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