1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * kernel/workqueue.c - generic async execution with shared worker pool
4  *
5  * Copyright (C) 2002		Ingo Molnar
6  *
7  *   Derived from the taskqueue/keventd code by:
8  *     David Woodhouse <dwmw2@infradead.org>
9  *     Andrew Morton
10  *     Kai Petzke <wpp@marie.physik.tu-berlin.de>
11  *     Theodore Ts'o <tytso@mit.edu>
12  *
13  * Made to use alloc_percpu by Christoph Lameter.
14  *
15  * Copyright (C) 2010		SUSE Linux Products GmbH
16  * Copyright (C) 2010		Tejun Heo <tj@kernel.org>
17  *
18  * This is the generic async execution mechanism.  Work items as are
19  * executed in process context.  The worker pool is shared and
20  * automatically managed.  There are two worker pools for each CPU (one for
21  * normal work items and the other for high priority ones) and some extra
22  * pools for workqueues which are not bound to any specific CPU - the
23  * number of these backing pools is dynamic.
24  *
25  * Please read Documentation/core-api/workqueue.rst for details.
26  */
27 
28 #include <linux/export.h>
29 #include <linux/kernel.h>
30 #include <linux/sched.h>
31 #include <linux/init.h>
32 #include <linux/signal.h>
33 #include <linux/completion.h>
34 #include <linux/workqueue.h>
35 #include <linux/slab.h>
36 #include <linux/cpu.h>
37 #include <linux/notifier.h>
38 #include <linux/kthread.h>
39 #include <linux/hardirq.h>
40 #include <linux/mempolicy.h>
41 #include <linux/freezer.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 #include <linux/sched/isolation.h>
52 #include <linux/nmi.h>
53 #include <linux/kvm_para.h>
54 
55 #include "workqueue_internal.h"
56 
57 enum {
58 	/*
59 	 * worker_pool flags
60 	 *
61 	 * A bound pool is either associated or disassociated with its CPU.
62 	 * While associated (!DISASSOCIATED), all workers are bound to the
63 	 * CPU and none has %WORKER_UNBOUND set and concurrency management
64 	 * is in effect.
65 	 *
66 	 * While DISASSOCIATED, the cpu may be offline and all workers have
67 	 * %WORKER_UNBOUND set and concurrency management disabled, and may
68 	 * be executing on any CPU.  The pool behaves as an unbound one.
69 	 *
70 	 * Note that DISASSOCIATED should be flipped only while holding
71 	 * wq_pool_attach_mutex to avoid changing binding state while
72 	 * worker_attach_to_pool() is in progress.
73 	 */
74 	POOL_MANAGER_ACTIVE	= 1 << 0,	/* being managed */
75 	POOL_DISASSOCIATED	= 1 << 2,	/* cpu can't serve workers */
76 
77 	/* worker flags */
78 	WORKER_DIE		= 1 << 1,	/* die die die */
79 	WORKER_IDLE		= 1 << 2,	/* is idle */
80 	WORKER_PREP		= 1 << 3,	/* preparing to run works */
81 	WORKER_CPU_INTENSIVE	= 1 << 6,	/* cpu intensive */
82 	WORKER_UNBOUND		= 1 << 7,	/* worker is unbound */
83 	WORKER_REBOUND		= 1 << 8,	/* worker was rebound */
84 
85 	WORKER_NOT_RUNNING	= WORKER_PREP | WORKER_CPU_INTENSIVE |
86 				  WORKER_UNBOUND | WORKER_REBOUND,
87 
88 	NR_STD_WORKER_POOLS	= 2,		/* # standard pools per cpu */
89 
90 	UNBOUND_POOL_HASH_ORDER	= 6,		/* hashed by pool->attrs */
91 	BUSY_WORKER_HASH_ORDER	= 6,		/* 64 pointers */
92 
93 	MAX_IDLE_WORKERS_RATIO	= 4,		/* 1/4 of busy can be idle */
94 	IDLE_WORKER_TIMEOUT	= 300 * HZ,	/* keep idle ones for 5 mins */
95 
96 	MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
97 						/* call for help after 10ms
98 						   (min two ticks) */
99 	MAYDAY_INTERVAL		= HZ / 10,	/* and then every 100ms */
100 	CREATE_COOLDOWN		= HZ,		/* time to breath after fail */
101 
102 	/*
103 	 * Rescue workers are used only on emergencies and shared by
104 	 * all cpus.  Give MIN_NICE.
105 	 */
106 	RESCUER_NICE_LEVEL	= MIN_NICE,
107 	HIGHPRI_NICE_LEVEL	= MIN_NICE,
108 
109 	WQ_NAME_LEN		= 24,
110 };
111 
112 /*
113  * Structure fields follow one of the following exclusion rules.
114  *
115  * I: Modifiable by initialization/destruction paths and read-only for
116  *    everyone else.
117  *
118  * P: Preemption protected.  Disabling preemption is enough and should
119  *    only be modified and accessed from the local cpu.
120  *
121  * L: pool->lock protected.  Access with pool->lock held.
122  *
123  * X: During normal operation, modification requires pool->lock and should
124  *    be done only from local cpu.  Either disabling preemption on local
125  *    cpu or grabbing pool->lock is enough for read access.  If
126  *    POOL_DISASSOCIATED is set, it's identical to L.
127  *
128  * A: wq_pool_attach_mutex protected.
129  *
130  * PL: wq_pool_mutex protected.
131  *
132  * PR: wq_pool_mutex protected for writes.  RCU protected for reads.
133  *
134  * PW: wq_pool_mutex and wq->mutex protected for writes.  Either for reads.
135  *
136  * PWR: wq_pool_mutex and wq->mutex protected for writes.  Either or
137  *      RCU for reads.
138  *
139  * WQ: wq->mutex protected.
140  *
141  * WR: wq->mutex protected for writes.  RCU protected for reads.
142  *
143  * MD: wq_mayday_lock protected.
144  */
145 
146 /* struct worker is defined in workqueue_internal.h */
147 
148 struct worker_pool {
149 	raw_spinlock_t		lock;		/* the pool lock */
150 	int			cpu;		/* I: the associated cpu */
151 	int			node;		/* I: the associated node ID */
152 	int			id;		/* I: pool ID */
153 	unsigned int		flags;		/* X: flags */
154 
155 	unsigned long		watchdog_ts;	/* L: watchdog timestamp */
156 
157 	/*
158 	 * The counter is incremented in a process context on the associated CPU
159 	 * w/ preemption disabled, and decremented or reset in the same context
160 	 * but w/ pool->lock held. The readers grab pool->lock and are
161 	 * guaranteed to see if the counter reached zero.
162 	 */
163 	int			nr_running;
164 
165 	struct list_head	worklist;	/* L: list of pending works */
166 
167 	int			nr_workers;	/* L: total number of workers */
168 	int			nr_idle;	/* L: currently idle workers */
169 
170 	struct list_head	idle_list;	/* L: list of idle workers */
171 	struct timer_list	idle_timer;	/* L: worker idle timeout */
172 	struct timer_list	mayday_timer;	/* L: SOS timer for workers */
173 
174 	/* a workers is either on busy_hash or idle_list, or the manager */
175 	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
176 						/* L: hash of busy workers */
177 
178 	struct worker		*manager;	/* L: purely informational */
179 	struct list_head	workers;	/* A: attached workers */
180 	struct completion	*detach_completion; /* all workers detached */
181 
182 	struct ida		worker_ida;	/* worker IDs for task name */
183 
184 	struct workqueue_attrs	*attrs;		/* I: worker attributes */
185 	struct hlist_node	hash_node;	/* PL: unbound_pool_hash node */
186 	int			refcnt;		/* PL: refcnt for unbound pools */
187 
188 	/*
189 	 * Destruction of pool is RCU protected to allow dereferences
190 	 * from get_work_pool().
191 	 */
192 	struct rcu_head		rcu;
193 };
194 
195 /*
196  * The per-pool workqueue.  While queued, the lower WORK_STRUCT_FLAG_BITS
197  * of work_struct->data are used for flags and the remaining high bits
198  * point to the pwq; thus, pwqs need to be aligned at two's power of the
199  * number of flag bits.
200  */
201 struct pool_workqueue {
202 	struct worker_pool	*pool;		/* I: the associated pool */
203 	struct workqueue_struct *wq;		/* I: the owning workqueue */
204 	int			work_color;	/* L: current color */
205 	int			flush_color;	/* L: flushing color */
206 	int			refcnt;		/* L: reference count */
207 	int			nr_in_flight[WORK_NR_COLORS];
208 						/* L: nr of in_flight works */
209 
210 	/*
211 	 * nr_active management and WORK_STRUCT_INACTIVE:
212 	 *
213 	 * When pwq->nr_active >= max_active, new work item is queued to
214 	 * pwq->inactive_works instead of pool->worklist and marked with
215 	 * WORK_STRUCT_INACTIVE.
216 	 *
217 	 * All work items marked with WORK_STRUCT_INACTIVE do not participate
218 	 * in pwq->nr_active and all work items in pwq->inactive_works are
219 	 * marked with WORK_STRUCT_INACTIVE.  But not all WORK_STRUCT_INACTIVE
220 	 * work items are in pwq->inactive_works.  Some of them are ready to
221 	 * run in pool->worklist or worker->scheduled.  Those work itmes are
222 	 * only struct wq_barrier which is used for flush_work() and should
223 	 * not participate in pwq->nr_active.  For non-barrier work item, it
224 	 * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
225 	 */
226 	int			nr_active;	/* L: nr of active works */
227 	int			max_active;	/* L: max active works */
228 	struct list_head	inactive_works;	/* L: inactive works */
229 	struct list_head	pwqs_node;	/* WR: node on wq->pwqs */
230 	struct list_head	mayday_node;	/* MD: node on wq->maydays */
231 
232 	/*
233 	 * Release of unbound pwq is punted to system_wq.  See put_pwq()
234 	 * and pwq_unbound_release_workfn() for details.  pool_workqueue
235 	 * itself is also RCU protected so that the first pwq can be
236 	 * determined without grabbing wq->mutex.
237 	 */
238 	struct work_struct	unbound_release_work;
239 	struct rcu_head		rcu;
240 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
241 
242 /*
243  * Structure used to wait for workqueue flush.
244  */
245 struct wq_flusher {
246 	struct list_head	list;		/* WQ: list of flushers */
247 	int			flush_color;	/* WQ: flush color waiting for */
248 	struct completion	done;		/* flush completion */
249 };
250 
251 struct wq_device;
252 
253 /*
254  * The externally visible workqueue.  It relays the issued work items to
255  * the appropriate worker_pool through its pool_workqueues.
256  */
257 struct workqueue_struct {
258 	struct list_head	pwqs;		/* WR: all pwqs of this wq */
259 	struct list_head	list;		/* PR: list of all workqueues */
260 
261 	struct mutex		mutex;		/* protects this wq */
262 	int			work_color;	/* WQ: current work color */
263 	int			flush_color;	/* WQ: current flush color */
264 	atomic_t		nr_pwqs_to_flush; /* flush in progress */
265 	struct wq_flusher	*first_flusher;	/* WQ: first flusher */
266 	struct list_head	flusher_queue;	/* WQ: flush waiters */
267 	struct list_head	flusher_overflow; /* WQ: flush overflow list */
268 
269 	struct list_head	maydays;	/* MD: pwqs requesting rescue */
270 	struct worker		*rescuer;	/* MD: rescue worker */
271 
272 	int			nr_drainers;	/* WQ: drain in progress */
273 	int			saved_max_active; /* WQ: saved pwq max_active */
274 
275 	struct workqueue_attrs	*unbound_attrs;	/* PW: only for unbound wqs */
276 	struct pool_workqueue	*dfl_pwq;	/* PW: only for unbound wqs */
277 
278 #ifdef CONFIG_SYSFS
279 	struct wq_device	*wq_dev;	/* I: for sysfs interface */
280 #endif
281 #ifdef CONFIG_LOCKDEP
282 	char			*lock_name;
283 	struct lock_class_key	key;
284 	struct lockdep_map	lockdep_map;
285 #endif
286 	char			name[WQ_NAME_LEN]; /* I: workqueue name */
287 
288 	/*
289 	 * Destruction of workqueue_struct is RCU protected to allow walking
290 	 * the workqueues list without grabbing wq_pool_mutex.
291 	 * This is used to dump all workqueues from sysrq.
292 	 */
293 	struct rcu_head		rcu;
294 
295 	/* hot fields used during command issue, aligned to cacheline */
296 	unsigned int		flags ____cacheline_aligned; /* WQ: WQ_* flags */
297 	struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
298 	struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
299 };
300 
301 static struct kmem_cache *pwq_cache;
302 
303 static cpumask_var_t *wq_numa_possible_cpumask;
304 					/* possible CPUs of each node */
305 
306 static bool wq_disable_numa;
307 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
308 
309 /* see the comment above the definition of WQ_POWER_EFFICIENT */
310 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
311 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
312 
313 static bool wq_online;			/* can kworkers be created yet? */
314 
315 static bool wq_numa_enabled;		/* unbound NUMA affinity enabled */
316 
317 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
318 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
319 
320 static DEFINE_MUTEX(wq_pool_mutex);	/* protects pools and workqueues list */
321 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
322 static DEFINE_RAW_SPINLOCK(wq_mayday_lock);	/* protects wq->maydays list */
323 /* wait for manager to go away */
324 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
325 
326 static LIST_HEAD(workqueues);		/* PR: list of all workqueues */
327 static bool workqueue_freezing;		/* PL: have wqs started freezing? */
328 
329 /* PL: allowable cpus for unbound wqs and work items */
330 static cpumask_var_t wq_unbound_cpumask;
331 
332 /* CPU where unbound work was last round robin scheduled from this CPU */
333 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
334 
335 /*
336  * Local execution of unbound work items is no longer guaranteed.  The
337  * following always forces round-robin CPU selection on unbound work items
338  * to uncover usages which depend on it.
339  */
340 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
341 static bool wq_debug_force_rr_cpu = true;
342 #else
343 static bool wq_debug_force_rr_cpu = false;
344 #endif
345 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
346 
347 /* the per-cpu worker pools */
348 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
349 
350 static DEFINE_IDR(worker_pool_idr);	/* PR: idr of all pools */
351 
352 /* PL: hash of all unbound pools keyed by pool->attrs */
353 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
354 
355 /* I: attributes used when instantiating standard unbound pools on demand */
356 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
357 
358 /* I: attributes used when instantiating ordered pools on demand */
359 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
360 
361 struct workqueue_struct *system_wq __read_mostly;
362 EXPORT_SYMBOL(system_wq);
363 struct workqueue_struct *system_highpri_wq __read_mostly;
364 EXPORT_SYMBOL_GPL(system_highpri_wq);
365 struct workqueue_struct *system_long_wq __read_mostly;
366 EXPORT_SYMBOL_GPL(system_long_wq);
367 struct workqueue_struct *system_unbound_wq __read_mostly;
368 EXPORT_SYMBOL_GPL(system_unbound_wq);
369 struct workqueue_struct *system_freezable_wq __read_mostly;
370 EXPORT_SYMBOL_GPL(system_freezable_wq);
371 struct workqueue_struct *system_power_efficient_wq __read_mostly;
372 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
373 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
374 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
375 
376 static int worker_thread(void *__worker);
377 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
378 static void show_pwq(struct pool_workqueue *pwq);
379 static void show_one_worker_pool(struct worker_pool *pool);
380 
381 #define CREATE_TRACE_POINTS
382 #include <trace/events/workqueue.h>
383 
384 #define assert_rcu_or_pool_mutex()					\
385 	RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&			\
386 			 !lockdep_is_held(&wq_pool_mutex),		\
387 			 "RCU or wq_pool_mutex should be held")
388 
389 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq)			\
390 	RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&			\
391 			 !lockdep_is_held(&wq->mutex) &&		\
392 			 !lockdep_is_held(&wq_pool_mutex),		\
393 			 "RCU, wq->mutex or wq_pool_mutex should be held")
394 
395 #define for_each_cpu_worker_pool(pool, cpu)				\
396 	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0];		\
397 	     (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
398 	     (pool)++)
399 
400 /**
401  * for_each_pool - iterate through all worker_pools in the system
402  * @pool: iteration cursor
403  * @pi: integer used for iteration
404  *
405  * This must be called either with wq_pool_mutex held or RCU read
406  * locked.  If the pool needs to be used beyond the locking in effect, the
407  * caller is responsible for guaranteeing that the pool stays online.
408  *
409  * The if/else clause exists only for the lockdep assertion and can be
410  * ignored.
411  */
412 #define for_each_pool(pool, pi)						\
413 	idr_for_each_entry(&worker_pool_idr, pool, pi)			\
414 		if (({ assert_rcu_or_pool_mutex(); false; })) { }	\
415 		else
416 
417 /**
418  * for_each_pool_worker - iterate through all workers of a worker_pool
419  * @worker: iteration cursor
420  * @pool: worker_pool to iterate workers of
421  *
422  * This must be called with wq_pool_attach_mutex.
423  *
424  * The if/else clause exists only for the lockdep assertion and can be
425  * ignored.
426  */
427 #define for_each_pool_worker(worker, pool)				\
428 	list_for_each_entry((worker), &(pool)->workers, node)		\
429 		if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
430 		else
431 
432 /**
433  * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
434  * @pwq: iteration cursor
435  * @wq: the target workqueue
436  *
437  * This must be called either with wq->mutex held or RCU read locked.
438  * If the pwq needs to be used beyond the locking in effect, the caller is
439  * responsible for guaranteeing that the pwq stays online.
440  *
441  * The if/else clause exists only for the lockdep assertion and can be
442  * ignored.
443  */
444 #define for_each_pwq(pwq, wq)						\
445 	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node,		\
446 				 lockdep_is_held(&(wq->mutex)))
447 
448 #ifdef CONFIG_DEBUG_OBJECTS_WORK
449 
450 static const struct debug_obj_descr work_debug_descr;
451 
work_debug_hint(void * addr)452 static void *work_debug_hint(void *addr)
453 {
454 	return ((struct work_struct *) addr)->func;
455 }
456 
work_is_static_object(void * addr)457 static bool work_is_static_object(void *addr)
458 {
459 	struct work_struct *work = addr;
460 
461 	return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
462 }
463 
464 /*
465  * fixup_init is called when:
466  * - an active object is initialized
467  */
work_fixup_init(void * addr,enum debug_obj_state state)468 static bool work_fixup_init(void *addr, enum debug_obj_state state)
469 {
470 	struct work_struct *work = addr;
471 
472 	switch (state) {
473 	case ODEBUG_STATE_ACTIVE:
474 		cancel_work_sync(work);
475 		debug_object_init(work, &work_debug_descr);
476 		return true;
477 	default:
478 		return false;
479 	}
480 }
481 
482 /*
483  * fixup_free is called when:
484  * - an active object is freed
485  */
work_fixup_free(void * addr,enum debug_obj_state state)486 static bool work_fixup_free(void *addr, enum debug_obj_state state)
487 {
488 	struct work_struct *work = addr;
489 
490 	switch (state) {
491 	case ODEBUG_STATE_ACTIVE:
492 		cancel_work_sync(work);
493 		debug_object_free(work, &work_debug_descr);
494 		return true;
495 	default:
496 		return false;
497 	}
498 }
499 
500 static const struct debug_obj_descr work_debug_descr = {
501 	.name		= "work_struct",
502 	.debug_hint	= work_debug_hint,
503 	.is_static_object = work_is_static_object,
504 	.fixup_init	= work_fixup_init,
505 	.fixup_free	= work_fixup_free,
506 };
507 
debug_work_activate(struct work_struct * work)508 static inline void debug_work_activate(struct work_struct *work)
509 {
510 	debug_object_activate(work, &work_debug_descr);
511 }
512 
debug_work_deactivate(struct work_struct * work)513 static inline void debug_work_deactivate(struct work_struct *work)
514 {
515 	debug_object_deactivate(work, &work_debug_descr);
516 }
517 
__init_work(struct work_struct * work,int onstack)518 void __init_work(struct work_struct *work, int onstack)
519 {
520 	if (onstack)
521 		debug_object_init_on_stack(work, &work_debug_descr);
522 	else
523 		debug_object_init(work, &work_debug_descr);
524 }
525 EXPORT_SYMBOL_GPL(__init_work);
526 
destroy_work_on_stack(struct work_struct * work)527 void destroy_work_on_stack(struct work_struct *work)
528 {
529 	debug_object_free(work, &work_debug_descr);
530 }
531 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
532 
destroy_delayed_work_on_stack(struct delayed_work * work)533 void destroy_delayed_work_on_stack(struct delayed_work *work)
534 {
535 	destroy_timer_on_stack(&work->timer);
536 	debug_object_free(&work->work, &work_debug_descr);
537 }
538 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
539 
540 #else
debug_work_activate(struct work_struct * work)541 static inline void debug_work_activate(struct work_struct *work) { }
debug_work_deactivate(struct work_struct * work)542 static inline void debug_work_deactivate(struct work_struct *work) { }
543 #endif
544 
545 /**
546  * worker_pool_assign_id - allocate ID and assign it to @pool
547  * @pool: the pool pointer of interest
548  *
549  * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
550  * successfully, -errno on failure.
551  */
worker_pool_assign_id(struct worker_pool * pool)552 static int worker_pool_assign_id(struct worker_pool *pool)
553 {
554 	int ret;
555 
556 	lockdep_assert_held(&wq_pool_mutex);
557 
558 	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
559 			GFP_KERNEL);
560 	if (ret >= 0) {
561 		pool->id = ret;
562 		return 0;
563 	}
564 	return ret;
565 }
566 
567 /**
568  * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
569  * @wq: the target workqueue
570  * @node: the node ID
571  *
572  * This must be called with any of wq_pool_mutex, wq->mutex or RCU
573  * read locked.
574  * If the pwq needs to be used beyond the locking in effect, the caller is
575  * responsible for guaranteeing that the pwq stays online.
576  *
577  * Return: The unbound pool_workqueue for @node.
578  */
unbound_pwq_by_node(struct workqueue_struct * wq,int node)579 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
580 						  int node)
581 {
582 	assert_rcu_or_wq_mutex_or_pool_mutex(wq);
583 
584 	/*
585 	 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
586 	 * delayed item is pending.  The plan is to keep CPU -> NODE
587 	 * mapping valid and stable across CPU on/offlines.  Once that
588 	 * happens, this workaround can be removed.
589 	 */
590 	if (unlikely(node == NUMA_NO_NODE))
591 		return wq->dfl_pwq;
592 
593 	return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
594 }
595 
work_color_to_flags(int color)596 static unsigned int work_color_to_flags(int color)
597 {
598 	return color << WORK_STRUCT_COLOR_SHIFT;
599 }
600 
get_work_color(unsigned long work_data)601 static int get_work_color(unsigned long work_data)
602 {
603 	return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
604 		((1 << WORK_STRUCT_COLOR_BITS) - 1);
605 }
606 
work_next_color(int color)607 static int work_next_color(int color)
608 {
609 	return (color + 1) % WORK_NR_COLORS;
610 }
611 
612 /*
613  * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
614  * contain the pointer to the queued pwq.  Once execution starts, the flag
615  * is cleared and the high bits contain OFFQ flags and pool ID.
616  *
617  * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
618  * and clear_work_data() can be used to set the pwq, pool or clear
619  * work->data.  These functions should only be called while the work is
620  * owned - ie. while the PENDING bit is set.
621  *
622  * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
623  * corresponding to a work.  Pool is available once the work has been
624  * queued anywhere after initialization until it is sync canceled.  pwq is
625  * available only while the work item is queued.
626  *
627  * %WORK_OFFQ_CANCELING is used to mark a work item which is being
628  * canceled.  While being canceled, a work item may have its PENDING set
629  * but stay off timer and worklist for arbitrarily long and nobody should
630  * try to steal the PENDING bit.
631  */
set_work_data(struct work_struct * work,unsigned long data,unsigned long flags)632 static inline void set_work_data(struct work_struct *work, unsigned long data,
633 				 unsigned long flags)
634 {
635 	WARN_ON_ONCE(!work_pending(work));
636 	atomic_long_set(&work->data, data | flags | work_static(work));
637 }
638 
set_work_pwq(struct work_struct * work,struct pool_workqueue * pwq,unsigned long extra_flags)639 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
640 			 unsigned long extra_flags)
641 {
642 	set_work_data(work, (unsigned long)pwq,
643 		      WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
644 }
645 
set_work_pool_and_keep_pending(struct work_struct * work,int pool_id)646 static void set_work_pool_and_keep_pending(struct work_struct *work,
647 					   int pool_id)
648 {
649 	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
650 		      WORK_STRUCT_PENDING);
651 }
652 
set_work_pool_and_clear_pending(struct work_struct * work,int pool_id)653 static void set_work_pool_and_clear_pending(struct work_struct *work,
654 					    int pool_id)
655 {
656 	/*
657 	 * The following wmb is paired with the implied mb in
658 	 * test_and_set_bit(PENDING) and ensures all updates to @work made
659 	 * here are visible to and precede any updates by the next PENDING
660 	 * owner.
661 	 */
662 	smp_wmb();
663 	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
664 	/*
665 	 * The following mb guarantees that previous clear of a PENDING bit
666 	 * will not be reordered with any speculative LOADS or STORES from
667 	 * work->current_func, which is executed afterwards.  This possible
668 	 * reordering can lead to a missed execution on attempt to queue
669 	 * the same @work.  E.g. consider this case:
670 	 *
671 	 *   CPU#0                         CPU#1
672 	 *   ----------------------------  --------------------------------
673 	 *
674 	 * 1  STORE event_indicated
675 	 * 2  queue_work_on() {
676 	 * 3    test_and_set_bit(PENDING)
677 	 * 4 }                             set_..._and_clear_pending() {
678 	 * 5                                 set_work_data() # clear bit
679 	 * 6                                 smp_mb()
680 	 * 7                               work->current_func() {
681 	 * 8				      LOAD event_indicated
682 	 *				   }
683 	 *
684 	 * Without an explicit full barrier speculative LOAD on line 8 can
685 	 * be executed before CPU#0 does STORE on line 1.  If that happens,
686 	 * CPU#0 observes the PENDING bit is still set and new execution of
687 	 * a @work is not queued in a hope, that CPU#1 will eventually
688 	 * finish the queued @work.  Meanwhile CPU#1 does not see
689 	 * event_indicated is set, because speculative LOAD was executed
690 	 * before actual STORE.
691 	 */
692 	smp_mb();
693 }
694 
clear_work_data(struct work_struct * work)695 static void clear_work_data(struct work_struct *work)
696 {
697 	smp_wmb();	/* see set_work_pool_and_clear_pending() */
698 	set_work_data(work, WORK_STRUCT_NO_POOL, 0);
699 }
700 
get_work_pwq(struct work_struct * work)701 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
702 {
703 	unsigned long data = atomic_long_read(&work->data);
704 
705 	if (data & WORK_STRUCT_PWQ)
706 		return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
707 	else
708 		return NULL;
709 }
710 
711 /**
712  * get_work_pool - return the worker_pool a given work was associated with
713  * @work: the work item of interest
714  *
715  * Pools are created and destroyed under wq_pool_mutex, and allows read
716  * access under RCU read lock.  As such, this function should be
717  * called under wq_pool_mutex or inside of a rcu_read_lock() region.
718  *
719  * All fields of the returned pool are accessible as long as the above
720  * mentioned locking is in effect.  If the returned pool needs to be used
721  * beyond the critical section, the caller is responsible for ensuring the
722  * returned pool is and stays online.
723  *
724  * Return: The worker_pool @work was last associated with.  %NULL if none.
725  */
get_work_pool(struct work_struct * work)726 static struct worker_pool *get_work_pool(struct work_struct *work)
727 {
728 	unsigned long data = atomic_long_read(&work->data);
729 	int pool_id;
730 
731 	assert_rcu_or_pool_mutex();
732 
733 	if (data & WORK_STRUCT_PWQ)
734 		return ((struct pool_workqueue *)
735 			(data & WORK_STRUCT_WQ_DATA_MASK))->pool;
736 
737 	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
738 	if (pool_id == WORK_OFFQ_POOL_NONE)
739 		return NULL;
740 
741 	return idr_find(&worker_pool_idr, pool_id);
742 }
743 
744 /**
745  * get_work_pool_id - return the worker pool ID a given work is associated with
746  * @work: the work item of interest
747  *
748  * Return: The worker_pool ID @work was last associated with.
749  * %WORK_OFFQ_POOL_NONE if none.
750  */
get_work_pool_id(struct work_struct * work)751 static int get_work_pool_id(struct work_struct *work)
752 {
753 	unsigned long data = atomic_long_read(&work->data);
754 
755 	if (data & WORK_STRUCT_PWQ)
756 		return ((struct pool_workqueue *)
757 			(data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
758 
759 	return data >> WORK_OFFQ_POOL_SHIFT;
760 }
761 
mark_work_canceling(struct work_struct * work)762 static void mark_work_canceling(struct work_struct *work)
763 {
764 	unsigned long pool_id = get_work_pool_id(work);
765 
766 	pool_id <<= WORK_OFFQ_POOL_SHIFT;
767 	set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
768 }
769 
work_is_canceling(struct work_struct * work)770 static bool work_is_canceling(struct work_struct *work)
771 {
772 	unsigned long data = atomic_long_read(&work->data);
773 
774 	return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
775 }
776 
777 /*
778  * Policy functions.  These define the policies on how the global worker
779  * pools are managed.  Unless noted otherwise, these functions assume that
780  * they're being called with pool->lock held.
781  */
782 
__need_more_worker(struct worker_pool * pool)783 static bool __need_more_worker(struct worker_pool *pool)
784 {
785 	return !pool->nr_running;
786 }
787 
788 /*
789  * Need to wake up a worker?  Called from anything but currently
790  * running workers.
791  *
792  * Note that, because unbound workers never contribute to nr_running, this
793  * function will always return %true for unbound pools as long as the
794  * worklist isn't empty.
795  */
need_more_worker(struct worker_pool * pool)796 static bool need_more_worker(struct worker_pool *pool)
797 {
798 	return !list_empty(&pool->worklist) && __need_more_worker(pool);
799 }
800 
801 /* Can I start working?  Called from busy but !running workers. */
may_start_working(struct worker_pool * pool)802 static bool may_start_working(struct worker_pool *pool)
803 {
804 	return pool->nr_idle;
805 }
806 
807 /* Do I need to keep working?  Called from currently running workers. */
keep_working(struct worker_pool * pool)808 static bool keep_working(struct worker_pool *pool)
809 {
810 	return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
811 }
812 
813 /* Do we need a new worker?  Called from manager. */
need_to_create_worker(struct worker_pool * pool)814 static bool need_to_create_worker(struct worker_pool *pool)
815 {
816 	return need_more_worker(pool) && !may_start_working(pool);
817 }
818 
819 /* Do we have too many workers and should some go away? */
too_many_workers(struct worker_pool * pool)820 static bool too_many_workers(struct worker_pool *pool)
821 {
822 	bool managing = pool->flags & POOL_MANAGER_ACTIVE;
823 	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
824 	int nr_busy = pool->nr_workers - nr_idle;
825 
826 	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
827 }
828 
829 /*
830  * Wake up functions.
831  */
832 
833 /* Return the first idle worker.  Called with pool->lock held. */
first_idle_worker(struct worker_pool * pool)834 static struct worker *first_idle_worker(struct worker_pool *pool)
835 {
836 	if (unlikely(list_empty(&pool->idle_list)))
837 		return NULL;
838 
839 	return list_first_entry(&pool->idle_list, struct worker, entry);
840 }
841 
842 /**
843  * wake_up_worker - wake up an idle worker
844  * @pool: worker pool to wake worker from
845  *
846  * Wake up the first idle worker of @pool.
847  *
848  * CONTEXT:
849  * raw_spin_lock_irq(pool->lock).
850  */
wake_up_worker(struct worker_pool * pool)851 static void wake_up_worker(struct worker_pool *pool)
852 {
853 	struct worker *worker = first_idle_worker(pool);
854 
855 	if (likely(worker))
856 		wake_up_process(worker->task);
857 }
858 
859 /**
860  * wq_worker_running - a worker is running again
861  * @task: task waking up
862  *
863  * This function is called when a worker returns from schedule()
864  */
wq_worker_running(struct task_struct * task)865 void wq_worker_running(struct task_struct *task)
866 {
867 	struct worker *worker = kthread_data(task);
868 
869 	if (!worker->sleeping)
870 		return;
871 
872 	/*
873 	 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
874 	 * and the nr_running increment below, we may ruin the nr_running reset
875 	 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
876 	 * pool. Protect against such race.
877 	 */
878 	preempt_disable();
879 	if (!(worker->flags & WORKER_NOT_RUNNING))
880 		worker->pool->nr_running++;
881 	preempt_enable();
882 	worker->sleeping = 0;
883 }
884 
885 /**
886  * wq_worker_sleeping - a worker is going to sleep
887  * @task: task going to sleep
888  *
889  * This function is called from schedule() when a busy worker is
890  * going to sleep.
891  */
wq_worker_sleeping(struct task_struct * task)892 void wq_worker_sleeping(struct task_struct *task)
893 {
894 	struct worker *worker = kthread_data(task);
895 	struct worker_pool *pool;
896 
897 	/*
898 	 * Rescuers, which may not have all the fields set up like normal
899 	 * workers, also reach here, let's not access anything before
900 	 * checking NOT_RUNNING.
901 	 */
902 	if (worker->flags & WORKER_NOT_RUNNING)
903 		return;
904 
905 	pool = worker->pool;
906 
907 	/* Return if preempted before wq_worker_running() was reached */
908 	if (worker->sleeping)
909 		return;
910 
911 	worker->sleeping = 1;
912 	raw_spin_lock_irq(&pool->lock);
913 
914 	/*
915 	 * Recheck in case unbind_workers() preempted us. We don't
916 	 * want to decrement nr_running after the worker is unbound
917 	 * and nr_running has been reset.
918 	 */
919 	if (worker->flags & WORKER_NOT_RUNNING) {
920 		raw_spin_unlock_irq(&pool->lock);
921 		return;
922 	}
923 
924 	pool->nr_running--;
925 	if (need_more_worker(pool))
926 		wake_up_worker(pool);
927 	raw_spin_unlock_irq(&pool->lock);
928 }
929 
930 /**
931  * wq_worker_last_func - retrieve worker's last work function
932  * @task: Task to retrieve last work function of.
933  *
934  * Determine the last function a worker executed. This is called from
935  * the scheduler to get a worker's last known identity.
936  *
937  * CONTEXT:
938  * raw_spin_lock_irq(rq->lock)
939  *
940  * This function is called during schedule() when a kworker is going
941  * to sleep. It's used by psi to identify aggregation workers during
942  * dequeuing, to allow periodic aggregation to shut-off when that
943  * worker is the last task in the system or cgroup to go to sleep.
944  *
945  * As this function doesn't involve any workqueue-related locking, it
946  * only returns stable values when called from inside the scheduler's
947  * queuing and dequeuing paths, when @task, which must be a kworker,
948  * is guaranteed to not be processing any works.
949  *
950  * Return:
951  * The last work function %current executed as a worker, NULL if it
952  * hasn't executed any work yet.
953  */
wq_worker_last_func(struct task_struct * task)954 work_func_t wq_worker_last_func(struct task_struct *task)
955 {
956 	struct worker *worker = kthread_data(task);
957 
958 	return worker->last_func;
959 }
960 
961 /**
962  * worker_set_flags - set worker flags and adjust nr_running accordingly
963  * @worker: self
964  * @flags: flags to set
965  *
966  * Set @flags in @worker->flags and adjust nr_running accordingly.
967  *
968  * CONTEXT:
969  * raw_spin_lock_irq(pool->lock)
970  */
worker_set_flags(struct worker * worker,unsigned int flags)971 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
972 {
973 	struct worker_pool *pool = worker->pool;
974 
975 	WARN_ON_ONCE(worker->task != current);
976 
977 	/* If transitioning into NOT_RUNNING, adjust nr_running. */
978 	if ((flags & WORKER_NOT_RUNNING) &&
979 	    !(worker->flags & WORKER_NOT_RUNNING)) {
980 		pool->nr_running--;
981 	}
982 
983 	worker->flags |= flags;
984 }
985 
986 /**
987  * worker_clr_flags - clear worker flags and adjust nr_running accordingly
988  * @worker: self
989  * @flags: flags to clear
990  *
991  * Clear @flags in @worker->flags and adjust nr_running accordingly.
992  *
993  * CONTEXT:
994  * raw_spin_lock_irq(pool->lock)
995  */
worker_clr_flags(struct worker * worker,unsigned int flags)996 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
997 {
998 	struct worker_pool *pool = worker->pool;
999 	unsigned int oflags = worker->flags;
1000 
1001 	WARN_ON_ONCE(worker->task != current);
1002 
1003 	worker->flags &= ~flags;
1004 
1005 	/*
1006 	 * If transitioning out of NOT_RUNNING, increment nr_running.  Note
1007 	 * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
1008 	 * of multiple flags, not a single flag.
1009 	 */
1010 	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1011 		if (!(worker->flags & WORKER_NOT_RUNNING))
1012 			pool->nr_running++;
1013 }
1014 
1015 /**
1016  * find_worker_executing_work - find worker which is executing a work
1017  * @pool: pool of interest
1018  * @work: work to find worker for
1019  *
1020  * Find a worker which is executing @work on @pool by searching
1021  * @pool->busy_hash which is keyed by the address of @work.  For a worker
1022  * to match, its current execution should match the address of @work and
1023  * its work function.  This is to avoid unwanted dependency between
1024  * unrelated work executions through a work item being recycled while still
1025  * being executed.
1026  *
1027  * This is a bit tricky.  A work item may be freed once its execution
1028  * starts and nothing prevents the freed area from being recycled for
1029  * another work item.  If the same work item address ends up being reused
1030  * before the original execution finishes, workqueue will identify the
1031  * recycled work item as currently executing and make it wait until the
1032  * current execution finishes, introducing an unwanted dependency.
1033  *
1034  * This function checks the work item address and work function to avoid
1035  * false positives.  Note that this isn't complete as one may construct a
1036  * work function which can introduce dependency onto itself through a
1037  * recycled work item.  Well, if somebody wants to shoot oneself in the
1038  * foot that badly, there's only so much we can do, and if such deadlock
1039  * actually occurs, it should be easy to locate the culprit work function.
1040  *
1041  * CONTEXT:
1042  * raw_spin_lock_irq(pool->lock).
1043  *
1044  * Return:
1045  * Pointer to worker which is executing @work if found, %NULL
1046  * otherwise.
1047  */
find_worker_executing_work(struct worker_pool * pool,struct work_struct * work)1048 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1049 						 struct work_struct *work)
1050 {
1051 	struct worker *worker;
1052 
1053 	hash_for_each_possible(pool->busy_hash, worker, hentry,
1054 			       (unsigned long)work)
1055 		if (worker->current_work == work &&
1056 		    worker->current_func == work->func)
1057 			return worker;
1058 
1059 	return NULL;
1060 }
1061 
1062 /**
1063  * move_linked_works - move linked works to a list
1064  * @work: start of series of works to be scheduled
1065  * @head: target list to append @work to
1066  * @nextp: out parameter for nested worklist walking
1067  *
1068  * Schedule linked works starting from @work to @head.  Work series to
1069  * be scheduled starts at @work and includes any consecutive work with
1070  * WORK_STRUCT_LINKED set in its predecessor.
1071  *
1072  * If @nextp is not NULL, it's updated to point to the next work of
1073  * the last scheduled work.  This allows move_linked_works() to be
1074  * nested inside outer list_for_each_entry_safe().
1075  *
1076  * CONTEXT:
1077  * raw_spin_lock_irq(pool->lock).
1078  */
move_linked_works(struct work_struct * work,struct list_head * head,struct work_struct ** nextp)1079 static void move_linked_works(struct work_struct *work, struct list_head *head,
1080 			      struct work_struct **nextp)
1081 {
1082 	struct work_struct *n;
1083 
1084 	/*
1085 	 * Linked worklist will always end before the end of the list,
1086 	 * use NULL for list head.
1087 	 */
1088 	list_for_each_entry_safe_from(work, n, NULL, entry) {
1089 		list_move_tail(&work->entry, head);
1090 		if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1091 			break;
1092 	}
1093 
1094 	/*
1095 	 * If we're already inside safe list traversal and have moved
1096 	 * multiple works to the scheduled queue, the next position
1097 	 * needs to be updated.
1098 	 */
1099 	if (nextp)
1100 		*nextp = n;
1101 }
1102 
1103 /**
1104  * get_pwq - get an extra reference on the specified pool_workqueue
1105  * @pwq: pool_workqueue to get
1106  *
1107  * Obtain an extra reference on @pwq.  The caller should guarantee that
1108  * @pwq has positive refcnt and be holding the matching pool->lock.
1109  */
get_pwq(struct pool_workqueue * pwq)1110 static void get_pwq(struct pool_workqueue *pwq)
1111 {
1112 	lockdep_assert_held(&pwq->pool->lock);
1113 	WARN_ON_ONCE(pwq->refcnt <= 0);
1114 	pwq->refcnt++;
1115 }
1116 
1117 /**
1118  * put_pwq - put a pool_workqueue reference
1119  * @pwq: pool_workqueue to put
1120  *
1121  * Drop a reference of @pwq.  If its refcnt reaches zero, schedule its
1122  * destruction.  The caller should be holding the matching pool->lock.
1123  */
put_pwq(struct pool_workqueue * pwq)1124 static void put_pwq(struct pool_workqueue *pwq)
1125 {
1126 	lockdep_assert_held(&pwq->pool->lock);
1127 	if (likely(--pwq->refcnt))
1128 		return;
1129 	if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1130 		return;
1131 	/*
1132 	 * @pwq can't be released under pool->lock, bounce to
1133 	 * pwq_unbound_release_workfn().  This never recurses on the same
1134 	 * pool->lock as this path is taken only for unbound workqueues and
1135 	 * the release work item is scheduled on a per-cpu workqueue.  To
1136 	 * avoid lockdep warning, unbound pool->locks are given lockdep
1137 	 * subclass of 1 in get_unbound_pool().
1138 	 */
1139 	schedule_work(&pwq->unbound_release_work);
1140 }
1141 
1142 /**
1143  * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1144  * @pwq: pool_workqueue to put (can be %NULL)
1145  *
1146  * put_pwq() with locking.  This function also allows %NULL @pwq.
1147  */
put_pwq_unlocked(struct pool_workqueue * pwq)1148 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1149 {
1150 	if (pwq) {
1151 		/*
1152 		 * As both pwqs and pools are RCU protected, the
1153 		 * following lock operations are safe.
1154 		 */
1155 		raw_spin_lock_irq(&pwq->pool->lock);
1156 		put_pwq(pwq);
1157 		raw_spin_unlock_irq(&pwq->pool->lock);
1158 	}
1159 }
1160 
pwq_activate_inactive_work(struct work_struct * work)1161 static void pwq_activate_inactive_work(struct work_struct *work)
1162 {
1163 	struct pool_workqueue *pwq = get_work_pwq(work);
1164 
1165 	trace_workqueue_activate_work(work);
1166 	if (list_empty(&pwq->pool->worklist))
1167 		pwq->pool->watchdog_ts = jiffies;
1168 	move_linked_works(work, &pwq->pool->worklist, NULL);
1169 	__clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work));
1170 	pwq->nr_active++;
1171 }
1172 
pwq_activate_first_inactive(struct pool_workqueue * pwq)1173 static void pwq_activate_first_inactive(struct pool_workqueue *pwq)
1174 {
1175 	struct work_struct *work = list_first_entry(&pwq->inactive_works,
1176 						    struct work_struct, entry);
1177 
1178 	pwq_activate_inactive_work(work);
1179 }
1180 
1181 /**
1182  * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1183  * @pwq: pwq of interest
1184  * @work_data: work_data of work which left the queue
1185  *
1186  * A work either has completed or is removed from pending queue,
1187  * decrement nr_in_flight of its pwq and handle workqueue flushing.
1188  *
1189  * CONTEXT:
1190  * raw_spin_lock_irq(pool->lock).
1191  */
pwq_dec_nr_in_flight(struct pool_workqueue * pwq,unsigned long work_data)1192 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1193 {
1194 	int color = get_work_color(work_data);
1195 
1196 	if (!(work_data & WORK_STRUCT_INACTIVE)) {
1197 		pwq->nr_active--;
1198 		if (!list_empty(&pwq->inactive_works)) {
1199 			/* one down, submit an inactive one */
1200 			if (pwq->nr_active < pwq->max_active)
1201 				pwq_activate_first_inactive(pwq);
1202 		}
1203 	}
1204 
1205 	pwq->nr_in_flight[color]--;
1206 
1207 	/* is flush in progress and are we at the flushing tip? */
1208 	if (likely(pwq->flush_color != color))
1209 		goto out_put;
1210 
1211 	/* are there still in-flight works? */
1212 	if (pwq->nr_in_flight[color])
1213 		goto out_put;
1214 
1215 	/* this pwq is done, clear flush_color */
1216 	pwq->flush_color = -1;
1217 
1218 	/*
1219 	 * If this was the last pwq, wake up the first flusher.  It
1220 	 * will handle the rest.
1221 	 */
1222 	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1223 		complete(&pwq->wq->first_flusher->done);
1224 out_put:
1225 	put_pwq(pwq);
1226 }
1227 
1228 /**
1229  * try_to_grab_pending - steal work item from worklist and disable irq
1230  * @work: work item to steal
1231  * @is_dwork: @work is a delayed_work
1232  * @flags: place to store irq state
1233  *
1234  * Try to grab PENDING bit of @work.  This function can handle @work in any
1235  * stable state - idle, on timer or on worklist.
1236  *
1237  * Return:
1238  *
1239  *  ========	================================================================
1240  *  1		if @work was pending and we successfully stole PENDING
1241  *  0		if @work was idle and we claimed PENDING
1242  *  -EAGAIN	if PENDING couldn't be grabbed at the moment, safe to busy-retry
1243  *  -ENOENT	if someone else is canceling @work, this state may persist
1244  *		for arbitrarily long
1245  *  ========	================================================================
1246  *
1247  * Note:
1248  * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1249  * interrupted while holding PENDING and @work off queue, irq must be
1250  * disabled on entry.  This, combined with delayed_work->timer being
1251  * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1252  *
1253  * On successful return, >= 0, irq is disabled and the caller is
1254  * responsible for releasing it using local_irq_restore(*@flags).
1255  *
1256  * This function is safe to call from any context including IRQ handler.
1257  */
try_to_grab_pending(struct work_struct * work,bool is_dwork,unsigned long * flags)1258 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1259 			       unsigned long *flags)
1260 {
1261 	struct worker_pool *pool;
1262 	struct pool_workqueue *pwq;
1263 
1264 	local_irq_save(*flags);
1265 
1266 	/* try to steal the timer if it exists */
1267 	if (is_dwork) {
1268 		struct delayed_work *dwork = to_delayed_work(work);
1269 
1270 		/*
1271 		 * dwork->timer is irqsafe.  If del_timer() fails, it's
1272 		 * guaranteed that the timer is not queued anywhere and not
1273 		 * running on the local CPU.
1274 		 */
1275 		if (likely(del_timer(&dwork->timer)))
1276 			return 1;
1277 	}
1278 
1279 	/* try to claim PENDING the normal way */
1280 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1281 		return 0;
1282 
1283 	rcu_read_lock();
1284 	/*
1285 	 * The queueing is in progress, or it is already queued. Try to
1286 	 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1287 	 */
1288 	pool = get_work_pool(work);
1289 	if (!pool)
1290 		goto fail;
1291 
1292 	raw_spin_lock(&pool->lock);
1293 	/*
1294 	 * work->data is guaranteed to point to pwq only while the work
1295 	 * item is queued on pwq->wq, and both updating work->data to point
1296 	 * to pwq on queueing and to pool on dequeueing are done under
1297 	 * pwq->pool->lock.  This in turn guarantees that, if work->data
1298 	 * points to pwq which is associated with a locked pool, the work
1299 	 * item is currently queued on that pool.
1300 	 */
1301 	pwq = get_work_pwq(work);
1302 	if (pwq && pwq->pool == pool) {
1303 		debug_work_deactivate(work);
1304 
1305 		/*
1306 		 * A cancelable inactive work item must be in the
1307 		 * pwq->inactive_works since a queued barrier can't be
1308 		 * canceled (see the comments in insert_wq_barrier()).
1309 		 *
1310 		 * An inactive work item cannot be grabbed directly because
1311 		 * it might have linked barrier work items which, if left
1312 		 * on the inactive_works list, will confuse pwq->nr_active
1313 		 * management later on and cause stall.  Make sure the work
1314 		 * item is activated before grabbing.
1315 		 */
1316 		if (*work_data_bits(work) & WORK_STRUCT_INACTIVE)
1317 			pwq_activate_inactive_work(work);
1318 
1319 		list_del_init(&work->entry);
1320 		pwq_dec_nr_in_flight(pwq, *work_data_bits(work));
1321 
1322 		/* work->data points to pwq iff queued, point to pool */
1323 		set_work_pool_and_keep_pending(work, pool->id);
1324 
1325 		raw_spin_unlock(&pool->lock);
1326 		rcu_read_unlock();
1327 		return 1;
1328 	}
1329 	raw_spin_unlock(&pool->lock);
1330 fail:
1331 	rcu_read_unlock();
1332 	local_irq_restore(*flags);
1333 	if (work_is_canceling(work))
1334 		return -ENOENT;
1335 	cpu_relax();
1336 	return -EAGAIN;
1337 }
1338 
1339 /**
1340  * insert_work - insert a work into a pool
1341  * @pwq: pwq @work belongs to
1342  * @work: work to insert
1343  * @head: insertion point
1344  * @extra_flags: extra WORK_STRUCT_* flags to set
1345  *
1346  * Insert @work which belongs to @pwq after @head.  @extra_flags is or'd to
1347  * work_struct flags.
1348  *
1349  * CONTEXT:
1350  * raw_spin_lock_irq(pool->lock).
1351  */
insert_work(struct pool_workqueue * pwq,struct work_struct * work,struct list_head * head,unsigned int extra_flags)1352 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1353 			struct list_head *head, unsigned int extra_flags)
1354 {
1355 	struct worker_pool *pool = pwq->pool;
1356 
1357 	/* record the work call stack in order to print it in KASAN reports */
1358 	kasan_record_aux_stack_noalloc(work);
1359 
1360 	/* we own @work, set data and link */
1361 	set_work_pwq(work, pwq, extra_flags);
1362 	list_add_tail(&work->entry, head);
1363 	get_pwq(pwq);
1364 
1365 	if (__need_more_worker(pool))
1366 		wake_up_worker(pool);
1367 }
1368 
1369 /*
1370  * Test whether @work is being queued from another work executing on the
1371  * same workqueue.
1372  */
is_chained_work(struct workqueue_struct * wq)1373 static bool is_chained_work(struct workqueue_struct *wq)
1374 {
1375 	struct worker *worker;
1376 
1377 	worker = current_wq_worker();
1378 	/*
1379 	 * Return %true iff I'm a worker executing a work item on @wq.  If
1380 	 * I'm @worker, it's safe to dereference it without locking.
1381 	 */
1382 	return worker && worker->current_pwq->wq == wq;
1383 }
1384 
1385 /*
1386  * When queueing an unbound work item to a wq, prefer local CPU if allowed
1387  * by wq_unbound_cpumask.  Otherwise, round robin among the allowed ones to
1388  * avoid perturbing sensitive tasks.
1389  */
wq_select_unbound_cpu(int cpu)1390 static int wq_select_unbound_cpu(int cpu)
1391 {
1392 	static bool printed_dbg_warning;
1393 	int new_cpu;
1394 
1395 	if (likely(!wq_debug_force_rr_cpu)) {
1396 		if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1397 			return cpu;
1398 	} else if (!printed_dbg_warning) {
1399 		pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1400 		printed_dbg_warning = true;
1401 	}
1402 
1403 	if (cpumask_empty(wq_unbound_cpumask))
1404 		return cpu;
1405 
1406 	new_cpu = __this_cpu_read(wq_rr_cpu_last);
1407 	new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1408 	if (unlikely(new_cpu >= nr_cpu_ids)) {
1409 		new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1410 		if (unlikely(new_cpu >= nr_cpu_ids))
1411 			return cpu;
1412 	}
1413 	__this_cpu_write(wq_rr_cpu_last, new_cpu);
1414 
1415 	return new_cpu;
1416 }
1417 
__queue_work(int cpu,struct workqueue_struct * wq,struct work_struct * work)1418 static void __queue_work(int cpu, struct workqueue_struct *wq,
1419 			 struct work_struct *work)
1420 {
1421 	struct pool_workqueue *pwq;
1422 	struct worker_pool *last_pool;
1423 	struct list_head *worklist;
1424 	unsigned int work_flags;
1425 	unsigned int req_cpu = cpu;
1426 
1427 	/*
1428 	 * While a work item is PENDING && off queue, a task trying to
1429 	 * steal the PENDING will busy-loop waiting for it to either get
1430 	 * queued or lose PENDING.  Grabbing PENDING and queueing should
1431 	 * happen with IRQ disabled.
1432 	 */
1433 	lockdep_assert_irqs_disabled();
1434 
1435 
1436 	/* if draining, only works from the same workqueue are allowed */
1437 	if (unlikely(wq->flags & __WQ_DRAINING) &&
1438 	    WARN_ON_ONCE(!is_chained_work(wq)))
1439 		return;
1440 	rcu_read_lock();
1441 retry:
1442 	/* pwq which will be used unless @work is executing elsewhere */
1443 	if (wq->flags & WQ_UNBOUND) {
1444 		if (req_cpu == WORK_CPU_UNBOUND)
1445 			cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1446 		pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1447 	} else {
1448 		if (req_cpu == WORK_CPU_UNBOUND)
1449 			cpu = raw_smp_processor_id();
1450 		pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1451 	}
1452 
1453 	/*
1454 	 * If @work was previously on a different pool, it might still be
1455 	 * running there, in which case the work needs to be queued on that
1456 	 * pool to guarantee non-reentrancy.
1457 	 */
1458 	last_pool = get_work_pool(work);
1459 	if (last_pool && last_pool != pwq->pool) {
1460 		struct worker *worker;
1461 
1462 		raw_spin_lock(&last_pool->lock);
1463 
1464 		worker = find_worker_executing_work(last_pool, work);
1465 
1466 		if (worker && worker->current_pwq->wq == wq) {
1467 			pwq = worker->current_pwq;
1468 		} else {
1469 			/* meh... not running there, queue here */
1470 			raw_spin_unlock(&last_pool->lock);
1471 			raw_spin_lock(&pwq->pool->lock);
1472 		}
1473 	} else {
1474 		raw_spin_lock(&pwq->pool->lock);
1475 	}
1476 
1477 	/*
1478 	 * pwq is determined and locked.  For unbound pools, we could have
1479 	 * raced with pwq release and it could already be dead.  If its
1480 	 * refcnt is zero, repeat pwq selection.  Note that pwqs never die
1481 	 * without another pwq replacing it in the numa_pwq_tbl or while
1482 	 * work items are executing on it, so the retrying is guaranteed to
1483 	 * make forward-progress.
1484 	 */
1485 	if (unlikely(!pwq->refcnt)) {
1486 		if (wq->flags & WQ_UNBOUND) {
1487 			raw_spin_unlock(&pwq->pool->lock);
1488 			cpu_relax();
1489 			goto retry;
1490 		}
1491 		/* oops */
1492 		WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1493 			  wq->name, cpu);
1494 	}
1495 
1496 	/* pwq determined, queue */
1497 	trace_workqueue_queue_work(req_cpu, pwq, work);
1498 
1499 	if (WARN_ON(!list_empty(&work->entry)))
1500 		goto out;
1501 
1502 	pwq->nr_in_flight[pwq->work_color]++;
1503 	work_flags = work_color_to_flags(pwq->work_color);
1504 
1505 	if (likely(pwq->nr_active < pwq->max_active)) {
1506 		trace_workqueue_activate_work(work);
1507 		pwq->nr_active++;
1508 		worklist = &pwq->pool->worklist;
1509 		if (list_empty(worklist))
1510 			pwq->pool->watchdog_ts = jiffies;
1511 	} else {
1512 		work_flags |= WORK_STRUCT_INACTIVE;
1513 		worklist = &pwq->inactive_works;
1514 	}
1515 
1516 	debug_work_activate(work);
1517 	insert_work(pwq, work, worklist, work_flags);
1518 
1519 out:
1520 	raw_spin_unlock(&pwq->pool->lock);
1521 	rcu_read_unlock();
1522 }
1523 
1524 /**
1525  * queue_work_on - queue work on specific cpu
1526  * @cpu: CPU number to execute work on
1527  * @wq: workqueue to use
1528  * @work: work to queue
1529  *
1530  * We queue the work to a specific CPU, the caller must ensure it
1531  * can't go away.  Callers that fail to ensure that the specified
1532  * CPU cannot go away will execute on a randomly chosen CPU.
1533  *
1534  * Return: %false if @work was already on a queue, %true otherwise.
1535  */
queue_work_on(int cpu,struct workqueue_struct * wq,struct work_struct * work)1536 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1537 		   struct work_struct *work)
1538 {
1539 	bool ret = false;
1540 	unsigned long flags;
1541 
1542 	local_irq_save(flags);
1543 
1544 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1545 		__queue_work(cpu, wq, work);
1546 		ret = true;
1547 	}
1548 
1549 	local_irq_restore(flags);
1550 	return ret;
1551 }
1552 EXPORT_SYMBOL(queue_work_on);
1553 
1554 /**
1555  * workqueue_select_cpu_near - Select a CPU based on NUMA node
1556  * @node: NUMA node ID that we want to select a CPU from
1557  *
1558  * This function will attempt to find a "random" cpu available on a given
1559  * node. If there are no CPUs available on the given node it will return
1560  * WORK_CPU_UNBOUND indicating that we should just schedule to any
1561  * available CPU if we need to schedule this work.
1562  */
workqueue_select_cpu_near(int node)1563 static int workqueue_select_cpu_near(int node)
1564 {
1565 	int cpu;
1566 
1567 	/* No point in doing this if NUMA isn't enabled for workqueues */
1568 	if (!wq_numa_enabled)
1569 		return WORK_CPU_UNBOUND;
1570 
1571 	/* Delay binding to CPU if node is not valid or online */
1572 	if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1573 		return WORK_CPU_UNBOUND;
1574 
1575 	/* Use local node/cpu if we are already there */
1576 	cpu = raw_smp_processor_id();
1577 	if (node == cpu_to_node(cpu))
1578 		return cpu;
1579 
1580 	/* Use "random" otherwise know as "first" online CPU of node */
1581 	cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1582 
1583 	/* If CPU is valid return that, otherwise just defer */
1584 	return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1585 }
1586 
1587 /**
1588  * queue_work_node - queue work on a "random" cpu for a given NUMA node
1589  * @node: NUMA node that we are targeting the work for
1590  * @wq: workqueue to use
1591  * @work: work to queue
1592  *
1593  * We queue the work to a "random" CPU within a given NUMA node. The basic
1594  * idea here is to provide a way to somehow associate work with a given
1595  * NUMA node.
1596  *
1597  * This function will only make a best effort attempt at getting this onto
1598  * the right NUMA node. If no node is requested or the requested node is
1599  * offline then we just fall back to standard queue_work behavior.
1600  *
1601  * Currently the "random" CPU ends up being the first available CPU in the
1602  * intersection of cpu_online_mask and the cpumask of the node, unless we
1603  * are running on the node. In that case we just use the current CPU.
1604  *
1605  * Return: %false if @work was already on a queue, %true otherwise.
1606  */
queue_work_node(int node,struct workqueue_struct * wq,struct work_struct * work)1607 bool queue_work_node(int node, struct workqueue_struct *wq,
1608 		     struct work_struct *work)
1609 {
1610 	unsigned long flags;
1611 	bool ret = false;
1612 
1613 	/*
1614 	 * This current implementation is specific to unbound workqueues.
1615 	 * Specifically we only return the first available CPU for a given
1616 	 * node instead of cycling through individual CPUs within the node.
1617 	 *
1618 	 * If this is used with a per-cpu workqueue then the logic in
1619 	 * workqueue_select_cpu_near would need to be updated to allow for
1620 	 * some round robin type logic.
1621 	 */
1622 	WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1623 
1624 	local_irq_save(flags);
1625 
1626 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1627 		int cpu = workqueue_select_cpu_near(node);
1628 
1629 		__queue_work(cpu, wq, work);
1630 		ret = true;
1631 	}
1632 
1633 	local_irq_restore(flags);
1634 	return ret;
1635 }
1636 EXPORT_SYMBOL_GPL(queue_work_node);
1637 
delayed_work_timer_fn(struct timer_list * t)1638 void delayed_work_timer_fn(struct timer_list *t)
1639 {
1640 	struct delayed_work *dwork = from_timer(dwork, t, timer);
1641 
1642 	/* should have been called from irqsafe timer with irq already off */
1643 	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
1644 }
1645 EXPORT_SYMBOL(delayed_work_timer_fn);
1646 
__queue_delayed_work(int cpu,struct workqueue_struct * wq,struct delayed_work * dwork,unsigned long delay)1647 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1648 				struct delayed_work *dwork, unsigned long delay)
1649 {
1650 	struct timer_list *timer = &dwork->timer;
1651 	struct work_struct *work = &dwork->work;
1652 
1653 	WARN_ON_ONCE(!wq);
1654 	WARN_ON_FUNCTION_MISMATCH(timer->function, delayed_work_timer_fn);
1655 	WARN_ON_ONCE(timer_pending(timer));
1656 	WARN_ON_ONCE(!list_empty(&work->entry));
1657 
1658 	/*
1659 	 * If @delay is 0, queue @dwork->work immediately.  This is for
1660 	 * both optimization and correctness.  The earliest @timer can
1661 	 * expire is on the closest next tick and delayed_work users depend
1662 	 * on that there's no such delay when @delay is 0.
1663 	 */
1664 	if (!delay) {
1665 		__queue_work(cpu, wq, &dwork->work);
1666 		return;
1667 	}
1668 
1669 	dwork->wq = wq;
1670 	dwork->cpu = cpu;
1671 	timer->expires = jiffies + delay;
1672 
1673 	if (unlikely(cpu != WORK_CPU_UNBOUND))
1674 		add_timer_on(timer, cpu);
1675 	else
1676 		add_timer(timer);
1677 }
1678 
1679 /**
1680  * queue_delayed_work_on - queue work on specific CPU after delay
1681  * @cpu: CPU number to execute work on
1682  * @wq: workqueue to use
1683  * @dwork: work to queue
1684  * @delay: number of jiffies to wait before queueing
1685  *
1686  * Return: %false if @work was already on a queue, %true otherwise.  If
1687  * @delay is zero and @dwork is idle, it will be scheduled for immediate
1688  * execution.
1689  */
queue_delayed_work_on(int cpu,struct workqueue_struct * wq,struct delayed_work * dwork,unsigned long delay)1690 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1691 			   struct delayed_work *dwork, unsigned long delay)
1692 {
1693 	struct work_struct *work = &dwork->work;
1694 	bool ret = false;
1695 	unsigned long flags;
1696 
1697 	/* read the comment in __queue_work() */
1698 	local_irq_save(flags);
1699 
1700 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1701 		__queue_delayed_work(cpu, wq, dwork, delay);
1702 		ret = true;
1703 	}
1704 
1705 	local_irq_restore(flags);
1706 	return ret;
1707 }
1708 EXPORT_SYMBOL(queue_delayed_work_on);
1709 
1710 /**
1711  * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1712  * @cpu: CPU number to execute work on
1713  * @wq: workqueue to use
1714  * @dwork: work to queue
1715  * @delay: number of jiffies to wait before queueing
1716  *
1717  * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1718  * modify @dwork's timer so that it expires after @delay.  If @delay is
1719  * zero, @work is guaranteed to be scheduled immediately regardless of its
1720  * current state.
1721  *
1722  * Return: %false if @dwork was idle and queued, %true if @dwork was
1723  * pending and its timer was modified.
1724  *
1725  * This function is safe to call from any context including IRQ handler.
1726  * See try_to_grab_pending() for details.
1727  */
mod_delayed_work_on(int cpu,struct workqueue_struct * wq,struct delayed_work * dwork,unsigned long delay)1728 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1729 			 struct delayed_work *dwork, unsigned long delay)
1730 {
1731 	unsigned long flags;
1732 	int ret;
1733 
1734 	do {
1735 		ret = try_to_grab_pending(&dwork->work, true, &flags);
1736 	} while (unlikely(ret == -EAGAIN));
1737 
1738 	if (likely(ret >= 0)) {
1739 		__queue_delayed_work(cpu, wq, dwork, delay);
1740 		local_irq_restore(flags);
1741 	}
1742 
1743 	/* -ENOENT from try_to_grab_pending() becomes %true */
1744 	return ret;
1745 }
1746 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1747 
rcu_work_rcufn(struct rcu_head * rcu)1748 static void rcu_work_rcufn(struct rcu_head *rcu)
1749 {
1750 	struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1751 
1752 	/* read the comment in __queue_work() */
1753 	local_irq_disable();
1754 	__queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1755 	local_irq_enable();
1756 }
1757 
1758 /**
1759  * queue_rcu_work - queue work after a RCU grace period
1760  * @wq: workqueue to use
1761  * @rwork: work to queue
1762  *
1763  * Return: %false if @rwork was already pending, %true otherwise.  Note
1764  * that a full RCU grace period is guaranteed only after a %true return.
1765  * While @rwork is guaranteed to be executed after a %false return, the
1766  * execution may happen before a full RCU grace period has passed.
1767  */
queue_rcu_work(struct workqueue_struct * wq,struct rcu_work * rwork)1768 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1769 {
1770 	struct work_struct *work = &rwork->work;
1771 
1772 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1773 		rwork->wq = wq;
1774 		call_rcu(&rwork->rcu, rcu_work_rcufn);
1775 		return true;
1776 	}
1777 
1778 	return false;
1779 }
1780 EXPORT_SYMBOL(queue_rcu_work);
1781 
1782 /**
1783  * worker_enter_idle - enter idle state
1784  * @worker: worker which is entering idle state
1785  *
1786  * @worker is entering idle state.  Update stats and idle timer if
1787  * necessary.
1788  *
1789  * LOCKING:
1790  * raw_spin_lock_irq(pool->lock).
1791  */
worker_enter_idle(struct worker * worker)1792 static void worker_enter_idle(struct worker *worker)
1793 {
1794 	struct worker_pool *pool = worker->pool;
1795 
1796 	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1797 	    WARN_ON_ONCE(!list_empty(&worker->entry) &&
1798 			 (worker->hentry.next || worker->hentry.pprev)))
1799 		return;
1800 
1801 	/* can't use worker_set_flags(), also called from create_worker() */
1802 	worker->flags |= WORKER_IDLE;
1803 	pool->nr_idle++;
1804 	worker->last_active = jiffies;
1805 
1806 	/* idle_list is LIFO */
1807 	list_add(&worker->entry, &pool->idle_list);
1808 
1809 	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1810 		mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1811 
1812 	/* Sanity check nr_running. */
1813 	WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1814 }
1815 
1816 /**
1817  * worker_leave_idle - leave idle state
1818  * @worker: worker which is leaving idle state
1819  *
1820  * @worker is leaving idle state.  Update stats.
1821  *
1822  * LOCKING:
1823  * raw_spin_lock_irq(pool->lock).
1824  */
worker_leave_idle(struct worker * worker)1825 static void worker_leave_idle(struct worker *worker)
1826 {
1827 	struct worker_pool *pool = worker->pool;
1828 
1829 	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1830 		return;
1831 	worker_clr_flags(worker, WORKER_IDLE);
1832 	pool->nr_idle--;
1833 	list_del_init(&worker->entry);
1834 }
1835 
alloc_worker(int node)1836 static struct worker *alloc_worker(int node)
1837 {
1838 	struct worker *worker;
1839 
1840 	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1841 	if (worker) {
1842 		INIT_LIST_HEAD(&worker->entry);
1843 		INIT_LIST_HEAD(&worker->scheduled);
1844 		INIT_LIST_HEAD(&worker->node);
1845 		/* on creation a worker is in !idle && prep state */
1846 		worker->flags = WORKER_PREP;
1847 	}
1848 	return worker;
1849 }
1850 
1851 /**
1852  * worker_attach_to_pool() - attach a worker to a pool
1853  * @worker: worker to be attached
1854  * @pool: the target pool
1855  *
1856  * Attach @worker to @pool.  Once attached, the %WORKER_UNBOUND flag and
1857  * cpu-binding of @worker are kept coordinated with the pool across
1858  * cpu-[un]hotplugs.
1859  */
worker_attach_to_pool(struct worker * worker,struct worker_pool * pool)1860 static void worker_attach_to_pool(struct worker *worker,
1861 				   struct worker_pool *pool)
1862 {
1863 	mutex_lock(&wq_pool_attach_mutex);
1864 
1865 	/*
1866 	 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1867 	 * stable across this function.  See the comments above the flag
1868 	 * definition for details.
1869 	 */
1870 	if (pool->flags & POOL_DISASSOCIATED)
1871 		worker->flags |= WORKER_UNBOUND;
1872 	else
1873 		kthread_set_per_cpu(worker->task, pool->cpu);
1874 
1875 	if (worker->rescue_wq)
1876 		set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1877 
1878 	list_add_tail(&worker->node, &pool->workers);
1879 	worker->pool = pool;
1880 
1881 	mutex_unlock(&wq_pool_attach_mutex);
1882 }
1883 
1884 /**
1885  * worker_detach_from_pool() - detach a worker from its pool
1886  * @worker: worker which is attached to its pool
1887  *
1888  * Undo the attaching which had been done in worker_attach_to_pool().  The
1889  * caller worker shouldn't access to the pool after detached except it has
1890  * other reference to the pool.
1891  */
worker_detach_from_pool(struct worker * worker)1892 static void worker_detach_from_pool(struct worker *worker)
1893 {
1894 	struct worker_pool *pool = worker->pool;
1895 	struct completion *detach_completion = NULL;
1896 
1897 	mutex_lock(&wq_pool_attach_mutex);
1898 
1899 	kthread_set_per_cpu(worker->task, -1);
1900 	list_del(&worker->node);
1901 	worker->pool = NULL;
1902 
1903 	if (list_empty(&pool->workers))
1904 		detach_completion = pool->detach_completion;
1905 	mutex_unlock(&wq_pool_attach_mutex);
1906 
1907 	/* clear leftover flags without pool->lock after it is detached */
1908 	worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1909 
1910 	if (detach_completion)
1911 		complete(detach_completion);
1912 }
1913 
1914 /**
1915  * create_worker - create a new workqueue worker
1916  * @pool: pool the new worker will belong to
1917  *
1918  * Create and start a new worker which is attached to @pool.
1919  *
1920  * CONTEXT:
1921  * Might sleep.  Does GFP_KERNEL allocations.
1922  *
1923  * Return:
1924  * Pointer to the newly created worker.
1925  */
create_worker(struct worker_pool * pool)1926 static struct worker *create_worker(struct worker_pool *pool)
1927 {
1928 	struct worker *worker;
1929 	int id;
1930 	char id_buf[16];
1931 
1932 	/* ID is needed to determine kthread name */
1933 	id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
1934 	if (id < 0)
1935 		return NULL;
1936 
1937 	worker = alloc_worker(pool->node);
1938 	if (!worker)
1939 		goto fail;
1940 
1941 	worker->id = id;
1942 
1943 	if (pool->cpu >= 0)
1944 		snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1945 			 pool->attrs->nice < 0  ? "H" : "");
1946 	else
1947 		snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1948 
1949 	worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1950 					      "kworker/%s", id_buf);
1951 	if (IS_ERR(worker->task))
1952 		goto fail;
1953 
1954 	set_user_nice(worker->task, pool->attrs->nice);
1955 	kthread_bind_mask(worker->task, pool->attrs->cpumask);
1956 
1957 	/* successful, attach the worker to the pool */
1958 	worker_attach_to_pool(worker, pool);
1959 
1960 	/* start the newly created worker */
1961 	raw_spin_lock_irq(&pool->lock);
1962 	worker->pool->nr_workers++;
1963 	worker_enter_idle(worker);
1964 	wake_up_process(worker->task);
1965 	raw_spin_unlock_irq(&pool->lock);
1966 
1967 	return worker;
1968 
1969 fail:
1970 	ida_free(&pool->worker_ida, id);
1971 	kfree(worker);
1972 	return NULL;
1973 }
1974 
1975 /**
1976  * destroy_worker - destroy a workqueue worker
1977  * @worker: worker to be destroyed
1978  *
1979  * Destroy @worker and adjust @pool stats accordingly.  The worker should
1980  * be idle.
1981  *
1982  * CONTEXT:
1983  * raw_spin_lock_irq(pool->lock).
1984  */
destroy_worker(struct worker * worker)1985 static void destroy_worker(struct worker *worker)
1986 {
1987 	struct worker_pool *pool = worker->pool;
1988 
1989 	lockdep_assert_held(&pool->lock);
1990 
1991 	/* sanity check frenzy */
1992 	if (WARN_ON(worker->current_work) ||
1993 	    WARN_ON(!list_empty(&worker->scheduled)) ||
1994 	    WARN_ON(!(worker->flags & WORKER_IDLE)))
1995 		return;
1996 
1997 	pool->nr_workers--;
1998 	pool->nr_idle--;
1999 
2000 	list_del_init(&worker->entry);
2001 	worker->flags |= WORKER_DIE;
2002 	wake_up_process(worker->task);
2003 }
2004 
idle_worker_timeout(struct timer_list * t)2005 static void idle_worker_timeout(struct timer_list *t)
2006 {
2007 	struct worker_pool *pool = from_timer(pool, t, idle_timer);
2008 
2009 	raw_spin_lock_irq(&pool->lock);
2010 
2011 	while (too_many_workers(pool)) {
2012 		struct worker *worker;
2013 		unsigned long expires;
2014 
2015 		/* idle_list is kept in LIFO order, check the last one */
2016 		worker = list_entry(pool->idle_list.prev, struct worker, entry);
2017 		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2018 
2019 		if (time_before(jiffies, expires)) {
2020 			mod_timer(&pool->idle_timer, expires);
2021 			break;
2022 		}
2023 
2024 		destroy_worker(worker);
2025 	}
2026 
2027 	raw_spin_unlock_irq(&pool->lock);
2028 }
2029 
send_mayday(struct work_struct * work)2030 static void send_mayday(struct work_struct *work)
2031 {
2032 	struct pool_workqueue *pwq = get_work_pwq(work);
2033 	struct workqueue_struct *wq = pwq->wq;
2034 
2035 	lockdep_assert_held(&wq_mayday_lock);
2036 
2037 	if (!wq->rescuer)
2038 		return;
2039 
2040 	/* mayday mayday mayday */
2041 	if (list_empty(&pwq->mayday_node)) {
2042 		/*
2043 		 * If @pwq is for an unbound wq, its base ref may be put at
2044 		 * any time due to an attribute change.  Pin @pwq until the
2045 		 * rescuer is done with it.
2046 		 */
2047 		get_pwq(pwq);
2048 		list_add_tail(&pwq->mayday_node, &wq->maydays);
2049 		wake_up_process(wq->rescuer->task);
2050 	}
2051 }
2052 
pool_mayday_timeout(struct timer_list * t)2053 static void pool_mayday_timeout(struct timer_list *t)
2054 {
2055 	struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2056 	struct work_struct *work;
2057 
2058 	raw_spin_lock_irq(&pool->lock);
2059 	raw_spin_lock(&wq_mayday_lock);		/* for wq->maydays */
2060 
2061 	if (need_to_create_worker(pool)) {
2062 		/*
2063 		 * We've been trying to create a new worker but
2064 		 * haven't been successful.  We might be hitting an
2065 		 * allocation deadlock.  Send distress signals to
2066 		 * rescuers.
2067 		 */
2068 		list_for_each_entry(work, &pool->worklist, entry)
2069 			send_mayday(work);
2070 	}
2071 
2072 	raw_spin_unlock(&wq_mayday_lock);
2073 	raw_spin_unlock_irq(&pool->lock);
2074 
2075 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2076 }
2077 
2078 /**
2079  * maybe_create_worker - create a new worker if necessary
2080  * @pool: pool to create a new worker for
2081  *
2082  * Create a new worker for @pool if necessary.  @pool is guaranteed to
2083  * have at least one idle worker on return from this function.  If
2084  * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2085  * sent to all rescuers with works scheduled on @pool to resolve
2086  * possible allocation deadlock.
2087  *
2088  * On return, need_to_create_worker() is guaranteed to be %false and
2089  * may_start_working() %true.
2090  *
2091  * LOCKING:
2092  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2093  * multiple times.  Does GFP_KERNEL allocations.  Called only from
2094  * manager.
2095  */
maybe_create_worker(struct worker_pool * pool)2096 static void maybe_create_worker(struct worker_pool *pool)
2097 __releases(&pool->lock)
2098 __acquires(&pool->lock)
2099 {
2100 restart:
2101 	raw_spin_unlock_irq(&pool->lock);
2102 
2103 	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2104 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2105 
2106 	while (true) {
2107 		if (create_worker(pool) || !need_to_create_worker(pool))
2108 			break;
2109 
2110 		schedule_timeout_interruptible(CREATE_COOLDOWN);
2111 
2112 		if (!need_to_create_worker(pool))
2113 			break;
2114 	}
2115 
2116 	del_timer_sync(&pool->mayday_timer);
2117 	raw_spin_lock_irq(&pool->lock);
2118 	/*
2119 	 * This is necessary even after a new worker was just successfully
2120 	 * created as @pool->lock was dropped and the new worker might have
2121 	 * already become busy.
2122 	 */
2123 	if (need_to_create_worker(pool))
2124 		goto restart;
2125 }
2126 
2127 /**
2128  * manage_workers - manage worker pool
2129  * @worker: self
2130  *
2131  * Assume the manager role and manage the worker pool @worker belongs
2132  * to.  At any given time, there can be only zero or one manager per
2133  * pool.  The exclusion is handled automatically by this function.
2134  *
2135  * The caller can safely start processing works on false return.  On
2136  * true return, it's guaranteed that need_to_create_worker() is false
2137  * and may_start_working() is true.
2138  *
2139  * CONTEXT:
2140  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2141  * multiple times.  Does GFP_KERNEL allocations.
2142  *
2143  * Return:
2144  * %false if the pool doesn't need management and the caller can safely
2145  * start processing works, %true if management function was performed and
2146  * the conditions that the caller verified before calling the function may
2147  * no longer be true.
2148  */
manage_workers(struct worker * worker)2149 static bool manage_workers(struct worker *worker)
2150 {
2151 	struct worker_pool *pool = worker->pool;
2152 
2153 	if (pool->flags & POOL_MANAGER_ACTIVE)
2154 		return false;
2155 
2156 	pool->flags |= POOL_MANAGER_ACTIVE;
2157 	pool->manager = worker;
2158 
2159 	maybe_create_worker(pool);
2160 
2161 	pool->manager = NULL;
2162 	pool->flags &= ~POOL_MANAGER_ACTIVE;
2163 	rcuwait_wake_up(&manager_wait);
2164 	return true;
2165 }
2166 
2167 /**
2168  * process_one_work - process single work
2169  * @worker: self
2170  * @work: work to process
2171  *
2172  * Process @work.  This function contains all the logics necessary to
2173  * process a single work including synchronization against and
2174  * interaction with other workers on the same cpu, queueing and
2175  * flushing.  As long as context requirement is met, any worker can
2176  * call this function to process a work.
2177  *
2178  * CONTEXT:
2179  * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2180  */
process_one_work(struct worker * worker,struct work_struct * work)2181 static void process_one_work(struct worker *worker, struct work_struct *work)
2182 __releases(&pool->lock)
2183 __acquires(&pool->lock)
2184 {
2185 	struct pool_workqueue *pwq = get_work_pwq(work);
2186 	struct worker_pool *pool = worker->pool;
2187 	bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2188 	unsigned long work_data;
2189 	struct worker *collision;
2190 #ifdef CONFIG_LOCKDEP
2191 	/*
2192 	 * It is permissible to free the struct work_struct from
2193 	 * inside the function that is called from it, this we need to
2194 	 * take into account for lockdep too.  To avoid bogus "held
2195 	 * lock freed" warnings as well as problems when looking into
2196 	 * work->lockdep_map, make a copy and use that here.
2197 	 */
2198 	struct lockdep_map lockdep_map;
2199 
2200 	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2201 #endif
2202 	/* ensure we're on the correct CPU */
2203 	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2204 		     raw_smp_processor_id() != pool->cpu);
2205 
2206 	/*
2207 	 * A single work shouldn't be executed concurrently by
2208 	 * multiple workers on a single cpu.  Check whether anyone is
2209 	 * already processing the work.  If so, defer the work to the
2210 	 * currently executing one.
2211 	 */
2212 	collision = find_worker_executing_work(pool, work);
2213 	if (unlikely(collision)) {
2214 		move_linked_works(work, &collision->scheduled, NULL);
2215 		return;
2216 	}
2217 
2218 	/* claim and dequeue */
2219 	debug_work_deactivate(work);
2220 	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2221 	worker->current_work = work;
2222 	worker->current_func = work->func;
2223 	worker->current_pwq = pwq;
2224 	work_data = *work_data_bits(work);
2225 	worker->current_color = get_work_color(work_data);
2226 
2227 	/*
2228 	 * Record wq name for cmdline and debug reporting, may get
2229 	 * overridden through set_worker_desc().
2230 	 */
2231 	strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2232 
2233 	list_del_init(&work->entry);
2234 
2235 	/*
2236 	 * CPU intensive works don't participate in concurrency management.
2237 	 * They're the scheduler's responsibility.  This takes @worker out
2238 	 * of concurrency management and the next code block will chain
2239 	 * execution of the pending work items.
2240 	 */
2241 	if (unlikely(cpu_intensive))
2242 		worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2243 
2244 	/*
2245 	 * Wake up another worker if necessary.  The condition is always
2246 	 * false for normal per-cpu workers since nr_running would always
2247 	 * be >= 1 at this point.  This is used to chain execution of the
2248 	 * pending work items for WORKER_NOT_RUNNING workers such as the
2249 	 * UNBOUND and CPU_INTENSIVE ones.
2250 	 */
2251 	if (need_more_worker(pool))
2252 		wake_up_worker(pool);
2253 
2254 	/*
2255 	 * Record the last pool and clear PENDING which should be the last
2256 	 * update to @work.  Also, do this inside @pool->lock so that
2257 	 * PENDING and queued state changes happen together while IRQ is
2258 	 * disabled.
2259 	 */
2260 	set_work_pool_and_clear_pending(work, pool->id);
2261 
2262 	raw_spin_unlock_irq(&pool->lock);
2263 
2264 	lock_map_acquire(&pwq->wq->lockdep_map);
2265 	lock_map_acquire(&lockdep_map);
2266 	/*
2267 	 * Strictly speaking we should mark the invariant state without holding
2268 	 * any locks, that is, before these two lock_map_acquire()'s.
2269 	 *
2270 	 * However, that would result in:
2271 	 *
2272 	 *   A(W1)
2273 	 *   WFC(C)
2274 	 *		A(W1)
2275 	 *		C(C)
2276 	 *
2277 	 * Which would create W1->C->W1 dependencies, even though there is no
2278 	 * actual deadlock possible. There are two solutions, using a
2279 	 * read-recursive acquire on the work(queue) 'locks', but this will then
2280 	 * hit the lockdep limitation on recursive locks, or simply discard
2281 	 * these locks.
2282 	 *
2283 	 * AFAICT there is no possible deadlock scenario between the
2284 	 * flush_work() and complete() primitives (except for single-threaded
2285 	 * workqueues), so hiding them isn't a problem.
2286 	 */
2287 	lockdep_invariant_state(true);
2288 	trace_workqueue_execute_start(work);
2289 	worker->current_func(work);
2290 	/*
2291 	 * While we must be careful to not use "work" after this, the trace
2292 	 * point will only record its address.
2293 	 */
2294 	trace_workqueue_execute_end(work, worker->current_func);
2295 	lock_map_release(&lockdep_map);
2296 	lock_map_release(&pwq->wq->lockdep_map);
2297 
2298 	if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2299 		pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2300 		       "     last function: %ps\n",
2301 		       current->comm, preempt_count(), task_pid_nr(current),
2302 		       worker->current_func);
2303 		debug_show_held_locks(current);
2304 		dump_stack();
2305 	}
2306 
2307 	/*
2308 	 * The following prevents a kworker from hogging CPU on !PREEMPTION
2309 	 * kernels, where a requeueing work item waiting for something to
2310 	 * happen could deadlock with stop_machine as such work item could
2311 	 * indefinitely requeue itself while all other CPUs are trapped in
2312 	 * stop_machine. At the same time, report a quiescent RCU state so
2313 	 * the same condition doesn't freeze RCU.
2314 	 */
2315 	cond_resched();
2316 
2317 	raw_spin_lock_irq(&pool->lock);
2318 
2319 	/* clear cpu intensive status */
2320 	if (unlikely(cpu_intensive))
2321 		worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2322 
2323 	/* tag the worker for identification in schedule() */
2324 	worker->last_func = worker->current_func;
2325 
2326 	/* we're done with it, release */
2327 	hash_del(&worker->hentry);
2328 	worker->current_work = NULL;
2329 	worker->current_func = NULL;
2330 	worker->current_pwq = NULL;
2331 	worker->current_color = INT_MAX;
2332 	pwq_dec_nr_in_flight(pwq, work_data);
2333 }
2334 
2335 /**
2336  * process_scheduled_works - process scheduled works
2337  * @worker: self
2338  *
2339  * Process all scheduled works.  Please note that the scheduled list
2340  * may change while processing a work, so this function repeatedly
2341  * fetches a work from the top and executes it.
2342  *
2343  * CONTEXT:
2344  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2345  * multiple times.
2346  */
process_scheduled_works(struct worker * worker)2347 static void process_scheduled_works(struct worker *worker)
2348 {
2349 	while (!list_empty(&worker->scheduled)) {
2350 		struct work_struct *work = list_first_entry(&worker->scheduled,
2351 						struct work_struct, entry);
2352 		process_one_work(worker, work);
2353 	}
2354 }
2355 
set_pf_worker(bool val)2356 static void set_pf_worker(bool val)
2357 {
2358 	mutex_lock(&wq_pool_attach_mutex);
2359 	if (val)
2360 		current->flags |= PF_WQ_WORKER;
2361 	else
2362 		current->flags &= ~PF_WQ_WORKER;
2363 	mutex_unlock(&wq_pool_attach_mutex);
2364 }
2365 
2366 /**
2367  * worker_thread - the worker thread function
2368  * @__worker: self
2369  *
2370  * The worker thread function.  All workers belong to a worker_pool -
2371  * either a per-cpu one or dynamic unbound one.  These workers process all
2372  * work items regardless of their specific target workqueue.  The only
2373  * exception is work items which belong to workqueues with a rescuer which
2374  * will be explained in rescuer_thread().
2375  *
2376  * Return: 0
2377  */
worker_thread(void * __worker)2378 static int worker_thread(void *__worker)
2379 {
2380 	struct worker *worker = __worker;
2381 	struct worker_pool *pool = worker->pool;
2382 
2383 	/* tell the scheduler that this is a workqueue worker */
2384 	set_pf_worker(true);
2385 woke_up:
2386 	raw_spin_lock_irq(&pool->lock);
2387 
2388 	/* am I supposed to die? */
2389 	if (unlikely(worker->flags & WORKER_DIE)) {
2390 		raw_spin_unlock_irq(&pool->lock);
2391 		WARN_ON_ONCE(!list_empty(&worker->entry));
2392 		set_pf_worker(false);
2393 
2394 		set_task_comm(worker->task, "kworker/dying");
2395 		ida_free(&pool->worker_ida, worker->id);
2396 		worker_detach_from_pool(worker);
2397 		kfree(worker);
2398 		return 0;
2399 	}
2400 
2401 	worker_leave_idle(worker);
2402 recheck:
2403 	/* no more worker necessary? */
2404 	if (!need_more_worker(pool))
2405 		goto sleep;
2406 
2407 	/* do we need to manage? */
2408 	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2409 		goto recheck;
2410 
2411 	/*
2412 	 * ->scheduled list can only be filled while a worker is
2413 	 * preparing to process a work or actually processing it.
2414 	 * Make sure nobody diddled with it while I was sleeping.
2415 	 */
2416 	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2417 
2418 	/*
2419 	 * Finish PREP stage.  We're guaranteed to have at least one idle
2420 	 * worker or that someone else has already assumed the manager
2421 	 * role.  This is where @worker starts participating in concurrency
2422 	 * management if applicable and concurrency management is restored
2423 	 * after being rebound.  See rebind_workers() for details.
2424 	 */
2425 	worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2426 
2427 	do {
2428 		struct work_struct *work =
2429 			list_first_entry(&pool->worklist,
2430 					 struct work_struct, entry);
2431 
2432 		pool->watchdog_ts = jiffies;
2433 
2434 		if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2435 			/* optimization path, not strictly necessary */
2436 			process_one_work(worker, work);
2437 			if (unlikely(!list_empty(&worker->scheduled)))
2438 				process_scheduled_works(worker);
2439 		} else {
2440 			move_linked_works(work, &worker->scheduled, NULL);
2441 			process_scheduled_works(worker);
2442 		}
2443 	} while (keep_working(pool));
2444 
2445 	worker_set_flags(worker, WORKER_PREP);
2446 sleep:
2447 	/*
2448 	 * pool->lock is held and there's no work to process and no need to
2449 	 * manage, sleep.  Workers are woken up only while holding
2450 	 * pool->lock or from local cpu, so setting the current state
2451 	 * before releasing pool->lock is enough to prevent losing any
2452 	 * event.
2453 	 */
2454 	worker_enter_idle(worker);
2455 	__set_current_state(TASK_IDLE);
2456 	raw_spin_unlock_irq(&pool->lock);
2457 	schedule();
2458 	goto woke_up;
2459 }
2460 
2461 /**
2462  * rescuer_thread - the rescuer thread function
2463  * @__rescuer: self
2464  *
2465  * Workqueue rescuer thread function.  There's one rescuer for each
2466  * workqueue which has WQ_MEM_RECLAIM set.
2467  *
2468  * Regular work processing on a pool may block trying to create a new
2469  * worker which uses GFP_KERNEL allocation which has slight chance of
2470  * developing into deadlock if some works currently on the same queue
2471  * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2472  * the problem rescuer solves.
2473  *
2474  * When such condition is possible, the pool summons rescuers of all
2475  * workqueues which have works queued on the pool and let them process
2476  * those works so that forward progress can be guaranteed.
2477  *
2478  * This should happen rarely.
2479  *
2480  * Return: 0
2481  */
rescuer_thread(void * __rescuer)2482 static int rescuer_thread(void *__rescuer)
2483 {
2484 	struct worker *rescuer = __rescuer;
2485 	struct workqueue_struct *wq = rescuer->rescue_wq;
2486 	struct list_head *scheduled = &rescuer->scheduled;
2487 	bool should_stop;
2488 
2489 	set_user_nice(current, RESCUER_NICE_LEVEL);
2490 
2491 	/*
2492 	 * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
2493 	 * doesn't participate in concurrency management.
2494 	 */
2495 	set_pf_worker(true);
2496 repeat:
2497 	set_current_state(TASK_IDLE);
2498 
2499 	/*
2500 	 * By the time the rescuer is requested to stop, the workqueue
2501 	 * shouldn't have any work pending, but @wq->maydays may still have
2502 	 * pwq(s) queued.  This can happen by non-rescuer workers consuming
2503 	 * all the work items before the rescuer got to them.  Go through
2504 	 * @wq->maydays processing before acting on should_stop so that the
2505 	 * list is always empty on exit.
2506 	 */
2507 	should_stop = kthread_should_stop();
2508 
2509 	/* see whether any pwq is asking for help */
2510 	raw_spin_lock_irq(&wq_mayday_lock);
2511 
2512 	while (!list_empty(&wq->maydays)) {
2513 		struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2514 					struct pool_workqueue, mayday_node);
2515 		struct worker_pool *pool = pwq->pool;
2516 		struct work_struct *work, *n;
2517 		bool first = true;
2518 
2519 		__set_current_state(TASK_RUNNING);
2520 		list_del_init(&pwq->mayday_node);
2521 
2522 		raw_spin_unlock_irq(&wq_mayday_lock);
2523 
2524 		worker_attach_to_pool(rescuer, pool);
2525 
2526 		raw_spin_lock_irq(&pool->lock);
2527 
2528 		/*
2529 		 * Slurp in all works issued via this workqueue and
2530 		 * process'em.
2531 		 */
2532 		WARN_ON_ONCE(!list_empty(scheduled));
2533 		list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2534 			if (get_work_pwq(work) == pwq) {
2535 				if (first)
2536 					pool->watchdog_ts = jiffies;
2537 				move_linked_works(work, scheduled, &n);
2538 			}
2539 			first = false;
2540 		}
2541 
2542 		if (!list_empty(scheduled)) {
2543 			process_scheduled_works(rescuer);
2544 
2545 			/*
2546 			 * The above execution of rescued work items could
2547 			 * have created more to rescue through
2548 			 * pwq_activate_first_inactive() or chained
2549 			 * queueing.  Let's put @pwq back on mayday list so
2550 			 * that such back-to-back work items, which may be
2551 			 * being used to relieve memory pressure, don't
2552 			 * incur MAYDAY_INTERVAL delay inbetween.
2553 			 */
2554 			if (pwq->nr_active && need_to_create_worker(pool)) {
2555 				raw_spin_lock(&wq_mayday_lock);
2556 				/*
2557 				 * Queue iff we aren't racing destruction
2558 				 * and somebody else hasn't queued it already.
2559 				 */
2560 				if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2561 					get_pwq(pwq);
2562 					list_add_tail(&pwq->mayday_node, &wq->maydays);
2563 				}
2564 				raw_spin_unlock(&wq_mayday_lock);
2565 			}
2566 		}
2567 
2568 		/*
2569 		 * Put the reference grabbed by send_mayday().  @pool won't
2570 		 * go away while we're still attached to it.
2571 		 */
2572 		put_pwq(pwq);
2573 
2574 		/*
2575 		 * Leave this pool.  If need_more_worker() is %true, notify a
2576 		 * regular worker; otherwise, we end up with 0 concurrency
2577 		 * and stalling the execution.
2578 		 */
2579 		if (need_more_worker(pool))
2580 			wake_up_worker(pool);
2581 
2582 		raw_spin_unlock_irq(&pool->lock);
2583 
2584 		worker_detach_from_pool(rescuer);
2585 
2586 		raw_spin_lock_irq(&wq_mayday_lock);
2587 	}
2588 
2589 	raw_spin_unlock_irq(&wq_mayday_lock);
2590 
2591 	if (should_stop) {
2592 		__set_current_state(TASK_RUNNING);
2593 		set_pf_worker(false);
2594 		return 0;
2595 	}
2596 
2597 	/* rescuers should never participate in concurrency management */
2598 	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2599 	schedule();
2600 	goto repeat;
2601 }
2602 
2603 /**
2604  * check_flush_dependency - check for flush dependency sanity
2605  * @target_wq: workqueue being flushed
2606  * @target_work: work item being flushed (NULL for workqueue flushes)
2607  *
2608  * %current is trying to flush the whole @target_wq or @target_work on it.
2609  * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2610  * reclaiming memory or running on a workqueue which doesn't have
2611  * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2612  * a deadlock.
2613  */
check_flush_dependency(struct workqueue_struct * target_wq,struct work_struct * target_work)2614 static void check_flush_dependency(struct workqueue_struct *target_wq,
2615 				   struct work_struct *target_work)
2616 {
2617 	work_func_t target_func = target_work ? target_work->func : NULL;
2618 	struct worker *worker;
2619 
2620 	if (target_wq->flags & WQ_MEM_RECLAIM)
2621 		return;
2622 
2623 	worker = current_wq_worker();
2624 
2625 	WARN_ONCE(current->flags & PF_MEMALLOC,
2626 		  "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2627 		  current->pid, current->comm, target_wq->name, target_func);
2628 	WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2629 			      (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2630 		  "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2631 		  worker->current_pwq->wq->name, worker->current_func,
2632 		  target_wq->name, target_func);
2633 }
2634 
2635 struct wq_barrier {
2636 	struct work_struct	work;
2637 	struct completion	done;
2638 	struct task_struct	*task;	/* purely informational */
2639 };
2640 
wq_barrier_func(struct work_struct * work)2641 static void wq_barrier_func(struct work_struct *work)
2642 {
2643 	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2644 	complete(&barr->done);
2645 }
2646 
2647 /**
2648  * insert_wq_barrier - insert a barrier work
2649  * @pwq: pwq to insert barrier into
2650  * @barr: wq_barrier to insert
2651  * @target: target work to attach @barr to
2652  * @worker: worker currently executing @target, NULL if @target is not executing
2653  *
2654  * @barr is linked to @target such that @barr is completed only after
2655  * @target finishes execution.  Please note that the ordering
2656  * guarantee is observed only with respect to @target and on the local
2657  * cpu.
2658  *
2659  * Currently, a queued barrier can't be canceled.  This is because
2660  * try_to_grab_pending() can't determine whether the work to be
2661  * grabbed is at the head of the queue and thus can't clear LINKED
2662  * flag of the previous work while there must be a valid next work
2663  * after a work with LINKED flag set.
2664  *
2665  * Note that when @worker is non-NULL, @target may be modified
2666  * underneath us, so we can't reliably determine pwq from @target.
2667  *
2668  * CONTEXT:
2669  * raw_spin_lock_irq(pool->lock).
2670  */
insert_wq_barrier(struct pool_workqueue * pwq,struct wq_barrier * barr,struct work_struct * target,struct worker * worker)2671 static void insert_wq_barrier(struct pool_workqueue *pwq,
2672 			      struct wq_barrier *barr,
2673 			      struct work_struct *target, struct worker *worker)
2674 {
2675 	unsigned int work_flags = 0;
2676 	unsigned int work_color;
2677 	struct list_head *head;
2678 
2679 	/*
2680 	 * debugobject calls are safe here even with pool->lock locked
2681 	 * as we know for sure that this will not trigger any of the
2682 	 * checks and call back into the fixup functions where we
2683 	 * might deadlock.
2684 	 */
2685 	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2686 	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2687 
2688 	init_completion_map(&barr->done, &target->lockdep_map);
2689 
2690 	barr->task = current;
2691 
2692 	/* The barrier work item does not participate in pwq->nr_active. */
2693 	work_flags |= WORK_STRUCT_INACTIVE;
2694 
2695 	/*
2696 	 * If @target is currently being executed, schedule the
2697 	 * barrier to the worker; otherwise, put it after @target.
2698 	 */
2699 	if (worker) {
2700 		head = worker->scheduled.next;
2701 		work_color = worker->current_color;
2702 	} else {
2703 		unsigned long *bits = work_data_bits(target);
2704 
2705 		head = target->entry.next;
2706 		/* there can already be other linked works, inherit and set */
2707 		work_flags |= *bits & WORK_STRUCT_LINKED;
2708 		work_color = get_work_color(*bits);
2709 		__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2710 	}
2711 
2712 	pwq->nr_in_flight[work_color]++;
2713 	work_flags |= work_color_to_flags(work_color);
2714 
2715 	debug_work_activate(&barr->work);
2716 	insert_work(pwq, &barr->work, head, work_flags);
2717 }
2718 
2719 /**
2720  * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2721  * @wq: workqueue being flushed
2722  * @flush_color: new flush color, < 0 for no-op
2723  * @work_color: new work color, < 0 for no-op
2724  *
2725  * Prepare pwqs for workqueue flushing.
2726  *
2727  * If @flush_color is non-negative, flush_color on all pwqs should be
2728  * -1.  If no pwq has in-flight commands at the specified color, all
2729  * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
2730  * has in flight commands, its pwq->flush_color is set to
2731  * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2732  * wakeup logic is armed and %true is returned.
2733  *
2734  * The caller should have initialized @wq->first_flusher prior to
2735  * calling this function with non-negative @flush_color.  If
2736  * @flush_color is negative, no flush color update is done and %false
2737  * is returned.
2738  *
2739  * If @work_color is non-negative, all pwqs should have the same
2740  * work_color which is previous to @work_color and all will be
2741  * advanced to @work_color.
2742  *
2743  * CONTEXT:
2744  * mutex_lock(wq->mutex).
2745  *
2746  * Return:
2747  * %true if @flush_color >= 0 and there's something to flush.  %false
2748  * otherwise.
2749  */
flush_workqueue_prep_pwqs(struct workqueue_struct * wq,int flush_color,int work_color)2750 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2751 				      int flush_color, int work_color)
2752 {
2753 	bool wait = false;
2754 	struct pool_workqueue *pwq;
2755 
2756 	if (flush_color >= 0) {
2757 		WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2758 		atomic_set(&wq->nr_pwqs_to_flush, 1);
2759 	}
2760 
2761 	for_each_pwq(pwq, wq) {
2762 		struct worker_pool *pool = pwq->pool;
2763 
2764 		raw_spin_lock_irq(&pool->lock);
2765 
2766 		if (flush_color >= 0) {
2767 			WARN_ON_ONCE(pwq->flush_color != -1);
2768 
2769 			if (pwq->nr_in_flight[flush_color]) {
2770 				pwq->flush_color = flush_color;
2771 				atomic_inc(&wq->nr_pwqs_to_flush);
2772 				wait = true;
2773 			}
2774 		}
2775 
2776 		if (work_color >= 0) {
2777 			WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2778 			pwq->work_color = work_color;
2779 		}
2780 
2781 		raw_spin_unlock_irq(&pool->lock);
2782 	}
2783 
2784 	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2785 		complete(&wq->first_flusher->done);
2786 
2787 	return wait;
2788 }
2789 
2790 /**
2791  * __flush_workqueue - ensure that any scheduled work has run to completion.
2792  * @wq: workqueue to flush
2793  *
2794  * This function sleeps until all work items which were queued on entry
2795  * have finished execution, but it is not livelocked by new incoming ones.
2796  */
__flush_workqueue(struct workqueue_struct * wq)2797 void __flush_workqueue(struct workqueue_struct *wq)
2798 {
2799 	struct wq_flusher this_flusher = {
2800 		.list = LIST_HEAD_INIT(this_flusher.list),
2801 		.flush_color = -1,
2802 		.done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2803 	};
2804 	int next_color;
2805 
2806 	if (WARN_ON(!wq_online))
2807 		return;
2808 
2809 	lock_map_acquire(&wq->lockdep_map);
2810 	lock_map_release(&wq->lockdep_map);
2811 
2812 	mutex_lock(&wq->mutex);
2813 
2814 	/*
2815 	 * Start-to-wait phase
2816 	 */
2817 	next_color = work_next_color(wq->work_color);
2818 
2819 	if (next_color != wq->flush_color) {
2820 		/*
2821 		 * Color space is not full.  The current work_color
2822 		 * becomes our flush_color and work_color is advanced
2823 		 * by one.
2824 		 */
2825 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2826 		this_flusher.flush_color = wq->work_color;
2827 		wq->work_color = next_color;
2828 
2829 		if (!wq->first_flusher) {
2830 			/* no flush in progress, become the first flusher */
2831 			WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2832 
2833 			wq->first_flusher = &this_flusher;
2834 
2835 			if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2836 						       wq->work_color)) {
2837 				/* nothing to flush, done */
2838 				wq->flush_color = next_color;
2839 				wq->first_flusher = NULL;
2840 				goto out_unlock;
2841 			}
2842 		} else {
2843 			/* wait in queue */
2844 			WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2845 			list_add_tail(&this_flusher.list, &wq->flusher_queue);
2846 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2847 		}
2848 	} else {
2849 		/*
2850 		 * Oops, color space is full, wait on overflow queue.
2851 		 * The next flush completion will assign us
2852 		 * flush_color and transfer to flusher_queue.
2853 		 */
2854 		list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2855 	}
2856 
2857 	check_flush_dependency(wq, NULL);
2858 
2859 	mutex_unlock(&wq->mutex);
2860 
2861 	wait_for_completion(&this_flusher.done);
2862 
2863 	/*
2864 	 * Wake-up-and-cascade phase
2865 	 *
2866 	 * First flushers are responsible for cascading flushes and
2867 	 * handling overflow.  Non-first flushers can simply return.
2868 	 */
2869 	if (READ_ONCE(wq->first_flusher) != &this_flusher)
2870 		return;
2871 
2872 	mutex_lock(&wq->mutex);
2873 
2874 	/* we might have raced, check again with mutex held */
2875 	if (wq->first_flusher != &this_flusher)
2876 		goto out_unlock;
2877 
2878 	WRITE_ONCE(wq->first_flusher, NULL);
2879 
2880 	WARN_ON_ONCE(!list_empty(&this_flusher.list));
2881 	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2882 
2883 	while (true) {
2884 		struct wq_flusher *next, *tmp;
2885 
2886 		/* complete all the flushers sharing the current flush color */
2887 		list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2888 			if (next->flush_color != wq->flush_color)
2889 				break;
2890 			list_del_init(&next->list);
2891 			complete(&next->done);
2892 		}
2893 
2894 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2895 			     wq->flush_color != work_next_color(wq->work_color));
2896 
2897 		/* this flush_color is finished, advance by one */
2898 		wq->flush_color = work_next_color(wq->flush_color);
2899 
2900 		/* one color has been freed, handle overflow queue */
2901 		if (!list_empty(&wq->flusher_overflow)) {
2902 			/*
2903 			 * Assign the same color to all overflowed
2904 			 * flushers, advance work_color and append to
2905 			 * flusher_queue.  This is the start-to-wait
2906 			 * phase for these overflowed flushers.
2907 			 */
2908 			list_for_each_entry(tmp, &wq->flusher_overflow, list)
2909 				tmp->flush_color = wq->work_color;
2910 
2911 			wq->work_color = work_next_color(wq->work_color);
2912 
2913 			list_splice_tail_init(&wq->flusher_overflow,
2914 					      &wq->flusher_queue);
2915 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2916 		}
2917 
2918 		if (list_empty(&wq->flusher_queue)) {
2919 			WARN_ON_ONCE(wq->flush_color != wq->work_color);
2920 			break;
2921 		}
2922 
2923 		/*
2924 		 * Need to flush more colors.  Make the next flusher
2925 		 * the new first flusher and arm pwqs.
2926 		 */
2927 		WARN_ON_ONCE(wq->flush_color == wq->work_color);
2928 		WARN_ON_ONCE(wq->flush_color != next->flush_color);
2929 
2930 		list_del_init(&next->list);
2931 		wq->first_flusher = next;
2932 
2933 		if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2934 			break;
2935 
2936 		/*
2937 		 * Meh... this color is already done, clear first
2938 		 * flusher and repeat cascading.
2939 		 */
2940 		wq->first_flusher = NULL;
2941 	}
2942 
2943 out_unlock:
2944 	mutex_unlock(&wq->mutex);
2945 }
2946 EXPORT_SYMBOL(__flush_workqueue);
2947 
2948 /**
2949  * drain_workqueue - drain a workqueue
2950  * @wq: workqueue to drain
2951  *
2952  * Wait until the workqueue becomes empty.  While draining is in progress,
2953  * only chain queueing is allowed.  IOW, only currently pending or running
2954  * work items on @wq can queue further work items on it.  @wq is flushed
2955  * repeatedly until it becomes empty.  The number of flushing is determined
2956  * by the depth of chaining and should be relatively short.  Whine if it
2957  * takes too long.
2958  */
drain_workqueue(struct workqueue_struct * wq)2959 void drain_workqueue(struct workqueue_struct *wq)
2960 {
2961 	unsigned int flush_cnt = 0;
2962 	struct pool_workqueue *pwq;
2963 
2964 	/*
2965 	 * __queue_work() needs to test whether there are drainers, is much
2966 	 * hotter than drain_workqueue() and already looks at @wq->flags.
2967 	 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2968 	 */
2969 	mutex_lock(&wq->mutex);
2970 	if (!wq->nr_drainers++)
2971 		wq->flags |= __WQ_DRAINING;
2972 	mutex_unlock(&wq->mutex);
2973 reflush:
2974 	__flush_workqueue(wq);
2975 
2976 	mutex_lock(&wq->mutex);
2977 
2978 	for_each_pwq(pwq, wq) {
2979 		bool drained;
2980 
2981 		raw_spin_lock_irq(&pwq->pool->lock);
2982 		drained = !pwq->nr_active && list_empty(&pwq->inactive_works);
2983 		raw_spin_unlock_irq(&pwq->pool->lock);
2984 
2985 		if (drained)
2986 			continue;
2987 
2988 		if (++flush_cnt == 10 ||
2989 		    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2990 			pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
2991 				wq->name, __func__, flush_cnt);
2992 
2993 		mutex_unlock(&wq->mutex);
2994 		goto reflush;
2995 	}
2996 
2997 	if (!--wq->nr_drainers)
2998 		wq->flags &= ~__WQ_DRAINING;
2999 	mutex_unlock(&wq->mutex);
3000 }
3001 EXPORT_SYMBOL_GPL(drain_workqueue);
3002 
start_flush_work(struct work_struct * work,struct wq_barrier * barr,bool from_cancel)3003 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
3004 			     bool from_cancel)
3005 {
3006 	struct worker *worker = NULL;
3007 	struct worker_pool *pool;
3008 	struct pool_workqueue *pwq;
3009 
3010 	might_sleep();
3011 
3012 	rcu_read_lock();
3013 	pool = get_work_pool(work);
3014 	if (!pool) {
3015 		rcu_read_unlock();
3016 		return false;
3017 	}
3018 
3019 	raw_spin_lock_irq(&pool->lock);
3020 	/* see the comment in try_to_grab_pending() with the same code */
3021 	pwq = get_work_pwq(work);
3022 	if (pwq) {
3023 		if (unlikely(pwq->pool != pool))
3024 			goto already_gone;
3025 	} else {
3026 		worker = find_worker_executing_work(pool, work);
3027 		if (!worker)
3028 			goto already_gone;
3029 		pwq = worker->current_pwq;
3030 	}
3031 
3032 	check_flush_dependency(pwq->wq, work);
3033 
3034 	insert_wq_barrier(pwq, barr, work, worker);
3035 	raw_spin_unlock_irq(&pool->lock);
3036 
3037 	/*
3038 	 * Force a lock recursion deadlock when using flush_work() inside a
3039 	 * single-threaded or rescuer equipped workqueue.
3040 	 *
3041 	 * For single threaded workqueues the deadlock happens when the work
3042 	 * is after the work issuing the flush_work(). For rescuer equipped
3043 	 * workqueues the deadlock happens when the rescuer stalls, blocking
3044 	 * forward progress.
3045 	 */
3046 	if (!from_cancel &&
3047 	    (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3048 		lock_map_acquire(&pwq->wq->lockdep_map);
3049 		lock_map_release(&pwq->wq->lockdep_map);
3050 	}
3051 	rcu_read_unlock();
3052 	return true;
3053 already_gone:
3054 	raw_spin_unlock_irq(&pool->lock);
3055 	rcu_read_unlock();
3056 	return false;
3057 }
3058 
__flush_work(struct work_struct * work,bool from_cancel)3059 static bool __flush_work(struct work_struct *work, bool from_cancel)
3060 {
3061 	struct wq_barrier barr;
3062 
3063 	if (WARN_ON(!wq_online))
3064 		return false;
3065 
3066 	if (WARN_ON(!work->func))
3067 		return false;
3068 
3069 	if (!from_cancel) {
3070 		lock_map_acquire(&work->lockdep_map);
3071 		lock_map_release(&work->lockdep_map);
3072 	}
3073 
3074 	if (start_flush_work(work, &barr, from_cancel)) {
3075 		wait_for_completion(&barr.done);
3076 		destroy_work_on_stack(&barr.work);
3077 		return true;
3078 	} else {
3079 		return false;
3080 	}
3081 }
3082 
3083 /**
3084  * flush_work - wait for a work to finish executing the last queueing instance
3085  * @work: the work to flush
3086  *
3087  * Wait until @work has finished execution.  @work is guaranteed to be idle
3088  * on return if it hasn't been requeued since flush started.
3089  *
3090  * Return:
3091  * %true if flush_work() waited for the work to finish execution,
3092  * %false if it was already idle.
3093  */
flush_work(struct work_struct * work)3094 bool flush_work(struct work_struct *work)
3095 {
3096 	return __flush_work(work, false);
3097 }
3098 EXPORT_SYMBOL_GPL(flush_work);
3099 
3100 struct cwt_wait {
3101 	wait_queue_entry_t		wait;
3102 	struct work_struct	*work;
3103 };
3104 
cwt_wakefn(wait_queue_entry_t * wait,unsigned mode,int sync,void * key)3105 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3106 {
3107 	struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3108 
3109 	if (cwait->work != key)
3110 		return 0;
3111 	return autoremove_wake_function(wait, mode, sync, key);
3112 }
3113 
__cancel_work_timer(struct work_struct * work,bool is_dwork)3114 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3115 {
3116 	static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3117 	unsigned long flags;
3118 	int ret;
3119 
3120 	do {
3121 		ret = try_to_grab_pending(work, is_dwork, &flags);
3122 		/*
3123 		 * If someone else is already canceling, wait for it to
3124 		 * finish.  flush_work() doesn't work for PREEMPT_NONE
3125 		 * because we may get scheduled between @work's completion
3126 		 * and the other canceling task resuming and clearing
3127 		 * CANCELING - flush_work() will return false immediately
3128 		 * as @work is no longer busy, try_to_grab_pending() will
3129 		 * return -ENOENT as @work is still being canceled and the
3130 		 * other canceling task won't be able to clear CANCELING as
3131 		 * we're hogging the CPU.
3132 		 *
3133 		 * Let's wait for completion using a waitqueue.  As this
3134 		 * may lead to the thundering herd problem, use a custom
3135 		 * wake function which matches @work along with exclusive
3136 		 * wait and wakeup.
3137 		 */
3138 		if (unlikely(ret == -ENOENT)) {
3139 			struct cwt_wait cwait;
3140 
3141 			init_wait(&cwait.wait);
3142 			cwait.wait.func = cwt_wakefn;
3143 			cwait.work = work;
3144 
3145 			prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3146 						  TASK_UNINTERRUPTIBLE);
3147 			if (work_is_canceling(work))
3148 				schedule();
3149 			finish_wait(&cancel_waitq, &cwait.wait);
3150 		}
3151 	} while (unlikely(ret < 0));
3152 
3153 	/* tell other tasks trying to grab @work to back off */
3154 	mark_work_canceling(work);
3155 	local_irq_restore(flags);
3156 
3157 	/*
3158 	 * This allows canceling during early boot.  We know that @work
3159 	 * isn't executing.
3160 	 */
3161 	if (wq_online)
3162 		__flush_work(work, true);
3163 
3164 	clear_work_data(work);
3165 
3166 	/*
3167 	 * Paired with prepare_to_wait() above so that either
3168 	 * waitqueue_active() is visible here or !work_is_canceling() is
3169 	 * visible there.
3170 	 */
3171 	smp_mb();
3172 	if (waitqueue_active(&cancel_waitq))
3173 		__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3174 
3175 	return ret;
3176 }
3177 
3178 /**
3179  * cancel_work_sync - cancel a work and wait for it to finish
3180  * @work: the work to cancel
3181  *
3182  * Cancel @work and wait for its execution to finish.  This function
3183  * can be used even if the work re-queues itself or migrates to
3184  * another workqueue.  On return from this function, @work is
3185  * guaranteed to be not pending or executing on any CPU.
3186  *
3187  * cancel_work_sync(&delayed_work->work) must not be used for
3188  * delayed_work's.  Use cancel_delayed_work_sync() instead.
3189  *
3190  * The caller must ensure that the workqueue on which @work was last
3191  * queued can't be destroyed before this function returns.
3192  *
3193  * Return:
3194  * %true if @work was pending, %false otherwise.
3195  */
cancel_work_sync(struct work_struct * work)3196 bool cancel_work_sync(struct work_struct *work)
3197 {
3198 	return __cancel_work_timer(work, false);
3199 }
3200 EXPORT_SYMBOL_GPL(cancel_work_sync);
3201 
3202 /**
3203  * flush_delayed_work - wait for a dwork to finish executing the last queueing
3204  * @dwork: the delayed work to flush
3205  *
3206  * Delayed timer is cancelled and the pending work is queued for
3207  * immediate execution.  Like flush_work(), this function only
3208  * considers the last queueing instance of @dwork.
3209  *
3210  * Return:
3211  * %true if flush_work() waited for the work to finish execution,
3212  * %false if it was already idle.
3213  */
flush_delayed_work(struct delayed_work * dwork)3214 bool flush_delayed_work(struct delayed_work *dwork)
3215 {
3216 	local_irq_disable();
3217 	if (del_timer_sync(&dwork->timer))
3218 		__queue_work(dwork->cpu, dwork->wq, &dwork->work);
3219 	local_irq_enable();
3220 	return flush_work(&dwork->work);
3221 }
3222 EXPORT_SYMBOL(flush_delayed_work);
3223 
3224 /**
3225  * flush_rcu_work - wait for a rwork to finish executing the last queueing
3226  * @rwork: the rcu work to flush
3227  *
3228  * Return:
3229  * %true if flush_rcu_work() waited for the work to finish execution,
3230  * %false if it was already idle.
3231  */
flush_rcu_work(struct rcu_work * rwork)3232 bool flush_rcu_work(struct rcu_work *rwork)
3233 {
3234 	if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3235 		rcu_barrier();
3236 		flush_work(&rwork->work);
3237 		return true;
3238 	} else {
3239 		return flush_work(&rwork->work);
3240 	}
3241 }
3242 EXPORT_SYMBOL(flush_rcu_work);
3243 
__cancel_work(struct work_struct * work,bool is_dwork)3244 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3245 {
3246 	unsigned long flags;
3247 	int ret;
3248 
3249 	do {
3250 		ret = try_to_grab_pending(work, is_dwork, &flags);
3251 	} while (unlikely(ret == -EAGAIN));
3252 
3253 	if (unlikely(ret < 0))
3254 		return false;
3255 
3256 	set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3257 	local_irq_restore(flags);
3258 	return ret;
3259 }
3260 
3261 /**
3262  * cancel_delayed_work - cancel a delayed work
3263  * @dwork: delayed_work to cancel
3264  *
3265  * Kill off a pending delayed_work.
3266  *
3267  * Return: %true if @dwork was pending and canceled; %false if it wasn't
3268  * pending.
3269  *
3270  * Note:
3271  * The work callback function may still be running on return, unless
3272  * it returns %true and the work doesn't re-arm itself.  Explicitly flush or
3273  * use cancel_delayed_work_sync() to wait on it.
3274  *
3275  * This function is safe to call from any context including IRQ handler.
3276  */
cancel_delayed_work(struct delayed_work * dwork)3277 bool cancel_delayed_work(struct delayed_work *dwork)
3278 {
3279 	return __cancel_work(&dwork->work, true);
3280 }
3281 EXPORT_SYMBOL(cancel_delayed_work);
3282 
3283 /**
3284  * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3285  * @dwork: the delayed work cancel
3286  *
3287  * This is cancel_work_sync() for delayed works.
3288  *
3289  * Return:
3290  * %true if @dwork was pending, %false otherwise.
3291  */
cancel_delayed_work_sync(struct delayed_work * dwork)3292 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3293 {
3294 	return __cancel_work_timer(&dwork->work, true);
3295 }
3296 EXPORT_SYMBOL(cancel_delayed_work_sync);
3297 
3298 /**
3299  * schedule_on_each_cpu - execute a function synchronously on each online CPU
3300  * @func: the function to call
3301  *
3302  * schedule_on_each_cpu() executes @func on each online CPU using the
3303  * system workqueue and blocks until all CPUs have completed.
3304  * schedule_on_each_cpu() is very slow.
3305  *
3306  * Return:
3307  * 0 on success, -errno on failure.
3308  */
schedule_on_each_cpu(work_func_t func)3309 int schedule_on_each_cpu(work_func_t func)
3310 {
3311 	int cpu;
3312 	struct work_struct __percpu *works;
3313 
3314 	works = alloc_percpu(struct work_struct);
3315 	if (!works)
3316 		return -ENOMEM;
3317 
3318 	cpus_read_lock();
3319 
3320 	for_each_online_cpu(cpu) {
3321 		struct work_struct *work = per_cpu_ptr(works, cpu);
3322 
3323 		INIT_WORK(work, func);
3324 		schedule_work_on(cpu, work);
3325 	}
3326 
3327 	for_each_online_cpu(cpu)
3328 		flush_work(per_cpu_ptr(works, cpu));
3329 
3330 	cpus_read_unlock();
3331 	free_percpu(works);
3332 	return 0;
3333 }
3334 
3335 /**
3336  * execute_in_process_context - reliably execute the routine with user context
3337  * @fn:		the function to execute
3338  * @ew:		guaranteed storage for the execute work structure (must
3339  *		be available when the work executes)
3340  *
3341  * Executes the function immediately if process context is available,
3342  * otherwise schedules the function for delayed execution.
3343  *
3344  * Return:	0 - function was executed
3345  *		1 - function was scheduled for execution
3346  */
execute_in_process_context(work_func_t fn,struct execute_work * ew)3347 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3348 {
3349 	if (!in_interrupt()) {
3350 		fn(&ew->work);
3351 		return 0;
3352 	}
3353 
3354 	INIT_WORK(&ew->work, fn);
3355 	schedule_work(&ew->work);
3356 
3357 	return 1;
3358 }
3359 EXPORT_SYMBOL_GPL(execute_in_process_context);
3360 
3361 /**
3362  * free_workqueue_attrs - free a workqueue_attrs
3363  * @attrs: workqueue_attrs to free
3364  *
3365  * Undo alloc_workqueue_attrs().
3366  */
free_workqueue_attrs(struct workqueue_attrs * attrs)3367 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3368 {
3369 	if (attrs) {
3370 		free_cpumask_var(attrs->cpumask);
3371 		kfree(attrs);
3372 	}
3373 }
3374 
3375 /**
3376  * alloc_workqueue_attrs - allocate a workqueue_attrs
3377  *
3378  * Allocate a new workqueue_attrs, initialize with default settings and
3379  * return it.
3380  *
3381  * Return: The allocated new workqueue_attr on success. %NULL on failure.
3382  */
alloc_workqueue_attrs(void)3383 struct workqueue_attrs *alloc_workqueue_attrs(void)
3384 {
3385 	struct workqueue_attrs *attrs;
3386 
3387 	attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3388 	if (!attrs)
3389 		goto fail;
3390 	if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3391 		goto fail;
3392 
3393 	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3394 	return attrs;
3395 fail:
3396 	free_workqueue_attrs(attrs);
3397 	return NULL;
3398 }
3399 
copy_workqueue_attrs(struct workqueue_attrs * to,const struct workqueue_attrs * from)3400 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3401 				 const struct workqueue_attrs *from)
3402 {
3403 	to->nice = from->nice;
3404 	cpumask_copy(to->cpumask, from->cpumask);
3405 	/*
3406 	 * Unlike hash and equality test, this function doesn't ignore
3407 	 * ->no_numa as it is used for both pool and wq attrs.  Instead,
3408 	 * get_unbound_pool() explicitly clears ->no_numa after copying.
3409 	 */
3410 	to->no_numa = from->no_numa;
3411 }
3412 
3413 /* hash value of the content of @attr */
wqattrs_hash(const struct workqueue_attrs * attrs)3414 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3415 {
3416 	u32 hash = 0;
3417 
3418 	hash = jhash_1word(attrs->nice, hash);
3419 	hash = jhash(cpumask_bits(attrs->cpumask),
3420 		     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3421 	return hash;
3422 }
3423 
3424 /* content equality test */
wqattrs_equal(const struct workqueue_attrs * a,const struct workqueue_attrs * b)3425 static bool wqattrs_equal(const struct workqueue_attrs *a,
3426 			  const struct workqueue_attrs *b)
3427 {
3428 	if (a->nice != b->nice)
3429 		return false;
3430 	if (!cpumask_equal(a->cpumask, b->cpumask))
3431 		return false;
3432 	return true;
3433 }
3434 
3435 /**
3436  * init_worker_pool - initialize a newly zalloc'd worker_pool
3437  * @pool: worker_pool to initialize
3438  *
3439  * Initialize a newly zalloc'd @pool.  It also allocates @pool->attrs.
3440  *
3441  * Return: 0 on success, -errno on failure.  Even on failure, all fields
3442  * inside @pool proper are initialized and put_unbound_pool() can be called
3443  * on @pool safely to release it.
3444  */
init_worker_pool(struct worker_pool * pool)3445 static int init_worker_pool(struct worker_pool *pool)
3446 {
3447 	raw_spin_lock_init(&pool->lock);
3448 	pool->id = -1;
3449 	pool->cpu = -1;
3450 	pool->node = NUMA_NO_NODE;
3451 	pool->flags |= POOL_DISASSOCIATED;
3452 	pool->watchdog_ts = jiffies;
3453 	INIT_LIST_HEAD(&pool->worklist);
3454 	INIT_LIST_HEAD(&pool->idle_list);
3455 	hash_init(pool->busy_hash);
3456 
3457 	timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3458 
3459 	timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3460 
3461 	INIT_LIST_HEAD(&pool->workers);
3462 
3463 	ida_init(&pool->worker_ida);
3464 	INIT_HLIST_NODE(&pool->hash_node);
3465 	pool->refcnt = 1;
3466 
3467 	/* shouldn't fail above this point */
3468 	pool->attrs = alloc_workqueue_attrs();
3469 	if (!pool->attrs)
3470 		return -ENOMEM;
3471 	return 0;
3472 }
3473 
3474 #ifdef CONFIG_LOCKDEP
wq_init_lockdep(struct workqueue_struct * wq)3475 static void wq_init_lockdep(struct workqueue_struct *wq)
3476 {
3477 	char *lock_name;
3478 
3479 	lockdep_register_key(&wq->key);
3480 	lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3481 	if (!lock_name)
3482 		lock_name = wq->name;
3483 
3484 	wq->lock_name = lock_name;
3485 	lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3486 }
3487 
wq_unregister_lockdep(struct workqueue_struct * wq)3488 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3489 {
3490 	lockdep_unregister_key(&wq->key);
3491 }
3492 
wq_free_lockdep(struct workqueue_struct * wq)3493 static void wq_free_lockdep(struct workqueue_struct *wq)
3494 {
3495 	if (wq->lock_name != wq->name)
3496 		kfree(wq->lock_name);
3497 }
3498 #else
wq_init_lockdep(struct workqueue_struct * wq)3499 static void wq_init_lockdep(struct workqueue_struct *wq)
3500 {
3501 }
3502 
wq_unregister_lockdep(struct workqueue_struct * wq)3503 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3504 {
3505 }
3506 
wq_free_lockdep(struct workqueue_struct * wq)3507 static void wq_free_lockdep(struct workqueue_struct *wq)
3508 {
3509 }
3510 #endif
3511 
rcu_free_wq(struct rcu_head * rcu)3512 static void rcu_free_wq(struct rcu_head *rcu)
3513 {
3514 	struct workqueue_struct *wq =
3515 		container_of(rcu, struct workqueue_struct, rcu);
3516 
3517 	wq_free_lockdep(wq);
3518 
3519 	if (!(wq->flags & WQ_UNBOUND))
3520 		free_percpu(wq->cpu_pwqs);
3521 	else
3522 		free_workqueue_attrs(wq->unbound_attrs);
3523 
3524 	kfree(wq);
3525 }
3526 
rcu_free_pool(struct rcu_head * rcu)3527 static void rcu_free_pool(struct rcu_head *rcu)
3528 {
3529 	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3530 
3531 	ida_destroy(&pool->worker_ida);
3532 	free_workqueue_attrs(pool->attrs);
3533 	kfree(pool);
3534 }
3535 
3536 /* This returns with the lock held on success (pool manager is inactive). */
wq_manager_inactive(struct worker_pool * pool)3537 static bool wq_manager_inactive(struct worker_pool *pool)
3538 {
3539 	raw_spin_lock_irq(&pool->lock);
3540 
3541 	if (pool->flags & POOL_MANAGER_ACTIVE) {
3542 		raw_spin_unlock_irq(&pool->lock);
3543 		return false;
3544 	}
3545 	return true;
3546 }
3547 
3548 /**
3549  * put_unbound_pool - put a worker_pool
3550  * @pool: worker_pool to put
3551  *
3552  * Put @pool.  If its refcnt reaches zero, it gets destroyed in RCU
3553  * safe manner.  get_unbound_pool() calls this function on its failure path
3554  * and this function should be able to release pools which went through,
3555  * successfully or not, init_worker_pool().
3556  *
3557  * Should be called with wq_pool_mutex held.
3558  */
put_unbound_pool(struct worker_pool * pool)3559 static void put_unbound_pool(struct worker_pool *pool)
3560 {
3561 	DECLARE_COMPLETION_ONSTACK(detach_completion);
3562 	struct worker *worker;
3563 
3564 	lockdep_assert_held(&wq_pool_mutex);
3565 
3566 	if (--pool->refcnt)
3567 		return;
3568 
3569 	/* sanity checks */
3570 	if (WARN_ON(!(pool->cpu < 0)) ||
3571 	    WARN_ON(!list_empty(&pool->worklist)))
3572 		return;
3573 
3574 	/* release id and unhash */
3575 	if (pool->id >= 0)
3576 		idr_remove(&worker_pool_idr, pool->id);
3577 	hash_del(&pool->hash_node);
3578 
3579 	/*
3580 	 * Become the manager and destroy all workers.  This prevents
3581 	 * @pool's workers from blocking on attach_mutex.  We're the last
3582 	 * manager and @pool gets freed with the flag set.
3583 	 * Because of how wq_manager_inactive() works, we will hold the
3584 	 * spinlock after a successful wait.
3585 	 */
3586 	rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool),
3587 			   TASK_UNINTERRUPTIBLE);
3588 	pool->flags |= POOL_MANAGER_ACTIVE;
3589 
3590 	while ((worker = first_idle_worker(pool)))
3591 		destroy_worker(worker);
3592 	WARN_ON(pool->nr_workers || pool->nr_idle);
3593 	raw_spin_unlock_irq(&pool->lock);
3594 
3595 	mutex_lock(&wq_pool_attach_mutex);
3596 	if (!list_empty(&pool->workers))
3597 		pool->detach_completion = &detach_completion;
3598 	mutex_unlock(&wq_pool_attach_mutex);
3599 
3600 	if (pool->detach_completion)
3601 		wait_for_completion(pool->detach_completion);
3602 
3603 	/* shut down the timers */
3604 	del_timer_sync(&pool->idle_timer);
3605 	del_timer_sync(&pool->mayday_timer);
3606 
3607 	/* RCU protected to allow dereferences from get_work_pool() */
3608 	call_rcu(&pool->rcu, rcu_free_pool);
3609 }
3610 
3611 /**
3612  * get_unbound_pool - get a worker_pool with the specified attributes
3613  * @attrs: the attributes of the worker_pool to get
3614  *
3615  * Obtain a worker_pool which has the same attributes as @attrs, bump the
3616  * reference count and return it.  If there already is a matching
3617  * worker_pool, it will be used; otherwise, this function attempts to
3618  * create a new one.
3619  *
3620  * Should be called with wq_pool_mutex held.
3621  *
3622  * Return: On success, a worker_pool with the same attributes as @attrs.
3623  * On failure, %NULL.
3624  */
get_unbound_pool(const struct workqueue_attrs * attrs)3625 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3626 {
3627 	u32 hash = wqattrs_hash(attrs);
3628 	struct worker_pool *pool;
3629 	int node;
3630 	int target_node = NUMA_NO_NODE;
3631 
3632 	lockdep_assert_held(&wq_pool_mutex);
3633 
3634 	/* do we already have a matching pool? */
3635 	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3636 		if (wqattrs_equal(pool->attrs, attrs)) {
3637 			pool->refcnt++;
3638 			return pool;
3639 		}
3640 	}
3641 
3642 	/* if cpumask is contained inside a NUMA node, we belong to that node */
3643 	if (wq_numa_enabled) {
3644 		for_each_node(node) {
3645 			if (cpumask_subset(attrs->cpumask,
3646 					   wq_numa_possible_cpumask[node])) {
3647 				target_node = node;
3648 				break;
3649 			}
3650 		}
3651 	}
3652 
3653 	/* nope, create a new one */
3654 	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3655 	if (!pool || init_worker_pool(pool) < 0)
3656 		goto fail;
3657 
3658 	lockdep_set_subclass(&pool->lock, 1);	/* see put_pwq() */
3659 	copy_workqueue_attrs(pool->attrs, attrs);
3660 	pool->node = target_node;
3661 
3662 	/*
3663 	 * no_numa isn't a worker_pool attribute, always clear it.  See
3664 	 * 'struct workqueue_attrs' comments for detail.
3665 	 */
3666 	pool->attrs->no_numa = false;
3667 
3668 	if (worker_pool_assign_id(pool) < 0)
3669 		goto fail;
3670 
3671 	/* create and start the initial worker */
3672 	if (wq_online && !create_worker(pool))
3673 		goto fail;
3674 
3675 	/* install */
3676 	hash_add(unbound_pool_hash, &pool->hash_node, hash);
3677 
3678 	return pool;
3679 fail:
3680 	if (pool)
3681 		put_unbound_pool(pool);
3682 	return NULL;
3683 }
3684 
rcu_free_pwq(struct rcu_head * rcu)3685 static void rcu_free_pwq(struct rcu_head *rcu)
3686 {
3687 	kmem_cache_free(pwq_cache,
3688 			container_of(rcu, struct pool_workqueue, rcu));
3689 }
3690 
3691 /*
3692  * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3693  * and needs to be destroyed.
3694  */
pwq_unbound_release_workfn(struct work_struct * work)3695 static void pwq_unbound_release_workfn(struct work_struct *work)
3696 {
3697 	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3698 						  unbound_release_work);
3699 	struct workqueue_struct *wq = pwq->wq;
3700 	struct worker_pool *pool = pwq->pool;
3701 	bool is_last = false;
3702 
3703 	/*
3704 	 * when @pwq is not linked, it doesn't hold any reference to the
3705 	 * @wq, and @wq is invalid to access.
3706 	 */
3707 	if (!list_empty(&pwq->pwqs_node)) {
3708 		if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3709 			return;
3710 
3711 		mutex_lock(&wq->mutex);
3712 		list_del_rcu(&pwq->pwqs_node);
3713 		is_last = list_empty(&wq->pwqs);
3714 		mutex_unlock(&wq->mutex);
3715 	}
3716 
3717 	mutex_lock(&wq_pool_mutex);
3718 	put_unbound_pool(pool);
3719 	mutex_unlock(&wq_pool_mutex);
3720 
3721 	call_rcu(&pwq->rcu, rcu_free_pwq);
3722 
3723 	/*
3724 	 * If we're the last pwq going away, @wq is already dead and no one
3725 	 * is gonna access it anymore.  Schedule RCU free.
3726 	 */
3727 	if (is_last) {
3728 		wq_unregister_lockdep(wq);
3729 		call_rcu(&wq->rcu, rcu_free_wq);
3730 	}
3731 }
3732 
3733 /**
3734  * pwq_adjust_max_active - update a pwq's max_active to the current setting
3735  * @pwq: target pool_workqueue
3736  *
3737  * If @pwq isn't freezing, set @pwq->max_active to the associated
3738  * workqueue's saved_max_active and activate inactive work items
3739  * accordingly.  If @pwq is freezing, clear @pwq->max_active to zero.
3740  */
pwq_adjust_max_active(struct pool_workqueue * pwq)3741 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3742 {
3743 	struct workqueue_struct *wq = pwq->wq;
3744 	bool freezable = wq->flags & WQ_FREEZABLE;
3745 	unsigned long flags;
3746 
3747 	/* for @wq->saved_max_active */
3748 	lockdep_assert_held(&wq->mutex);
3749 
3750 	/* fast exit for non-freezable wqs */
3751 	if (!freezable && pwq->max_active == wq->saved_max_active)
3752 		return;
3753 
3754 	/* this function can be called during early boot w/ irq disabled */
3755 	raw_spin_lock_irqsave(&pwq->pool->lock, flags);
3756 
3757 	/*
3758 	 * During [un]freezing, the caller is responsible for ensuring that
3759 	 * this function is called at least once after @workqueue_freezing
3760 	 * is updated and visible.
3761 	 */
3762 	if (!freezable || !workqueue_freezing) {
3763 		bool kick = false;
3764 
3765 		pwq->max_active = wq->saved_max_active;
3766 
3767 		while (!list_empty(&pwq->inactive_works) &&
3768 		       pwq->nr_active < pwq->max_active) {
3769 			pwq_activate_first_inactive(pwq);
3770 			kick = true;
3771 		}
3772 
3773 		/*
3774 		 * Need to kick a worker after thawed or an unbound wq's
3775 		 * max_active is bumped. In realtime scenarios, always kicking a
3776 		 * worker will cause interference on the isolated cpu cores, so
3777 		 * let's kick iff work items were activated.
3778 		 */
3779 		if (kick)
3780 			wake_up_worker(pwq->pool);
3781 	} else {
3782 		pwq->max_active = 0;
3783 	}
3784 
3785 	raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
3786 }
3787 
3788 /* initialize newly allocated @pwq which is associated with @wq and @pool */
init_pwq(struct pool_workqueue * pwq,struct workqueue_struct * wq,struct worker_pool * pool)3789 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3790 		     struct worker_pool *pool)
3791 {
3792 	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3793 
3794 	memset(pwq, 0, sizeof(*pwq));
3795 
3796 	pwq->pool = pool;
3797 	pwq->wq = wq;
3798 	pwq->flush_color = -1;
3799 	pwq->refcnt = 1;
3800 	INIT_LIST_HEAD(&pwq->inactive_works);
3801 	INIT_LIST_HEAD(&pwq->pwqs_node);
3802 	INIT_LIST_HEAD(&pwq->mayday_node);
3803 	INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3804 }
3805 
3806 /* sync @pwq with the current state of its associated wq and link it */
link_pwq(struct pool_workqueue * pwq)3807 static void link_pwq(struct pool_workqueue *pwq)
3808 {
3809 	struct workqueue_struct *wq = pwq->wq;
3810 
3811 	lockdep_assert_held(&wq->mutex);
3812 
3813 	/* may be called multiple times, ignore if already linked */
3814 	if (!list_empty(&pwq->pwqs_node))
3815 		return;
3816 
3817 	/* set the matching work_color */
3818 	pwq->work_color = wq->work_color;
3819 
3820 	/* sync max_active to the current setting */
3821 	pwq_adjust_max_active(pwq);
3822 
3823 	/* link in @pwq */
3824 	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3825 }
3826 
3827 /* 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)3828 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3829 					const struct workqueue_attrs *attrs)
3830 {
3831 	struct worker_pool *pool;
3832 	struct pool_workqueue *pwq;
3833 
3834 	lockdep_assert_held(&wq_pool_mutex);
3835 
3836 	pool = get_unbound_pool(attrs);
3837 	if (!pool)
3838 		return NULL;
3839 
3840 	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3841 	if (!pwq) {
3842 		put_unbound_pool(pool);
3843 		return NULL;
3844 	}
3845 
3846 	init_pwq(pwq, wq, pool);
3847 	return pwq;
3848 }
3849 
3850 /**
3851  * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3852  * @attrs: the wq_attrs of the default pwq of the target workqueue
3853  * @node: the target NUMA node
3854  * @cpu_going_down: if >= 0, the CPU to consider as offline
3855  * @cpumask: outarg, the resulting cpumask
3856  *
3857  * Calculate the cpumask a workqueue with @attrs should use on @node.  If
3858  * @cpu_going_down is >= 0, that cpu is considered offline during
3859  * calculation.  The result is stored in @cpumask.
3860  *
3861  * If NUMA affinity is not enabled, @attrs->cpumask is always used.  If
3862  * enabled and @node has online CPUs requested by @attrs, the returned
3863  * cpumask is the intersection of the possible CPUs of @node and
3864  * @attrs->cpumask.
3865  *
3866  * The caller is responsible for ensuring that the cpumask of @node stays
3867  * stable.
3868  *
3869  * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3870  * %false if equal.
3871  */
wq_calc_node_cpumask(const struct workqueue_attrs * attrs,int node,int cpu_going_down,cpumask_t * cpumask)3872 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3873 				 int cpu_going_down, cpumask_t *cpumask)
3874 {
3875 	if (!wq_numa_enabled || attrs->no_numa)
3876 		goto use_dfl;
3877 
3878 	/* does @node have any online CPUs @attrs wants? */
3879 	cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3880 	if (cpu_going_down >= 0)
3881 		cpumask_clear_cpu(cpu_going_down, cpumask);
3882 
3883 	if (cpumask_empty(cpumask))
3884 		goto use_dfl;
3885 
3886 	/* yeap, return possible CPUs in @node that @attrs wants */
3887 	cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3888 
3889 	if (cpumask_empty(cpumask)) {
3890 		pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3891 				"possible intersect\n");
3892 		return false;
3893 	}
3894 
3895 	return !cpumask_equal(cpumask, attrs->cpumask);
3896 
3897 use_dfl:
3898 	cpumask_copy(cpumask, attrs->cpumask);
3899 	return false;
3900 }
3901 
3902 /* 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)3903 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3904 						   int node,
3905 						   struct pool_workqueue *pwq)
3906 {
3907 	struct pool_workqueue *old_pwq;
3908 
3909 	lockdep_assert_held(&wq_pool_mutex);
3910 	lockdep_assert_held(&wq->mutex);
3911 
3912 	/* link_pwq() can handle duplicate calls */
3913 	link_pwq(pwq);
3914 
3915 	old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3916 	rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3917 	return old_pwq;
3918 }
3919 
3920 /* context to store the prepared attrs & pwqs before applying */
3921 struct apply_wqattrs_ctx {
3922 	struct workqueue_struct	*wq;		/* target workqueue */
3923 	struct workqueue_attrs	*attrs;		/* attrs to apply */
3924 	struct list_head	list;		/* queued for batching commit */
3925 	struct pool_workqueue	*dfl_pwq;
3926 	struct pool_workqueue	*pwq_tbl[];
3927 };
3928 
3929 /* free the resources after success or abort */
apply_wqattrs_cleanup(struct apply_wqattrs_ctx * ctx)3930 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3931 {
3932 	if (ctx) {
3933 		int node;
3934 
3935 		for_each_node(node)
3936 			put_pwq_unlocked(ctx->pwq_tbl[node]);
3937 		put_pwq_unlocked(ctx->dfl_pwq);
3938 
3939 		free_workqueue_attrs(ctx->attrs);
3940 
3941 		kfree(ctx);
3942 	}
3943 }
3944 
3945 /* allocate the attrs and pwqs for later installation */
3946 static struct apply_wqattrs_ctx *
apply_wqattrs_prepare(struct workqueue_struct * wq,const struct workqueue_attrs * attrs)3947 apply_wqattrs_prepare(struct workqueue_struct *wq,
3948 		      const struct workqueue_attrs *attrs)
3949 {
3950 	struct apply_wqattrs_ctx *ctx;
3951 	struct workqueue_attrs *new_attrs, *tmp_attrs;
3952 	int node;
3953 
3954 	lockdep_assert_held(&wq_pool_mutex);
3955 
3956 	ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3957 
3958 	new_attrs = alloc_workqueue_attrs();
3959 	tmp_attrs = alloc_workqueue_attrs();
3960 	if (!ctx || !new_attrs || !tmp_attrs)
3961 		goto out_free;
3962 
3963 	/*
3964 	 * Calculate the attrs of the default pwq.
3965 	 * If the user configured cpumask doesn't overlap with the
3966 	 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3967 	 */
3968 	copy_workqueue_attrs(new_attrs, attrs);
3969 	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3970 	if (unlikely(cpumask_empty(new_attrs->cpumask)))
3971 		cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3972 
3973 	/*
3974 	 * We may create multiple pwqs with differing cpumasks.  Make a
3975 	 * copy of @new_attrs which will be modified and used to obtain
3976 	 * pools.
3977 	 */
3978 	copy_workqueue_attrs(tmp_attrs, new_attrs);
3979 
3980 	/*
3981 	 * If something goes wrong during CPU up/down, we'll fall back to
3982 	 * the default pwq covering whole @attrs->cpumask.  Always create
3983 	 * it even if we don't use it immediately.
3984 	 */
3985 	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3986 	if (!ctx->dfl_pwq)
3987 		goto out_free;
3988 
3989 	for_each_node(node) {
3990 		if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3991 			ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3992 			if (!ctx->pwq_tbl[node])
3993 				goto out_free;
3994 		} else {
3995 			ctx->dfl_pwq->refcnt++;
3996 			ctx->pwq_tbl[node] = ctx->dfl_pwq;
3997 		}
3998 	}
3999 
4000 	/* save the user configured attrs and sanitize it. */
4001 	copy_workqueue_attrs(new_attrs, attrs);
4002 	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
4003 	ctx->attrs = new_attrs;
4004 
4005 	ctx->wq = wq;
4006 	free_workqueue_attrs(tmp_attrs);
4007 	return ctx;
4008 
4009 out_free:
4010 	free_workqueue_attrs(tmp_attrs);
4011 	free_workqueue_attrs(new_attrs);
4012 	apply_wqattrs_cleanup(ctx);
4013 	return NULL;
4014 }
4015 
4016 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
apply_wqattrs_commit(struct apply_wqattrs_ctx * ctx)4017 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4018 {
4019 	int node;
4020 
4021 	/* all pwqs have been created successfully, let's install'em */
4022 	mutex_lock(&ctx->wq->mutex);
4023 
4024 	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4025 
4026 	/* save the previous pwq and install the new one */
4027 	for_each_node(node)
4028 		ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
4029 							  ctx->pwq_tbl[node]);
4030 
4031 	/* @dfl_pwq might not have been used, ensure it's linked */
4032 	link_pwq(ctx->dfl_pwq);
4033 	swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
4034 
4035 	mutex_unlock(&ctx->wq->mutex);
4036 }
4037 
apply_wqattrs_lock(void)4038 static void apply_wqattrs_lock(void)
4039 {
4040 	/* CPUs should stay stable across pwq creations and installations */
4041 	cpus_read_lock();
4042 	mutex_lock(&wq_pool_mutex);
4043 }
4044 
apply_wqattrs_unlock(void)4045 static void apply_wqattrs_unlock(void)
4046 {
4047 	mutex_unlock(&wq_pool_mutex);
4048 	cpus_read_unlock();
4049 }
4050 
apply_workqueue_attrs_locked(struct workqueue_struct * wq,const struct workqueue_attrs * attrs)4051 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4052 					const struct workqueue_attrs *attrs)
4053 {
4054 	struct apply_wqattrs_ctx *ctx;
4055 
4056 	/* only unbound workqueues can change attributes */
4057 	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4058 		return -EINVAL;
4059 
4060 	/* creating multiple pwqs breaks ordering guarantee */
4061 	if (!list_empty(&wq->pwqs)) {
4062 		if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4063 			return -EINVAL;
4064 
4065 		wq->flags &= ~__WQ_ORDERED;
4066 	}
4067 
4068 	ctx = apply_wqattrs_prepare(wq, attrs);
4069 	if (!ctx)
4070 		return -ENOMEM;
4071 
4072 	/* the ctx has been prepared successfully, let's commit it */
4073 	apply_wqattrs_commit(ctx);
4074 	apply_wqattrs_cleanup(ctx);
4075 
4076 	return 0;
4077 }
4078 
4079 /**
4080  * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4081  * @wq: the target workqueue
4082  * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4083  *
4084  * Apply @attrs to an unbound workqueue @wq.  Unless disabled, on NUMA
4085  * machines, this function maps a separate pwq to each NUMA node with
4086  * possibles CPUs in @attrs->cpumask so that work items are affine to the
4087  * NUMA node it was issued on.  Older pwqs are released as in-flight work
4088  * items finish.  Note that a work item which repeatedly requeues itself
4089  * back-to-back will stay on its current pwq.
4090  *
4091  * Performs GFP_KERNEL allocations.
4092  *
4093  * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
4094  *
4095  * Return: 0 on success and -errno on failure.
4096  */
apply_workqueue_attrs(struct workqueue_struct * wq,const struct workqueue_attrs * attrs)4097 int apply_workqueue_attrs(struct workqueue_struct *wq,
4098 			  const struct workqueue_attrs *attrs)
4099 {
4100 	int ret;
4101 
4102 	lockdep_assert_cpus_held();
4103 
4104 	mutex_lock(&wq_pool_mutex);
4105 	ret = apply_workqueue_attrs_locked(wq, attrs);
4106 	mutex_unlock(&wq_pool_mutex);
4107 
4108 	return ret;
4109 }
4110 
4111 /**
4112  * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4113  * @wq: the target workqueue
4114  * @cpu: the CPU coming up or going down
4115  * @online: whether @cpu is coming up or going down
4116  *
4117  * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4118  * %CPU_DOWN_FAILED.  @cpu is being hot[un]plugged, update NUMA affinity of
4119  * @wq accordingly.
4120  *
4121  * If NUMA affinity can't be adjusted due to memory allocation failure, it
4122  * falls back to @wq->dfl_pwq which may not be optimal but is always
4123  * correct.
4124  *
4125  * Note that when the last allowed CPU of a NUMA node goes offline for a
4126  * workqueue with a cpumask spanning multiple nodes, the workers which were
4127  * already executing the work items for the workqueue will lose their CPU
4128  * affinity and may execute on any CPU.  This is similar to how per-cpu
4129  * workqueues behave on CPU_DOWN.  If a workqueue user wants strict
4130  * affinity, it's the user's responsibility to flush the work item from
4131  * CPU_DOWN_PREPARE.
4132  */
wq_update_unbound_numa(struct workqueue_struct * wq,int cpu,bool online)4133 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4134 				   bool online)
4135 {
4136 	int node = cpu_to_node(cpu);
4137 	int cpu_off = online ? -1 : cpu;
4138 	struct pool_workqueue *old_pwq = NULL, *pwq;
4139 	struct workqueue_attrs *target_attrs;
4140 	cpumask_t *cpumask;
4141 
4142 	lockdep_assert_held(&wq_pool_mutex);
4143 
4144 	if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4145 	    wq->unbound_attrs->no_numa)
4146 		return;
4147 
4148 	/*
4149 	 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4150 	 * Let's use a preallocated one.  The following buf is protected by
4151 	 * CPU hotplug exclusion.
4152 	 */
4153 	target_attrs = wq_update_unbound_numa_attrs_buf;
4154 	cpumask = target_attrs->cpumask;
4155 
4156 	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4157 	pwq = unbound_pwq_by_node(wq, node);
4158 
4159 	/*
4160 	 * Let's determine what needs to be done.  If the target cpumask is
4161 	 * different from the default pwq's, we need to compare it to @pwq's
4162 	 * and create a new one if they don't match.  If the target cpumask
4163 	 * equals the default pwq's, the default pwq should be used.
4164 	 */
4165 	if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4166 		if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4167 			return;
4168 	} else {
4169 		goto use_dfl_pwq;
4170 	}
4171 
4172 	/* create a new pwq */
4173 	pwq = alloc_unbound_pwq(wq, target_attrs);
4174 	if (!pwq) {
4175 		pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4176 			wq->name);
4177 		goto use_dfl_pwq;
4178 	}
4179 
4180 	/* Install the new pwq. */
4181 	mutex_lock(&wq->mutex);
4182 	old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4183 	goto out_unlock;
4184 
4185 use_dfl_pwq:
4186 	mutex_lock(&wq->mutex);
4187 	raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
4188 	get_pwq(wq->dfl_pwq);
4189 	raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4190 	old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4191 out_unlock:
4192 	mutex_unlock(&wq->mutex);
4193 	put_pwq_unlocked(old_pwq);
4194 }
4195 
alloc_and_link_pwqs(struct workqueue_struct * wq)4196 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4197 {
4198 	bool highpri = wq->flags & WQ_HIGHPRI;
4199 	int cpu, ret;
4200 
4201 	if (!(wq->flags & WQ_UNBOUND)) {
4202 		wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4203 		if (!wq->cpu_pwqs)
4204 			return -ENOMEM;
4205 
4206 		for_each_possible_cpu(cpu) {
4207 			struct pool_workqueue *pwq =
4208 				per_cpu_ptr(wq->cpu_pwqs, cpu);
4209 			struct worker_pool *cpu_pools =
4210 				per_cpu(cpu_worker_pools, cpu);
4211 
4212 			init_pwq(pwq, wq, &cpu_pools[highpri]);
4213 
4214 			mutex_lock(&wq->mutex);
4215 			link_pwq(pwq);
4216 			mutex_unlock(&wq->mutex);
4217 		}
4218 		return 0;
4219 	}
4220 
4221 	cpus_read_lock();
4222 	if (wq->flags & __WQ_ORDERED) {
4223 		ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4224 		/* there should only be single pwq for ordering guarantee */
4225 		WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4226 			      wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4227 		     "ordering guarantee broken for workqueue %s\n", wq->name);
4228 	} else {
4229 		ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4230 	}
4231 	cpus_read_unlock();
4232 
4233 	return ret;
4234 }
4235 
wq_clamp_max_active(int max_active,unsigned int flags,const char * name)4236 static int wq_clamp_max_active(int max_active, unsigned int flags,
4237 			       const char *name)
4238 {
4239 	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4240 
4241 	if (max_active < 1 || max_active > lim)
4242 		pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4243 			max_active, name, 1, lim);
4244 
4245 	return clamp_val(max_active, 1, lim);
4246 }
4247 
4248 /*
4249  * Workqueues which may be used during memory reclaim should have a rescuer
4250  * to guarantee forward progress.
4251  */
init_rescuer(struct workqueue_struct * wq)4252 static int init_rescuer(struct workqueue_struct *wq)
4253 {
4254 	struct worker *rescuer;
4255 	int ret;
4256 
4257 	if (!(wq->flags & WQ_MEM_RECLAIM))
4258 		return 0;
4259 
4260 	rescuer = alloc_worker(NUMA_NO_NODE);
4261 	if (!rescuer)
4262 		return -ENOMEM;
4263 
4264 	rescuer->rescue_wq = wq;
4265 	rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4266 	if (IS_ERR(rescuer->task)) {
4267 		ret = PTR_ERR(rescuer->task);
4268 		kfree(rescuer);
4269 		return ret;
4270 	}
4271 
4272 	wq->rescuer = rescuer;
4273 	kthread_bind_mask(rescuer->task, cpu_possible_mask);
4274 	wake_up_process(rescuer->task);
4275 
4276 	return 0;
4277 }
4278 
4279 __printf(1, 4)
alloc_workqueue(const char * fmt,unsigned int flags,int max_active,...)4280 struct workqueue_struct *alloc_workqueue(const char *fmt,
4281 					 unsigned int flags,
4282 					 int max_active, ...)
4283 {
4284 	size_t tbl_size = 0;
4285 	va_list args;
4286 	struct workqueue_struct *wq;
4287 	struct pool_workqueue *pwq;
4288 
4289 	/*
4290 	 * Unbound && max_active == 1 used to imply ordered, which is no
4291 	 * longer the case on NUMA machines due to per-node pools.  While
4292 	 * alloc_ordered_workqueue() is the right way to create an ordered
4293 	 * workqueue, keep the previous behavior to avoid subtle breakages
4294 	 * on NUMA.
4295 	 */
4296 	if ((flags & WQ_UNBOUND) && max_active == 1)
4297 		flags |= __WQ_ORDERED;
4298 
4299 	/* see the comment above the definition of WQ_POWER_EFFICIENT */
4300 	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4301 		flags |= WQ_UNBOUND;
4302 
4303 	/* allocate wq and format name */
4304 	if (flags & WQ_UNBOUND)
4305 		tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4306 
4307 	wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4308 	if (!wq)
4309 		return NULL;
4310 
4311 	if (flags & WQ_UNBOUND) {
4312 		wq->unbound_attrs = alloc_workqueue_attrs();
4313 		if (!wq->unbound_attrs)
4314 			goto err_free_wq;
4315 	}
4316 
4317 	va_start(args, max_active);
4318 	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4319 	va_end(args);
4320 
4321 	max_active = max_active ?: WQ_DFL_ACTIVE;
4322 	max_active = wq_clamp_max_active(max_active, flags, wq->name);
4323 
4324 	/* init wq */
4325 	wq->flags = flags;
4326 	wq->saved_max_active = max_active;
4327 	mutex_init(&wq->mutex);
4328 	atomic_set(&wq->nr_pwqs_to_flush, 0);
4329 	INIT_LIST_HEAD(&wq->pwqs);
4330 	INIT_LIST_HEAD(&wq->flusher_queue);
4331 	INIT_LIST_HEAD(&wq->flusher_overflow);
4332 	INIT_LIST_HEAD(&wq->maydays);
4333 
4334 	wq_init_lockdep(wq);
4335 	INIT_LIST_HEAD(&wq->list);
4336 
4337 	if (alloc_and_link_pwqs(wq) < 0)
4338 		goto err_unreg_lockdep;
4339 
4340 	if (wq_online && init_rescuer(wq) < 0)
4341 		goto err_destroy;
4342 
4343 	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4344 		goto err_destroy;
4345 
4346 	/*
4347 	 * wq_pool_mutex protects global freeze state and workqueues list.
4348 	 * Grab it, adjust max_active and add the new @wq to workqueues
4349 	 * list.
4350 	 */
4351 	mutex_lock(&wq_pool_mutex);
4352 
4353 	mutex_lock(&wq->mutex);
4354 	for_each_pwq(pwq, wq)
4355 		pwq_adjust_max_active(pwq);
4356 	mutex_unlock(&wq->mutex);
4357 
4358 	list_add_tail_rcu(&wq->list, &workqueues);
4359 
4360 	mutex_unlock(&wq_pool_mutex);
4361 
4362 	return wq;
4363 
4364 err_unreg_lockdep:
4365 	wq_unregister_lockdep(wq);
4366 	wq_free_lockdep(wq);
4367 err_free_wq:
4368 	free_workqueue_attrs(wq->unbound_attrs);
4369 	kfree(wq);
4370 	return NULL;
4371 err_destroy:
4372 	destroy_workqueue(wq);
4373 	return NULL;
4374 }
4375 EXPORT_SYMBOL_GPL(alloc_workqueue);
4376 
pwq_busy(struct pool_workqueue * pwq)4377 static bool pwq_busy(struct pool_workqueue *pwq)
4378 {
4379 	int i;
4380 
4381 	for (i = 0; i < WORK_NR_COLORS; i++)
4382 		if (pwq->nr_in_flight[i])
4383 			return true;
4384 
4385 	if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4386 		return true;
4387 	if (pwq->nr_active || !list_empty(&pwq->inactive_works))
4388 		return true;
4389 
4390 	return false;
4391 }
4392 
4393 /**
4394  * destroy_workqueue - safely terminate a workqueue
4395  * @wq: target workqueue
4396  *
4397  * Safely destroy a workqueue. All work currently pending will be done first.
4398  */
destroy_workqueue(struct workqueue_struct * wq)4399 void destroy_workqueue(struct workqueue_struct *wq)
4400 {
4401 	struct pool_workqueue *pwq;
4402 	int node;
4403 
4404 	/*
4405 	 * Remove it from sysfs first so that sanity check failure doesn't
4406 	 * lead to sysfs name conflicts.
4407 	 */
4408 	workqueue_sysfs_unregister(wq);
4409 
4410 	/* drain it before proceeding with destruction */
4411 	drain_workqueue(wq);
4412 
4413 	/* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4414 	if (wq->rescuer) {
4415 		struct worker *rescuer = wq->rescuer;
4416 
4417 		/* this prevents new queueing */
4418 		raw_spin_lock_irq(&wq_mayday_lock);
4419 		wq->rescuer = NULL;
4420 		raw_spin_unlock_irq(&wq_mayday_lock);
4421 
4422 		/* rescuer will empty maydays list before exiting */
4423 		kthread_stop(rescuer->task);
4424 		kfree(rescuer);
4425 	}
4426 
4427 	/*
4428 	 * Sanity checks - grab all the locks so that we wait for all
4429 	 * in-flight operations which may do put_pwq().
4430 	 */
4431 	mutex_lock(&wq_pool_mutex);
4432 	mutex_lock(&wq->mutex);
4433 	for_each_pwq(pwq, wq) {
4434 		raw_spin_lock_irq(&pwq->pool->lock);
4435 		if (WARN_ON(pwq_busy(pwq))) {
4436 			pr_warn("%s: %s has the following busy pwq\n",
4437 				__func__, wq->name);
4438 			show_pwq(pwq);
4439 			raw_spin_unlock_irq(&pwq->pool->lock);
4440 			mutex_unlock(&wq->mutex);
4441 			mutex_unlock(&wq_pool_mutex);
4442 			show_one_workqueue(wq);
4443 			return;
4444 		}
4445 		raw_spin_unlock_irq(&pwq->pool->lock);
4446 	}
4447 	mutex_unlock(&wq->mutex);
4448 
4449 	/*
4450 	 * wq list is used to freeze wq, remove from list after
4451 	 * flushing is complete in case freeze races us.
4452 	 */
4453 	list_del_rcu(&wq->list);
4454 	mutex_unlock(&wq_pool_mutex);
4455 
4456 	if (!(wq->flags & WQ_UNBOUND)) {
4457 		wq_unregister_lockdep(wq);
4458 		/*
4459 		 * The base ref is never dropped on per-cpu pwqs.  Directly
4460 		 * schedule RCU free.
4461 		 */
4462 		call_rcu(&wq->rcu, rcu_free_wq);
4463 	} else {
4464 		/*
4465 		 * We're the sole accessor of @wq at this point.  Directly
4466 		 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4467 		 * @wq will be freed when the last pwq is released.
4468 		 */
4469 		for_each_node(node) {
4470 			pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4471 			RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4472 			put_pwq_unlocked(pwq);
4473 		}
4474 
4475 		/*
4476 		 * Put dfl_pwq.  @wq may be freed any time after dfl_pwq is
4477 		 * put.  Don't access it afterwards.
4478 		 */
4479 		pwq = wq->dfl_pwq;
4480 		wq->dfl_pwq = NULL;
4481 		put_pwq_unlocked(pwq);
4482 	}
4483 }
4484 EXPORT_SYMBOL_GPL(destroy_workqueue);
4485 
4486 /**
4487  * workqueue_set_max_active - adjust max_active of a workqueue
4488  * @wq: target workqueue
4489  * @max_active: new max_active value.
4490  *
4491  * Set max_active of @wq to @max_active.
4492  *
4493  * CONTEXT:
4494  * Don't call from IRQ context.
4495  */
workqueue_set_max_active(struct workqueue_struct * wq,int max_active)4496 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4497 {
4498 	struct pool_workqueue *pwq;
4499 
4500 	/* disallow meddling with max_active for ordered workqueues */
4501 	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4502 		return;
4503 
4504 	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4505 
4506 	mutex_lock(&wq->mutex);
4507 
4508 	wq->flags &= ~__WQ_ORDERED;
4509 	wq->saved_max_active = max_active;
4510 
4511 	for_each_pwq(pwq, wq)
4512 		pwq_adjust_max_active(pwq);
4513 
4514 	mutex_unlock(&wq->mutex);
4515 }
4516 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4517 
4518 /**
4519  * current_work - retrieve %current task's work struct
4520  *
4521  * Determine if %current task is a workqueue worker and what it's working on.
4522  * Useful to find out the context that the %current task is running in.
4523  *
4524  * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4525  */
current_work(void)4526 struct work_struct *current_work(void)
4527 {
4528 	struct worker *worker = current_wq_worker();
4529 
4530 	return worker ? worker->current_work : NULL;
4531 }
4532 EXPORT_SYMBOL(current_work);
4533 
4534 /**
4535  * current_is_workqueue_rescuer - is %current workqueue rescuer?
4536  *
4537  * Determine whether %current is a workqueue rescuer.  Can be used from
4538  * work functions to determine whether it's being run off the rescuer task.
4539  *
4540  * Return: %true if %current is a workqueue rescuer. %false otherwise.
4541  */
current_is_workqueue_rescuer(void)4542 bool current_is_workqueue_rescuer(void)
4543 {
4544 	struct worker *worker = current_wq_worker();
4545 
4546 	return worker && worker->rescue_wq;
4547 }
4548 
4549 /**
4550  * workqueue_congested - test whether a workqueue is congested
4551  * @cpu: CPU in question
4552  * @wq: target workqueue
4553  *
4554  * Test whether @wq's cpu workqueue for @cpu is congested.  There is
4555  * no synchronization around this function and the test result is
4556  * unreliable and only useful as advisory hints or for debugging.
4557  *
4558  * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4559  * Note that both per-cpu and unbound workqueues may be associated with
4560  * multiple pool_workqueues which have separate congested states.  A
4561  * workqueue being congested on one CPU doesn't mean the workqueue is also
4562  * contested on other CPUs / NUMA nodes.
4563  *
4564  * Return:
4565  * %true if congested, %false otherwise.
4566  */
workqueue_congested(int cpu,struct workqueue_struct * wq)4567 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4568 {
4569 	struct pool_workqueue *pwq;
4570 	bool ret;
4571 
4572 	rcu_read_lock();
4573 	preempt_disable();
4574 
4575 	if (cpu == WORK_CPU_UNBOUND)
4576 		cpu = smp_processor_id();
4577 
4578 	if (!(wq->flags & WQ_UNBOUND))
4579 		pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4580 	else
4581 		pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4582 
4583 	ret = !list_empty(&pwq->inactive_works);
4584 	preempt_enable();
4585 	rcu_read_unlock();
4586 
4587 	return ret;
4588 }
4589 EXPORT_SYMBOL_GPL(workqueue_congested);
4590 
4591 /**
4592  * work_busy - test whether a work is currently pending or running
4593  * @work: the work to be tested
4594  *
4595  * Test whether @work is currently pending or running.  There is no
4596  * synchronization around this function and the test result is
4597  * unreliable and only useful as advisory hints or for debugging.
4598  *
4599  * Return:
4600  * OR'd bitmask of WORK_BUSY_* bits.
4601  */
work_busy(struct work_struct * work)4602 unsigned int work_busy(struct work_struct *work)
4603 {
4604 	struct worker_pool *pool;
4605 	unsigned long flags;
4606 	unsigned int ret = 0;
4607 
4608 	if (work_pending(work))
4609 		ret |= WORK_BUSY_PENDING;
4610 
4611 	rcu_read_lock();
4612 	pool = get_work_pool(work);
4613 	if (pool) {
4614 		raw_spin_lock_irqsave(&pool->lock, flags);
4615 		if (find_worker_executing_work(pool, work))
4616 			ret |= WORK_BUSY_RUNNING;
4617 		raw_spin_unlock_irqrestore(&pool->lock, flags);
4618 	}
4619 	rcu_read_unlock();
4620 
4621 	return ret;
4622 }
4623 EXPORT_SYMBOL_GPL(work_busy);
4624 
4625 /**
4626  * set_worker_desc - set description for the current work item
4627  * @fmt: printf-style format string
4628  * @...: arguments for the format string
4629  *
4630  * This function can be called by a running work function to describe what
4631  * the work item is about.  If the worker task gets dumped, this
4632  * information will be printed out together to help debugging.  The
4633  * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4634  */
set_worker_desc(const char * fmt,...)4635 void set_worker_desc(const char *fmt, ...)
4636 {
4637 	struct worker *worker = current_wq_worker();
4638 	va_list args;
4639 
4640 	if (worker) {
4641 		va_start(args, fmt);
4642 		vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4643 		va_end(args);
4644 	}
4645 }
4646 EXPORT_SYMBOL_GPL(set_worker_desc);
4647 
4648 /**
4649  * print_worker_info - print out worker information and description
4650  * @log_lvl: the log level to use when printing
4651  * @task: target task
4652  *
4653  * If @task is a worker and currently executing a work item, print out the
4654  * name of the workqueue being serviced and worker description set with
4655  * set_worker_desc() by the currently executing work item.
4656  *
4657  * This function can be safely called on any task as long as the
4658  * task_struct itself is accessible.  While safe, this function isn't
4659  * synchronized and may print out mixups or garbages of limited length.
4660  */
print_worker_info(const char * log_lvl,struct task_struct * task)4661 void print_worker_info(const char *log_lvl, struct task_struct *task)
4662 {
4663 	work_func_t *fn = NULL;
4664 	char name[WQ_NAME_LEN] = { };
4665 	char desc[WORKER_DESC_LEN] = { };
4666 	struct pool_workqueue *pwq = NULL;
4667 	struct workqueue_struct *wq = NULL;
4668 	struct worker *worker;
4669 
4670 	if (!(task->flags & PF_WQ_WORKER))
4671 		return;
4672 
4673 	/*
4674 	 * This function is called without any synchronization and @task
4675 	 * could be in any state.  Be careful with dereferences.
4676 	 */
4677 	worker = kthread_probe_data(task);
4678 
4679 	/*
4680 	 * Carefully copy the associated workqueue's workfn, name and desc.
4681 	 * Keep the original last '\0' in case the original is garbage.
4682 	 */
4683 	copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
4684 	copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
4685 	copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
4686 	copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
4687 	copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
4688 
4689 	if (fn || name[0] || desc[0]) {
4690 		printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
4691 		if (strcmp(name, desc))
4692 			pr_cont(" (%s)", desc);
4693 		pr_cont("\n");
4694 	}
4695 }
4696 
pr_cont_pool_info(struct worker_pool * pool)4697 static void pr_cont_pool_info(struct worker_pool *pool)
4698 {
4699 	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4700 	if (pool->node != NUMA_NO_NODE)
4701 		pr_cont(" node=%d", pool->node);
4702 	pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4703 }
4704 
pr_cont_work(bool comma,struct work_struct * work)4705 static void pr_cont_work(bool comma, struct work_struct *work)
4706 {
4707 	if (work->func == wq_barrier_func) {
4708 		struct wq_barrier *barr;
4709 
4710 		barr = container_of(work, struct wq_barrier, work);
4711 
4712 		pr_cont("%s BAR(%d)", comma ? "," : "",
4713 			task_pid_nr(barr->task));
4714 	} else {
4715 		pr_cont("%s %ps", comma ? "," : "", work->func);
4716 	}
4717 }
4718 
show_pwq(struct pool_workqueue * pwq)4719 static void show_pwq(struct pool_workqueue *pwq)
4720 {
4721 	struct worker_pool *pool = pwq->pool;
4722 	struct work_struct *work;
4723 	struct worker *worker;
4724 	bool has_in_flight = false, has_pending = false;
4725 	int bkt;
4726 
4727 	pr_info("  pwq %d:", pool->id);
4728 	pr_cont_pool_info(pool);
4729 
4730 	pr_cont(" active=%d/%d refcnt=%d%s\n",
4731 		pwq->nr_active, pwq->max_active, pwq->refcnt,
4732 		!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4733 
4734 	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4735 		if (worker->current_pwq == pwq) {
4736 			has_in_flight = true;
4737 			break;
4738 		}
4739 	}
4740 	if (has_in_flight) {
4741 		bool comma = false;
4742 
4743 		pr_info("    in-flight:");
4744 		hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4745 			if (worker->current_pwq != pwq)
4746 				continue;
4747 
4748 			pr_cont("%s %d%s:%ps", comma ? "," : "",
4749 				task_pid_nr(worker->task),
4750 				worker->rescue_wq ? "(RESCUER)" : "",
4751 				worker->current_func);
4752 			list_for_each_entry(work, &worker->scheduled, entry)
4753 				pr_cont_work(false, work);
4754 			comma = true;
4755 		}
4756 		pr_cont("\n");
4757 	}
4758 
4759 	list_for_each_entry(work, &pool->worklist, entry) {
4760 		if (get_work_pwq(work) == pwq) {
4761 			has_pending = true;
4762 			break;
4763 		}
4764 	}
4765 	if (has_pending) {
4766 		bool comma = false;
4767 
4768 		pr_info("    pending:");
4769 		list_for_each_entry(work, &pool->worklist, entry) {
4770 			if (get_work_pwq(work) != pwq)
4771 				continue;
4772 
4773 			pr_cont_work(comma, work);
4774 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4775 		}
4776 		pr_cont("\n");
4777 	}
4778 
4779 	if (!list_empty(&pwq->inactive_works)) {
4780 		bool comma = false;
4781 
4782 		pr_info("    inactive:");
4783 		list_for_each_entry(work, &pwq->inactive_works, entry) {
4784 			pr_cont_work(comma, work);
4785 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4786 		}
4787 		pr_cont("\n");
4788 	}
4789 }
4790 
4791 /**
4792  * show_one_workqueue - dump state of specified workqueue
4793  * @wq: workqueue whose state will be printed
4794  */
show_one_workqueue(struct workqueue_struct * wq)4795 void show_one_workqueue(struct workqueue_struct *wq)
4796 {
4797 	struct pool_workqueue *pwq;
4798 	bool idle = true;
4799 	unsigned long flags;
4800 
4801 	for_each_pwq(pwq, wq) {
4802 		if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
4803 			idle = false;
4804 			break;
4805 		}
4806 	}
4807 	if (idle) /* Nothing to print for idle workqueue */
4808 		return;
4809 
4810 	pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4811 
4812 	for_each_pwq(pwq, wq) {
4813 		raw_spin_lock_irqsave(&pwq->pool->lock, flags);
4814 		if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
4815 			/*
4816 			 * Defer printing to avoid deadlocks in console
4817 			 * drivers that queue work while holding locks
4818 			 * also taken in their write paths.
4819 			 */
4820 			printk_deferred_enter();
4821 			show_pwq(pwq);
4822 			printk_deferred_exit();
4823 		}
4824 		raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
4825 		/*
4826 		 * We could be printing a lot from atomic context, e.g.
4827 		 * sysrq-t -> show_all_workqueues(). Avoid triggering
4828 		 * hard lockup.
4829 		 */
4830 		touch_nmi_watchdog();
4831 	}
4832 
4833 }
4834 
4835 /**
4836  * show_one_worker_pool - dump state of specified worker pool
4837  * @pool: worker pool whose state will be printed
4838  */
show_one_worker_pool(struct worker_pool * pool)4839 static void show_one_worker_pool(struct worker_pool *pool)
4840 {
4841 	struct worker *worker;
4842 	bool first = true;
4843 	unsigned long flags;
4844 
4845 	raw_spin_lock_irqsave(&pool->lock, flags);
4846 	if (pool->nr_workers == pool->nr_idle)
4847 		goto next_pool;
4848 	/*
4849 	 * Defer printing to avoid deadlocks in console drivers that
4850 	 * queue work while holding locks also taken in their write
4851 	 * paths.
4852 	 */
4853 	printk_deferred_enter();
4854 	pr_info("pool %d:", pool->id);
4855 	pr_cont_pool_info(pool);
4856 	pr_cont(" hung=%us workers=%d",
4857 		jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4858 		pool->nr_workers);
4859 	if (pool->manager)
4860 		pr_cont(" manager: %d",
4861 			task_pid_nr(pool->manager->task));
4862 	list_for_each_entry(worker, &pool->idle_list, entry) {
4863 		pr_cont(" %s%d", first ? "idle: " : "",
4864 			task_pid_nr(worker->task));
4865 		first = false;
4866 	}
4867 	pr_cont("\n");
4868 	printk_deferred_exit();
4869 next_pool:
4870 	raw_spin_unlock_irqrestore(&pool->lock, flags);
4871 	/*
4872 	 * We could be printing a lot from atomic context, e.g.
4873 	 * sysrq-t -> show_all_workqueues(). Avoid triggering
4874 	 * hard lockup.
4875 	 */
4876 	touch_nmi_watchdog();
4877 
4878 }
4879 
4880 /**
4881  * show_all_workqueues - dump workqueue state
4882  *
4883  * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4884  * all busy workqueues and pools.
4885  */
show_all_workqueues(void)4886 void show_all_workqueues(void)
4887 {
4888 	struct workqueue_struct *wq;
4889 	struct worker_pool *pool;
4890 	int pi;
4891 
4892 	rcu_read_lock();
4893 
4894 	pr_info("Showing busy workqueues and worker pools:\n");
4895 
4896 	list_for_each_entry_rcu(wq, &workqueues, list)
4897 		show_one_workqueue(wq);
4898 
4899 	for_each_pool(pool, pi)
4900 		show_one_worker_pool(pool);
4901 
4902 	rcu_read_unlock();
4903 }
4904 
4905 /* used to show worker information through /proc/PID/{comm,stat,status} */
wq_worker_comm(char * buf,size_t size,struct task_struct * task)4906 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4907 {
4908 	int off;
4909 
4910 	/* always show the actual comm */
4911 	off = strscpy(buf, task->comm, size);
4912 	if (off < 0)
4913 		return;
4914 
4915 	/* stabilize PF_WQ_WORKER and worker pool association */
4916 	mutex_lock(&wq_pool_attach_mutex);
4917 
4918 	if (task->flags & PF_WQ_WORKER) {
4919 		struct worker *worker = kthread_data(task);
4920 		struct worker_pool *pool = worker->pool;
4921 
4922 		if (pool) {
4923 			raw_spin_lock_irq(&pool->lock);
4924 			/*
4925 			 * ->desc tracks information (wq name or
4926 			 * set_worker_desc()) for the latest execution.  If
4927 			 * current, prepend '+', otherwise '-'.
4928 			 */
4929 			if (worker->desc[0] != '\0') {
4930 				if (worker->current_work)
4931 					scnprintf(buf + off, size - off, "+%s",
4932 						  worker->desc);
4933 				else
4934 					scnprintf(buf + off, size - off, "-%s",
4935 						  worker->desc);
4936 			}
4937 			raw_spin_unlock_irq(&pool->lock);
4938 		}
4939 	}
4940 
4941 	mutex_unlock(&wq_pool_attach_mutex);
4942 }
4943 
4944 #ifdef CONFIG_SMP
4945 
4946 /*
4947  * CPU hotplug.
4948  *
4949  * There are two challenges in supporting CPU hotplug.  Firstly, there
4950  * are a lot of assumptions on strong associations among work, pwq and
4951  * pool which make migrating pending and scheduled works very
4952  * difficult to implement without impacting hot paths.  Secondly,
4953  * worker pools serve mix of short, long and very long running works making
4954  * blocked draining impractical.
4955  *
4956  * This is solved by allowing the pools to be disassociated from the CPU
4957  * running as an unbound one and allowing it to be reattached later if the
4958  * cpu comes back online.
4959  */
4960 
unbind_workers(int cpu)4961 static void unbind_workers(int cpu)
4962 {
4963 	struct worker_pool *pool;
4964 	struct worker *worker;
4965 
4966 	for_each_cpu_worker_pool(pool, cpu) {
4967 		mutex_lock(&wq_pool_attach_mutex);
4968 		raw_spin_lock_irq(&pool->lock);
4969 
4970 		/*
4971 		 * We've blocked all attach/detach operations. Make all workers
4972 		 * unbound and set DISASSOCIATED.  Before this, all workers
4973 		 * must be on the cpu.  After this, they may become diasporas.
4974 		 * And the preemption disabled section in their sched callbacks
4975 		 * are guaranteed to see WORKER_UNBOUND since the code here
4976 		 * is on the same cpu.
4977 		 */
4978 		for_each_pool_worker(worker, pool)
4979 			worker->flags |= WORKER_UNBOUND;
4980 
4981 		pool->flags |= POOL_DISASSOCIATED;
4982 
4983 		/*
4984 		 * The handling of nr_running in sched callbacks are disabled
4985 		 * now.  Zap nr_running.  After this, nr_running stays zero and
4986 		 * need_more_worker() and keep_working() are always true as
4987 		 * long as the worklist is not empty.  This pool now behaves as
4988 		 * an unbound (in terms of concurrency management) pool which
4989 		 * are served by workers tied to the pool.
4990 		 */
4991 		pool->nr_running = 0;
4992 
4993 		/*
4994 		 * With concurrency management just turned off, a busy
4995 		 * worker blocking could lead to lengthy stalls.  Kick off
4996 		 * unbound chain execution of currently pending work items.
4997 		 */
4998 		wake_up_worker(pool);
4999 
5000 		raw_spin_unlock_irq(&pool->lock);
5001 
5002 		for_each_pool_worker(worker, pool) {
5003 			kthread_set_per_cpu(worker->task, -1);
5004 			if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
5005 				WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
5006 			else
5007 				WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
5008 		}
5009 
5010 		mutex_unlock(&wq_pool_attach_mutex);
5011 	}
5012 }
5013 
5014 /**
5015  * rebind_workers - rebind all workers of a pool to the associated CPU
5016  * @pool: pool of interest
5017  *
5018  * @pool->cpu is coming online.  Rebind all workers to the CPU.
5019  */
rebind_workers(struct worker_pool * pool)5020 static void rebind_workers(struct worker_pool *pool)
5021 {
5022 	struct worker *worker;
5023 
5024 	lockdep_assert_held(&wq_pool_attach_mutex);
5025 
5026 	/*
5027 	 * Restore CPU affinity of all workers.  As all idle workers should
5028 	 * be on the run-queue of the associated CPU before any local
5029 	 * wake-ups for concurrency management happen, restore CPU affinity
5030 	 * of all workers first and then clear UNBOUND.  As we're called
5031 	 * from CPU_ONLINE, the following shouldn't fail.
5032 	 */
5033 	for_each_pool_worker(worker, pool) {
5034 		kthread_set_per_cpu(worker->task, pool->cpu);
5035 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
5036 						  pool->attrs->cpumask) < 0);
5037 	}
5038 
5039 	raw_spin_lock_irq(&pool->lock);
5040 
5041 	pool->flags &= ~POOL_DISASSOCIATED;
5042 
5043 	for_each_pool_worker(worker, pool) {
5044 		unsigned int worker_flags = worker->flags;
5045 
5046 		/*
5047 		 * We want to clear UNBOUND but can't directly call
5048 		 * worker_clr_flags() or adjust nr_running.  Atomically
5049 		 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5050 		 * @worker will clear REBOUND using worker_clr_flags() when
5051 		 * it initiates the next execution cycle thus restoring
5052 		 * concurrency management.  Note that when or whether
5053 		 * @worker clears REBOUND doesn't affect correctness.
5054 		 *
5055 		 * WRITE_ONCE() is necessary because @worker->flags may be
5056 		 * tested without holding any lock in
5057 		 * wq_worker_running().  Without it, NOT_RUNNING test may
5058 		 * fail incorrectly leading to premature concurrency
5059 		 * management operations.
5060 		 */
5061 		WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
5062 		worker_flags |= WORKER_REBOUND;
5063 		worker_flags &= ~WORKER_UNBOUND;
5064 		WRITE_ONCE(worker->flags, worker_flags);
5065 	}
5066 
5067 	raw_spin_unlock_irq(&pool->lock);
5068 }
5069 
5070 /**
5071  * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5072  * @pool: unbound pool of interest
5073  * @cpu: the CPU which is coming up
5074  *
5075  * An unbound pool may end up with a cpumask which doesn't have any online
5076  * CPUs.  When a worker of such pool get scheduled, the scheduler resets
5077  * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
5078  * online CPU before, cpus_allowed of all its workers should be restored.
5079  */
restore_unbound_workers_cpumask(struct worker_pool * pool,int cpu)5080 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5081 {
5082 	static cpumask_t cpumask;
5083 	struct worker *worker;
5084 
5085 	lockdep_assert_held(&wq_pool_attach_mutex);
5086 
5087 	/* is @cpu allowed for @pool? */
5088 	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5089 		return;
5090 
5091 	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5092 
5093 	/* as we're called from CPU_ONLINE, the following shouldn't fail */
5094 	for_each_pool_worker(worker, pool)
5095 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5096 }
5097 
workqueue_prepare_cpu(unsigned int cpu)5098 int workqueue_prepare_cpu(unsigned int cpu)
5099 {
5100 	struct worker_pool *pool;
5101 
5102 	for_each_cpu_worker_pool(pool, cpu) {
5103 		if (pool->nr_workers)
5104 			continue;
5105 		if (!create_worker(pool))
5106 			return -ENOMEM;
5107 	}
5108 	return 0;
5109 }
5110 
workqueue_online_cpu(unsigned int cpu)5111 int workqueue_online_cpu(unsigned int cpu)
5112 {
5113 	struct worker_pool *pool;
5114 	struct workqueue_struct *wq;
5115 	int pi;
5116 
5117 	mutex_lock(&wq_pool_mutex);
5118 
5119 	for_each_pool(pool, pi) {
5120 		mutex_lock(&wq_pool_attach_mutex);
5121 
5122 		if (pool->cpu == cpu)
5123 			rebind_workers(pool);
5124 		else if (pool->cpu < 0)
5125 			restore_unbound_workers_cpumask(pool, cpu);
5126 
5127 		mutex_unlock(&wq_pool_attach_mutex);
5128 	}
5129 
5130 	/* update NUMA affinity of unbound workqueues */
5131 	list_for_each_entry(wq, &workqueues, list)
5132 		wq_update_unbound_numa(wq, cpu, true);
5133 
5134 	mutex_unlock(&wq_pool_mutex);
5135 	return 0;
5136 }
5137 
workqueue_offline_cpu(unsigned int cpu)5138 int workqueue_offline_cpu(unsigned int cpu)
5139 {
5140 	struct workqueue_struct *wq;
5141 
5142 	/* unbinding per-cpu workers should happen on the local CPU */
5143 	if (WARN_ON(cpu != smp_processor_id()))
5144 		return -1;
5145 
5146 	unbind_workers(cpu);
5147 
5148 	/* update NUMA affinity of unbound workqueues */
5149 	mutex_lock(&wq_pool_mutex);
5150 	list_for_each_entry(wq, &workqueues, list)
5151 		wq_update_unbound_numa(wq, cpu, false);
5152 	mutex_unlock(&wq_pool_mutex);
5153 
5154 	return 0;
5155 }
5156 
5157 struct work_for_cpu {
5158 	struct work_struct work;
5159 	long (*fn)(void *);
5160 	void *arg;
5161 	long ret;
5162 };
5163 
work_for_cpu_fn(struct work_struct * work)5164 static void work_for_cpu_fn(struct work_struct *work)
5165 {
5166 	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5167 
5168 	wfc->ret = wfc->fn(wfc->arg);
5169 }
5170 
5171 /**
5172  * work_on_cpu - run a function in thread context on a particular cpu
5173  * @cpu: the cpu to run on
5174  * @fn: the function to run
5175  * @arg: the function arg
5176  *
5177  * It is up to the caller to ensure that the cpu doesn't go offline.
5178  * The caller must not hold any locks which would prevent @fn from completing.
5179  *
5180  * Return: The value @fn returns.
5181  */
work_on_cpu(int cpu,long (* fn)(void *),void * arg)5182 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5183 {
5184 	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5185 
5186 	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5187 	schedule_work_on(cpu, &wfc.work);
5188 	flush_work(&wfc.work);
5189 	destroy_work_on_stack(&wfc.work);
5190 	return wfc.ret;
5191 }
5192 EXPORT_SYMBOL_GPL(work_on_cpu);
5193 
5194 /**
5195  * work_on_cpu_safe - run a function in thread context on a particular cpu
5196  * @cpu: the cpu to run on
5197  * @fn:  the function to run
5198  * @arg: the function argument
5199  *
5200  * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5201  * any locks which would prevent @fn from completing.
5202  *
5203  * Return: The value @fn returns.
5204  */
work_on_cpu_safe(int cpu,long (* fn)(void *),void * arg)5205 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5206 {
5207 	long ret = -ENODEV;
5208 
5209 	cpus_read_lock();
5210 	if (cpu_online(cpu))
5211 		ret = work_on_cpu(cpu, fn, arg);
5212 	cpus_read_unlock();
5213 	return ret;
5214 }
5215 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5216 #endif /* CONFIG_SMP */
5217 
5218 #ifdef CONFIG_FREEZER
5219 
5220 /**
5221  * freeze_workqueues_begin - begin freezing workqueues
5222  *
5223  * Start freezing workqueues.  After this function returns, all freezable
5224  * workqueues will queue new works to their inactive_works list instead of
5225  * pool->worklist.
5226  *
5227  * CONTEXT:
5228  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5229  */
freeze_workqueues_begin(void)5230 void freeze_workqueues_begin(void)
5231 {
5232 	struct workqueue_struct *wq;
5233 	struct pool_workqueue *pwq;
5234 
5235 	mutex_lock(&wq_pool_mutex);
5236 
5237 	WARN_ON_ONCE(workqueue_freezing);
5238 	workqueue_freezing = true;
5239 
5240 	list_for_each_entry(wq, &workqueues, list) {
5241 		mutex_lock(&wq->mutex);
5242 		for_each_pwq(pwq, wq)
5243 			pwq_adjust_max_active(pwq);
5244 		mutex_unlock(&wq->mutex);
5245 	}
5246 
5247 	mutex_unlock(&wq_pool_mutex);
5248 }
5249 
5250 /**
5251  * freeze_workqueues_busy - are freezable workqueues still busy?
5252  *
5253  * Check whether freezing is complete.  This function must be called
5254  * between freeze_workqueues_begin() and thaw_workqueues().
5255  *
5256  * CONTEXT:
5257  * Grabs and releases wq_pool_mutex.
5258  *
5259  * Return:
5260  * %true if some freezable workqueues are still busy.  %false if freezing
5261  * is complete.
5262  */
freeze_workqueues_busy(void)5263 bool freeze_workqueues_busy(void)
5264 {
5265 	bool busy = false;
5266 	struct workqueue_struct *wq;
5267 	struct pool_workqueue *pwq;
5268 
5269 	mutex_lock(&wq_pool_mutex);
5270 
5271 	WARN_ON_ONCE(!workqueue_freezing);
5272 
5273 	list_for_each_entry(wq, &workqueues, list) {
5274 		if (!(wq->flags & WQ_FREEZABLE))
5275 			continue;
5276 		/*
5277 		 * nr_active is monotonically decreasing.  It's safe
5278 		 * to peek without lock.
5279 		 */
5280 		rcu_read_lock();
5281 		for_each_pwq(pwq, wq) {
5282 			WARN_ON_ONCE(pwq->nr_active < 0);
5283 			if (pwq->nr_active) {
5284 				busy = true;
5285 				rcu_read_unlock();
5286 				goto out_unlock;
5287 			}
5288 		}
5289 		rcu_read_unlock();
5290 	}
5291 out_unlock:
5292 	mutex_unlock(&wq_pool_mutex);
5293 	return busy;
5294 }
5295 
5296 /**
5297  * thaw_workqueues - thaw workqueues
5298  *
5299  * Thaw workqueues.  Normal queueing is restored and all collected
5300  * frozen works are transferred to their respective pool worklists.
5301  *
5302  * CONTEXT:
5303  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5304  */
thaw_workqueues(void)5305 void thaw_workqueues(void)
5306 {
5307 	struct workqueue_struct *wq;
5308 	struct pool_workqueue *pwq;
5309 
5310 	mutex_lock(&wq_pool_mutex);
5311 
5312 	if (!workqueue_freezing)
5313 		goto out_unlock;
5314 
5315 	workqueue_freezing = false;
5316 
5317 	/* restore max_active and repopulate worklist */
5318 	list_for_each_entry(wq, &workqueues, list) {
5319 		mutex_lock(&wq->mutex);
5320 		for_each_pwq(pwq, wq)
5321 			pwq_adjust_max_active(pwq);
5322 		mutex_unlock(&wq->mutex);
5323 	}
5324 
5325 out_unlock:
5326 	mutex_unlock(&wq_pool_mutex);
5327 }
5328 #endif /* CONFIG_FREEZER */
5329 
workqueue_apply_unbound_cpumask(void)5330 static int workqueue_apply_unbound_cpumask(void)
5331 {
5332 	LIST_HEAD(ctxs);
5333 	int ret = 0;
5334 	struct workqueue_struct *wq;
5335 	struct apply_wqattrs_ctx *ctx, *n;
5336 
5337 	lockdep_assert_held(&wq_pool_mutex);
5338 
5339 	list_for_each_entry(wq, &workqueues, list) {
5340 		if (!(wq->flags & WQ_UNBOUND))
5341 			continue;
5342 		/* creating multiple pwqs breaks ordering guarantee */
5343 		if (wq->flags & __WQ_ORDERED)
5344 			continue;
5345 
5346 		ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5347 		if (!ctx) {
5348 			ret = -ENOMEM;
5349 			break;
5350 		}
5351 
5352 		list_add_tail(&ctx->list, &ctxs);
5353 	}
5354 
5355 	list_for_each_entry_safe(ctx, n, &ctxs, list) {
5356 		if (!ret)
5357 			apply_wqattrs_commit(ctx);
5358 		apply_wqattrs_cleanup(ctx);
5359 	}
5360 
5361 	return ret;
5362 }
5363 
5364 /**
5365  *  workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5366  *  @cpumask: the cpumask to set
5367  *
5368  *  The low-level workqueues cpumask is a global cpumask that limits
5369  *  the affinity of all unbound workqueues.  This function check the @cpumask
5370  *  and apply it to all unbound workqueues and updates all pwqs of them.
5371  *
5372  *  Return:	0	- Success
5373  *  		-EINVAL	- Invalid @cpumask
5374  *  		-ENOMEM	- Failed to allocate memory for attrs or pwqs.
5375  */
workqueue_set_unbound_cpumask(cpumask_var_t cpumask)5376 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5377 {
5378 	int ret = -EINVAL;
5379 	cpumask_var_t saved_cpumask;
5380 
5381 	/*
5382 	 * Not excluding isolated cpus on purpose.
5383 	 * If the user wishes to include them, we allow that.
5384 	 */
5385 	cpumask_and(cpumask, cpumask, cpu_possible_mask);
5386 	if (!cpumask_empty(cpumask)) {
5387 		apply_wqattrs_lock();
5388 		if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
5389 			ret = 0;
5390 			goto out_unlock;
5391 		}
5392 
5393 		if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) {
5394 			ret = -ENOMEM;
5395 			goto out_unlock;
5396 		}
5397 
5398 		/* save the old wq_unbound_cpumask. */
5399 		cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5400 
5401 		/* update wq_unbound_cpumask at first and apply it to wqs. */
5402 		cpumask_copy(wq_unbound_cpumask, cpumask);
5403 		ret = workqueue_apply_unbound_cpumask();
5404 
5405 		/* restore the wq_unbound_cpumask when failed. */
5406 		if (ret < 0)
5407 			cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5408 
5409 		free_cpumask_var(saved_cpumask);
5410 out_unlock:
5411 		apply_wqattrs_unlock();
5412 	}
5413 
5414 	return ret;
5415 }
5416 
5417 #ifdef CONFIG_SYSFS
5418 /*
5419  * Workqueues with WQ_SYSFS flag set is visible to userland via
5420  * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
5421  * following attributes.
5422  *
5423  *  per_cpu	RO bool	: whether the workqueue is per-cpu or unbound
5424  *  max_active	RW int	: maximum number of in-flight work items
5425  *
5426  * Unbound workqueues have the following extra attributes.
5427  *
5428  *  pool_ids	RO int	: the associated pool IDs for each node
5429  *  nice	RW int	: nice value of the workers
5430  *  cpumask	RW mask	: bitmask of allowed CPUs for the workers
5431  *  numa	RW bool	: whether enable NUMA affinity
5432  */
5433 struct wq_device {
5434 	struct workqueue_struct		*wq;
5435 	struct device			dev;
5436 };
5437 
dev_to_wq(struct device * dev)5438 static struct workqueue_struct *dev_to_wq(struct device *dev)
5439 {
5440 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5441 
5442 	return wq_dev->wq;
5443 }
5444 
per_cpu_show(struct device * dev,struct device_attribute * attr,char * buf)5445 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5446 			    char *buf)
5447 {
5448 	struct workqueue_struct *wq = dev_to_wq(dev);
5449 
5450 	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5451 }
5452 static DEVICE_ATTR_RO(per_cpu);
5453 
max_active_show(struct device * dev,struct device_attribute * attr,char * buf)5454 static ssize_t max_active_show(struct device *dev,
5455 			       struct device_attribute *attr, char *buf)
5456 {
5457 	struct workqueue_struct *wq = dev_to_wq(dev);
5458 
5459 	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5460 }
5461 
max_active_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)5462 static ssize_t max_active_store(struct device *dev,
5463 				struct device_attribute *attr, const char *buf,
5464 				size_t count)
5465 {
5466 	struct workqueue_struct *wq = dev_to_wq(dev);
5467 	int val;
5468 
5469 	if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5470 		return -EINVAL;
5471 
5472 	workqueue_set_max_active(wq, val);
5473 	return count;
5474 }
5475 static DEVICE_ATTR_RW(max_active);
5476 
5477 static struct attribute *wq_sysfs_attrs[] = {
5478 	&dev_attr_per_cpu.attr,
5479 	&dev_attr_max_active.attr,
5480 	NULL,
5481 };
5482 ATTRIBUTE_GROUPS(wq_sysfs);
5483 
wq_pool_ids_show(struct device * dev,struct device_attribute * attr,char * buf)5484 static ssize_t wq_pool_ids_show(struct device *dev,
5485 				struct device_attribute *attr, char *buf)
5486 {
5487 	struct workqueue_struct *wq = dev_to_wq(dev);
5488 	const char *delim = "";
5489 	int node, written = 0;
5490 
5491 	cpus_read_lock();
5492 	rcu_read_lock();
5493 	for_each_node(node) {
5494 		written += scnprintf(buf + written, PAGE_SIZE - written,
5495 				     "%s%d:%d", delim, node,
5496 				     unbound_pwq_by_node(wq, node)->pool->id);
5497 		delim = " ";
5498 	}
5499 	written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5500 	rcu_read_unlock();
5501 	cpus_read_unlock();
5502 
5503 	return written;
5504 }
5505 
wq_nice_show(struct device * dev,struct device_attribute * attr,char * buf)5506 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5507 			    char *buf)
5508 {
5509 	struct workqueue_struct *wq = dev_to_wq(dev);
5510 	int written;
5511 
5512 	mutex_lock(&wq->mutex);
5513 	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5514 	mutex_unlock(&wq->mutex);
5515 
5516 	return written;
5517 }
5518 
5519 /* prepare workqueue_attrs for sysfs store operations */
wq_sysfs_prep_attrs(struct workqueue_struct * wq)5520 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5521 {
5522 	struct workqueue_attrs *attrs;
5523 
5524 	lockdep_assert_held(&wq_pool_mutex);
5525 
5526 	attrs = alloc_workqueue_attrs();
5527 	if (!attrs)
5528 		return NULL;
5529 
5530 	copy_workqueue_attrs(attrs, wq->unbound_attrs);
5531 	return attrs;
5532 }
5533 
wq_nice_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)5534 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5535 			     const char *buf, size_t count)
5536 {
5537 	struct workqueue_struct *wq = dev_to_wq(dev);
5538 	struct workqueue_attrs *attrs;
5539 	int ret = -ENOMEM;
5540 
5541 	apply_wqattrs_lock();
5542 
5543 	attrs = wq_sysfs_prep_attrs(wq);
5544 	if (!attrs)
5545 		goto out_unlock;
5546 
5547 	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5548 	    attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5549 		ret = apply_workqueue_attrs_locked(wq, attrs);
5550 	else
5551 		ret = -EINVAL;
5552 
5553 out_unlock:
5554 	apply_wqattrs_unlock();
5555 	free_workqueue_attrs(attrs);
5556 	return ret ?: count;
5557 }
5558 
wq_cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)5559 static ssize_t wq_cpumask_show(struct device *dev,
5560 			       struct device_attribute *attr, char *buf)
5561 {
5562 	struct workqueue_struct *wq = dev_to_wq(dev);
5563 	int written;
5564 
5565 	mutex_lock(&wq->mutex);
5566 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5567 			    cpumask_pr_args(wq->unbound_attrs->cpumask));
5568 	mutex_unlock(&wq->mutex);
5569 	return written;
5570 }
5571 
wq_cpumask_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)5572 static ssize_t wq_cpumask_store(struct device *dev,
5573 				struct device_attribute *attr,
5574 				const char *buf, size_t count)
5575 {
5576 	struct workqueue_struct *wq = dev_to_wq(dev);
5577 	struct workqueue_attrs *attrs;
5578 	int ret = -ENOMEM;
5579 
5580 	apply_wqattrs_lock();
5581 
5582 	attrs = wq_sysfs_prep_attrs(wq);
5583 	if (!attrs)
5584 		goto out_unlock;
5585 
5586 	ret = cpumask_parse(buf, attrs->cpumask);
5587 	if (!ret)
5588 		ret = apply_workqueue_attrs_locked(wq, attrs);
5589 
5590 out_unlock:
5591 	apply_wqattrs_unlock();
5592 	free_workqueue_attrs(attrs);
5593 	return ret ?: count;
5594 }
5595 
wq_numa_show(struct device * dev,struct device_attribute * attr,char * buf)5596 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5597 			    char *buf)
5598 {
5599 	struct workqueue_struct *wq = dev_to_wq(dev);
5600 	int written;
5601 
5602 	mutex_lock(&wq->mutex);
5603 	written = scnprintf(buf, PAGE_SIZE, "%d\n",
5604 			    !wq->unbound_attrs->no_numa);
5605 	mutex_unlock(&wq->mutex);
5606 
5607 	return written;
5608 }
5609 
wq_numa_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)5610 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5611 			     const char *buf, size_t count)
5612 {
5613 	struct workqueue_struct *wq = dev_to_wq(dev);
5614 	struct workqueue_attrs *attrs;
5615 	int v, ret = -ENOMEM;
5616 
5617 	apply_wqattrs_lock();
5618 
5619 	attrs = wq_sysfs_prep_attrs(wq);
5620 	if (!attrs)
5621 		goto out_unlock;
5622 
5623 	ret = -EINVAL;
5624 	if (sscanf(buf, "%d", &v) == 1) {
5625 		attrs->no_numa = !v;
5626 		ret = apply_workqueue_attrs_locked(wq, attrs);
5627 	}
5628 
5629 out_unlock:
5630 	apply_wqattrs_unlock();
5631 	free_workqueue_attrs(attrs);
5632 	return ret ?: count;
5633 }
5634 
5635 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5636 	__ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5637 	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5638 	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5639 	__ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5640 	__ATTR_NULL,
5641 };
5642 
5643 static struct bus_type wq_subsys = {
5644 	.name				= "workqueue",
5645 	.dev_groups			= wq_sysfs_groups,
5646 };
5647 
wq_unbound_cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)5648 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5649 		struct device_attribute *attr, char *buf)
5650 {
5651 	int written;
5652 
5653 	mutex_lock(&wq_pool_mutex);
5654 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5655 			    cpumask_pr_args(wq_unbound_cpumask));
5656 	mutex_unlock(&wq_pool_mutex);
5657 
5658 	return written;
5659 }
5660 
wq_unbound_cpumask_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)5661 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5662 		struct device_attribute *attr, const char *buf, size_t count)
5663 {
5664 	cpumask_var_t cpumask;
5665 	int ret;
5666 
5667 	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5668 		return -ENOMEM;
5669 
5670 	ret = cpumask_parse(buf, cpumask);
5671 	if (!ret)
5672 		ret = workqueue_set_unbound_cpumask(cpumask);
5673 
5674 	free_cpumask_var(cpumask);
5675 	return ret ? ret : count;
5676 }
5677 
5678 static struct device_attribute wq_sysfs_cpumask_attr =
5679 	__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5680 	       wq_unbound_cpumask_store);
5681 
wq_sysfs_init(void)5682 static int __init wq_sysfs_init(void)
5683 {
5684 	int err;
5685 
5686 	err = subsys_virtual_register(&wq_subsys, NULL);
5687 	if (err)
5688 		return err;
5689 
5690 	return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5691 }
5692 core_initcall(wq_sysfs_init);
5693 
wq_device_release(struct device * dev)5694 static void wq_device_release(struct device *dev)
5695 {
5696 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5697 
5698 	kfree(wq_dev);
5699 }
5700 
5701 /**
5702  * workqueue_sysfs_register - make a workqueue visible in sysfs
5703  * @wq: the workqueue to register
5704  *
5705  * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5706  * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5707  * which is the preferred method.
5708  *
5709  * Workqueue user should use this function directly iff it wants to apply
5710  * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5711  * apply_workqueue_attrs() may race against userland updating the
5712  * attributes.
5713  *
5714  * Return: 0 on success, -errno on failure.
5715  */
workqueue_sysfs_register(struct workqueue_struct * wq)5716 int workqueue_sysfs_register(struct workqueue_struct *wq)
5717 {
5718 	struct wq_device *wq_dev;
5719 	int ret;
5720 
5721 	/*
5722 	 * Adjusting max_active or creating new pwqs by applying
5723 	 * attributes breaks ordering guarantee.  Disallow exposing ordered
5724 	 * workqueues.
5725 	 */
5726 	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5727 		return -EINVAL;
5728 
5729 	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5730 	if (!wq_dev)
5731 		return -ENOMEM;
5732 
5733 	wq_dev->wq = wq;
5734 	wq_dev->dev.bus = &wq_subsys;
5735 	wq_dev->dev.release = wq_device_release;
5736 	dev_set_name(&wq_dev->dev, "%s", wq->name);
5737 
5738 	/*
5739 	 * unbound_attrs are created separately.  Suppress uevent until
5740 	 * everything is ready.
5741 	 */
5742 	dev_set_uevent_suppress(&wq_dev->dev, true);
5743 
5744 	ret = device_register(&wq_dev->dev);
5745 	if (ret) {
5746 		put_device(&wq_dev->dev);
5747 		wq->wq_dev = NULL;
5748 		return ret;
5749 	}
5750 
5751 	if (wq->flags & WQ_UNBOUND) {
5752 		struct device_attribute *attr;
5753 
5754 		for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5755 			ret = device_create_file(&wq_dev->dev, attr);
5756 			if (ret) {
5757 				device_unregister(&wq_dev->dev);
5758 				wq->wq_dev = NULL;
5759 				return ret;
5760 			}
5761 		}
5762 	}
5763 
5764 	dev_set_uevent_suppress(&wq_dev->dev, false);
5765 	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5766 	return 0;
5767 }
5768 
5769 /**
5770  * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5771  * @wq: the workqueue to unregister
5772  *
5773  * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5774  */
workqueue_sysfs_unregister(struct workqueue_struct * wq)5775 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5776 {
5777 	struct wq_device *wq_dev = wq->wq_dev;
5778 
5779 	if (!wq->wq_dev)
5780 		return;
5781 
5782 	wq->wq_dev = NULL;
5783 	device_unregister(&wq_dev->dev);
5784 }
5785 #else	/* CONFIG_SYSFS */
workqueue_sysfs_unregister(struct workqueue_struct * wq)5786 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)	{ }
5787 #endif	/* CONFIG_SYSFS */
5788 
5789 /*
5790  * Workqueue watchdog.
5791  *
5792  * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5793  * flush dependency, a concurrency managed work item which stays RUNNING
5794  * indefinitely.  Workqueue stalls can be very difficult to debug as the
5795  * usual warning mechanisms don't trigger and internal workqueue state is
5796  * largely opaque.
5797  *
5798  * Workqueue watchdog monitors all worker pools periodically and dumps
5799  * state if some pools failed to make forward progress for a while where
5800  * forward progress is defined as the first item on ->worklist changing.
5801  *
5802  * This mechanism is controlled through the kernel parameter
5803  * "workqueue.watchdog_thresh" which can be updated at runtime through the
5804  * corresponding sysfs parameter file.
5805  */
5806 #ifdef CONFIG_WQ_WATCHDOG
5807 
5808 static unsigned long wq_watchdog_thresh = 30;
5809 static struct timer_list wq_watchdog_timer;
5810 
5811 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5812 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5813 
wq_watchdog_reset_touched(void)5814 static void wq_watchdog_reset_touched(void)
5815 {
5816 	int cpu;
5817 
5818 	wq_watchdog_touched = jiffies;
5819 	for_each_possible_cpu(cpu)
5820 		per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5821 }
5822 
wq_watchdog_timer_fn(struct timer_list * unused)5823 static void wq_watchdog_timer_fn(struct timer_list *unused)
5824 {
5825 	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5826 	bool lockup_detected = false;
5827 	unsigned long now = jiffies;
5828 	struct worker_pool *pool;
5829 	int pi;
5830 
5831 	if (!thresh)
5832 		return;
5833 
5834 	rcu_read_lock();
5835 
5836 	for_each_pool(pool, pi) {
5837 		unsigned long pool_ts, touched, ts;
5838 
5839 		if (list_empty(&pool->worklist))
5840 			continue;
5841 
5842 		/*
5843 		 * If a virtual machine is stopped by the host it can look to
5844 		 * the watchdog like a stall.
5845 		 */
5846 		kvm_check_and_clear_guest_paused();
5847 
5848 		/* get the latest of pool and touched timestamps */
5849 		if (pool->cpu >= 0)
5850 			touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
5851 		else
5852 			touched = READ_ONCE(wq_watchdog_touched);
5853 		pool_ts = READ_ONCE(pool->watchdog_ts);
5854 
5855 		if (time_after(pool_ts, touched))
5856 			ts = pool_ts;
5857 		else
5858 			ts = touched;
5859 
5860 		/* did we stall? */
5861 		if (time_after(now, ts + thresh)) {
5862 			lockup_detected = true;
5863 			pr_emerg("BUG: workqueue lockup - pool");
5864 			pr_cont_pool_info(pool);
5865 			pr_cont(" stuck for %us!\n",
5866 				jiffies_to_msecs(now - pool_ts) / 1000);
5867 		}
5868 	}
5869 
5870 	rcu_read_unlock();
5871 
5872 	if (lockup_detected)
5873 		show_all_workqueues();
5874 
5875 	wq_watchdog_reset_touched();
5876 	mod_timer(&wq_watchdog_timer, jiffies + thresh);
5877 }
5878 
wq_watchdog_touch(int cpu)5879 notrace void wq_watchdog_touch(int cpu)
5880 {
5881 	if (cpu >= 0)
5882 		per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5883 
5884 	wq_watchdog_touched = jiffies;
5885 }
5886 
wq_watchdog_set_thresh(unsigned long thresh)5887 static void wq_watchdog_set_thresh(unsigned long thresh)
5888 {
5889 	wq_watchdog_thresh = 0;
5890 	del_timer_sync(&wq_watchdog_timer);
5891 
5892 	if (thresh) {
5893 		wq_watchdog_thresh = thresh;
5894 		wq_watchdog_reset_touched();
5895 		mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5896 	}
5897 }
5898 
wq_watchdog_param_set_thresh(const char * val,const struct kernel_param * kp)5899 static int wq_watchdog_param_set_thresh(const char *val,
5900 					const struct kernel_param *kp)
5901 {
5902 	unsigned long thresh;
5903 	int ret;
5904 
5905 	ret = kstrtoul(val, 0, &thresh);
5906 	if (ret)
5907 		return ret;
5908 
5909 	if (system_wq)
5910 		wq_watchdog_set_thresh(thresh);
5911 	else
5912 		wq_watchdog_thresh = thresh;
5913 
5914 	return 0;
5915 }
5916 
5917 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5918 	.set	= wq_watchdog_param_set_thresh,
5919 	.get	= param_get_ulong,
5920 };
5921 
5922 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5923 		0644);
5924 
wq_watchdog_init(void)5925 static void wq_watchdog_init(void)
5926 {
5927 	timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5928 	wq_watchdog_set_thresh(wq_watchdog_thresh);
5929 }
5930 
5931 #else	/* CONFIG_WQ_WATCHDOG */
5932 
wq_watchdog_init(void)5933 static inline void wq_watchdog_init(void) { }
5934 
5935 #endif	/* CONFIG_WQ_WATCHDOG */
5936 
wq_numa_init(void)5937 static void __init wq_numa_init(void)
5938 {
5939 	cpumask_var_t *tbl;
5940 	int node, cpu;
5941 
5942 	if (num_possible_nodes() <= 1)
5943 		return;
5944 
5945 	if (wq_disable_numa) {
5946 		pr_info("workqueue: NUMA affinity support disabled\n");
5947 		return;
5948 	}
5949 
5950 	for_each_possible_cpu(cpu) {
5951 		if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) {
5952 			pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5953 			return;
5954 		}
5955 	}
5956 
5957 	wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
5958 	BUG_ON(!wq_update_unbound_numa_attrs_buf);
5959 
5960 	/*
5961 	 * We want masks of possible CPUs of each node which isn't readily
5962 	 * available.  Build one from cpu_to_node() which should have been
5963 	 * fully initialized by now.
5964 	 */
5965 	tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5966 	BUG_ON(!tbl);
5967 
5968 	for_each_node(node)
5969 		BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5970 				node_online(node) ? node : NUMA_NO_NODE));
5971 
5972 	for_each_possible_cpu(cpu) {
5973 		node = cpu_to_node(cpu);
5974 		cpumask_set_cpu(cpu, tbl[node]);
5975 	}
5976 
5977 	wq_numa_possible_cpumask = tbl;
5978 	wq_numa_enabled = true;
5979 }
5980 
5981 /**
5982  * workqueue_init_early - early init for workqueue subsystem
5983  *
5984  * This is the first half of two-staged workqueue subsystem initialization
5985  * and invoked as soon as the bare basics - memory allocation, cpumasks and
5986  * idr are up.  It sets up all the data structures and system workqueues
5987  * and allows early boot code to create workqueues and queue/cancel work
5988  * items.  Actual work item execution starts only after kthreads can be
5989  * created and scheduled right before early initcalls.
5990  */
workqueue_init_early(void)5991 void __init workqueue_init_early(void)
5992 {
5993 	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5994 	int i, cpu;
5995 
5996 	BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5997 
5998 	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5999 	cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ));
6000 	cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN));
6001 
6002 	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
6003 
6004 	/* initialize CPU pools */
6005 	for_each_possible_cpu(cpu) {
6006 		struct worker_pool *pool;
6007 
6008 		i = 0;
6009 		for_each_cpu_worker_pool(pool, cpu) {
6010 			BUG_ON(init_worker_pool(pool));
6011 			pool->cpu = cpu;
6012 			cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
6013 			pool->attrs->nice = std_nice[i++];
6014 			pool->node = cpu_to_node(cpu);
6015 
6016 			/* alloc pool ID */
6017 			mutex_lock(&wq_pool_mutex);
6018 			BUG_ON(worker_pool_assign_id(pool));
6019 			mutex_unlock(&wq_pool_mutex);
6020 		}
6021 	}
6022 
6023 	/* create default unbound and ordered wq attrs */
6024 	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
6025 		struct workqueue_attrs *attrs;
6026 
6027 		BUG_ON(!(attrs = alloc_workqueue_attrs()));
6028 		attrs->nice = std_nice[i];
6029 		unbound_std_wq_attrs[i] = attrs;
6030 
6031 		/*
6032 		 * An ordered wq should have only one pwq as ordering is
6033 		 * guaranteed by max_active which is enforced by pwqs.
6034 		 * Turn off NUMA so that dfl_pwq is used for all nodes.
6035 		 */
6036 		BUG_ON(!(attrs = alloc_workqueue_attrs()));
6037 		attrs->nice = std_nice[i];
6038 		attrs->no_numa = true;
6039 		ordered_wq_attrs[i] = attrs;
6040 	}
6041 
6042 	system_wq = alloc_workqueue("events", 0, 0);
6043 	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
6044 	system_long_wq = alloc_workqueue("events_long", 0, 0);
6045 	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
6046 					    WQ_UNBOUND_MAX_ACTIVE);
6047 	system_freezable_wq = alloc_workqueue("events_freezable",
6048 					      WQ_FREEZABLE, 0);
6049 	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
6050 					      WQ_POWER_EFFICIENT, 0);
6051 	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
6052 					      WQ_FREEZABLE | WQ_POWER_EFFICIENT,
6053 					      0);
6054 	BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
6055 	       !system_unbound_wq || !system_freezable_wq ||
6056 	       !system_power_efficient_wq ||
6057 	       !system_freezable_power_efficient_wq);
6058 }
6059 
6060 /**
6061  * workqueue_init - bring workqueue subsystem fully online
6062  *
6063  * This is the latter half of two-staged workqueue subsystem initialization
6064  * and invoked as soon as kthreads can be created and scheduled.
6065  * Workqueues have been created and work items queued on them, but there
6066  * are no kworkers executing the work items yet.  Populate the worker pools
6067  * with the initial workers and enable future kworker creations.
6068  */
workqueue_init(void)6069 void __init workqueue_init(void)
6070 {
6071 	struct workqueue_struct *wq;
6072 	struct worker_pool *pool;
6073 	int cpu, bkt;
6074 
6075 	/*
6076 	 * It'd be simpler to initialize NUMA in workqueue_init_early() but
6077 	 * CPU to node mapping may not be available that early on some
6078 	 * archs such as power and arm64.  As per-cpu pools created
6079 	 * previously could be missing node hint and unbound pools NUMA
6080 	 * affinity, fix them up.
6081 	 *
6082 	 * Also, while iterating workqueues, create rescuers if requested.
6083 	 */
6084 	wq_numa_init();
6085 
6086 	mutex_lock(&wq_pool_mutex);
6087 
6088 	for_each_possible_cpu(cpu) {
6089 		for_each_cpu_worker_pool(pool, cpu) {
6090 			pool->node = cpu_to_node(cpu);
6091 		}
6092 	}
6093 
6094 	list_for_each_entry(wq, &workqueues, list) {
6095 		wq_update_unbound_numa(wq, smp_processor_id(), true);
6096 		WARN(init_rescuer(wq),
6097 		     "workqueue: failed to create early rescuer for %s",
6098 		     wq->name);
6099 	}
6100 
6101 	mutex_unlock(&wq_pool_mutex);
6102 
6103 	/* create the initial workers */
6104 	for_each_online_cpu(cpu) {
6105 		for_each_cpu_worker_pool(pool, cpu) {
6106 			pool->flags &= ~POOL_DISASSOCIATED;
6107 			BUG_ON(!create_worker(pool));
6108 		}
6109 	}
6110 
6111 	hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6112 		BUG_ON(!create_worker(pool));
6113 
6114 	wq_online = true;
6115 	wq_watchdog_init();
6116 }
6117 
6118 /*
6119  * Despite the naming, this is a no-op function which is here only for avoiding
6120  * link error. Since compile-time warning may fail to catch, we will need to
6121  * emit run-time warning from __flush_workqueue().
6122  */
__warn_flushing_systemwide_wq(void)6123 void __warn_flushing_systemwide_wq(void) { }
6124 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
6125