->flags |= WORKER_UNBOUND;
4484 
4485 		pool->flags |= POOL_DISASSOCIATED;
4486 
4487 		spin_unlock_irq(&pool->lock);
4488 		mutex_unlock(&pool->manager_mutex);
4489 
4490 		/*
4491 		 * Call schedule() so that we cross rq->lock and thus can
4492 		 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4493 		 * This is necessary as scheduler callbacks may be invoked
4494 		 * from other cpus.
4495 		 */
4496 		schedule();
4497 
4498 		/*
4499 		 * Sched callbacks are disabled now.  Zap nr_running.
4500 		 * After this, nr_running stays zero and need_more_worker()
4501 		 * and keep_working() are always true as long as the
4502 		 * worklist is not empty.  This pool now behaves as an
4503 		 * unbound (in terms of concurrency management) pool which
4504 		 * are served by workers tied to the pool.
4505 		 */
4506 		atomic_set(&pool->nr_running, 0);
4507 
4508 		/*
4509 		 * With concurrency management just turned off, a busy
4510 		 * worker blocking could lead to lengthy stalls.  Kick off
4511 		 * unbound chain execution of currently pending work items.
4512 		 */
4513 		spin_lock_irq(&pool->lock);
4514 		wake_up_worker(pool);
4515 		spin_unlock_irq(&pool->lock);
4516 	}
4517 }
4518 
4519 /**
4520  * rebind_workers - rebind all workers of a pool to the associated CPU
4521  * @pool: pool of interest
4522  *
4523  * @pool->cpu is coming online.  Rebind all workers to the CPU.
4524  */
4525 static void rebind_workers(struct worker_pool *pool)
4526 {
4527 	struct worker *worker;
4528 	int wi;
4529 
4530 	lockdep_assert_held(&pool->manager_mutex);
4531 
4532 	/*
4533 	 * Restore CPU affinity of all workers.  As all idle workers should
4534 	 * be on the run-queue of the associated CPU before any local
4535 	 * wake-ups for concurrency management happen, restore CPU affinty
4536 	 * of all workers first and then clear UNBOUND.  As we're called
4537 	 * from CPU_ONLINE, the following shouldn't fail.
4538 	 */
4539 	for_each_pool_worker(worker, wi, pool)
4540 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4541 						  pool->attrs->cpumask) < 0);
4542 
4543 	spin_lock_irq(&pool->lock);
4544 
4545 	for_each_pool_worker(worker, wi, pool) {
4546 		unsigned int worker_flags = worker->flags;
4547 
4548 		/*
4549 		 * A bound idle worker should actually be on the runqueue
4550 		 * of the associated CPU for local wake-ups targeting it to
4551 		 * work.  Kick all idle workers so that they migrate to the
4552 		 * associated CPU.  Doing this in the same loop as
4553 		 * replacing UNBOUND with REBOUND is safe as no worker will
4554 		 * be bound before @pool->lock is released.
4555 		 */
4556 		if (worker_flags & WORKER_IDLE)
4557 			wake_up_process(worker->task);
4558 
4559 		/*
4560 		 * We want to clear UNBOUND but can't directly call
4561 		 * worker_clr_flags() or adjust nr_running.  Atomically
4562 		 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4563 		 * @worker will clear REBOUND using worker_clr_flags() when
4564 		 * it initiates the next execution cycle thus restoring
4565 		 * concurrency management.  Note that when or whether
4566 		 * @worker clears REBOUND doesn't affect correctness.
4567 		 *
4568 		 * ACCESS_ONCE() is necessary because @worker->flags may be
4569 		 * tested without holding any lock in
4570 		 * wq_worker_waking_up().  Without it, NOT_RUNNING test may
4571 		 * fail incorrectly leading to premature concurrency
4572 		 * management operations.
4573 		 */
4574 		WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4575 		worker_flags |= WORKER_REBOUND;
4576 		worker_flags &= ~WORKER_UNBOUND;
4577 		ACCESS_ONCE(worker->flags) = worker_flags;
4578 	}
4579 
4580 	spin_unlock_irq(&pool->lock);
4581 }
4582 
4583 /**
4584  * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4585  * @pool: unbound pool of interest
4586  * @cpu: the CPU which is coming up
4587  *
4588  * An unbound pool may end up with a cpumask which doesn't have any online
4589  * CPUs.  When a worker of such pool get scheduled, the scheduler resets
4590  * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
4591  * online CPU before, cpus_allowed of all its workers should be restored.
4592  */
4593 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4594 {
4595 	static cpumask_t cpumask;
4596 	struct worker *worker;
4597 	int wi;
4598 
4599 	lockdep_assert_held(&pool->manager_mutex);
4600 
4601 	/* is @cpu allowed for @pool? */
4602 	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4603 		return;
4604 
4605 	/* is @cpu the only online CPU? */
4606 	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4607 	if (cpumask_weight(&cpumask) != 1)
4608 		return;
4609 
4610 	/* as we're called from CPU_ONLINE, the following shouldn't fail */
4611 	for_each_pool_worker(worker, wi, pool)
4612 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4613 						  pool->attrs->cpumask) < 0);
4614 }
4615 
4616 /*
4617  * Workqueues should be brought up before normal priority CPU notifiers.
4618  * This will be registered high priority CPU notifier.
4619  */
4620 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
4621 					       unsigned long action,
4622 					       void *hcpu)
4623 {
4624 	int cpu = (unsigned long)hcpu;
4625 	struct worker_pool *pool;
4626 	struct workqueue_struct *wq;
4627 	int pi;
4628 
4629 	switch (action & ~CPU_TASKS_FROZEN) {
4630 	case CPU_UP_PREPARE:
4631 		for_each_cpu_worker_pool(pool, cpu) {
4632 			if (pool->nr_workers)
4633 				continue;
4634 			if (create_and_start_worker(pool) < 0)
4635 				return NOTIFY_BAD;
4636 		}
4637 		break;
4638 
4639 	case CPU_DOWN_FAILED:
4640 	case CPU_ONLINE:
4641 		mutex_lock(&wq_pool_mutex);
4642 
4643 		for_each_pool(pool, pi) {
4644 			mutex_lock(&pool->manager_mutex);
4645 
4646 			if (pool->cpu == cpu) {
4647 				spin_lock_irq(&pool->lock);
4648 				pool->flags &= ~POOL_DISASSOCIATED;
4649 				spin_unlock_irq(&pool->lock);
4650 
4651 				rebind_workers(pool);
4652 			} else if (pool->cpu < 0) {
4653 				restore_unbound_workers_cpumask(pool, cpu);
4654 			}
4655 
4656 			mutex_unlock(&pool->manager_mutex);
4657 		}
4658 
4659 		/* update NUMA affinity of unbound workqueues */
4660 		list_for_each_entry(wq, &workqueues, list)
4661 			wq_update_unbound_numa(wq, cpu, true);
4662 
4663 		mutex_unlock(&wq_pool_mutex);
4664 		break;
4665 	}
4666 	return NOTIFY_OK;
4667 }
4668 
4669 /*
4670  * Workqueues should be brought down after normal priority CPU notifiers.
4671  * This will be registered as low priority CPU notifier.
4672  */
4673 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
4674 						 unsigned long action,
4675 						 void *hcpu)
4676 {
4677 	int cpu = (unsigned long)hcpu;
4678 	struct work_struct unbind_work;
4679 	struct workqueue_struct *wq;
4680 
4681 	switch (action & ~CPU_TASKS_FROZEN) {
4682 	case CPU_DOWN_PREPARE:
4683 		/* unbinding per-cpu workers should happen on the local CPU */
4684 		INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4685 		queue_work_on(cpu, system_highpri_wq, &unbind_work);
4686 
4687 		/* update NUMA affinity of unbound workqueues */
4688 		mutex_lock(&wq_pool_mutex);
4689 		list_for_each_entry(wq, &workqueues, list)
4690 			wq_update_unbound_numa(wq, cpu, false);
4691 		mutex_unlock(&wq_pool_mutex);
4692 
4693 		/* wait for per-cpu unbinding to finish */
4694 		flush_work(&unbind_work);
4695 		break;
4696 	}
4697 	return NOTIFY_OK;
4698 }
4699 
4700 #ifdef CONFIG_SMP
4701 
4702 struct work_for_cpu {
4703 	struct work_struct work;
4704 	long (*fn)(void *);
4705 	void *arg;
4706 	long ret;
4707 };
4708 
4709 static void work_for_cpu_fn(struct work_struct *work)
4710 {
4711 	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4712 
4713 	wfc->ret = wfc->fn(wfc->arg);
4714 }
4715 
4716 /**
4717  * work_on_cpu - run a function in user context on a particular cpu
4718  * @cpu: the cpu to run on
4719  * @fn: the function to run
4720  * @arg: the function arg
4721  *
4722  * This will return the value @fn returns.
4723  * It is up to the caller to ensure that the cpu doesn't go offline.
4724  * The caller must not hold any locks which would prevent @fn from completing.
4725  */
4726 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4727 {
4728 	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4729 
4730 	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4731 	schedule_work_on(cpu, &wfc.work);
4732 	flush_work(&wfc.work);
4733 	return wfc.ret;
4734 }
4735 EXPORT_SYMBOL_GPL(work_on_cpu);
4736 #endif /* CONFIG_SMP */
4737 
4738 #ifdef CONFIG_FREEZER
4739 
4740 /**
4741  * freeze_workqueues_begin - begin freezing workqueues
4742  *
4743  * Start freezing workqueues.  After this function returns, all freezable
4744  * workqueues will queue new works to their delayed_works list instead of
4745  * pool->worklist.
4746  *
4747  * CONTEXT:
4748  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4749  */
4750 void freeze_workqueues_begin(void)
4751 {
4752 	struct worker_pool *pool;
4753 	struct workqueue_struct *wq;
4754 	struct pool_workqueue *pwq;
4755 	int pi;
4756 
4757 	mutex_lock(&wq_pool_mutex);
4758 
4759 	WARN_ON_ONCE(workqueue_freezing);
4760 	workqueue_freezing = true;
4761 
4762 	/* set FREEZING */
4763 	for_each_pool(pool, pi) {
4764 		spin_lock_irq(&pool->lock);
4765 		WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4766 		pool->flags |= POOL_FREEZING;
4767 		spin_unlock_irq(&pool->lock);
4768 	}
4769 
4770 	list_for_each_entry(wq, &workqueues, list) {
4771 		mutex_lock(&wq->mutex);
4772 		for_each_pwq(pwq, wq)
4773 			pwq_adjust_max_active(pwq);
4774 		mutex_unlock(&wq->mutex);
4775 	}
4776 
4777 	mutex_unlock(&wq_pool_mutex);
4778 }
4779 
4780 /**
4781  * freeze_workqueues_busy - are freezable workqueues still busy?
4782  *
4783  * Check whether freezing is complete.  This function must be called
4784  * between freeze_workqueues_begin() and thaw_workqueues().
4785  *
4786  * CONTEXT:
4787  * Grabs and releases wq_pool_mutex.
4788  *
4789  * RETURNS:
4790  * %true if some freezable workqueues are still busy.  %false if freezing
4791  * is complete.
4792  */
4793 bool freeze_workqueues_busy(void)
4794 {
4795 	bool busy = false;
4796 	struct workqueue_struct *wq;
4797 	struct pool_workqueue *pwq;
4798 
4799 	mutex_lock(&wq_pool_mutex);
4800 
4801 	WARN_ON_ONCE(!workqueue_freezing);
4802 
4803 	list_for_each_entry(wq, &workqueues, list) {
4804 		if (!(wq->flags & WQ_FREEZABLE))
4805 			continue;
4806 		/*
4807 		 * nr_active is monotonically decreasing.  It's safe
4808 		 * to peek without lock.
4809 		 */
4810 		rcu_read_lock_sched();
4811 		for_each_pwq(pwq, wq) {
4812 			WARN_ON_ONCE(pwq->nr_active < 0);
4813 			if (pwq->nr_active) {
4814 				busy = true;
4815 				rcu_read_unlock_sched();
4816 				goto out_unlock;
4817 			}
4818 		}
4819 		rcu_read_unlock_sched();
4820 	}
4821 out_unlock:
4822 	mutex_unlock(&wq_pool_mutex);
4823 	return busy;
4824 }
4825 
4826 /**
4827  * thaw_workqueues - thaw workqueues
4828  *
4829  * Thaw workqueues.  Normal queueing is restored and all collected
4830  * frozen works are transferred to their respective pool worklists.
4831  *
4832  * CONTEXT:
4833  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4834  */
4835 void thaw_workqueues(void)
4836 {
4837 	struct workqueue_struct *wq;
4838 	struct pool_workqueue *pwq;
4839 	struct worker_pool *pool;
4840 	int pi;
4841 
4842 	mutex_lock(&wq_pool_mutex);
4843 
4844 	if (!workqueue_freezing)
4845 		goto out_unlock;
4846 
4847 	/* clear FREEZING */
4848 	for_each_pool(pool, pi) {
4849 		spin_lock_irq(&pool->lock);
4850 		WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4851 		pool->flags &= ~POOL_FREEZING;
4852 		spin_unlock_irq(&pool->lock);
4853 	}
4854 
4855 	/* restore max_active and repopulate worklist */
4856 	list_for_each_entry(wq, &workqueues, list) {
4857 		mutex_lock(&wq->mutex);
4858 		for_each_pwq(pwq, wq)
4859 			pwq_adjust_max_active(pwq);
4860 		mutex_unlock(&wq->mutex);
4861 	}
4862 
4863 	workqueue_freezing = false;
4864 out_unlock:
4865 	mutex_unlock(&wq_pool_mutex);
4866 }
4867 #endif /* CONFIG_FREEZER */
4868 
4869 static void __init wq_numa_init(void)
4870 {
4871 	cpumask_var_t *tbl;
4872 	int node, cpu;
4873 
4874 	/* determine NUMA pwq table len - highest node id + 1 */
4875 	for_each_node(node)
4876 		wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4877 
4878 	if (num_possible_nodes() <= 1)
4879 		return;
4880 
4881 	if (wq_disable_numa) {
4882 		pr_info("workqueue: NUMA affinity support disabled\n");
4883 		return;
4884 	}
4885 
4886 	wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4887 	BUG_ON(!wq_update_unbound_numa_attrs_buf);
4888 
4889 	/*
4890 	 * We want masks of possible CPUs of each node which isn't readily
4891 	 * available.  Build one from cpu_to_node() which should have been
4892 	 * fully initialized by now.
4893 	 */
4894 	tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4895 	BUG_ON(!tbl);
4896 
4897 	for_each_node(node)
4898 		BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, node));
4899 
4900 	for_each_possible_cpu(cpu) {
4901 		node = cpu_to_node(cpu);
4902 		if (WARN_ON(node == NUMA_NO_NODE)) {
4903 			pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4904 			/* happens iff arch is bonkers, let's just proceed */
4905 			return;
4906 		}
4907 		cpumask_set_cpu(cpu, tbl[node]);
4908 	}
4909 
4910 	wq_numa_possible_cpumask = tbl;
4911 	wq_numa_enabled = true;
4912 }
4913 
4914 static int __init init_workqueues(void)
4915 {
4916 	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4917 	int i, cpu;
4918 
4919 	/* make sure we have enough bits for OFFQ pool ID */
4920 	BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4921 		     WORK_CPU_END * NR_STD_WORKER_POOLS);
4922 
4923 	WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4924 
4925 	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4926 
4927 	cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4928 	hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4929 
4930 	wq_numa_init();
4931 
4932 	/* initialize CPU pools */
4933 	for_each_possible_cpu(cpu) {
4934 		struct worker_pool *pool;
4935 
4936 		i = 0;
4937 		for_each_cpu_worker_pool(pool, cpu) {
4938 			BUG_ON(init_worker_pool(pool));
4939 			pool->cpu = cpu;
4940 			cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4941 			pool->attrs->nice = std_nice[i++];
4942 			pool->node = cpu_to_node(cpu);
4943 
4944 			/* alloc pool ID */
4945 			mutex_lock(&wq_pool_mutex);
4946 			BUG_ON(worker_pool_assign_id(pool));
4947 			mutex_unlock(&wq_pool_mutex);
4948 		}
4949 	}
4950 
4951 	/* create the initial worker */
4952 	for_each_online_cpu(cpu) {
4953 		struct worker_pool *pool;
4954 
4955 		for_each_cpu_worker_pool(pool, cpu) {
4956 			pool->flags &= ~POOL_DISASSOCIATED;
4957 			BUG_ON(create_and_start_worker(pool) < 0);
4958 		}
4959 	}
4960 
4961 	/* create default unbound wq attrs */
4962 	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4963 		struct workqueue_attrs *attrs;
4964 
4965 		BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4966 		attrs->nice = std_nice[i];
4967 		unbound_std_wq_attrs[i] = attrs;
4968 	}
4969 
4970 	system_wq = alloc_workqueue("events", 0, 0);
4971 	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4972 	system_long_wq = alloc_workqueue("events_long", 0, 0);
4973 	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4974 					    WQ_UNBOUND_MAX_ACTIVE);
4975 	system_freezable_wq = alloc_workqueue("events_freezable",
4976 					      WQ_FREEZABLE, 0);
4977 	BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4978 	       !system_unbound_wq || !system_freezable_wq);
4979 	return 0;
4980 }
4981 early_initcall(init_workqueues);
4982