1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * kernel/workqueue.c - generic async execution with shared worker pool
4  *
5  * Copyright (C) 2002		Ingo Molnar
6  *
7  *   Derived from the taskqueue/keventd code by:
8  *     David Woodhouse <dwmw2@infradead.org>
9  *     Andrew Morton
10  *     Kai Petzke <wpp@marie.physik.tu-berlin.de>
11  *     Theodore Ts'o <tytso@mit.edu>
12  *
13  * Made to use alloc_percpu by Christoph Lameter.
14  *
15  * Copyright (C) 2010		SUSE Linux Products GmbH
16  * Copyright (C) 2010		Tejun Heo <tj@kernel.org>
17  *
18  * This is the generic async execution mechanism.  Work items as are
19  * executed in process context.  The worker pool is shared and
20  * automatically managed.  There are two worker pools for each CPU (one for
21  * normal work items and the other for high priority ones) and some extra
22  * pools for workqueues which are not bound to any specific CPU - the
23  * number of these backing pools is dynamic.
24  *
25  * Please read Documentation/core-api/workqueue.rst for details.
26  */
27 
28 #include <linux/export.h>
29 #include <linux/kernel.h>
30 #include <linux/sched.h>
31 #include <linux/init.h>
32 #include <linux/signal.h>
33 #include <linux/completion.h>
34 #include <linux/workqueue.h>
35 #include <linux/slab.h>
36 #include <linux/cpu.h>
37 #include <linux/notifier.h>
38 #include <linux/kthread.h>
39 #include <linux/hardirq.h>
40 #include <linux/mempolicy.h>
41 #include <linux/freezer.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 #include <linux/sched/isolation.h>
52 #include <linux/sched/debug.h>
53 #include <linux/nmi.h>
54 #include <linux/kvm_para.h>
55 #include <linux/delay.h>
56 
57 #include "workqueue_internal.h"
58 
59 enum {
60 	/*
61 	 * worker_pool flags
62 	 *
63 	 * A bound pool is either associated or disassociated with its CPU.
64 	 * While associated (!DISASSOCIATED), all workers are bound to the
65 	 * CPU and none has %WORKER_UNBOUND set and concurrency management
66 	 * is in effect.
67 	 *
68 	 * While DISASSOCIATED, the cpu may be offline and all workers have
69 	 * %WORKER_UNBOUND set and concurrency management disabled, and may
70 	 * be executing on any CPU.  The pool behaves as an unbound one.
71 	 *
72 	 * Note that DISASSOCIATED should be flipped only while holding
73 	 * wq_pool_attach_mutex to avoid changing binding state while
74 	 * worker_attach_to_pool() is in progress.
75 	 */
76 	POOL_MANAGER_ACTIVE	= 1 << 0,	/* being managed */
77 	POOL_DISASSOCIATED	= 1 << 2,	/* cpu can't serve workers */
78 
79 	/* worker flags */
80 	WORKER_DIE		= 1 << 1,	/* die die die */
81 	WORKER_IDLE		= 1 << 2,	/* is idle */
82 	WORKER_PREP		= 1 << 3,	/* preparing to run works */
83 	WORKER_CPU_INTENSIVE	= 1 << 6,	/* cpu intensive */
84 	WORKER_UNBOUND		= 1 << 7,	/* worker is unbound */
85 	WORKER_REBOUND		= 1 << 8,	/* worker was rebound */
86 
87 	WORKER_NOT_RUNNING	= WORKER_PREP | WORKER_CPU_INTENSIVE |
88 				  WORKER_UNBOUND | WORKER_REBOUND,
89 
90 	NR_STD_WORKER_POOLS	= 2,		/* # standard pools per cpu */
91 
92 	UNBOUND_POOL_HASH_ORDER	= 6,		/* hashed by pool->attrs */
93 	BUSY_WORKER_HASH_ORDER	= 6,		/* 64 pointers */
94 
95 	MAX_IDLE_WORKERS_RATIO	= 4,		/* 1/4 of busy can be idle */
96 	IDLE_WORKER_TIMEOUT	= 300 * HZ,	/* keep idle ones for 5 mins */
97 
98 	MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
99 						/* call for help after 10ms
100 						   (min two ticks) */
101 	MAYDAY_INTERVAL		= HZ / 10,	/* and then every 100ms */
102 	CREATE_COOLDOWN		= HZ,		/* time to breath after fail */
103 
104 	/*
105 	 * Rescue workers are used only on emergencies and shared by
106 	 * all cpus.  Give MIN_NICE.
107 	 */
108 	RESCUER_NICE_LEVEL	= MIN_NICE,
109 	HIGHPRI_NICE_LEVEL	= MIN_NICE,
110 
111 	WQ_NAME_LEN		= 24,
112 };
113 
114 /*
115  * Structure fields follow one of the following exclusion rules.
116  *
117  * I: Modifiable by initialization/destruction paths and read-only for
118  *    everyone else.
119  *
120  * P: Preemption protected.  Disabling preemption is enough and should
121  *    only be modified and accessed from the local cpu.
122  *
123  * L: pool->lock protected.  Access with pool->lock held.
124  *
125  * K: Only modified by worker while holding pool->lock. Can be safely read by
126  *    self, while holding pool->lock or from IRQ context if %current is the
127  *    kworker.
128  *
129  * S: Only modified by worker self.
130  *
131  * A: wq_pool_attach_mutex protected.
132  *
133  * PL: wq_pool_mutex protected.
134  *
135  * PR: wq_pool_mutex protected for writes.  RCU protected for reads.
136  *
137  * PW: wq_pool_mutex and wq->mutex protected for writes.  Either for reads.
138  *
139  * PWR: wq_pool_mutex and wq->mutex protected for writes.  Either or
140  *      RCU for reads.
141  *
142  * WQ: wq->mutex protected.
143  *
144  * WR: wq->mutex protected for writes.  RCU protected for reads.
145  *
146  * MD: wq_mayday_lock protected.
147  *
148  * WD: Used internally by the watchdog.
149  */
150 
151 /* struct worker is defined in workqueue_internal.h */
152 
153 struct worker_pool {
154 	raw_spinlock_t		lock;		/* the pool lock */
155 	int			cpu;		/* I: the associated cpu */
156 	int			node;		/* I: the associated node ID */
157 	int			id;		/* I: pool ID */
158 	unsigned int		flags;		/* L: flags */
159 
160 	unsigned long		watchdog_ts;	/* L: watchdog timestamp */
161 	bool			cpu_stall;	/* WD: stalled cpu bound pool */
162 
163 	/*
164 	 * The counter is incremented in a process context on the associated CPU
165 	 * w/ preemption disabled, and decremented or reset in the same context
166 	 * but w/ pool->lock held. The readers grab pool->lock and are
167 	 * guaranteed to see if the counter reached zero.
168 	 */
169 	int			nr_running;
170 
171 	struct list_head	worklist;	/* L: list of pending works */
172 
173 	int			nr_workers;	/* L: total number of workers */
174 	int			nr_idle;	/* L: currently idle workers */
175 
176 	struct list_head	idle_list;	/* L: list of idle workers */
177 	struct timer_list	idle_timer;	/* L: worker idle timeout */
178 	struct work_struct      idle_cull_work; /* L: worker idle cleanup */
179 
180 	struct timer_list	mayday_timer;	  /* L: SOS timer for workers */
181 
182 	/* a workers is either on busy_hash or idle_list, or the manager */
183 	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
184 						/* L: hash of busy workers */
185 
186 	struct worker		*manager;	/* L: purely informational */
187 	struct list_head	workers;	/* A: attached workers */
188 	struct list_head        dying_workers;  /* A: workers about to die */
189 	struct completion	*detach_completion; /* all workers detached */
190 
191 	struct ida		worker_ida;	/* worker IDs for task name */
192 
193 	struct workqueue_attrs	*attrs;		/* I: worker attributes */
194 	struct hlist_node	hash_node;	/* PL: unbound_pool_hash node */
195 	int			refcnt;		/* PL: refcnt for unbound pools */
196 
197 	/*
198 	 * Destruction of pool is RCU protected to allow dereferences
199 	 * from get_work_pool().
200 	 */
201 	struct rcu_head		rcu;
202 };
203 
204 /*
205  * Per-pool_workqueue statistics. These can be monitored using
206  * tools/workqueue/wq_monitor.py.
207  */
208 enum pool_workqueue_stats {
209 	PWQ_STAT_STARTED,	/* work items started execution */
210 	PWQ_STAT_COMPLETED,	/* work items completed execution */
211 	PWQ_STAT_CPU_TIME,	/* total CPU time consumed */
212 	PWQ_STAT_CPU_INTENSIVE,	/* wq_cpu_intensive_thresh_us violations */
213 	PWQ_STAT_CM_WAKEUP,	/* concurrency-management worker wakeups */
214 	PWQ_STAT_REPATRIATED,	/* unbound workers brought back into scope */
215 	PWQ_STAT_MAYDAY,	/* maydays to rescuer */
216 	PWQ_STAT_RESCUED,	/* linked work items executed by rescuer */
217 
218 	PWQ_NR_STATS,
219 };
220 
221 /*
222  * The per-pool workqueue.  While queued, the lower WORK_STRUCT_FLAG_BITS
223  * of work_struct->data are used for flags and the remaining high bits
224  * point to the pwq; thus, pwqs need to be aligned at two's power of the
225  * number of flag bits.
226  */
227 struct pool_workqueue {
228 	struct worker_pool	*pool;		/* I: the associated pool */
229 	struct workqueue_struct *wq;		/* I: the owning workqueue */
230 	int			work_color;	/* L: current color */
231 	int			flush_color;	/* L: flushing color */
232 	int			refcnt;		/* L: reference count */
233 	int			nr_in_flight[WORK_NR_COLORS];
234 						/* L: nr of in_flight works */
235 
236 	/*
237 	 * nr_active management and WORK_STRUCT_INACTIVE:
238 	 *
239 	 * When pwq->nr_active >= max_active, new work item is queued to
240 	 * pwq->inactive_works instead of pool->worklist and marked with
241 	 * WORK_STRUCT_INACTIVE.
242 	 *
243 	 * All work items marked with WORK_STRUCT_INACTIVE do not participate
244 	 * in pwq->nr_active and all work items in pwq->inactive_works are
245 	 * marked with WORK_STRUCT_INACTIVE.  But not all WORK_STRUCT_INACTIVE
246 	 * work items are in pwq->inactive_works.  Some of them are ready to
247 	 * run in pool->worklist or worker->scheduled.  Those work itmes are
248 	 * only struct wq_barrier which is used for flush_work() and should
249 	 * not participate in pwq->nr_active.  For non-barrier work item, it
250 	 * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
251 	 */
252 	int			nr_active;	/* L: nr of active works */
253 	int			max_active;	/* L: max active works */
254 	struct list_head	inactive_works;	/* L: inactive works */
255 	struct list_head	pwqs_node;	/* WR: node on wq->pwqs */
256 	struct list_head	mayday_node;	/* MD: node on wq->maydays */
257 
258 	u64			stats[PWQ_NR_STATS];
259 
260 	/*
261 	 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
262 	 * and pwq_release_workfn() for details. pool_workqueue itself is also
263 	 * RCU protected so that the first pwq can be determined without
264 	 * grabbing wq->mutex.
265 	 */
266 	struct kthread_work	release_work;
267 	struct rcu_head		rcu;
268 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
269 
270 /*
271  * Structure used to wait for workqueue flush.
272  */
273 struct wq_flusher {
274 	struct list_head	list;		/* WQ: list of flushers */
275 	int			flush_color;	/* WQ: flush color waiting for */
276 	struct completion	done;		/* flush completion */
277 };
278 
279 struct wq_device;
280 
281 /*
282  * The externally visible workqueue.  It relays the issued work items to
283  * the appropriate worker_pool through its pool_workqueues.
284  */
285 struct workqueue_struct {
286 	struct list_head	pwqs;		/* WR: all pwqs of this wq */
287 	struct list_head	list;		/* PR: list of all workqueues */
288 
289 	struct mutex		mutex;		/* protects this wq */
290 	int			work_color;	/* WQ: current work color */
291 	int			flush_color;	/* WQ: current flush color */
292 	atomic_t		nr_pwqs_to_flush; /* flush in progress */
293 	struct wq_flusher	*first_flusher;	/* WQ: first flusher */
294 	struct list_head	flusher_queue;	/* WQ: flush waiters */
295 	struct list_head	flusher_overflow; /* WQ: flush overflow list */
296 
297 	struct list_head	maydays;	/* MD: pwqs requesting rescue */
298 	struct worker		*rescuer;	/* MD: rescue worker */
299 
300 	int			nr_drainers;	/* WQ: drain in progress */
301 	int			saved_max_active; /* WQ: saved pwq max_active */
302 
303 	struct workqueue_attrs	*unbound_attrs;	/* PW: only for unbound wqs */
304 	struct pool_workqueue	*dfl_pwq;	/* PW: only for unbound wqs */
305 
306 #ifdef CONFIG_SYSFS
307 	struct wq_device	*wq_dev;	/* I: for sysfs interface */
308 #endif
309 #ifdef CONFIG_LOCKDEP
310 	char			*lock_name;
311 	struct lock_class_key	key;
312 	struct lockdep_map	lockdep_map;
313 #endif
314 	char			name[WQ_NAME_LEN]; /* I: workqueue name */
315 
316 	/*
317 	 * Destruction of workqueue_struct is RCU protected to allow walking
318 	 * the workqueues list without grabbing wq_pool_mutex.
319 	 * This is used to dump all workqueues from sysrq.
320 	 */
321 	struct rcu_head		rcu;
322 
323 	/* hot fields used during command issue, aligned to cacheline */
324 	unsigned int		flags ____cacheline_aligned; /* WQ: WQ_* flags */
325 	struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
326 };
327 
328 static struct kmem_cache *pwq_cache;
329 
330 /*
331  * Each pod type describes how CPUs should be grouped for unbound workqueues.
332  * See the comment above workqueue_attrs->affn_scope.
333  */
334 struct wq_pod_type {
335 	int			nr_pods;	/* number of pods */
336 	cpumask_var_t		*pod_cpus;	/* pod -> cpus */
337 	int			*pod_node;	/* pod -> node */
338 	int			*cpu_pod;	/* cpu -> pod */
339 };
340 
341 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
342 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
343 
344 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
345 	[WQ_AFFN_DFL]			= "default",
346 	[WQ_AFFN_CPU]			= "cpu",
347 	[WQ_AFFN_SMT]			= "smt",
348 	[WQ_AFFN_CACHE]			= "cache",
349 	[WQ_AFFN_NUMA]			= "numa",
350 	[WQ_AFFN_SYSTEM]		= "system",
351 };
352 
353 /*
354  * Per-cpu work items which run for longer than the following threshold are
355  * automatically considered CPU intensive and excluded from concurrency
356  * management to prevent them from noticeably delaying other per-cpu work items.
357  * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
358  * The actual value is initialized in wq_cpu_intensive_thresh_init().
359  */
360 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
361 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
362 
363 /* see the comment above the definition of WQ_POWER_EFFICIENT */
364 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
365 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
366 
367 static bool wq_online;			/* can kworkers be created yet? */
368 
369 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
370 static struct workqueue_attrs *wq_update_pod_attrs_buf;
371 
372 static DEFINE_MUTEX(wq_pool_mutex);	/* protects pools and workqueues list */
373 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
374 static DEFINE_RAW_SPINLOCK(wq_mayday_lock);	/* protects wq->maydays list */
375 /* wait for manager to go away */
376 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
377 
378 static LIST_HEAD(workqueues);		/* PR: list of all workqueues */
379 static bool workqueue_freezing;		/* PL: have wqs started freezing? */
380 
381 /* PL&A: allowable cpus for unbound wqs and work items */
382 static cpumask_var_t wq_unbound_cpumask;
383 
384 /* for further constrain wq_unbound_cpumask by cmdline parameter*/
385 static struct cpumask wq_cmdline_cpumask __initdata;
386 
387 /* CPU where unbound work was last round robin scheduled from this CPU */
388 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
389 
390 /*
391  * Local execution of unbound work items is no longer guaranteed.  The
392  * following always forces round-robin CPU selection on unbound work items
393  * to uncover usages which depend on it.
394  */
395 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
396 static bool wq_debug_force_rr_cpu = true;
397 #else
398 static bool wq_debug_force_rr_cpu = false;
399 #endif
400 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
401 
402 /* the per-cpu worker pools */
403 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
404 
405 static DEFINE_IDR(worker_pool_idr);	/* PR: idr of all pools */
406 
407 /* PL: hash of all unbound pools keyed by pool->attrs */
408 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
409 
410 /* I: attributes used when instantiating standard unbound pools on demand */
411 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
412 
413 /* I: attributes used when instantiating ordered pools on demand */
414 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
415 
416 /*
417  * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
418  * process context while holding a pool lock. Bounce to a dedicated kthread
419  * worker to avoid A-A deadlocks.
420  */
421 static struct kthread_worker *pwq_release_worker;
422 
423 struct workqueue_struct *system_wq __read_mostly;
424 EXPORT_SYMBOL(system_wq);
425 struct workqueue_struct *system_highpri_wq __read_mostly;
426 EXPORT_SYMBOL_GPL(system_highpri_wq);
427 struct workqueue_struct *system_long_wq __read_mostly;
428 EXPORT_SYMBOL_GPL(system_long_wq);
429 struct workqueue_struct *system_unbound_wq __read_mostly;
430 EXPORT_SYMBOL_GPL(system_unbound_wq);
431 struct workqueue_struct *system_freezable_wq __read_mostly;
432 EXPORT_SYMBOL_GPL(system_freezable_wq);
433 struct workqueue_struct *system_power_efficient_wq __read_mostly;
434 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
435 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
436 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
437 
438 static int worker_thread(void *__worker);
439 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
440 static void show_pwq(struct pool_workqueue *pwq);
441 static void show_one_worker_pool(struct worker_pool *pool);
442 
443 #define CREATE_TRACE_POINTS
444 #include <trace/events/workqueue.h>
445 
446 #define assert_rcu_or_pool_mutex()					\
447 	RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&			\
448 			 !lockdep_is_held(&wq_pool_mutex),		\
449 			 "RCU or wq_pool_mutex should be held")
450 
451 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq)			\
452 	RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&			\
453 			 !lockdep_is_held(&wq->mutex) &&		\
454 			 !lockdep_is_held(&wq_pool_mutex),		\
455 			 "RCU, wq->mutex or wq_pool_mutex should be held")
456 
457 #define for_each_cpu_worker_pool(pool, cpu)				\
458 	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0];		\
459 	     (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
460 	     (pool)++)
461 
462 /**
463  * for_each_pool - iterate through all worker_pools in the system
464  * @pool: iteration cursor
465  * @pi: integer used for iteration
466  *
467  * This must be called either with wq_pool_mutex held or RCU read
468  * locked.  If the pool needs to be used beyond the locking in effect, the
469  * caller is responsible for guaranteeing that the pool stays online.
470  *
471  * The if/else clause exists only for the lockdep assertion and can be
472  * ignored.
473  */
474 #define for_each_pool(pool, pi)						\
475 	idr_for_each_entry(&worker_pool_idr, pool, pi)			\
476 		if (({ assert_rcu_or_pool_mutex(); false; })) { }	\
477 		else
478 
479 /**
480  * for_each_pool_worker - iterate through all workers of a worker_pool
481  * @worker: iteration cursor
482  * @pool: worker_pool to iterate workers of
483  *
484  * This must be called with wq_pool_attach_mutex.
485  *
486  * The if/else clause exists only for the lockdep assertion and can be
487  * ignored.
488  */
489 #define for_each_pool_worker(worker, pool)				\
490 	list_for_each_entry((worker), &(pool)->workers, node)		\
491 		if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
492 		else
493 
494 /**
495  * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
496  * @pwq: iteration cursor
497  * @wq: the target workqueue
498  *
499  * This must be called either with wq->mutex held or RCU read locked.
500  * If the pwq needs to be used beyond the locking in effect, the caller is
501  * responsible for guaranteeing that the pwq stays online.
502  *
503  * The if/else clause exists only for the lockdep assertion and can be
504  * ignored.
505  */
506 #define for_each_pwq(pwq, wq)						\
507 	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node,		\
508 				 lockdep_is_held(&(wq->mutex)))
509 
510 #ifdef CONFIG_DEBUG_OBJECTS_WORK
511 
512 static const struct debug_obj_descr work_debug_descr;
513 
work_debug_hint(void * addr)514 static void *work_debug_hint(void *addr)
515 {
516 	return ((struct work_struct *) addr)->func;
517 }
518 
work_is_static_object(void * addr)519 static bool work_is_static_object(void *addr)
520 {
521 	struct work_struct *work = addr;
522 
523 	return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
524 }
525 
526 /*
527  * fixup_init is called when:
528  * - an active object is initialized
529  */
work_fixup_init(void * addr,enum debug_obj_state state)530 static bool work_fixup_init(void *addr, enum debug_obj_state state)
531 {
532 	struct work_struct *work = addr;
533 
534 	switch (state) {
535 	case ODEBUG_STATE_ACTIVE:
536 		cancel_work_sync(work);
537 		debug_object_init(work, &work_debug_descr);
538 		return true;
539 	default:
540 		return false;
541 	}
542 }
543 
544 /*
545  * fixup_free is called when:
546  * - an active object is freed
547  */
work_fixup_free(void * addr,enum debug_obj_state state)548 static bool work_fixup_free(void *addr, enum debug_obj_state state)
549 {
550 	struct work_struct *work = addr;
551 
552 	switch (state) {
553 	case ODEBUG_STATE_ACTIVE:
554 		cancel_work_sync(work);
555 		debug_object_free(work, &work_debug_descr);
556 		return true;
557 	default:
558 		return false;
559 	}
560 }
561 
562 static const struct debug_obj_descr work_debug_descr = {
563 	.name		= "work_struct",
564 	.debug_hint	= work_debug_hint,
565 	.is_static_object = work_is_static_object,
566 	.fixup_init	= work_fixup_init,
567 	.fixup_free	= work_fixup_free,
568 };
569 
debug_work_activate(struct work_struct * work)570 static inline void debug_work_activate(struct work_struct *work)
571 {
572 	debug_object_activate(work, &work_debug_descr);
573 }
574 
debug_work_deactivate(struct work_struct * work)575 static inline void debug_work_deactivate(struct work_struct *work)
576 {
577 	debug_object_deactivate(work, &work_debug_descr);
578 }
579 
__init_work(struct work_struct * work,int onstack)580 void __init_work(struct work_struct *work, int onstack)
581 {
582 	if (onstack)
583 		debug_object_init_on_stack(work, &work_debug_descr);
584 	else
585 		debug_object_init(work, &work_debug_descr);
586 }
587 EXPORT_SYMBOL_GPL(__init_work);
588 
destroy_work_on_stack(struct work_struct * work)589 void destroy_work_on_stack(struct work_struct *work)
590 {
591 	debug_object_free(work, &work_debug_descr);
592 }
593 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
594 
destroy_delayed_work_on_stack(struct delayed_work * work)595 void destroy_delayed_work_on_stack(struct delayed_work *work)
596 {
597 	destroy_timer_on_stack(&work->timer);
598 	debug_object_free(&work->work, &work_debug_descr);
599 }
600 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
601 
602 #else
debug_work_activate(struct work_struct * work)603 static inline void debug_work_activate(struct work_struct *work) { }
debug_work_deactivate(struct work_struct * work)604 static inline void debug_work_deactivate(struct work_struct *work) { }
605 #endif
606 
607 /**
608  * worker_pool_assign_id - allocate ID and assign it to @pool
609  * @pool: the pool pointer of interest
610  *
611  * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
612  * successfully, -errno on failure.
613  */
worker_pool_assign_id(struct worker_pool * pool)614 static int worker_pool_assign_id(struct worker_pool *pool)
615 {
616 	int ret;
617 
618 	lockdep_assert_held(&wq_pool_mutex);
619 
620 	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
621 			GFP_KERNEL);
622 	if (ret >= 0) {
623 		pool->id = ret;
624 		return 0;
625 	}
626 	return ret;
627 }
628 
work_color_to_flags(int color)629 static unsigned int work_color_to_flags(int color)
630 {
631 	return color << WORK_STRUCT_COLOR_SHIFT;
632 }
633 
get_work_color(unsigned long work_data)634 static int get_work_color(unsigned long work_data)
635 {
636 	return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
637 		((1 << WORK_STRUCT_COLOR_BITS) - 1);
638 }
639 
work_next_color(int color)640 static int work_next_color(int color)
641 {
642 	return (color + 1) % WORK_NR_COLORS;
643 }
644 
645 /*
646  * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
647  * contain the pointer to the queued pwq.  Once execution starts, the flag
648  * is cleared and the high bits contain OFFQ flags and pool ID.
649  *
650  * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
651  * and clear_work_data() can be used to set the pwq, pool or clear
652  * work->data.  These functions should only be called while the work is
653  * owned - ie. while the PENDING bit is set.
654  *
655  * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
656  * corresponding to a work.  Pool is available once the work has been
657  * queued anywhere after initialization until it is sync canceled.  pwq is
658  * available only while the work item is queued.
659  *
660  * %WORK_OFFQ_CANCELING is used to mark a work item which is being
661  * canceled.  While being canceled, a work item may have its PENDING set
662  * but stay off timer and worklist for arbitrarily long and nobody should
663  * try to steal the PENDING bit.
664  */
set_work_data(struct work_struct * work,unsigned long data,unsigned long flags)665 static inline void set_work_data(struct work_struct *work, unsigned long data,
666 				 unsigned long flags)
667 {
668 	WARN_ON_ONCE(!work_pending(work));
669 	atomic_long_set(&work->data, data | flags | work_static(work));
670 }
671 
set_work_pwq(struct work_struct * work,struct pool_workqueue * pwq,unsigned long extra_flags)672 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
673 			 unsigned long extra_flags)
674 {
675 	set_work_data(work, (unsigned long)pwq,
676 		      WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
677 }
678 
set_work_pool_and_keep_pending(struct work_struct * work,int pool_id)679 static void set_work_pool_and_keep_pending(struct work_struct *work,
680 					   int pool_id)
681 {
682 	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
683 		      WORK_STRUCT_PENDING);
684 }
685 
set_work_pool_and_clear_pending(struct work_struct * work,int pool_id)686 static void set_work_pool_and_clear_pending(struct work_struct *work,
687 					    int pool_id)
688 {
689 	/*
690 	 * The following wmb is paired with the implied mb in
691 	 * test_and_set_bit(PENDING) and ensures all updates to @work made
692 	 * here are visible to and precede any updates by the next PENDING
693 	 * owner.
694 	 */
695 	smp_wmb();
696 	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
697 	/*
698 	 * The following mb guarantees that previous clear of a PENDING bit
699 	 * will not be reordered with any speculative LOADS or STORES from
700 	 * work->current_func, which is executed afterwards.  This possible
701 	 * reordering can lead to a missed execution on attempt to queue
702 	 * the same @work.  E.g. consider this case:
703 	 *
704 	 *   CPU#0                         CPU#1
705 	 *   ----------------------------  --------------------------------
706 	 *
707 	 * 1  STORE event_indicated
708 	 * 2  queue_work_on() {
709 	 * 3    test_and_set_bit(PENDING)
710 	 * 4 }                             set_..._and_clear_pending() {
711 	 * 5                                 set_work_data() # clear bit
712 	 * 6                                 smp_mb()
713 	 * 7                               work->current_func() {
714 	 * 8				      LOAD event_indicated
715 	 *				   }
716 	 *
717 	 * Without an explicit full barrier speculative LOAD on line 8 can
718 	 * be executed before CPU#0 does STORE on line 1.  If that happens,
719 	 * CPU#0 observes the PENDING bit is still set and new execution of
720 	 * a @work is not queued in a hope, that CPU#1 will eventually
721 	 * finish the queued @work.  Meanwhile CPU#1 does not see
722 	 * event_indicated is set, because speculative LOAD was executed
723 	 * before actual STORE.
724 	 */
725 	smp_mb();
726 }
727 
clear_work_data(struct work_struct * work)728 static void clear_work_data(struct work_struct *work)
729 {
730 	smp_wmb();	/* see set_work_pool_and_clear_pending() */
731 	set_work_data(work, WORK_STRUCT_NO_POOL, 0);
732 }
733 
work_struct_pwq(unsigned long data)734 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
735 {
736 	return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
737 }
738 
get_work_pwq(struct work_struct * work)739 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
740 {
741 	unsigned long data = atomic_long_read(&work->data);
742 
743 	if (data & WORK_STRUCT_PWQ)
744 		return work_struct_pwq(data);
745 	else
746 		return NULL;
747 }
748 
749 /**
750  * get_work_pool - return the worker_pool a given work was associated with
751  * @work: the work item of interest
752  *
753  * Pools are created and destroyed under wq_pool_mutex, and allows read
754  * access under RCU read lock.  As such, this function should be
755  * called under wq_pool_mutex or inside of a rcu_read_lock() region.
756  *
757  * All fields of the returned pool are accessible as long as the above
758  * mentioned locking is in effect.  If the returned pool needs to be used
759  * beyond the critical section, the caller is responsible for ensuring the
760  * returned pool is and stays online.
761  *
762  * Return: The worker_pool @work was last associated with.  %NULL if none.
763  */
get_work_pool(struct work_struct * work)764 static struct worker_pool *get_work_pool(struct work_struct *work)
765 {
766 	unsigned long data = atomic_long_read(&work->data);
767 	int pool_id;
768 
769 	assert_rcu_or_pool_mutex();
770 
771 	if (data & WORK_STRUCT_PWQ)
772 		return work_struct_pwq(data)->pool;
773 
774 	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
775 	if (pool_id == WORK_OFFQ_POOL_NONE)
776 		return NULL;
777 
778 	return idr_find(&worker_pool_idr, pool_id);
779 }
780 
781 /**
782  * get_work_pool_id - return the worker pool ID a given work is associated with
783  * @work: the work item of interest
784  *
785  * Return: The worker_pool ID @work was last associated with.
786  * %WORK_OFFQ_POOL_NONE if none.
787  */
get_work_pool_id(struct work_struct * work)788 static int get_work_pool_id(struct work_struct *work)
789 {
790 	unsigned long data = atomic_long_read(&work->data);
791 
792 	if (data & WORK_STRUCT_PWQ)
793 		return work_struct_pwq(data)->pool->id;
794 
795 	return data >> WORK_OFFQ_POOL_SHIFT;
796 }
797 
mark_work_canceling(struct work_struct * work)798 static void mark_work_canceling(struct work_struct *work)
799 {
800 	unsigned long pool_id = get_work_pool_id(work);
801 
802 	pool_id <<= WORK_OFFQ_POOL_SHIFT;
803 	set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
804 }
805 
work_is_canceling(struct work_struct * work)806 static bool work_is_canceling(struct work_struct *work)
807 {
808 	unsigned long data = atomic_long_read(&work->data);
809 
810 	return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
811 }
812 
813 /*
814  * Policy functions.  These define the policies on how the global worker
815  * pools are managed.  Unless noted otherwise, these functions assume that
816  * they're being called with pool->lock held.
817  */
818 
819 /*
820  * Need to wake up a worker?  Called from anything but currently
821  * running workers.
822  *
823  * Note that, because unbound workers never contribute to nr_running, this
824  * function will always return %true for unbound pools as long as the
825  * worklist isn't empty.
826  */
need_more_worker(struct worker_pool * pool)827 static bool need_more_worker(struct worker_pool *pool)
828 {
829 	return !list_empty(&pool->worklist) && !pool->nr_running;
830 }
831 
832 /* Can I start working?  Called from busy but !running workers. */
may_start_working(struct worker_pool * pool)833 static bool may_start_working(struct worker_pool *pool)
834 {
835 	return pool->nr_idle;
836 }
837 
838 /* Do I need to keep working?  Called from currently running workers. */
keep_working(struct worker_pool * pool)839 static bool keep_working(struct worker_pool *pool)
840 {
841 	return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
842 }
843 
844 /* Do we need a new worker?  Called from manager. */
need_to_create_worker(struct worker_pool * pool)845 static bool need_to_create_worker(struct worker_pool *pool)
846 {
847 	return need_more_worker(pool) && !may_start_working(pool);
848 }
849 
850 /* Do we have too many workers and should some go away? */
too_many_workers(struct worker_pool * pool)851 static bool too_many_workers(struct worker_pool *pool)
852 {
853 	bool managing = pool->flags & POOL_MANAGER_ACTIVE;
854 	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
855 	int nr_busy = pool->nr_workers - nr_idle;
856 
857 	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
858 }
859 
860 /**
861  * worker_set_flags - set worker flags and adjust nr_running accordingly
862  * @worker: self
863  * @flags: flags to set
864  *
865  * Set @flags in @worker->flags and adjust nr_running accordingly.
866  */
worker_set_flags(struct worker * worker,unsigned int flags)867 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
868 {
869 	struct worker_pool *pool = worker->pool;
870 
871 	lockdep_assert_held(&pool->lock);
872 
873 	/* If transitioning into NOT_RUNNING, adjust nr_running. */
874 	if ((flags & WORKER_NOT_RUNNING) &&
875 	    !(worker->flags & WORKER_NOT_RUNNING)) {
876 		pool->nr_running--;
877 	}
878 
879 	worker->flags |= flags;
880 }
881 
882 /**
883  * worker_clr_flags - clear worker flags and adjust nr_running accordingly
884  * @worker: self
885  * @flags: flags to clear
886  *
887  * Clear @flags in @worker->flags and adjust nr_running accordingly.
888  */
worker_clr_flags(struct worker * worker,unsigned int flags)889 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
890 {
891 	struct worker_pool *pool = worker->pool;
892 	unsigned int oflags = worker->flags;
893 
894 	lockdep_assert_held(&pool->lock);
895 
896 	worker->flags &= ~flags;
897 
898 	/*
899 	 * If transitioning out of NOT_RUNNING, increment nr_running.  Note
900 	 * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
901 	 * of multiple flags, not a single flag.
902 	 */
903 	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
904 		if (!(worker->flags & WORKER_NOT_RUNNING))
905 			pool->nr_running++;
906 }
907 
908 /* Return the first idle worker.  Called with pool->lock held. */
first_idle_worker(struct worker_pool * pool)909 static struct worker *first_idle_worker(struct worker_pool *pool)
910 {
911 	if (unlikely(list_empty(&pool->idle_list)))
912 		return NULL;
913 
914 	return list_first_entry(&pool->idle_list, struct worker, entry);
915 }
916 
917 /**
918  * worker_enter_idle - enter idle state
919  * @worker: worker which is entering idle state
920  *
921  * @worker is entering idle state.  Update stats and idle timer if
922  * necessary.
923  *
924  * LOCKING:
925  * raw_spin_lock_irq(pool->lock).
926  */
worker_enter_idle(struct worker * worker)927 static void worker_enter_idle(struct worker *worker)
928 {
929 	struct worker_pool *pool = worker->pool;
930 
931 	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
932 	    WARN_ON_ONCE(!list_empty(&worker->entry) &&
933 			 (worker->hentry.next || worker->hentry.pprev)))
934 		return;
935 
936 	/* can't use worker_set_flags(), also called from create_worker() */
937 	worker->flags |= WORKER_IDLE;
938 	pool->nr_idle++;
939 	worker->last_active = jiffies;
940 
941 	/* idle_list is LIFO */
942 	list_add(&worker->entry, &pool->idle_list);
943 
944 	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
945 		mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
946 
947 	/* Sanity check nr_running. */
948 	WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
949 }
950 
951 /**
952  * worker_leave_idle - leave idle state
953  * @worker: worker which is leaving idle state
954  *
955  * @worker is leaving idle state.  Update stats.
956  *
957  * LOCKING:
958  * raw_spin_lock_irq(pool->lock).
959  */
worker_leave_idle(struct worker * worker)960 static void worker_leave_idle(struct worker *worker)
961 {
962 	struct worker_pool *pool = worker->pool;
963 
964 	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
965 		return;
966 	worker_clr_flags(worker, WORKER_IDLE);
967 	pool->nr_idle--;
968 	list_del_init(&worker->entry);
969 }
970 
971 /**
972  * find_worker_executing_work - find worker which is executing a work
973  * @pool: pool of interest
974  * @work: work to find worker for
975  *
976  * Find a worker which is executing @work on @pool by searching
977  * @pool->busy_hash which is keyed by the address of @work.  For a worker
978  * to match, its current execution should match the address of @work and
979  * its work function.  This is to avoid unwanted dependency between
980  * unrelated work executions through a work item being recycled while still
981  * being executed.
982  *
983  * This is a bit tricky.  A work item may be freed once its execution
984  * starts and nothing prevents the freed area from being recycled for
985  * another work item.  If the same work item address ends up being reused
986  * before the original execution finishes, workqueue will identify the
987  * recycled work item as currently executing and make it wait until the
988  * current execution finishes, introducing an unwanted dependency.
989  *
990  * This function checks the work item address and work function to avoid
991  * false positives.  Note that this isn't complete as one may construct a
992  * work function which can introduce dependency onto itself through a
993  * recycled work item.  Well, if somebody wants to shoot oneself in the
994  * foot that badly, there's only so much we can do, and if such deadlock
995  * actually occurs, it should be easy to locate the culprit work function.
996  *
997  * CONTEXT:
998  * raw_spin_lock_irq(pool->lock).
999  *
1000  * Return:
1001  * Pointer to worker which is executing @work if found, %NULL
1002  * otherwise.
1003  */
find_worker_executing_work(struct worker_pool * pool,struct work_struct * work)1004 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1005 						 struct work_struct *work)
1006 {
1007 	struct worker *worker;
1008 
1009 	hash_for_each_possible(pool->busy_hash, worker, hentry,
1010 			       (unsigned long)work)
1011 		if (worker->current_work == work &&
1012 		    worker->current_func == work->func)
1013 			return worker;
1014 
1015 	return NULL;
1016 }
1017 
1018 /**
1019  * move_linked_works - move linked works to a list
1020  * @work: start of series of works to be scheduled
1021  * @head: target list to append @work to
1022  * @nextp: out parameter for nested worklist walking
1023  *
1024  * Schedule linked works starting from @work to @head. Work series to be
1025  * scheduled starts at @work and includes any consecutive work with
1026  * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1027  * @nextp.
1028  *
1029  * CONTEXT:
1030  * raw_spin_lock_irq(pool->lock).
1031  */
move_linked_works(struct work_struct * work,struct list_head * head,struct work_struct ** nextp)1032 static void move_linked_works(struct work_struct *work, struct list_head *head,
1033 			      struct work_struct **nextp)
1034 {
1035 	struct work_struct *n;
1036 
1037 	/*
1038 	 * Linked worklist will always end before the end of the list,
1039 	 * use NULL for list head.
1040 	 */
1041 	list_for_each_entry_safe_from(work, n, NULL, entry) {
1042 		list_move_tail(&work->entry, head);
1043 		if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1044 			break;
1045 	}
1046 
1047 	/*
1048 	 * If we're already inside safe list traversal and have moved
1049 	 * multiple works to the scheduled queue, the next position
1050 	 * needs to be updated.
1051 	 */
1052 	if (nextp)
1053 		*nextp = n;
1054 }
1055 
1056 /**
1057  * assign_work - assign a work item and its linked work items to a worker
1058  * @work: work to assign
1059  * @worker: worker to assign to
1060  * @nextp: out parameter for nested worklist walking
1061  *
1062  * Assign @work and its linked work items to @worker. If @work is already being
1063  * executed by another worker in the same pool, it'll be punted there.
1064  *
1065  * If @nextp is not NULL, it's updated to point to the next work of the last
1066  * scheduled work. This allows assign_work() to be nested inside
1067  * list_for_each_entry_safe().
1068  *
1069  * Returns %true if @work was successfully assigned to @worker. %false if @work
1070  * was punted to another worker already executing it.
1071  */
assign_work(struct work_struct * work,struct worker * worker,struct work_struct ** nextp)1072 static bool assign_work(struct work_struct *work, struct worker *worker,
1073 			struct work_struct **nextp)
1074 {
1075 	struct worker_pool *pool = worker->pool;
1076 	struct worker *collision;
1077 
1078 	lockdep_assert_held(&pool->lock);
1079 
1080 	/*
1081 	 * A single work shouldn't be executed concurrently by multiple workers.
1082 	 * __queue_work() ensures that @work doesn't jump to a different pool
1083 	 * while still running in the previous pool. Here, we should ensure that
1084 	 * @work is not executed concurrently by multiple workers from the same
1085 	 * pool. Check whether anyone is already processing the work. If so,
1086 	 * defer the work to the currently executing one.
1087 	 */
1088 	collision = find_worker_executing_work(pool, work);
1089 	if (unlikely(collision)) {
1090 		move_linked_works(work, &collision->scheduled, nextp);
1091 		return false;
1092 	}
1093 
1094 	move_linked_works(work, &worker->scheduled, nextp);
1095 	return true;
1096 }
1097 
1098 /**
1099  * kick_pool - wake up an idle worker if necessary
1100  * @pool: pool to kick
1101  *
1102  * @pool may have pending work items. Wake up worker if necessary. Returns
1103  * whether a worker was woken up.
1104  */
kick_pool(struct worker_pool * pool)1105 static bool kick_pool(struct worker_pool *pool)
1106 {
1107 	struct worker *worker = first_idle_worker(pool);
1108 	struct task_struct *p;
1109 
1110 	lockdep_assert_held(&pool->lock);
1111 
1112 	if (!need_more_worker(pool) || !worker)
1113 		return false;
1114 
1115 	p = worker->task;
1116 
1117 #ifdef CONFIG_SMP
1118 	/*
1119 	 * Idle @worker is about to execute @work and waking up provides an
1120 	 * opportunity to migrate @worker at a lower cost by setting the task's
1121 	 * wake_cpu field. Let's see if we want to move @worker to improve
1122 	 * execution locality.
1123 	 *
1124 	 * We're waking the worker that went idle the latest and there's some
1125 	 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1126 	 * so, setting the wake_cpu won't do anything. As this is a best-effort
1127 	 * optimization and the race window is narrow, let's leave as-is for
1128 	 * now. If this becomes pronounced, we can skip over workers which are
1129 	 * still on cpu when picking an idle worker.
1130 	 *
1131 	 * If @pool has non-strict affinity, @worker might have ended up outside
1132 	 * its affinity scope. Repatriate.
1133 	 */
1134 	if (!pool->attrs->affn_strict &&
1135 	    !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1136 		struct work_struct *work = list_first_entry(&pool->worklist,
1137 						struct work_struct, entry);
1138 		p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask);
1139 		get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1140 	}
1141 #endif
1142 	wake_up_process(p);
1143 	return true;
1144 }
1145 
1146 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1147 
1148 /*
1149  * Concurrency-managed per-cpu work items that hog CPU for longer than
1150  * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1151  * which prevents them from stalling other concurrency-managed work items. If a
1152  * work function keeps triggering this mechanism, it's likely that the work item
1153  * should be using an unbound workqueue instead.
1154  *
1155  * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1156  * and report them so that they can be examined and converted to use unbound
1157  * workqueues as appropriate. To avoid flooding the console, each violating work
1158  * function is tracked and reported with exponential backoff.
1159  */
1160 #define WCI_MAX_ENTS 128
1161 
1162 struct wci_ent {
1163 	work_func_t		func;
1164 	atomic64_t		cnt;
1165 	struct hlist_node	hash_node;
1166 };
1167 
1168 static struct wci_ent wci_ents[WCI_MAX_ENTS];
1169 static int wci_nr_ents;
1170 static DEFINE_RAW_SPINLOCK(wci_lock);
1171 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1172 
wci_find_ent(work_func_t func)1173 static struct wci_ent *wci_find_ent(work_func_t func)
1174 {
1175 	struct wci_ent *ent;
1176 
1177 	hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1178 				   (unsigned long)func) {
1179 		if (ent->func == func)
1180 			return ent;
1181 	}
1182 	return NULL;
1183 }
1184 
wq_cpu_intensive_report(work_func_t func)1185 static void wq_cpu_intensive_report(work_func_t func)
1186 {
1187 	struct wci_ent *ent;
1188 
1189 restart:
1190 	ent = wci_find_ent(func);
1191 	if (ent) {
1192 		u64 cnt;
1193 
1194 		/*
1195 		 * Start reporting from the fourth time and back off
1196 		 * exponentially.
1197 		 */
1198 		cnt = atomic64_inc_return_relaxed(&ent->cnt);
1199 		if (cnt >= 4 && is_power_of_2(cnt))
1200 			printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1201 					ent->func, wq_cpu_intensive_thresh_us,
1202 					atomic64_read(&ent->cnt));
1203 		return;
1204 	}
1205 
1206 	/*
1207 	 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1208 	 * is exhausted, something went really wrong and we probably made enough
1209 	 * noise already.
1210 	 */
1211 	if (wci_nr_ents >= WCI_MAX_ENTS)
1212 		return;
1213 
1214 	raw_spin_lock(&wci_lock);
1215 
1216 	if (wci_nr_ents >= WCI_MAX_ENTS) {
1217 		raw_spin_unlock(&wci_lock);
1218 		return;
1219 	}
1220 
1221 	if (wci_find_ent(func)) {
1222 		raw_spin_unlock(&wci_lock);
1223 		goto restart;
1224 	}
1225 
1226 	ent = &wci_ents[wci_nr_ents++];
1227 	ent->func = func;
1228 	atomic64_set(&ent->cnt, 1);
1229 	hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1230 
1231 	raw_spin_unlock(&wci_lock);
1232 }
1233 
1234 #else	/* CONFIG_WQ_CPU_INTENSIVE_REPORT */
wq_cpu_intensive_report(work_func_t func)1235 static void wq_cpu_intensive_report(work_func_t func) {}
1236 #endif	/* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1237 
1238 /**
1239  * wq_worker_running - a worker is running again
1240  * @task: task waking up
1241  *
1242  * This function is called when a worker returns from schedule()
1243  */
wq_worker_running(struct task_struct * task)1244 void wq_worker_running(struct task_struct *task)
1245 {
1246 	struct worker *worker = kthread_data(task);
1247 
1248 	if (!READ_ONCE(worker->sleeping))
1249 		return;
1250 
1251 	/*
1252 	 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1253 	 * and the nr_running increment below, we may ruin the nr_running reset
1254 	 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1255 	 * pool. Protect against such race.
1256 	 */
1257 	preempt_disable();
1258 	if (!(worker->flags & WORKER_NOT_RUNNING))
1259 		worker->pool->nr_running++;
1260 	preempt_enable();
1261 
1262 	/*
1263 	 * CPU intensive auto-detection cares about how long a work item hogged
1264 	 * CPU without sleeping. Reset the starting timestamp on wakeup.
1265 	 */
1266 	worker->current_at = worker->task->se.sum_exec_runtime;
1267 
1268 	WRITE_ONCE(worker->sleeping, 0);
1269 }
1270 
1271 /**
1272  * wq_worker_sleeping - a worker is going to sleep
1273  * @task: task going to sleep
1274  *
1275  * This function is called from schedule() when a busy worker is
1276  * going to sleep.
1277  */
wq_worker_sleeping(struct task_struct * task)1278 void wq_worker_sleeping(struct task_struct *task)
1279 {
1280 	struct worker *worker = kthread_data(task);
1281 	struct worker_pool *pool;
1282 
1283 	/*
1284 	 * Rescuers, which may not have all the fields set up like normal
1285 	 * workers, also reach here, let's not access anything before
1286 	 * checking NOT_RUNNING.
1287 	 */
1288 	if (worker->flags & WORKER_NOT_RUNNING)
1289 		return;
1290 
1291 	pool = worker->pool;
1292 
1293 	/* Return if preempted before wq_worker_running() was reached */
1294 	if (READ_ONCE(worker->sleeping))
1295 		return;
1296 
1297 	WRITE_ONCE(worker->sleeping, 1);
1298 	raw_spin_lock_irq(&pool->lock);
1299 
1300 	/*
1301 	 * Recheck in case unbind_workers() preempted us. We don't
1302 	 * want to decrement nr_running after the worker is unbound
1303 	 * and nr_running has been reset.
1304 	 */
1305 	if (worker->flags & WORKER_NOT_RUNNING) {
1306 		raw_spin_unlock_irq(&pool->lock);
1307 		return;
1308 	}
1309 
1310 	pool->nr_running--;
1311 	if (kick_pool(pool))
1312 		worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1313 
1314 	raw_spin_unlock_irq(&pool->lock);
1315 }
1316 
1317 /**
1318  * wq_worker_tick - a scheduler tick occurred while a kworker is running
1319  * @task: task currently running
1320  *
1321  * Called from scheduler_tick(). We're in the IRQ context and the current
1322  * worker's fields which follow the 'K' locking rule can be accessed safely.
1323  */
wq_worker_tick(struct task_struct * task)1324 void wq_worker_tick(struct task_struct *task)
1325 {
1326 	struct worker *worker = kthread_data(task);
1327 	struct pool_workqueue *pwq = worker->current_pwq;
1328 	struct worker_pool *pool = worker->pool;
1329 
1330 	if (!pwq)
1331 		return;
1332 
1333 	pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1334 
1335 	if (!wq_cpu_intensive_thresh_us)
1336 		return;
1337 
1338 	/*
1339 	 * If the current worker is concurrency managed and hogged the CPU for
1340 	 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1341 	 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1342 	 *
1343 	 * Set @worker->sleeping means that @worker is in the process of
1344 	 * switching out voluntarily and won't be contributing to
1345 	 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1346 	 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1347 	 * double decrements. The task is releasing the CPU anyway. Let's skip.
1348 	 * We probably want to make this prettier in the future.
1349 	 */
1350 	if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1351 	    worker->task->se.sum_exec_runtime - worker->current_at <
1352 	    wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1353 		return;
1354 
1355 	raw_spin_lock(&pool->lock);
1356 
1357 	worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1358 	wq_cpu_intensive_report(worker->current_func);
1359 	pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1360 
1361 	if (kick_pool(pool))
1362 		pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1363 
1364 	raw_spin_unlock(&pool->lock);
1365 }
1366 
1367 /**
1368  * wq_worker_last_func - retrieve worker's last work function
1369  * @task: Task to retrieve last work function of.
1370  *
1371  * Determine the last function a worker executed. This is called from
1372  * the scheduler to get a worker's last known identity.
1373  *
1374  * CONTEXT:
1375  * raw_spin_lock_irq(rq->lock)
1376  *
1377  * This function is called during schedule() when a kworker is going
1378  * to sleep. It's used by psi to identify aggregation workers during
1379  * dequeuing, to allow periodic aggregation to shut-off when that
1380  * worker is the last task in the system or cgroup to go to sleep.
1381  *
1382  * As this function doesn't involve any workqueue-related locking, it
1383  * only returns stable values when called from inside the scheduler's
1384  * queuing and dequeuing paths, when @task, which must be a kworker,
1385  * is guaranteed to not be processing any works.
1386  *
1387  * Return:
1388  * The last work function %current executed as a worker, NULL if it
1389  * hasn't executed any work yet.
1390  */
wq_worker_last_func(struct task_struct * task)1391 work_func_t wq_worker_last_func(struct task_struct *task)
1392 {
1393 	struct worker *worker = kthread_data(task);
1394 
1395 	return worker->last_func;
1396 }
1397 
1398 /**
1399  * get_pwq - get an extra reference on the specified pool_workqueue
1400  * @pwq: pool_workqueue to get
1401  *
1402  * Obtain an extra reference on @pwq.  The caller should guarantee that
1403  * @pwq has positive refcnt and be holding the matching pool->lock.
1404  */
get_pwq(struct pool_workqueue * pwq)1405 static void get_pwq(struct pool_workqueue *pwq)
1406 {
1407 	lockdep_assert_held(&pwq->pool->lock);
1408 	WARN_ON_ONCE(pwq->refcnt <= 0);
1409 	pwq->refcnt++;
1410 }
1411 
1412 /**
1413  * put_pwq - put a pool_workqueue reference
1414  * @pwq: pool_workqueue to put
1415  *
1416  * Drop a reference of @pwq.  If its refcnt reaches zero, schedule its
1417  * destruction.  The caller should be holding the matching pool->lock.
1418  */
put_pwq(struct pool_workqueue * pwq)1419 static void put_pwq(struct pool_workqueue *pwq)
1420 {
1421 	lockdep_assert_held(&pwq->pool->lock);
1422 	if (likely(--pwq->refcnt))
1423 		return;
1424 	/*
1425 	 * @pwq can't be released under pool->lock, bounce to a dedicated
1426 	 * kthread_worker to avoid A-A deadlocks.
1427 	 */
1428 	kthread_queue_work(pwq_release_worker, &pwq->release_work);
1429 }
1430 
1431 /**
1432  * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1433  * @pwq: pool_workqueue to put (can be %NULL)
1434  *
1435  * put_pwq() with locking.  This function also allows %NULL @pwq.
1436  */
put_pwq_unlocked(struct pool_workqueue * pwq)1437 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1438 {
1439 	if (pwq) {
1440 		/*
1441 		 * As both pwqs and pools are RCU protected, the
1442 		 * following lock operations are safe.
1443 		 */
1444 		raw_spin_lock_irq(&pwq->pool->lock);
1445 		put_pwq(pwq);
1446 		raw_spin_unlock_irq(&pwq->pool->lock);
1447 	}
1448 }
1449 
pwq_activate_inactive_work(struct work_struct * work)1450 static void pwq_activate_inactive_work(struct work_struct *work)
1451 {
1452 	struct pool_workqueue *pwq = get_work_pwq(work);
1453 
1454 	trace_workqueue_activate_work(work);
1455 	if (list_empty(&pwq->pool->worklist))
1456 		pwq->pool->watchdog_ts = jiffies;
1457 	move_linked_works(work, &pwq->pool->worklist, NULL);
1458 	__clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work));
1459 	pwq->nr_active++;
1460 }
1461 
pwq_activate_first_inactive(struct pool_workqueue * pwq)1462 static void pwq_activate_first_inactive(struct pool_workqueue *pwq)
1463 {
1464 	struct work_struct *work = list_first_entry(&pwq->inactive_works,
1465 						    struct work_struct, entry);
1466 
1467 	pwq_activate_inactive_work(work);
1468 }
1469 
1470 /**
1471  * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1472  * @pwq: pwq of interest
1473  * @work_data: work_data of work which left the queue
1474  *
1475  * A work either has completed or is removed from pending queue,
1476  * decrement nr_in_flight of its pwq and handle workqueue flushing.
1477  *
1478  * CONTEXT:
1479  * raw_spin_lock_irq(pool->lock).
1480  */
pwq_dec_nr_in_flight(struct pool_workqueue * pwq,unsigned long work_data)1481 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1482 {
1483 	int color = get_work_color(work_data);
1484 
1485 	if (!(work_data & WORK_STRUCT_INACTIVE)) {
1486 		pwq->nr_active--;
1487 		if (!list_empty(&pwq->inactive_works)) {
1488 			/* one down, submit an inactive one */
1489 			if (pwq->nr_active < pwq->max_active)
1490 				pwq_activate_first_inactive(pwq);
1491 		}
1492 	}
1493 
1494 	pwq->nr_in_flight[color]--;
1495 
1496 	/* is flush in progress and are we at the flushing tip? */
1497 	if (likely(pwq->flush_color != color))
1498 		goto out_put;
1499 
1500 	/* are there still in-flight works? */
1501 	if (pwq->nr_in_flight[color])
1502 		goto out_put;
1503 
1504 	/* this pwq is done, clear flush_color */
1505 	pwq->flush_color = -1;
1506 
1507 	/*
1508 	 * If this was the last pwq, wake up the first flusher.  It
1509 	 * will handle the rest.
1510 	 */
1511 	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1512 		complete(&pwq->wq->first_flusher->done);
1513 out_put:
1514 	put_pwq(pwq);
1515 }
1516 
1517 /**
1518  * try_to_grab_pending - steal work item from worklist and disable irq
1519  * @work: work item to steal
1520  * @is_dwork: @work is a delayed_work
1521  * @flags: place to store irq state
1522  *
1523  * Try to grab PENDING bit of @work.  This function can handle @work in any
1524  * stable state - idle, on timer or on worklist.
1525  *
1526  * Return:
1527  *
1528  *  ========	================================================================
1529  *  1		if @work was pending and we successfully stole PENDING
1530  *  0		if @work was idle and we claimed PENDING
1531  *  -EAGAIN	if PENDING couldn't be grabbed at the moment, safe to busy-retry
1532  *  -ENOENT	if someone else is canceling @work, this state may persist
1533  *		for arbitrarily long
1534  *  ========	================================================================
1535  *
1536  * Note:
1537  * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1538  * interrupted while holding PENDING and @work off queue, irq must be
1539  * disabled on entry.  This, combined with delayed_work->timer being
1540  * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1541  *
1542  * On successful return, >= 0, irq is disabled and the caller is
1543  * responsible for releasing it using local_irq_restore(*@flags).
1544  *
1545  * This function is safe to call from any context including IRQ handler.
1546  */
try_to_grab_pending(struct work_struct * work,bool is_dwork,unsigned long * flags)1547 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1548 			       unsigned long *flags)
1549 {
1550 	struct worker_pool *pool;
1551 	struct pool_workqueue *pwq;
1552 
1553 	local_irq_save(*flags);
1554 
1555 	/* try to steal the timer if it exists */
1556 	if (is_dwork) {
1557 		struct delayed_work *dwork = to_delayed_work(work);
1558 
1559 		/*
1560 		 * dwork->timer is irqsafe.  If del_timer() fails, it's
1561 		 * guaranteed that the timer is not queued anywhere and not
1562 		 * running on the local CPU.
1563 		 */
1564 		if (likely(del_timer(&dwork->timer)))
1565 			return 1;
1566 	}
1567 
1568 	/* try to claim PENDING the normal way */
1569 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1570 		return 0;
1571 
1572 	rcu_read_lock();
1573 	/*
1574 	 * The queueing is in progress, or it is already queued. Try to
1575 	 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1576 	 */
1577 	pool = get_work_pool(work);
1578 	if (!pool)
1579 		goto fail;
1580 
1581 	raw_spin_lock(&pool->lock);
1582 	/*
1583 	 * work->data is guaranteed to point to pwq only while the work
1584 	 * item is queued on pwq->wq, and both updating work->data to point
1585 	 * to pwq on queueing and to pool on dequeueing are done under
1586 	 * pwq->pool->lock.  This in turn guarantees that, if work->data
1587 	 * points to pwq which is associated with a locked pool, the work
1588 	 * item is currently queued on that pool.
1589 	 */
1590 	pwq = get_work_pwq(work);
1591 	if (pwq && pwq->pool == pool) {
1592 		debug_work_deactivate(work);
1593 
1594 		/*
1595 		 * A cancelable inactive work item must be in the
1596 		 * pwq->inactive_works since a queued barrier can't be
1597 		 * canceled (see the comments in insert_wq_barrier()).
1598 		 *
1599 		 * An inactive work item cannot be grabbed directly because
1600 		 * it might have linked barrier work items which, if left
1601 		 * on the inactive_works list, will confuse pwq->nr_active
1602 		 * management later on and cause stall.  Make sure the work
1603 		 * item is activated before grabbing.
1604 		 */
1605 		if (*work_data_bits(work) & WORK_STRUCT_INACTIVE)
1606 			pwq_activate_inactive_work(work);
1607 
1608 		list_del_init(&work->entry);
1609 		pwq_dec_nr_in_flight(pwq, *work_data_bits(work));
1610 
1611 		/* work->data points to pwq iff queued, point to pool */
1612 		set_work_pool_and_keep_pending(work, pool->id);
1613 
1614 		raw_spin_unlock(&pool->lock);
1615 		rcu_read_unlock();
1616 		return 1;
1617 	}
1618 	raw_spin_unlock(&pool->lock);
1619 fail:
1620 	rcu_read_unlock();
1621 	local_irq_restore(*flags);
1622 	if (work_is_canceling(work))
1623 		return -ENOENT;
1624 	cpu_relax();
1625 	return -EAGAIN;
1626 }
1627 
1628 /**
1629  * insert_work - insert a work into a pool
1630  * @pwq: pwq @work belongs to
1631  * @work: work to insert
1632  * @head: insertion point
1633  * @extra_flags: extra WORK_STRUCT_* flags to set
1634  *
1635  * Insert @work which belongs to @pwq after @head.  @extra_flags is or'd to
1636  * work_struct flags.
1637  *
1638  * CONTEXT:
1639  * raw_spin_lock_irq(pool->lock).
1640  */
insert_work(struct pool_workqueue * pwq,struct work_struct * work,struct list_head * head,unsigned int extra_flags)1641 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1642 			struct list_head *head, unsigned int extra_flags)
1643 {
1644 	debug_work_activate(work);
1645 
1646 	/* record the work call stack in order to print it in KASAN reports */
1647 	kasan_record_aux_stack_noalloc(work);
1648 
1649 	/* we own @work, set data and link */
1650 	set_work_pwq(work, pwq, extra_flags);
1651 	list_add_tail(&work->entry, head);
1652 	get_pwq(pwq);
1653 }
1654 
1655 /*
1656  * Test whether @work is being queued from another work executing on the
1657  * same workqueue.
1658  */
is_chained_work(struct workqueue_struct * wq)1659 static bool is_chained_work(struct workqueue_struct *wq)
1660 {
1661 	struct worker *worker;
1662 
1663 	worker = current_wq_worker();
1664 	/*
1665 	 * Return %true iff I'm a worker executing a work item on @wq.  If
1666 	 * I'm @worker, it's safe to dereference it without locking.
1667 	 */
1668 	return worker && worker->current_pwq->wq == wq;
1669 }
1670 
1671 /*
1672  * When queueing an unbound work item to a wq, prefer local CPU if allowed
1673  * by wq_unbound_cpumask.  Otherwise, round robin among the allowed ones to
1674  * avoid perturbing sensitive tasks.
1675  */
wq_select_unbound_cpu(int cpu)1676 static int wq_select_unbound_cpu(int cpu)
1677 {
1678 	int new_cpu;
1679 
1680 	if (likely(!wq_debug_force_rr_cpu)) {
1681 		if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1682 			return cpu;
1683 	} else {
1684 		pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
1685 	}
1686 
1687 	new_cpu = __this_cpu_read(wq_rr_cpu_last);
1688 	new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1689 	if (unlikely(new_cpu >= nr_cpu_ids)) {
1690 		new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1691 		if (unlikely(new_cpu >= nr_cpu_ids))
1692 			return cpu;
1693 	}
1694 	__this_cpu_write(wq_rr_cpu_last, new_cpu);
1695 
1696 	return new_cpu;
1697 }
1698 
__queue_work(int cpu,struct workqueue_struct * wq,struct work_struct * work)1699 static void __queue_work(int cpu, struct workqueue_struct *wq,
1700 			 struct work_struct *work)
1701 {
1702 	struct pool_workqueue *pwq;
1703 	struct worker_pool *last_pool, *pool;
1704 	unsigned int work_flags;
1705 	unsigned int req_cpu = cpu;
1706 
1707 	/*
1708 	 * While a work item is PENDING && off queue, a task trying to
1709 	 * steal the PENDING will busy-loop waiting for it to either get
1710 	 * queued or lose PENDING.  Grabbing PENDING and queueing should
1711 	 * happen with IRQ disabled.
1712 	 */
1713 	lockdep_assert_irqs_disabled();
1714 
1715 
1716 	/*
1717 	 * For a draining wq, only works from the same workqueue are
1718 	 * allowed. The __WQ_DESTROYING helps to spot the issue that
1719 	 * queues a new work item to a wq after destroy_workqueue(wq).
1720 	 */
1721 	if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
1722 		     WARN_ON_ONCE(!is_chained_work(wq))))
1723 		return;
1724 	rcu_read_lock();
1725 retry:
1726 	/* pwq which will be used unless @work is executing elsewhere */
1727 	if (req_cpu == WORK_CPU_UNBOUND) {
1728 		if (wq->flags & WQ_UNBOUND)
1729 			cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1730 		else
1731 			cpu = raw_smp_processor_id();
1732 	}
1733 
1734 	pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
1735 	pool = pwq->pool;
1736 
1737 	/*
1738 	 * If @work was previously on a different pool, it might still be
1739 	 * running there, in which case the work needs to be queued on that
1740 	 * pool to guarantee non-reentrancy.
1741 	 */
1742 	last_pool = get_work_pool(work);
1743 	if (last_pool && last_pool != pool) {
1744 		struct worker *worker;
1745 
1746 		raw_spin_lock(&last_pool->lock);
1747 
1748 		worker = find_worker_executing_work(last_pool, work);
1749 
1750 		if (worker && worker->current_pwq->wq == wq) {
1751 			pwq = worker->current_pwq;
1752 			pool = pwq->pool;
1753 			WARN_ON_ONCE(pool != last_pool);
1754 		} else {
1755 			/* meh... not running there, queue here */
1756 			raw_spin_unlock(&last_pool->lock);
1757 			raw_spin_lock(&pool->lock);
1758 		}
1759 	} else {
1760 		raw_spin_lock(&pool->lock);
1761 	}
1762 
1763 	/*
1764 	 * pwq is determined and locked. For unbound pools, we could have raced
1765 	 * with pwq release and it could already be dead. If its refcnt is zero,
1766 	 * repeat pwq selection. Note that unbound pwqs never die without
1767 	 * another pwq replacing it in cpu_pwq or while work items are executing
1768 	 * on it, so the retrying is guaranteed to make forward-progress.
1769 	 */
1770 	if (unlikely(!pwq->refcnt)) {
1771 		if (wq->flags & WQ_UNBOUND) {
1772 			raw_spin_unlock(&pool->lock);
1773 			cpu_relax();
1774 			goto retry;
1775 		}
1776 		/* oops */
1777 		WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1778 			  wq->name, cpu);
1779 	}
1780 
1781 	/* pwq determined, queue */
1782 	trace_workqueue_queue_work(req_cpu, pwq, work);
1783 
1784 	if (WARN_ON(!list_empty(&work->entry)))
1785 		goto out;
1786 
1787 	pwq->nr_in_flight[pwq->work_color]++;
1788 	work_flags = work_color_to_flags(pwq->work_color);
1789 
1790 	if (likely(pwq->nr_active < pwq->max_active)) {
1791 		if (list_empty(&pool->worklist))
1792 			pool->watchdog_ts = jiffies;
1793 
1794 		trace_workqueue_activate_work(work);
1795 		pwq->nr_active++;
1796 		insert_work(pwq, work, &pool->worklist, work_flags);
1797 		kick_pool(pool);
1798 	} else {
1799 		work_flags |= WORK_STRUCT_INACTIVE;
1800 		insert_work(pwq, work, &pwq->inactive_works, work_flags);
1801 	}
1802 
1803 out:
1804 	raw_spin_unlock(&pool->lock);
1805 	rcu_read_unlock();
1806 }
1807 
1808 /**
1809  * queue_work_on - queue work on specific cpu
1810  * @cpu: CPU number to execute work on
1811  * @wq: workqueue to use
1812  * @work: work to queue
1813  *
1814  * We queue the work to a specific CPU, the caller must ensure it
1815  * can't go away.  Callers that fail to ensure that the specified
1816  * CPU cannot go away will execute on a randomly chosen CPU.
1817  * But note well that callers specifying a CPU that never has been
1818  * online will get a splat.
1819  *
1820  * Return: %false if @work was already on a queue, %true otherwise.
1821  */
queue_work_on(int cpu,struct workqueue_struct * wq,struct work_struct * work)1822 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1823 		   struct work_struct *work)
1824 {
1825 	bool ret = false;
1826 	unsigned long flags;
1827 
1828 	local_irq_save(flags);
1829 
1830 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1831 		__queue_work(cpu, wq, work);
1832 		ret = true;
1833 	}
1834 
1835 	local_irq_restore(flags);
1836 	return ret;
1837 }
1838 EXPORT_SYMBOL(queue_work_on);
1839 
1840 /**
1841  * select_numa_node_cpu - Select a CPU based on NUMA node
1842  * @node: NUMA node ID that we want to select a CPU from
1843  *
1844  * This function will attempt to find a "random" cpu available on a given
1845  * node. If there are no CPUs available on the given node it will return
1846  * WORK_CPU_UNBOUND indicating that we should just schedule to any
1847  * available CPU if we need to schedule this work.
1848  */
select_numa_node_cpu(int node)1849 static int select_numa_node_cpu(int node)
1850 {
1851 	int cpu;
1852 
1853 	/* Delay binding to CPU if node is not valid or online */
1854 	if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1855 		return WORK_CPU_UNBOUND;
1856 
1857 	/* Use local node/cpu if we are already there */
1858 	cpu = raw_smp_processor_id();
1859 	if (node == cpu_to_node(cpu))
1860 		return cpu;
1861 
1862 	/* Use "random" otherwise know as "first" online CPU of node */
1863 	cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1864 
1865 	/* If CPU is valid return that, otherwise just defer */
1866 	return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1867 }
1868 
1869 /**
1870  * queue_work_node - queue work on a "random" cpu for a given NUMA node
1871  * @node: NUMA node that we are targeting the work for
1872  * @wq: workqueue to use
1873  * @work: work to queue
1874  *
1875  * We queue the work to a "random" CPU within a given NUMA node. The basic
1876  * idea here is to provide a way to somehow associate work with a given
1877  * NUMA node.
1878  *
1879  * This function will only make a best effort attempt at getting this onto
1880  * the right NUMA node. If no node is requested or the requested node is
1881  * offline then we just fall back to standard queue_work behavior.
1882  *
1883  * Currently the "random" CPU ends up being the first available CPU in the
1884  * intersection of cpu_online_mask and the cpumask of the node, unless we
1885  * are running on the node. In that case we just use the current CPU.
1886  *
1887  * Return: %false if @work was already on a queue, %true otherwise.
1888  */
queue_work_node(int node,struct workqueue_struct * wq,struct work_struct * work)1889 bool queue_work_node(int node, struct workqueue_struct *wq,
1890 		     struct work_struct *work)
1891 {
1892 	unsigned long flags;
1893 	bool ret = false;
1894 
1895 	/*
1896 	 * This current implementation is specific to unbound workqueues.
1897 	 * Specifically we only return the first available CPU for a given
1898 	 * node instead of cycling through individual CPUs within the node.
1899 	 *
1900 	 * If this is used with a per-cpu workqueue then the logic in
1901 	 * workqueue_select_cpu_near would need to be updated to allow for
1902 	 * some round robin type logic.
1903 	 */
1904 	WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1905 
1906 	local_irq_save(flags);
1907 
1908 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1909 		int cpu = select_numa_node_cpu(node);
1910 
1911 		__queue_work(cpu, wq, work);
1912 		ret = true;
1913 	}
1914 
1915 	local_irq_restore(flags);
1916 	return ret;
1917 }
1918 EXPORT_SYMBOL_GPL(queue_work_node);
1919 
delayed_work_timer_fn(struct timer_list * t)1920 void delayed_work_timer_fn(struct timer_list *t)
1921 {
1922 	struct delayed_work *dwork = from_timer(dwork, t, timer);
1923 
1924 	/* should have been called from irqsafe timer with irq already off */
1925 	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
1926 }
1927 EXPORT_SYMBOL(delayed_work_timer_fn);
1928 
__queue_delayed_work(int cpu,struct workqueue_struct * wq,struct delayed_work * dwork,unsigned long delay)1929 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1930 				struct delayed_work *dwork, unsigned long delay)
1931 {
1932 	struct timer_list *timer = &dwork->timer;
1933 	struct work_struct *work = &dwork->work;
1934 
1935 	WARN_ON_ONCE(!wq);
1936 	WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1937 	WARN_ON_ONCE(timer_pending(timer));
1938 	WARN_ON_ONCE(!list_empty(&work->entry));
1939 
1940 	/*
1941 	 * If @delay is 0, queue @dwork->work immediately.  This is for
1942 	 * both optimization and correctness.  The earliest @timer can
1943 	 * expire is on the closest next tick and delayed_work users depend
1944 	 * on that there's no such delay when @delay is 0.
1945 	 */
1946 	if (!delay) {
1947 		__queue_work(cpu, wq, &dwork->work);
1948 		return;
1949 	}
1950 
1951 	dwork->wq = wq;
1952 	dwork->cpu = cpu;
1953 	timer->expires = jiffies + delay;
1954 
1955 	if (unlikely(cpu != WORK_CPU_UNBOUND))
1956 		add_timer_on(timer, cpu);
1957 	else
1958 		add_timer(timer);
1959 }
1960 
1961 /**
1962  * queue_delayed_work_on - queue work on specific CPU after delay
1963  * @cpu: CPU number to execute work on
1964  * @wq: workqueue to use
1965  * @dwork: work to queue
1966  * @delay: number of jiffies to wait before queueing
1967  *
1968  * Return: %false if @work was already on a queue, %true otherwise.  If
1969  * @delay is zero and @dwork is idle, it will be scheduled for immediate
1970  * execution.
1971  */
queue_delayed_work_on(int cpu,struct workqueue_struct * wq,struct delayed_work * dwork,unsigned long delay)1972 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1973 			   struct delayed_work *dwork, unsigned long delay)
1974 {
1975 	struct work_struct *work = &dwork->work;
1976 	bool ret = false;
1977 	unsigned long flags;
1978 
1979 	/* read the comment in __queue_work() */
1980 	local_irq_save(flags);
1981 
1982 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1983 		__queue_delayed_work(cpu, wq, dwork, delay);
1984 		ret = true;
1985 	}
1986 
1987 	local_irq_restore(flags);
1988 	return ret;
1989 }
1990 EXPORT_SYMBOL(queue_delayed_work_on);
1991 
1992 /**
1993  * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1994  * @cpu: CPU number to execute work on
1995  * @wq: workqueue to use
1996  * @dwork: work to queue
1997  * @delay: number of jiffies to wait before queueing
1998  *
1999  * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2000  * modify @dwork's timer so that it expires after @delay.  If @delay is
2001  * zero, @work is guaranteed to be scheduled immediately regardless of its
2002  * current state.
2003  *
2004  * Return: %false if @dwork was idle and queued, %true if @dwork was
2005  * pending and its timer was modified.
2006  *
2007  * This function is safe to call from any context including IRQ handler.
2008  * See try_to_grab_pending() for details.
2009  */
mod_delayed_work_on(int cpu,struct workqueue_struct * wq,struct delayed_work * dwork,unsigned long delay)2010 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2011 			 struct delayed_work *dwork, unsigned long delay)
2012 {
2013 	unsigned long flags;
2014 	int ret;
2015 
2016 	do {
2017 		ret = try_to_grab_pending(&dwork->work, true, &flags);
2018 	} while (unlikely(ret == -EAGAIN));
2019 
2020 	if (likely(ret >= 0)) {
2021 		__queue_delayed_work(cpu, wq, dwork, delay);
2022 		local_irq_restore(flags);
2023 	}
2024 
2025 	/* -ENOENT from try_to_grab_pending() becomes %true */
2026 	return ret;
2027 }
2028 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2029 
rcu_work_rcufn(struct rcu_head * rcu)2030 static void rcu_work_rcufn(struct rcu_head *rcu)
2031 {
2032 	struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2033 
2034 	/* read the comment in __queue_work() */
2035 	local_irq_disable();
2036 	__queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2037 	local_irq_enable();
2038 }
2039 
2040 /**
2041  * queue_rcu_work - queue work after a RCU grace period
2042  * @wq: workqueue to use
2043  * @rwork: work to queue
2044  *
2045  * Return: %false if @rwork was already pending, %true otherwise.  Note
2046  * that a full RCU grace period is guaranteed only after a %true return.
2047  * While @rwork is guaranteed to be executed after a %false return, the
2048  * execution may happen before a full RCU grace period has passed.
2049  */
queue_rcu_work(struct workqueue_struct * wq,struct rcu_work * rwork)2050 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2051 {
2052 	struct work_struct *work = &rwork->work;
2053 
2054 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2055 		rwork->wq = wq;
2056 		call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2057 		return true;
2058 	}
2059 
2060 	return false;
2061 }
2062 EXPORT_SYMBOL(queue_rcu_work);
2063 
alloc_worker(int node)2064 static struct worker *alloc_worker(int node)
2065 {
2066 	struct worker *worker;
2067 
2068 	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2069 	if (worker) {
2070 		INIT_LIST_HEAD(&worker->entry);
2071 		INIT_LIST_HEAD(&worker->scheduled);
2072 		INIT_LIST_HEAD(&worker->node);
2073 		/* on creation a worker is in !idle && prep state */
2074 		worker->flags = WORKER_PREP;
2075 	}
2076 	return worker;
2077 }
2078 
pool_allowed_cpus(struct worker_pool * pool)2079 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2080 {
2081 	if (pool->cpu < 0 && pool->attrs->affn_strict)
2082 		return pool->attrs->__pod_cpumask;
2083 	else
2084 		return pool->attrs->cpumask;
2085 }
2086 
2087 /**
2088  * worker_attach_to_pool() - attach a worker to a pool
2089  * @worker: worker to be attached
2090  * @pool: the target pool
2091  *
2092  * Attach @worker to @pool.  Once attached, the %WORKER_UNBOUND flag and
2093  * cpu-binding of @worker are kept coordinated with the pool across
2094  * cpu-[un]hotplugs.
2095  */
worker_attach_to_pool(struct worker * worker,struct worker_pool * pool)2096 static void worker_attach_to_pool(struct worker *worker,
2097 				   struct worker_pool *pool)
2098 {
2099 	mutex_lock(&wq_pool_attach_mutex);
2100 
2101 	/*
2102 	 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
2103 	 * stable across this function.  See the comments above the flag
2104 	 * definition for details.
2105 	 */
2106 	if (pool->flags & POOL_DISASSOCIATED)
2107 		worker->flags |= WORKER_UNBOUND;
2108 	else
2109 		kthread_set_per_cpu(worker->task, pool->cpu);
2110 
2111 	if (worker->rescue_wq)
2112 		set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2113 
2114 	list_add_tail(&worker->node, &pool->workers);
2115 	worker->pool = pool;
2116 
2117 	mutex_unlock(&wq_pool_attach_mutex);
2118 }
2119 
2120 /**
2121  * worker_detach_from_pool() - detach a worker from its pool
2122  * @worker: worker which is attached to its pool
2123  *
2124  * Undo the attaching which had been done in worker_attach_to_pool().  The
2125  * caller worker shouldn't access to the pool after detached except it has
2126  * other reference to the pool.
2127  */
worker_detach_from_pool(struct worker * worker)2128 static void worker_detach_from_pool(struct worker *worker)
2129 {
2130 	struct worker_pool *pool = worker->pool;
2131 	struct completion *detach_completion = NULL;
2132 
2133 	mutex_lock(&wq_pool_attach_mutex);
2134 
2135 	kthread_set_per_cpu(worker->task, -1);
2136 	list_del(&worker->node);
2137 	worker->pool = NULL;
2138 
2139 	if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
2140 		detach_completion = pool->detach_completion;
2141 	mutex_unlock(&wq_pool_attach_mutex);
2142 
2143 	/* clear leftover flags without pool->lock after it is detached */
2144 	worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2145 
2146 	if (detach_completion)
2147 		complete(detach_completion);
2148 }
2149 
2150 /**
2151  * create_worker - create a new workqueue worker
2152  * @pool: pool the new worker will belong to
2153  *
2154  * Create and start a new worker which is attached to @pool.
2155  *
2156  * CONTEXT:
2157  * Might sleep.  Does GFP_KERNEL allocations.
2158  *
2159  * Return:
2160  * Pointer to the newly created worker.
2161  */
create_worker(struct worker_pool * pool)2162 static struct worker *create_worker(struct worker_pool *pool)
2163 {
2164 	struct worker *worker;
2165 	int id;
2166 	char id_buf[23];
2167 
2168 	/* ID is needed to determine kthread name */
2169 	id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2170 	if (id < 0) {
2171 		pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2172 			    ERR_PTR(id));
2173 		return NULL;
2174 	}
2175 
2176 	worker = alloc_worker(pool->node);
2177 	if (!worker) {
2178 		pr_err_once("workqueue: Failed to allocate a worker\n");
2179 		goto fail;
2180 	}
2181 
2182 	worker->id = id;
2183 
2184 	if (pool->cpu >= 0)
2185 		snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
2186 			 pool->attrs->nice < 0  ? "H" : "");
2187 	else
2188 		snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2189 
2190 	worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
2191 					      "kworker/%s", id_buf);
2192 	if (IS_ERR(worker->task)) {
2193 		if (PTR_ERR(worker->task) == -EINTR) {
2194 			pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2195 			       id_buf);
2196 		} else {
2197 			pr_err_once("workqueue: Failed to create a worker thread: %pe",
2198 				    worker->task);
2199 		}
2200 		goto fail;
2201 	}
2202 
2203 	set_user_nice(worker->task, pool->attrs->nice);
2204 	kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2205 
2206 	/* successful, attach the worker to the pool */
2207 	worker_attach_to_pool(worker, pool);
2208 
2209 	/* start the newly created worker */
2210 	raw_spin_lock_irq(&pool->lock);
2211 
2212 	worker->pool->nr_workers++;
2213 	worker_enter_idle(worker);
2214 	kick_pool(pool);
2215 
2216 	/*
2217 	 * @worker is waiting on a completion in kthread() and will trigger hung
2218 	 * check if not woken up soon. As kick_pool() might not have waken it
2219 	 * up, wake it up explicitly once more.
2220 	 */
2221 	wake_up_process(worker->task);
2222 
2223 	raw_spin_unlock_irq(&pool->lock);
2224 
2225 	return worker;
2226 
2227 fail:
2228 	ida_free(&pool->worker_ida, id);
2229 	kfree(worker);
2230 	return NULL;
2231 }
2232 
unbind_worker(struct worker * worker)2233 static void unbind_worker(struct worker *worker)
2234 {
2235 	lockdep_assert_held(&wq_pool_attach_mutex);
2236 
2237 	kthread_set_per_cpu(worker->task, -1);
2238 	if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2239 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2240 	else
2241 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2242 }
2243 
wake_dying_workers(struct list_head * cull_list)2244 static void wake_dying_workers(struct list_head *cull_list)
2245 {
2246 	struct worker *worker, *tmp;
2247 
2248 	list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2249 		list_del_init(&worker->entry);
2250 		unbind_worker(worker);
2251 		/*
2252 		 * If the worker was somehow already running, then it had to be
2253 		 * in pool->idle_list when set_worker_dying() happened or we
2254 		 * wouldn't have gotten here.
2255 		 *
2256 		 * Thus, the worker must either have observed the WORKER_DIE
2257 		 * flag, or have set its state to TASK_IDLE. Either way, the
2258 		 * below will be observed by the worker and is safe to do
2259 		 * outside of pool->lock.
2260 		 */
2261 		wake_up_process(worker->task);
2262 	}
2263 }
2264 
2265 /**
2266  * set_worker_dying - Tag a worker for destruction
2267  * @worker: worker to be destroyed
2268  * @list: transfer worker away from its pool->idle_list and into list
2269  *
2270  * Tag @worker for destruction and adjust @pool stats accordingly.  The worker
2271  * should be idle.
2272  *
2273  * CONTEXT:
2274  * raw_spin_lock_irq(pool->lock).
2275  */
set_worker_dying(struct worker * worker,struct list_head * list)2276 static void set_worker_dying(struct worker *worker, struct list_head *list)
2277 {
2278 	struct worker_pool *pool = worker->pool;
2279 
2280 	lockdep_assert_held(&pool->lock);
2281 	lockdep_assert_held(&wq_pool_attach_mutex);
2282 
2283 	/* sanity check frenzy */
2284 	if (WARN_ON(worker->current_work) ||
2285 	    WARN_ON(!list_empty(&worker->scheduled)) ||
2286 	    WARN_ON(!(worker->flags & WORKER_IDLE)))
2287 		return;
2288 
2289 	pool->nr_workers--;
2290 	pool->nr_idle--;
2291 
2292 	worker->flags |= WORKER_DIE;
2293 
2294 	list_move(&worker->entry, list);
2295 	list_move(&worker->node, &pool->dying_workers);
2296 }
2297 
2298 /**
2299  * idle_worker_timeout - check if some idle workers can now be deleted.
2300  * @t: The pool's idle_timer that just expired
2301  *
2302  * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2303  * worker_leave_idle(), as a worker flicking between idle and active while its
2304  * pool is at the too_many_workers() tipping point would cause too much timer
2305  * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2306  * it expire and re-evaluate things from there.
2307  */
idle_worker_timeout(struct timer_list * t)2308 static void idle_worker_timeout(struct timer_list *t)
2309 {
2310 	struct worker_pool *pool = from_timer(pool, t, idle_timer);
2311 	bool do_cull = false;
2312 
2313 	if (work_pending(&pool->idle_cull_work))
2314 		return;
2315 
2316 	raw_spin_lock_irq(&pool->lock);
2317 
2318 	if (too_many_workers(pool)) {
2319 		struct worker *worker;
2320 		unsigned long expires;
2321 
2322 		/* idle_list is kept in LIFO order, check the last one */
2323 		worker = list_entry(pool->idle_list.prev, struct worker, entry);
2324 		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2325 		do_cull = !time_before(jiffies, expires);
2326 
2327 		if (!do_cull)
2328 			mod_timer(&pool->idle_timer, expires);
2329 	}
2330 	raw_spin_unlock_irq(&pool->lock);
2331 
2332 	if (do_cull)
2333 		queue_work(system_unbound_wq, &pool->idle_cull_work);
2334 }
2335 
2336 /**
2337  * idle_cull_fn - cull workers that have been idle for too long.
2338  * @work: the pool's work for handling these idle workers
2339  *
2340  * This goes through a pool's idle workers and gets rid of those that have been
2341  * idle for at least IDLE_WORKER_TIMEOUT seconds.
2342  *
2343  * We don't want to disturb isolated CPUs because of a pcpu kworker being
2344  * culled, so this also resets worker affinity. This requires a sleepable
2345  * context, hence the split between timer callback and work item.
2346  */
idle_cull_fn(struct work_struct * work)2347 static void idle_cull_fn(struct work_struct *work)
2348 {
2349 	struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2350 	LIST_HEAD(cull_list);
2351 
2352 	/*
2353 	 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2354 	 * cannot proceed beyong worker_detach_from_pool() in its self-destruct
2355 	 * path. This is required as a previously-preempted worker could run after
2356 	 * set_worker_dying() has happened but before wake_dying_workers() did.
2357 	 */
2358 	mutex_lock(&wq_pool_attach_mutex);
2359 	raw_spin_lock_irq(&pool->lock);
2360 
2361 	while (too_many_workers(pool)) {
2362 		struct worker *worker;
2363 		unsigned long expires;
2364 
2365 		worker = list_entry(pool->idle_list.prev, struct worker, entry);
2366 		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2367 
2368 		if (time_before(jiffies, expires)) {
2369 			mod_timer(&pool->idle_timer, expires);
2370 			break;
2371 		}
2372 
2373 		set_worker_dying(worker, &cull_list);
2374 	}
2375 
2376 	raw_spin_unlock_irq(&pool->lock);
2377 	wake_dying_workers(&cull_list);
2378 	mutex_unlock(&wq_pool_attach_mutex);
2379 }
2380 
send_mayday(struct work_struct * work)2381 static void send_mayday(struct work_struct *work)
2382 {
2383 	struct pool_workqueue *pwq = get_work_pwq(work);
2384 	struct workqueue_struct *wq = pwq->wq;
2385 
2386 	lockdep_assert_held(&wq_mayday_lock);
2387 
2388 	if (!wq->rescuer)
2389 		return;
2390 
2391 	/* mayday mayday mayday */
2392 	if (list_empty(&pwq->mayday_node)) {
2393 		/*
2394 		 * If @pwq is for an unbound wq, its base ref may be put at
2395 		 * any time due to an attribute change.  Pin @pwq until the
2396 		 * rescuer is done with it.
2397 		 */
2398 		get_pwq(pwq);
2399 		list_add_tail(&pwq->mayday_node, &wq->maydays);
2400 		wake_up_process(wq->rescuer->task);
2401 		pwq->stats[PWQ_STAT_MAYDAY]++;
2402 	}
2403 }
2404 
pool_mayday_timeout(struct timer_list * t)2405 static void pool_mayday_timeout(struct timer_list *t)
2406 {
2407 	struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2408 	struct work_struct *work;
2409 
2410 	raw_spin_lock_irq(&pool->lock);
2411 	raw_spin_lock(&wq_mayday_lock);		/* for wq->maydays */
2412 
2413 	if (need_to_create_worker(pool)) {
2414 		/*
2415 		 * We've been trying to create a new worker but
2416 		 * haven't been successful.  We might be hitting an
2417 		 * allocation deadlock.  Send distress signals to
2418 		 * rescuers.
2419 		 */
2420 		list_for_each_entry(work, &pool->worklist, entry)
2421 			send_mayday(work);
2422 	}
2423 
2424 	raw_spin_unlock(&wq_mayday_lock);
2425 	raw_spin_unlock_irq(&pool->lock);
2426 
2427 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2428 }
2429 
2430 /**
2431  * maybe_create_worker - create a new worker if necessary
2432  * @pool: pool to create a new worker for
2433  *
2434  * Create a new worker for @pool if necessary.  @pool is guaranteed to
2435  * have at least one idle worker on return from this function.  If
2436  * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2437  * sent to all rescuers with works scheduled on @pool to resolve
2438  * possible allocation deadlock.
2439  *
2440  * On return, need_to_create_worker() is guaranteed to be %false and
2441  * may_start_working() %true.
2442  *
2443  * LOCKING:
2444  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2445  * multiple times.  Does GFP_KERNEL allocations.  Called only from
2446  * manager.
2447  */
maybe_create_worker(struct worker_pool * pool)2448 static void maybe_create_worker(struct worker_pool *pool)
2449 __releases(&pool->lock)
2450 __acquires(&pool->lock)
2451 {
2452 restart:
2453 	raw_spin_unlock_irq(&pool->lock);
2454 
2455 	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2456 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2457 
2458 	while (true) {
2459 		if (create_worker(pool) || !need_to_create_worker(pool))
2460 			break;
2461 
2462 		schedule_timeout_interruptible(CREATE_COOLDOWN);
2463 
2464 		if (!need_to_create_worker(pool))
2465 			break;
2466 	}
2467 
2468 	del_timer_sync(&pool->mayday_timer);
2469 	raw_spin_lock_irq(&pool->lock);
2470 	/*
2471 	 * This is necessary even after a new worker was just successfully
2472 	 * created as @pool->lock was dropped and the new worker might have
2473 	 * already become busy.
2474 	 */
2475 	if (need_to_create_worker(pool))
2476 		goto restart;
2477 }
2478 
2479 /**
2480  * manage_workers - manage worker pool
2481  * @worker: self
2482  *
2483  * Assume the manager role and manage the worker pool @worker belongs
2484  * to.  At any given time, there can be only zero or one manager per
2485  * pool.  The exclusion is handled automatically by this function.
2486  *
2487  * The caller can safely start processing works on false return.  On
2488  * true return, it's guaranteed that need_to_create_worker() is false
2489  * and may_start_working() is true.
2490  *
2491  * CONTEXT:
2492  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2493  * multiple times.  Does GFP_KERNEL allocations.
2494  *
2495  * Return:
2496  * %false if the pool doesn't need management and the caller can safely
2497  * start processing works, %true if management function was performed and
2498  * the conditions that the caller verified before calling the function may
2499  * no longer be true.
2500  */
manage_workers(struct worker * worker)2501 static bool manage_workers(struct worker *worker)
2502 {
2503 	struct worker_pool *pool = worker->pool;
2504 
2505 	if (pool->flags & POOL_MANAGER_ACTIVE)
2506 		return false;
2507 
2508 	pool->flags |= POOL_MANAGER_ACTIVE;
2509 	pool->manager = worker;
2510 
2511 	maybe_create_worker(pool);
2512 
2513 	pool->manager = NULL;
2514 	pool->flags &= ~POOL_MANAGER_ACTIVE;
2515 	rcuwait_wake_up(&manager_wait);
2516 	return true;
2517 }
2518 
2519 /**
2520  * process_one_work - process single work
2521  * @worker: self
2522  * @work: work to process
2523  *
2524  * Process @work.  This function contains all the logics necessary to
2525  * process a single work including synchronization against and
2526  * interaction with other workers on the same cpu, queueing and
2527  * flushing.  As long as context requirement is met, any worker can
2528  * call this function to process a work.
2529  *
2530  * CONTEXT:
2531  * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2532  */
process_one_work(struct worker * worker,struct work_struct * work)2533 static void process_one_work(struct worker *worker, struct work_struct *work)
2534 __releases(&pool->lock)
2535 __acquires(&pool->lock)
2536 {
2537 	struct pool_workqueue *pwq = get_work_pwq(work);
2538 	struct worker_pool *pool = worker->pool;
2539 	unsigned long work_data;
2540 #ifdef CONFIG_LOCKDEP
2541 	/*
2542 	 * It is permissible to free the struct work_struct from
2543 	 * inside the function that is called from it, this we need to
2544 	 * take into account for lockdep too.  To avoid bogus "held
2545 	 * lock freed" warnings as well as problems when looking into
2546 	 * work->lockdep_map, make a copy and use that here.
2547 	 */
2548 	struct lockdep_map lockdep_map;
2549 
2550 	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2551 #endif
2552 	/* ensure we're on the correct CPU */
2553 	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2554 		     raw_smp_processor_id() != pool->cpu);
2555 
2556 	/* claim and dequeue */
2557 	debug_work_deactivate(work);
2558 	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2559 	worker->current_work = work;
2560 	worker->current_func = work->func;
2561 	worker->current_pwq = pwq;
2562 	worker->current_at = worker->task->se.sum_exec_runtime;
2563 	work_data = *work_data_bits(work);
2564 	worker->current_color = get_work_color(work_data);
2565 
2566 	/*
2567 	 * Record wq name for cmdline and debug reporting, may get
2568 	 * overridden through set_worker_desc().
2569 	 */
2570 	strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2571 
2572 	list_del_init(&work->entry);
2573 
2574 	/*
2575 	 * CPU intensive works don't participate in concurrency management.
2576 	 * They're the scheduler's responsibility.  This takes @worker out
2577 	 * of concurrency management and the next code block will chain
2578 	 * execution of the pending work items.
2579 	 */
2580 	if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
2581 		worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2582 
2583 	/*
2584 	 * Kick @pool if necessary. It's always noop for per-cpu worker pools
2585 	 * since nr_running would always be >= 1 at this point. This is used to
2586 	 * chain execution of the pending work items for WORKER_NOT_RUNNING
2587 	 * workers such as the UNBOUND and CPU_INTENSIVE ones.
2588 	 */
2589 	kick_pool(pool);
2590 
2591 	/*
2592 	 * Record the last pool and clear PENDING which should be the last
2593 	 * update to @work.  Also, do this inside @pool->lock so that
2594 	 * PENDING and queued state changes happen together while IRQ is
2595 	 * disabled.
2596 	 */
2597 	set_work_pool_and_clear_pending(work, pool->id);
2598 
2599 	pwq->stats[PWQ_STAT_STARTED]++;
2600 	raw_spin_unlock_irq(&pool->lock);
2601 
2602 	lock_map_acquire(&pwq->wq->lockdep_map);
2603 	lock_map_acquire(&lockdep_map);
2604 	/*
2605 	 * Strictly speaking we should mark the invariant state without holding
2606 	 * any locks, that is, before these two lock_map_acquire()'s.
2607 	 *
2608 	 * However, that would result in:
2609 	 *
2610 	 *   A(W1)
2611 	 *   WFC(C)
2612 	 *		A(W1)
2613 	 *		C(C)
2614 	 *
2615 	 * Which would create W1->C->W1 dependencies, even though there is no
2616 	 * actual deadlock possible. There are two solutions, using a
2617 	 * read-recursive acquire on the work(queue) 'locks', but this will then
2618 	 * hit the lockdep limitation on recursive locks, or simply discard
2619 	 * these locks.
2620 	 *
2621 	 * AFAICT there is no possible deadlock scenario between the
2622 	 * flush_work() and complete() primitives (except for single-threaded
2623 	 * workqueues), so hiding them isn't a problem.
2624 	 */
2625 	lockdep_invariant_state(true);
2626 	trace_workqueue_execute_start(work);
2627 	worker->current_func(work);
2628 	/*
2629 	 * While we must be careful to not use "work" after this, the trace
2630 	 * point will only record its address.
2631 	 */
2632 	trace_workqueue_execute_end(work, worker->current_func);
2633 	pwq->stats[PWQ_STAT_COMPLETED]++;
2634 	lock_map_release(&lockdep_map);
2635 	lock_map_release(&pwq->wq->lockdep_map);
2636 
2637 	if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2638 		pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2639 		       "     last function: %ps\n",
2640 		       current->comm, preempt_count(), task_pid_nr(current),
2641 		       worker->current_func);
2642 		debug_show_held_locks(current);
2643 		dump_stack();
2644 	}
2645 
2646 	/*
2647 	 * The following prevents a kworker from hogging CPU on !PREEMPTION
2648 	 * kernels, where a requeueing work item waiting for something to
2649 	 * happen could deadlock with stop_machine as such work item could
2650 	 * indefinitely requeue itself while all other CPUs are trapped in
2651 	 * stop_machine. At the same time, report a quiescent RCU state so
2652 	 * the same condition doesn't freeze RCU.
2653 	 */
2654 	cond_resched();
2655 
2656 	raw_spin_lock_irq(&pool->lock);
2657 
2658 	/*
2659 	 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
2660 	 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
2661 	 * wq_cpu_intensive_thresh_us. Clear it.
2662 	 */
2663 	worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2664 
2665 	/* tag the worker for identification in schedule() */
2666 	worker->last_func = worker->current_func;
2667 
2668 	/* we're done with it, release */
2669 	hash_del(&worker->hentry);
2670 	worker->current_work = NULL;
2671 	worker->current_func = NULL;
2672 	worker->current_pwq = NULL;
2673 	worker->current_color = INT_MAX;
2674 	pwq_dec_nr_in_flight(pwq, work_data);
2675 }
2676 
2677 /**
2678  * process_scheduled_works - process scheduled works
2679  * @worker: self
2680  *
2681  * Process all scheduled works.  Please note that the scheduled list
2682  * may change while processing a work, so this function repeatedly
2683  * fetches a work from the top and executes it.
2684  *
2685  * CONTEXT:
2686  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2687  * multiple times.
2688  */
process_scheduled_works(struct worker * worker)2689 static void process_scheduled_works(struct worker *worker)
2690 {
2691 	struct work_struct *work;
2692 	bool first = true;
2693 
2694 	while ((work = list_first_entry_or_null(&worker->scheduled,
2695 						struct work_struct, entry))) {
2696 		if (first) {
2697 			worker->pool->watchdog_ts = jiffies;
2698 			first = false;
2699 		}
2700 		process_one_work(worker, work);
2701 	}
2702 }
2703 
set_pf_worker(bool val)2704 static void set_pf_worker(bool val)
2705 {
2706 	mutex_lock(&wq_pool_attach_mutex);
2707 	if (val)
2708 		current->flags |= PF_WQ_WORKER;
2709 	else
2710 		current->flags &= ~PF_WQ_WORKER;
2711 	mutex_unlock(&wq_pool_attach_mutex);
2712 }
2713 
2714 /**
2715  * worker_thread - the worker thread function
2716  * @__worker: self
2717  *
2718  * The worker thread function.  All workers belong to a worker_pool -
2719  * either a per-cpu one or dynamic unbound one.  These workers process all
2720  * work items regardless of their specific target workqueue.  The only
2721  * exception is work items which belong to workqueues with a rescuer which
2722  * will be explained in rescuer_thread().
2723  *
2724  * Return: 0
2725  */
worker_thread(void * __worker)2726 static int worker_thread(void *__worker)
2727 {
2728 	struct worker *worker = __worker;
2729 	struct worker_pool *pool = worker->pool;
2730 
2731 	/* tell the scheduler that this is a workqueue worker */
2732 	set_pf_worker(true);
2733 woke_up:
2734 	raw_spin_lock_irq(&pool->lock);
2735 
2736 	/* am I supposed to die? */
2737 	if (unlikely(worker->flags & WORKER_DIE)) {
2738 		raw_spin_unlock_irq(&pool->lock);
2739 		set_pf_worker(false);
2740 
2741 		set_task_comm(worker->task, "kworker/dying");
2742 		ida_free(&pool->worker_ida, worker->id);
2743 		worker_detach_from_pool(worker);
2744 		WARN_ON_ONCE(!list_empty(&worker->entry));
2745 		kfree(worker);
2746 		return 0;
2747 	}
2748 
2749 	worker_leave_idle(worker);
2750 recheck:
2751 	/* no more worker necessary? */
2752 	if (!need_more_worker(pool))
2753 		goto sleep;
2754 
2755 	/* do we need to manage? */
2756 	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2757 		goto recheck;
2758 
2759 	/*
2760 	 * ->scheduled list can only be filled while a worker is
2761 	 * preparing to process a work or actually processing it.
2762 	 * Make sure nobody diddled with it while I was sleeping.
2763 	 */
2764 	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2765 
2766 	/*
2767 	 * Finish PREP stage.  We're guaranteed to have at least one idle
2768 	 * worker or that someone else has already assumed the manager
2769 	 * role.  This is where @worker starts participating in concurrency
2770 	 * management if applicable and concurrency management is restored
2771 	 * after being rebound.  See rebind_workers() for details.
2772 	 */
2773 	worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2774 
2775 	do {
2776 		struct work_struct *work =
2777 			list_first_entry(&pool->worklist,
2778 					 struct work_struct, entry);
2779 
2780 		if (assign_work(work, worker, NULL))
2781 			process_scheduled_works(worker);
2782 	} while (keep_working(pool));
2783 
2784 	worker_set_flags(worker, WORKER_PREP);
2785 sleep:
2786 	/*
2787 	 * pool->lock is held and there's no work to process and no need to
2788 	 * manage, sleep.  Workers are woken up only while holding
2789 	 * pool->lock or from local cpu, so setting the current state
2790 	 * before releasing pool->lock is enough to prevent losing any
2791 	 * event.
2792 	 */
2793 	worker_enter_idle(worker);
2794 	__set_current_state(TASK_IDLE);
2795 	raw_spin_unlock_irq(&pool->lock);
2796 	schedule();
2797 	goto woke_up;
2798 }
2799 
2800 /**
2801  * rescuer_thread - the rescuer thread function
2802  * @__rescuer: self
2803  *
2804  * Workqueue rescuer thread function.  There's one rescuer for each
2805  * workqueue which has WQ_MEM_RECLAIM set.
2806  *
2807  * Regular work processing on a pool may block trying to create a new
2808  * worker which uses GFP_KERNEL allocation which has slight chance of
2809  * developing into deadlock if some works currently on the same queue
2810  * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2811  * the problem rescuer solves.
2812  *
2813  * When such condition is possible, the pool summons rescuers of all
2814  * workqueues which have works queued on the pool and let them process
2815  * those works so that forward progress can be guaranteed.
2816  *
2817  * This should happen rarely.
2818  *
2819  * Return: 0
2820  */
rescuer_thread(void * __rescuer)2821 static int rescuer_thread(void *__rescuer)
2822 {
2823 	struct worker *rescuer = __rescuer;
2824 	struct workqueue_struct *wq = rescuer->rescue_wq;
2825 	bool should_stop;
2826 
2827 	set_user_nice(current, RESCUER_NICE_LEVEL);
2828 
2829 	/*
2830 	 * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
2831 	 * doesn't participate in concurrency management.
2832 	 */
2833 	set_pf_worker(true);
2834 repeat:
2835 	set_current_state(TASK_IDLE);
2836 
2837 	/*
2838 	 * By the time the rescuer is requested to stop, the workqueue
2839 	 * shouldn't have any work pending, but @wq->maydays may still have
2840 	 * pwq(s) queued.  This can happen by non-rescuer workers consuming
2841 	 * all the work items before the rescuer got to them.  Go through
2842 	 * @wq->maydays processing before acting on should_stop so that the
2843 	 * list is always empty on exit.
2844 	 */
2845 	should_stop = kthread_should_stop();
2846 
2847 	/* see whether any pwq is asking for help */
2848 	raw_spin_lock_irq(&wq_mayday_lock);
2849 
2850 	while (!list_empty(&wq->maydays)) {
2851 		struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2852 					struct pool_workqueue, mayday_node);
2853 		struct worker_pool *pool = pwq->pool;
2854 		struct work_struct *work, *n;
2855 
2856 		__set_current_state(TASK_RUNNING);
2857 		list_del_init(&pwq->mayday_node);
2858 
2859 		raw_spin_unlock_irq(&wq_mayday_lock);
2860 
2861 		worker_attach_to_pool(rescuer, pool);
2862 
2863 		raw_spin_lock_irq(&pool->lock);
2864 
2865 		/*
2866 		 * Slurp in all works issued via this workqueue and
2867 		 * process'em.
2868 		 */
2869 		WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2870 		list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2871 			if (get_work_pwq(work) == pwq &&
2872 			    assign_work(work, rescuer, &n))
2873 				pwq->stats[PWQ_STAT_RESCUED]++;
2874 		}
2875 
2876 		if (!list_empty(&rescuer->scheduled)) {
2877 			process_scheduled_works(rescuer);
2878 
2879 			/*
2880 			 * The above execution of rescued work items could
2881 			 * have created more to rescue through
2882 			 * pwq_activate_first_inactive() or chained
2883 			 * queueing.  Let's put @pwq back on mayday list so
2884 			 * that such back-to-back work items, which may be
2885 			 * being used to relieve memory pressure, don't
2886 			 * incur MAYDAY_INTERVAL delay inbetween.
2887 			 */
2888 			if (pwq->nr_active && need_to_create_worker(pool)) {
2889 				raw_spin_lock(&wq_mayday_lock);
2890 				/*
2891 				 * Queue iff we aren't racing destruction
2892 				 * and somebody else hasn't queued it already.
2893 				 */
2894 				if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2895 					get_pwq(pwq);
2896 					list_add_tail(&pwq->mayday_node, &wq->maydays);
2897 				}
2898 				raw_spin_unlock(&wq_mayday_lock);
2899 			}
2900 		}
2901 
2902 		/*
2903 		 * Put the reference grabbed by send_mayday().  @pool won't
2904 		 * go away while we're still attached to it.
2905 		 */
2906 		put_pwq(pwq);
2907 
2908 		/*
2909 		 * Leave this pool. Notify regular workers; otherwise, we end up
2910 		 * with 0 concurrency and stalling the execution.
2911 		 */
2912 		kick_pool(pool);
2913 
2914 		raw_spin_unlock_irq(&pool->lock);
2915 
2916 		worker_detach_from_pool(rescuer);
2917 
2918 		raw_spin_lock_irq(&wq_mayday_lock);
2919 	}
2920 
2921 	raw_spin_unlock_irq(&wq_mayday_lock);
2922 
2923 	if (should_stop) {
2924 		__set_current_state(TASK_RUNNING);
2925 		set_pf_worker(false);
2926 		return 0;
2927 	}
2928 
2929 	/* rescuers should never participate in concurrency management */
2930 	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2931 	schedule();
2932 	goto repeat;
2933 }
2934 
2935 /**
2936  * check_flush_dependency - check for flush dependency sanity
2937  * @target_wq: workqueue being flushed
2938  * @target_work: work item being flushed (NULL for workqueue flushes)
2939  *
2940  * %current is trying to flush the whole @target_wq or @target_work on it.
2941  * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2942  * reclaiming memory or running on a workqueue which doesn't have
2943  * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2944  * a deadlock.
2945  */
check_flush_dependency(struct workqueue_struct * target_wq,struct work_struct * target_work)2946 static void check_flush_dependency(struct workqueue_struct *target_wq,
2947 				   struct work_struct *target_work)
2948 {
2949 	work_func_t target_func = target_work ? target_work->func : NULL;
2950 	struct worker *worker;
2951 
2952 	if (target_wq->flags & WQ_MEM_RECLAIM)
2953 		return;
2954 
2955 	worker = current_wq_worker();
2956 
2957 	WARN_ONCE(current->flags & PF_MEMALLOC,
2958 		  "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2959 		  current->pid, current->comm, target_wq->name, target_func);
2960 	WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2961 			      (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2962 		  "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2963 		  worker->current_pwq->wq->name, worker->current_func,
2964 		  target_wq->name, target_func);
2965 }
2966 
2967 struct wq_barrier {
2968 	struct work_struct	work;
2969 	struct completion	done;
2970 	struct task_struct	*task;	/* purely informational */
2971 };
2972 
wq_barrier_func(struct work_struct * work)2973 static void wq_barrier_func(struct work_struct *work)
2974 {
2975 	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2976 	complete(&barr->done);
2977 }
2978 
2979 /**
2980  * insert_wq_barrier - insert a barrier work
2981  * @pwq: pwq to insert barrier into
2982  * @barr: wq_barrier to insert
2983  * @target: target work to attach @barr to
2984  * @worker: worker currently executing @target, NULL if @target is not executing
2985  *
2986  * @barr is linked to @target such that @barr is completed only after
2987  * @target finishes execution.  Please note that the ordering
2988  * guarantee is observed only with respect to @target and on the local
2989  * cpu.
2990  *
2991  * Currently, a queued barrier can't be canceled.  This is because
2992  * try_to_grab_pending() can't determine whether the work to be
2993  * grabbed is at the head of the queue and thus can't clear LINKED
2994  * flag of the previous work while there must be a valid next work
2995  * after a work with LINKED flag set.
2996  *
2997  * Note that when @worker is non-NULL, @target may be modified
2998  * underneath us, so we can't reliably determine pwq from @target.
2999  *
3000  * CONTEXT:
3001  * raw_spin_lock_irq(pool->lock).
3002  */
insert_wq_barrier(struct pool_workqueue * pwq,struct wq_barrier * barr,struct work_struct * target,struct worker * worker)3003 static void insert_wq_barrier(struct pool_workqueue *pwq,
3004 			      struct wq_barrier *barr,
3005 			      struct work_struct *target, struct worker *worker)
3006 {
3007 	unsigned int work_flags = 0;
3008 	unsigned int work_color;
3009 	struct list_head *head;
3010 
3011 	/*
3012 	 * debugobject calls are safe here even with pool->lock locked
3013 	 * as we know for sure that this will not trigger any of the
3014 	 * checks and call back into the fixup functions where we
3015 	 * might deadlock.
3016 	 */
3017 	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
3018 	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3019 
3020 	init_completion_map(&barr->done, &target->lockdep_map);
3021 
3022 	barr->task = current;
3023 
3024 	/* The barrier work item does not participate in pwq->nr_active. */
3025 	work_flags |= WORK_STRUCT_INACTIVE;
3026 
3027 	/*
3028 	 * If @target is currently being executed, schedule the
3029 	 * barrier to the worker; otherwise, put it after @target.
3030 	 */
3031 	if (worker) {
3032 		head = worker->scheduled.next;
3033 		work_color = worker->current_color;
3034 	} else {
3035 		unsigned long *bits = work_data_bits(target);
3036 
3037 		head = target->entry.next;
3038 		/* there can already be other linked works, inherit and set */
3039 		work_flags |= *bits & WORK_STRUCT_LINKED;
3040 		work_color = get_work_color(*bits);
3041 		__set_bit(WORK_STRUCT_LINKED_BIT, bits);
3042 	}
3043 
3044 	pwq->nr_in_flight[work_color]++;
3045 	work_flags |= work_color_to_flags(work_color);
3046 
3047 	insert_work(pwq, &barr->work, head, work_flags);
3048 }
3049 
3050 /**
3051  * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3052  * @wq: workqueue being flushed
3053  * @flush_color: new flush color, < 0 for no-op
3054  * @work_color: new work color, < 0 for no-op
3055  *
3056  * Prepare pwqs for workqueue flushing.
3057  *
3058  * If @flush_color is non-negative, flush_color on all pwqs should be
3059  * -1.  If no pwq has in-flight commands at the specified color, all
3060  * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
3061  * has in flight commands, its pwq->flush_color is set to
3062  * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3063  * wakeup logic is armed and %true is returned.
3064  *
3065  * The caller should have initialized @wq->first_flusher prior to
3066  * calling this function with non-negative @flush_color.  If
3067  * @flush_color is negative, no flush color update is done and %false
3068  * is returned.
3069  *
3070  * If @work_color is non-negative, all pwqs should have the same
3071  * work_color which is previous to @work_color and all will be
3072  * advanced to @work_color.
3073  *
3074  * CONTEXT:
3075  * mutex_lock(wq->mutex).
3076  *
3077  * Return:
3078  * %true if @flush_color >= 0 and there's something to flush.  %false
3079  * otherwise.
3080  */
flush_workqueue_prep_pwqs(struct workqueue_struct * wq,int flush_color,int work_color)3081 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3082 				      int flush_color, int work_color)
3083 {
3084 	bool wait = false;
3085 	struct pool_workqueue *pwq;
3086 
3087 	if (flush_color >= 0) {
3088 		WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3089 		atomic_set(&wq->nr_pwqs_to_flush, 1);
3090 	}
3091 
3092 	for_each_pwq(pwq, wq) {
3093 		struct worker_pool *pool = pwq->pool;
3094 
3095 		raw_spin_lock_irq(&pool->lock);
3096 
3097 		if (flush_color >= 0) {
3098 			WARN_ON_ONCE(pwq->flush_color != -1);
3099 
3100 			if (pwq->nr_in_flight[flush_color]) {
3101 				pwq->flush_color = flush_color;
3102 				atomic_inc(&wq->nr_pwqs_to_flush);
3103 				wait = true;
3104 			}
3105 		}
3106 
3107 		if (work_color >= 0) {
3108 			WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3109 			pwq->work_color = work_color;
3110 		}
3111 
3112 		raw_spin_unlock_irq(&pool->lock);
3113 	}
3114 
3115 	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3116 		complete(&wq->first_flusher->done);
3117 
3118 	return wait;
3119 }
3120 
3121 /**
3122  * __flush_workqueue - ensure that any scheduled work has run to completion.
3123  * @wq: workqueue to flush
3124  *
3125  * This function sleeps until all work items which were queued on entry
3126  * have finished execution, but it is not livelocked by new incoming ones.
3127  */
__flush_workqueue(struct workqueue_struct * wq)3128 void __flush_workqueue(struct workqueue_struct *wq)
3129 {
3130 	struct wq_flusher this_flusher = {
3131 		.list = LIST_HEAD_INIT(this_flusher.list),
3132 		.flush_color = -1,
3133 		.done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3134 	};
3135 	int next_color;
3136 
3137 	if (WARN_ON(!wq_online))
3138 		return;
3139 
3140 	lock_map_acquire(&wq->lockdep_map);
3141 	lock_map_release(&wq->lockdep_map);
3142 
3143 	mutex_lock(&wq->mutex);
3144 
3145 	/*
3146 	 * Start-to-wait phase
3147 	 */
3148 	next_color = work_next_color(wq->work_color);
3149 
3150 	if (next_color != wq->flush_color) {
3151 		/*
3152 		 * Color space is not full.  The current work_color
3153 		 * becomes our flush_color and work_color is advanced
3154 		 * by one.
3155 		 */
3156 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3157 		this_flusher.flush_color = wq->work_color;
3158 		wq->work_color = next_color;
3159 
3160 		if (!wq->first_flusher) {
3161 			/* no flush in progress, become the first flusher */
3162 			WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3163 
3164 			wq->first_flusher = &this_flusher;
3165 
3166 			if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3167 						       wq->work_color)) {
3168 				/* nothing to flush, done */
3169 				wq->flush_color = next_color;
3170 				wq->first_flusher = NULL;
3171 				goto out_unlock;
3172 			}
3173 		} else {
3174 			/* wait in queue */
3175 			WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3176 			list_add_tail(&this_flusher.list, &wq->flusher_queue);
3177 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3178 		}
3179 	} else {
3180 		/*
3181 		 * Oops, color space is full, wait on overflow queue.
3182 		 * The next flush completion will assign us
3183 		 * flush_color and transfer to flusher_queue.
3184 		 */
3185 		list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3186 	}
3187 
3188 	check_flush_dependency(wq, NULL);
3189 
3190 	mutex_unlock(&wq->mutex);
3191 
3192 	wait_for_completion(&this_flusher.done);
3193 
3194 	/*
3195 	 * Wake-up-and-cascade phase
3196 	 *
3197 	 * First flushers are responsible for cascading flushes and
3198 	 * handling overflow.  Non-first flushers can simply return.
3199 	 */
3200 	if (READ_ONCE(wq->first_flusher) != &this_flusher)
3201 		return;
3202 
3203 	mutex_lock(&wq->mutex);
3204 
3205 	/* we might have raced, check again with mutex held */
3206 	if (wq->first_flusher != &this_flusher)
3207 		goto out_unlock;
3208 
3209 	WRITE_ONCE(wq->first_flusher, NULL);
3210 
3211 	WARN_ON_ONCE(!list_empty(&this_flusher.list));
3212 	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3213 
3214 	while (true) {
3215 		struct wq_flusher *next, *tmp;
3216 
3217 		/* complete all the flushers sharing the current flush color */
3218 		list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3219 			if (next->flush_color != wq->flush_color)
3220 				break;
3221 			list_del_init(&next->list);
3222 			complete(&next->done);
3223 		}
3224 
3225 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
3226 			     wq->flush_color != work_next_color(wq->work_color));
3227 
3228 		/* this flush_color is finished, advance by one */
3229 		wq->flush_color = work_next_color(wq->flush_color);
3230 
3231 		/* one color has been freed, handle overflow queue */
3232 		if (!list_empty(&wq->flusher_overflow)) {
3233 			/*
3234 			 * Assign the same color to all overflowed
3235 			 * flushers, advance work_color and append to
3236 			 * flusher_queue.  This is the start-to-wait
3237 			 * phase for these overflowed flushers.
3238 			 */
3239 			list_for_each_entry(tmp, &wq->flusher_overflow, list)
3240 				tmp->flush_color = wq->work_color;
3241 
3242 			wq->work_color = work_next_color(wq->work_color);
3243 
3244 			list_splice_tail_init(&wq->flusher_overflow,
3245 					      &wq->flusher_queue);
3246 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3247 		}
3248 
3249 		if (list_empty(&wq->flusher_queue)) {
3250 			WARN_ON_ONCE(wq->flush_color != wq->work_color);
3251 			break;
3252 		}
3253 
3254 		/*
3255 		 * Need to flush more colors.  Make the next flusher
3256 		 * the new first flusher and arm pwqs.
3257 		 */
3258 		WARN_ON_ONCE(wq->flush_color == wq->work_color);
3259 		WARN_ON_ONCE(wq->flush_color != next->flush_color);
3260 
3261 		list_del_init(&next->list);
3262 		wq->first_flusher = next;
3263 
3264 		if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
3265 			break;
3266 
3267 		/*
3268 		 * Meh... this color is already done, clear first
3269 		 * flusher and repeat cascading.
3270 		 */
3271 		wq->first_flusher = NULL;
3272 	}
3273 
3274 out_unlock:
3275 	mutex_unlock(&wq->mutex);
3276 }
3277 EXPORT_SYMBOL(__flush_workqueue);
3278 
3279 /**
3280  * drain_workqueue - drain a workqueue
3281  * @wq: workqueue to drain
3282  *
3283  * Wait until the workqueue becomes empty.  While draining is in progress,
3284  * only chain queueing is allowed.  IOW, only currently pending or running
3285  * work items on @wq can queue further work items on it.  @wq is flushed
3286  * repeatedly until it becomes empty.  The number of flushing is determined
3287  * by the depth of chaining and should be relatively short.  Whine if it
3288  * takes too long.
3289  */
drain_workqueue(struct workqueue_struct * wq)3290 void drain_workqueue(struct workqueue_struct *wq)
3291 {
3292 	unsigned int flush_cnt = 0;
3293 	struct pool_workqueue *pwq;
3294 
3295 	/*
3296 	 * __queue_work() needs to test whether there are drainers, is much
3297 	 * hotter than drain_workqueue() and already looks at @wq->flags.
3298 	 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
3299 	 */
3300 	mutex_lock(&wq->mutex);
3301 	if (!wq->nr_drainers++)
3302 		wq->flags |= __WQ_DRAINING;
3303 	mutex_unlock(&wq->mutex);
3304 reflush:
3305 	__flush_workqueue(wq);
3306 
3307 	mutex_lock(&wq->mutex);
3308 
3309 	for_each_pwq(pwq, wq) {
3310 		bool drained;
3311 
3312 		raw_spin_lock_irq(&pwq->pool->lock);
3313 		drained = !pwq->nr_active && list_empty(&pwq->inactive_works);
3314 		raw_spin_unlock_irq(&pwq->pool->lock);
3315 
3316 		if (drained)
3317 			continue;
3318 
3319 		if (++flush_cnt == 10 ||
3320 		    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3321 			pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
3322 				wq->name, __func__, flush_cnt);
3323 
3324 		mutex_unlock(&wq->mutex);
3325 		goto reflush;
3326 	}
3327 
3328 	if (!--wq->nr_drainers)
3329 		wq->flags &= ~__WQ_DRAINING;
3330 	mutex_unlock(&wq->mutex);
3331 }
3332 EXPORT_SYMBOL_GPL(drain_workqueue);
3333 
start_flush_work(struct work_struct * work,struct wq_barrier * barr,bool from_cancel)3334 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
3335 			     bool from_cancel)
3336 {
3337 	struct worker *worker = NULL;
3338 	struct worker_pool *pool;
3339 	struct pool_workqueue *pwq;
3340 
3341 	might_sleep();
3342 
3343 	rcu_read_lock();
3344 	pool = get_work_pool(work);
3345 	if (!pool) {
3346 		rcu_read_unlock();
3347 		return false;
3348 	}
3349 
3350 	raw_spin_lock_irq(&pool->lock);
3351 	/* see the comment in try_to_grab_pending() with the same code */
3352 	pwq = get_work_pwq(work);
3353 	if (pwq) {
3354 		if (unlikely(pwq->pool != pool))
3355 			goto already_gone;
3356 	} else {
3357 		worker = find_worker_executing_work(pool, work);
3358 		if (!worker)
3359 			goto already_gone;
3360 		pwq = worker->current_pwq;
3361 	}
3362 
3363 	check_flush_dependency(pwq->wq, work);
3364 
3365 	insert_wq_barrier(pwq, barr, work, worker);
3366 	raw_spin_unlock_irq(&pool->lock);
3367 
3368 	/*
3369 	 * Force a lock recursion deadlock when using flush_work() inside a
3370 	 * single-threaded or rescuer equipped workqueue.
3371 	 *
3372 	 * For single threaded workqueues the deadlock happens when the work
3373 	 * is after the work issuing the flush_work(). For rescuer equipped
3374 	 * workqueues the deadlock happens when the rescuer stalls, blocking
3375 	 * forward progress.
3376 	 */
3377 	if (!from_cancel &&
3378 	    (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3379 		lock_map_acquire(&pwq->wq->lockdep_map);
3380 		lock_map_release(&pwq->wq->lockdep_map);
3381 	}
3382 	rcu_read_unlock();
3383 	return true;
3384 already_gone:
3385 	raw_spin_unlock_irq(&pool->lock);
3386 	rcu_read_unlock();
3387 	return false;
3388 }
3389 
__flush_work(struct work_struct * work,bool from_cancel)3390 static bool __flush_work(struct work_struct *work, bool from_cancel)
3391 {
3392 	struct wq_barrier barr;
3393 
3394 	if (WARN_ON(!wq_online))
3395 		return false;
3396 
3397 	if (WARN_ON(!work->func))
3398 		return false;
3399 
3400 	lock_map_acquire(&work->lockdep_map);
3401 	lock_map_release(&work->lockdep_map);
3402 
3403 	if (start_flush_work(work, &barr, from_cancel)) {
3404 		wait_for_completion(&barr.done);
3405 		destroy_work_on_stack(&barr.work);
3406 		return true;
3407 	} else {
3408 		return false;
3409 	}
3410 }
3411 
3412 /**
3413  * flush_work - wait for a work to finish executing the last queueing instance
3414  * @work: the work to flush
3415  *
3416  * Wait until @work has finished execution.  @work is guaranteed to be idle
3417  * on return if it hasn't been requeued since flush started.
3418  *
3419  * Return:
3420  * %true if flush_work() waited for the work to finish execution,
3421  * %false if it was already idle.
3422  */
flush_work(struct work_struct * work)3423 bool flush_work(struct work_struct *work)
3424 {
3425 	return __flush_work(work, false);
3426 }
3427 EXPORT_SYMBOL_GPL(flush_work);
3428 
3429 struct cwt_wait {
3430 	wait_queue_entry_t		wait;
3431 	struct work_struct	*work;
3432 };
3433 
cwt_wakefn(wait_queue_entry_t * wait,unsigned mode,int sync,void * key)3434 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3435 {
3436 	struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3437 
3438 	if (cwait->work != key)
3439 		return 0;
3440 	return autoremove_wake_function(wait, mode, sync, key);
3441 }
3442 
__cancel_work_timer(struct work_struct * work,bool is_dwork)3443 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3444 {
3445 	static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3446 	unsigned long flags;
3447 	int ret;
3448 
3449 	do {
3450 		ret = try_to_grab_pending(work, is_dwork, &flags);
3451 		/*
3452 		 * If someone else is already canceling, wait for it to
3453 		 * finish.  flush_work() doesn't work for PREEMPT_NONE
3454 		 * because we may get scheduled between @work's completion
3455 		 * and the other canceling task resuming and clearing
3456 		 * CANCELING - flush_work() will return false immediately
3457 		 * as @work is no longer busy, try_to_grab_pending() will
3458 		 * return -ENOENT as @work is still being canceled and the
3459 		 * other canceling task won't be able to clear CANCELING as
3460 		 * we're hogging the CPU.
3461 		 *
3462 		 * Let's wait for completion using a waitqueue.  As this
3463 		 * may lead to the thundering herd problem, use a custom
3464 		 * wake function which matches @work along with exclusive
3465 		 * wait and wakeup.
3466 		 */
3467 		if (unlikely(ret == -ENOENT)) {
3468 			struct cwt_wait cwait;
3469 
3470 			init_wait(&cwait.wait);
3471 			cwait.wait.func = cwt_wakefn;
3472 			cwait.work = work;
3473 
3474 			prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3475 						  TASK_UNINTERRUPTIBLE);
3476 			if (work_is_canceling(work))
3477 				schedule();
3478 			finish_wait(&cancel_waitq, &cwait.wait);
3479 		}
3480 	} while (unlikely(ret < 0));
3481 
3482 	/* tell other tasks trying to grab @work to back off */
3483 	mark_work_canceling(work);
3484 	local_irq_restore(flags);
3485 
3486 	/*
3487 	 * This allows canceling during early boot.  We know that @work
3488 	 * isn't executing.
3489 	 */
3490 	if (wq_online)
3491 		__flush_work(work, true);
3492 
3493 	clear_work_data(work);
3494 
3495 	/*
3496 	 * Paired with prepare_to_wait() above so that either
3497 	 * waitqueue_active() is visible here or !work_is_canceling() is
3498 	 * visible there.
3499 	 */
3500 	smp_mb();
3501 	if (waitqueue_active(&cancel_waitq))
3502 		__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3503 
3504 	return ret;
3505 }
3506 
3507 /**
3508  * cancel_work_sync - cancel a work and wait for it to finish
3509  * @work: the work to cancel
3510  *
3511  * Cancel @work and wait for its execution to finish.  This function
3512  * can be used even if the work re-queues itself or migrates to
3513  * another workqueue.  On return from this function, @work is
3514  * guaranteed to be not pending or executing on any CPU.
3515  *
3516  * cancel_work_sync(&delayed_work->work) must not be used for
3517  * delayed_work's.  Use cancel_delayed_work_sync() instead.
3518  *
3519  * The caller must ensure that the workqueue on which @work was last
3520  * queued can't be destroyed before this function returns.
3521  *
3522  * Return:
3523  * %true if @work was pending, %false otherwise.
3524  */
cancel_work_sync(struct work_struct * work)3525 bool cancel_work_sync(struct work_struct *work)
3526 {
3527 	return __cancel_work_timer(work, false);
3528 }
3529 EXPORT_SYMBOL_GPL(cancel_work_sync);
3530 
3531 /**
3532  * flush_delayed_work - wait for a dwork to finish executing the last queueing
3533  * @dwork: the delayed work to flush
3534  *
3535  * Delayed timer is cancelled and the pending work is queued for
3536  * immediate execution.  Like flush_work(), this function only
3537  * considers the last queueing instance of @dwork.
3538  *
3539  * Return:
3540  * %true if flush_work() waited for the work to finish execution,
3541  * %false if it was already idle.
3542  */
flush_delayed_work(struct delayed_work * dwork)3543 bool flush_delayed_work(struct delayed_work *dwork)
3544 {
3545 	local_irq_disable();
3546 	if (del_timer_sync(&dwork->timer))
3547 		__queue_work(dwork->cpu, dwork->wq, &dwork->work);
3548 	local_irq_enable();
3549 	return flush_work(&dwork->work);
3550 }
3551 EXPORT_SYMBOL(flush_delayed_work);
3552 
3553 /**
3554  * flush_rcu_work - wait for a rwork to finish executing the last queueing
3555  * @rwork: the rcu work to flush
3556  *
3557  * Return:
3558  * %true if flush_rcu_work() waited for the work to finish execution,
3559  * %false if it was already idle.
3560  */
flush_rcu_work(struct rcu_work * rwork)3561 bool flush_rcu_work(struct rcu_work *rwork)
3562 {
3563 	if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3564 		rcu_barrier();
3565 		flush_work(&rwork->work);
3566 		return true;
3567 	} else {
3568 		return flush_work(&rwork->work);
3569 	}
3570 }
3571 EXPORT_SYMBOL(flush_rcu_work);
3572 
__cancel_work(struct work_struct * work,bool is_dwork)3573 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3574 {
3575 	unsigned long flags;
3576 	int ret;
3577 
3578 	do {
3579 		ret = try_to_grab_pending(work, is_dwork, &flags);
3580 	} while (unlikely(ret == -EAGAIN));
3581 
3582 	if (unlikely(ret < 0))
3583 		return false;
3584 
3585 	set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3586 	local_irq_restore(flags);
3587 	return ret;
3588 }
3589 
3590 /*
3591  * See cancel_delayed_work()
3592  */
cancel_work(struct work_struct * work)3593 bool cancel_work(struct work_struct *work)
3594 {
3595 	return __cancel_work(work, false);
3596 }
3597 EXPORT_SYMBOL(cancel_work);
3598 
3599 /**
3600  * cancel_delayed_work - cancel a delayed work
3601  * @dwork: delayed_work to cancel
3602  *
3603  * Kill off a pending delayed_work.
3604  *
3605  * Return: %true if @dwork was pending and canceled; %false if it wasn't
3606  * pending.
3607  *
3608  * Note:
3609  * The work callback function may still be running on return, unless
3610  * it returns %true and the work doesn't re-arm itself.  Explicitly flush or
3611  * use cancel_delayed_work_sync() to wait on it.
3612  *
3613  * This function is safe to call from any context including IRQ handler.
3614  */
cancel_delayed_work(struct delayed_work * dwork)3615 bool cancel_delayed_work(struct delayed_work *dwork)
3616 {
3617 	return __cancel_work(&dwork->work, true);
3618 }
3619 EXPORT_SYMBOL(cancel_delayed_work);
3620 
3621 /**
3622  * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3623  * @dwork: the delayed work cancel
3624  *
3625  * This is cancel_work_sync() for delayed works.
3626  *
3627  * Return:
3628  * %true if @dwork was pending, %false otherwise.
3629  */
cancel_delayed_work_sync(struct delayed_work * dwork)3630 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3631 {
3632 	return __cancel_work_timer(&dwork->work, true);
3633 }
3634 EXPORT_SYMBOL(cancel_delayed_work_sync);
3635 
3636 /**
3637  * schedule_on_each_cpu - execute a function synchronously on each online CPU
3638  * @func: the function to call
3639  *
3640  * schedule_on_each_cpu() executes @func on each online CPU using the
3641  * system workqueue and blocks until all CPUs have completed.
3642  * schedule_on_each_cpu() is very slow.
3643  *
3644  * Return:
3645  * 0 on success, -errno on failure.
3646  */
schedule_on_each_cpu(work_func_t func)3647 int schedule_on_each_cpu(work_func_t func)
3648 {
3649 	int cpu;
3650 	struct work_struct __percpu *works;
3651 
3652 	works = alloc_percpu(struct work_struct);
3653 	if (!works)
3654 		return -ENOMEM;
3655 
3656 	cpus_read_lock();
3657 
3658 	for_each_online_cpu(cpu) {
3659 		struct work_struct *work = per_cpu_ptr(works, cpu);
3660 
3661 		INIT_WORK(work, func);
3662 		schedule_work_on(cpu, work);
3663 	}
3664 
3665 	for_each_online_cpu(cpu)
3666 		flush_work(per_cpu_ptr(works, cpu));
3667 
3668 	cpus_read_unlock();
3669 	free_percpu(works);
3670 	return 0;
3671 }
3672 
3673 /**
3674  * execute_in_process_context - reliably execute the routine with user context
3675  * @fn:		the function to execute
3676  * @ew:		guaranteed storage for the execute work structure (must
3677  *		be available when the work executes)
3678  *
3679  * Executes the function immediately if process context is available,
3680  * otherwise schedules the function for delayed execution.
3681  *
3682  * Return:	0 - function was executed
3683  *		1 - function was scheduled for execution
3684  */
execute_in_process_context(work_func_t fn,struct execute_work * ew)3685 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3686 {
3687 	if (!in_interrupt()) {
3688 		fn(&ew->work);
3689 		return 0;
3690 	}
3691 
3692 	INIT_WORK(&ew->work, fn);
3693 	schedule_work(&ew->work);
3694 
3695 	return 1;
3696 }
3697 EXPORT_SYMBOL_GPL(execute_in_process_context);
3698 
3699 /**
3700  * free_workqueue_attrs - free a workqueue_attrs
3701  * @attrs: workqueue_attrs to free
3702  *
3703  * Undo alloc_workqueue_attrs().
3704  */
free_workqueue_attrs(struct workqueue_attrs * attrs)3705 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3706 {
3707 	if (attrs) {
3708 		free_cpumask_var(attrs->cpumask);
3709 		free_cpumask_var(attrs->__pod_cpumask);
3710 		kfree(attrs);
3711 	}
3712 }
3713 
3714 /**
3715  * alloc_workqueue_attrs - allocate a workqueue_attrs
3716  *
3717  * Allocate a new workqueue_attrs, initialize with default settings and
3718  * return it.
3719  *
3720  * Return: The allocated new workqueue_attr on success. %NULL on failure.
3721  */
alloc_workqueue_attrs(void)3722 struct workqueue_attrs *alloc_workqueue_attrs(void)
3723 {
3724 	struct workqueue_attrs *attrs;
3725 
3726 	attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3727 	if (!attrs)
3728 		goto fail;
3729 	if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3730 		goto fail;
3731 	if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
3732 		goto fail;
3733 
3734 	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3735 	attrs->affn_scope = WQ_AFFN_DFL;
3736 	return attrs;
3737 fail:
3738 	free_workqueue_attrs(attrs);
3739 	return NULL;
3740 }
3741 
copy_workqueue_attrs(struct workqueue_attrs * to,const struct workqueue_attrs * from)3742 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3743 				 const struct workqueue_attrs *from)
3744 {
3745 	to->nice = from->nice;
3746 	cpumask_copy(to->cpumask, from->cpumask);
3747 	cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
3748 	to->affn_strict = from->affn_strict;
3749 
3750 	/*
3751 	 * Unlike hash and equality test, copying shouldn't ignore wq-only
3752 	 * fields as copying is used for both pool and wq attrs. Instead,
3753 	 * get_unbound_pool() explicitly clears the fields.
3754 	 */
3755 	to->affn_scope = from->affn_scope;
3756 	to->ordered = from->ordered;
3757 }
3758 
3759 /*
3760  * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
3761  * comments in 'struct workqueue_attrs' definition.
3762  */
wqattrs_clear_for_pool(struct workqueue_attrs * attrs)3763 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
3764 {
3765 	attrs->affn_scope = WQ_AFFN_NR_TYPES;
3766 	attrs->ordered = false;
3767 }
3768 
3769 /* hash value of the content of @attr */
wqattrs_hash(const struct workqueue_attrs * attrs)3770 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3771 {
3772 	u32 hash = 0;
3773 
3774 	hash = jhash_1word(attrs->nice, hash);
3775 	hash = jhash(cpumask_bits(attrs->cpumask),
3776 		     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3777 	hash = jhash(cpumask_bits(attrs->__pod_cpumask),
3778 		     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3779 	hash = jhash_1word(attrs->affn_strict, hash);
3780 	return hash;
3781 }
3782 
3783 /* content equality test */
wqattrs_equal(const struct workqueue_attrs * a,const struct workqueue_attrs * b)3784 static bool wqattrs_equal(const struct workqueue_attrs *a,
3785 			  const struct workqueue_attrs *b)
3786 {
3787 	if (a->nice != b->nice)
3788 		return false;
3789 	if (!cpumask_equal(a->cpumask, b->cpumask))
3790 		return false;
3791 	if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
3792 		return false;
3793 	if (a->affn_strict != b->affn_strict)
3794 		return false;
3795 	return true;
3796 }
3797 
3798 /* Update @attrs with actually available CPUs */
wqattrs_actualize_cpumask(struct workqueue_attrs * attrs,const cpumask_t * unbound_cpumask)3799 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
3800 				      const cpumask_t *unbound_cpumask)
3801 {
3802 	/*
3803 	 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
3804 	 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
3805 	 * @unbound_cpumask.
3806 	 */
3807 	cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
3808 	if (unlikely(cpumask_empty(attrs->cpumask)))
3809 		cpumask_copy(attrs->cpumask, unbound_cpumask);
3810 }
3811 
3812 /* find wq_pod_type to use for @attrs */
3813 static const struct wq_pod_type *
wqattrs_pod_type(const struct workqueue_attrs * attrs)3814 wqattrs_pod_type(const struct workqueue_attrs *attrs)
3815 {
3816 	enum wq_affn_scope scope;
3817 	struct wq_pod_type *pt;
3818 
3819 	/* to synchronize access to wq_affn_dfl */
3820 	lockdep_assert_held(&wq_pool_mutex);
3821 
3822 	if (attrs->affn_scope == WQ_AFFN_DFL)
3823 		scope = wq_affn_dfl;
3824 	else
3825 		scope = attrs->affn_scope;
3826 
3827 	pt = &wq_pod_types[scope];
3828 
3829 	if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
3830 	    likely(pt->nr_pods))
3831 		return pt;
3832 
3833 	/*
3834 	 * Before workqueue_init_topology(), only SYSTEM is available which is
3835 	 * initialized in workqueue_init_early().
3836 	 */
3837 	pt = &wq_pod_types[WQ_AFFN_SYSTEM];
3838 	BUG_ON(!pt->nr_pods);
3839 	return pt;
3840 }
3841 
3842 /**
3843  * init_worker_pool - initialize a newly zalloc'd worker_pool
3844  * @pool: worker_pool to initialize
3845  *
3846  * Initialize a newly zalloc'd @pool.  It also allocates @pool->attrs.
3847  *
3848  * Return: 0 on success, -errno on failure.  Even on failure, all fields
3849  * inside @pool proper are initialized and put_unbound_pool() can be called
3850  * on @pool safely to release it.
3851  */
init_worker_pool(struct worker_pool * pool)3852 static int init_worker_pool(struct worker_pool *pool)
3853 {
3854 	raw_spin_lock_init(&pool->lock);
3855 	pool->id = -1;
3856 	pool->cpu = -1;
3857 	pool->node = NUMA_NO_NODE;
3858 	pool->flags |= POOL_DISASSOCIATED;
3859 	pool->watchdog_ts = jiffies;
3860 	INIT_LIST_HEAD(&pool->worklist);
3861 	INIT_LIST_HEAD(&pool->idle_list);
3862 	hash_init(pool->busy_hash);
3863 
3864 	timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3865 	INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
3866 
3867 	timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3868 
3869 	INIT_LIST_HEAD(&pool->workers);
3870 	INIT_LIST_HEAD(&pool->dying_workers);
3871 
3872 	ida_init(&pool->worker_ida);
3873 	INIT_HLIST_NODE(&pool->hash_node);
3874 	pool->refcnt = 1;
3875 
3876 	/* shouldn't fail above this point */
3877 	pool->attrs = alloc_workqueue_attrs();
3878 	if (!pool->attrs)
3879 		return -ENOMEM;
3880 
3881 	wqattrs_clear_for_pool(pool->attrs);
3882 
3883 	return 0;
3884 }
3885 
3886 #ifdef CONFIG_LOCKDEP
wq_init_lockdep(struct workqueue_struct * wq)3887 static void wq_init_lockdep(struct workqueue_struct *wq)
3888 {
3889 	char *lock_name;
3890 
3891 	lockdep_register_key(&wq->key);
3892 	lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3893 	if (!lock_name)
3894 		lock_name = wq->name;
3895 
3896 	wq->lock_name = lock_name;
3897 	lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3898 }
3899 
wq_unregister_lockdep(struct workqueue_struct * wq)3900 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3901 {
3902 	lockdep_unregister_key(&wq->key);
3903 }
3904 
wq_free_lockdep(struct workqueue_struct * wq)3905 static void wq_free_lockdep(struct workqueue_struct *wq)
3906 {
3907 	if (wq->lock_name != wq->name)
3908 		kfree(wq->lock_name);
3909 }
3910 #else
wq_init_lockdep(struct workqueue_struct * wq)3911 static void wq_init_lockdep(struct workqueue_struct *wq)
3912 {
3913 }
3914 
wq_unregister_lockdep(struct workqueue_struct * wq)3915 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3916 {
3917 }
3918 
wq_free_lockdep(struct workqueue_struct * wq)3919 static void wq_free_lockdep(struct workqueue_struct *wq)
3920 {
3921 }
3922 #endif
3923 
rcu_free_wq(struct rcu_head * rcu)3924 static void rcu_free_wq(struct rcu_head *rcu)
3925 {
3926 	struct workqueue_struct *wq =
3927 		container_of(rcu, struct workqueue_struct, rcu);
3928 
3929 	wq_free_lockdep(wq);
3930 	free_percpu(wq->cpu_pwq);
3931 	free_workqueue_attrs(wq->unbound_attrs);
3932 	kfree(wq);
3933 }
3934 
rcu_free_pool(struct rcu_head * rcu)3935 static void rcu_free_pool(struct rcu_head *rcu)
3936 {
3937 	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3938 
3939 	ida_destroy(&pool->worker_ida);
3940 	free_workqueue_attrs(pool->attrs);
3941 	kfree(pool);
3942 }
3943 
3944 /**
3945  * put_unbound_pool - put a worker_pool
3946  * @pool: worker_pool to put
3947  *
3948  * Put @pool.  If its refcnt reaches zero, it gets destroyed in RCU
3949  * safe manner.  get_unbound_pool() calls this function on its failure path
3950  * and this function should be able to release pools which went through,
3951  * successfully or not, init_worker_pool().
3952  *
3953  * Should be called with wq_pool_mutex held.
3954  */
put_unbound_pool(struct worker_pool * pool)3955 static void put_unbound_pool(struct worker_pool *pool)
3956 {
3957 	DECLARE_COMPLETION_ONSTACK(detach_completion);
3958 	struct worker *worker;
3959 	LIST_HEAD(cull_list);
3960 
3961 	lockdep_assert_held(&wq_pool_mutex);
3962 
3963 	if (--pool->refcnt)
3964 		return;
3965 
3966 	/* sanity checks */
3967 	if (WARN_ON(!(pool->cpu < 0)) ||
3968 	    WARN_ON(!list_empty(&pool->worklist)))
3969 		return;
3970 
3971 	/* release id and unhash */
3972 	if (pool->id >= 0)
3973 		idr_remove(&worker_pool_idr, pool->id);
3974 	hash_del(&pool->hash_node);
3975 
3976 	/*
3977 	 * Become the manager and destroy all workers.  This prevents
3978 	 * @pool's workers from blocking on attach_mutex.  We're the last
3979 	 * manager and @pool gets freed with the flag set.
3980 	 *
3981 	 * Having a concurrent manager is quite unlikely to happen as we can
3982 	 * only get here with
3983 	 *   pwq->refcnt == pool->refcnt == 0
3984 	 * which implies no work queued to the pool, which implies no worker can
3985 	 * become the manager. However a worker could have taken the role of
3986 	 * manager before the refcnts dropped to 0, since maybe_create_worker()
3987 	 * drops pool->lock
3988 	 */
3989 	while (true) {
3990 		rcuwait_wait_event(&manager_wait,
3991 				   !(pool->flags & POOL_MANAGER_ACTIVE),
3992 				   TASK_UNINTERRUPTIBLE);
3993 
3994 		mutex_lock(&wq_pool_attach_mutex);
3995 		raw_spin_lock_irq(&pool->lock);
3996 		if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
3997 			pool->flags |= POOL_MANAGER_ACTIVE;
3998 			break;
3999 		}
4000 		raw_spin_unlock_irq(&pool->lock);
4001 		mutex_unlock(&wq_pool_attach_mutex);
4002 	}
4003 
4004 	while ((worker = first_idle_worker(pool)))
4005 		set_worker_dying(worker, &cull_list);
4006 	WARN_ON(pool->nr_workers || pool->nr_idle);
4007 	raw_spin_unlock_irq(&pool->lock);
4008 
4009 	wake_dying_workers(&cull_list);
4010 
4011 	if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
4012 		pool->detach_completion = &detach_completion;
4013 	mutex_unlock(&wq_pool_attach_mutex);
4014 
4015 	if (pool->detach_completion)
4016 		wait_for_completion(pool->detach_completion);
4017 
4018 	/* shut down the timers */
4019 	del_timer_sync(&pool->idle_timer);
4020 	cancel_work_sync(&pool->idle_cull_work);
4021 	del_timer_sync(&pool->mayday_timer);
4022 
4023 	/* RCU protected to allow dereferences from get_work_pool() */
4024 	call_rcu(&pool->rcu, rcu_free_pool);
4025 }
4026 
4027 /**
4028  * get_unbound_pool - get a worker_pool with the specified attributes
4029  * @attrs: the attributes of the worker_pool to get
4030  *
4031  * Obtain a worker_pool which has the same attributes as @attrs, bump the
4032  * reference count and return it.  If there already is a matching
4033  * worker_pool, it will be used; otherwise, this function attempts to
4034  * create a new one.
4035  *
4036  * Should be called with wq_pool_mutex held.
4037  *
4038  * Return: On success, a worker_pool with the same attributes as @attrs.
4039  * On failure, %NULL.
4040  */
get_unbound_pool(const struct workqueue_attrs * attrs)4041 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4042 {
4043 	struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4044 	u32 hash = wqattrs_hash(attrs);
4045 	struct worker_pool *pool;
4046 	int pod, node = NUMA_NO_NODE;
4047 
4048 	lockdep_assert_held(&wq_pool_mutex);
4049 
4050 	/* do we already have a matching pool? */
4051 	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4052 		if (wqattrs_equal(pool->attrs, attrs)) {
4053 			pool->refcnt++;
4054 			return pool;
4055 		}
4056 	}
4057 
4058 	/* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4059 	for (pod = 0; pod < pt->nr_pods; pod++) {
4060 		if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4061 			node = pt->pod_node[pod];
4062 			break;
4063 		}
4064 	}
4065 
4066 	/* nope, create a new one */
4067 	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4068 	if (!pool || init_worker_pool(pool) < 0)
4069 		goto fail;
4070 
4071 	pool->node = node;
4072 	copy_workqueue_attrs(pool->attrs, attrs);
4073 	wqattrs_clear_for_pool(pool->attrs);
4074 
4075 	if (worker_pool_assign_id(pool) < 0)
4076 		goto fail;
4077 
4078 	/* create and start the initial worker */
4079 	if (wq_online && !create_worker(pool))
4080 		goto fail;
4081 
4082 	/* install */
4083 	hash_add(unbound_pool_hash, &pool->hash_node, hash);
4084 
4085 	return pool;
4086 fail:
4087 	if (pool)
4088 		put_unbound_pool(pool);
4089 	return NULL;
4090 }
4091 
rcu_free_pwq(struct rcu_head * rcu)4092 static void rcu_free_pwq(struct rcu_head *rcu)
4093 {
4094 	kmem_cache_free(pwq_cache,
4095 			container_of(rcu, struct pool_workqueue, rcu));
4096 }
4097 
4098 /*
4099  * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
4100  * refcnt and needs to be destroyed.
4101  */
pwq_release_workfn(struct kthread_work * work)4102 static void pwq_release_workfn(struct kthread_work *work)
4103 {
4104 	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
4105 						  release_work);
4106 	struct workqueue_struct *wq = pwq->wq;
4107 	struct worker_pool *pool = pwq->pool;
4108 	bool is_last = false;
4109 
4110 	/*
4111 	 * When @pwq is not linked, it doesn't hold any reference to the
4112 	 * @wq, and @wq is invalid to access.
4113 	 */
4114 	if (!list_empty(&pwq->pwqs_node)) {
4115 		mutex_lock(&wq->mutex);
4116 		list_del_rcu(&pwq->pwqs_node);
4117 		is_last = list_empty(&wq->pwqs);
4118 		mutex_unlock(&wq->mutex);
4119 	}
4120 
4121 	if (wq->flags & WQ_UNBOUND) {
4122 		mutex_lock(&wq_pool_mutex);
4123 		put_unbound_pool(pool);
4124 		mutex_unlock(&wq_pool_mutex);
4125 	}
4126 
4127 	call_rcu(&pwq->rcu, rcu_free_pwq);
4128 
4129 	/*
4130 	 * If we're the last pwq going away, @wq is already dead and no one
4131 	 * is gonna access it anymore.  Schedule RCU free.
4132 	 */
4133 	if (is_last) {
4134 		wq_unregister_lockdep(wq);
4135 		call_rcu(&wq->rcu, rcu_free_wq);
4136 	}
4137 }
4138 
4139 /**
4140  * pwq_adjust_max_active - update a pwq's max_active to the current setting
4141  * @pwq: target pool_workqueue
4142  *
4143  * If @pwq isn't freezing, set @pwq->max_active to the associated
4144  * workqueue's saved_max_active and activate inactive work items
4145  * accordingly.  If @pwq is freezing, clear @pwq->max_active to zero.
4146  */
pwq_adjust_max_active(struct pool_workqueue * pwq)4147 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
4148 {
4149 	struct workqueue_struct *wq = pwq->wq;
4150 	bool freezable = wq->flags & WQ_FREEZABLE;
4151 	unsigned long flags;
4152 
4153 	/* for @wq->saved_max_active */
4154 	lockdep_assert_held(&wq->mutex);
4155 
4156 	/* fast exit for non-freezable wqs */
4157 	if (!freezable && pwq->max_active == wq->saved_max_active)
4158 		return;
4159 
4160 	/* this function can be called during early boot w/ irq disabled */
4161 	raw_spin_lock_irqsave(&pwq->pool->lock, flags);
4162 
4163 	/*
4164 	 * During [un]freezing, the caller is responsible for ensuring that
4165 	 * this function is called at least once after @workqueue_freezing
4166 	 * is updated and visible.
4167 	 */
4168 	if (!freezable || !workqueue_freezing) {
4169 		pwq->max_active = wq->saved_max_active;
4170 
4171 		while (!list_empty(&pwq->inactive_works) &&
4172 		       pwq->nr_active < pwq->max_active)
4173 			pwq_activate_first_inactive(pwq);
4174 
4175 		kick_pool(pwq->pool);
4176 	} else {
4177 		pwq->max_active = 0;
4178 	}
4179 
4180 	raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
4181 }
4182 
4183 /* initialize newly allocated @pwq which is associated with @wq and @pool */
init_pwq(struct pool_workqueue * pwq,struct workqueue_struct * wq,struct worker_pool * pool)4184 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
4185 		     struct worker_pool *pool)
4186 {
4187 	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
4188 
4189 	memset(pwq, 0, sizeof(*pwq));
4190 
4191 	pwq->pool = pool;
4192 	pwq->wq = wq;
4193 	pwq->flush_color = -1;
4194 	pwq->refcnt = 1;
4195 	INIT_LIST_HEAD(&pwq->inactive_works);
4196 	INIT_LIST_HEAD(&pwq->pwqs_node);
4197 	INIT_LIST_HEAD(&pwq->mayday_node);
4198 	kthread_init_work(&pwq->release_work, pwq_release_workfn);
4199 }
4200 
4201 /* sync @pwq with the current state of its associated wq and link it */
link_pwq(struct pool_workqueue * pwq)4202 static void link_pwq(struct pool_workqueue *pwq)
4203 {
4204 	struct workqueue_struct *wq = pwq->wq;
4205 
4206 	lockdep_assert_held(&wq->mutex);
4207 
4208 	/* may be called multiple times, ignore if already linked */
4209 	if (!list_empty(&pwq->pwqs_node))
4210 		return;
4211 
4212 	/* set the matching work_color */
4213 	pwq->work_color = wq->work_color;
4214 
4215 	/* sync max_active to the current setting */
4216 	pwq_adjust_max_active(pwq);
4217 
4218 	/* link in @pwq */
4219 	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
4220 }
4221 
4222 /* 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)4223 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
4224 					const struct workqueue_attrs *attrs)
4225 {
4226 	struct worker_pool *pool;
4227 	struct pool_workqueue *pwq;
4228 
4229 	lockdep_assert_held(&wq_pool_mutex);
4230 
4231 	pool = get_unbound_pool(attrs);
4232 	if (!pool)
4233 		return NULL;
4234 
4235 	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
4236 	if (!pwq) {
4237 		put_unbound_pool(pool);
4238 		return NULL;
4239 	}
4240 
4241 	init_pwq(pwq, wq, pool);
4242 	return pwq;
4243 }
4244 
4245 /**
4246  * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
4247  * @attrs: the wq_attrs of the default pwq of the target workqueue
4248  * @cpu: the target CPU
4249  * @cpu_going_down: if >= 0, the CPU to consider as offline
4250  *
4251  * Calculate the cpumask a workqueue with @attrs should use on @pod. If
4252  * @cpu_going_down is >= 0, that cpu is considered offline during calculation.
4253  * The result is stored in @attrs->__pod_cpumask.
4254  *
4255  * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
4256  * and @pod has online CPUs requested by @attrs, the returned cpumask is the
4257  * intersection of the possible CPUs of @pod and @attrs->cpumask.
4258  *
4259  * The caller is responsible for ensuring that the cpumask of @pod stays stable.
4260  */
wq_calc_pod_cpumask(struct workqueue_attrs * attrs,int cpu,int cpu_going_down)4261 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
4262 				int cpu_going_down)
4263 {
4264 	const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
4265 	int pod = pt->cpu_pod[cpu];
4266 
4267 	/* does @pod have any online CPUs @attrs wants? */
4268 	cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
4269 	cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
4270 	if (cpu_going_down >= 0)
4271 		cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
4272 
4273 	if (cpumask_empty(attrs->__pod_cpumask)) {
4274 		cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
4275 		return;
4276 	}
4277 
4278 	/* yeap, return possible CPUs in @pod that @attrs wants */
4279 	cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
4280 
4281 	if (cpumask_empty(attrs->__pod_cpumask))
4282 		pr_warn_once("WARNING: workqueue cpumask: online intersect > "
4283 				"possible intersect\n");
4284 }
4285 
4286 /* install @pwq into @wq's cpu_pwq and return the old pwq */
install_unbound_pwq(struct workqueue_struct * wq,int cpu,struct pool_workqueue * pwq)4287 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
4288 					int cpu, struct pool_workqueue *pwq)
4289 {
4290 	struct pool_workqueue *old_pwq;
4291 
4292 	lockdep_assert_held(&wq_pool_mutex);
4293 	lockdep_assert_held(&wq->mutex);
4294 
4295 	/* link_pwq() can handle duplicate calls */
4296 	link_pwq(pwq);
4297 
4298 	old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
4299 	rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq);
4300 	return old_pwq;
4301 }
4302 
4303 /* context to store the prepared attrs & pwqs before applying */
4304 struct apply_wqattrs_ctx {
4305 	struct workqueue_struct	*wq;		/* target workqueue */
4306 	struct workqueue_attrs	*attrs;		/* attrs to apply */
4307 	struct list_head	list;		/* queued for batching commit */
4308 	struct pool_workqueue	*dfl_pwq;
4309 	struct pool_workqueue	*pwq_tbl[];
4310 };
4311 
4312 /* free the resources after success or abort */
apply_wqattrs_cleanup(struct apply_wqattrs_ctx * ctx)4313 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
4314 {
4315 	if (ctx) {
4316 		int cpu;
4317 
4318 		for_each_possible_cpu(cpu)
4319 			put_pwq_unlocked(ctx->pwq_tbl[cpu]);
4320 		put_pwq_unlocked(ctx->dfl_pwq);
4321 
4322 		free_workqueue_attrs(ctx->attrs);
4323 
4324 		kfree(ctx);
4325 	}
4326 }
4327 
4328 /* allocate the attrs and pwqs for later installation */
4329 static struct apply_wqattrs_ctx *
apply_wqattrs_prepare(struct workqueue_struct * wq,const struct workqueue_attrs * attrs,const cpumask_var_t unbound_cpumask)4330 apply_wqattrs_prepare(struct workqueue_struct *wq,
4331 		      const struct workqueue_attrs *attrs,
4332 		      const cpumask_var_t unbound_cpumask)
4333 {
4334 	struct apply_wqattrs_ctx *ctx;
4335 	struct workqueue_attrs *new_attrs;
4336 	int cpu;
4337 
4338 	lockdep_assert_held(&wq_pool_mutex);
4339 
4340 	if (WARN_ON(attrs->affn_scope < 0 ||
4341 		    attrs->affn_scope >= WQ_AFFN_NR_TYPES))
4342 		return ERR_PTR(-EINVAL);
4343 
4344 	ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
4345 
4346 	new_attrs = alloc_workqueue_attrs();
4347 	if (!ctx || !new_attrs)
4348 		goto out_free;
4349 
4350 	/*
4351 	 * If something goes wrong during CPU up/down, we'll fall back to
4352 	 * the default pwq covering whole @attrs->cpumask.  Always create
4353 	 * it even if we don't use it immediately.
4354 	 */
4355 	copy_workqueue_attrs(new_attrs, attrs);
4356 	wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
4357 	cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4358 	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
4359 	if (!ctx->dfl_pwq)
4360 		goto out_free;
4361 
4362 	for_each_possible_cpu(cpu) {
4363 		if (new_attrs->ordered) {
4364 			ctx->dfl_pwq->refcnt++;
4365 			ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
4366 		} else {
4367 			wq_calc_pod_cpumask(new_attrs, cpu, -1);
4368 			ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
4369 			if (!ctx->pwq_tbl[cpu])
4370 				goto out_free;
4371 		}
4372 	}
4373 
4374 	/* save the user configured attrs and sanitize it. */
4375 	copy_workqueue_attrs(new_attrs, attrs);
4376 	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
4377 	cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4378 	ctx->attrs = new_attrs;
4379 
4380 	ctx->wq = wq;
4381 	return ctx;
4382 
4383 out_free:
4384 	free_workqueue_attrs(new_attrs);
4385 	apply_wqattrs_cleanup(ctx);
4386 	return ERR_PTR(-ENOMEM);
4387 }
4388 
4389 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
apply_wqattrs_commit(struct apply_wqattrs_ctx * ctx)4390 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4391 {
4392 	int cpu;
4393 
4394 	/* all pwqs have been created successfully, let's install'em */
4395 	mutex_lock(&ctx->wq->mutex);
4396 
4397 	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4398 
4399 	/* save the previous pwq and install the new one */
4400 	for_each_possible_cpu(cpu)
4401 		ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
4402 							ctx->pwq_tbl[cpu]);
4403 
4404 	/* @dfl_pwq might not have been used, ensure it's linked */
4405 	link_pwq(ctx->dfl_pwq);
4406 	swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
4407 
4408 	mutex_unlock(&ctx->wq->mutex);
4409 }
4410 
apply_wqattrs_lock(void)4411 static void apply_wqattrs_lock(void)
4412 {
4413 	/* CPUs should stay stable across pwq creations and installations */
4414 	cpus_read_lock();
4415 	mutex_lock(&wq_pool_mutex);
4416 }
4417 
apply_wqattrs_unlock(void)4418 static void apply_wqattrs_unlock(void)
4419 {
4420 	mutex_unlock(&wq_pool_mutex);
4421 	cpus_read_unlock();
4422 }
4423 
apply_workqueue_attrs_locked(struct workqueue_struct * wq,const struct workqueue_attrs * attrs)4424 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4425 					const struct workqueue_attrs *attrs)
4426 {
4427 	struct apply_wqattrs_ctx *ctx;
4428 
4429 	/* only unbound workqueues can change attributes */
4430 	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4431 		return -EINVAL;
4432 
4433 	/* creating multiple pwqs breaks ordering guarantee */
4434 	if (!list_empty(&wq->pwqs)) {
4435 		if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4436 			return -EINVAL;
4437 
4438 		wq->flags &= ~__WQ_ORDERED;
4439 	}
4440 
4441 	ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
4442 	if (IS_ERR(ctx))
4443 		return PTR_ERR(ctx);
4444 
4445 	/* the ctx has been prepared successfully, let's commit it */
4446 	apply_wqattrs_commit(ctx);
4447 	apply_wqattrs_cleanup(ctx);
4448 
4449 	return 0;
4450 }
4451 
4452 /**
4453  * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4454  * @wq: the target workqueue
4455  * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4456  *
4457  * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
4458  * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
4459  * work items are affine to the pod it was issued on. Older pwqs are released as
4460  * in-flight work items finish. Note that a work item which repeatedly requeues
4461  * itself back-to-back will stay on its current pwq.
4462  *
4463  * Performs GFP_KERNEL allocations.
4464  *
4465  * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
4466  *
4467  * Return: 0 on success and -errno on failure.
4468  */
apply_workqueue_attrs(struct workqueue_struct * wq,const struct workqueue_attrs * attrs)4469 int apply_workqueue_attrs(struct workqueue_struct *wq,
4470 			  const struct workqueue_attrs *attrs)
4471 {
4472 	int ret;
4473 
4474 	lockdep_assert_cpus_held();
4475 
4476 	mutex_lock(&wq_pool_mutex);
4477 	ret = apply_workqueue_attrs_locked(wq, attrs);
4478 	mutex_unlock(&wq_pool_mutex);
4479 
4480 	return ret;
4481 }
4482 
4483 /**
4484  * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
4485  * @wq: the target workqueue
4486  * @cpu: the CPU to update pool association for
4487  * @hotplug_cpu: the CPU coming up or going down
4488  * @online: whether @cpu is coming up or going down
4489  *
4490  * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4491  * %CPU_DOWN_FAILED.  @cpu is being hot[un]plugged, update pod affinity of
4492  * @wq accordingly.
4493  *
4494  *
4495  * If pod affinity can't be adjusted due to memory allocation failure, it falls
4496  * back to @wq->dfl_pwq which may not be optimal but is always correct.
4497  *
4498  * Note that when the last allowed CPU of a pod goes offline for a workqueue
4499  * with a cpumask spanning multiple pods, the workers which were already
4500  * executing the work items for the workqueue will lose their CPU affinity and
4501  * may execute on any CPU. This is similar to how per-cpu workqueues behave on
4502  * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
4503  * responsibility to flush the work item from CPU_DOWN_PREPARE.
4504  */
wq_update_pod(struct workqueue_struct * wq,int cpu,int hotplug_cpu,bool online)4505 static void wq_update_pod(struct workqueue_struct *wq, int cpu,
4506 			  int hotplug_cpu, bool online)
4507 {
4508 	int off_cpu = online ? -1 : hotplug_cpu;
4509 	struct pool_workqueue *old_pwq = NULL, *pwq;
4510 	struct workqueue_attrs *target_attrs;
4511 
4512 	lockdep_assert_held(&wq_pool_mutex);
4513 
4514 	if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
4515 		return;
4516 
4517 	/*
4518 	 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4519 	 * Let's use a preallocated one.  The following buf is protected by
4520 	 * CPU hotplug exclusion.
4521 	 */
4522 	target_attrs = wq_update_pod_attrs_buf;
4523 
4524 	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4525 	wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
4526 
4527 	/* nothing to do if the target cpumask matches the current pwq */
4528 	wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
4529 	pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu),
4530 					lockdep_is_held(&wq_pool_mutex));
4531 	if (wqattrs_equal(target_attrs, pwq->pool->attrs))
4532 		return;
4533 
4534 	/* create a new pwq */
4535 	pwq = alloc_unbound_pwq(wq, target_attrs);
4536 	if (!pwq) {
4537 		pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
4538 			wq->name);
4539 		goto use_dfl_pwq;
4540 	}
4541 
4542 	/* Install the new pwq. */
4543 	mutex_lock(&wq->mutex);
4544 	old_pwq = install_unbound_pwq(wq, cpu, pwq);
4545 	goto out_unlock;
4546 
4547 use_dfl_pwq:
4548 	mutex_lock(&wq->mutex);
4549 	raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
4550 	get_pwq(wq->dfl_pwq);
4551 	raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4552 	old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq);
4553 out_unlock:
4554 	mutex_unlock(&wq->mutex);
4555 	put_pwq_unlocked(old_pwq);
4556 }
4557 
alloc_and_link_pwqs(struct workqueue_struct * wq)4558 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4559 {
4560 	bool highpri = wq->flags & WQ_HIGHPRI;
4561 	int cpu, ret;
4562 
4563 	wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
4564 	if (!wq->cpu_pwq)
4565 		goto enomem;
4566 
4567 	if (!(wq->flags & WQ_UNBOUND)) {
4568 		for_each_possible_cpu(cpu) {
4569 			struct pool_workqueue **pwq_p =
4570 				per_cpu_ptr(wq->cpu_pwq, cpu);
4571 			struct worker_pool *pool =
4572 				&(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]);
4573 
4574 			*pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
4575 						       pool->node);
4576 			if (!*pwq_p)
4577 				goto enomem;
4578 
4579 			init_pwq(*pwq_p, wq, pool);
4580 
4581 			mutex_lock(&wq->mutex);
4582 			link_pwq(*pwq_p);
4583 			mutex_unlock(&wq->mutex);
4584 		}
4585 		return 0;
4586 	}
4587 
4588 	cpus_read_lock();
4589 	if (wq->flags & __WQ_ORDERED) {
4590 		ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4591 		/* there should only be single pwq for ordering guarantee */
4592 		WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4593 			      wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4594 		     "ordering guarantee broken for workqueue %s\n", wq->name);
4595 	} else {
4596 		ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4597 	}
4598 	cpus_read_unlock();
4599 
4600 	/* for unbound pwq, flush the pwq_release_worker ensures that the
4601 	 * pwq_release_workfn() completes before calling kfree(wq).
4602 	 */
4603 	if (ret)
4604 		kthread_flush_worker(pwq_release_worker);
4605 
4606 	return ret;
4607 
4608 enomem:
4609 	if (wq->cpu_pwq) {
4610 		for_each_possible_cpu(cpu) {
4611 			struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
4612 
4613 			if (pwq)
4614 				kmem_cache_free(pwq_cache, pwq);
4615 		}
4616 		free_percpu(wq->cpu_pwq);
4617 		wq->cpu_pwq = NULL;
4618 	}
4619 	return -ENOMEM;
4620 }
4621 
wq_clamp_max_active(int max_active,unsigned int flags,const char * name)4622 static int wq_clamp_max_active(int max_active, unsigned int flags,
4623 			       const char *name)
4624 {
4625 	if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
4626 		pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4627 			max_active, name, 1, WQ_MAX_ACTIVE);
4628 
4629 	return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
4630 }
4631 
4632 /*
4633  * Workqueues which may be used during memory reclaim should have a rescuer
4634  * to guarantee forward progress.
4635  */
init_rescuer(struct workqueue_struct * wq)4636 static int init_rescuer(struct workqueue_struct *wq)
4637 {
4638 	struct worker *rescuer;
4639 	int ret;
4640 
4641 	if (!(wq->flags & WQ_MEM_RECLAIM))
4642 		return 0;
4643 
4644 	rescuer = alloc_worker(NUMA_NO_NODE);
4645 	if (!rescuer) {
4646 		pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
4647 		       wq->name);
4648 		return -ENOMEM;
4649 	}
4650 
4651 	rescuer->rescue_wq = wq;
4652 	rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
4653 	if (IS_ERR(rescuer->task)) {
4654 		ret = PTR_ERR(rescuer->task);
4655 		pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
4656 		       wq->name, ERR_PTR(ret));
4657 		kfree(rescuer);
4658 		return ret;
4659 	}
4660 
4661 	wq->rescuer = rescuer;
4662 	kthread_bind_mask(rescuer->task, cpu_possible_mask);
4663 	wake_up_process(rescuer->task);
4664 
4665 	return 0;
4666 }
4667 
4668 __printf(1, 4)
alloc_workqueue(const char * fmt,unsigned int flags,int max_active,...)4669 struct workqueue_struct *alloc_workqueue(const char *fmt,
4670 					 unsigned int flags,
4671 					 int max_active, ...)
4672 {
4673 	va_list args;
4674 	struct workqueue_struct *wq;
4675 	struct pool_workqueue *pwq;
4676 
4677 	/*
4678 	 * Unbound && max_active == 1 used to imply ordered, which is no longer
4679 	 * the case on many machines due to per-pod pools. While
4680 	 * alloc_ordered_workqueue() is the right way to create an ordered
4681 	 * workqueue, keep the previous behavior to avoid subtle breakages.
4682 	 */
4683 	if ((flags & WQ_UNBOUND) && max_active == 1)
4684 		flags |= __WQ_ORDERED;
4685 
4686 	/* see the comment above the definition of WQ_POWER_EFFICIENT */
4687 	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4688 		flags |= WQ_UNBOUND;
4689 
4690 	/* allocate wq and format name */
4691 	wq = kzalloc(sizeof(*wq), GFP_KERNEL);
4692 	if (!wq)
4693 		return NULL;
4694 
4695 	if (flags & WQ_UNBOUND) {
4696 		wq->unbound_attrs = alloc_workqueue_attrs();
4697 		if (!wq->unbound_attrs)
4698 			goto err_free_wq;
4699 	}
4700 
4701 	va_start(args, max_active);
4702 	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4703 	va_end(args);
4704 
4705 	max_active = max_active ?: WQ_DFL_ACTIVE;
4706 	max_active = wq_clamp_max_active(max_active, flags, wq->name);
4707 
4708 	/* init wq */
4709 	wq->flags = flags;
4710 	wq->saved_max_active = max_active;
4711 	mutex_init(&wq->mutex);
4712 	atomic_set(&wq->nr_pwqs_to_flush, 0);
4713 	INIT_LIST_HEAD(&wq->pwqs);
4714 	INIT_LIST_HEAD(&wq->flusher_queue);
4715 	INIT_LIST_HEAD(&wq->flusher_overflow);
4716 	INIT_LIST_HEAD(&wq->maydays);
4717 
4718 	wq_init_lockdep(wq);
4719 	INIT_LIST_HEAD(&wq->list);
4720 
4721 	if (alloc_and_link_pwqs(wq) < 0)
4722 		goto err_unreg_lockdep;
4723 
4724 	if (wq_online && init_rescuer(wq) < 0)
4725 		goto err_destroy;
4726 
4727 	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4728 		goto err_destroy;
4729 
4730 	/*
4731 	 * wq_pool_mutex protects global freeze state and workqueues list.
4732 	 * Grab it, adjust max_active and add the new @wq to workqueues
4733 	 * list.
4734 	 */
4735 	mutex_lock(&wq_pool_mutex);
4736 
4737 	mutex_lock(&wq->mutex);
4738 	for_each_pwq(pwq, wq)
4739 		pwq_adjust_max_active(pwq);
4740 	mutex_unlock(&wq->mutex);
4741 
4742 	list_add_tail_rcu(&wq->list, &workqueues);
4743 
4744 	mutex_unlock(&wq_pool_mutex);
4745 
4746 	return wq;
4747 
4748 err_unreg_lockdep:
4749 	wq_unregister_lockdep(wq);
4750 	wq_free_lockdep(wq);
4751 err_free_wq:
4752 	free_workqueue_attrs(wq->unbound_attrs);
4753 	kfree(wq);
4754 	return NULL;
4755 err_destroy:
4756 	destroy_workqueue(wq);
4757 	return NULL;
4758 }
4759 EXPORT_SYMBOL_GPL(alloc_workqueue);
4760 
pwq_busy(struct pool_workqueue * pwq)4761 static bool pwq_busy(struct pool_workqueue *pwq)
4762 {
4763 	int i;
4764 
4765 	for (i = 0; i < WORK_NR_COLORS; i++)
4766 		if (pwq->nr_in_flight[i])
4767 			return true;
4768 
4769 	if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4770 		return true;
4771 	if (pwq->nr_active || !list_empty(&pwq->inactive_works))
4772 		return true;
4773 
4774 	return false;
4775 }
4776 
4777 /**
4778  * destroy_workqueue - safely terminate a workqueue
4779  * @wq: target workqueue
4780  *
4781  * Safely destroy a workqueue. All work currently pending will be done first.
4782  */
destroy_workqueue(struct workqueue_struct * wq)4783 void destroy_workqueue(struct workqueue_struct *wq)
4784 {
4785 	struct pool_workqueue *pwq;
4786 	int cpu;
4787 
4788 	/*
4789 	 * Remove it from sysfs first so that sanity check failure doesn't
4790 	 * lead to sysfs name conflicts.
4791 	 */
4792 	workqueue_sysfs_unregister(wq);
4793 
4794 	/* mark the workqueue destruction is in progress */
4795 	mutex_lock(&wq->mutex);
4796 	wq->flags |= __WQ_DESTROYING;
4797 	mutex_unlock(&wq->mutex);
4798 
4799 	/* drain it before proceeding with destruction */
4800 	drain_workqueue(wq);
4801 
4802 	/* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4803 	if (wq->rescuer) {
4804 		struct worker *rescuer = wq->rescuer;
4805 
4806 		/* this prevents new queueing */
4807 		raw_spin_lock_irq(&wq_mayday_lock);
4808 		wq->rescuer = NULL;
4809 		raw_spin_unlock_irq(&wq_mayday_lock);
4810 
4811 		/* rescuer will empty maydays list before exiting */
4812 		kthread_stop(rescuer->task);
4813 		kfree(rescuer);
4814 	}
4815 
4816 	/*
4817 	 * Sanity checks - grab all the locks so that we wait for all
4818 	 * in-flight operations which may do put_pwq().
4819 	 */
4820 	mutex_lock(&wq_pool_mutex);
4821 	mutex_lock(&wq->mutex);
4822 	for_each_pwq(pwq, wq) {
4823 		raw_spin_lock_irq(&pwq->pool->lock);
4824 		if (WARN_ON(pwq_busy(pwq))) {
4825 			pr_warn("%s: %s has the following busy pwq\n",
4826 				__func__, wq->name);
4827 			show_pwq(pwq);
4828 			raw_spin_unlock_irq(&pwq->pool->lock);
4829 			mutex_unlock(&wq->mutex);
4830 			mutex_unlock(&wq_pool_mutex);
4831 			show_one_workqueue(wq);
4832 			return;
4833 		}
4834 		raw_spin_unlock_irq(&pwq->pool->lock);
4835 	}
4836 	mutex_unlock(&wq->mutex);
4837 
4838 	/*
4839 	 * wq list is used to freeze wq, remove from list after
4840 	 * flushing is complete in case freeze races us.
4841 	 */
4842 	list_del_rcu(&wq->list);
4843 	mutex_unlock(&wq_pool_mutex);
4844 
4845 	/*
4846 	 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
4847 	 * to put the base refs. @wq will be auto-destroyed from the last
4848 	 * pwq_put. RCU read lock prevents @wq from going away from under us.
4849 	 */
4850 	rcu_read_lock();
4851 
4852 	for_each_possible_cpu(cpu) {
4853 		pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
4854 		RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL);
4855 		put_pwq_unlocked(pwq);
4856 	}
4857 
4858 	put_pwq_unlocked(wq->dfl_pwq);
4859 	wq->dfl_pwq = NULL;
4860 
4861 	rcu_read_unlock();
4862 }
4863 EXPORT_SYMBOL_GPL(destroy_workqueue);
4864 
4865 /**
4866  * workqueue_set_max_active - adjust max_active of a workqueue
4867  * @wq: target workqueue
4868  * @max_active: new max_active value.
4869  *
4870  * Set max_active of @wq to @max_active.
4871  *
4872  * CONTEXT:
4873  * Don't call from IRQ context.
4874  */
workqueue_set_max_active(struct workqueue_struct * wq,int max_active)4875 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4876 {
4877 	struct pool_workqueue *pwq;
4878 
4879 	/* disallow meddling with max_active for ordered workqueues */
4880 	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4881 		return;
4882 
4883 	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4884 
4885 	mutex_lock(&wq->mutex);
4886 
4887 	wq->flags &= ~__WQ_ORDERED;
4888 	wq->saved_max_active = max_active;
4889 
4890 	for_each_pwq(pwq, wq)
4891 		pwq_adjust_max_active(pwq);
4892 
4893 	mutex_unlock(&wq->mutex);
4894 }
4895 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4896 
4897 /**
4898  * current_work - retrieve %current task's work struct
4899  *
4900  * Determine if %current task is a workqueue worker and what it's working on.
4901  * Useful to find out the context that the %current task is running in.
4902  *
4903  * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4904  */
current_work(void)4905 struct work_struct *current_work(void)
4906 {
4907 	struct worker *worker = current_wq_worker();
4908 
4909 	return worker ? worker->current_work : NULL;
4910 }
4911 EXPORT_SYMBOL(current_work);
4912 
4913 /**
4914  * current_is_workqueue_rescuer - is %current workqueue rescuer?
4915  *
4916  * Determine whether %current is a workqueue rescuer.  Can be used from
4917  * work functions to determine whether it's being run off the rescuer task.
4918  *
4919  * Return: %true if %current is a workqueue rescuer. %false otherwise.
4920  */
current_is_workqueue_rescuer(void)4921 bool current_is_workqueue_rescuer(void)
4922 {
4923 	struct worker *worker = current_wq_worker();
4924 
4925 	return worker && worker->rescue_wq;
4926 }
4927 
4928 /**
4929  * workqueue_congested - test whether a workqueue is congested
4930  * @cpu: CPU in question
4931  * @wq: target workqueue
4932  *
4933  * Test whether @wq's cpu workqueue for @cpu is congested.  There is
4934  * no synchronization around this function and the test result is
4935  * unreliable and only useful as advisory hints or for debugging.
4936  *
4937  * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4938  *
4939  * With the exception of ordered workqueues, all workqueues have per-cpu
4940  * pool_workqueues, each with its own congested state. A workqueue being
4941  * congested on one CPU doesn't mean that the workqueue is contested on any
4942  * other CPUs.
4943  *
4944  * Return:
4945  * %true if congested, %false otherwise.
4946  */
workqueue_congested(int cpu,struct workqueue_struct * wq)4947 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4948 {
4949 	struct pool_workqueue *pwq;
4950 	bool ret;
4951 
4952 	rcu_read_lock();
4953 	preempt_disable();
4954 
4955 	if (cpu == WORK_CPU_UNBOUND)
4956 		cpu = smp_processor_id();
4957 
4958 	pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
4959 	ret = !list_empty(&pwq->inactive_works);
4960 
4961 	preempt_enable();
4962 	rcu_read_unlock();
4963 
4964 	return ret;
4965 }
4966 EXPORT_SYMBOL_GPL(workqueue_congested);
4967 
4968 /**
4969  * work_busy - test whether a work is currently pending or running
4970  * @work: the work to be tested
4971  *
4972  * Test whether @work is currently pending or running.  There is no
4973  * synchronization around this function and the test result is
4974  * unreliable and only useful as advisory hints or for debugging.
4975  *
4976  * Return:
4977  * OR'd bitmask of WORK_BUSY_* bits.
4978  */
work_busy(struct work_struct * work)4979 unsigned int work_busy(struct work_struct *work)
4980 {
4981 	struct worker_pool *pool;
4982 	unsigned long flags;
4983 	unsigned int ret = 0;
4984 
4985 	if (work_pending(work))
4986 		ret |= WORK_BUSY_PENDING;
4987 
4988 	rcu_read_lock();
4989 	pool = get_work_pool(work);
4990 	if (pool) {
4991 		raw_spin_lock_irqsave(&pool->lock, flags);
4992 		if (find_worker_executing_work(pool, work))
4993 			ret |= WORK_BUSY_RUNNING;
4994 		raw_spin_unlock_irqrestore(&pool->lock, flags);
4995 	}
4996 	rcu_read_unlock();
4997 
4998 	return ret;
4999 }
5000 EXPORT_SYMBOL_GPL(work_busy);
5001 
5002 /**
5003  * set_worker_desc - set description for the current work item
5004  * @fmt: printf-style format string
5005  * @...: arguments for the format string
5006  *
5007  * This function can be called by a running work function to describe what
5008  * the work item is about.  If the worker task gets dumped, this
5009  * information will be printed out together to help debugging.  The
5010  * description can be at most WORKER_DESC_LEN including the trailing '\0'.
5011  */
set_worker_desc(const char * fmt,...)5012 void set_worker_desc(const char *fmt, ...)
5013 {
5014 	struct worker *worker = current_wq_worker();
5015 	va_list args;
5016 
5017 	if (worker) {
5018 		va_start(args, fmt);
5019 		vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
5020 		va_end(args);
5021 	}
5022 }
5023 EXPORT_SYMBOL_GPL(set_worker_desc);
5024 
5025 /**
5026  * print_worker_info - print out worker information and description
5027  * @log_lvl: the log level to use when printing
5028  * @task: target task
5029  *
5030  * If @task is a worker and currently executing a work item, print out the
5031  * name of the workqueue being serviced and worker description set with
5032  * set_worker_desc() by the currently executing work item.
5033  *
5034  * This function can be safely called on any task as long as the
5035  * task_struct itself is accessible.  While safe, this function isn't
5036  * synchronized and may print out mixups or garbages of limited length.
5037  */
print_worker_info(const char * log_lvl,struct task_struct * task)5038 void print_worker_info(const char *log_lvl, struct task_struct *task)
5039 {
5040 	work_func_t *fn = NULL;
5041 	char name[WQ_NAME_LEN] = { };
5042 	char desc[WORKER_DESC_LEN] = { };
5043 	struct pool_workqueue *pwq = NULL;
5044 	struct workqueue_struct *wq = NULL;
5045 	struct worker *worker;
5046 
5047 	if (!(task->flags & PF_WQ_WORKER))
5048 		return;
5049 
5050 	/*
5051 	 * This function is called without any synchronization and @task
5052 	 * could be in any state.  Be careful with dereferences.
5053 	 */
5054 	worker = kthread_probe_data(task);
5055 
5056 	/*
5057 	 * Carefully copy the associated workqueue's workfn, name and desc.
5058 	 * Keep the original last '\0' in case the original is garbage.
5059 	 */
5060 	copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
5061 	copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
5062 	copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
5063 	copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
5064 	copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
5065 
5066 	if (fn || name[0] || desc[0]) {
5067 		printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
5068 		if (strcmp(name, desc))
5069 			pr_cont(" (%s)", desc);
5070 		pr_cont("\n");
5071 	}
5072 }
5073 
pr_cont_pool_info(struct worker_pool * pool)5074 static void pr_cont_pool_info(struct worker_pool *pool)
5075 {
5076 	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
5077 	if (pool->node != NUMA_NO_NODE)
5078 		pr_cont(" node=%d", pool->node);
5079 	pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
5080 }
5081 
5082 struct pr_cont_work_struct {
5083 	bool comma;
5084 	work_func_t func;
5085 	long ctr;
5086 };
5087 
pr_cont_work_flush(bool comma,work_func_t func,struct pr_cont_work_struct * pcwsp)5088 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
5089 {
5090 	if (!pcwsp->ctr)
5091 		goto out_record;
5092 	if (func == pcwsp->func) {
5093 		pcwsp->ctr++;
5094 		return;
5095 	}
5096 	if (pcwsp->ctr == 1)
5097 		pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
5098 	else
5099 		pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
5100 	pcwsp->ctr = 0;
5101 out_record:
5102 	if ((long)func == -1L)
5103 		return;
5104 	pcwsp->comma = comma;
5105 	pcwsp->func = func;
5106 	pcwsp->ctr = 1;
5107 }
5108 
pr_cont_work(bool comma,struct work_struct * work,struct pr_cont_work_struct * pcwsp)5109 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
5110 {
5111 	if (work->func == wq_barrier_func) {
5112 		struct wq_barrier *barr;
5113 
5114 		barr = container_of(work, struct wq_barrier, work);
5115 
5116 		pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5117 		pr_cont("%s BAR(%d)", comma ? "," : "",
5118 			task_pid_nr(barr->task));
5119 	} else {
5120 		if (!comma)
5121 			pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5122 		pr_cont_work_flush(comma, work->func, pcwsp);
5123 	}
5124 }
5125 
show_pwq(struct pool_workqueue * pwq)5126 static void show_pwq(struct pool_workqueue *pwq)
5127 {
5128 	struct pr_cont_work_struct pcws = { .ctr = 0, };
5129 	struct worker_pool *pool = pwq->pool;
5130 	struct work_struct *work;
5131 	struct worker *worker;
5132 	bool has_in_flight = false, has_pending = false;
5133 	int bkt;
5134 
5135 	pr_info("  pwq %d:", pool->id);
5136 	pr_cont_pool_info(pool);
5137 
5138 	pr_cont(" active=%d/%d refcnt=%d%s\n",
5139 		pwq->nr_active, pwq->max_active, pwq->refcnt,
5140 		!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
5141 
5142 	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5143 		if (worker->current_pwq == pwq) {
5144 			has_in_flight = true;
5145 			break;
5146 		}
5147 	}
5148 	if (has_in_flight) {
5149 		bool comma = false;
5150 
5151 		pr_info("    in-flight:");
5152 		hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5153 			if (worker->current_pwq != pwq)
5154 				continue;
5155 
5156 			pr_cont("%s %d%s:%ps", comma ? "," : "",
5157 				task_pid_nr(worker->task),
5158 				worker->rescue_wq ? "(RESCUER)" : "",
5159 				worker->current_func);
5160 			list_for_each_entry(work, &worker->scheduled, entry)
5161 				pr_cont_work(false, work, &pcws);
5162 			pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5163 			comma = true;
5164 		}
5165 		pr_cont("\n");
5166 	}
5167 
5168 	list_for_each_entry(work, &pool->worklist, entry) {
5169 		if (get_work_pwq(work) == pwq) {
5170 			has_pending = true;
5171 			break;
5172 		}
5173 	}
5174 	if (has_pending) {
5175 		bool comma = false;
5176 
5177 		pr_info("    pending:");
5178 		list_for_each_entry(work, &pool->worklist, entry) {
5179 			if (get_work_pwq(work) != pwq)
5180 				continue;
5181 
5182 			pr_cont_work(comma, work, &pcws);
5183 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5184 		}
5185 		pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5186 		pr_cont("\n");
5187 	}
5188 
5189 	if (!list_empty(&pwq->inactive_works)) {
5190 		bool comma = false;
5191 
5192 		pr_info("    inactive:");
5193 		list_for_each_entry(work, &pwq->inactive_works, entry) {
5194 			pr_cont_work(comma, work, &pcws);
5195 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5196 		}
5197 		pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5198 		pr_cont("\n");
5199 	}
5200 }
5201 
5202 /**
5203  * show_one_workqueue - dump state of specified workqueue
5204  * @wq: workqueue whose state will be printed
5205  */
show_one_workqueue(struct workqueue_struct * wq)5206 void show_one_workqueue(struct workqueue_struct *wq)
5207 {
5208 	struct pool_workqueue *pwq;
5209 	bool idle = true;
5210 	unsigned long flags;
5211 
5212 	for_each_pwq(pwq, wq) {
5213 		if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
5214 			idle = false;
5215 			break;
5216 		}
5217 	}
5218 	if (idle) /* Nothing to print for idle workqueue */
5219 		return;
5220 
5221 	pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
5222 
5223 	for_each_pwq(pwq, wq) {
5224 		raw_spin_lock_irqsave(&pwq->pool->lock, flags);
5225 		if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
5226 			/*
5227 			 * Defer printing to avoid deadlocks in console
5228 			 * drivers that queue work while holding locks
5229 			 * also taken in their write paths.
5230 			 */
5231 			printk_deferred_enter();
5232 			show_pwq(pwq);
5233 			printk_deferred_exit();
5234 		}
5235 		raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
5236 		/*
5237 		 * We could be printing a lot from atomic context, e.g.
5238 		 * sysrq-t -> show_all_workqueues(). Avoid triggering
5239 		 * hard lockup.
5240 		 */
5241 		touch_nmi_watchdog();
5242 	}
5243 
5244 }
5245 
5246 /**
5247  * show_one_worker_pool - dump state of specified worker pool
5248  * @pool: worker pool whose state will be printed
5249  */
show_one_worker_pool(struct worker_pool * pool)5250 static void show_one_worker_pool(struct worker_pool *pool)
5251 {
5252 	struct worker *worker;
5253 	bool first = true;
5254 	unsigned long flags;
5255 	unsigned long hung = 0;
5256 
5257 	raw_spin_lock_irqsave(&pool->lock, flags);
5258 	if (pool->nr_workers == pool->nr_idle)
5259 		goto next_pool;
5260 
5261 	/* How long the first pending work is waiting for a worker. */
5262 	if (!list_empty(&pool->worklist))
5263 		hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
5264 
5265 	/*
5266 	 * Defer printing to avoid deadlocks in console drivers that
5267 	 * queue work while holding locks also taken in their write
5268 	 * paths.
5269 	 */
5270 	printk_deferred_enter();
5271 	pr_info("pool %d:", pool->id);
5272 	pr_cont_pool_info(pool);
5273 	pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
5274 	if (pool->manager)
5275 		pr_cont(" manager: %d",
5276 			task_pid_nr(pool->manager->task));
5277 	list_for_each_entry(worker, &pool->idle_list, entry) {
5278 		pr_cont(" %s%d", first ? "idle: " : "",
5279 			task_pid_nr(worker->task));
5280 		first = false;
5281 	}
5282 	pr_cont("\n");
5283 	printk_deferred_exit();
5284 next_pool:
5285 	raw_spin_unlock_irqrestore(&pool->lock, flags);
5286 	/*
5287 	 * We could be printing a lot from atomic context, e.g.
5288 	 * sysrq-t -> show_all_workqueues(). Avoid triggering
5289 	 * hard lockup.
5290 	 */
5291 	touch_nmi_watchdog();
5292 
5293 }
5294 
5295 /**
5296  * show_all_workqueues - dump workqueue state
5297  *
5298  * Called from a sysrq handler and prints out all busy workqueues and pools.
5299  */
show_all_workqueues(void)5300 void show_all_workqueues(void)
5301 {
5302 	struct workqueue_struct *wq;
5303 	struct worker_pool *pool;
5304 	int pi;
5305 
5306 	rcu_read_lock();
5307 
5308 	pr_info("Showing busy workqueues and worker pools:\n");
5309 
5310 	list_for_each_entry_rcu(wq, &workqueues, list)
5311 		show_one_workqueue(wq);
5312 
5313 	for_each_pool(pool, pi)
5314 		show_one_worker_pool(pool);
5315 
5316 	rcu_read_unlock();
5317 }
5318 
5319 /**
5320  * show_freezable_workqueues - dump freezable workqueue state
5321  *
5322  * Called from try_to_freeze_tasks() and prints out all freezable workqueues
5323  * still busy.
5324  */
show_freezable_workqueues(void)5325 void show_freezable_workqueues(void)
5326 {
5327 	struct workqueue_struct *wq;
5328 
5329 	rcu_read_lock();
5330 
5331 	pr_info("Showing freezable workqueues that are still busy:\n");
5332 
5333 	list_for_each_entry_rcu(wq, &workqueues, list) {
5334 		if (!(wq->flags & WQ_FREEZABLE))
5335 			continue;
5336 		show_one_workqueue(wq);
5337 	}
5338 
5339 	rcu_read_unlock();
5340 }
5341 
5342 /* used to show worker information through /proc/PID/{comm,stat,status} */
wq_worker_comm(char * buf,size_t size,struct task_struct * task)5343 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
5344 {
5345 	int off;
5346 
5347 	/* always show the actual comm */
5348 	off = strscpy(buf, task->comm, size);
5349 	if (off < 0)
5350 		return;
5351 
5352 	/* stabilize PF_WQ_WORKER and worker pool association */
5353 	mutex_lock(&wq_pool_attach_mutex);
5354 
5355 	if (task->flags & PF_WQ_WORKER) {
5356 		struct worker *worker = kthread_data(task);
5357 		struct worker_pool *pool = worker->pool;
5358 
5359 		if (pool) {
5360 			raw_spin_lock_irq(&pool->lock);
5361 			/*
5362 			 * ->desc tracks information (wq name or
5363 			 * set_worker_desc()) for the latest execution.  If
5364 			 * current, prepend '+', otherwise '-'.
5365 			 */
5366 			if (worker->desc[0] != '\0') {
5367 				if (worker->current_work)
5368 					scnprintf(buf + off, size - off, "+%s",
5369 						  worker->desc);
5370 				else
5371 					scnprintf(buf + off, size - off, "-%s",
5372 						  worker->desc);
5373 			}
5374 			raw_spin_unlock_irq(&pool->lock);
5375 		}
5376 	}
5377 
5378 	mutex_unlock(&wq_pool_attach_mutex);
5379 }
5380 
5381 #ifdef CONFIG_SMP
5382 
5383 /*
5384  * CPU hotplug.
5385  *
5386  * There are two challenges in supporting CPU hotplug.  Firstly, there
5387  * are a lot of assumptions on strong associations among work, pwq and
5388  * pool which make migrating pending and scheduled works very
5389  * difficult to implement without impacting hot paths.  Secondly,
5390  * worker pools serve mix of short, long and very long running works making
5391  * blocked draining impractical.
5392  *
5393  * This is solved by allowing the pools to be disassociated from the CPU
5394  * running as an unbound one and allowing it to be reattached later if the
5395  * cpu comes back online.
5396  */
5397 
unbind_workers(int cpu)5398 static void unbind_workers(int cpu)
5399 {
5400 	struct worker_pool *pool;
5401 	struct worker *worker;
5402 
5403 	for_each_cpu_worker_pool(pool, cpu) {
5404 		mutex_lock(&wq_pool_attach_mutex);
5405 		raw_spin_lock_irq(&pool->lock);
5406 
5407 		/*
5408 		 * We've blocked all attach/detach operations. Make all workers
5409 		 * unbound and set DISASSOCIATED.  Before this, all workers
5410 		 * must be on the cpu.  After this, they may become diasporas.
5411 		 * And the preemption disabled section in their sched callbacks
5412 		 * are guaranteed to see WORKER_UNBOUND since the code here
5413 		 * is on the same cpu.
5414 		 */
5415 		for_each_pool_worker(worker, pool)
5416 			worker->flags |= WORKER_UNBOUND;
5417 
5418 		pool->flags |= POOL_DISASSOCIATED;
5419 
5420 		/*
5421 		 * The handling of nr_running in sched callbacks are disabled
5422 		 * now.  Zap nr_running.  After this, nr_running stays zero and
5423 		 * need_more_worker() and keep_working() are always true as
5424 		 * long as the worklist is not empty.  This pool now behaves as
5425 		 * an unbound (in terms of concurrency management) pool which
5426 		 * are served by workers tied to the pool.
5427 		 */
5428 		pool->nr_running = 0;
5429 
5430 		/*
5431 		 * With concurrency management just turned off, a busy
5432 		 * worker blocking could lead to lengthy stalls.  Kick off
5433 		 * unbound chain execution of currently pending work items.
5434 		 */
5435 		kick_pool(pool);
5436 
5437 		raw_spin_unlock_irq(&pool->lock);
5438 
5439 		for_each_pool_worker(worker, pool)
5440 			unbind_worker(worker);
5441 
5442 		mutex_unlock(&wq_pool_attach_mutex);
5443 	}
5444 }
5445 
5446 /**
5447  * rebind_workers - rebind all workers of a pool to the associated CPU
5448  * @pool: pool of interest
5449  *
5450  * @pool->cpu is coming online.  Rebind all workers to the CPU.
5451  */
rebind_workers(struct worker_pool * pool)5452 static void rebind_workers(struct worker_pool *pool)
5453 {
5454 	struct worker *worker;
5455 
5456 	lockdep_assert_held(&wq_pool_attach_mutex);
5457 
5458 	/*
5459 	 * Restore CPU affinity of all workers.  As all idle workers should
5460 	 * be on the run-queue of the associated CPU before any local
5461 	 * wake-ups for concurrency management happen, restore CPU affinity
5462 	 * of all workers first and then clear UNBOUND.  As we're called
5463 	 * from CPU_ONLINE, the following shouldn't fail.
5464 	 */
5465 	for_each_pool_worker(worker, pool) {
5466 		kthread_set_per_cpu(worker->task, pool->cpu);
5467 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
5468 						  pool_allowed_cpus(pool)) < 0);
5469 	}
5470 
5471 	raw_spin_lock_irq(&pool->lock);
5472 
5473 	pool->flags &= ~POOL_DISASSOCIATED;
5474 
5475 	for_each_pool_worker(worker, pool) {
5476 		unsigned int worker_flags = worker->flags;
5477 
5478 		/*
5479 		 * We want to clear UNBOUND but can't directly call
5480 		 * worker_clr_flags() or adjust nr_running.  Atomically
5481 		 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5482 		 * @worker will clear REBOUND using worker_clr_flags() when
5483 		 * it initiates the next execution cycle thus restoring
5484 		 * concurrency management.  Note that when or whether
5485 		 * @worker clears REBOUND doesn't affect correctness.
5486 		 *
5487 		 * WRITE_ONCE() is necessary because @worker->flags may be
5488 		 * tested without holding any lock in
5489 		 * wq_worker_running().  Without it, NOT_RUNNING test may
5490 		 * fail incorrectly leading to premature concurrency
5491 		 * management operations.
5492 		 */
5493 		WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
5494 		worker_flags |= WORKER_REBOUND;
5495 		worker_flags &= ~WORKER_UNBOUND;
5496 		WRITE_ONCE(worker->flags, worker_flags);
5497 	}
5498 
5499 	raw_spin_unlock_irq(&pool->lock);
5500 }
5501 
5502 /**
5503  * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5504  * @pool: unbound pool of interest
5505  * @cpu: the CPU which is coming up
5506  *
5507  * An unbound pool may end up with a cpumask which doesn't have any online
5508  * CPUs.  When a worker of such pool get scheduled, the scheduler resets
5509  * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
5510  * online CPU before, cpus_allowed of all its workers should be restored.
5511  */
restore_unbound_workers_cpumask(struct worker_pool * pool,int cpu)5512 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5513 {
5514 	static cpumask_t cpumask;
5515 	struct worker *worker;
5516 
5517 	lockdep_assert_held(&wq_pool_attach_mutex);
5518 
5519 	/* is @cpu allowed for @pool? */
5520 	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5521 		return;
5522 
5523 	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5524 
5525 	/* as we're called from CPU_ONLINE, the following shouldn't fail */
5526 	for_each_pool_worker(worker, pool)
5527 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5528 }
5529 
workqueue_prepare_cpu(unsigned int cpu)5530 int workqueue_prepare_cpu(unsigned int cpu)
5531 {
5532 	struct worker_pool *pool;
5533 
5534 	for_each_cpu_worker_pool(pool, cpu) {
5535 		if (pool->nr_workers)
5536 			continue;
5537 		if (!create_worker(pool))
5538 			return -ENOMEM;
5539 	}
5540 	return 0;
5541 }
5542 
workqueue_online_cpu(unsigned int cpu)5543 int workqueue_online_cpu(unsigned int cpu)
5544 {
5545 	struct worker_pool *pool;
5546 	struct workqueue_struct *wq;
5547 	int pi;
5548 
5549 	mutex_lock(&wq_pool_mutex);
5550 
5551 	for_each_pool(pool, pi) {
5552 		mutex_lock(&wq_pool_attach_mutex);
5553 
5554 		if (pool->cpu == cpu)
5555 			rebind_workers(pool);
5556 		else if (pool->cpu < 0)
5557 			restore_unbound_workers_cpumask(pool, cpu);
5558 
5559 		mutex_unlock(&wq_pool_attach_mutex);
5560 	}
5561 
5562 	/* update pod affinity of unbound workqueues */
5563 	list_for_each_entry(wq, &workqueues, list) {
5564 		struct workqueue_attrs *attrs = wq->unbound_attrs;
5565 
5566 		if (attrs) {
5567 			const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5568 			int tcpu;
5569 
5570 			for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
5571 				wq_update_pod(wq, tcpu, cpu, true);
5572 		}
5573 	}
5574 
5575 	mutex_unlock(&wq_pool_mutex);
5576 	return 0;
5577 }
5578 
workqueue_offline_cpu(unsigned int cpu)5579 int workqueue_offline_cpu(unsigned int cpu)
5580 {
5581 	struct workqueue_struct *wq;
5582 
5583 	/* unbinding per-cpu workers should happen on the local CPU */
5584 	if (WARN_ON(cpu != smp_processor_id()))
5585 		return -1;
5586 
5587 	unbind_workers(cpu);
5588 
5589 	/* update pod affinity of unbound workqueues */
5590 	mutex_lock(&wq_pool_mutex);
5591 	list_for_each_entry(wq, &workqueues, list) {
5592 		struct workqueue_attrs *attrs = wq->unbound_attrs;
5593 
5594 		if (attrs) {
5595 			const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5596 			int tcpu;
5597 
5598 			for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
5599 				wq_update_pod(wq, tcpu, cpu, false);
5600 		}
5601 	}
5602 	mutex_unlock(&wq_pool_mutex);
5603 
5604 	return 0;
5605 }
5606 
5607 struct work_for_cpu {
5608 	struct work_struct work;
5609 	long (*fn)(void *);
5610 	void *arg;
5611 	long ret;
5612 };
5613 
work_for_cpu_fn(struct work_struct * work)5614 static void work_for_cpu_fn(struct work_struct *work)
5615 {
5616 	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5617 
5618 	wfc->ret = wfc->fn(wfc->arg);
5619 }
5620 
5621 /**
5622  * work_on_cpu_key - run a function in thread context on a particular cpu
5623  * @cpu: the cpu to run on
5624  * @fn: the function to run
5625  * @arg: the function arg
5626  * @key: The lock class key for lock debugging purposes
5627  *
5628  * It is up to the caller to ensure that the cpu doesn't go offline.
5629  * The caller must not hold any locks which would prevent @fn from completing.
5630  *
5631  * Return: The value @fn returns.
5632  */
work_on_cpu_key(int cpu,long (* fn)(void *),void * arg,struct lock_class_key * key)5633 long work_on_cpu_key(int cpu, long (*fn)(void *),
5634 		     void *arg, struct lock_class_key *key)
5635 {
5636 	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5637 
5638 	INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
5639 	schedule_work_on(cpu, &wfc.work);
5640 	flush_work(&wfc.work);
5641 	destroy_work_on_stack(&wfc.work);
5642 	return wfc.ret;
5643 }
5644 EXPORT_SYMBOL_GPL(work_on_cpu_key);
5645 
5646 /**
5647  * work_on_cpu_safe_key - run a function in thread context on a particular cpu
5648  * @cpu: the cpu to run on
5649  * @fn:  the function to run
5650  * @arg: the function argument
5651  * @key: The lock class key for lock debugging purposes
5652  *
5653  * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5654  * any locks which would prevent @fn from completing.
5655  *
5656  * Return: The value @fn returns.
5657  */
work_on_cpu_safe_key(int cpu,long (* fn)(void *),void * arg,struct lock_class_key * key)5658 long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
5659 			  void *arg, struct lock_class_key *key)
5660 {
5661 	long ret = -ENODEV;
5662 
5663 	cpus_read_lock();
5664 	if (cpu_online(cpu))
5665 		ret = work_on_cpu_key(cpu, fn, arg, key);
5666 	cpus_read_unlock();
5667 	return ret;
5668 }
5669 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
5670 #endif /* CONFIG_SMP */
5671 
5672 #ifdef CONFIG_FREEZER
5673 
5674 /**
5675  * freeze_workqueues_begin - begin freezing workqueues
5676  *
5677  * Start freezing workqueues.  After this function returns, all freezable
5678  * workqueues will queue new works to their inactive_works list instead of
5679  * pool->worklist.
5680  *
5681  * CONTEXT:
5682  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5683  */
freeze_workqueues_begin(void)5684 void freeze_workqueues_begin(void)
5685 {
5686 	struct workqueue_struct *wq;
5687 	struct pool_workqueue *pwq;
5688 
5689 	mutex_lock(&wq_pool_mutex);
5690 
5691 	WARN_ON_ONCE(workqueue_freezing);
5692 	workqueue_freezing = true;
5693 
5694 	list_for_each_entry(wq, &workqueues, list) {
5695 		mutex_lock(&wq->mutex);
5696 		for_each_pwq(pwq, wq)
5697 			pwq_adjust_max_active(pwq);
5698 		mutex_unlock(&wq->mutex);
5699 	}
5700 
5701 	mutex_unlock(&wq_pool_mutex);
5702 }
5703 
5704 /**
5705  * freeze_workqueues_busy - are freezable workqueues still busy?
5706  *
5707  * Check whether freezing is complete.  This function must be called
5708  * between freeze_workqueues_begin() and thaw_workqueues().
5709  *
5710  * CONTEXT:
5711  * Grabs and releases wq_pool_mutex.
5712  *
5713  * Return:
5714  * %true if some freezable workqueues are still busy.  %false if freezing
5715  * is complete.
5716  */
freeze_workqueues_busy(void)5717 bool freeze_workqueues_busy(void)
5718 {
5719 	bool busy = false;
5720 	struct workqueue_struct *wq;
5721 	struct pool_workqueue *pwq;
5722 
5723 	mutex_lock(&wq_pool_mutex);
5724 
5725 	WARN_ON_ONCE(!workqueue_freezing);
5726 
5727 	list_for_each_entry(wq, &workqueues, list) {
5728 		if (!(wq->flags & WQ_FREEZABLE))
5729 			continue;
5730 		/*
5731 		 * nr_active is monotonically decreasing.  It's safe
5732 		 * to peek without lock.
5733 		 */
5734 		rcu_read_lock();
5735 		for_each_pwq(pwq, wq) {
5736 			WARN_ON_ONCE(pwq->nr_active < 0);
5737 			if (pwq->nr_active) {
5738 				busy = true;
5739 				rcu_read_unlock();
5740 				goto out_unlock;
5741 			}
5742 		}
5743 		rcu_read_unlock();
5744 	}
5745 out_unlock:
5746 	mutex_unlock(&wq_pool_mutex);
5747 	return busy;
5748 }
5749 
5750 /**
5751  * thaw_workqueues - thaw workqueues
5752  *
5753  * Thaw workqueues.  Normal queueing is restored and all collected
5754  * frozen works are transferred to their respective pool worklists.
5755  *
5756  * CONTEXT:
5757  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5758  */
thaw_workqueues(void)5759 void thaw_workqueues(void)
5760 {
5761 	struct workqueue_struct *wq;
5762 	struct pool_workqueue *pwq;
5763 
5764 	mutex_lock(&wq_pool_mutex);
5765 
5766 	if (!workqueue_freezing)
5767 		goto out_unlock;
5768 
5769 	workqueue_freezing = false;
5770 
5771 	/* restore max_active and repopulate worklist */
5772 	list_for_each_entry(wq, &workqueues, list) {
5773 		mutex_lock(&wq->mutex);
5774 		for_each_pwq(pwq, wq)
5775 			pwq_adjust_max_active(pwq);
5776 		mutex_unlock(&wq->mutex);
5777 	}
5778 
5779 out_unlock:
5780 	mutex_unlock(&wq_pool_mutex);
5781 }
5782 #endif /* CONFIG_FREEZER */
5783 
workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)5784 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
5785 {
5786 	LIST_HEAD(ctxs);
5787 	int ret = 0;
5788 	struct workqueue_struct *wq;
5789 	struct apply_wqattrs_ctx *ctx, *n;
5790 
5791 	lockdep_assert_held(&wq_pool_mutex);
5792 
5793 	list_for_each_entry(wq, &workqueues, list) {
5794 		if (!(wq->flags & WQ_UNBOUND))
5795 			continue;
5796 		/* creating multiple pwqs breaks ordering guarantee */
5797 		if (wq->flags & __WQ_ORDERED)
5798 			continue;
5799 
5800 		ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
5801 		if (IS_ERR(ctx)) {
5802 			ret = PTR_ERR(ctx);
5803 			break;
5804 		}
5805 
5806 		list_add_tail(&ctx->list, &ctxs);
5807 	}
5808 
5809 	list_for_each_entry_safe(ctx, n, &ctxs, list) {
5810 		if (!ret)
5811 			apply_wqattrs_commit(ctx);
5812 		apply_wqattrs_cleanup(ctx);
5813 	}
5814 
5815 	if (!ret) {
5816 		mutex_lock(&wq_pool_attach_mutex);
5817 		cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
5818 		mutex_unlock(&wq_pool_attach_mutex);
5819 	}
5820 	return ret;
5821 }
5822 
5823 /**
5824  *  workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5825  *  @cpumask: the cpumask to set
5826  *
5827  *  The low-level workqueues cpumask is a global cpumask that limits
5828  *  the affinity of all unbound workqueues.  This function check the @cpumask
5829  *  and apply it to all unbound workqueues and updates all pwqs of them.
5830  *
5831  *  Return:	0	- Success
5832  *  		-EINVAL	- Invalid @cpumask
5833  *  		-ENOMEM	- Failed to allocate memory for attrs or pwqs.
5834  */
workqueue_set_unbound_cpumask(cpumask_var_t cpumask)5835 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5836 {
5837 	int ret = -EINVAL;
5838 
5839 	/*
5840 	 * Not excluding isolated cpus on purpose.
5841 	 * If the user wishes to include them, we allow that.
5842 	 */
5843 	cpumask_and(cpumask, cpumask, cpu_possible_mask);
5844 	if (!cpumask_empty(cpumask)) {
5845 		apply_wqattrs_lock();
5846 		if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
5847 			ret = 0;
5848 			goto out_unlock;
5849 		}
5850 
5851 		ret = workqueue_apply_unbound_cpumask(cpumask);
5852 
5853 out_unlock:
5854 		apply_wqattrs_unlock();
5855 	}
5856 
5857 	return ret;
5858 }
5859 
parse_affn_scope(const char * val)5860 static int parse_affn_scope(const char *val)
5861 {
5862 	int i;
5863 
5864 	for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
5865 		if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
5866 			return i;
5867 	}
5868 	return -EINVAL;
5869 }
5870 
wq_affn_dfl_set(const char * val,const struct kernel_param * kp)5871 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
5872 {
5873 	struct workqueue_struct *wq;
5874 	int affn, cpu;
5875 
5876 	affn = parse_affn_scope(val);
5877 	if (affn < 0)
5878 		return affn;
5879 	if (affn == WQ_AFFN_DFL)
5880 		return -EINVAL;
5881 
5882 	cpus_read_lock();
5883 	mutex_lock(&wq_pool_mutex);
5884 
5885 	wq_affn_dfl = affn;
5886 
5887 	list_for_each_entry(wq, &workqueues, list) {
5888 		for_each_online_cpu(cpu) {
5889 			wq_update_pod(wq, cpu, cpu, true);
5890 		}
5891 	}
5892 
5893 	mutex_unlock(&wq_pool_mutex);
5894 	cpus_read_unlock();
5895 
5896 	return 0;
5897 }
5898 
wq_affn_dfl_get(char * buffer,const struct kernel_param * kp)5899 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
5900 {
5901 	return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
5902 }
5903 
5904 static const struct kernel_param_ops wq_affn_dfl_ops = {
5905 	.set	= wq_affn_dfl_set,
5906 	.get	= wq_affn_dfl_get,
5907 };
5908 
5909 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
5910 
5911 #ifdef CONFIG_SYSFS
5912 /*
5913  * Workqueues with WQ_SYSFS flag set is visible to userland via
5914  * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
5915  * following attributes.
5916  *
5917  *  per_cpu		RO bool	: whether the workqueue is per-cpu or unbound
5918  *  max_active		RW int	: maximum number of in-flight work items
5919  *
5920  * Unbound workqueues have the following extra attributes.
5921  *
5922  *  nice		RW int	: nice value of the workers
5923  *  cpumask		RW mask	: bitmask of allowed CPUs for the workers
5924  *  affinity_scope	RW str  : worker CPU affinity scope (cache, numa, none)
5925  *  affinity_strict	RW bool : worker CPU affinity is strict
5926  */
5927 struct wq_device {
5928 	struct workqueue_struct		*wq;
5929 	struct device			dev;
5930 };
5931 
dev_to_wq(struct device * dev)5932 static struct workqueue_struct *dev_to_wq(struct device *dev)
5933 {
5934 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5935 
5936 	return wq_dev->wq;
5937 }
5938 
per_cpu_show(struct device * dev,struct device_attribute * attr,char * buf)5939 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5940 			    char *buf)
5941 {
5942 	struct workqueue_struct *wq = dev_to_wq(dev);
5943 
5944 	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5945 }
5946 static DEVICE_ATTR_RO(per_cpu);
5947 
max_active_show(struct device * dev,struct device_attribute * attr,char * buf)5948 static ssize_t max_active_show(struct device *dev,
5949 			       struct device_attribute *attr, char *buf)
5950 {
5951 	struct workqueue_struct *wq = dev_to_wq(dev);
5952 
5953 	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5954 }
5955 
max_active_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)5956 static ssize_t max_active_store(struct device *dev,
5957 				struct device_attribute *attr, const char *buf,
5958 				size_t count)
5959 {
5960 	struct workqueue_struct *wq = dev_to_wq(dev);
5961 	int val;
5962 
5963 	if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5964 		return -EINVAL;
5965 
5966 	workqueue_set_max_active(wq, val);
5967 	return count;
5968 }
5969 static DEVICE_ATTR_RW(max_active);
5970 
5971 static struct attribute *wq_sysfs_attrs[] = {
5972 	&dev_attr_per_cpu.attr,
5973 	&dev_attr_max_active.attr,
5974 	NULL,
5975 };
5976 ATTRIBUTE_GROUPS(wq_sysfs);
5977 
wq_nice_show(struct device * dev,struct device_attribute * attr,char * buf)5978 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5979 			    char *buf)
5980 {
5981 	struct workqueue_struct *wq = dev_to_wq(dev);
5982 	int written;
5983 
5984 	mutex_lock(&wq->mutex);
5985 	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5986 	mutex_unlock(&wq->mutex);
5987 
5988 	return written;
5989 }
5990 
5991 /* prepare workqueue_attrs for sysfs store operations */
wq_sysfs_prep_attrs(struct workqueue_struct * wq)5992 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5993 {
5994 	struct workqueue_attrs *attrs;
5995 
5996 	lockdep_assert_held(&wq_pool_mutex);
5997 
5998 	attrs = alloc_workqueue_attrs();
5999 	if (!attrs)
6000 		return NULL;
6001 
6002 	copy_workqueue_attrs(attrs, wq->unbound_attrs);
6003 	return attrs;
6004 }
6005 
wq_nice_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)6006 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
6007 			     const char *buf, size_t count)
6008 {
6009 	struct workqueue_struct *wq = dev_to_wq(dev);
6010 	struct workqueue_attrs *attrs;
6011 	int ret = -ENOMEM;
6012 
6013 	apply_wqattrs_lock();
6014 
6015 	attrs = wq_sysfs_prep_attrs(wq);
6016 	if (!attrs)
6017 		goto out_unlock;
6018 
6019 	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
6020 	    attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
6021 		ret = apply_workqueue_attrs_locked(wq, attrs);
6022 	else
6023 		ret = -EINVAL;
6024 
6025 out_unlock:
6026 	apply_wqattrs_unlock();
6027 	free_workqueue_attrs(attrs);
6028 	return ret ?: count;
6029 }
6030 
wq_cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)6031 static ssize_t wq_cpumask_show(struct device *dev,
6032 			       struct device_attribute *attr, char *buf)
6033 {
6034 	struct workqueue_struct *wq = dev_to_wq(dev);
6035 	int written;
6036 
6037 	mutex_lock(&wq->mutex);
6038 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6039 			    cpumask_pr_args(wq->unbound_attrs->cpumask));
6040 	mutex_unlock(&wq->mutex);
6041 	return written;
6042 }
6043 
wq_cpumask_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)6044 static ssize_t wq_cpumask_store(struct device *dev,
6045 				struct device_attribute *attr,
6046 				const char *buf, size_t count)
6047 {
6048 	struct workqueue_struct *wq = dev_to_wq(dev);
6049 	struct workqueue_attrs *attrs;
6050 	int ret = -ENOMEM;
6051 
6052 	apply_wqattrs_lock();
6053 
6054 	attrs = wq_sysfs_prep_attrs(wq);
6055 	if (!attrs)
6056 		goto out_unlock;
6057 
6058 	ret = cpumask_parse(buf, attrs->cpumask);
6059 	if (!ret)
6060 		ret = apply_workqueue_attrs_locked(wq, attrs);
6061 
6062 out_unlock:
6063 	apply_wqattrs_unlock();
6064 	free_workqueue_attrs(attrs);
6065 	return ret ?: count;
6066 }
6067 
wq_affn_scope_show(struct device * dev,struct device_attribute * attr,char * buf)6068 static ssize_t wq_affn_scope_show(struct device *dev,
6069 				  struct device_attribute *attr, char *buf)
6070 {
6071 	struct workqueue_struct *wq = dev_to_wq(dev);
6072 	int written;
6073 
6074 	mutex_lock(&wq->mutex);
6075 	if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
6076 		written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
6077 				    wq_affn_names[WQ_AFFN_DFL],
6078 				    wq_affn_names[wq_affn_dfl]);
6079 	else
6080 		written = scnprintf(buf, PAGE_SIZE, "%s\n",
6081 				    wq_affn_names[wq->unbound_attrs->affn_scope]);
6082 	mutex_unlock(&wq->mutex);
6083 
6084 	return written;
6085 }
6086 
wq_affn_scope_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)6087 static ssize_t wq_affn_scope_store(struct device *dev,
6088 				   struct device_attribute *attr,
6089 				   const char *buf, size_t count)
6090 {
6091 	struct workqueue_struct *wq = dev_to_wq(dev);
6092 	struct workqueue_attrs *attrs;
6093 	int affn, ret = -ENOMEM;
6094 
6095 	affn = parse_affn_scope(buf);
6096 	if (affn < 0)
6097 		return affn;
6098 
6099 	apply_wqattrs_lock();
6100 	attrs = wq_sysfs_prep_attrs(wq);
6101 	if (attrs) {
6102 		attrs->affn_scope = affn;
6103 		ret = apply_workqueue_attrs_locked(wq, attrs);
6104 	}
6105 	apply_wqattrs_unlock();
6106 	free_workqueue_attrs(attrs);
6107 	return ret ?: count;
6108 }
6109 
wq_affinity_strict_show(struct device * dev,struct device_attribute * attr,char * buf)6110 static ssize_t wq_affinity_strict_show(struct device *dev,
6111 				       struct device_attribute *attr, char *buf)
6112 {
6113 	struct workqueue_struct *wq = dev_to_wq(dev);
6114 
6115 	return scnprintf(buf, PAGE_SIZE, "%d\n",
6116 			 wq->unbound_attrs->affn_strict);
6117 }
6118 
wq_affinity_strict_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)6119 static ssize_t wq_affinity_strict_store(struct device *dev,
6120 					struct device_attribute *attr,
6121 					const char *buf, size_t count)
6122 {
6123 	struct workqueue_struct *wq = dev_to_wq(dev);
6124 	struct workqueue_attrs *attrs;
6125 	int v, ret = -ENOMEM;
6126 
6127 	if (sscanf(buf, "%d", &v) != 1)
6128 		return -EINVAL;
6129 
6130 	apply_wqattrs_lock();
6131 	attrs = wq_sysfs_prep_attrs(wq);
6132 	if (attrs) {
6133 		attrs->affn_strict = (bool)v;
6134 		ret = apply_workqueue_attrs_locked(wq, attrs);
6135 	}
6136 	apply_wqattrs_unlock();
6137 	free_workqueue_attrs(attrs);
6138 	return ret ?: count;
6139 }
6140 
6141 static struct device_attribute wq_sysfs_unbound_attrs[] = {
6142 	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
6143 	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
6144 	__ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
6145 	__ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
6146 	__ATTR_NULL,
6147 };
6148 
6149 static struct bus_type wq_subsys = {
6150 	.name				= "workqueue",
6151 	.dev_groups			= wq_sysfs_groups,
6152 };
6153 
wq_unbound_cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)6154 static ssize_t wq_unbound_cpumask_show(struct device *dev,
6155 		struct device_attribute *attr, char *buf)
6156 {
6157 	int written;
6158 
6159 	mutex_lock(&wq_pool_mutex);
6160 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6161 			    cpumask_pr_args(wq_unbound_cpumask));
6162 	mutex_unlock(&wq_pool_mutex);
6163 
6164 	return written;
6165 }
6166 
wq_unbound_cpumask_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)6167 static ssize_t wq_unbound_cpumask_store(struct device *dev,
6168 		struct device_attribute *attr, const char *buf, size_t count)
6169 {
6170 	cpumask_var_t cpumask;
6171 	int ret;
6172 
6173 	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6174 		return -ENOMEM;
6175 
6176 	ret = cpumask_parse(buf, cpumask);
6177 	if (!ret)
6178 		ret = workqueue_set_unbound_cpumask(cpumask);
6179 
6180 	free_cpumask_var(cpumask);
6181 	return ret ? ret : count;
6182 }
6183 
6184 static struct device_attribute wq_sysfs_cpumask_attr =
6185 	__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
6186 	       wq_unbound_cpumask_store);
6187 
wq_sysfs_init(void)6188 static int __init wq_sysfs_init(void)
6189 {
6190 	struct device *dev_root;
6191 	int err;
6192 
6193 	err = subsys_virtual_register(&wq_subsys, NULL);
6194 	if (err)
6195 		return err;
6196 
6197 	dev_root = bus_get_dev_root(&wq_subsys);
6198 	if (dev_root) {
6199 		err = device_create_file(dev_root, &wq_sysfs_cpumask_attr);
6200 		put_device(dev_root);
6201 	}
6202 	return err;
6203 }
6204 core_initcall(wq_sysfs_init);
6205 
wq_device_release(struct device * dev)6206 static void wq_device_release(struct device *dev)
6207 {
6208 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6209 
6210 	kfree(wq_dev);
6211 }
6212 
6213 /**
6214  * workqueue_sysfs_register - make a workqueue visible in sysfs
6215  * @wq: the workqueue to register
6216  *
6217  * Expose @wq in sysfs under /sys/bus/workqueue/devices.
6218  * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
6219  * which is the preferred method.
6220  *
6221  * Workqueue user should use this function directly iff it wants to apply
6222  * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
6223  * apply_workqueue_attrs() may race against userland updating the
6224  * attributes.
6225  *
6226  * Return: 0 on success, -errno on failure.
6227  */
workqueue_sysfs_register(struct workqueue_struct * wq)6228 int workqueue_sysfs_register(struct workqueue_struct *wq)
6229 {
6230 	struct wq_device *wq_dev;
6231 	int ret;
6232 
6233 	/*
6234 	 * Adjusting max_active or creating new pwqs by applying
6235 	 * attributes breaks ordering guarantee.  Disallow exposing ordered
6236 	 * workqueues.
6237 	 */
6238 	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
6239 		return -EINVAL;
6240 
6241 	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
6242 	if (!wq_dev)
6243 		return -ENOMEM;
6244 
6245 	wq_dev->wq = wq;
6246 	wq_dev->dev.bus = &wq_subsys;
6247 	wq_dev->dev.release = wq_device_release;
6248 	dev_set_name(&wq_dev->dev, "%s", wq->name);
6249 
6250 	/*
6251 	 * unbound_attrs are created separately.  Suppress uevent until
6252 	 * everything is ready.
6253 	 */
6254 	dev_set_uevent_suppress(&wq_dev->dev, true);
6255 
6256 	ret = device_register(&wq_dev->dev);
6257 	if (ret) {
6258 		put_device(&wq_dev->dev);
6259 		wq->wq_dev = NULL;
6260 		return ret;
6261 	}
6262 
6263 	if (wq->flags & WQ_UNBOUND) {
6264 		struct device_attribute *attr;
6265 
6266 		for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
6267 			ret = device_create_file(&wq_dev->dev, attr);
6268 			if (ret) {
6269 				device_unregister(&wq_dev->dev);
6270 				wq->wq_dev = NULL;
6271 				return ret;
6272 			}
6273 		}
6274 	}
6275 
6276 	dev_set_uevent_suppress(&wq_dev->dev, false);
6277 	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
6278 	return 0;
6279 }
6280 
6281 /**
6282  * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
6283  * @wq: the workqueue to unregister
6284  *
6285  * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
6286  */
workqueue_sysfs_unregister(struct workqueue_struct * wq)6287 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
6288 {
6289 	struct wq_device *wq_dev = wq->wq_dev;
6290 
6291 	if (!wq->wq_dev)
6292 		return;
6293 
6294 	wq->wq_dev = NULL;
6295 	device_unregister(&wq_dev->dev);
6296 }
6297 #else	/* CONFIG_SYSFS */
workqueue_sysfs_unregister(struct workqueue_struct * wq)6298 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)	{ }
6299 #endif	/* CONFIG_SYSFS */
6300 
6301 /*
6302  * Workqueue watchdog.
6303  *
6304  * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
6305  * flush dependency, a concurrency managed work item which stays RUNNING
6306  * indefinitely.  Workqueue stalls can be very difficult to debug as the
6307  * usual warning mechanisms don't trigger and internal workqueue state is
6308  * largely opaque.
6309  *
6310  * Workqueue watchdog monitors all worker pools periodically and dumps
6311  * state if some pools failed to make forward progress for a while where
6312  * forward progress is defined as the first item on ->worklist changing.
6313  *
6314  * This mechanism is controlled through the kernel parameter
6315  * "workqueue.watchdog_thresh" which can be updated at runtime through the
6316  * corresponding sysfs parameter file.
6317  */
6318 #ifdef CONFIG_WQ_WATCHDOG
6319 
6320 static unsigned long wq_watchdog_thresh = 30;
6321 static struct timer_list wq_watchdog_timer;
6322 
6323 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
6324 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
6325 
6326 /*
6327  * Show workers that might prevent the processing of pending work items.
6328  * The only candidates are CPU-bound workers in the running state.
6329  * Pending work items should be handled by another idle worker
6330  * in all other situations.
6331  */
show_cpu_pool_hog(struct worker_pool * pool)6332 static void show_cpu_pool_hog(struct worker_pool *pool)
6333 {
6334 	struct worker *worker;
6335 	unsigned long flags;
6336 	int bkt;
6337 
6338 	raw_spin_lock_irqsave(&pool->lock, flags);
6339 
6340 	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6341 		if (task_is_running(worker->task)) {
6342 			/*
6343 			 * Defer printing to avoid deadlocks in console
6344 			 * drivers that queue work while holding locks
6345 			 * also taken in their write paths.
6346 			 */
6347 			printk_deferred_enter();
6348 
6349 			pr_info("pool %d:\n", pool->id);
6350 			sched_show_task(worker->task);
6351 
6352 			printk_deferred_exit();
6353 		}
6354 	}
6355 
6356 	raw_spin_unlock_irqrestore(&pool->lock, flags);
6357 }
6358 
show_cpu_pools_hogs(void)6359 static void show_cpu_pools_hogs(void)
6360 {
6361 	struct worker_pool *pool;
6362 	int pi;
6363 
6364 	pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
6365 
6366 	rcu_read_lock();
6367 
6368 	for_each_pool(pool, pi) {
6369 		if (pool->cpu_stall)
6370 			show_cpu_pool_hog(pool);
6371 
6372 	}
6373 
6374 	rcu_read_unlock();
6375 }
6376 
wq_watchdog_reset_touched(void)6377 static void wq_watchdog_reset_touched(void)
6378 {
6379 	int cpu;
6380 
6381 	wq_watchdog_touched = jiffies;
6382 	for_each_possible_cpu(cpu)
6383 		per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
6384 }
6385 
wq_watchdog_timer_fn(struct timer_list * unused)6386 static void wq_watchdog_timer_fn(struct timer_list *unused)
6387 {
6388 	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
6389 	bool lockup_detected = false;
6390 	bool cpu_pool_stall = false;
6391 	unsigned long now = jiffies;
6392 	struct worker_pool *pool;
6393 	int pi;
6394 
6395 	if (!thresh)
6396 		return;
6397 
6398 	rcu_read_lock();
6399 
6400 	for_each_pool(pool, pi) {
6401 		unsigned long pool_ts, touched, ts;
6402 
6403 		pool->cpu_stall = false;
6404 		if (list_empty(&pool->worklist))
6405 			continue;
6406 
6407 		/*
6408 		 * If a virtual machine is stopped by the host it can look to
6409 		 * the watchdog like a stall.
6410 		 */
6411 		kvm_check_and_clear_guest_paused();
6412 
6413 		/* get the latest of pool and touched timestamps */
6414 		if (pool->cpu >= 0)
6415 			touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
6416 		else
6417 			touched = READ_ONCE(wq_watchdog_touched);
6418 		pool_ts = READ_ONCE(pool->watchdog_ts);
6419 
6420 		if (time_after(pool_ts, touched))
6421 			ts = pool_ts;
6422 		else
6423 			ts = touched;
6424 
6425 		/* did we stall? */
6426 		if (time_after(now, ts + thresh)) {
6427 			lockup_detected = true;
6428 			if (pool->cpu >= 0) {
6429 				pool->cpu_stall = true;
6430 				cpu_pool_stall = true;
6431 			}
6432 			pr_emerg("BUG: workqueue lockup - pool");
6433 			pr_cont_pool_info(pool);
6434 			pr_cont(" stuck for %us!\n",
6435 				jiffies_to_msecs(now - pool_ts) / 1000);
6436 		}
6437 
6438 
6439 	}
6440 
6441 	rcu_read_unlock();
6442 
6443 	if (lockup_detected)
6444 		show_all_workqueues();
6445 
6446 	if (cpu_pool_stall)
6447 		show_cpu_pools_hogs();
6448 
6449 	wq_watchdog_reset_touched();
6450 	mod_timer(&wq_watchdog_timer, jiffies + thresh);
6451 }
6452 
wq_watchdog_touch(int cpu)6453 notrace void wq_watchdog_touch(int cpu)
6454 {
6455 	if (cpu >= 0)
6456 		per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
6457 
6458 	wq_watchdog_touched = jiffies;
6459 }
6460 
wq_watchdog_set_thresh(unsigned long thresh)6461 static void wq_watchdog_set_thresh(unsigned long thresh)
6462 {
6463 	wq_watchdog_thresh = 0;
6464 	del_timer_sync(&wq_watchdog_timer);
6465 
6466 	if (thresh) {
6467 		wq_watchdog_thresh = thresh;
6468 		wq_watchdog_reset_touched();
6469 		mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
6470 	}
6471 }
6472 
wq_watchdog_param_set_thresh(const char * val,const struct kernel_param * kp)6473 static int wq_watchdog_param_set_thresh(const char *val,
6474 					const struct kernel_param *kp)
6475 {
6476 	unsigned long thresh;
6477 	int ret;
6478 
6479 	ret = kstrtoul(val, 0, &thresh);
6480 	if (ret)
6481 		return ret;
6482 
6483 	if (system_wq)
6484 		wq_watchdog_set_thresh(thresh);
6485 	else
6486 		wq_watchdog_thresh = thresh;
6487 
6488 	return 0;
6489 }
6490 
6491 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
6492 	.set	= wq_watchdog_param_set_thresh,
6493 	.get	= param_get_ulong,
6494 };
6495 
6496 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
6497 		0644);
6498 
wq_watchdog_init(void)6499 static void wq_watchdog_init(void)
6500 {
6501 	timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
6502 	wq_watchdog_set_thresh(wq_watchdog_thresh);
6503 }
6504 
6505 #else	/* CONFIG_WQ_WATCHDOG */
6506 
wq_watchdog_init(void)6507 static inline void wq_watchdog_init(void) { }
6508 
6509 #endif	/* CONFIG_WQ_WATCHDOG */
6510 
restrict_unbound_cpumask(const char * name,const struct cpumask * mask)6511 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
6512 {
6513 	if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
6514 		pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
6515 			cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
6516 		return;
6517 	}
6518 
6519 	cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
6520 }
6521 
6522 /**
6523  * workqueue_init_early - early init for workqueue subsystem
6524  *
6525  * This is the first step of three-staged workqueue subsystem initialization and
6526  * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
6527  * up. It sets up all the data structures and system workqueues and allows early
6528  * boot code to create workqueues and queue/cancel work items. Actual work item
6529  * execution starts only after kthreads can be created and scheduled right
6530  * before early initcalls.
6531  */
workqueue_init_early(void)6532 void __init workqueue_init_early(void)
6533 {
6534 	struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
6535 	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
6536 	int i, cpu;
6537 
6538 	BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
6539 
6540 	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
6541 	cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
6542 	restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
6543 	restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
6544 	if (!cpumask_empty(&wq_cmdline_cpumask))
6545 		restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
6546 
6547 	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
6548 
6549 	wq_update_pod_attrs_buf = alloc_workqueue_attrs();
6550 	BUG_ON(!wq_update_pod_attrs_buf);
6551 
6552 	/* initialize WQ_AFFN_SYSTEM pods */
6553 	pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
6554 	pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
6555 	pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
6556 	BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
6557 
6558 	BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
6559 
6560 	pt->nr_pods = 1;
6561 	cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
6562 	pt->pod_node[0] = NUMA_NO_NODE;
6563 	pt->cpu_pod[0] = 0;
6564 
6565 	/* initialize CPU pools */
6566 	for_each_possible_cpu(cpu) {
6567 		struct worker_pool *pool;
6568 
6569 		i = 0;
6570 		for_each_cpu_worker_pool(pool, cpu) {
6571 			BUG_ON(init_worker_pool(pool));
6572 			pool->cpu = cpu;
6573 			cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
6574 			cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
6575 			pool->attrs->nice = std_nice[i++];
6576 			pool->attrs->affn_strict = true;
6577 			pool->node = cpu_to_node(cpu);
6578 
6579 			/* alloc pool ID */
6580 			mutex_lock(&wq_pool_mutex);
6581 			BUG_ON(worker_pool_assign_id(pool));
6582 			mutex_unlock(&wq_pool_mutex);
6583 		}
6584 	}
6585 
6586 	/* create default unbound and ordered wq attrs */
6587 	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
6588 		struct workqueue_attrs *attrs;
6589 
6590 		BUG_ON(!(attrs = alloc_workqueue_attrs()));
6591 		attrs->nice = std_nice[i];
6592 		unbound_std_wq_attrs[i] = attrs;
6593 
6594 		/*
6595 		 * An ordered wq should have only one pwq as ordering is
6596 		 * guaranteed by max_active which is enforced by pwqs.
6597 		 */
6598 		BUG_ON(!(attrs = alloc_workqueue_attrs()));
6599 		attrs->nice = std_nice[i];
6600 		attrs->ordered = true;
6601 		ordered_wq_attrs[i] = attrs;
6602 	}
6603 
6604 	system_wq = alloc_workqueue("events", 0, 0);
6605 	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
6606 	system_long_wq = alloc_workqueue("events_long", 0, 0);
6607 	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
6608 					    WQ_MAX_ACTIVE);
6609 	system_freezable_wq = alloc_workqueue("events_freezable",
6610 					      WQ_FREEZABLE, 0);
6611 	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
6612 					      WQ_POWER_EFFICIENT, 0);
6613 	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
6614 					      WQ_FREEZABLE | WQ_POWER_EFFICIENT,
6615 					      0);
6616 	BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
6617 	       !system_unbound_wq || !system_freezable_wq ||
6618 	       !system_power_efficient_wq ||
6619 	       !system_freezable_power_efficient_wq);
6620 }
6621 
wq_cpu_intensive_thresh_init(void)6622 static void __init wq_cpu_intensive_thresh_init(void)
6623 {
6624 	unsigned long thresh;
6625 	unsigned long bogo;
6626 
6627 	pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
6628 	BUG_ON(IS_ERR(pwq_release_worker));
6629 
6630 	/* if the user set it to a specific value, keep it */
6631 	if (wq_cpu_intensive_thresh_us != ULONG_MAX)
6632 		return;
6633 
6634 	/*
6635 	 * The default of 10ms is derived from the fact that most modern (as of
6636 	 * 2023) processors can do a lot in 10ms and that it's just below what
6637 	 * most consider human-perceivable. However, the kernel also runs on a
6638 	 * lot slower CPUs including microcontrollers where the threshold is way
6639 	 * too low.
6640 	 *
6641 	 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
6642 	 * This is by no means accurate but it doesn't have to be. The mechanism
6643 	 * is still useful even when the threshold is fully scaled up. Also, as
6644 	 * the reports would usually be applicable to everyone, some machines
6645 	 * operating on longer thresholds won't significantly diminish their
6646 	 * usefulness.
6647 	 */
6648 	thresh = 10 * USEC_PER_MSEC;
6649 
6650 	/* see init/calibrate.c for lpj -> BogoMIPS calculation */
6651 	bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
6652 	if (bogo < 4000)
6653 		thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
6654 
6655 	pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
6656 		 loops_per_jiffy, bogo, thresh);
6657 
6658 	wq_cpu_intensive_thresh_us = thresh;
6659 }
6660 
6661 /**
6662  * workqueue_init - bring workqueue subsystem fully online
6663  *
6664  * This is the second step of three-staged workqueue subsystem initialization
6665  * and invoked as soon as kthreads can be created and scheduled. Workqueues have
6666  * been created and work items queued on them, but there are no kworkers
6667  * executing the work items yet. Populate the worker pools with the initial
6668  * workers and enable future kworker creations.
6669  */
workqueue_init(void)6670 void __init workqueue_init(void)
6671 {
6672 	struct workqueue_struct *wq;
6673 	struct worker_pool *pool;
6674 	int cpu, bkt;
6675 
6676 	wq_cpu_intensive_thresh_init();
6677 
6678 	mutex_lock(&wq_pool_mutex);
6679 
6680 	/*
6681 	 * Per-cpu pools created earlier could be missing node hint. Fix them
6682 	 * up. Also, create a rescuer for workqueues that requested it.
6683 	 */
6684 	for_each_possible_cpu(cpu) {
6685 		for_each_cpu_worker_pool(pool, cpu) {
6686 			pool->node = cpu_to_node(cpu);
6687 		}
6688 	}
6689 
6690 	list_for_each_entry(wq, &workqueues, list) {
6691 		WARN(init_rescuer(wq),
6692 		     "workqueue: failed to create early rescuer for %s",
6693 		     wq->name);
6694 	}
6695 
6696 	mutex_unlock(&wq_pool_mutex);
6697 
6698 	/* create the initial workers */
6699 	for_each_online_cpu(cpu) {
6700 		for_each_cpu_worker_pool(pool, cpu) {
6701 			pool->flags &= ~POOL_DISASSOCIATED;
6702 			BUG_ON(!create_worker(pool));
6703 		}
6704 	}
6705 
6706 	hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6707 		BUG_ON(!create_worker(pool));
6708 
6709 	wq_online = true;
6710 	wq_watchdog_init();
6711 }
6712 
6713 /*
6714  * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
6715  * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
6716  * and consecutive pod ID. The rest of @pt is initialized accordingly.
6717  */
init_pod_type(struct wq_pod_type * pt,bool (* cpus_share_pod)(int,int))6718 static void __init init_pod_type(struct wq_pod_type *pt,
6719 				 bool (*cpus_share_pod)(int, int))
6720 {
6721 	int cur, pre, cpu, pod;
6722 
6723 	pt->nr_pods = 0;
6724 
6725 	/* init @pt->cpu_pod[] according to @cpus_share_pod() */
6726 	pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
6727 	BUG_ON(!pt->cpu_pod);
6728 
6729 	for_each_possible_cpu(cur) {
6730 		for_each_possible_cpu(pre) {
6731 			if (pre >= cur) {
6732 				pt->cpu_pod[cur] = pt->nr_pods++;
6733 				break;
6734 			}
6735 			if (cpus_share_pod(cur, pre)) {
6736 				pt->cpu_pod[cur] = pt->cpu_pod[pre];
6737 				break;
6738 			}
6739 		}
6740 	}
6741 
6742 	/* init the rest to match @pt->cpu_pod[] */
6743 	pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
6744 	pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
6745 	BUG_ON(!pt->pod_cpus || !pt->pod_node);
6746 
6747 	for (pod = 0; pod < pt->nr_pods; pod++)
6748 		BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
6749 
6750 	for_each_possible_cpu(cpu) {
6751 		cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
6752 		pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
6753 	}
6754 }
6755 
cpus_dont_share(int cpu0,int cpu1)6756 static bool __init cpus_dont_share(int cpu0, int cpu1)
6757 {
6758 	return false;
6759 }
6760 
cpus_share_smt(int cpu0,int cpu1)6761 static bool __init cpus_share_smt(int cpu0, int cpu1)
6762 {
6763 #ifdef CONFIG_SCHED_SMT
6764 	return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
6765 #else
6766 	return false;
6767 #endif
6768 }
6769 
cpus_share_numa(int cpu0,int cpu1)6770 static bool __init cpus_share_numa(int cpu0, int cpu1)
6771 {
6772 	return cpu_to_node(cpu0) == cpu_to_node(cpu1);
6773 }
6774 
6775 /**
6776  * workqueue_init_topology - initialize CPU pods for unbound workqueues
6777  *
6778  * This is the third step of there-staged workqueue subsystem initialization and
6779  * invoked after SMP and topology information are fully initialized. It
6780  * initializes the unbound CPU pods accordingly.
6781  */
workqueue_init_topology(void)6782 void __init workqueue_init_topology(void)
6783 {
6784 	struct workqueue_struct *wq;
6785 	int cpu;
6786 
6787 	init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
6788 	init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
6789 	init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
6790 	init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
6791 
6792 	mutex_lock(&wq_pool_mutex);
6793 
6794 	/*
6795 	 * Workqueues allocated earlier would have all CPUs sharing the default
6796 	 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
6797 	 * combinations to apply per-pod sharing.
6798 	 */
6799 	list_for_each_entry(wq, &workqueues, list) {
6800 		for_each_online_cpu(cpu) {
6801 			wq_update_pod(wq, cpu, cpu, true);
6802 		}
6803 	}
6804 
6805 	mutex_unlock(&wq_pool_mutex);
6806 }
6807 
__warn_flushing_systemwide_wq(void)6808 void __warn_flushing_systemwide_wq(void)
6809 {
6810 	pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
6811 	dump_stack();
6812 }
6813 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
6814 
workqueue_unbound_cpus_setup(char * str)6815 static int __init workqueue_unbound_cpus_setup(char *str)
6816 {
6817 	if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
6818 		cpumask_clear(&wq_cmdline_cpumask);
6819 		pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
6820 	}
6821 
6822 	return 1;
6823 }
6824 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);
6825