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