xref: /linux/kernel/workqueue.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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/interrupt.h>
33 #include <linux/signal.h>
34 #include <linux/completion.h>
35 #include <linux/workqueue.h>
36 #include <linux/slab.h>
37 #include <linux/cpu.h>
38 #include <linux/notifier.h>
39 #include <linux/kthread.h>
40 #include <linux/hardirq.h>
41 #include <linux/mempolicy.h>
42 #include <linux/freezer.h>
43 #include <linux/debug_locks.h>
44 #include <linux/lockdep.h>
45 #include <linux/idr.h>
46 #include <linux/jhash.h>
47 #include <linux/hashtable.h>
48 #include <linux/rculist.h>
49 #include <linux/nodemask.h>
50 #include <linux/moduleparam.h>
51 #include <linux/uaccess.h>
52 #include <linux/sched/isolation.h>
53 #include <linux/sched/debug.h>
54 #include <linux/nmi.h>
55 #include <linux/kvm_para.h>
56 #include <linux/delay.h>
57 #include <linux/irq_work.h>
58 
59 #include "workqueue_internal.h"
60 
61 enum worker_pool_flags {
62 	/*
63 	 * worker_pool flags
64 	 *
65 	 * A bound pool is either associated or disassociated with its CPU.
66 	 * While associated (!DISASSOCIATED), all workers are bound to the
67 	 * CPU and none has %WORKER_UNBOUND set and concurrency management
68 	 * is in effect.
69 	 *
70 	 * While DISASSOCIATED, the cpu may be offline and all workers have
71 	 * %WORKER_UNBOUND set and concurrency management disabled, and may
72 	 * be executing on any CPU.  The pool behaves as an unbound one.
73 	 *
74 	 * Note that DISASSOCIATED should be flipped only while holding
75 	 * wq_pool_attach_mutex to avoid changing binding state while
76 	 * worker_attach_to_pool() is in progress.
77 	 *
78 	 * As there can only be one concurrent BH execution context per CPU, a
79 	 * BH pool is per-CPU and always DISASSOCIATED.
80 	 */
81 	POOL_BH			= 1 << 0,	/* is a BH pool */
82 	POOL_MANAGER_ACTIVE	= 1 << 1,	/* being managed */
83 	POOL_DISASSOCIATED	= 1 << 2,	/* cpu can't serve workers */
84 	POOL_BH_DRAINING	= 1 << 3,	/* draining after CPU offline */
85 };
86 
87 enum worker_flags {
88 	/* worker flags */
89 	WORKER_DIE		= 1 << 1,	/* die die die */
90 	WORKER_IDLE		= 1 << 2,	/* is idle */
91 	WORKER_PREP		= 1 << 3,	/* preparing to run works */
92 	WORKER_CPU_INTENSIVE	= 1 << 6,	/* cpu intensive */
93 	WORKER_UNBOUND		= 1 << 7,	/* worker is unbound */
94 	WORKER_REBOUND		= 1 << 8,	/* worker was rebound */
95 
96 	WORKER_NOT_RUNNING	= WORKER_PREP | WORKER_CPU_INTENSIVE |
97 				  WORKER_UNBOUND | WORKER_REBOUND,
98 };
99 
100 enum work_cancel_flags {
101 	WORK_CANCEL_DELAYED	= 1 << 0,	/* canceling a delayed_work */
102 	WORK_CANCEL_DISABLE	= 1 << 1,	/* canceling to disable */
103 };
104 
105 enum wq_internal_consts {
106 	NR_STD_WORKER_POOLS	= 2,		/* # standard pools per cpu */
107 
108 	UNBOUND_POOL_HASH_ORDER	= 6,		/* hashed by pool->attrs */
109 	BUSY_WORKER_HASH_ORDER	= 6,		/* 64 pointers */
110 
111 	MAX_IDLE_WORKERS_RATIO	= 4,		/* 1/4 of busy can be idle */
112 	IDLE_WORKER_TIMEOUT	= 300 * HZ,	/* keep idle ones for 5 mins */
113 
114 	MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
115 						/* call for help after 10ms
116 						   (min two ticks) */
117 	MAYDAY_INTERVAL		= HZ / 10,	/* and then every 100ms */
118 	CREATE_COOLDOWN		= HZ,		/* time to breath after fail */
119 
120 	/*
121 	 * Rescue workers are used only on emergencies and shared by
122 	 * all cpus.  Give MIN_NICE.
123 	 */
124 	RESCUER_NICE_LEVEL	= MIN_NICE,
125 	HIGHPRI_NICE_LEVEL	= MIN_NICE,
126 
127 	WQ_NAME_LEN		= 32,
128 	WORKER_ID_LEN		= 10 + WQ_NAME_LEN, /* "kworker/R-" + WQ_NAME_LEN */
129 };
130 
131 /*
132  * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and
133  * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because
134  * msecs_to_jiffies() can't be an initializer.
135  */
136 #define BH_WORKER_JIFFIES	msecs_to_jiffies(2)
137 #define BH_WORKER_RESTARTS	10
138 
139 /*
140  * Structure fields follow one of the following exclusion rules.
141  *
142  * I: Modifiable by initialization/destruction paths and read-only for
143  *    everyone else.
144  *
145  * P: Preemption protected.  Disabling preemption is enough and should
146  *    only be modified and accessed from the local cpu.
147  *
148  * L: pool->lock protected.  Access with pool->lock held.
149  *
150  * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for
151  *     reads.
152  *
153  * K: Only modified by worker while holding pool->lock. Can be safely read by
154  *    self, while holding pool->lock or from IRQ context if %current is the
155  *    kworker.
156  *
157  * S: Only modified by worker self.
158  *
159  * A: wq_pool_attach_mutex protected.
160  *
161  * PL: wq_pool_mutex protected.
162  *
163  * PR: wq_pool_mutex protected for writes.  RCU protected for reads.
164  *
165  * PW: wq_pool_mutex and wq->mutex protected for writes.  Either for reads.
166  *
167  * PWR: wq_pool_mutex and wq->mutex protected for writes.  Either or
168  *      RCU for reads.
169  *
170  * WQ: wq->mutex protected.
171  *
172  * WR: wq->mutex protected for writes.  RCU protected for reads.
173  *
174  * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read
175  *     with READ_ONCE() without locking.
176  *
177  * MD: wq_mayday_lock protected.
178  *
179  * WD: Used internally by the watchdog.
180  */
181 
182 /* struct worker is defined in workqueue_internal.h */
183 
184 struct worker_pool {
185 	raw_spinlock_t		lock;		/* the pool lock */
186 	int			cpu;		/* I: the associated cpu */
187 	int			node;		/* I: the associated node ID */
188 	int			id;		/* I: pool ID */
189 	unsigned int		flags;		/* L: flags */
190 
191 	unsigned long		watchdog_ts;	/* L: watchdog timestamp */
192 	bool			cpu_stall;	/* WD: stalled cpu bound pool */
193 
194 	/*
195 	 * The counter is incremented in a process context on the associated CPU
196 	 * w/ preemption disabled, and decremented or reset in the same context
197 	 * but w/ pool->lock held. The readers grab pool->lock and are
198 	 * guaranteed to see if the counter reached zero.
199 	 */
200 	int			nr_running;
201 
202 	struct list_head	worklist;	/* L: list of pending works */
203 
204 	int			nr_workers;	/* L: total number of workers */
205 	int			nr_idle;	/* L: currently idle workers */
206 
207 	struct list_head	idle_list;	/* L: list of idle workers */
208 	struct timer_list	idle_timer;	/* L: worker idle timeout */
209 	struct work_struct      idle_cull_work; /* L: worker idle cleanup */
210 
211 	struct timer_list	mayday_timer;	  /* L: SOS timer for workers */
212 
213 	/* a workers is either on busy_hash or idle_list, or the manager */
214 	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
215 						/* L: hash of busy workers */
216 
217 	struct worker		*manager;	/* L: purely informational */
218 	struct list_head	workers;	/* A: attached workers */
219 
220 	struct ida		worker_ida;	/* worker IDs for task name */
221 
222 	struct workqueue_attrs	*attrs;		/* I: worker attributes */
223 	struct hlist_node	hash_node;	/* PL: unbound_pool_hash node */
224 	int			refcnt;		/* PL: refcnt for unbound pools */
225 
226 	/*
227 	 * Destruction of pool is RCU protected to allow dereferences
228 	 * from get_work_pool().
229 	 */
230 	struct rcu_head		rcu;
231 };
232 
233 /*
234  * Per-pool_workqueue statistics. These can be monitored using
235  * tools/workqueue/wq_monitor.py.
236  */
237 enum pool_workqueue_stats {
238 	PWQ_STAT_STARTED,	/* work items started execution */
239 	PWQ_STAT_COMPLETED,	/* work items completed execution */
240 	PWQ_STAT_CPU_TIME,	/* total CPU time consumed */
241 	PWQ_STAT_CPU_INTENSIVE,	/* wq_cpu_intensive_thresh_us violations */
242 	PWQ_STAT_CM_WAKEUP,	/* concurrency-management worker wakeups */
243 	PWQ_STAT_REPATRIATED,	/* unbound workers brought back into scope */
244 	PWQ_STAT_MAYDAY,	/* maydays to rescuer */
245 	PWQ_STAT_RESCUED,	/* linked work items executed by rescuer */
246 
247 	PWQ_NR_STATS,
248 };
249 
250 /*
251  * The per-pool workqueue.  While queued, bits below WORK_PWQ_SHIFT
252  * of work_struct->data are used for flags and the remaining high bits
253  * point to the pwq; thus, pwqs need to be aligned at two's power of the
254  * number of flag bits.
255  */
256 struct pool_workqueue {
257 	struct worker_pool	*pool;		/* I: the associated pool */
258 	struct workqueue_struct *wq;		/* I: the owning workqueue */
259 	int			work_color;	/* L: current color */
260 	int			flush_color;	/* L: flushing color */
261 	int			refcnt;		/* L: reference count */
262 	int			nr_in_flight[WORK_NR_COLORS];
263 						/* L: nr of in_flight works */
264 	bool			plugged;	/* L: execution suspended */
265 
266 	/*
267 	 * nr_active management and WORK_STRUCT_INACTIVE:
268 	 *
269 	 * When pwq->nr_active >= max_active, new work item is queued to
270 	 * pwq->inactive_works instead of pool->worklist and marked with
271 	 * WORK_STRUCT_INACTIVE.
272 	 *
273 	 * All work items marked with WORK_STRUCT_INACTIVE do not participate in
274 	 * nr_active and all work items in pwq->inactive_works are marked with
275 	 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are
276 	 * in pwq->inactive_works. Some of them are ready to run in
277 	 * pool->worklist or worker->scheduled. Those work itmes are only struct
278 	 * wq_barrier which is used for flush_work() and should not participate
279 	 * in nr_active. For non-barrier work item, it is marked with
280 	 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
281 	 */
282 	int			nr_active;	/* L: nr of active works */
283 	struct list_head	inactive_works;	/* L: inactive works */
284 	struct list_head	pending_node;	/* LN: node on wq_node_nr_active->pending_pwqs */
285 	struct list_head	pwqs_node;	/* WR: node on wq->pwqs */
286 	struct list_head	mayday_node;	/* MD: node on wq->maydays */
287 
288 	u64			stats[PWQ_NR_STATS];
289 
290 	/*
291 	 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
292 	 * and pwq_release_workfn() for details. pool_workqueue itself is also
293 	 * RCU protected so that the first pwq can be determined without
294 	 * grabbing wq->mutex.
295 	 */
296 	struct kthread_work	release_work;
297 	struct rcu_head		rcu;
298 } __aligned(1 << WORK_STRUCT_PWQ_SHIFT);
299 
300 /*
301  * Structure used to wait for workqueue flush.
302  */
303 struct wq_flusher {
304 	struct list_head	list;		/* WQ: list of flushers */
305 	int			flush_color;	/* WQ: flush color waiting for */
306 	struct completion	done;		/* flush completion */
307 };
308 
309 struct wq_device;
310 
311 /*
312  * Unlike in a per-cpu workqueue where max_active limits its concurrency level
313  * on each CPU, in an unbound workqueue, max_active applies to the whole system.
314  * As sharing a single nr_active across multiple sockets can be very expensive,
315  * the counting and enforcement is per NUMA node.
316  *
317  * The following struct is used to enforce per-node max_active. When a pwq wants
318  * to start executing a work item, it should increment ->nr using
319  * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over
320  * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish
321  * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in
322  * round-robin order.
323  */
324 struct wq_node_nr_active {
325 	int			max;		/* per-node max_active */
326 	atomic_t		nr;		/* per-node nr_active */
327 	raw_spinlock_t		lock;		/* nests inside pool locks */
328 	struct list_head	pending_pwqs;	/* LN: pwqs with inactive works */
329 };
330 
331 /*
332  * The externally visible workqueue.  It relays the issued work items to
333  * the appropriate worker_pool through its pool_workqueues.
334  */
335 struct workqueue_struct {
336 	struct list_head	pwqs;		/* WR: all pwqs of this wq */
337 	struct list_head	list;		/* PR: list of all workqueues */
338 
339 	struct mutex		mutex;		/* protects this wq */
340 	int			work_color;	/* WQ: current work color */
341 	int			flush_color;	/* WQ: current flush color */
342 	atomic_t		nr_pwqs_to_flush; /* flush in progress */
343 	struct wq_flusher	*first_flusher;	/* WQ: first flusher */
344 	struct list_head	flusher_queue;	/* WQ: flush waiters */
345 	struct list_head	flusher_overflow; /* WQ: flush overflow list */
346 
347 	struct list_head	maydays;	/* MD: pwqs requesting rescue */
348 	struct worker		*rescuer;	/* MD: rescue worker */
349 
350 	int			nr_drainers;	/* WQ: drain in progress */
351 
352 	/* See alloc_workqueue() function comment for info on min/max_active */
353 	int			max_active;	/* WO: max active works */
354 	int			min_active;	/* WO: min active works */
355 	int			saved_max_active; /* WQ: saved max_active */
356 	int			saved_min_active; /* WQ: saved min_active */
357 
358 	struct workqueue_attrs	*unbound_attrs;	/* PW: only for unbound wqs */
359 	struct pool_workqueue __rcu *dfl_pwq;   /* PW: only for unbound wqs */
360 
361 #ifdef CONFIG_SYSFS
362 	struct wq_device	*wq_dev;	/* I: for sysfs interface */
363 #endif
364 #ifdef CONFIG_LOCKDEP
365 	char			*lock_name;
366 	struct lock_class_key	key;
367 	struct lockdep_map	__lockdep_map;
368 	struct lockdep_map	*lockdep_map;
369 #endif
370 	char			name[WQ_NAME_LEN]; /* I: workqueue name */
371 
372 	/*
373 	 * Destruction of workqueue_struct is RCU protected to allow walking
374 	 * the workqueues list without grabbing wq_pool_mutex.
375 	 * This is used to dump all workqueues from sysrq.
376 	 */
377 	struct rcu_head		rcu;
378 
379 	/* hot fields used during command issue, aligned to cacheline */
380 	unsigned int		flags ____cacheline_aligned; /* WQ: WQ_* flags */
381 	struct pool_workqueue __rcu * __percpu *cpu_pwq; /* I: per-cpu pwqs */
382 	struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */
383 };
384 
385 /*
386  * Each pod type describes how CPUs should be grouped for unbound workqueues.
387  * See the comment above workqueue_attrs->affn_scope.
388  */
389 struct wq_pod_type {
390 	int			nr_pods;	/* number of pods */
391 	cpumask_var_t		*pod_cpus;	/* pod -> cpus */
392 	int			*pod_node;	/* pod -> node */
393 	int			*cpu_pod;	/* cpu -> pod */
394 };
395 
396 struct work_offq_data {
397 	u32			pool_id;
398 	u32			disable;
399 	u32			flags;
400 };
401 
402 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
403 	[WQ_AFFN_DFL]		= "default",
404 	[WQ_AFFN_CPU]		= "cpu",
405 	[WQ_AFFN_SMT]		= "smt",
406 	[WQ_AFFN_CACHE]		= "cache",
407 	[WQ_AFFN_NUMA]		= "numa",
408 	[WQ_AFFN_SYSTEM]	= "system",
409 };
410 
411 /*
412  * Per-cpu work items which run for longer than the following threshold are
413  * automatically considered CPU intensive and excluded from concurrency
414  * management to prevent them from noticeably delaying other per-cpu work items.
415  * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
416  * The actual value is initialized in wq_cpu_intensive_thresh_init().
417  */
418 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
419 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
420 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
421 static unsigned int wq_cpu_intensive_warning_thresh = 4;
422 module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644);
423 #endif
424 
425 /* see the comment above the definition of WQ_POWER_EFFICIENT */
426 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
427 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
428 
429 static bool wq_online;			/* can kworkers be created yet? */
430 static bool wq_topo_initialized __read_mostly = false;
431 
432 static struct kmem_cache *pwq_cache;
433 
434 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
435 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
436 
437 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
438 static struct workqueue_attrs *unbound_wq_update_pwq_attrs_buf;
439 
440 static DEFINE_MUTEX(wq_pool_mutex);	/* protects pools and workqueues list */
441 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
442 static DEFINE_RAW_SPINLOCK(wq_mayday_lock);	/* protects wq->maydays list */
443 /* wait for manager to go away */
444 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
445 
446 static LIST_HEAD(workqueues);		/* PR: list of all workqueues */
447 static bool workqueue_freezing;		/* PL: have wqs started freezing? */
448 
449 /* PL: mirror the cpu_online_mask excluding the CPU in the midst of hotplugging */
450 static cpumask_var_t wq_online_cpumask;
451 
452 /* PL&A: allowable cpus for unbound wqs and work items */
453 static cpumask_var_t wq_unbound_cpumask;
454 
455 /* PL: user requested unbound cpumask via sysfs */
456 static cpumask_var_t wq_requested_unbound_cpumask;
457 
458 /* PL: isolated cpumask to be excluded from unbound cpumask */
459 static cpumask_var_t wq_isolated_cpumask;
460 
461 /* for further constrain wq_unbound_cpumask by cmdline parameter*/
462 static struct cpumask wq_cmdline_cpumask __initdata;
463 
464 /* CPU where unbound work was last round robin scheduled from this CPU */
465 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
466 
467 /*
468  * Local execution of unbound work items is no longer guaranteed.  The
469  * following always forces round-robin CPU selection on unbound work items
470  * to uncover usages which depend on it.
471  */
472 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
473 static bool wq_debug_force_rr_cpu = true;
474 #else
475 static bool wq_debug_force_rr_cpu = false;
476 #endif
477 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
478 
479 /* to raise softirq for the BH worker pools on other CPUs */
480 static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS], bh_pool_irq_works);
481 
482 /* the BH worker pools */
483 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], bh_worker_pools);
484 
485 /* the per-cpu worker pools */
486 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
487 
488 static DEFINE_IDR(worker_pool_idr);	/* PR: idr of all pools */
489 
490 /* PL: hash of all unbound pools keyed by pool->attrs */
491 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
492 
493 /* I: attributes used when instantiating standard unbound pools on demand */
494 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
495 
496 /* I: attributes used when instantiating ordered pools on demand */
497 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
498 
499 /*
500  * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
501  * process context while holding a pool lock. Bounce to a dedicated kthread
502  * worker to avoid A-A deadlocks.
503  */
504 static struct kthread_worker *pwq_release_worker __ro_after_init;
505 
506 struct workqueue_struct *system_wq __ro_after_init;
507 EXPORT_SYMBOL(system_wq);
508 struct workqueue_struct *system_highpri_wq __ro_after_init;
509 EXPORT_SYMBOL_GPL(system_highpri_wq);
510 struct workqueue_struct *system_long_wq __ro_after_init;
511 EXPORT_SYMBOL_GPL(system_long_wq);
512 struct workqueue_struct *system_unbound_wq __ro_after_init;
513 EXPORT_SYMBOL_GPL(system_unbound_wq);
514 struct workqueue_struct *system_freezable_wq __ro_after_init;
515 EXPORT_SYMBOL_GPL(system_freezable_wq);
516 struct workqueue_struct *system_power_efficient_wq __ro_after_init;
517 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
518 struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
519 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
520 struct workqueue_struct *system_bh_wq;
521 EXPORT_SYMBOL_GPL(system_bh_wq);
522 struct workqueue_struct *system_bh_highpri_wq;
523 EXPORT_SYMBOL_GPL(system_bh_highpri_wq);
524 
525 static int worker_thread(void *__worker);
526 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
527 static void show_pwq(struct pool_workqueue *pwq);
528 static void show_one_worker_pool(struct worker_pool *pool);
529 
530 #define CREATE_TRACE_POINTS
531 #include <trace/events/workqueue.h>
532 
533 #define assert_rcu_or_pool_mutex()					\
534 	RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() &&			\
535 			 !lockdep_is_held(&wq_pool_mutex),		\
536 			 "RCU or wq_pool_mutex should be held")
537 
538 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq)			\
539 	RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() &&			\
540 			 !lockdep_is_held(&wq->mutex) &&		\
541 			 !lockdep_is_held(&wq_pool_mutex),		\
542 			 "RCU, wq->mutex or wq_pool_mutex should be held")
543 
544 #define for_each_bh_worker_pool(pool, cpu)				\
545 	for ((pool) = &per_cpu(bh_worker_pools, cpu)[0];		\
546 	     (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
547 	     (pool)++)
548 
549 #define for_each_cpu_worker_pool(pool, cpu)				\
550 	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0];		\
551 	     (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
552 	     (pool)++)
553 
554 /**
555  * for_each_pool - iterate through all worker_pools in the system
556  * @pool: iteration cursor
557  * @pi: integer used for iteration
558  *
559  * This must be called either with wq_pool_mutex held or RCU read
560  * locked.  If the pool needs to be used beyond the locking in effect, the
561  * caller is responsible for guaranteeing that the pool stays online.
562  *
563  * The if/else clause exists only for the lockdep assertion and can be
564  * ignored.
565  */
566 #define for_each_pool(pool, pi)						\
567 	idr_for_each_entry(&worker_pool_idr, pool, pi)			\
568 		if (({ assert_rcu_or_pool_mutex(); false; })) { }	\
569 		else
570 
571 /**
572  * for_each_pool_worker - iterate through all workers of a worker_pool
573  * @worker: iteration cursor
574  * @pool: worker_pool to iterate workers of
575  *
576  * This must be called with wq_pool_attach_mutex.
577  *
578  * The if/else clause exists only for the lockdep assertion and can be
579  * ignored.
580  */
581 #define for_each_pool_worker(worker, pool)				\
582 	list_for_each_entry((worker), &(pool)->workers, node)		\
583 		if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
584 		else
585 
586 /**
587  * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
588  * @pwq: iteration cursor
589  * @wq: the target workqueue
590  *
591  * This must be called either with wq->mutex held or RCU read locked.
592  * If the pwq needs to be used beyond the locking in effect, the caller is
593  * responsible for guaranteeing that the pwq stays online.
594  *
595  * The if/else clause exists only for the lockdep assertion and can be
596  * ignored.
597  */
598 #define for_each_pwq(pwq, wq)						\
599 	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node,		\
600 				 lockdep_is_held(&(wq->mutex)))
601 
602 #ifdef CONFIG_DEBUG_OBJECTS_WORK
603 
604 static const struct debug_obj_descr work_debug_descr;
605 
606 static void *work_debug_hint(void *addr)
607 {
608 	return ((struct work_struct *) addr)->func;
609 }
610 
611 static bool work_is_static_object(void *addr)
612 {
613 	struct work_struct *work = addr;
614 
615 	return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
616 }
617 
618 /*
619  * fixup_init is called when:
620  * - an active object is initialized
621  */
622 static bool work_fixup_init(void *addr, enum debug_obj_state state)
623 {
624 	struct work_struct *work = addr;
625 
626 	switch (state) {
627 	case ODEBUG_STATE_ACTIVE:
628 		cancel_work_sync(work);
629 		debug_object_init(work, &work_debug_descr);
630 		return true;
631 	default:
632 		return false;
633 	}
634 }
635 
636 /*
637  * fixup_free is called when:
638  * - an active object is freed
639  */
640 static bool work_fixup_free(void *addr, enum debug_obj_state state)
641 {
642 	struct work_struct *work = addr;
643 
644 	switch (state) {
645 	case ODEBUG_STATE_ACTIVE:
646 		cancel_work_sync(work);
647 		debug_object_free(work, &work_debug_descr);
648 		return true;
649 	default:
650 		return false;
651 	}
652 }
653 
654 static const struct debug_obj_descr work_debug_descr = {
655 	.name		= "work_struct",
656 	.debug_hint	= work_debug_hint,
657 	.is_static_object = work_is_static_object,
658 	.fixup_init	= work_fixup_init,
659 	.fixup_free	= work_fixup_free,
660 };
661 
662 static inline void debug_work_activate(struct work_struct *work)
663 {
664 	debug_object_activate(work, &work_debug_descr);
665 }
666 
667 static inline void debug_work_deactivate(struct work_struct *work)
668 {
669 	debug_object_deactivate(work, &work_debug_descr);
670 }
671 
672 void __init_work(struct work_struct *work, int onstack)
673 {
674 	if (onstack)
675 		debug_object_init_on_stack(work, &work_debug_descr);
676 	else
677 		debug_object_init(work, &work_debug_descr);
678 }
679 EXPORT_SYMBOL_GPL(__init_work);
680 
681 void destroy_work_on_stack(struct work_struct *work)
682 {
683 	debug_object_free(work, &work_debug_descr);
684 }
685 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
686 
687 void destroy_delayed_work_on_stack(struct delayed_work *work)
688 {
689 	destroy_timer_on_stack(&work->timer);
690 	debug_object_free(&work->work, &work_debug_descr);
691 }
692 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
693 
694 #else
695 static inline void debug_work_activate(struct work_struct *work) { }
696 static inline void debug_work_deactivate(struct work_struct *work) { }
697 #endif
698 
699 /**
700  * worker_pool_assign_id - allocate ID and assign it to @pool
701  * @pool: the pool pointer of interest
702  *
703  * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
704  * successfully, -errno on failure.
705  */
706 static int worker_pool_assign_id(struct worker_pool *pool)
707 {
708 	int ret;
709 
710 	lockdep_assert_held(&wq_pool_mutex);
711 
712 	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
713 			GFP_KERNEL);
714 	if (ret >= 0) {
715 		pool->id = ret;
716 		return 0;
717 	}
718 	return ret;
719 }
720 
721 static struct pool_workqueue __rcu **
722 unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
723 {
724        if (cpu >= 0)
725                return per_cpu_ptr(wq->cpu_pwq, cpu);
726        else
727                return &wq->dfl_pwq;
728 }
729 
730 /* @cpu < 0 for dfl_pwq */
731 static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu)
732 {
733 	return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
734 				     lockdep_is_held(&wq_pool_mutex) ||
735 				     lockdep_is_held(&wq->mutex));
736 }
737 
738 /**
739  * unbound_effective_cpumask - effective cpumask of an unbound workqueue
740  * @wq: workqueue of interest
741  *
742  * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which
743  * is masked with wq_unbound_cpumask to determine the effective cpumask. The
744  * default pwq is always mapped to the pool with the current effective cpumask.
745  */
746 static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq)
747 {
748 	return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask;
749 }
750 
751 static unsigned int work_color_to_flags(int color)
752 {
753 	return color << WORK_STRUCT_COLOR_SHIFT;
754 }
755 
756 static int get_work_color(unsigned long work_data)
757 {
758 	return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
759 		((1 << WORK_STRUCT_COLOR_BITS) - 1);
760 }
761 
762 static int work_next_color(int color)
763 {
764 	return (color + 1) % WORK_NR_COLORS;
765 }
766 
767 static unsigned long pool_offq_flags(struct worker_pool *pool)
768 {
769 	return (pool->flags & POOL_BH) ? WORK_OFFQ_BH : 0;
770 }
771 
772 /*
773  * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
774  * contain the pointer to the queued pwq.  Once execution starts, the flag
775  * is cleared and the high bits contain OFFQ flags and pool ID.
776  *
777  * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling()
778  * can be used to set the pwq, pool or clear work->data. These functions should
779  * only be called while the work is owned - ie. while the PENDING bit is set.
780  *
781  * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
782  * corresponding to a work.  Pool is available once the work has been
783  * queued anywhere after initialization until it is sync canceled.  pwq is
784  * available only while the work item is queued.
785  */
786 static inline void set_work_data(struct work_struct *work, unsigned long data)
787 {
788 	WARN_ON_ONCE(!work_pending(work));
789 	atomic_long_set(&work->data, data | work_static(work));
790 }
791 
792 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
793 			 unsigned long flags)
794 {
795 	set_work_data(work, (unsigned long)pwq | WORK_STRUCT_PENDING |
796 		      WORK_STRUCT_PWQ | flags);
797 }
798 
799 static void set_work_pool_and_keep_pending(struct work_struct *work,
800 					   int pool_id, unsigned long flags)
801 {
802 	set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
803 		      WORK_STRUCT_PENDING | flags);
804 }
805 
806 static void set_work_pool_and_clear_pending(struct work_struct *work,
807 					    int pool_id, unsigned long flags)
808 {
809 	/*
810 	 * The following wmb is paired with the implied mb in
811 	 * test_and_set_bit(PENDING) and ensures all updates to @work made
812 	 * here are visible to and precede any updates by the next PENDING
813 	 * owner.
814 	 */
815 	smp_wmb();
816 	set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
817 		      flags);
818 	/*
819 	 * The following mb guarantees that previous clear of a PENDING bit
820 	 * will not be reordered with any speculative LOADS or STORES from
821 	 * work->current_func, which is executed afterwards.  This possible
822 	 * reordering can lead to a missed execution on attempt to queue
823 	 * the same @work.  E.g. consider this case:
824 	 *
825 	 *   CPU#0                         CPU#1
826 	 *   ----------------------------  --------------------------------
827 	 *
828 	 * 1  STORE event_indicated
829 	 * 2  queue_work_on() {
830 	 * 3    test_and_set_bit(PENDING)
831 	 * 4 }                             set_..._and_clear_pending() {
832 	 * 5                                 set_work_data() # clear bit
833 	 * 6                                 smp_mb()
834 	 * 7                               work->current_func() {
835 	 * 8				      LOAD event_indicated
836 	 *				   }
837 	 *
838 	 * Without an explicit full barrier speculative LOAD on line 8 can
839 	 * be executed before CPU#0 does STORE on line 1.  If that happens,
840 	 * CPU#0 observes the PENDING bit is still set and new execution of
841 	 * a @work is not queued in a hope, that CPU#1 will eventually
842 	 * finish the queued @work.  Meanwhile CPU#1 does not see
843 	 * event_indicated is set, because speculative LOAD was executed
844 	 * before actual STORE.
845 	 */
846 	smp_mb();
847 }
848 
849 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
850 {
851 	return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK);
852 }
853 
854 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
855 {
856 	unsigned long data = atomic_long_read(&work->data);
857 
858 	if (data & WORK_STRUCT_PWQ)
859 		return work_struct_pwq(data);
860 	else
861 		return NULL;
862 }
863 
864 /**
865  * get_work_pool - return the worker_pool a given work was associated with
866  * @work: the work item of interest
867  *
868  * Pools are created and destroyed under wq_pool_mutex, and allows read
869  * access under RCU read lock.  As such, this function should be
870  * called under wq_pool_mutex or inside of a rcu_read_lock() region.
871  *
872  * All fields of the returned pool are accessible as long as the above
873  * mentioned locking is in effect.  If the returned pool needs to be used
874  * beyond the critical section, the caller is responsible for ensuring the
875  * returned pool is and stays online.
876  *
877  * Return: The worker_pool @work was last associated with.  %NULL if none.
878  */
879 static struct worker_pool *get_work_pool(struct work_struct *work)
880 {
881 	unsigned long data = atomic_long_read(&work->data);
882 	int pool_id;
883 
884 	assert_rcu_or_pool_mutex();
885 
886 	if (data & WORK_STRUCT_PWQ)
887 		return work_struct_pwq(data)->pool;
888 
889 	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
890 	if (pool_id == WORK_OFFQ_POOL_NONE)
891 		return NULL;
892 
893 	return idr_find(&worker_pool_idr, pool_id);
894 }
895 
896 static unsigned long shift_and_mask(unsigned long v, u32 shift, u32 bits)
897 {
898 	return (v >> shift) & ((1U << bits) - 1);
899 }
900 
901 static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data)
902 {
903 	WARN_ON_ONCE(data & WORK_STRUCT_PWQ);
904 
905 	offqd->pool_id = shift_and_mask(data, WORK_OFFQ_POOL_SHIFT,
906 					WORK_OFFQ_POOL_BITS);
907 	offqd->disable = shift_and_mask(data, WORK_OFFQ_DISABLE_SHIFT,
908 					WORK_OFFQ_DISABLE_BITS);
909 	offqd->flags = data & WORK_OFFQ_FLAG_MASK;
910 }
911 
912 static unsigned long work_offqd_pack_flags(struct work_offq_data *offqd)
913 {
914 	return ((unsigned long)offqd->disable << WORK_OFFQ_DISABLE_SHIFT) |
915 		((unsigned long)offqd->flags);
916 }
917 
918 /*
919  * Policy functions.  These define the policies on how the global worker
920  * pools are managed.  Unless noted otherwise, these functions assume that
921  * they're being called with pool->lock held.
922  */
923 
924 /*
925  * Need to wake up a worker?  Called from anything but currently
926  * running workers.
927  *
928  * Note that, because unbound workers never contribute to nr_running, this
929  * function will always return %true for unbound pools as long as the
930  * worklist isn't empty.
931  */
932 static bool need_more_worker(struct worker_pool *pool)
933 {
934 	return !list_empty(&pool->worklist) && !pool->nr_running;
935 }
936 
937 /* Can I start working?  Called from busy but !running workers. */
938 static bool may_start_working(struct worker_pool *pool)
939 {
940 	return pool->nr_idle;
941 }
942 
943 /* Do I need to keep working?  Called from currently running workers. */
944 static bool keep_working(struct worker_pool *pool)
945 {
946 	return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
947 }
948 
949 /* Do we need a new worker?  Called from manager. */
950 static bool need_to_create_worker(struct worker_pool *pool)
951 {
952 	return need_more_worker(pool) && !may_start_working(pool);
953 }
954 
955 /* Do we have too many workers and should some go away? */
956 static bool too_many_workers(struct worker_pool *pool)
957 {
958 	bool managing = pool->flags & POOL_MANAGER_ACTIVE;
959 	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
960 	int nr_busy = pool->nr_workers - nr_idle;
961 
962 	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
963 }
964 
965 /**
966  * worker_set_flags - set worker flags and adjust nr_running accordingly
967  * @worker: self
968  * @flags: flags to set
969  *
970  * Set @flags in @worker->flags and adjust nr_running accordingly.
971  */
972 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
973 {
974 	struct worker_pool *pool = worker->pool;
975 
976 	lockdep_assert_held(&pool->lock);
977 
978 	/* If transitioning into NOT_RUNNING, adjust nr_running. */
979 	if ((flags & WORKER_NOT_RUNNING) &&
980 	    !(worker->flags & WORKER_NOT_RUNNING)) {
981 		pool->nr_running--;
982 	}
983 
984 	worker->flags |= flags;
985 }
986 
987 /**
988  * worker_clr_flags - clear worker flags and adjust nr_running accordingly
989  * @worker: self
990  * @flags: flags to clear
991  *
992  * Clear @flags in @worker->flags and adjust nr_running accordingly.
993  */
994 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
995 {
996 	struct worker_pool *pool = worker->pool;
997 	unsigned int oflags = worker->flags;
998 
999 	lockdep_assert_held(&pool->lock);
1000 
1001 	worker->flags &= ~flags;
1002 
1003 	/*
1004 	 * If transitioning out of NOT_RUNNING, increment nr_running.  Note
1005 	 * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
1006 	 * of multiple flags, not a single flag.
1007 	 */
1008 	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1009 		if (!(worker->flags & WORKER_NOT_RUNNING))
1010 			pool->nr_running++;
1011 }
1012 
1013 /* Return the first idle worker.  Called with pool->lock held. */
1014 static struct worker *first_idle_worker(struct worker_pool *pool)
1015 {
1016 	if (unlikely(list_empty(&pool->idle_list)))
1017 		return NULL;
1018 
1019 	return list_first_entry(&pool->idle_list, struct worker, entry);
1020 }
1021 
1022 /**
1023  * worker_enter_idle - enter idle state
1024  * @worker: worker which is entering idle state
1025  *
1026  * @worker is entering idle state.  Update stats and idle timer if
1027  * necessary.
1028  *
1029  * LOCKING:
1030  * raw_spin_lock_irq(pool->lock).
1031  */
1032 static void worker_enter_idle(struct worker *worker)
1033 {
1034 	struct worker_pool *pool = worker->pool;
1035 
1036 	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1037 	    WARN_ON_ONCE(!list_empty(&worker->entry) &&
1038 			 (worker->hentry.next || worker->hentry.pprev)))
1039 		return;
1040 
1041 	/* can't use worker_set_flags(), also called from create_worker() */
1042 	worker->flags |= WORKER_IDLE;
1043 	pool->nr_idle++;
1044 	worker->last_active = jiffies;
1045 
1046 	/* idle_list is LIFO */
1047 	list_add(&worker->entry, &pool->idle_list);
1048 
1049 	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1050 		mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1051 
1052 	/* Sanity check nr_running. */
1053 	WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1054 }
1055 
1056 /**
1057  * worker_leave_idle - leave idle state
1058  * @worker: worker which is leaving idle state
1059  *
1060  * @worker is leaving idle state.  Update stats.
1061  *
1062  * LOCKING:
1063  * raw_spin_lock_irq(pool->lock).
1064  */
1065 static void worker_leave_idle(struct worker *worker)
1066 {
1067 	struct worker_pool *pool = worker->pool;
1068 
1069 	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1070 		return;
1071 	worker_clr_flags(worker, WORKER_IDLE);
1072 	pool->nr_idle--;
1073 	list_del_init(&worker->entry);
1074 }
1075 
1076 /**
1077  * find_worker_executing_work - find worker which is executing a work
1078  * @pool: pool of interest
1079  * @work: work to find worker for
1080  *
1081  * Find a worker which is executing @work on @pool by searching
1082  * @pool->busy_hash which is keyed by the address of @work.  For a worker
1083  * to match, its current execution should match the address of @work and
1084  * its work function.  This is to avoid unwanted dependency between
1085  * unrelated work executions through a work item being recycled while still
1086  * being executed.
1087  *
1088  * This is a bit tricky.  A work item may be freed once its execution
1089  * starts and nothing prevents the freed area from being recycled for
1090  * another work item.  If the same work item address ends up being reused
1091  * before the original execution finishes, workqueue will identify the
1092  * recycled work item as currently executing and make it wait until the
1093  * current execution finishes, introducing an unwanted dependency.
1094  *
1095  * This function checks the work item address and work function to avoid
1096  * false positives.  Note that this isn't complete as one may construct a
1097  * work function which can introduce dependency onto itself through a
1098  * recycled work item.  Well, if somebody wants to shoot oneself in the
1099  * foot that badly, there's only so much we can do, and if such deadlock
1100  * actually occurs, it should be easy to locate the culprit work function.
1101  *
1102  * CONTEXT:
1103  * raw_spin_lock_irq(pool->lock).
1104  *
1105  * Return:
1106  * Pointer to worker which is executing @work if found, %NULL
1107  * otherwise.
1108  */
1109 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1110 						 struct work_struct *work)
1111 {
1112 	struct worker *worker;
1113 
1114 	hash_for_each_possible(pool->busy_hash, worker, hentry,
1115 			       (unsigned long)work)
1116 		if (worker->current_work == work &&
1117 		    worker->current_func == work->func)
1118 			return worker;
1119 
1120 	return NULL;
1121 }
1122 
1123 /**
1124  * move_linked_works - move linked works to a list
1125  * @work: start of series of works to be scheduled
1126  * @head: target list to append @work to
1127  * @nextp: out parameter for nested worklist walking
1128  *
1129  * Schedule linked works starting from @work to @head. Work series to be
1130  * scheduled starts at @work and includes any consecutive work with
1131  * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1132  * @nextp.
1133  *
1134  * CONTEXT:
1135  * raw_spin_lock_irq(pool->lock).
1136  */
1137 static void move_linked_works(struct work_struct *work, struct list_head *head,
1138 			      struct work_struct **nextp)
1139 {
1140 	struct work_struct *n;
1141 
1142 	/*
1143 	 * Linked worklist will always end before the end of the list,
1144 	 * use NULL for list head.
1145 	 */
1146 	list_for_each_entry_safe_from(work, n, NULL, entry) {
1147 		list_move_tail(&work->entry, head);
1148 		if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1149 			break;
1150 	}
1151 
1152 	/*
1153 	 * If we're already inside safe list traversal and have moved
1154 	 * multiple works to the scheduled queue, the next position
1155 	 * needs to be updated.
1156 	 */
1157 	if (nextp)
1158 		*nextp = n;
1159 }
1160 
1161 /**
1162  * assign_work - assign a work item and its linked work items to a worker
1163  * @work: work to assign
1164  * @worker: worker to assign to
1165  * @nextp: out parameter for nested worklist walking
1166  *
1167  * Assign @work and its linked work items to @worker. If @work is already being
1168  * executed by another worker in the same pool, it'll be punted there.
1169  *
1170  * If @nextp is not NULL, it's updated to point to the next work of the last
1171  * scheduled work. This allows assign_work() to be nested inside
1172  * list_for_each_entry_safe().
1173  *
1174  * Returns %true if @work was successfully assigned to @worker. %false if @work
1175  * was punted to another worker already executing it.
1176  */
1177 static bool assign_work(struct work_struct *work, struct worker *worker,
1178 			struct work_struct **nextp)
1179 {
1180 	struct worker_pool *pool = worker->pool;
1181 	struct worker *collision;
1182 
1183 	lockdep_assert_held(&pool->lock);
1184 
1185 	/*
1186 	 * A single work shouldn't be executed concurrently by multiple workers.
1187 	 * __queue_work() ensures that @work doesn't jump to a different pool
1188 	 * while still running in the previous pool. Here, we should ensure that
1189 	 * @work is not executed concurrently by multiple workers from the same
1190 	 * pool. Check whether anyone is already processing the work. If so,
1191 	 * defer the work to the currently executing one.
1192 	 */
1193 	collision = find_worker_executing_work(pool, work);
1194 	if (unlikely(collision)) {
1195 		move_linked_works(work, &collision->scheduled, nextp);
1196 		return false;
1197 	}
1198 
1199 	move_linked_works(work, &worker->scheduled, nextp);
1200 	return true;
1201 }
1202 
1203 static struct irq_work *bh_pool_irq_work(struct worker_pool *pool)
1204 {
1205 	int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0;
1206 
1207 	return &per_cpu(bh_pool_irq_works, pool->cpu)[high];
1208 }
1209 
1210 static void kick_bh_pool(struct worker_pool *pool)
1211 {
1212 #ifdef CONFIG_SMP
1213 	/* see drain_dead_softirq_workfn() for BH_DRAINING */
1214 	if (unlikely(pool->cpu != smp_processor_id() &&
1215 		     !(pool->flags & POOL_BH_DRAINING))) {
1216 		irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu);
1217 		return;
1218 	}
1219 #endif
1220 	if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
1221 		raise_softirq_irqoff(HI_SOFTIRQ);
1222 	else
1223 		raise_softirq_irqoff(TASKLET_SOFTIRQ);
1224 }
1225 
1226 /**
1227  * kick_pool - wake up an idle worker if necessary
1228  * @pool: pool to kick
1229  *
1230  * @pool may have pending work items. Wake up worker if necessary. Returns
1231  * whether a worker was woken up.
1232  */
1233 static bool kick_pool(struct worker_pool *pool)
1234 {
1235 	struct worker *worker = first_idle_worker(pool);
1236 	struct task_struct *p;
1237 
1238 	lockdep_assert_held(&pool->lock);
1239 
1240 	if (!need_more_worker(pool) || !worker)
1241 		return false;
1242 
1243 	if (pool->flags & POOL_BH) {
1244 		kick_bh_pool(pool);
1245 		return true;
1246 	}
1247 
1248 	p = worker->task;
1249 
1250 #ifdef CONFIG_SMP
1251 	/*
1252 	 * Idle @worker is about to execute @work and waking up provides an
1253 	 * opportunity to migrate @worker at a lower cost by setting the task's
1254 	 * wake_cpu field. Let's see if we want to move @worker to improve
1255 	 * execution locality.
1256 	 *
1257 	 * We're waking the worker that went idle the latest and there's some
1258 	 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1259 	 * so, setting the wake_cpu won't do anything. As this is a best-effort
1260 	 * optimization and the race window is narrow, let's leave as-is for
1261 	 * now. If this becomes pronounced, we can skip over workers which are
1262 	 * still on cpu when picking an idle worker.
1263 	 *
1264 	 * If @pool has non-strict affinity, @worker might have ended up outside
1265 	 * its affinity scope. Repatriate.
1266 	 */
1267 	if (!pool->attrs->affn_strict &&
1268 	    !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1269 		struct work_struct *work = list_first_entry(&pool->worklist,
1270 						struct work_struct, entry);
1271 		int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask,
1272 							  cpu_online_mask);
1273 		if (wake_cpu < nr_cpu_ids) {
1274 			p->wake_cpu = wake_cpu;
1275 			get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1276 		}
1277 	}
1278 #endif
1279 	wake_up_process(p);
1280 	return true;
1281 }
1282 
1283 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1284 
1285 /*
1286  * Concurrency-managed per-cpu work items that hog CPU for longer than
1287  * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1288  * which prevents them from stalling other concurrency-managed work items. If a
1289  * work function keeps triggering this mechanism, it's likely that the work item
1290  * should be using an unbound workqueue instead.
1291  *
1292  * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1293  * and report them so that they can be examined and converted to use unbound
1294  * workqueues as appropriate. To avoid flooding the console, each violating work
1295  * function is tracked and reported with exponential backoff.
1296  */
1297 #define WCI_MAX_ENTS 128
1298 
1299 struct wci_ent {
1300 	work_func_t		func;
1301 	atomic64_t		cnt;
1302 	struct hlist_node	hash_node;
1303 };
1304 
1305 static struct wci_ent wci_ents[WCI_MAX_ENTS];
1306 static int wci_nr_ents;
1307 static DEFINE_RAW_SPINLOCK(wci_lock);
1308 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1309 
1310 static struct wci_ent *wci_find_ent(work_func_t func)
1311 {
1312 	struct wci_ent *ent;
1313 
1314 	hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1315 				   (unsigned long)func) {
1316 		if (ent->func == func)
1317 			return ent;
1318 	}
1319 	return NULL;
1320 }
1321 
1322 static void wq_cpu_intensive_report(work_func_t func)
1323 {
1324 	struct wci_ent *ent;
1325 
1326 restart:
1327 	ent = wci_find_ent(func);
1328 	if (ent) {
1329 		u64 cnt;
1330 
1331 		/*
1332 		 * Start reporting from the warning_thresh and back off
1333 		 * exponentially.
1334 		 */
1335 		cnt = atomic64_inc_return_relaxed(&ent->cnt);
1336 		if (wq_cpu_intensive_warning_thresh &&
1337 		    cnt >= wq_cpu_intensive_warning_thresh &&
1338 		    is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh))
1339 			printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1340 					ent->func, wq_cpu_intensive_thresh_us,
1341 					atomic64_read(&ent->cnt));
1342 		return;
1343 	}
1344 
1345 	/*
1346 	 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1347 	 * is exhausted, something went really wrong and we probably made enough
1348 	 * noise already.
1349 	 */
1350 	if (wci_nr_ents >= WCI_MAX_ENTS)
1351 		return;
1352 
1353 	raw_spin_lock(&wci_lock);
1354 
1355 	if (wci_nr_ents >= WCI_MAX_ENTS) {
1356 		raw_spin_unlock(&wci_lock);
1357 		return;
1358 	}
1359 
1360 	if (wci_find_ent(func)) {
1361 		raw_spin_unlock(&wci_lock);
1362 		goto restart;
1363 	}
1364 
1365 	ent = &wci_ents[wci_nr_ents++];
1366 	ent->func = func;
1367 	atomic64_set(&ent->cnt, 0);
1368 	hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1369 
1370 	raw_spin_unlock(&wci_lock);
1371 
1372 	goto restart;
1373 }
1374 
1375 #else	/* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1376 static void wq_cpu_intensive_report(work_func_t func) {}
1377 #endif	/* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1378 
1379 /**
1380  * wq_worker_running - a worker is running again
1381  * @task: task waking up
1382  *
1383  * This function is called when a worker returns from schedule()
1384  */
1385 void wq_worker_running(struct task_struct *task)
1386 {
1387 	struct worker *worker = kthread_data(task);
1388 
1389 	if (!READ_ONCE(worker->sleeping))
1390 		return;
1391 
1392 	/*
1393 	 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1394 	 * and the nr_running increment below, we may ruin the nr_running reset
1395 	 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1396 	 * pool. Protect against such race.
1397 	 */
1398 	preempt_disable();
1399 	if (!(worker->flags & WORKER_NOT_RUNNING))
1400 		worker->pool->nr_running++;
1401 	preempt_enable();
1402 
1403 	/*
1404 	 * CPU intensive auto-detection cares about how long a work item hogged
1405 	 * CPU without sleeping. Reset the starting timestamp on wakeup.
1406 	 */
1407 	worker->current_at = worker->task->se.sum_exec_runtime;
1408 
1409 	WRITE_ONCE(worker->sleeping, 0);
1410 }
1411 
1412 /**
1413  * wq_worker_sleeping - a worker is going to sleep
1414  * @task: task going to sleep
1415  *
1416  * This function is called from schedule() when a busy worker is
1417  * going to sleep.
1418  */
1419 void wq_worker_sleeping(struct task_struct *task)
1420 {
1421 	struct worker *worker = kthread_data(task);
1422 	struct worker_pool *pool;
1423 
1424 	/*
1425 	 * Rescuers, which may not have all the fields set up like normal
1426 	 * workers, also reach here, let's not access anything before
1427 	 * checking NOT_RUNNING.
1428 	 */
1429 	if (worker->flags & WORKER_NOT_RUNNING)
1430 		return;
1431 
1432 	pool = worker->pool;
1433 
1434 	/* Return if preempted before wq_worker_running() was reached */
1435 	if (READ_ONCE(worker->sleeping))
1436 		return;
1437 
1438 	WRITE_ONCE(worker->sleeping, 1);
1439 	raw_spin_lock_irq(&pool->lock);
1440 
1441 	/*
1442 	 * Recheck in case unbind_workers() preempted us. We don't
1443 	 * want to decrement nr_running after the worker is unbound
1444 	 * and nr_running has been reset.
1445 	 */
1446 	if (worker->flags & WORKER_NOT_RUNNING) {
1447 		raw_spin_unlock_irq(&pool->lock);
1448 		return;
1449 	}
1450 
1451 	pool->nr_running--;
1452 	if (kick_pool(pool))
1453 		worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1454 
1455 	raw_spin_unlock_irq(&pool->lock);
1456 }
1457 
1458 /**
1459  * wq_worker_tick - a scheduler tick occurred while a kworker is running
1460  * @task: task currently running
1461  *
1462  * Called from sched_tick(). We're in the IRQ context and the current
1463  * worker's fields which follow the 'K' locking rule can be accessed safely.
1464  */
1465 void wq_worker_tick(struct task_struct *task)
1466 {
1467 	struct worker *worker = kthread_data(task);
1468 	struct pool_workqueue *pwq = worker->current_pwq;
1469 	struct worker_pool *pool = worker->pool;
1470 
1471 	if (!pwq)
1472 		return;
1473 
1474 	pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1475 
1476 	if (!wq_cpu_intensive_thresh_us)
1477 		return;
1478 
1479 	/*
1480 	 * If the current worker is concurrency managed and hogged the CPU for
1481 	 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1482 	 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1483 	 *
1484 	 * Set @worker->sleeping means that @worker is in the process of
1485 	 * switching out voluntarily and won't be contributing to
1486 	 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1487 	 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1488 	 * double decrements. The task is releasing the CPU anyway. Let's skip.
1489 	 * We probably want to make this prettier in the future.
1490 	 */
1491 	if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1492 	    worker->task->se.sum_exec_runtime - worker->current_at <
1493 	    wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1494 		return;
1495 
1496 	raw_spin_lock(&pool->lock);
1497 
1498 	worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1499 	wq_cpu_intensive_report(worker->current_func);
1500 	pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1501 
1502 	if (kick_pool(pool))
1503 		pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1504 
1505 	raw_spin_unlock(&pool->lock);
1506 }
1507 
1508 /**
1509  * wq_worker_last_func - retrieve worker's last work function
1510  * @task: Task to retrieve last work function of.
1511  *
1512  * Determine the last function a worker executed. This is called from
1513  * the scheduler to get a worker's last known identity.
1514  *
1515  * CONTEXT:
1516  * raw_spin_lock_irq(rq->lock)
1517  *
1518  * This function is called during schedule() when a kworker is going
1519  * to sleep. It's used by psi to identify aggregation workers during
1520  * dequeuing, to allow periodic aggregation to shut-off when that
1521  * worker is the last task in the system or cgroup to go to sleep.
1522  *
1523  * As this function doesn't involve any workqueue-related locking, it
1524  * only returns stable values when called from inside the scheduler's
1525  * queuing and dequeuing paths, when @task, which must be a kworker,
1526  * is guaranteed to not be processing any works.
1527  *
1528  * Return:
1529  * The last work function %current executed as a worker, NULL if it
1530  * hasn't executed any work yet.
1531  */
1532 work_func_t wq_worker_last_func(struct task_struct *task)
1533 {
1534 	struct worker *worker = kthread_data(task);
1535 
1536 	return worker->last_func;
1537 }
1538 
1539 /**
1540  * wq_node_nr_active - Determine wq_node_nr_active to use
1541  * @wq: workqueue of interest
1542  * @node: NUMA node, can be %NUMA_NO_NODE
1543  *
1544  * Determine wq_node_nr_active to use for @wq on @node. Returns:
1545  *
1546  * - %NULL for per-cpu workqueues as they don't need to use shared nr_active.
1547  *
1548  * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1549  *
1550  * - Otherwise, node_nr_active[@node].
1551  */
1552 static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq,
1553 						   int node)
1554 {
1555 	if (!(wq->flags & WQ_UNBOUND))
1556 		return NULL;
1557 
1558 	if (node == NUMA_NO_NODE)
1559 		node = nr_node_ids;
1560 
1561 	return wq->node_nr_active[node];
1562 }
1563 
1564 /**
1565  * wq_update_node_max_active - Update per-node max_actives to use
1566  * @wq: workqueue to update
1567  * @off_cpu: CPU that's going down, -1 if a CPU is not going down
1568  *
1569  * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1570  * distributed among nodes according to the proportions of numbers of online
1571  * cpus. The result is always between @wq->min_active and max_active.
1572  */
1573 static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1574 {
1575 	struct cpumask *effective = unbound_effective_cpumask(wq);
1576 	int min_active = READ_ONCE(wq->min_active);
1577 	int max_active = READ_ONCE(wq->max_active);
1578 	int total_cpus, node;
1579 
1580 	lockdep_assert_held(&wq->mutex);
1581 
1582 	if (!wq_topo_initialized)
1583 		return;
1584 
1585 	if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective))
1586 		off_cpu = -1;
1587 
1588 	total_cpus = cpumask_weight_and(effective, cpu_online_mask);
1589 	if (off_cpu >= 0)
1590 		total_cpus--;
1591 
1592 	/* If all CPUs of the wq get offline, use the default values */
1593 	if (unlikely(!total_cpus)) {
1594 		for_each_node(node)
1595 			wq_node_nr_active(wq, node)->max = min_active;
1596 
1597 		wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1598 		return;
1599 	}
1600 
1601 	for_each_node(node) {
1602 		int node_cpus;
1603 
1604 		node_cpus = cpumask_weight_and(effective, cpumask_of_node(node));
1605 		if (off_cpu >= 0 && cpu_to_node(off_cpu) == node)
1606 			node_cpus--;
1607 
1608 		wq_node_nr_active(wq, node)->max =
1609 			clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1610 			      min_active, max_active);
1611 	}
1612 
1613 	wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1614 }
1615 
1616 /**
1617  * get_pwq - get an extra reference on the specified pool_workqueue
1618  * @pwq: pool_workqueue to get
1619  *
1620  * Obtain an extra reference on @pwq.  The caller should guarantee that
1621  * @pwq has positive refcnt and be holding the matching pool->lock.
1622  */
1623 static void get_pwq(struct pool_workqueue *pwq)
1624 {
1625 	lockdep_assert_held(&pwq->pool->lock);
1626 	WARN_ON_ONCE(pwq->refcnt <= 0);
1627 	pwq->refcnt++;
1628 }
1629 
1630 /**
1631  * put_pwq - put a pool_workqueue reference
1632  * @pwq: pool_workqueue to put
1633  *
1634  * Drop a reference of @pwq.  If its refcnt reaches zero, schedule its
1635  * destruction.  The caller should be holding the matching pool->lock.
1636  */
1637 static void put_pwq(struct pool_workqueue *pwq)
1638 {
1639 	lockdep_assert_held(&pwq->pool->lock);
1640 	if (likely(--pwq->refcnt))
1641 		return;
1642 	/*
1643 	 * @pwq can't be released under pool->lock, bounce to a dedicated
1644 	 * kthread_worker to avoid A-A deadlocks.
1645 	 */
1646 	kthread_queue_work(pwq_release_worker, &pwq->release_work);
1647 }
1648 
1649 /**
1650  * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1651  * @pwq: pool_workqueue to put (can be %NULL)
1652  *
1653  * put_pwq() with locking.  This function also allows %NULL @pwq.
1654  */
1655 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1656 {
1657 	if (pwq) {
1658 		/*
1659 		 * As both pwqs and pools are RCU protected, the
1660 		 * following lock operations are safe.
1661 		 */
1662 		raw_spin_lock_irq(&pwq->pool->lock);
1663 		put_pwq(pwq);
1664 		raw_spin_unlock_irq(&pwq->pool->lock);
1665 	}
1666 }
1667 
1668 static bool pwq_is_empty(struct pool_workqueue *pwq)
1669 {
1670 	return !pwq->nr_active && list_empty(&pwq->inactive_works);
1671 }
1672 
1673 static void __pwq_activate_work(struct pool_workqueue *pwq,
1674 				struct work_struct *work)
1675 {
1676 	unsigned long *wdb = work_data_bits(work);
1677 
1678 	WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1679 	trace_workqueue_activate_work(work);
1680 	if (list_empty(&pwq->pool->worklist))
1681 		pwq->pool->watchdog_ts = jiffies;
1682 	move_linked_works(work, &pwq->pool->worklist, NULL);
1683 	__clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1684 }
1685 
1686 static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1687 {
1688 	int max = READ_ONCE(nna->max);
1689 
1690 	while (true) {
1691 		int old, tmp;
1692 
1693 		old = atomic_read(&nna->nr);
1694 		if (old >= max)
1695 			return false;
1696 		tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1);
1697 		if (tmp == old)
1698 			return true;
1699 	}
1700 }
1701 
1702 /**
1703  * pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1704  * @pwq: pool_workqueue of interest
1705  * @fill: max_active may have increased, try to increase concurrency level
1706  *
1707  * Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1708  * successfully obtained. %false otherwise.
1709  */
1710 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1711 {
1712 	struct workqueue_struct *wq = pwq->wq;
1713 	struct worker_pool *pool = pwq->pool;
1714 	struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node);
1715 	bool obtained = false;
1716 
1717 	lockdep_assert_held(&pool->lock);
1718 
1719 	if (!nna) {
1720 		/* BH or per-cpu workqueue, pwq->nr_active is sufficient */
1721 		obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1722 		goto out;
1723 	}
1724 
1725 	if (unlikely(pwq->plugged))
1726 		return false;
1727 
1728 	/*
1729 	 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1730 	 * already waiting on $nna, pwq_dec_nr_active() will maintain the
1731 	 * concurrency level. Don't jump the line.
1732 	 *
1733 	 * We need to ignore the pending test after max_active has increased as
1734 	 * pwq_dec_nr_active() can only maintain the concurrency level but not
1735 	 * increase it. This is indicated by @fill.
1736 	 */
1737 	if (!list_empty(&pwq->pending_node) && likely(!fill))
1738 		goto out;
1739 
1740 	obtained = tryinc_node_nr_active(nna);
1741 	if (obtained)
1742 		goto out;
1743 
1744 	/*
1745 	 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1746 	 * and try again. The smp_mb() is paired with the implied memory barrier
1747 	 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1748 	 * we see the decremented $nna->nr or they see non-empty
1749 	 * $nna->pending_pwqs.
1750 	 */
1751 	raw_spin_lock(&nna->lock);
1752 
1753 	if (list_empty(&pwq->pending_node))
1754 		list_add_tail(&pwq->pending_node, &nna->pending_pwqs);
1755 	else if (likely(!fill))
1756 		goto out_unlock;
1757 
1758 	smp_mb();
1759 
1760 	obtained = tryinc_node_nr_active(nna);
1761 
1762 	/*
1763 	 * If @fill, @pwq might have already been pending. Being spuriously
1764 	 * pending in cold paths doesn't affect anything. Let's leave it be.
1765 	 */
1766 	if (obtained && likely(!fill))
1767 		list_del_init(&pwq->pending_node);
1768 
1769 out_unlock:
1770 	raw_spin_unlock(&nna->lock);
1771 out:
1772 	if (obtained)
1773 		pwq->nr_active++;
1774 	return obtained;
1775 }
1776 
1777 /**
1778  * pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1779  * @pwq: pool_workqueue of interest
1780  * @fill: max_active may have increased, try to increase concurrency level
1781  *
1782  * Activate the first inactive work item of @pwq if available and allowed by
1783  * max_active limit.
1784  *
1785  * Returns %true if an inactive work item has been activated. %false if no
1786  * inactive work item is found or max_active limit is reached.
1787  */
1788 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1789 {
1790 	struct work_struct *work =
1791 		list_first_entry_or_null(&pwq->inactive_works,
1792 					 struct work_struct, entry);
1793 
1794 	if (work && pwq_tryinc_nr_active(pwq, fill)) {
1795 		__pwq_activate_work(pwq, work);
1796 		return true;
1797 	} else {
1798 		return false;
1799 	}
1800 }
1801 
1802 /**
1803  * unplug_oldest_pwq - unplug the oldest pool_workqueue
1804  * @wq: workqueue_struct where its oldest pwq is to be unplugged
1805  *
1806  * This function should only be called for ordered workqueues where only the
1807  * oldest pwq is unplugged, the others are plugged to suspend execution to
1808  * ensure proper work item ordering::
1809  *
1810  *    dfl_pwq --------------+     [P] - plugged
1811  *                          |
1812  *                          v
1813  *    pwqs -> A -> B [P] -> C [P] (newest)
1814  *            |    |        |
1815  *            1    3        5
1816  *            |    |        |
1817  *            2    4        6
1818  *
1819  * When the oldest pwq is drained and removed, this function should be called
1820  * to unplug the next oldest one to start its work item execution. Note that
1821  * pwq's are linked into wq->pwqs with the oldest first, so the first one in
1822  * the list is the oldest.
1823  */
1824 static void unplug_oldest_pwq(struct workqueue_struct *wq)
1825 {
1826 	struct pool_workqueue *pwq;
1827 
1828 	lockdep_assert_held(&wq->mutex);
1829 
1830 	/* Caller should make sure that pwqs isn't empty before calling */
1831 	pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue,
1832 				       pwqs_node);
1833 	raw_spin_lock_irq(&pwq->pool->lock);
1834 	if (pwq->plugged) {
1835 		pwq->plugged = false;
1836 		if (pwq_activate_first_inactive(pwq, true))
1837 			kick_pool(pwq->pool);
1838 	}
1839 	raw_spin_unlock_irq(&pwq->pool->lock);
1840 }
1841 
1842 /**
1843  * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1844  * @nna: wq_node_nr_active to activate a pending pwq for
1845  * @caller_pool: worker_pool the caller is locking
1846  *
1847  * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1848  * @caller_pool may be unlocked and relocked to lock other worker_pools.
1849  */
1850 static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1851 				      struct worker_pool *caller_pool)
1852 {
1853 	struct worker_pool *locked_pool = caller_pool;
1854 	struct pool_workqueue *pwq;
1855 	struct work_struct *work;
1856 
1857 	lockdep_assert_held(&caller_pool->lock);
1858 
1859 	raw_spin_lock(&nna->lock);
1860 retry:
1861 	pwq = list_first_entry_or_null(&nna->pending_pwqs,
1862 				       struct pool_workqueue, pending_node);
1863 	if (!pwq)
1864 		goto out_unlock;
1865 
1866 	/*
1867 	 * If @pwq is for a different pool than @locked_pool, we need to lock
1868 	 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1869 	 * / lock dance. For that, we also need to release @nna->lock as it's
1870 	 * nested inside pool locks.
1871 	 */
1872 	if (pwq->pool != locked_pool) {
1873 		raw_spin_unlock(&locked_pool->lock);
1874 		locked_pool = pwq->pool;
1875 		if (!raw_spin_trylock(&locked_pool->lock)) {
1876 			raw_spin_unlock(&nna->lock);
1877 			raw_spin_lock(&locked_pool->lock);
1878 			raw_spin_lock(&nna->lock);
1879 			goto retry;
1880 		}
1881 	}
1882 
1883 	/*
1884 	 * $pwq may not have any inactive work items due to e.g. cancellations.
1885 	 * Drop it from pending_pwqs and see if there's another one.
1886 	 */
1887 	work = list_first_entry_or_null(&pwq->inactive_works,
1888 					struct work_struct, entry);
1889 	if (!work) {
1890 		list_del_init(&pwq->pending_node);
1891 		goto retry;
1892 	}
1893 
1894 	/*
1895 	 * Acquire an nr_active count and activate the inactive work item. If
1896 	 * $pwq still has inactive work items, rotate it to the end of the
1897 	 * pending_pwqs so that we round-robin through them. This means that
1898 	 * inactive work items are not activated in queueing order which is fine
1899 	 * given that there has never been any ordering across different pwqs.
1900 	 */
1901 	if (likely(tryinc_node_nr_active(nna))) {
1902 		pwq->nr_active++;
1903 		__pwq_activate_work(pwq, work);
1904 
1905 		if (list_empty(&pwq->inactive_works))
1906 			list_del_init(&pwq->pending_node);
1907 		else
1908 			list_move_tail(&pwq->pending_node, &nna->pending_pwqs);
1909 
1910 		/* if activating a foreign pool, make sure it's running */
1911 		if (pwq->pool != caller_pool)
1912 			kick_pool(pwq->pool);
1913 	}
1914 
1915 out_unlock:
1916 	raw_spin_unlock(&nna->lock);
1917 	if (locked_pool != caller_pool) {
1918 		raw_spin_unlock(&locked_pool->lock);
1919 		raw_spin_lock(&caller_pool->lock);
1920 	}
1921 }
1922 
1923 /**
1924  * pwq_dec_nr_active - Retire an active count
1925  * @pwq: pool_workqueue of interest
1926  *
1927  * Decrement @pwq's nr_active and try to activate the first inactive work item.
1928  * For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
1929  */
1930 static void pwq_dec_nr_active(struct pool_workqueue *pwq)
1931 {
1932 	struct worker_pool *pool = pwq->pool;
1933 	struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node);
1934 
1935 	lockdep_assert_held(&pool->lock);
1936 
1937 	/*
1938 	 * @pwq->nr_active should be decremented for both percpu and unbound
1939 	 * workqueues.
1940 	 */
1941 	pwq->nr_active--;
1942 
1943 	/*
1944 	 * For a percpu workqueue, it's simple. Just need to kick the first
1945 	 * inactive work item on @pwq itself.
1946 	 */
1947 	if (!nna) {
1948 		pwq_activate_first_inactive(pwq, false);
1949 		return;
1950 	}
1951 
1952 	/*
1953 	 * If @pwq is for an unbound workqueue, it's more complicated because
1954 	 * multiple pwqs and pools may be sharing the nr_active count. When a
1955 	 * pwq needs to wait for an nr_active count, it puts itself on
1956 	 * $nna->pending_pwqs. The following atomic_dec_return()'s implied
1957 	 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
1958 	 * guarantee that either we see non-empty pending_pwqs or they see
1959 	 * decremented $nna->nr.
1960 	 *
1961 	 * $nna->max may change as CPUs come online/offline and @pwq->wq's
1962 	 * max_active gets updated. However, it is guaranteed to be equal to or
1963 	 * larger than @pwq->wq->min_active which is above zero unless freezing.
1964 	 * This maintains the forward progress guarantee.
1965 	 */
1966 	if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max))
1967 		return;
1968 
1969 	if (!list_empty(&nna->pending_pwqs))
1970 		node_activate_pending_pwq(nna, pool);
1971 }
1972 
1973 /**
1974  * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1975  * @pwq: pwq of interest
1976  * @work_data: work_data of work which left the queue
1977  *
1978  * A work either has completed or is removed from pending queue,
1979  * decrement nr_in_flight of its pwq and handle workqueue flushing.
1980  *
1981  * NOTE:
1982  * For unbound workqueues, this function may temporarily drop @pwq->pool->lock
1983  * and thus should be called after all other state updates for the in-flight
1984  * work item is complete.
1985  *
1986  * CONTEXT:
1987  * raw_spin_lock_irq(pool->lock).
1988  */
1989 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1990 {
1991 	int color = get_work_color(work_data);
1992 
1993 	if (!(work_data & WORK_STRUCT_INACTIVE))
1994 		pwq_dec_nr_active(pwq);
1995 
1996 	pwq->nr_in_flight[color]--;
1997 
1998 	/* is flush in progress and are we at the flushing tip? */
1999 	if (likely(pwq->flush_color != color))
2000 		goto out_put;
2001 
2002 	/* are there still in-flight works? */
2003 	if (pwq->nr_in_flight[color])
2004 		goto out_put;
2005 
2006 	/* this pwq is done, clear flush_color */
2007 	pwq->flush_color = -1;
2008 
2009 	/*
2010 	 * If this was the last pwq, wake up the first flusher.  It
2011 	 * will handle the rest.
2012 	 */
2013 	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2014 		complete(&pwq->wq->first_flusher->done);
2015 out_put:
2016 	put_pwq(pwq);
2017 }
2018 
2019 /**
2020  * try_to_grab_pending - steal work item from worklist and disable irq
2021  * @work: work item to steal
2022  * @cflags: %WORK_CANCEL_ flags
2023  * @irq_flags: place to store irq state
2024  *
2025  * Try to grab PENDING bit of @work.  This function can handle @work in any
2026  * stable state - idle, on timer or on worklist.
2027  *
2028  * Return:
2029  *
2030  *  ========	================================================================
2031  *  1		if @work was pending and we successfully stole PENDING
2032  *  0		if @work was idle and we claimed PENDING
2033  *  -EAGAIN	if PENDING couldn't be grabbed at the moment, safe to busy-retry
2034  *  ========	================================================================
2035  *
2036  * Note:
2037  * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
2038  * interrupted while holding PENDING and @work off queue, irq must be
2039  * disabled on entry.  This, combined with delayed_work->timer being
2040  * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2041  *
2042  * On successful return, >= 0, irq is disabled and the caller is
2043  * responsible for releasing it using local_irq_restore(*@irq_flags).
2044  *
2045  * This function is safe to call from any context including IRQ handler.
2046  */
2047 static int try_to_grab_pending(struct work_struct *work, u32 cflags,
2048 			       unsigned long *irq_flags)
2049 {
2050 	struct worker_pool *pool;
2051 	struct pool_workqueue *pwq;
2052 
2053 	local_irq_save(*irq_flags);
2054 
2055 	/* try to steal the timer if it exists */
2056 	if (cflags & WORK_CANCEL_DELAYED) {
2057 		struct delayed_work *dwork = to_delayed_work(work);
2058 
2059 		/*
2060 		 * dwork->timer is irqsafe.  If del_timer() fails, it's
2061 		 * guaranteed that the timer is not queued anywhere and not
2062 		 * running on the local CPU.
2063 		 */
2064 		if (likely(del_timer(&dwork->timer)))
2065 			return 1;
2066 	}
2067 
2068 	/* try to claim PENDING the normal way */
2069 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2070 		return 0;
2071 
2072 	rcu_read_lock();
2073 	/*
2074 	 * The queueing is in progress, or it is already queued. Try to
2075 	 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2076 	 */
2077 	pool = get_work_pool(work);
2078 	if (!pool)
2079 		goto fail;
2080 
2081 	raw_spin_lock(&pool->lock);
2082 	/*
2083 	 * work->data is guaranteed to point to pwq only while the work
2084 	 * item is queued on pwq->wq, and both updating work->data to point
2085 	 * to pwq on queueing and to pool on dequeueing are done under
2086 	 * pwq->pool->lock.  This in turn guarantees that, if work->data
2087 	 * points to pwq which is associated with a locked pool, the work
2088 	 * item is currently queued on that pool.
2089 	 */
2090 	pwq = get_work_pwq(work);
2091 	if (pwq && pwq->pool == pool) {
2092 		unsigned long work_data = *work_data_bits(work);
2093 
2094 		debug_work_deactivate(work);
2095 
2096 		/*
2097 		 * A cancelable inactive work item must be in the
2098 		 * pwq->inactive_works since a queued barrier can't be
2099 		 * canceled (see the comments in insert_wq_barrier()).
2100 		 *
2101 		 * An inactive work item cannot be deleted directly because
2102 		 * it might have linked barrier work items which, if left
2103 		 * on the inactive_works list, will confuse pwq->nr_active
2104 		 * management later on and cause stall.  Move the linked
2105 		 * barrier work items to the worklist when deleting the grabbed
2106 		 * item. Also keep WORK_STRUCT_INACTIVE in work_data, so that
2107 		 * it doesn't participate in nr_active management in later
2108 		 * pwq_dec_nr_in_flight().
2109 		 */
2110 		if (work_data & WORK_STRUCT_INACTIVE)
2111 			move_linked_works(work, &pwq->pool->worklist, NULL);
2112 
2113 		list_del_init(&work->entry);
2114 
2115 		/*
2116 		 * work->data points to pwq iff queued. Let's point to pool. As
2117 		 * this destroys work->data needed by the next step, stash it.
2118 		 */
2119 		set_work_pool_and_keep_pending(work, pool->id,
2120 					       pool_offq_flags(pool));
2121 
2122 		/* must be the last step, see the function comment */
2123 		pwq_dec_nr_in_flight(pwq, work_data);
2124 
2125 		raw_spin_unlock(&pool->lock);
2126 		rcu_read_unlock();
2127 		return 1;
2128 	}
2129 	raw_spin_unlock(&pool->lock);
2130 fail:
2131 	rcu_read_unlock();
2132 	local_irq_restore(*irq_flags);
2133 	return -EAGAIN;
2134 }
2135 
2136 /**
2137  * work_grab_pending - steal work item from worklist and disable irq
2138  * @work: work item to steal
2139  * @cflags: %WORK_CANCEL_ flags
2140  * @irq_flags: place to store IRQ state
2141  *
2142  * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer
2143  * or on worklist.
2144  *
2145  * Can be called from any context. IRQ is disabled on return with IRQ state
2146  * stored in *@irq_flags. The caller is responsible for re-enabling it using
2147  * local_irq_restore().
2148  *
2149  * Returns %true if @work was pending. %false if idle.
2150  */
2151 static bool work_grab_pending(struct work_struct *work, u32 cflags,
2152 			      unsigned long *irq_flags)
2153 {
2154 	int ret;
2155 
2156 	while (true) {
2157 		ret = try_to_grab_pending(work, cflags, irq_flags);
2158 		if (ret >= 0)
2159 			return ret;
2160 		cpu_relax();
2161 	}
2162 }
2163 
2164 /**
2165  * insert_work - insert a work into a pool
2166  * @pwq: pwq @work belongs to
2167  * @work: work to insert
2168  * @head: insertion point
2169  * @extra_flags: extra WORK_STRUCT_* flags to set
2170  *
2171  * Insert @work which belongs to @pwq after @head.  @extra_flags is or'd to
2172  * work_struct flags.
2173  *
2174  * CONTEXT:
2175  * raw_spin_lock_irq(pool->lock).
2176  */
2177 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2178 			struct list_head *head, unsigned int extra_flags)
2179 {
2180 	debug_work_activate(work);
2181 
2182 	/* record the work call stack in order to print it in KASAN reports */
2183 	kasan_record_aux_stack_noalloc(work);
2184 
2185 	/* we own @work, set data and link */
2186 	set_work_pwq(work, pwq, extra_flags);
2187 	list_add_tail(&work->entry, head);
2188 	get_pwq(pwq);
2189 }
2190 
2191 /*
2192  * Test whether @work is being queued from another work executing on the
2193  * same workqueue.
2194  */
2195 static bool is_chained_work(struct workqueue_struct *wq)
2196 {
2197 	struct worker *worker;
2198 
2199 	worker = current_wq_worker();
2200 	/*
2201 	 * Return %true iff I'm a worker executing a work item on @wq.  If
2202 	 * I'm @worker, it's safe to dereference it without locking.
2203 	 */
2204 	return worker && worker->current_pwq->wq == wq;
2205 }
2206 
2207 /*
2208  * When queueing an unbound work item to a wq, prefer local CPU if allowed
2209  * by wq_unbound_cpumask.  Otherwise, round robin among the allowed ones to
2210  * avoid perturbing sensitive tasks.
2211  */
2212 static int wq_select_unbound_cpu(int cpu)
2213 {
2214 	int new_cpu;
2215 
2216 	if (likely(!wq_debug_force_rr_cpu)) {
2217 		if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2218 			return cpu;
2219 	} else {
2220 		pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2221 	}
2222 
2223 	new_cpu = __this_cpu_read(wq_rr_cpu_last);
2224 	new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2225 	if (unlikely(new_cpu >= nr_cpu_ids)) {
2226 		new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2227 		if (unlikely(new_cpu >= nr_cpu_ids))
2228 			return cpu;
2229 	}
2230 	__this_cpu_write(wq_rr_cpu_last, new_cpu);
2231 
2232 	return new_cpu;
2233 }
2234 
2235 static void __queue_work(int cpu, struct workqueue_struct *wq,
2236 			 struct work_struct *work)
2237 {
2238 	struct pool_workqueue *pwq;
2239 	struct worker_pool *last_pool, *pool;
2240 	unsigned int work_flags;
2241 	unsigned int req_cpu = cpu;
2242 
2243 	/*
2244 	 * While a work item is PENDING && off queue, a task trying to
2245 	 * steal the PENDING will busy-loop waiting for it to either get
2246 	 * queued or lose PENDING.  Grabbing PENDING and queueing should
2247 	 * happen with IRQ disabled.
2248 	 */
2249 	lockdep_assert_irqs_disabled();
2250 
2251 	/*
2252 	 * For a draining wq, only works from the same workqueue are
2253 	 * allowed. The __WQ_DESTROYING helps to spot the issue that
2254 	 * queues a new work item to a wq after destroy_workqueue(wq).
2255 	 */
2256 	if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
2257 		     WARN_ON_ONCE(!is_chained_work(wq))))
2258 		return;
2259 	rcu_read_lock();
2260 retry:
2261 	/* pwq which will be used unless @work is executing elsewhere */
2262 	if (req_cpu == WORK_CPU_UNBOUND) {
2263 		if (wq->flags & WQ_UNBOUND)
2264 			cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2265 		else
2266 			cpu = raw_smp_processor_id();
2267 	}
2268 
2269 	pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2270 	pool = pwq->pool;
2271 
2272 	/*
2273 	 * If @work was previously on a different pool, it might still be
2274 	 * running there, in which case the work needs to be queued on that
2275 	 * pool to guarantee non-reentrancy.
2276 	 *
2277 	 * For ordered workqueue, work items must be queued on the newest pwq
2278 	 * for accurate order management.  Guaranteed order also guarantees
2279 	 * non-reentrancy.  See the comments above unplug_oldest_pwq().
2280 	 */
2281 	last_pool = get_work_pool(work);
2282 	if (last_pool && last_pool != pool && !(wq->flags & __WQ_ORDERED)) {
2283 		struct worker *worker;
2284 
2285 		raw_spin_lock(&last_pool->lock);
2286 
2287 		worker = find_worker_executing_work(last_pool, work);
2288 
2289 		if (worker && worker->current_pwq->wq == wq) {
2290 			pwq = worker->current_pwq;
2291 			pool = pwq->pool;
2292 			WARN_ON_ONCE(pool != last_pool);
2293 		} else {
2294 			/* meh... not running there, queue here */
2295 			raw_spin_unlock(&last_pool->lock);
2296 			raw_spin_lock(&pool->lock);
2297 		}
2298 	} else {
2299 		raw_spin_lock(&pool->lock);
2300 	}
2301 
2302 	/*
2303 	 * pwq is determined and locked. For unbound pools, we could have raced
2304 	 * with pwq release and it could already be dead. If its refcnt is zero,
2305 	 * repeat pwq selection. Note that unbound pwqs never die without
2306 	 * another pwq replacing it in cpu_pwq or while work items are executing
2307 	 * on it, so the retrying is guaranteed to make forward-progress.
2308 	 */
2309 	if (unlikely(!pwq->refcnt)) {
2310 		if (wq->flags & WQ_UNBOUND) {
2311 			raw_spin_unlock(&pool->lock);
2312 			cpu_relax();
2313 			goto retry;
2314 		}
2315 		/* oops */
2316 		WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2317 			  wq->name, cpu);
2318 	}
2319 
2320 	/* pwq determined, queue */
2321 	trace_workqueue_queue_work(req_cpu, pwq, work);
2322 
2323 	if (WARN_ON(!list_empty(&work->entry)))
2324 		goto out;
2325 
2326 	pwq->nr_in_flight[pwq->work_color]++;
2327 	work_flags = work_color_to_flags(pwq->work_color);
2328 
2329 	/*
2330 	 * Limit the number of concurrently active work items to max_active.
2331 	 * @work must also queue behind existing inactive work items to maintain
2332 	 * ordering when max_active changes. See wq_adjust_max_active().
2333 	 */
2334 	if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) {
2335 		if (list_empty(&pool->worklist))
2336 			pool->watchdog_ts = jiffies;
2337 
2338 		trace_workqueue_activate_work(work);
2339 		insert_work(pwq, work, &pool->worklist, work_flags);
2340 		kick_pool(pool);
2341 	} else {
2342 		work_flags |= WORK_STRUCT_INACTIVE;
2343 		insert_work(pwq, work, &pwq->inactive_works, work_flags);
2344 	}
2345 
2346 out:
2347 	raw_spin_unlock(&pool->lock);
2348 	rcu_read_unlock();
2349 }
2350 
2351 static bool clear_pending_if_disabled(struct work_struct *work)
2352 {
2353 	unsigned long data = *work_data_bits(work);
2354 	struct work_offq_data offqd;
2355 
2356 	if (likely((data & WORK_STRUCT_PWQ) ||
2357 		   !(data & WORK_OFFQ_DISABLE_MASK)))
2358 		return false;
2359 
2360 	work_offqd_unpack(&offqd, data);
2361 	set_work_pool_and_clear_pending(work, offqd.pool_id,
2362 					work_offqd_pack_flags(&offqd));
2363 	return true;
2364 }
2365 
2366 /**
2367  * queue_work_on - queue work on specific cpu
2368  * @cpu: CPU number to execute work on
2369  * @wq: workqueue to use
2370  * @work: work to queue
2371  *
2372  * We queue the work to a specific CPU, the caller must ensure it
2373  * can't go away.  Callers that fail to ensure that the specified
2374  * CPU cannot go away will execute on a randomly chosen CPU.
2375  * But note well that callers specifying a CPU that never has been
2376  * online will get a splat.
2377  *
2378  * Return: %false if @work was already on a queue, %true otherwise.
2379  */
2380 bool queue_work_on(int cpu, struct workqueue_struct *wq,
2381 		   struct work_struct *work)
2382 {
2383 	bool ret = false;
2384 	unsigned long irq_flags;
2385 
2386 	local_irq_save(irq_flags);
2387 
2388 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2389 	    !clear_pending_if_disabled(work)) {
2390 		__queue_work(cpu, wq, work);
2391 		ret = true;
2392 	}
2393 
2394 	local_irq_restore(irq_flags);
2395 	return ret;
2396 }
2397 EXPORT_SYMBOL(queue_work_on);
2398 
2399 /**
2400  * select_numa_node_cpu - Select a CPU based on NUMA node
2401  * @node: NUMA node ID that we want to select a CPU from
2402  *
2403  * This function will attempt to find a "random" cpu available on a given
2404  * node. If there are no CPUs available on the given node it will return
2405  * WORK_CPU_UNBOUND indicating that we should just schedule to any
2406  * available CPU if we need to schedule this work.
2407  */
2408 static int select_numa_node_cpu(int node)
2409 {
2410 	int cpu;
2411 
2412 	/* Delay binding to CPU if node is not valid or online */
2413 	if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2414 		return WORK_CPU_UNBOUND;
2415 
2416 	/* Use local node/cpu if we are already there */
2417 	cpu = raw_smp_processor_id();
2418 	if (node == cpu_to_node(cpu))
2419 		return cpu;
2420 
2421 	/* Use "random" otherwise know as "first" online CPU of node */
2422 	cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2423 
2424 	/* If CPU is valid return that, otherwise just defer */
2425 	return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2426 }
2427 
2428 /**
2429  * queue_work_node - queue work on a "random" cpu for a given NUMA node
2430  * @node: NUMA node that we are targeting the work for
2431  * @wq: workqueue to use
2432  * @work: work to queue
2433  *
2434  * We queue the work to a "random" CPU within a given NUMA node. The basic
2435  * idea here is to provide a way to somehow associate work with a given
2436  * NUMA node.
2437  *
2438  * This function will only make a best effort attempt at getting this onto
2439  * the right NUMA node. If no node is requested or the requested node is
2440  * offline then we just fall back to standard queue_work behavior.
2441  *
2442  * Currently the "random" CPU ends up being the first available CPU in the
2443  * intersection of cpu_online_mask and the cpumask of the node, unless we
2444  * are running on the node. In that case we just use the current CPU.
2445  *
2446  * Return: %false if @work was already on a queue, %true otherwise.
2447  */
2448 bool queue_work_node(int node, struct workqueue_struct *wq,
2449 		     struct work_struct *work)
2450 {
2451 	unsigned long irq_flags;
2452 	bool ret = false;
2453 
2454 	/*
2455 	 * This current implementation is specific to unbound workqueues.
2456 	 * Specifically we only return the first available CPU for a given
2457 	 * node instead of cycling through individual CPUs within the node.
2458 	 *
2459 	 * If this is used with a per-cpu workqueue then the logic in
2460 	 * workqueue_select_cpu_near would need to be updated to allow for
2461 	 * some round robin type logic.
2462 	 */
2463 	WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2464 
2465 	local_irq_save(irq_flags);
2466 
2467 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2468 	    !clear_pending_if_disabled(work)) {
2469 		int cpu = select_numa_node_cpu(node);
2470 
2471 		__queue_work(cpu, wq, work);
2472 		ret = true;
2473 	}
2474 
2475 	local_irq_restore(irq_flags);
2476 	return ret;
2477 }
2478 EXPORT_SYMBOL_GPL(queue_work_node);
2479 
2480 void delayed_work_timer_fn(struct timer_list *t)
2481 {
2482 	struct delayed_work *dwork = from_timer(dwork, t, timer);
2483 
2484 	/* should have been called from irqsafe timer with irq already off */
2485 	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
2486 }
2487 EXPORT_SYMBOL(delayed_work_timer_fn);
2488 
2489 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2490 				struct delayed_work *dwork, unsigned long delay)
2491 {
2492 	struct timer_list *timer = &dwork->timer;
2493 	struct work_struct *work = &dwork->work;
2494 
2495 	WARN_ON_ONCE(!wq);
2496 	WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2497 	WARN_ON_ONCE(timer_pending(timer));
2498 	WARN_ON_ONCE(!list_empty(&work->entry));
2499 
2500 	/*
2501 	 * If @delay is 0, queue @dwork->work immediately.  This is for
2502 	 * both optimization and correctness.  The earliest @timer can
2503 	 * expire is on the closest next tick and delayed_work users depend
2504 	 * on that there's no such delay when @delay is 0.
2505 	 */
2506 	if (!delay) {
2507 		__queue_work(cpu, wq, &dwork->work);
2508 		return;
2509 	}
2510 
2511 	dwork->wq = wq;
2512 	dwork->cpu = cpu;
2513 	timer->expires = jiffies + delay;
2514 
2515 	if (housekeeping_enabled(HK_TYPE_TIMER)) {
2516 		/* If the current cpu is a housekeeping cpu, use it. */
2517 		cpu = smp_processor_id();
2518 		if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER))
2519 			cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
2520 		add_timer_on(timer, cpu);
2521 	} else {
2522 		if (likely(cpu == WORK_CPU_UNBOUND))
2523 			add_timer_global(timer);
2524 		else
2525 			add_timer_on(timer, cpu);
2526 	}
2527 }
2528 
2529 /**
2530  * queue_delayed_work_on - queue work on specific CPU after delay
2531  * @cpu: CPU number to execute work on
2532  * @wq: workqueue to use
2533  * @dwork: work to queue
2534  * @delay: number of jiffies to wait before queueing
2535  *
2536  * Return: %false if @work was already on a queue, %true otherwise.  If
2537  * @delay is zero and @dwork is idle, it will be scheduled for immediate
2538  * execution.
2539  */
2540 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2541 			   struct delayed_work *dwork, unsigned long delay)
2542 {
2543 	struct work_struct *work = &dwork->work;
2544 	bool ret = false;
2545 	unsigned long irq_flags;
2546 
2547 	/* read the comment in __queue_work() */
2548 	local_irq_save(irq_flags);
2549 
2550 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2551 	    !clear_pending_if_disabled(work)) {
2552 		__queue_delayed_work(cpu, wq, dwork, delay);
2553 		ret = true;
2554 	}
2555 
2556 	local_irq_restore(irq_flags);
2557 	return ret;
2558 }
2559 EXPORT_SYMBOL(queue_delayed_work_on);
2560 
2561 /**
2562  * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2563  * @cpu: CPU number to execute work on
2564  * @wq: workqueue to use
2565  * @dwork: work to queue
2566  * @delay: number of jiffies to wait before queueing
2567  *
2568  * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2569  * modify @dwork's timer so that it expires after @delay.  If @delay is
2570  * zero, @work is guaranteed to be scheduled immediately regardless of its
2571  * current state.
2572  *
2573  * Return: %false if @dwork was idle and queued, %true if @dwork was
2574  * pending and its timer was modified.
2575  *
2576  * This function is safe to call from any context including IRQ handler.
2577  * See try_to_grab_pending() for details.
2578  */
2579 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2580 			 struct delayed_work *dwork, unsigned long delay)
2581 {
2582 	unsigned long irq_flags;
2583 	bool ret;
2584 
2585 	ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags);
2586 
2587 	if (!clear_pending_if_disabled(&dwork->work))
2588 		__queue_delayed_work(cpu, wq, dwork, delay);
2589 
2590 	local_irq_restore(irq_flags);
2591 	return ret;
2592 }
2593 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2594 
2595 static void rcu_work_rcufn(struct rcu_head *rcu)
2596 {
2597 	struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2598 
2599 	/* read the comment in __queue_work() */
2600 	local_irq_disable();
2601 	__queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2602 	local_irq_enable();
2603 }
2604 
2605 /**
2606  * queue_rcu_work - queue work after a RCU grace period
2607  * @wq: workqueue to use
2608  * @rwork: work to queue
2609  *
2610  * Return: %false if @rwork was already pending, %true otherwise.  Note
2611  * that a full RCU grace period is guaranteed only after a %true return.
2612  * While @rwork is guaranteed to be executed after a %false return, the
2613  * execution may happen before a full RCU grace period has passed.
2614  */
2615 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2616 {
2617 	struct work_struct *work = &rwork->work;
2618 
2619 	/*
2620 	 * rcu_work can't be canceled or disabled. Warn if the user reached
2621 	 * inside @rwork and disabled the inner work.
2622 	 */
2623 	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2624 	    !WARN_ON_ONCE(clear_pending_if_disabled(work))) {
2625 		rwork->wq = wq;
2626 		call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2627 		return true;
2628 	}
2629 
2630 	return false;
2631 }
2632 EXPORT_SYMBOL(queue_rcu_work);
2633 
2634 static struct worker *alloc_worker(int node)
2635 {
2636 	struct worker *worker;
2637 
2638 	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2639 	if (worker) {
2640 		INIT_LIST_HEAD(&worker->entry);
2641 		INIT_LIST_HEAD(&worker->scheduled);
2642 		INIT_LIST_HEAD(&worker->node);
2643 		/* on creation a worker is in !idle && prep state */
2644 		worker->flags = WORKER_PREP;
2645 	}
2646 	return worker;
2647 }
2648 
2649 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2650 {
2651 	if (pool->cpu < 0 && pool->attrs->affn_strict)
2652 		return pool->attrs->__pod_cpumask;
2653 	else
2654 		return pool->attrs->cpumask;
2655 }
2656 
2657 /**
2658  * worker_attach_to_pool() - attach a worker to a pool
2659  * @worker: worker to be attached
2660  * @pool: the target pool
2661  *
2662  * Attach @worker to @pool.  Once attached, the %WORKER_UNBOUND flag and
2663  * cpu-binding of @worker are kept coordinated with the pool across
2664  * cpu-[un]hotplugs.
2665  */
2666 static void worker_attach_to_pool(struct worker *worker,
2667 				  struct worker_pool *pool)
2668 {
2669 	mutex_lock(&wq_pool_attach_mutex);
2670 
2671 	/*
2672 	 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable
2673 	 * across this function. See the comments above the flag definition for
2674 	 * details. BH workers are, while per-CPU, always DISASSOCIATED.
2675 	 */
2676 	if (pool->flags & POOL_DISASSOCIATED) {
2677 		worker->flags |= WORKER_UNBOUND;
2678 	} else {
2679 		WARN_ON_ONCE(pool->flags & POOL_BH);
2680 		kthread_set_per_cpu(worker->task, pool->cpu);
2681 	}
2682 
2683 	if (worker->rescue_wq)
2684 		set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2685 
2686 	list_add_tail(&worker->node, &pool->workers);
2687 	worker->pool = pool;
2688 
2689 	mutex_unlock(&wq_pool_attach_mutex);
2690 }
2691 
2692 static void unbind_worker(struct worker *worker)
2693 {
2694 	lockdep_assert_held(&wq_pool_attach_mutex);
2695 
2696 	kthread_set_per_cpu(worker->task, -1);
2697 	if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2698 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2699 	else
2700 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2701 }
2702 
2703 
2704 static void detach_worker(struct worker *worker)
2705 {
2706 	lockdep_assert_held(&wq_pool_attach_mutex);
2707 
2708 	unbind_worker(worker);
2709 	list_del(&worker->node);
2710 }
2711 
2712 /**
2713  * worker_detach_from_pool() - detach a worker from its pool
2714  * @worker: worker which is attached to its pool
2715  *
2716  * Undo the attaching which had been done in worker_attach_to_pool().  The
2717  * caller worker shouldn't access to the pool after detached except it has
2718  * other reference to the pool.
2719  */
2720 static void worker_detach_from_pool(struct worker *worker)
2721 {
2722 	struct worker_pool *pool = worker->pool;
2723 
2724 	/* there is one permanent BH worker per CPU which should never detach */
2725 	WARN_ON_ONCE(pool->flags & POOL_BH);
2726 
2727 	mutex_lock(&wq_pool_attach_mutex);
2728 	detach_worker(worker);
2729 	worker->pool = NULL;
2730 	mutex_unlock(&wq_pool_attach_mutex);
2731 
2732 	/* clear leftover flags without pool->lock after it is detached */
2733 	worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2734 }
2735 
2736 static int format_worker_id(char *buf, size_t size, struct worker *worker,
2737 			    struct worker_pool *pool)
2738 {
2739 	if (worker->rescue_wq)
2740 		return scnprintf(buf, size, "kworker/R-%s",
2741 				 worker->rescue_wq->name);
2742 
2743 	if (pool) {
2744 		if (pool->cpu >= 0)
2745 			return scnprintf(buf, size, "kworker/%d:%d%s",
2746 					 pool->cpu, worker->id,
2747 					 pool->attrs->nice < 0  ? "H" : "");
2748 		else
2749 			return scnprintf(buf, size, "kworker/u%d:%d",
2750 					 pool->id, worker->id);
2751 	} else {
2752 		return scnprintf(buf, size, "kworker/dying");
2753 	}
2754 }
2755 
2756 /**
2757  * create_worker - create a new workqueue worker
2758  * @pool: pool the new worker will belong to
2759  *
2760  * Create and start a new worker which is attached to @pool.
2761  *
2762  * CONTEXT:
2763  * Might sleep.  Does GFP_KERNEL allocations.
2764  *
2765  * Return:
2766  * Pointer to the newly created worker.
2767  */
2768 static struct worker *create_worker(struct worker_pool *pool)
2769 {
2770 	struct worker *worker;
2771 	int id;
2772 
2773 	/* ID is needed to determine kthread name */
2774 	id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2775 	if (id < 0) {
2776 		pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2777 			    ERR_PTR(id));
2778 		return NULL;
2779 	}
2780 
2781 	worker = alloc_worker(pool->node);
2782 	if (!worker) {
2783 		pr_err_once("workqueue: Failed to allocate a worker\n");
2784 		goto fail;
2785 	}
2786 
2787 	worker->id = id;
2788 
2789 	if (!(pool->flags & POOL_BH)) {
2790 		char id_buf[WORKER_ID_LEN];
2791 
2792 		format_worker_id(id_buf, sizeof(id_buf), worker, pool);
2793 		worker->task = kthread_create_on_node(worker_thread, worker,
2794 						      pool->node, "%s", id_buf);
2795 		if (IS_ERR(worker->task)) {
2796 			if (PTR_ERR(worker->task) == -EINTR) {
2797 				pr_err("workqueue: Interrupted when creating a worker thread \"%s\"\n",
2798 				       id_buf);
2799 			} else {
2800 				pr_err_once("workqueue: Failed to create a worker thread: %pe",
2801 					    worker->task);
2802 			}
2803 			goto fail;
2804 		}
2805 
2806 		set_user_nice(worker->task, pool->attrs->nice);
2807 		kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2808 	}
2809 
2810 	/* successful, attach the worker to the pool */
2811 	worker_attach_to_pool(worker, pool);
2812 
2813 	/* start the newly created worker */
2814 	raw_spin_lock_irq(&pool->lock);
2815 
2816 	worker->pool->nr_workers++;
2817 	worker_enter_idle(worker);
2818 
2819 	/*
2820 	 * @worker is waiting on a completion in kthread() and will trigger hung
2821 	 * check if not woken up soon. As kick_pool() is noop if @pool is empty,
2822 	 * wake it up explicitly.
2823 	 */
2824 	if (worker->task)
2825 		wake_up_process(worker->task);
2826 
2827 	raw_spin_unlock_irq(&pool->lock);
2828 
2829 	return worker;
2830 
2831 fail:
2832 	ida_free(&pool->worker_ida, id);
2833 	kfree(worker);
2834 	return NULL;
2835 }
2836 
2837 static void detach_dying_workers(struct list_head *cull_list)
2838 {
2839 	struct worker *worker;
2840 
2841 	list_for_each_entry(worker, cull_list, entry)
2842 		detach_worker(worker);
2843 }
2844 
2845 static void reap_dying_workers(struct list_head *cull_list)
2846 {
2847 	struct worker *worker, *tmp;
2848 
2849 	list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2850 		list_del_init(&worker->entry);
2851 		kthread_stop_put(worker->task);
2852 		kfree(worker);
2853 	}
2854 }
2855 
2856 /**
2857  * set_worker_dying - Tag a worker for destruction
2858  * @worker: worker to be destroyed
2859  * @list: transfer worker away from its pool->idle_list and into list
2860  *
2861  * Tag @worker for destruction and adjust @pool stats accordingly.  The worker
2862  * should be idle.
2863  *
2864  * CONTEXT:
2865  * raw_spin_lock_irq(pool->lock).
2866  */
2867 static void set_worker_dying(struct worker *worker, struct list_head *list)
2868 {
2869 	struct worker_pool *pool = worker->pool;
2870 
2871 	lockdep_assert_held(&pool->lock);
2872 	lockdep_assert_held(&wq_pool_attach_mutex);
2873 
2874 	/* sanity check frenzy */
2875 	if (WARN_ON(worker->current_work) ||
2876 	    WARN_ON(!list_empty(&worker->scheduled)) ||
2877 	    WARN_ON(!(worker->flags & WORKER_IDLE)))
2878 		return;
2879 
2880 	pool->nr_workers--;
2881 	pool->nr_idle--;
2882 
2883 	worker->flags |= WORKER_DIE;
2884 
2885 	list_move(&worker->entry, list);
2886 
2887 	/* get an extra task struct reference for later kthread_stop_put() */
2888 	get_task_struct(worker->task);
2889 }
2890 
2891 /**
2892  * idle_worker_timeout - check if some idle workers can now be deleted.
2893  * @t: The pool's idle_timer that just expired
2894  *
2895  * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2896  * worker_leave_idle(), as a worker flicking between idle and active while its
2897  * pool is at the too_many_workers() tipping point would cause too much timer
2898  * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2899  * it expire and re-evaluate things from there.
2900  */
2901 static void idle_worker_timeout(struct timer_list *t)
2902 {
2903 	struct worker_pool *pool = from_timer(pool, t, idle_timer);
2904 	bool do_cull = false;
2905 
2906 	if (work_pending(&pool->idle_cull_work))
2907 		return;
2908 
2909 	raw_spin_lock_irq(&pool->lock);
2910 
2911 	if (too_many_workers(pool)) {
2912 		struct worker *worker;
2913 		unsigned long expires;
2914 
2915 		/* idle_list is kept in LIFO order, check the last one */
2916 		worker = list_last_entry(&pool->idle_list, struct worker, entry);
2917 		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2918 		do_cull = !time_before(jiffies, expires);
2919 
2920 		if (!do_cull)
2921 			mod_timer(&pool->idle_timer, expires);
2922 	}
2923 	raw_spin_unlock_irq(&pool->lock);
2924 
2925 	if (do_cull)
2926 		queue_work(system_unbound_wq, &pool->idle_cull_work);
2927 }
2928 
2929 /**
2930  * idle_cull_fn - cull workers that have been idle for too long.
2931  * @work: the pool's work for handling these idle workers
2932  *
2933  * This goes through a pool's idle workers and gets rid of those that have been
2934  * idle for at least IDLE_WORKER_TIMEOUT seconds.
2935  *
2936  * We don't want to disturb isolated CPUs because of a pcpu kworker being
2937  * culled, so this also resets worker affinity. This requires a sleepable
2938  * context, hence the split between timer callback and work item.
2939  */
2940 static void idle_cull_fn(struct work_struct *work)
2941 {
2942 	struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2943 	LIST_HEAD(cull_list);
2944 
2945 	/*
2946 	 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2947 	 * cannot proceed beyong set_pf_worker() in its self-destruct path.
2948 	 * This is required as a previously-preempted worker could run after
2949 	 * set_worker_dying() has happened but before detach_dying_workers() did.
2950 	 */
2951 	mutex_lock(&wq_pool_attach_mutex);
2952 	raw_spin_lock_irq(&pool->lock);
2953 
2954 	while (too_many_workers(pool)) {
2955 		struct worker *worker;
2956 		unsigned long expires;
2957 
2958 		worker = list_last_entry(&pool->idle_list, struct worker, entry);
2959 		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2960 
2961 		if (time_before(jiffies, expires)) {
2962 			mod_timer(&pool->idle_timer, expires);
2963 			break;
2964 		}
2965 
2966 		set_worker_dying(worker, &cull_list);
2967 	}
2968 
2969 	raw_spin_unlock_irq(&pool->lock);
2970 	detach_dying_workers(&cull_list);
2971 	mutex_unlock(&wq_pool_attach_mutex);
2972 
2973 	reap_dying_workers(&cull_list);
2974 }
2975 
2976 static void send_mayday(struct work_struct *work)
2977 {
2978 	struct pool_workqueue *pwq = get_work_pwq(work);
2979 	struct workqueue_struct *wq = pwq->wq;
2980 
2981 	lockdep_assert_held(&wq_mayday_lock);
2982 
2983 	if (!wq->rescuer)
2984 		return;
2985 
2986 	/* mayday mayday mayday */
2987 	if (list_empty(&pwq->mayday_node)) {
2988 		/*
2989 		 * If @pwq is for an unbound wq, its base ref may be put at
2990 		 * any time due to an attribute change.  Pin @pwq until the
2991 		 * rescuer is done with it.
2992 		 */
2993 		get_pwq(pwq);
2994 		list_add_tail(&pwq->mayday_node, &wq->maydays);
2995 		wake_up_process(wq->rescuer->task);
2996 		pwq->stats[PWQ_STAT_MAYDAY]++;
2997 	}
2998 }
2999 
3000 static void pool_mayday_timeout(struct timer_list *t)
3001 {
3002 	struct worker_pool *pool = from_timer(pool, t, mayday_timer);
3003 	struct work_struct *work;
3004 
3005 	raw_spin_lock_irq(&pool->lock);
3006 	raw_spin_lock(&wq_mayday_lock);		/* for wq->maydays */
3007 
3008 	if (need_to_create_worker(pool)) {
3009 		/*
3010 		 * We've been trying to create a new worker but
3011 		 * haven't been successful.  We might be hitting an
3012 		 * allocation deadlock.  Send distress signals to
3013 		 * rescuers.
3014 		 */
3015 		list_for_each_entry(work, &pool->worklist, entry)
3016 			send_mayday(work);
3017 	}
3018 
3019 	raw_spin_unlock(&wq_mayday_lock);
3020 	raw_spin_unlock_irq(&pool->lock);
3021 
3022 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3023 }
3024 
3025 /**
3026  * maybe_create_worker - create a new worker if necessary
3027  * @pool: pool to create a new worker for
3028  *
3029  * Create a new worker for @pool if necessary.  @pool is guaranteed to
3030  * have at least one idle worker on return from this function.  If
3031  * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3032  * sent to all rescuers with works scheduled on @pool to resolve
3033  * possible allocation deadlock.
3034  *
3035  * On return, need_to_create_worker() is guaranteed to be %false and
3036  * may_start_working() %true.
3037  *
3038  * LOCKING:
3039  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3040  * multiple times.  Does GFP_KERNEL allocations.  Called only from
3041  * manager.
3042  */
3043 static void maybe_create_worker(struct worker_pool *pool)
3044 __releases(&pool->lock)
3045 __acquires(&pool->lock)
3046 {
3047 restart:
3048 	raw_spin_unlock_irq(&pool->lock);
3049 
3050 	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3051 	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3052 
3053 	while (true) {
3054 		if (create_worker(pool) || !need_to_create_worker(pool))
3055 			break;
3056 
3057 		schedule_timeout_interruptible(CREATE_COOLDOWN);
3058 
3059 		if (!need_to_create_worker(pool))
3060 			break;
3061 	}
3062 
3063 	del_timer_sync(&pool->mayday_timer);
3064 	raw_spin_lock_irq(&pool->lock);
3065 	/*
3066 	 * This is necessary even after a new worker was just successfully
3067 	 * created as @pool->lock was dropped and the new worker might have
3068 	 * already become busy.
3069 	 */
3070 	if (need_to_create_worker(pool))
3071 		goto restart;
3072 }
3073 
3074 /**
3075  * manage_workers - manage worker pool
3076  * @worker: self
3077  *
3078  * Assume the manager role and manage the worker pool @worker belongs
3079  * to.  At any given time, there can be only zero or one manager per
3080  * pool.  The exclusion is handled automatically by this function.
3081  *
3082  * The caller can safely start processing works on false return.  On
3083  * true return, it's guaranteed that need_to_create_worker() is false
3084  * and may_start_working() is true.
3085  *
3086  * CONTEXT:
3087  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3088  * multiple times.  Does GFP_KERNEL allocations.
3089  *
3090  * Return:
3091  * %false if the pool doesn't need management and the caller can safely
3092  * start processing works, %true if management function was performed and
3093  * the conditions that the caller verified before calling the function may
3094  * no longer be true.
3095  */
3096 static bool manage_workers(struct worker *worker)
3097 {
3098 	struct worker_pool *pool = worker->pool;
3099 
3100 	if (pool->flags & POOL_MANAGER_ACTIVE)
3101 		return false;
3102 
3103 	pool->flags |= POOL_MANAGER_ACTIVE;
3104 	pool->manager = worker;
3105 
3106 	maybe_create_worker(pool);
3107 
3108 	pool->manager = NULL;
3109 	pool->flags &= ~POOL_MANAGER_ACTIVE;
3110 	rcuwait_wake_up(&manager_wait);
3111 	return true;
3112 }
3113 
3114 /**
3115  * process_one_work - process single work
3116  * @worker: self
3117  * @work: work to process
3118  *
3119  * Process @work.  This function contains all the logics necessary to
3120  * process a single work including synchronization against and
3121  * interaction with other workers on the same cpu, queueing and
3122  * flushing.  As long as context requirement is met, any worker can
3123  * call this function to process a work.
3124  *
3125  * CONTEXT:
3126  * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3127  */
3128 static void process_one_work(struct worker *worker, struct work_struct *work)
3129 __releases(&pool->lock)
3130 __acquires(&pool->lock)
3131 {
3132 	struct pool_workqueue *pwq = get_work_pwq(work);
3133 	struct worker_pool *pool = worker->pool;
3134 	unsigned long work_data;
3135 	int lockdep_start_depth, rcu_start_depth;
3136 	bool bh_draining = pool->flags & POOL_BH_DRAINING;
3137 #ifdef CONFIG_LOCKDEP
3138 	/*
3139 	 * It is permissible to free the struct work_struct from
3140 	 * inside the function that is called from it, this we need to
3141 	 * take into account for lockdep too.  To avoid bogus "held
3142 	 * lock freed" warnings as well as problems when looking into
3143 	 * work->lockdep_map, make a copy and use that here.
3144 	 */
3145 	struct lockdep_map lockdep_map;
3146 
3147 	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3148 #endif
3149 	/* ensure we're on the correct CPU */
3150 	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3151 		     raw_smp_processor_id() != pool->cpu);
3152 
3153 	/* claim and dequeue */
3154 	debug_work_deactivate(work);
3155 	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3156 	worker->current_work = work;
3157 	worker->current_func = work->func;
3158 	worker->current_pwq = pwq;
3159 	if (worker->task)
3160 		worker->current_at = worker->task->se.sum_exec_runtime;
3161 	work_data = *work_data_bits(work);
3162 	worker->current_color = get_work_color(work_data);
3163 
3164 	/*
3165 	 * Record wq name for cmdline and debug reporting, may get
3166 	 * overridden through set_worker_desc().
3167 	 */
3168 	strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3169 
3170 	list_del_init(&work->entry);
3171 
3172 	/*
3173 	 * CPU intensive works don't participate in concurrency management.
3174 	 * They're the scheduler's responsibility.  This takes @worker out
3175 	 * of concurrency management and the next code block will chain
3176 	 * execution of the pending work items.
3177 	 */
3178 	if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3179 		worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3180 
3181 	/*
3182 	 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3183 	 * since nr_running would always be >= 1 at this point. This is used to
3184 	 * chain execution of the pending work items for WORKER_NOT_RUNNING
3185 	 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3186 	 */
3187 	kick_pool(pool);
3188 
3189 	/*
3190 	 * Record the last pool and clear PENDING which should be the last
3191 	 * update to @work.  Also, do this inside @pool->lock so that
3192 	 * PENDING and queued state changes happen together while IRQ is
3193 	 * disabled.
3194 	 */
3195 	set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool));
3196 
3197 	pwq->stats[PWQ_STAT_STARTED]++;
3198 	raw_spin_unlock_irq(&pool->lock);
3199 
3200 	rcu_start_depth = rcu_preempt_depth();
3201 	lockdep_start_depth = lockdep_depth(current);
3202 	/* see drain_dead_softirq_workfn() */
3203 	if (!bh_draining)
3204 		lock_map_acquire(pwq->wq->lockdep_map);
3205 	lock_map_acquire(&lockdep_map);
3206 	/*
3207 	 * Strictly speaking we should mark the invariant state without holding
3208 	 * any locks, that is, before these two lock_map_acquire()'s.
3209 	 *
3210 	 * However, that would result in:
3211 	 *
3212 	 *   A(W1)
3213 	 *   WFC(C)
3214 	 *		A(W1)
3215 	 *		C(C)
3216 	 *
3217 	 * Which would create W1->C->W1 dependencies, even though there is no
3218 	 * actual deadlock possible. There are two solutions, using a
3219 	 * read-recursive acquire on the work(queue) 'locks', but this will then
3220 	 * hit the lockdep limitation on recursive locks, or simply discard
3221 	 * these locks.
3222 	 *
3223 	 * AFAICT there is no possible deadlock scenario between the
3224 	 * flush_work() and complete() primitives (except for single-threaded
3225 	 * workqueues), so hiding them isn't a problem.
3226 	 */
3227 	lockdep_invariant_state(true);
3228 	trace_workqueue_execute_start(work);
3229 	worker->current_func(work);
3230 	/*
3231 	 * While we must be careful to not use "work" after this, the trace
3232 	 * point will only record its address.
3233 	 */
3234 	trace_workqueue_execute_end(work, worker->current_func);
3235 	pwq->stats[PWQ_STAT_COMPLETED]++;
3236 	lock_map_release(&lockdep_map);
3237 	if (!bh_draining)
3238 		lock_map_release(pwq->wq->lockdep_map);
3239 
3240 	if (unlikely((worker->task && in_atomic()) ||
3241 		     lockdep_depth(current) != lockdep_start_depth ||
3242 		     rcu_preempt_depth() != rcu_start_depth)) {
3243 		pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n"
3244 		       "     preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n",
3245 		       current->comm, task_pid_nr(current), preempt_count(),
3246 		       lockdep_start_depth, lockdep_depth(current),
3247 		       rcu_start_depth, rcu_preempt_depth(),
3248 		       worker->current_func);
3249 		debug_show_held_locks(current);
3250 		dump_stack();
3251 	}
3252 
3253 	/*
3254 	 * The following prevents a kworker from hogging CPU on !PREEMPTION
3255 	 * kernels, where a requeueing work item waiting for something to
3256 	 * happen could deadlock with stop_machine as such work item could
3257 	 * indefinitely requeue itself while all other CPUs are trapped in
3258 	 * stop_machine. At the same time, report a quiescent RCU state so
3259 	 * the same condition doesn't freeze RCU.
3260 	 */
3261 	if (worker->task)
3262 		cond_resched();
3263 
3264 	raw_spin_lock_irq(&pool->lock);
3265 
3266 	/*
3267 	 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3268 	 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3269 	 * wq_cpu_intensive_thresh_us. Clear it.
3270 	 */
3271 	worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3272 
3273 	/* tag the worker for identification in schedule() */
3274 	worker->last_func = worker->current_func;
3275 
3276 	/* we're done with it, release */
3277 	hash_del(&worker->hentry);
3278 	worker->current_work = NULL;
3279 	worker->current_func = NULL;
3280 	worker->current_pwq = NULL;
3281 	worker->current_color = INT_MAX;
3282 
3283 	/* must be the last step, see the function comment */
3284 	pwq_dec_nr_in_flight(pwq, work_data);
3285 }
3286 
3287 /**
3288  * process_scheduled_works - process scheduled works
3289  * @worker: self
3290  *
3291  * Process all scheduled works.  Please note that the scheduled list
3292  * may change while processing a work, so this function repeatedly
3293  * fetches a work from the top and executes it.
3294  *
3295  * CONTEXT:
3296  * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3297  * multiple times.
3298  */
3299 static void process_scheduled_works(struct worker *worker)
3300 {
3301 	struct work_struct *work;
3302 	bool first = true;
3303 
3304 	while ((work = list_first_entry_or_null(&worker->scheduled,
3305 						struct work_struct, entry))) {
3306 		if (first) {
3307 			worker->pool->watchdog_ts = jiffies;
3308 			first = false;
3309 		}
3310 		process_one_work(worker, work);
3311 	}
3312 }
3313 
3314 static void set_pf_worker(bool val)
3315 {
3316 	mutex_lock(&wq_pool_attach_mutex);
3317 	if (val)
3318 		current->flags |= PF_WQ_WORKER;
3319 	else
3320 		current->flags &= ~PF_WQ_WORKER;
3321 	mutex_unlock(&wq_pool_attach_mutex);
3322 }
3323 
3324 /**
3325  * worker_thread - the worker thread function
3326  * @__worker: self
3327  *
3328  * The worker thread function.  All workers belong to a worker_pool -
3329  * either a per-cpu one or dynamic unbound one.  These workers process all
3330  * work items regardless of their specific target workqueue.  The only
3331  * exception is work items which belong to workqueues with a rescuer which
3332  * will be explained in rescuer_thread().
3333  *
3334  * Return: 0
3335  */
3336 static int worker_thread(void *__worker)
3337 {
3338 	struct worker *worker = __worker;
3339 	struct worker_pool *pool = worker->pool;
3340 
3341 	/* tell the scheduler that this is a workqueue worker */
3342 	set_pf_worker(true);
3343 woke_up:
3344 	raw_spin_lock_irq(&pool->lock);
3345 
3346 	/* am I supposed to die? */
3347 	if (unlikely(worker->flags & WORKER_DIE)) {
3348 		raw_spin_unlock_irq(&pool->lock);
3349 		set_pf_worker(false);
3350 		/*
3351 		 * The worker is dead and PF_WQ_WORKER is cleared, worker->pool
3352 		 * shouldn't be accessed, reset it to NULL in case otherwise.
3353 		 */
3354 		worker->pool = NULL;
3355 		ida_free(&pool->worker_ida, worker->id);
3356 		return 0;
3357 	}
3358 
3359 	worker_leave_idle(worker);
3360 recheck:
3361 	/* no more worker necessary? */
3362 	if (!need_more_worker(pool))
3363 		goto sleep;
3364 
3365 	/* do we need to manage? */
3366 	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3367 		goto recheck;
3368 
3369 	/*
3370 	 * ->scheduled list can only be filled while a worker is
3371 	 * preparing to process a work or actually processing it.
3372 	 * Make sure nobody diddled with it while I was sleeping.
3373 	 */
3374 	WARN_ON_ONCE(!list_empty(&worker->scheduled));
3375 
3376 	/*
3377 	 * Finish PREP stage.  We're guaranteed to have at least one idle
3378 	 * worker or that someone else has already assumed the manager
3379 	 * role.  This is where @worker starts participating in concurrency
3380 	 * management if applicable and concurrency management is restored
3381 	 * after being rebound.  See rebind_workers() for details.
3382 	 */
3383 	worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3384 
3385 	do {
3386 		struct work_struct *work =
3387 			list_first_entry(&pool->worklist,
3388 					 struct work_struct, entry);
3389 
3390 		if (assign_work(work, worker, NULL))
3391 			process_scheduled_works(worker);
3392 	} while (keep_working(pool));
3393 
3394 	worker_set_flags(worker, WORKER_PREP);
3395 sleep:
3396 	/*
3397 	 * pool->lock is held and there's no work to process and no need to
3398 	 * manage, sleep.  Workers are woken up only while holding
3399 	 * pool->lock or from local cpu, so setting the current state
3400 	 * before releasing pool->lock is enough to prevent losing any
3401 	 * event.
3402 	 */
3403 	worker_enter_idle(worker);
3404 	__set_current_state(TASK_IDLE);
3405 	raw_spin_unlock_irq(&pool->lock);
3406 	schedule();
3407 	goto woke_up;
3408 }
3409 
3410 /**
3411  * rescuer_thread - the rescuer thread function
3412  * @__rescuer: self
3413  *
3414  * Workqueue rescuer thread function.  There's one rescuer for each
3415  * workqueue which has WQ_MEM_RECLAIM set.
3416  *
3417  * Regular work processing on a pool may block trying to create a new
3418  * worker which uses GFP_KERNEL allocation which has slight chance of
3419  * developing into deadlock if some works currently on the same queue
3420  * need to be processed to satisfy the GFP_KERNEL allocation.  This is
3421  * the problem rescuer solves.
3422  *
3423  * When such condition is possible, the pool summons rescuers of all
3424  * workqueues which have works queued on the pool and let them process
3425  * those works so that forward progress can be guaranteed.
3426  *
3427  * This should happen rarely.
3428  *
3429  * Return: 0
3430  */
3431 static int rescuer_thread(void *__rescuer)
3432 {
3433 	struct worker *rescuer = __rescuer;
3434 	struct workqueue_struct *wq = rescuer->rescue_wq;
3435 	bool should_stop;
3436 
3437 	set_user_nice(current, RESCUER_NICE_LEVEL);
3438 
3439 	/*
3440 	 * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
3441 	 * doesn't participate in concurrency management.
3442 	 */
3443 	set_pf_worker(true);
3444 repeat:
3445 	set_current_state(TASK_IDLE);
3446 
3447 	/*
3448 	 * By the time the rescuer is requested to stop, the workqueue
3449 	 * shouldn't have any work pending, but @wq->maydays may still have
3450 	 * pwq(s) queued.  This can happen by non-rescuer workers consuming
3451 	 * all the work items before the rescuer got to them.  Go through
3452 	 * @wq->maydays processing before acting on should_stop so that the
3453 	 * list is always empty on exit.
3454 	 */
3455 	should_stop = kthread_should_stop();
3456 
3457 	/* see whether any pwq is asking for help */
3458 	raw_spin_lock_irq(&wq_mayday_lock);
3459 
3460 	while (!list_empty(&wq->maydays)) {
3461 		struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3462 					struct pool_workqueue, mayday_node);
3463 		struct worker_pool *pool = pwq->pool;
3464 		struct work_struct *work, *n;
3465 
3466 		__set_current_state(TASK_RUNNING);
3467 		list_del_init(&pwq->mayday_node);
3468 
3469 		raw_spin_unlock_irq(&wq_mayday_lock);
3470 
3471 		worker_attach_to_pool(rescuer, pool);
3472 
3473 		raw_spin_lock_irq(&pool->lock);
3474 
3475 		/*
3476 		 * Slurp in all works issued via this workqueue and
3477 		 * process'em.
3478 		 */
3479 		WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3480 		list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3481 			if (get_work_pwq(work) == pwq &&
3482 			    assign_work(work, rescuer, &n))
3483 				pwq->stats[PWQ_STAT_RESCUED]++;
3484 		}
3485 
3486 		if (!list_empty(&rescuer->scheduled)) {
3487 			process_scheduled_works(rescuer);
3488 
3489 			/*
3490 			 * The above execution of rescued work items could
3491 			 * have created more to rescue through
3492 			 * pwq_activate_first_inactive() or chained
3493 			 * queueing.  Let's put @pwq back on mayday list so
3494 			 * that such back-to-back work items, which may be
3495 			 * being used to relieve memory pressure, don't
3496 			 * incur MAYDAY_INTERVAL delay inbetween.
3497 			 */
3498 			if (pwq->nr_active && need_to_create_worker(pool)) {
3499 				raw_spin_lock(&wq_mayday_lock);
3500 				/*
3501 				 * Queue iff we aren't racing destruction
3502 				 * and somebody else hasn't queued it already.
3503 				 */
3504 				if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3505 					get_pwq(pwq);
3506 					list_add_tail(&pwq->mayday_node, &wq->maydays);
3507 				}
3508 				raw_spin_unlock(&wq_mayday_lock);
3509 			}
3510 		}
3511 
3512 		/*
3513 		 * Put the reference grabbed by send_mayday().  @pool won't
3514 		 * go away while we're still attached to it.
3515 		 */
3516 		put_pwq(pwq);
3517 
3518 		/*
3519 		 * Leave this pool. Notify regular workers; otherwise, we end up
3520 		 * with 0 concurrency and stalling the execution.
3521 		 */
3522 		kick_pool(pool);
3523 
3524 		raw_spin_unlock_irq(&pool->lock);
3525 
3526 		worker_detach_from_pool(rescuer);
3527 
3528 		raw_spin_lock_irq(&wq_mayday_lock);
3529 	}
3530 
3531 	raw_spin_unlock_irq(&wq_mayday_lock);
3532 
3533 	if (should_stop) {
3534 		__set_current_state(TASK_RUNNING);
3535 		set_pf_worker(false);
3536 		return 0;
3537 	}
3538 
3539 	/* rescuers should never participate in concurrency management */
3540 	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3541 	schedule();
3542 	goto repeat;
3543 }
3544 
3545 static void bh_worker(struct worker *worker)
3546 {
3547 	struct worker_pool *pool = worker->pool;
3548 	int nr_restarts = BH_WORKER_RESTARTS;
3549 	unsigned long end = jiffies + BH_WORKER_JIFFIES;
3550 
3551 	raw_spin_lock_irq(&pool->lock);
3552 	worker_leave_idle(worker);
3553 
3554 	/*
3555 	 * This function follows the structure of worker_thread(). See there for
3556 	 * explanations on each step.
3557 	 */
3558 	if (!need_more_worker(pool))
3559 		goto done;
3560 
3561 	WARN_ON_ONCE(!list_empty(&worker->scheduled));
3562 	worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3563 
3564 	do {
3565 		struct work_struct *work =
3566 			list_first_entry(&pool->worklist,
3567 					 struct work_struct, entry);
3568 
3569 		if (assign_work(work, worker, NULL))
3570 			process_scheduled_works(worker);
3571 	} while (keep_working(pool) &&
3572 		 --nr_restarts && time_before(jiffies, end));
3573 
3574 	worker_set_flags(worker, WORKER_PREP);
3575 done:
3576 	worker_enter_idle(worker);
3577 	kick_pool(pool);
3578 	raw_spin_unlock_irq(&pool->lock);
3579 }
3580 
3581 /*
3582  * TODO: Convert all tasklet users to workqueue and use softirq directly.
3583  *
3584  * This is currently called from tasklet[_hi]action() and thus is also called
3585  * whenever there are tasklets to run. Let's do an early exit if there's nothing
3586  * queued. Once conversion from tasklet is complete, the need_more_worker() test
3587  * can be dropped.
3588  *
3589  * After full conversion, we'll add worker->softirq_action, directly use the
3590  * softirq action and obtain the worker pointer from the softirq_action pointer.
3591  */
3592 void workqueue_softirq_action(bool highpri)
3593 {
3594 	struct worker_pool *pool =
3595 		&per_cpu(bh_worker_pools, smp_processor_id())[highpri];
3596 	if (need_more_worker(pool))
3597 		bh_worker(list_first_entry(&pool->workers, struct worker, node));
3598 }
3599 
3600 struct wq_drain_dead_softirq_work {
3601 	struct work_struct	work;
3602 	struct worker_pool	*pool;
3603 	struct completion	done;
3604 };
3605 
3606 static void drain_dead_softirq_workfn(struct work_struct *work)
3607 {
3608 	struct wq_drain_dead_softirq_work *dead_work =
3609 		container_of(work, struct wq_drain_dead_softirq_work, work);
3610 	struct worker_pool *pool = dead_work->pool;
3611 	bool repeat;
3612 
3613 	/*
3614 	 * @pool's CPU is dead and we want to execute its still pending work
3615 	 * items from this BH work item which is running on a different CPU. As
3616 	 * its CPU is dead, @pool can't be kicked and, as work execution path
3617 	 * will be nested, a lockdep annotation needs to be suppressed. Mark
3618 	 * @pool with %POOL_BH_DRAINING for the special treatments.
3619 	 */
3620 	raw_spin_lock_irq(&pool->lock);
3621 	pool->flags |= POOL_BH_DRAINING;
3622 	raw_spin_unlock_irq(&pool->lock);
3623 
3624 	bh_worker(list_first_entry(&pool->workers, struct worker, node));
3625 
3626 	raw_spin_lock_irq(&pool->lock);
3627 	pool->flags &= ~POOL_BH_DRAINING;
3628 	repeat = need_more_worker(pool);
3629 	raw_spin_unlock_irq(&pool->lock);
3630 
3631 	/*
3632 	 * bh_worker() might hit consecutive execution limit and bail. If there
3633 	 * still are pending work items, reschedule self and return so that we
3634 	 * don't hog this CPU's BH.
3635 	 */
3636 	if (repeat) {
3637 		if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3638 			queue_work(system_bh_highpri_wq, work);
3639 		else
3640 			queue_work(system_bh_wq, work);
3641 	} else {
3642 		complete(&dead_work->done);
3643 	}
3644 }
3645 
3646 /*
3647  * @cpu is dead. Drain the remaining BH work items on the current CPU. It's
3648  * possible to allocate dead_work per CPU and avoid flushing. However, then we
3649  * have to worry about draining overlapping with CPU coming back online or
3650  * nesting (one CPU's dead_work queued on another CPU which is also dead and so
3651  * on). Let's keep it simple and drain them synchronously. These are BH work
3652  * items which shouldn't be requeued on the same pool. Shouldn't take long.
3653  */
3654 void workqueue_softirq_dead(unsigned int cpu)
3655 {
3656 	int i;
3657 
3658 	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
3659 		struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i];
3660 		struct wq_drain_dead_softirq_work dead_work;
3661 
3662 		if (!need_more_worker(pool))
3663 			continue;
3664 
3665 		INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn);
3666 		dead_work.pool = pool;
3667 		init_completion(&dead_work.done);
3668 
3669 		if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3670 			queue_work(system_bh_highpri_wq, &dead_work.work);
3671 		else
3672 			queue_work(system_bh_wq, &dead_work.work);
3673 
3674 		wait_for_completion(&dead_work.done);
3675 		destroy_work_on_stack(&dead_work.work);
3676 	}
3677 }
3678 
3679 /**
3680  * check_flush_dependency - check for flush dependency sanity
3681  * @target_wq: workqueue being flushed
3682  * @target_work: work item being flushed (NULL for workqueue flushes)
3683  *
3684  * %current is trying to flush the whole @target_wq or @target_work on it.
3685  * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3686  * reclaiming memory or running on a workqueue which doesn't have
3687  * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3688  * a deadlock.
3689  */
3690 static void check_flush_dependency(struct workqueue_struct *target_wq,
3691 				   struct work_struct *target_work)
3692 {
3693 	work_func_t target_func = target_work ? target_work->func : NULL;
3694 	struct worker *worker;
3695 
3696 	if (target_wq->flags & WQ_MEM_RECLAIM)
3697 		return;
3698 
3699 	worker = current_wq_worker();
3700 
3701 	WARN_ONCE(current->flags & PF_MEMALLOC,
3702 		  "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3703 		  current->pid, current->comm, target_wq->name, target_func);
3704 	WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3705 			      (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3706 		  "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3707 		  worker->current_pwq->wq->name, worker->current_func,
3708 		  target_wq->name, target_func);
3709 }
3710 
3711 struct wq_barrier {
3712 	struct work_struct	work;
3713 	struct completion	done;
3714 	struct task_struct	*task;	/* purely informational */
3715 };
3716 
3717 static void wq_barrier_func(struct work_struct *work)
3718 {
3719 	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3720 	complete(&barr->done);
3721 }
3722 
3723 /**
3724  * insert_wq_barrier - insert a barrier work
3725  * @pwq: pwq to insert barrier into
3726  * @barr: wq_barrier to insert
3727  * @target: target work to attach @barr to
3728  * @worker: worker currently executing @target, NULL if @target is not executing
3729  *
3730  * @barr is linked to @target such that @barr is completed only after
3731  * @target finishes execution.  Please note that the ordering
3732  * guarantee is observed only with respect to @target and on the local
3733  * cpu.
3734  *
3735  * Currently, a queued barrier can't be canceled.  This is because
3736  * try_to_grab_pending() can't determine whether the work to be
3737  * grabbed is at the head of the queue and thus can't clear LINKED
3738  * flag of the previous work while there must be a valid next work
3739  * after a work with LINKED flag set.
3740  *
3741  * Note that when @worker is non-NULL, @target may be modified
3742  * underneath us, so we can't reliably determine pwq from @target.
3743  *
3744  * CONTEXT:
3745  * raw_spin_lock_irq(pool->lock).
3746  */
3747 static void insert_wq_barrier(struct pool_workqueue *pwq,
3748 			      struct wq_barrier *barr,
3749 			      struct work_struct *target, struct worker *worker)
3750 {
3751 	static __maybe_unused struct lock_class_key bh_key, thr_key;
3752 	unsigned int work_flags = 0;
3753 	unsigned int work_color;
3754 	struct list_head *head;
3755 
3756 	/*
3757 	 * debugobject calls are safe here even with pool->lock locked
3758 	 * as we know for sure that this will not trigger any of the
3759 	 * checks and call back into the fixup functions where we
3760 	 * might deadlock.
3761 	 *
3762 	 * BH and threaded workqueues need separate lockdep keys to avoid
3763 	 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W}
3764 	 * usage".
3765 	 */
3766 	INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func,
3767 			      (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key);
3768 	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3769 
3770 	init_completion_map(&barr->done, &target->lockdep_map);
3771 
3772 	barr->task = current;
3773 
3774 	/* The barrier work item does not participate in nr_active. */
3775 	work_flags |= WORK_STRUCT_INACTIVE;
3776 
3777 	/*
3778 	 * If @target is currently being executed, schedule the
3779 	 * barrier to the worker; otherwise, put it after @target.
3780 	 */
3781 	if (worker) {
3782 		head = worker->scheduled.next;
3783 		work_color = worker->current_color;
3784 	} else {
3785 		unsigned long *bits = work_data_bits(target);
3786 
3787 		head = target->entry.next;
3788 		/* there can already be other linked works, inherit and set */
3789 		work_flags |= *bits & WORK_STRUCT_LINKED;
3790 		work_color = get_work_color(*bits);
3791 		__set_bit(WORK_STRUCT_LINKED_BIT, bits);
3792 	}
3793 
3794 	pwq->nr_in_flight[work_color]++;
3795 	work_flags |= work_color_to_flags(work_color);
3796 
3797 	insert_work(pwq, &barr->work, head, work_flags);
3798 }
3799 
3800 /**
3801  * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3802  * @wq: workqueue being flushed
3803  * @flush_color: new flush color, < 0 for no-op
3804  * @work_color: new work color, < 0 for no-op
3805  *
3806  * Prepare pwqs for workqueue flushing.
3807  *
3808  * If @flush_color is non-negative, flush_color on all pwqs should be
3809  * -1.  If no pwq has in-flight commands at the specified color, all
3810  * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
3811  * has in flight commands, its pwq->flush_color is set to
3812  * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3813  * wakeup logic is armed and %true is returned.
3814  *
3815  * The caller should have initialized @wq->first_flusher prior to
3816  * calling this function with non-negative @flush_color.  If
3817  * @flush_color is negative, no flush color update is done and %false
3818  * is returned.
3819  *
3820  * If @work_color is non-negative, all pwqs should have the same
3821  * work_color which is previous to @work_color and all will be
3822  * advanced to @work_color.
3823  *
3824  * CONTEXT:
3825  * mutex_lock(wq->mutex).
3826  *
3827  * Return:
3828  * %true if @flush_color >= 0 and there's something to flush.  %false
3829  * otherwise.
3830  */
3831 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3832 				      int flush_color, int work_color)
3833 {
3834 	bool wait = false;
3835 	struct pool_workqueue *pwq;
3836 	struct worker_pool *current_pool = NULL;
3837 
3838 	if (flush_color >= 0) {
3839 		WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3840 		atomic_set(&wq->nr_pwqs_to_flush, 1);
3841 	}
3842 
3843 	/*
3844 	 * For unbound workqueue, pwqs will map to only a few pools.
3845 	 * Most of the time, pwqs within the same pool will be linked
3846 	 * sequentially to wq->pwqs by cpu index. So in the majority
3847 	 * of pwq iters, the pool is the same, only doing lock/unlock
3848 	 * if the pool has changed. This can largely reduce expensive
3849 	 * lock operations.
3850 	 */
3851 	for_each_pwq(pwq, wq) {
3852 		if (current_pool != pwq->pool) {
3853 			if (likely(current_pool))
3854 				raw_spin_unlock_irq(&current_pool->lock);
3855 			current_pool = pwq->pool;
3856 			raw_spin_lock_irq(&current_pool->lock);
3857 		}
3858 
3859 		if (flush_color >= 0) {
3860 			WARN_ON_ONCE(pwq->flush_color != -1);
3861 
3862 			if (pwq->nr_in_flight[flush_color]) {
3863 				pwq->flush_color = flush_color;
3864 				atomic_inc(&wq->nr_pwqs_to_flush);
3865 				wait = true;
3866 			}
3867 		}
3868 
3869 		if (work_color >= 0) {
3870 			WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3871 			pwq->work_color = work_color;
3872 		}
3873 
3874 	}
3875 
3876 	if (current_pool)
3877 		raw_spin_unlock_irq(&current_pool->lock);
3878 
3879 	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3880 		complete(&wq->first_flusher->done);
3881 
3882 	return wait;
3883 }
3884 
3885 static void touch_wq_lockdep_map(struct workqueue_struct *wq)
3886 {
3887 #ifdef CONFIG_LOCKDEP
3888 	if (unlikely(!wq->lockdep_map))
3889 		return;
3890 
3891 	if (wq->flags & WQ_BH)
3892 		local_bh_disable();
3893 
3894 	lock_map_acquire(wq->lockdep_map);
3895 	lock_map_release(wq->lockdep_map);
3896 
3897 	if (wq->flags & WQ_BH)
3898 		local_bh_enable();
3899 #endif
3900 }
3901 
3902 static void touch_work_lockdep_map(struct work_struct *work,
3903 				   struct workqueue_struct *wq)
3904 {
3905 #ifdef CONFIG_LOCKDEP
3906 	if (wq->flags & WQ_BH)
3907 		local_bh_disable();
3908 
3909 	lock_map_acquire(&work->lockdep_map);
3910 	lock_map_release(&work->lockdep_map);
3911 
3912 	if (wq->flags & WQ_BH)
3913 		local_bh_enable();
3914 #endif
3915 }
3916 
3917 /**
3918  * __flush_workqueue - ensure that any scheduled work has run to completion.
3919  * @wq: workqueue to flush
3920  *
3921  * This function sleeps until all work items which were queued on entry
3922  * have finished execution, but it is not livelocked by new incoming ones.
3923  */
3924 void __flush_workqueue(struct workqueue_struct *wq)
3925 {
3926 	struct wq_flusher this_flusher = {
3927 		.list = LIST_HEAD_INIT(this_flusher.list),
3928 		.flush_color = -1,
3929 		.done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, (*wq->lockdep_map)),
3930 	};
3931 	int next_color;
3932 
3933 	if (WARN_ON(!wq_online))
3934 		return;
3935 
3936 	touch_wq_lockdep_map(wq);
3937 
3938 	mutex_lock(&wq->mutex);
3939 
3940 	/*
3941 	 * Start-to-wait phase
3942 	 */
3943 	next_color = work_next_color(wq->work_color);
3944 
3945 	if (next_color != wq->flush_color) {
3946 		/*
3947 		 * Color space is not full.  The current work_color
3948 		 * becomes our flush_color and work_color is advanced
3949 		 * by one.
3950 		 */
3951 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3952 		this_flusher.flush_color = wq->work_color;
3953 		wq->work_color = next_color;
3954 
3955 		if (!wq->first_flusher) {
3956 			/* no flush in progress, become the first flusher */
3957 			WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3958 
3959 			wq->first_flusher = &this_flusher;
3960 
3961 			if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3962 						       wq->work_color)) {
3963 				/* nothing to flush, done */
3964 				wq->flush_color = next_color;
3965 				wq->first_flusher = NULL;
3966 				goto out_unlock;
3967 			}
3968 		} else {
3969 			/* wait in queue */
3970 			WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3971 			list_add_tail(&this_flusher.list, &wq->flusher_queue);
3972 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3973 		}
3974 	} else {
3975 		/*
3976 		 * Oops, color space is full, wait on overflow queue.
3977 		 * The next flush completion will assign us
3978 		 * flush_color and transfer to flusher_queue.
3979 		 */
3980 		list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3981 	}
3982 
3983 	check_flush_dependency(wq, NULL);
3984 
3985 	mutex_unlock(&wq->mutex);
3986 
3987 	wait_for_completion(&this_flusher.done);
3988 
3989 	/*
3990 	 * Wake-up-and-cascade phase
3991 	 *
3992 	 * First flushers are responsible for cascading flushes and
3993 	 * handling overflow.  Non-first flushers can simply return.
3994 	 */
3995 	if (READ_ONCE(wq->first_flusher) != &this_flusher)
3996 		return;
3997 
3998 	mutex_lock(&wq->mutex);
3999 
4000 	/* we might have raced, check again with mutex held */
4001 	if (wq->first_flusher != &this_flusher)
4002 		goto out_unlock;
4003 
4004 	WRITE_ONCE(wq->first_flusher, NULL);
4005 
4006 	WARN_ON_ONCE(!list_empty(&this_flusher.list));
4007 	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
4008 
4009 	while (true) {
4010 		struct wq_flusher *next, *tmp;
4011 
4012 		/* complete all the flushers sharing the current flush color */
4013 		list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
4014 			if (next->flush_color != wq->flush_color)
4015 				break;
4016 			list_del_init(&next->list);
4017 			complete(&next->done);
4018 		}
4019 
4020 		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4021 			     wq->flush_color != work_next_color(wq->work_color));
4022 
4023 		/* this flush_color is finished, advance by one */
4024 		wq->flush_color = work_next_color(wq->flush_color);
4025 
4026 		/* one color has been freed, handle overflow queue */
4027 		if (!list_empty(&wq->flusher_overflow)) {
4028 			/*
4029 			 * Assign the same color to all overflowed
4030 			 * flushers, advance work_color and append to
4031 			 * flusher_queue.  This is the start-to-wait
4032 			 * phase for these overflowed flushers.
4033 			 */
4034 			list_for_each_entry(tmp, &wq->flusher_overflow, list)
4035 				tmp->flush_color = wq->work_color;
4036 
4037 			wq->work_color = work_next_color(wq->work_color);
4038 
4039 			list_splice_tail_init(&wq->flusher_overflow,
4040 					      &wq->flusher_queue);
4041 			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4042 		}
4043 
4044 		if (list_empty(&wq->flusher_queue)) {
4045 			WARN_ON_ONCE(wq->flush_color != wq->work_color);
4046 			break;
4047 		}
4048 
4049 		/*
4050 		 * Need to flush more colors.  Make the next flusher
4051 		 * the new first flusher and arm pwqs.
4052 		 */
4053 		WARN_ON_ONCE(wq->flush_color == wq->work_color);
4054 		WARN_ON_ONCE(wq->flush_color != next->flush_color);
4055 
4056 		list_del_init(&next->list);
4057 		wq->first_flusher = next;
4058 
4059 		if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4060 			break;
4061 
4062 		/*
4063 		 * Meh... this color is already done, clear first
4064 		 * flusher and repeat cascading.
4065 		 */
4066 		wq->first_flusher = NULL;
4067 	}
4068 
4069 out_unlock:
4070 	mutex_unlock(&wq->mutex);
4071 }
4072 EXPORT_SYMBOL(__flush_workqueue);
4073 
4074 /**
4075  * drain_workqueue - drain a workqueue
4076  * @wq: workqueue to drain
4077  *
4078  * Wait until the workqueue becomes empty.  While draining is in progress,
4079  * only chain queueing is allowed.  IOW, only currently pending or running
4080  * work items on @wq can queue further work items on it.  @wq is flushed
4081  * repeatedly until it becomes empty.  The number of flushing is determined
4082  * by the depth of chaining and should be relatively short.  Whine if it
4083  * takes too long.
4084  */
4085 void drain_workqueue(struct workqueue_struct *wq)
4086 {
4087 	unsigned int flush_cnt = 0;
4088 	struct pool_workqueue *pwq;
4089 
4090 	/*
4091 	 * __queue_work() needs to test whether there are drainers, is much
4092 	 * hotter than drain_workqueue() and already looks at @wq->flags.
4093 	 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4094 	 */
4095 	mutex_lock(&wq->mutex);
4096 	if (!wq->nr_drainers++)
4097 		wq->flags |= __WQ_DRAINING;
4098 	mutex_unlock(&wq->mutex);
4099 reflush:
4100 	__flush_workqueue(wq);
4101 
4102 	mutex_lock(&wq->mutex);
4103 
4104 	for_each_pwq(pwq, wq) {
4105 		bool drained;
4106 
4107 		raw_spin_lock_irq(&pwq->pool->lock);
4108 		drained = pwq_is_empty(pwq);
4109 		raw_spin_unlock_irq(&pwq->pool->lock);
4110 
4111 		if (drained)
4112 			continue;
4113 
4114 		if (++flush_cnt == 10 ||
4115 		    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4116 			pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
4117 				wq->name, __func__, flush_cnt);
4118 
4119 		mutex_unlock(&wq->mutex);
4120 		goto reflush;
4121 	}
4122 
4123 	if (!--wq->nr_drainers)
4124 		wq->flags &= ~__WQ_DRAINING;
4125 	mutex_unlock(&wq->mutex);
4126 }
4127 EXPORT_SYMBOL_GPL(drain_workqueue);
4128 
4129 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
4130 			     bool from_cancel)
4131 {
4132 	struct worker *worker = NULL;
4133 	struct worker_pool *pool;
4134 	struct pool_workqueue *pwq;
4135 	struct workqueue_struct *wq;
4136 
4137 	rcu_read_lock();
4138 	pool = get_work_pool(work);
4139 	if (!pool) {
4140 		rcu_read_unlock();
4141 		return false;
4142 	}
4143 
4144 	raw_spin_lock_irq(&pool->lock);
4145 	/* see the comment in try_to_grab_pending() with the same code */
4146 	pwq = get_work_pwq(work);
4147 	if (pwq) {
4148 		if (unlikely(pwq->pool != pool))
4149 			goto already_gone;
4150 	} else {
4151 		worker = find_worker_executing_work(pool, work);
4152 		if (!worker)
4153 			goto already_gone;
4154 		pwq = worker->current_pwq;
4155 	}
4156 
4157 	wq = pwq->wq;
4158 	check_flush_dependency(wq, work);
4159 
4160 	insert_wq_barrier(pwq, barr, work, worker);
4161 	raw_spin_unlock_irq(&pool->lock);
4162 
4163 	touch_work_lockdep_map(work, wq);
4164 
4165 	/*
4166 	 * Force a lock recursion deadlock when using flush_work() inside a
4167 	 * single-threaded or rescuer equipped workqueue.
4168 	 *
4169 	 * For single threaded workqueues the deadlock happens when the work
4170 	 * is after the work issuing the flush_work(). For rescuer equipped
4171 	 * workqueues the deadlock happens when the rescuer stalls, blocking
4172 	 * forward progress.
4173 	 */
4174 	if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer))
4175 		touch_wq_lockdep_map(wq);
4176 
4177 	rcu_read_unlock();
4178 	return true;
4179 already_gone:
4180 	raw_spin_unlock_irq(&pool->lock);
4181 	rcu_read_unlock();
4182 	return false;
4183 }
4184 
4185 static bool __flush_work(struct work_struct *work, bool from_cancel)
4186 {
4187 	struct wq_barrier barr;
4188 
4189 	if (WARN_ON(!wq_online))
4190 		return false;
4191 
4192 	if (WARN_ON(!work->func))
4193 		return false;
4194 
4195 	if (!start_flush_work(work, &barr, from_cancel))
4196 		return false;
4197 
4198 	/*
4199 	 * start_flush_work() returned %true. If @from_cancel is set, we know
4200 	 * that @work must have been executing during start_flush_work() and
4201 	 * can't currently be queued. Its data must contain OFFQ bits. If @work
4202 	 * was queued on a BH workqueue, we also know that it was running in the
4203 	 * BH context and thus can be busy-waited.
4204 	 */
4205 	if (from_cancel) {
4206 		unsigned long data = *work_data_bits(work);
4207 
4208 		if (!WARN_ON_ONCE(data & WORK_STRUCT_PWQ) &&
4209 		    (data & WORK_OFFQ_BH)) {
4210 			/*
4211 			 * On RT, prevent a live lock when %current preempted
4212 			 * soft interrupt processing or prevents ksoftirqd from
4213 			 * running by keeping flipping BH. If the BH work item
4214 			 * runs on a different CPU then this has no effect other
4215 			 * than doing the BH disable/enable dance for nothing.
4216 			 * This is copied from
4217 			 * kernel/softirq.c::tasklet_unlock_spin_wait().
4218 			 */
4219 			while (!try_wait_for_completion(&barr.done)) {
4220 				if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
4221 					local_bh_disable();
4222 					local_bh_enable();
4223 				} else {
4224 					cpu_relax();
4225 				}
4226 			}
4227 			goto out_destroy;
4228 		}
4229 	}
4230 
4231 	wait_for_completion(&barr.done);
4232 
4233 out_destroy:
4234 	destroy_work_on_stack(&barr.work);
4235 	return true;
4236 }
4237 
4238 /**
4239  * flush_work - wait for a work to finish executing the last queueing instance
4240  * @work: the work to flush
4241  *
4242  * Wait until @work has finished execution.  @work is guaranteed to be idle
4243  * on return if it hasn't been requeued since flush started.
4244  *
4245  * Return:
4246  * %true if flush_work() waited for the work to finish execution,
4247  * %false if it was already idle.
4248  */
4249 bool flush_work(struct work_struct *work)
4250 {
4251 	might_sleep();
4252 	return __flush_work(work, false);
4253 }
4254 EXPORT_SYMBOL_GPL(flush_work);
4255 
4256 /**
4257  * flush_delayed_work - wait for a dwork to finish executing the last queueing
4258  * @dwork: the delayed work to flush
4259  *
4260  * Delayed timer is cancelled and the pending work is queued for
4261  * immediate execution.  Like flush_work(), this function only
4262  * considers the last queueing instance of @dwork.
4263  *
4264  * Return:
4265  * %true if flush_work() waited for the work to finish execution,
4266  * %false if it was already idle.
4267  */
4268 bool flush_delayed_work(struct delayed_work *dwork)
4269 {
4270 	local_irq_disable();
4271 	if (del_timer_sync(&dwork->timer))
4272 		__queue_work(dwork->cpu, dwork->wq, &dwork->work);
4273 	local_irq_enable();
4274 	return flush_work(&dwork->work);
4275 }
4276 EXPORT_SYMBOL(flush_delayed_work);
4277 
4278 /**
4279  * flush_rcu_work - wait for a rwork to finish executing the last queueing
4280  * @rwork: the rcu work to flush
4281  *
4282  * Return:
4283  * %true if flush_rcu_work() waited for the work to finish execution,
4284  * %false if it was already idle.
4285  */
4286 bool flush_rcu_work(struct rcu_work *rwork)
4287 {
4288 	if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4289 		rcu_barrier();
4290 		flush_work(&rwork->work);
4291 		return true;
4292 	} else {
4293 		return flush_work(&rwork->work);
4294 	}
4295 }
4296 EXPORT_SYMBOL(flush_rcu_work);
4297 
4298 static void work_offqd_disable(struct work_offq_data *offqd)
4299 {
4300 	const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1;
4301 
4302 	if (likely(offqd->disable < max))
4303 		offqd->disable++;
4304 	else
4305 		WARN_ONCE(true, "workqueue: work disable count overflowed\n");
4306 }
4307 
4308 static void work_offqd_enable(struct work_offq_data *offqd)
4309 {
4310 	if (likely(offqd->disable > 0))
4311 		offqd->disable--;
4312 	else
4313 		WARN_ONCE(true, "workqueue: work disable count underflowed\n");
4314 }
4315 
4316 static bool __cancel_work(struct work_struct *work, u32 cflags)
4317 {
4318 	struct work_offq_data offqd;
4319 	unsigned long irq_flags;
4320 	int ret;
4321 
4322 	ret = work_grab_pending(work, cflags, &irq_flags);
4323 
4324 	work_offqd_unpack(&offqd, *work_data_bits(work));
4325 
4326 	if (cflags & WORK_CANCEL_DISABLE)
4327 		work_offqd_disable(&offqd);
4328 
4329 	set_work_pool_and_clear_pending(work, offqd.pool_id,
4330 					work_offqd_pack_flags(&offqd));
4331 	local_irq_restore(irq_flags);
4332 	return ret;
4333 }
4334 
4335 static bool __cancel_work_sync(struct work_struct *work, u32 cflags)
4336 {
4337 	bool ret;
4338 
4339 	ret = __cancel_work(work, cflags | WORK_CANCEL_DISABLE);
4340 
4341 	if (*work_data_bits(work) & WORK_OFFQ_BH)
4342 		WARN_ON_ONCE(in_hardirq());
4343 	else
4344 		might_sleep();
4345 
4346 	/*
4347 	 * Skip __flush_work() during early boot when we know that @work isn't
4348 	 * executing. This allows canceling during early boot.
4349 	 */
4350 	if (wq_online)
4351 		__flush_work(work, true);
4352 
4353 	if (!(cflags & WORK_CANCEL_DISABLE))
4354 		enable_work(work);
4355 
4356 	return ret;
4357 }
4358 
4359 /*
4360  * See cancel_delayed_work()
4361  */
4362 bool cancel_work(struct work_struct *work)
4363 {
4364 	return __cancel_work(work, 0);
4365 }
4366 EXPORT_SYMBOL(cancel_work);
4367 
4368 /**
4369  * cancel_work_sync - cancel a work and wait for it to finish
4370  * @work: the work to cancel
4371  *
4372  * Cancel @work and wait for its execution to finish. This function can be used
4373  * even if the work re-queues itself or migrates to another workqueue. On return
4374  * from this function, @work is guaranteed to be not pending or executing on any
4375  * CPU as long as there aren't racing enqueues.
4376  *
4377  * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's.
4378  * Use cancel_delayed_work_sync() instead.
4379  *
4380  * Must be called from a sleepable context if @work was last queued on a non-BH
4381  * workqueue. Can also be called from non-hardirq atomic contexts including BH
4382  * if @work was last queued on a BH workqueue.
4383  *
4384  * Returns %true if @work was pending, %false otherwise.
4385  */
4386 bool cancel_work_sync(struct work_struct *work)
4387 {
4388 	return __cancel_work_sync(work, 0);
4389 }
4390 EXPORT_SYMBOL_GPL(cancel_work_sync);
4391 
4392 /**
4393  * cancel_delayed_work - cancel a delayed work
4394  * @dwork: delayed_work to cancel
4395  *
4396  * Kill off a pending delayed_work.
4397  *
4398  * Return: %true if @dwork was pending and canceled; %false if it wasn't
4399  * pending.
4400  *
4401  * Note:
4402  * The work callback function may still be running on return, unless
4403  * it returns %true and the work doesn't re-arm itself.  Explicitly flush or
4404  * use cancel_delayed_work_sync() to wait on it.
4405  *
4406  * This function is safe to call from any context including IRQ handler.
4407  */
4408 bool cancel_delayed_work(struct delayed_work *dwork)
4409 {
4410 	return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED);
4411 }
4412 EXPORT_SYMBOL(cancel_delayed_work);
4413 
4414 /**
4415  * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4416  * @dwork: the delayed work cancel
4417  *
4418  * This is cancel_work_sync() for delayed works.
4419  *
4420  * Return:
4421  * %true if @dwork was pending, %false otherwise.
4422  */
4423 bool cancel_delayed_work_sync(struct delayed_work *dwork)
4424 {
4425 	return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED);
4426 }
4427 EXPORT_SYMBOL(cancel_delayed_work_sync);
4428 
4429 /**
4430  * disable_work - Disable and cancel a work item
4431  * @work: work item to disable
4432  *
4433  * Disable @work by incrementing its disable count and cancel it if currently
4434  * pending. As long as the disable count is non-zero, any attempt to queue @work
4435  * will fail and return %false. The maximum supported disable depth is 2 to the
4436  * power of %WORK_OFFQ_DISABLE_BITS, currently 65536.
4437  *
4438  * Can be called from any context. Returns %true if @work was pending, %false
4439  * otherwise.
4440  */
4441 bool disable_work(struct work_struct *work)
4442 {
4443 	return __cancel_work(work, WORK_CANCEL_DISABLE);
4444 }
4445 EXPORT_SYMBOL_GPL(disable_work);
4446 
4447 /**
4448  * disable_work_sync - Disable, cancel and drain a work item
4449  * @work: work item to disable
4450  *
4451  * Similar to disable_work() but also wait for @work to finish if currently
4452  * executing.
4453  *
4454  * Must be called from a sleepable context if @work was last queued on a non-BH
4455  * workqueue. Can also be called from non-hardirq atomic contexts including BH
4456  * if @work was last queued on a BH workqueue.
4457  *
4458  * Returns %true if @work was pending, %false otherwise.
4459  */
4460 bool disable_work_sync(struct work_struct *work)
4461 {
4462 	return __cancel_work_sync(work, WORK_CANCEL_DISABLE);
4463 }
4464 EXPORT_SYMBOL_GPL(disable_work_sync);
4465 
4466 /**
4467  * enable_work - Enable a work item
4468  * @work: work item to enable
4469  *
4470  * Undo disable_work[_sync]() by decrementing @work's disable count. @work can
4471  * only be queued if its disable count is 0.
4472  *
4473  * Can be called from any context. Returns %true if the disable count reached 0.
4474  * Otherwise, %false.
4475  */
4476 bool enable_work(struct work_struct *work)
4477 {
4478 	struct work_offq_data offqd;
4479 	unsigned long irq_flags;
4480 
4481 	work_grab_pending(work, 0, &irq_flags);
4482 
4483 	work_offqd_unpack(&offqd, *work_data_bits(work));
4484 	work_offqd_enable(&offqd);
4485 	set_work_pool_and_clear_pending(work, offqd.pool_id,
4486 					work_offqd_pack_flags(&offqd));
4487 	local_irq_restore(irq_flags);
4488 
4489 	return !offqd.disable;
4490 }
4491 EXPORT_SYMBOL_GPL(enable_work);
4492 
4493 /**
4494  * disable_delayed_work - Disable and cancel a delayed work item
4495  * @dwork: delayed work item to disable
4496  *
4497  * disable_work() for delayed work items.
4498  */
4499 bool disable_delayed_work(struct delayed_work *dwork)
4500 {
4501 	return __cancel_work(&dwork->work,
4502 			     WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4503 }
4504 EXPORT_SYMBOL_GPL(disable_delayed_work);
4505 
4506 /**
4507  * disable_delayed_work_sync - Disable, cancel and drain a delayed work item
4508  * @dwork: delayed work item to disable
4509  *
4510  * disable_work_sync() for delayed work items.
4511  */
4512 bool disable_delayed_work_sync(struct delayed_work *dwork)
4513 {
4514 	return __cancel_work_sync(&dwork->work,
4515 				  WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4516 }
4517 EXPORT_SYMBOL_GPL(disable_delayed_work_sync);
4518 
4519 /**
4520  * enable_delayed_work - Enable a delayed work item
4521  * @dwork: delayed work item to enable
4522  *
4523  * enable_work() for delayed work items.
4524  */
4525 bool enable_delayed_work(struct delayed_work *dwork)
4526 {
4527 	return enable_work(&dwork->work);
4528 }
4529 EXPORT_SYMBOL_GPL(enable_delayed_work);
4530 
4531 /**
4532  * schedule_on_each_cpu - execute a function synchronously on each online CPU
4533  * @func: the function to call
4534  *
4535  * schedule_on_each_cpu() executes @func on each online CPU using the
4536  * system workqueue and blocks until all CPUs have completed.
4537  * schedule_on_each_cpu() is very slow.
4538  *
4539  * Return:
4540  * 0 on success, -errno on failure.
4541  */
4542 int schedule_on_each_cpu(work_func_t func)
4543 {
4544 	int cpu;
4545 	struct work_struct __percpu *works;
4546 
4547 	works = alloc_percpu(struct work_struct);
4548 	if (!works)
4549 		return -ENOMEM;
4550 
4551 	cpus_read_lock();
4552 
4553 	for_each_online_cpu(cpu) {
4554 		struct work_struct *work = per_cpu_ptr(works, cpu);
4555 
4556 		INIT_WORK(work, func);
4557 		schedule_work_on(cpu, work);
4558 	}
4559 
4560 	for_each_online_cpu(cpu)
4561 		flush_work(per_cpu_ptr(works, cpu));
4562 
4563 	cpus_read_unlock();
4564 	free_percpu(works);
4565 	return 0;
4566 }
4567 
4568 /**
4569  * execute_in_process_context - reliably execute the routine with user context
4570  * @fn:		the function to execute
4571  * @ew:		guaranteed storage for the execute work structure (must
4572  *		be available when the work executes)
4573  *
4574  * Executes the function immediately if process context is available,
4575  * otherwise schedules the function for delayed execution.
4576  *
4577  * Return:	0 - function was executed
4578  *		1 - function was scheduled for execution
4579  */
4580 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4581 {
4582 	if (!in_interrupt()) {
4583 		fn(&ew->work);
4584 		return 0;
4585 	}
4586 
4587 	INIT_WORK(&ew->work, fn);
4588 	schedule_work(&ew->work);
4589 
4590 	return 1;
4591 }
4592 EXPORT_SYMBOL_GPL(execute_in_process_context);
4593 
4594 /**
4595  * free_workqueue_attrs - free a workqueue_attrs
4596  * @attrs: workqueue_attrs to free
4597  *
4598  * Undo alloc_workqueue_attrs().
4599  */
4600 void free_workqueue_attrs(struct workqueue_attrs *attrs)
4601 {
4602 	if (attrs) {
4603 		free_cpumask_var(attrs->cpumask);
4604 		free_cpumask_var(attrs->__pod_cpumask);
4605 		kfree(attrs);
4606 	}
4607 }
4608 
4609 /**
4610  * alloc_workqueue_attrs - allocate a workqueue_attrs
4611  *
4612  * Allocate a new workqueue_attrs, initialize with default settings and
4613  * return it.
4614  *
4615  * Return: The allocated new workqueue_attr on success. %NULL on failure.
4616  */
4617 struct workqueue_attrs *alloc_workqueue_attrs(void)
4618 {
4619 	struct workqueue_attrs *attrs;
4620 
4621 	attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4622 	if (!attrs)
4623 		goto fail;
4624 	if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4625 		goto fail;
4626 	if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4627 		goto fail;
4628 
4629 	cpumask_copy(attrs->cpumask, cpu_possible_mask);
4630 	attrs->affn_scope = WQ_AFFN_DFL;
4631 	return attrs;
4632 fail:
4633 	free_workqueue_attrs(attrs);
4634 	return NULL;
4635 }
4636 
4637 static void copy_workqueue_attrs(struct workqueue_attrs *to,
4638 				 const struct workqueue_attrs *from)
4639 {
4640 	to->nice = from->nice;
4641 	cpumask_copy(to->cpumask, from->cpumask);
4642 	cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4643 	to->affn_strict = from->affn_strict;
4644 
4645 	/*
4646 	 * Unlike hash and equality test, copying shouldn't ignore wq-only
4647 	 * fields as copying is used for both pool and wq attrs. Instead,
4648 	 * get_unbound_pool() explicitly clears the fields.
4649 	 */
4650 	to->affn_scope = from->affn_scope;
4651 	to->ordered = from->ordered;
4652 }
4653 
4654 /*
4655  * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4656  * comments in 'struct workqueue_attrs' definition.
4657  */
4658 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4659 {
4660 	attrs->affn_scope = WQ_AFFN_NR_TYPES;
4661 	attrs->ordered = false;
4662 	if (attrs->affn_strict)
4663 		cpumask_copy(attrs->cpumask, cpu_possible_mask);
4664 }
4665 
4666 /* hash value of the content of @attr */
4667 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4668 {
4669 	u32 hash = 0;
4670 
4671 	hash = jhash_1word(attrs->nice, hash);
4672 	hash = jhash_1word(attrs->affn_strict, hash);
4673 	hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4674 		     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4675 	if (!attrs->affn_strict)
4676 		hash = jhash(cpumask_bits(attrs->cpumask),
4677 			     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4678 	return hash;
4679 }
4680 
4681 /* content equality test */
4682 static bool wqattrs_equal(const struct workqueue_attrs *a,
4683 			  const struct workqueue_attrs *b)
4684 {
4685 	if (a->nice != b->nice)
4686 		return false;
4687 	if (a->affn_strict != b->affn_strict)
4688 		return false;
4689 	if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4690 		return false;
4691 	if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask))
4692 		return false;
4693 	return true;
4694 }
4695 
4696 /* Update @attrs with actually available CPUs */
4697 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4698 				      const cpumask_t *unbound_cpumask)
4699 {
4700 	/*
4701 	 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4702 	 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4703 	 * @unbound_cpumask.
4704 	 */
4705 	cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4706 	if (unlikely(cpumask_empty(attrs->cpumask)))
4707 		cpumask_copy(attrs->cpumask, unbound_cpumask);
4708 }
4709 
4710 /* find wq_pod_type to use for @attrs */
4711 static const struct wq_pod_type *
4712 wqattrs_pod_type(const struct workqueue_attrs *attrs)
4713 {
4714 	enum wq_affn_scope scope;
4715 	struct wq_pod_type *pt;
4716 
4717 	/* to synchronize access to wq_affn_dfl */
4718 	lockdep_assert_held(&wq_pool_mutex);
4719 
4720 	if (attrs->affn_scope == WQ_AFFN_DFL)
4721 		scope = wq_affn_dfl;
4722 	else
4723 		scope = attrs->affn_scope;
4724 
4725 	pt = &wq_pod_types[scope];
4726 
4727 	if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4728 	    likely(pt->nr_pods))
4729 		return pt;
4730 
4731 	/*
4732 	 * Before workqueue_init_topology(), only SYSTEM is available which is
4733 	 * initialized in workqueue_init_early().
4734 	 */
4735 	pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4736 	BUG_ON(!pt->nr_pods);
4737 	return pt;
4738 }
4739 
4740 /**
4741  * init_worker_pool - initialize a newly zalloc'd worker_pool
4742  * @pool: worker_pool to initialize
4743  *
4744  * Initialize a newly zalloc'd @pool.  It also allocates @pool->attrs.
4745  *
4746  * Return: 0 on success, -errno on failure.  Even on failure, all fields
4747  * inside @pool proper are initialized and put_unbound_pool() can be called
4748  * on @pool safely to release it.
4749  */
4750 static int init_worker_pool(struct worker_pool *pool)
4751 {
4752 	raw_spin_lock_init(&pool->lock);
4753 	pool->id = -1;
4754 	pool->cpu = -1;
4755 	pool->node = NUMA_NO_NODE;
4756 	pool->flags |= POOL_DISASSOCIATED;
4757 	pool->watchdog_ts = jiffies;
4758 	INIT_LIST_HEAD(&pool->worklist);
4759 	INIT_LIST_HEAD(&pool->idle_list);
4760 	hash_init(pool->busy_hash);
4761 
4762 	timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4763 	INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4764 
4765 	timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4766 
4767 	INIT_LIST_HEAD(&pool->workers);
4768 
4769 	ida_init(&pool->worker_ida);
4770 	INIT_HLIST_NODE(&pool->hash_node);
4771 	pool->refcnt = 1;
4772 
4773 	/* shouldn't fail above this point */
4774 	pool->attrs = alloc_workqueue_attrs();
4775 	if (!pool->attrs)
4776 		return -ENOMEM;
4777 
4778 	wqattrs_clear_for_pool(pool->attrs);
4779 
4780 	return 0;
4781 }
4782 
4783 #ifdef CONFIG_LOCKDEP
4784 static void wq_init_lockdep(struct workqueue_struct *wq)
4785 {
4786 	char *lock_name;
4787 
4788 	lockdep_register_key(&wq->key);
4789 	lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4790 	if (!lock_name)
4791 		lock_name = wq->name;
4792 
4793 	wq->lock_name = lock_name;
4794 	wq->lockdep_map = &wq->__lockdep_map;
4795 	lockdep_init_map(wq->lockdep_map, lock_name, &wq->key, 0);
4796 }
4797 
4798 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4799 {
4800 	if (wq->lockdep_map != &wq->__lockdep_map)
4801 		return;
4802 
4803 	lockdep_unregister_key(&wq->key);
4804 }
4805 
4806 static void wq_free_lockdep(struct workqueue_struct *wq)
4807 {
4808 	if (wq->lockdep_map != &wq->__lockdep_map)
4809 		return;
4810 
4811 	if (wq->lock_name != wq->name)
4812 		kfree(wq->lock_name);
4813 }
4814 #else
4815 static void wq_init_lockdep(struct workqueue_struct *wq)
4816 {
4817 }
4818 
4819 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4820 {
4821 }
4822 
4823 static void wq_free_lockdep(struct workqueue_struct *wq)
4824 {
4825 }
4826 #endif
4827 
4828 static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
4829 {
4830 	int node;
4831 
4832 	for_each_node(node) {
4833 		kfree(nna_ar[node]);
4834 		nna_ar[node] = NULL;
4835 	}
4836 
4837 	kfree(nna_ar[nr_node_ids]);
4838 	nna_ar[nr_node_ids] = NULL;
4839 }
4840 
4841 static void init_node_nr_active(struct wq_node_nr_active *nna)
4842 {
4843 	nna->max = WQ_DFL_MIN_ACTIVE;
4844 	atomic_set(&nna->nr, 0);
4845 	raw_spin_lock_init(&nna->lock);
4846 	INIT_LIST_HEAD(&nna->pending_pwqs);
4847 }
4848 
4849 /*
4850  * Each node's nr_active counter will be accessed mostly from its own node and
4851  * should be allocated in the node.
4852  */
4853 static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
4854 {
4855 	struct wq_node_nr_active *nna;
4856 	int node;
4857 
4858 	for_each_node(node) {
4859 		nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
4860 		if (!nna)
4861 			goto err_free;
4862 		init_node_nr_active(nna);
4863 		nna_ar[node] = nna;
4864 	}
4865 
4866 	/* [nr_node_ids] is used as the fallback */
4867 	nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
4868 	if (!nna)
4869 		goto err_free;
4870 	init_node_nr_active(nna);
4871 	nna_ar[nr_node_ids] = nna;
4872 
4873 	return 0;
4874 
4875 err_free:
4876 	free_node_nr_active(nna_ar);
4877 	return -ENOMEM;
4878 }
4879 
4880 static void rcu_free_wq(struct rcu_head *rcu)
4881 {
4882 	struct workqueue_struct *wq =
4883 		container_of(rcu, struct workqueue_struct, rcu);
4884 
4885 	if (wq->flags & WQ_UNBOUND)
4886 		free_node_nr_active(wq->node_nr_active);
4887 
4888 	wq_free_lockdep(wq);
4889 	free_percpu(wq->cpu_pwq);
4890 	free_workqueue_attrs(wq->unbound_attrs);
4891 	kfree(wq);
4892 }
4893 
4894 static void rcu_free_pool(struct rcu_head *rcu)
4895 {
4896 	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4897 
4898 	ida_destroy(&pool->worker_ida);
4899 	free_workqueue_attrs(pool->attrs);
4900 	kfree(pool);
4901 }
4902 
4903 /**
4904  * put_unbound_pool - put a worker_pool
4905  * @pool: worker_pool to put
4906  *
4907  * Put @pool.  If its refcnt reaches zero, it gets destroyed in RCU
4908  * safe manner.  get_unbound_pool() calls this function on its failure path
4909  * and this function should be able to release pools which went through,
4910  * successfully or not, init_worker_pool().
4911  *
4912  * Should be called with wq_pool_mutex held.
4913  */
4914 static void put_unbound_pool(struct worker_pool *pool)
4915 {
4916 	struct worker *worker;
4917 	LIST_HEAD(cull_list);
4918 
4919 	lockdep_assert_held(&wq_pool_mutex);
4920 
4921 	if (--pool->refcnt)
4922 		return;
4923 
4924 	/* sanity checks */
4925 	if (WARN_ON(!(pool->cpu < 0)) ||
4926 	    WARN_ON(!list_empty(&pool->worklist)))
4927 		return;
4928 
4929 	/* release id and unhash */
4930 	if (pool->id >= 0)
4931 		idr_remove(&worker_pool_idr, pool->id);
4932 	hash_del(&pool->hash_node);
4933 
4934 	/*
4935 	 * Become the manager and destroy all workers.  This prevents
4936 	 * @pool's workers from blocking on attach_mutex.  We're the last
4937 	 * manager and @pool gets freed with the flag set.
4938 	 *
4939 	 * Having a concurrent manager is quite unlikely to happen as we can
4940 	 * only get here with
4941 	 *   pwq->refcnt == pool->refcnt == 0
4942 	 * which implies no work queued to the pool, which implies no worker can
4943 	 * become the manager. However a worker could have taken the role of
4944 	 * manager before the refcnts dropped to 0, since maybe_create_worker()
4945 	 * drops pool->lock
4946 	 */
4947 	while (true) {
4948 		rcuwait_wait_event(&manager_wait,
4949 				   !(pool->flags & POOL_MANAGER_ACTIVE),
4950 				   TASK_UNINTERRUPTIBLE);
4951 
4952 		mutex_lock(&wq_pool_attach_mutex);
4953 		raw_spin_lock_irq(&pool->lock);
4954 		if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4955 			pool->flags |= POOL_MANAGER_ACTIVE;
4956 			break;
4957 		}
4958 		raw_spin_unlock_irq(&pool->lock);
4959 		mutex_unlock(&wq_pool_attach_mutex);
4960 	}
4961 
4962 	while ((worker = first_idle_worker(pool)))
4963 		set_worker_dying(worker, &cull_list);
4964 	WARN_ON(pool->nr_workers || pool->nr_idle);
4965 	raw_spin_unlock_irq(&pool->lock);
4966 
4967 	detach_dying_workers(&cull_list);
4968 
4969 	mutex_unlock(&wq_pool_attach_mutex);
4970 
4971 	reap_dying_workers(&cull_list);
4972 
4973 	/* shut down the timers */
4974 	del_timer_sync(&pool->idle_timer);
4975 	cancel_work_sync(&pool->idle_cull_work);
4976 	del_timer_sync(&pool->mayday_timer);
4977 
4978 	/* RCU protected to allow dereferences from get_work_pool() */
4979 	call_rcu(&pool->rcu, rcu_free_pool);
4980 }
4981 
4982 /**
4983  * get_unbound_pool - get a worker_pool with the specified attributes
4984  * @attrs: the attributes of the worker_pool to get
4985  *
4986  * Obtain a worker_pool which has the same attributes as @attrs, bump the
4987  * reference count and return it.  If there already is a matching
4988  * worker_pool, it will be used; otherwise, this function attempts to
4989  * create a new one.
4990  *
4991  * Should be called with wq_pool_mutex held.
4992  *
4993  * Return: On success, a worker_pool with the same attributes as @attrs.
4994  * On failure, %NULL.
4995  */
4996 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4997 {
4998 	struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4999 	u32 hash = wqattrs_hash(attrs);
5000 	struct worker_pool *pool;
5001 	int pod, node = NUMA_NO_NODE;
5002 
5003 	lockdep_assert_held(&wq_pool_mutex);
5004 
5005 	/* do we already have a matching pool? */
5006 	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
5007 		if (wqattrs_equal(pool->attrs, attrs)) {
5008 			pool->refcnt++;
5009 			return pool;
5010 		}
5011 	}
5012 
5013 	/* If __pod_cpumask is contained inside a NUMA pod, that's our node */
5014 	for (pod = 0; pod < pt->nr_pods; pod++) {
5015 		if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
5016 			node = pt->pod_node[pod];
5017 			break;
5018 		}
5019 	}
5020 
5021 	/* nope, create a new one */
5022 	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
5023 	if (!pool || init_worker_pool(pool) < 0)
5024 		goto fail;
5025 
5026 	pool->node = node;
5027 	copy_workqueue_attrs(pool->attrs, attrs);
5028 	wqattrs_clear_for_pool(pool->attrs);
5029 
5030 	if (worker_pool_assign_id(pool) < 0)
5031 		goto fail;
5032 
5033 	/* create and start the initial worker */
5034 	if (wq_online && !create_worker(pool))
5035 		goto fail;
5036 
5037 	/* install */
5038 	hash_add(unbound_pool_hash, &pool->hash_node, hash);
5039 
5040 	return pool;
5041 fail:
5042 	if (pool)
5043 		put_unbound_pool(pool);
5044 	return NULL;
5045 }
5046 
5047 /*
5048  * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
5049  * refcnt and needs to be destroyed.
5050  */
5051 static void pwq_release_workfn(struct kthread_work *work)
5052 {
5053 	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
5054 						  release_work);
5055 	struct workqueue_struct *wq = pwq->wq;
5056 	struct worker_pool *pool = pwq->pool;
5057 	bool is_last = false;
5058 
5059 	/*
5060 	 * When @pwq is not linked, it doesn't hold any reference to the
5061 	 * @wq, and @wq is invalid to access.
5062 	 */
5063 	if (!list_empty(&pwq->pwqs_node)) {
5064 		mutex_lock(&wq->mutex);
5065 		list_del_rcu(&pwq->pwqs_node);
5066 		is_last = list_empty(&wq->pwqs);
5067 
5068 		/*
5069 		 * For ordered workqueue with a plugged dfl_pwq, restart it now.
5070 		 */
5071 		if (!is_last && (wq->flags & __WQ_ORDERED))
5072 			unplug_oldest_pwq(wq);
5073 
5074 		mutex_unlock(&wq->mutex);
5075 	}
5076 
5077 	if (wq->flags & WQ_UNBOUND) {
5078 		mutex_lock(&wq_pool_mutex);
5079 		put_unbound_pool(pool);
5080 		mutex_unlock(&wq_pool_mutex);
5081 	}
5082 
5083 	if (!list_empty(&pwq->pending_node)) {
5084 		struct wq_node_nr_active *nna =
5085 			wq_node_nr_active(pwq->wq, pwq->pool->node);
5086 
5087 		raw_spin_lock_irq(&nna->lock);
5088 		list_del_init(&pwq->pending_node);
5089 		raw_spin_unlock_irq(&nna->lock);
5090 	}
5091 
5092 	kfree_rcu(pwq, rcu);
5093 
5094 	/*
5095 	 * If we're the last pwq going away, @wq is already dead and no one
5096 	 * is gonna access it anymore.  Schedule RCU free.
5097 	 */
5098 	if (is_last) {
5099 		wq_unregister_lockdep(wq);
5100 		call_rcu(&wq->rcu, rcu_free_wq);
5101 	}
5102 }
5103 
5104 /* initialize newly allocated @pwq which is associated with @wq and @pool */
5105 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
5106 		     struct worker_pool *pool)
5107 {
5108 	BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK);
5109 
5110 	memset(pwq, 0, sizeof(*pwq));
5111 
5112 	pwq->pool = pool;
5113 	pwq->wq = wq;
5114 	pwq->flush_color = -1;
5115 	pwq->refcnt = 1;
5116 	INIT_LIST_HEAD(&pwq->inactive_works);
5117 	INIT_LIST_HEAD(&pwq->pending_node);
5118 	INIT_LIST_HEAD(&pwq->pwqs_node);
5119 	INIT_LIST_HEAD(&pwq->mayday_node);
5120 	kthread_init_work(&pwq->release_work, pwq_release_workfn);
5121 }
5122 
5123 /* sync @pwq with the current state of its associated wq and link it */
5124 static void link_pwq(struct pool_workqueue *pwq)
5125 {
5126 	struct workqueue_struct *wq = pwq->wq;
5127 
5128 	lockdep_assert_held(&wq->mutex);
5129 
5130 	/* may be called multiple times, ignore if already linked */
5131 	if (!list_empty(&pwq->pwqs_node))
5132 		return;
5133 
5134 	/* set the matching work_color */
5135 	pwq->work_color = wq->work_color;
5136 
5137 	/* link in @pwq */
5138 	list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs);
5139 }
5140 
5141 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5142 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
5143 					const struct workqueue_attrs *attrs)
5144 {
5145 	struct worker_pool *pool;
5146 	struct pool_workqueue *pwq;
5147 
5148 	lockdep_assert_held(&wq_pool_mutex);
5149 
5150 	pool = get_unbound_pool(attrs);
5151 	if (!pool)
5152 		return NULL;
5153 
5154 	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5155 	if (!pwq) {
5156 		put_unbound_pool(pool);
5157 		return NULL;
5158 	}
5159 
5160 	init_pwq(pwq, wq, pool);
5161 	return pwq;
5162 }
5163 
5164 static void apply_wqattrs_lock(void)
5165 {
5166 	mutex_lock(&wq_pool_mutex);
5167 }
5168 
5169 static void apply_wqattrs_unlock(void)
5170 {
5171 	mutex_unlock(&wq_pool_mutex);
5172 }
5173 
5174 /**
5175  * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
5176  * @attrs: the wq_attrs of the default pwq of the target workqueue
5177  * @cpu: the target CPU
5178  *
5179  * Calculate the cpumask a workqueue with @attrs should use on @pod.
5180  * The result is stored in @attrs->__pod_cpumask.
5181  *
5182  * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
5183  * and @pod has online CPUs requested by @attrs, the returned cpumask is the
5184  * intersection of the possible CPUs of @pod and @attrs->cpumask.
5185  *
5186  * The caller is responsible for ensuring that the cpumask of @pod stays stable.
5187  */
5188 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu)
5189 {
5190 	const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5191 	int pod = pt->cpu_pod[cpu];
5192 
5193 	/* calculate possible CPUs in @pod that @attrs wants */
5194 	cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
5195 	/* does @pod have any online CPUs @attrs wants? */
5196 	if (!cpumask_intersects(attrs->__pod_cpumask, wq_online_cpumask)) {
5197 		cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
5198 		return;
5199 	}
5200 }
5201 
5202 /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
5203 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
5204 					int cpu, struct pool_workqueue *pwq)
5205 {
5206 	struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
5207 	struct pool_workqueue *old_pwq;
5208 
5209 	lockdep_assert_held(&wq_pool_mutex);
5210 	lockdep_assert_held(&wq->mutex);
5211 
5212 	/* link_pwq() can handle duplicate calls */
5213 	link_pwq(pwq);
5214 
5215 	old_pwq = rcu_access_pointer(*slot);
5216 	rcu_assign_pointer(*slot, pwq);
5217 	return old_pwq;
5218 }
5219 
5220 /* context to store the prepared attrs & pwqs before applying */
5221 struct apply_wqattrs_ctx {
5222 	struct workqueue_struct	*wq;		/* target workqueue */
5223 	struct workqueue_attrs	*attrs;		/* attrs to apply */
5224 	struct list_head	list;		/* queued for batching commit */
5225 	struct pool_workqueue	*dfl_pwq;
5226 	struct pool_workqueue	*pwq_tbl[];
5227 };
5228 
5229 /* free the resources after success or abort */
5230 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5231 {
5232 	if (ctx) {
5233 		int cpu;
5234 
5235 		for_each_possible_cpu(cpu)
5236 			put_pwq_unlocked(ctx->pwq_tbl[cpu]);
5237 		put_pwq_unlocked(ctx->dfl_pwq);
5238 
5239 		free_workqueue_attrs(ctx->attrs);
5240 
5241 		kfree(ctx);
5242 	}
5243 }
5244 
5245 /* allocate the attrs and pwqs for later installation */
5246 static struct apply_wqattrs_ctx *
5247 apply_wqattrs_prepare(struct workqueue_struct *wq,
5248 		      const struct workqueue_attrs *attrs,
5249 		      const cpumask_var_t unbound_cpumask)
5250 {
5251 	struct apply_wqattrs_ctx *ctx;
5252 	struct workqueue_attrs *new_attrs;
5253 	int cpu;
5254 
5255 	lockdep_assert_held(&wq_pool_mutex);
5256 
5257 	if (WARN_ON(attrs->affn_scope < 0 ||
5258 		    attrs->affn_scope >= WQ_AFFN_NR_TYPES))
5259 		return ERR_PTR(-EINVAL);
5260 
5261 	ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
5262 
5263 	new_attrs = alloc_workqueue_attrs();
5264 	if (!ctx || !new_attrs)
5265 		goto out_free;
5266 
5267 	/*
5268 	 * If something goes wrong during CPU up/down, we'll fall back to
5269 	 * the default pwq covering whole @attrs->cpumask.  Always create
5270 	 * it even if we don't use it immediately.
5271 	 */
5272 	copy_workqueue_attrs(new_attrs, attrs);
5273 	wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
5274 	cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5275 	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
5276 	if (!ctx->dfl_pwq)
5277 		goto out_free;
5278 
5279 	for_each_possible_cpu(cpu) {
5280 		if (new_attrs->ordered) {
5281 			ctx->dfl_pwq->refcnt++;
5282 			ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
5283 		} else {
5284 			wq_calc_pod_cpumask(new_attrs, cpu);
5285 			ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
5286 			if (!ctx->pwq_tbl[cpu])
5287 				goto out_free;
5288 		}
5289 	}
5290 
5291 	/* save the user configured attrs and sanitize it. */
5292 	copy_workqueue_attrs(new_attrs, attrs);
5293 	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
5294 	cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5295 	ctx->attrs = new_attrs;
5296 
5297 	/*
5298 	 * For initialized ordered workqueues, there should only be one pwq
5299 	 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution
5300 	 * of newly queued work items until execution of older work items in
5301 	 * the old pwq's have completed.
5302 	 */
5303 	if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))
5304 		ctx->dfl_pwq->plugged = true;
5305 
5306 	ctx->wq = wq;
5307 	return ctx;
5308 
5309 out_free:
5310 	free_workqueue_attrs(new_attrs);
5311 	apply_wqattrs_cleanup(ctx);
5312 	return ERR_PTR(-ENOMEM);
5313 }
5314 
5315 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
5316 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5317 {
5318 	int cpu;
5319 
5320 	/* all pwqs have been created successfully, let's install'em */
5321 	mutex_lock(&ctx->wq->mutex);
5322 
5323 	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
5324 
5325 	/* save the previous pwqs and install the new ones */
5326 	for_each_possible_cpu(cpu)
5327 		ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
5328 							ctx->pwq_tbl[cpu]);
5329 	ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq);
5330 
5331 	/* update node_nr_active->max */
5332 	wq_update_node_max_active(ctx->wq, -1);
5333 
5334 	/* rescuer needs to respect wq cpumask changes */
5335 	if (ctx->wq->rescuer)
5336 		set_cpus_allowed_ptr(ctx->wq->rescuer->task,
5337 				     unbound_effective_cpumask(ctx->wq));
5338 
5339 	mutex_unlock(&ctx->wq->mutex);
5340 }
5341 
5342 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5343 					const struct workqueue_attrs *attrs)
5344 {
5345 	struct apply_wqattrs_ctx *ctx;
5346 
5347 	/* only unbound workqueues can change attributes */
5348 	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5349 		return -EINVAL;
5350 
5351 	ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
5352 	if (IS_ERR(ctx))
5353 		return PTR_ERR(ctx);
5354 
5355 	/* the ctx has been prepared successfully, let's commit it */
5356 	apply_wqattrs_commit(ctx);
5357 	apply_wqattrs_cleanup(ctx);
5358 
5359 	return 0;
5360 }
5361 
5362 /**
5363  * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5364  * @wq: the target workqueue
5365  * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5366  *
5367  * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
5368  * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
5369  * work items are affine to the pod it was issued on. Older pwqs are released as
5370  * in-flight work items finish. Note that a work item which repeatedly requeues
5371  * itself back-to-back will stay on its current pwq.
5372  *
5373  * Performs GFP_KERNEL allocations.
5374  *
5375  * Return: 0 on success and -errno on failure.
5376  */
5377 int apply_workqueue_attrs(struct workqueue_struct *wq,
5378 			  const struct workqueue_attrs *attrs)
5379 {
5380 	int ret;
5381 
5382 	mutex_lock(&wq_pool_mutex);
5383 	ret = apply_workqueue_attrs_locked(wq, attrs);
5384 	mutex_unlock(&wq_pool_mutex);
5385 
5386 	return ret;
5387 }
5388 
5389 /**
5390  * unbound_wq_update_pwq - update a pwq slot for CPU hot[un]plug
5391  * @wq: the target workqueue
5392  * @cpu: the CPU to update the pwq slot for
5393  *
5394  * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5395  * %CPU_DOWN_FAILED.  @cpu is in the same pod of the CPU being hot[un]plugged.
5396  *
5397  *
5398  * If pod affinity can't be adjusted due to memory allocation failure, it falls
5399  * back to @wq->dfl_pwq which may not be optimal but is always correct.
5400  *
5401  * Note that when the last allowed CPU of a pod goes offline for a workqueue
5402  * with a cpumask spanning multiple pods, the workers which were already
5403  * executing the work items for the workqueue will lose their CPU affinity and
5404  * may execute on any CPU. This is similar to how per-cpu workqueues behave on
5405  * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
5406  * responsibility to flush the work item from CPU_DOWN_PREPARE.
5407  */
5408 static void unbound_wq_update_pwq(struct workqueue_struct *wq, int cpu)
5409 {
5410 	struct pool_workqueue *old_pwq = NULL, *pwq;
5411 	struct workqueue_attrs *target_attrs;
5412 
5413 	lockdep_assert_held(&wq_pool_mutex);
5414 
5415 	if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
5416 		return;
5417 
5418 	/*
5419 	 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5420 	 * Let's use a preallocated one.  The following buf is protected by
5421 	 * CPU hotplug exclusion.
5422 	 */
5423 	target_attrs = unbound_wq_update_pwq_attrs_buf;
5424 
5425 	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5426 	wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
5427 
5428 	/* nothing to do if the target cpumask matches the current pwq */
5429 	wq_calc_pod_cpumask(target_attrs, cpu);
5430 	if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs))
5431 		return;
5432 
5433 	/* create a new pwq */
5434 	pwq = alloc_unbound_pwq(wq, target_attrs);
5435 	if (!pwq) {
5436 		pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
5437 			wq->name);
5438 		goto use_dfl_pwq;
5439 	}
5440 
5441 	/* Install the new pwq. */
5442 	mutex_lock(&wq->mutex);
5443 	old_pwq = install_unbound_pwq(wq, cpu, pwq);
5444 	goto out_unlock;
5445 
5446 use_dfl_pwq:
5447 	mutex_lock(&wq->mutex);
5448 	pwq = unbound_pwq(wq, -1);
5449 	raw_spin_lock_irq(&pwq->pool->lock);
5450 	get_pwq(pwq);
5451 	raw_spin_unlock_irq(&pwq->pool->lock);
5452 	old_pwq = install_unbound_pwq(wq, cpu, pwq);
5453 out_unlock:
5454 	mutex_unlock(&wq->mutex);
5455 	put_pwq_unlocked(old_pwq);
5456 }
5457 
5458 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5459 {
5460 	bool highpri = wq->flags & WQ_HIGHPRI;
5461 	int cpu, ret;
5462 
5463 	lockdep_assert_held(&wq_pool_mutex);
5464 
5465 	wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
5466 	if (!wq->cpu_pwq)
5467 		goto enomem;
5468 
5469 	if (!(wq->flags & WQ_UNBOUND)) {
5470 		struct worker_pool __percpu *pools;
5471 
5472 		if (wq->flags & WQ_BH)
5473 			pools = bh_worker_pools;
5474 		else
5475 			pools = cpu_worker_pools;
5476 
5477 		for_each_possible_cpu(cpu) {
5478 			struct pool_workqueue **pwq_p;
5479 			struct worker_pool *pool;
5480 
5481 			pool = &(per_cpu_ptr(pools, cpu)[highpri]);
5482 			pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu);
5483 
5484 			*pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5485 						       pool->node);
5486 			if (!*pwq_p)
5487 				goto enomem;
5488 
5489 			init_pwq(*pwq_p, wq, pool);
5490 
5491 			mutex_lock(&wq->mutex);
5492 			link_pwq(*pwq_p);
5493 			mutex_unlock(&wq->mutex);
5494 		}
5495 		return 0;
5496 	}
5497 
5498 	if (wq->flags & __WQ_ORDERED) {
5499 		struct pool_workqueue *dfl_pwq;
5500 
5501 		ret = apply_workqueue_attrs_locked(wq, ordered_wq_attrs[highpri]);
5502 		/* there should only be single pwq for ordering guarantee */
5503 		dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5504 		WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node ||
5505 			      wq->pwqs.prev != &dfl_pwq->pwqs_node),
5506 		     "ordering guarantee broken for workqueue %s\n", wq->name);
5507 	} else {
5508 		ret = apply_workqueue_attrs_locked(wq, unbound_std_wq_attrs[highpri]);
5509 	}
5510 
5511 	return ret;
5512 
5513 enomem:
5514 	if (wq->cpu_pwq) {
5515 		for_each_possible_cpu(cpu) {
5516 			struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5517 
5518 			if (pwq)
5519 				kmem_cache_free(pwq_cache, pwq);
5520 		}
5521 		free_percpu(wq->cpu_pwq);
5522 		wq->cpu_pwq = NULL;
5523 	}
5524 	return -ENOMEM;
5525 }
5526 
5527 static int wq_clamp_max_active(int max_active, unsigned int flags,
5528 			       const char *name)
5529 {
5530 	if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5531 		pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5532 			max_active, name, 1, WQ_MAX_ACTIVE);
5533 
5534 	return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5535 }
5536 
5537 /*
5538  * Workqueues which may be used during memory reclaim should have a rescuer
5539  * to guarantee forward progress.
5540  */
5541 static int init_rescuer(struct workqueue_struct *wq)
5542 {
5543 	struct worker *rescuer;
5544 	char id_buf[WORKER_ID_LEN];
5545 	int ret;
5546 
5547 	lockdep_assert_held(&wq_pool_mutex);
5548 
5549 	if (!(wq->flags & WQ_MEM_RECLAIM))
5550 		return 0;
5551 
5552 	rescuer = alloc_worker(NUMA_NO_NODE);
5553 	if (!rescuer) {
5554 		pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5555 		       wq->name);
5556 		return -ENOMEM;
5557 	}
5558 
5559 	rescuer->rescue_wq = wq;
5560 	format_worker_id(id_buf, sizeof(id_buf), rescuer, NULL);
5561 
5562 	rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", id_buf);
5563 	if (IS_ERR(rescuer->task)) {
5564 		ret = PTR_ERR(rescuer->task);
5565 		pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5566 		       wq->name, ERR_PTR(ret));
5567 		kfree(rescuer);
5568 		return ret;
5569 	}
5570 
5571 	wq->rescuer = rescuer;
5572 	if (wq->flags & WQ_UNBOUND)
5573 		kthread_bind_mask(rescuer->task, unbound_effective_cpumask(wq));
5574 	else
5575 		kthread_bind_mask(rescuer->task, cpu_possible_mask);
5576 	wake_up_process(rescuer->task);
5577 
5578 	return 0;
5579 }
5580 
5581 /**
5582  * wq_adjust_max_active - update a wq's max_active to the current setting
5583  * @wq: target workqueue
5584  *
5585  * If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5586  * activate inactive work items accordingly. If @wq is freezing, clear
5587  * @wq->max_active to zero.
5588  */
5589 static void wq_adjust_max_active(struct workqueue_struct *wq)
5590 {
5591 	bool activated;
5592 	int new_max, new_min;
5593 
5594 	lockdep_assert_held(&wq->mutex);
5595 
5596 	if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5597 		new_max = 0;
5598 		new_min = 0;
5599 	} else {
5600 		new_max = wq->saved_max_active;
5601 		new_min = wq->saved_min_active;
5602 	}
5603 
5604 	if (wq->max_active == new_max && wq->min_active == new_min)
5605 		return;
5606 
5607 	/*
5608 	 * Update @wq->max/min_active and then kick inactive work items if more
5609 	 * active work items are allowed. This doesn't break work item ordering
5610 	 * because new work items are always queued behind existing inactive
5611 	 * work items if there are any.
5612 	 */
5613 	WRITE_ONCE(wq->max_active, new_max);
5614 	WRITE_ONCE(wq->min_active, new_min);
5615 
5616 	if (wq->flags & WQ_UNBOUND)
5617 		wq_update_node_max_active(wq, -1);
5618 
5619 	if (new_max == 0)
5620 		return;
5621 
5622 	/*
5623 	 * Round-robin through pwq's activating the first inactive work item
5624 	 * until max_active is filled.
5625 	 */
5626 	do {
5627 		struct pool_workqueue *pwq;
5628 
5629 		activated = false;
5630 		for_each_pwq(pwq, wq) {
5631 			unsigned long irq_flags;
5632 
5633 			/* can be called during early boot w/ irq disabled */
5634 			raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
5635 			if (pwq_activate_first_inactive(pwq, true)) {
5636 				activated = true;
5637 				kick_pool(pwq->pool);
5638 			}
5639 			raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
5640 		}
5641 	} while (activated);
5642 }
5643 
5644 static struct workqueue_struct *__alloc_workqueue(const char *fmt,
5645 						  unsigned int flags,
5646 						  int max_active, va_list args)
5647 {
5648 	struct workqueue_struct *wq;
5649 	size_t wq_size;
5650 	int name_len;
5651 
5652 	if (flags & WQ_BH) {
5653 		if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS))
5654 			return NULL;
5655 		if (WARN_ON_ONCE(max_active))
5656 			return NULL;
5657 	}
5658 
5659 	/* see the comment above the definition of WQ_POWER_EFFICIENT */
5660 	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5661 		flags |= WQ_UNBOUND;
5662 
5663 	/* allocate wq and format name */
5664 	if (flags & WQ_UNBOUND)
5665 		wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5666 	else
5667 		wq_size = sizeof(*wq);
5668 
5669 	wq = kzalloc(wq_size, GFP_KERNEL);
5670 	if (!wq)
5671 		return NULL;
5672 
5673 	if (flags & WQ_UNBOUND) {
5674 		wq->unbound_attrs = alloc_workqueue_attrs();
5675 		if (!wq->unbound_attrs)
5676 			goto err_free_wq;
5677 	}
5678 
5679 	name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5680 
5681 	if (name_len >= WQ_NAME_LEN)
5682 		pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5683 			     wq->name);
5684 
5685 	if (flags & WQ_BH) {
5686 		/*
5687 		 * BH workqueues always share a single execution context per CPU
5688 		 * and don't impose any max_active limit.
5689 		 */
5690 		max_active = INT_MAX;
5691 	} else {
5692 		max_active = max_active ?: WQ_DFL_ACTIVE;
5693 		max_active = wq_clamp_max_active(max_active, flags, wq->name);
5694 	}
5695 
5696 	/* init wq */
5697 	wq->flags = flags;
5698 	wq->max_active = max_active;
5699 	wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5700 	wq->saved_max_active = wq->max_active;
5701 	wq->saved_min_active = wq->min_active;
5702 	mutex_init(&wq->mutex);
5703 	atomic_set(&wq->nr_pwqs_to_flush, 0);
5704 	INIT_LIST_HEAD(&wq->pwqs);
5705 	INIT_LIST_HEAD(&wq->flusher_queue);
5706 	INIT_LIST_HEAD(&wq->flusher_overflow);
5707 	INIT_LIST_HEAD(&wq->maydays);
5708 
5709 	INIT_LIST_HEAD(&wq->list);
5710 
5711 	if (flags & WQ_UNBOUND) {
5712 		if (alloc_node_nr_active(wq->node_nr_active) < 0)
5713 			goto err_free_wq;
5714 	}
5715 
5716 	/*
5717 	 * wq_pool_mutex protects the workqueues list, allocations of PWQs,
5718 	 * and the global freeze state.
5719 	 */
5720 	apply_wqattrs_lock();
5721 
5722 	if (alloc_and_link_pwqs(wq) < 0)
5723 		goto err_unlock_free_node_nr_active;
5724 
5725 	mutex_lock(&wq->mutex);
5726 	wq_adjust_max_active(wq);
5727 	mutex_unlock(&wq->mutex);
5728 
5729 	list_add_tail_rcu(&wq->list, &workqueues);
5730 
5731 	if (wq_online && init_rescuer(wq) < 0)
5732 		goto err_unlock_destroy;
5733 
5734 	apply_wqattrs_unlock();
5735 
5736 	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5737 		goto err_destroy;
5738 
5739 	return wq;
5740 
5741 err_unlock_free_node_nr_active:
5742 	apply_wqattrs_unlock();
5743 	/*
5744 	 * Failed alloc_and_link_pwqs() may leave pending pwq->release_work,
5745 	 * flushing the pwq_release_worker ensures that the pwq_release_workfn()
5746 	 * completes before calling kfree(wq).
5747 	 */
5748 	if (wq->flags & WQ_UNBOUND) {
5749 		kthread_flush_worker(pwq_release_worker);
5750 		free_node_nr_active(wq->node_nr_active);
5751 	}
5752 err_free_wq:
5753 	free_workqueue_attrs(wq->unbound_attrs);
5754 	kfree(wq);
5755 	return NULL;
5756 err_unlock_destroy:
5757 	apply_wqattrs_unlock();
5758 err_destroy:
5759 	destroy_workqueue(wq);
5760 	return NULL;
5761 }
5762 
5763 __printf(1, 4)
5764 struct workqueue_struct *alloc_workqueue(const char *fmt,
5765 					 unsigned int flags,
5766 					 int max_active, ...)
5767 {
5768 	struct workqueue_struct *wq;
5769 	va_list args;
5770 
5771 	va_start(args, max_active);
5772 	wq = __alloc_workqueue(fmt, flags, max_active, args);
5773 	va_end(args);
5774 	if (!wq)
5775 		return NULL;
5776 
5777 	wq_init_lockdep(wq);
5778 
5779 	return wq;
5780 }
5781 EXPORT_SYMBOL_GPL(alloc_workqueue);
5782 
5783 #ifdef CONFIG_LOCKDEP
5784 __printf(1, 5)
5785 struct workqueue_struct *
5786 alloc_workqueue_lockdep_map(const char *fmt, unsigned int flags,
5787 			    int max_active, struct lockdep_map *lockdep_map, ...)
5788 {
5789 	struct workqueue_struct *wq;
5790 	va_list args;
5791 
5792 	va_start(args, lockdep_map);
5793 	wq = __alloc_workqueue(fmt, flags, max_active, args);
5794 	va_end(args);
5795 	if (!wq)
5796 		return NULL;
5797 
5798 	wq->lockdep_map = lockdep_map;
5799 
5800 	return wq;
5801 }
5802 EXPORT_SYMBOL_GPL(alloc_workqueue_lockdep_map);
5803 #endif
5804 
5805 static bool pwq_busy(struct pool_workqueue *pwq)
5806 {
5807 	int i;
5808 
5809 	for (i = 0; i < WORK_NR_COLORS; i++)
5810 		if (pwq->nr_in_flight[i])
5811 			return true;
5812 
5813 	if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
5814 		return true;
5815 	if (!pwq_is_empty(pwq))
5816 		return true;
5817 
5818 	return false;
5819 }
5820 
5821 /**
5822  * destroy_workqueue - safely terminate a workqueue
5823  * @wq: target workqueue
5824  *
5825  * Safely destroy a workqueue. All work currently pending will be done first.
5826  */
5827 void destroy_workqueue(struct workqueue_struct *wq)
5828 {
5829 	struct pool_workqueue *pwq;
5830 	int cpu;
5831 
5832 	/*
5833 	 * Remove it from sysfs first so that sanity check failure doesn't
5834 	 * lead to sysfs name conflicts.
5835 	 */
5836 	workqueue_sysfs_unregister(wq);
5837 
5838 	/* mark the workqueue destruction is in progress */
5839 	mutex_lock(&wq->mutex);
5840 	wq->flags |= __WQ_DESTROYING;
5841 	mutex_unlock(&wq->mutex);
5842 
5843 	/* drain it before proceeding with destruction */
5844 	drain_workqueue(wq);
5845 
5846 	/* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5847 	if (wq->rescuer) {
5848 		struct worker *rescuer = wq->rescuer;
5849 
5850 		/* this prevents new queueing */
5851 		raw_spin_lock_irq(&wq_mayday_lock);
5852 		wq->rescuer = NULL;
5853 		raw_spin_unlock_irq(&wq_mayday_lock);
5854 
5855 		/* rescuer will empty maydays list before exiting */
5856 		kthread_stop(rescuer->task);
5857 		kfree(rescuer);
5858 	}
5859 
5860 	/*
5861 	 * Sanity checks - grab all the locks so that we wait for all
5862 	 * in-flight operations which may do put_pwq().
5863 	 */
5864 	mutex_lock(&wq_pool_mutex);
5865 	mutex_lock(&wq->mutex);
5866 	for_each_pwq(pwq, wq) {
5867 		raw_spin_lock_irq(&pwq->pool->lock);
5868 		if (WARN_ON(pwq_busy(pwq))) {
5869 			pr_warn("%s: %s has the following busy pwq\n",
5870 				__func__, wq->name);
5871 			show_pwq(pwq);
5872 			raw_spin_unlock_irq(&pwq->pool->lock);
5873 			mutex_unlock(&wq->mutex);
5874 			mutex_unlock(&wq_pool_mutex);
5875 			show_one_workqueue(wq);
5876 			return;
5877 		}
5878 		raw_spin_unlock_irq(&pwq->pool->lock);
5879 	}
5880 	mutex_unlock(&wq->mutex);
5881 
5882 	/*
5883 	 * wq list is used to freeze wq, remove from list after
5884 	 * flushing is complete in case freeze races us.
5885 	 */
5886 	list_del_rcu(&wq->list);
5887 	mutex_unlock(&wq_pool_mutex);
5888 
5889 	/*
5890 	 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5891 	 * to put the base refs. @wq will be auto-destroyed from the last
5892 	 * pwq_put. RCU read lock prevents @wq from going away from under us.
5893 	 */
5894 	rcu_read_lock();
5895 
5896 	for_each_possible_cpu(cpu) {
5897 		put_pwq_unlocked(unbound_pwq(wq, cpu));
5898 		RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
5899 	}
5900 
5901 	put_pwq_unlocked(unbound_pwq(wq, -1));
5902 	RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
5903 
5904 	rcu_read_unlock();
5905 }
5906 EXPORT_SYMBOL_GPL(destroy_workqueue);
5907 
5908 /**
5909  * workqueue_set_max_active - adjust max_active of a workqueue
5910  * @wq: target workqueue
5911  * @max_active: new max_active value.
5912  *
5913  * Set max_active of @wq to @max_active. See the alloc_workqueue() function
5914  * comment.
5915  *
5916  * CONTEXT:
5917  * Don't call from IRQ context.
5918  */
5919 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5920 {
5921 	/* max_active doesn't mean anything for BH workqueues */
5922 	if (WARN_ON(wq->flags & WQ_BH))
5923 		return;
5924 	/* disallow meddling with max_active for ordered workqueues */
5925 	if (WARN_ON(wq->flags & __WQ_ORDERED))
5926 		return;
5927 
5928 	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5929 
5930 	mutex_lock(&wq->mutex);
5931 
5932 	wq->saved_max_active = max_active;
5933 	if (wq->flags & WQ_UNBOUND)
5934 		wq->saved_min_active = min(wq->saved_min_active, max_active);
5935 
5936 	wq_adjust_max_active(wq);
5937 
5938 	mutex_unlock(&wq->mutex);
5939 }
5940 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5941 
5942 /**
5943  * workqueue_set_min_active - adjust min_active of an unbound workqueue
5944  * @wq: target unbound workqueue
5945  * @min_active: new min_active value
5946  *
5947  * Set min_active of an unbound workqueue. Unlike other types of workqueues, an
5948  * unbound workqueue is not guaranteed to be able to process max_active
5949  * interdependent work items. Instead, an unbound workqueue is guaranteed to be
5950  * able to process min_active number of interdependent work items which is
5951  * %WQ_DFL_MIN_ACTIVE by default.
5952  *
5953  * Use this function to adjust the min_active value between 0 and the current
5954  * max_active.
5955  */
5956 void workqueue_set_min_active(struct workqueue_struct *wq, int min_active)
5957 {
5958 	/* min_active is only meaningful for non-ordered unbound workqueues */
5959 	if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) !=
5960 		    WQ_UNBOUND))
5961 		return;
5962 
5963 	mutex_lock(&wq->mutex);
5964 	wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active);
5965 	wq_adjust_max_active(wq);
5966 	mutex_unlock(&wq->mutex);
5967 }
5968 
5969 /**
5970  * current_work - retrieve %current task's work struct
5971  *
5972  * Determine if %current task is a workqueue worker and what it's working on.
5973  * Useful to find out the context that the %current task is running in.
5974  *
5975  * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5976  */
5977 struct work_struct *current_work(void)
5978 {
5979 	struct worker *worker = current_wq_worker();
5980 
5981 	return worker ? worker->current_work : NULL;
5982 }
5983 EXPORT_SYMBOL(current_work);
5984 
5985 /**
5986  * current_is_workqueue_rescuer - is %current workqueue rescuer?
5987  *
5988  * Determine whether %current is a workqueue rescuer.  Can be used from
5989  * work functions to determine whether it's being run off the rescuer task.
5990  *
5991  * Return: %true if %current is a workqueue rescuer. %false otherwise.
5992  */
5993 bool current_is_workqueue_rescuer(void)
5994 {
5995 	struct worker *worker = current_wq_worker();
5996 
5997 	return worker && worker->rescue_wq;
5998 }
5999 
6000 /**
6001  * workqueue_congested - test whether a workqueue is congested
6002  * @cpu: CPU in question
6003  * @wq: target workqueue
6004  *
6005  * Test whether @wq's cpu workqueue for @cpu is congested.  There is
6006  * no synchronization around this function and the test result is
6007  * unreliable and only useful as advisory hints or for debugging.
6008  *
6009  * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
6010  *
6011  * With the exception of ordered workqueues, all workqueues have per-cpu
6012  * pool_workqueues, each with its own congested state. A workqueue being
6013  * congested on one CPU doesn't mean that the workqueue is contested on any
6014  * other CPUs.
6015  *
6016  * Return:
6017  * %true if congested, %false otherwise.
6018  */
6019 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
6020 {
6021 	struct pool_workqueue *pwq;
6022 	bool ret;
6023 
6024 	rcu_read_lock();
6025 	preempt_disable();
6026 
6027 	if (cpu == WORK_CPU_UNBOUND)
6028 		cpu = smp_processor_id();
6029 
6030 	pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
6031 	ret = !list_empty(&pwq->inactive_works);
6032 
6033 	preempt_enable();
6034 	rcu_read_unlock();
6035 
6036 	return ret;
6037 }
6038 EXPORT_SYMBOL_GPL(workqueue_congested);
6039 
6040 /**
6041  * work_busy - test whether a work is currently pending or running
6042  * @work: the work to be tested
6043  *
6044  * Test whether @work is currently pending or running.  There is no
6045  * synchronization around this function and the test result is
6046  * unreliable and only useful as advisory hints or for debugging.
6047  *
6048  * Return:
6049  * OR'd bitmask of WORK_BUSY_* bits.
6050  */
6051 unsigned int work_busy(struct work_struct *work)
6052 {
6053 	struct worker_pool *pool;
6054 	unsigned long irq_flags;
6055 	unsigned int ret = 0;
6056 
6057 	if (work_pending(work))
6058 		ret |= WORK_BUSY_PENDING;
6059 
6060 	rcu_read_lock();
6061 	pool = get_work_pool(work);
6062 	if (pool) {
6063 		raw_spin_lock_irqsave(&pool->lock, irq_flags);
6064 		if (find_worker_executing_work(pool, work))
6065 			ret |= WORK_BUSY_RUNNING;
6066 		raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6067 	}
6068 	rcu_read_unlock();
6069 
6070 	return ret;
6071 }
6072 EXPORT_SYMBOL_GPL(work_busy);
6073 
6074 /**
6075  * set_worker_desc - set description for the current work item
6076  * @fmt: printf-style format string
6077  * @...: arguments for the format string
6078  *
6079  * This function can be called by a running work function to describe what
6080  * the work item is about.  If the worker task gets dumped, this
6081  * information will be printed out together to help debugging.  The
6082  * description can be at most WORKER_DESC_LEN including the trailing '\0'.
6083  */
6084 void set_worker_desc(const char *fmt, ...)
6085 {
6086 	struct worker *worker = current_wq_worker();
6087 	va_list args;
6088 
6089 	if (worker) {
6090 		va_start(args, fmt);
6091 		vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
6092 		va_end(args);
6093 	}
6094 }
6095 EXPORT_SYMBOL_GPL(set_worker_desc);
6096 
6097 /**
6098  * print_worker_info - print out worker information and description
6099  * @log_lvl: the log level to use when printing
6100  * @task: target task
6101  *
6102  * If @task is a worker and currently executing a work item, print out the
6103  * name of the workqueue being serviced and worker description set with
6104  * set_worker_desc() by the currently executing work item.
6105  *
6106  * This function can be safely called on any task as long as the
6107  * task_struct itself is accessible.  While safe, this function isn't
6108  * synchronized and may print out mixups or garbages of limited length.
6109  */
6110 void print_worker_info(const char *log_lvl, struct task_struct *task)
6111 {
6112 	work_func_t *fn = NULL;
6113 	char name[WQ_NAME_LEN] = { };
6114 	char desc[WORKER_DESC_LEN] = { };
6115 	struct pool_workqueue *pwq = NULL;
6116 	struct workqueue_struct *wq = NULL;
6117 	struct worker *worker;
6118 
6119 	if (!(task->flags & PF_WQ_WORKER))
6120 		return;
6121 
6122 	/*
6123 	 * This function is called without any synchronization and @task
6124 	 * could be in any state.  Be careful with dereferences.
6125 	 */
6126 	worker = kthread_probe_data(task);
6127 
6128 	/*
6129 	 * Carefully copy the associated workqueue's workfn, name and desc.
6130 	 * Keep the original last '\0' in case the original is garbage.
6131 	 */
6132 	copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
6133 	copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
6134 	copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
6135 	copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
6136 	copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
6137 
6138 	if (fn || name[0] || desc[0]) {
6139 		printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
6140 		if (strcmp(name, desc))
6141 			pr_cont(" (%s)", desc);
6142 		pr_cont("\n");
6143 	}
6144 }
6145 
6146 static void pr_cont_pool_info(struct worker_pool *pool)
6147 {
6148 	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
6149 	if (pool->node != NUMA_NO_NODE)
6150 		pr_cont(" node=%d", pool->node);
6151 	pr_cont(" flags=0x%x", pool->flags);
6152 	if (pool->flags & POOL_BH)
6153 		pr_cont(" bh%s",
6154 			pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6155 	else
6156 		pr_cont(" nice=%d", pool->attrs->nice);
6157 }
6158 
6159 static void pr_cont_worker_id(struct worker *worker)
6160 {
6161 	struct worker_pool *pool = worker->pool;
6162 
6163 	if (pool->flags & WQ_BH)
6164 		pr_cont("bh%s",
6165 			pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6166 	else
6167 		pr_cont("%d%s", task_pid_nr(worker->task),
6168 			worker->rescue_wq ? "(RESCUER)" : "");
6169 }
6170 
6171 struct pr_cont_work_struct {
6172 	bool comma;
6173 	work_func_t func;
6174 	long ctr;
6175 };
6176 
6177 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
6178 {
6179 	if (!pcwsp->ctr)
6180 		goto out_record;
6181 	if (func == pcwsp->func) {
6182 		pcwsp->ctr++;
6183 		return;
6184 	}
6185 	if (pcwsp->ctr == 1)
6186 		pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
6187 	else
6188 		pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
6189 	pcwsp->ctr = 0;
6190 out_record:
6191 	if ((long)func == -1L)
6192 		return;
6193 	pcwsp->comma = comma;
6194 	pcwsp->func = func;
6195 	pcwsp->ctr = 1;
6196 }
6197 
6198 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
6199 {
6200 	if (work->func == wq_barrier_func) {
6201 		struct wq_barrier *barr;
6202 
6203 		barr = container_of(work, struct wq_barrier, work);
6204 
6205 		pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6206 		pr_cont("%s BAR(%d)", comma ? "," : "",
6207 			task_pid_nr(barr->task));
6208 	} else {
6209 		if (!comma)
6210 			pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6211 		pr_cont_work_flush(comma, work->func, pcwsp);
6212 	}
6213 }
6214 
6215 static void show_pwq(struct pool_workqueue *pwq)
6216 {
6217 	struct pr_cont_work_struct pcws = { .ctr = 0, };
6218 	struct worker_pool *pool = pwq->pool;
6219 	struct work_struct *work;
6220 	struct worker *worker;
6221 	bool has_in_flight = false, has_pending = false;
6222 	int bkt;
6223 
6224 	pr_info("  pwq %d:", pool->id);
6225 	pr_cont_pool_info(pool);
6226 
6227 	pr_cont(" active=%d refcnt=%d%s\n",
6228 		pwq->nr_active, pwq->refcnt,
6229 		!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6230 
6231 	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6232 		if (worker->current_pwq == pwq) {
6233 			has_in_flight = true;
6234 			break;
6235 		}
6236 	}
6237 	if (has_in_flight) {
6238 		bool comma = false;
6239 
6240 		pr_info("    in-flight:");
6241 		hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6242 			if (worker->current_pwq != pwq)
6243 				continue;
6244 
6245 			pr_cont(" %s", comma ? "," : "");
6246 			pr_cont_worker_id(worker);
6247 			pr_cont(":%ps", worker->current_func);
6248 			list_for_each_entry(work, &worker->scheduled, entry)
6249 				pr_cont_work(false, work, &pcws);
6250 			pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6251 			comma = true;
6252 		}
6253 		pr_cont("\n");
6254 	}
6255 
6256 	list_for_each_entry(work, &pool->worklist, entry) {
6257 		if (get_work_pwq(work) == pwq) {
6258 			has_pending = true;
6259 			break;
6260 		}
6261 	}
6262 	if (has_pending) {
6263 		bool comma = false;
6264 
6265 		pr_info("    pending:");
6266 		list_for_each_entry(work, &pool->worklist, entry) {
6267 			if (get_work_pwq(work) != pwq)
6268 				continue;
6269 
6270 			pr_cont_work(comma, work, &pcws);
6271 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6272 		}
6273 		pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6274 		pr_cont("\n");
6275 	}
6276 
6277 	if (!list_empty(&pwq->inactive_works)) {
6278 		bool comma = false;
6279 
6280 		pr_info("    inactive:");
6281 		list_for_each_entry(work, &pwq->inactive_works, entry) {
6282 			pr_cont_work(comma, work, &pcws);
6283 			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6284 		}
6285 		pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6286 		pr_cont("\n");
6287 	}
6288 }
6289 
6290 /**
6291  * show_one_workqueue - dump state of specified workqueue
6292  * @wq: workqueue whose state will be printed
6293  */
6294 void show_one_workqueue(struct workqueue_struct *wq)
6295 {
6296 	struct pool_workqueue *pwq;
6297 	bool idle = true;
6298 	unsigned long irq_flags;
6299 
6300 	for_each_pwq(pwq, wq) {
6301 		if (!pwq_is_empty(pwq)) {
6302 			idle = false;
6303 			break;
6304 		}
6305 	}
6306 	if (idle) /* Nothing to print for idle workqueue */
6307 		return;
6308 
6309 	pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6310 
6311 	for_each_pwq(pwq, wq) {
6312 		raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
6313 		if (!pwq_is_empty(pwq)) {
6314 			/*
6315 			 * Defer printing to avoid deadlocks in console
6316 			 * drivers that queue work while holding locks
6317 			 * also taken in their write paths.
6318 			 */
6319 			printk_deferred_enter();
6320 			show_pwq(pwq);
6321 			printk_deferred_exit();
6322 		}
6323 		raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
6324 		/*
6325 		 * We could be printing a lot from atomic context, e.g.
6326 		 * sysrq-t -> show_all_workqueues(). Avoid triggering
6327 		 * hard lockup.
6328 		 */
6329 		touch_nmi_watchdog();
6330 	}
6331 
6332 }
6333 
6334 /**
6335  * show_one_worker_pool - dump state of specified worker pool
6336  * @pool: worker pool whose state will be printed
6337  */
6338 static void show_one_worker_pool(struct worker_pool *pool)
6339 {
6340 	struct worker *worker;
6341 	bool first = true;
6342 	unsigned long irq_flags;
6343 	unsigned long hung = 0;
6344 
6345 	raw_spin_lock_irqsave(&pool->lock, irq_flags);
6346 	if (pool->nr_workers == pool->nr_idle)
6347 		goto next_pool;
6348 
6349 	/* How long the first pending work is waiting for a worker. */
6350 	if (!list_empty(&pool->worklist))
6351 		hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
6352 
6353 	/*
6354 	 * Defer printing to avoid deadlocks in console drivers that
6355 	 * queue work while holding locks also taken in their write
6356 	 * paths.
6357 	 */
6358 	printk_deferred_enter();
6359 	pr_info("pool %d:", pool->id);
6360 	pr_cont_pool_info(pool);
6361 	pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
6362 	if (pool->manager)
6363 		pr_cont(" manager: %d",
6364 			task_pid_nr(pool->manager->task));
6365 	list_for_each_entry(worker, &pool->idle_list, entry) {
6366 		pr_cont(" %s", first ? "idle: " : "");
6367 		pr_cont_worker_id(worker);
6368 		first = false;
6369 	}
6370 	pr_cont("\n");
6371 	printk_deferred_exit();
6372 next_pool:
6373 	raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6374 	/*
6375 	 * We could be printing a lot from atomic context, e.g.
6376 	 * sysrq-t -> show_all_workqueues(). Avoid triggering
6377 	 * hard lockup.
6378 	 */
6379 	touch_nmi_watchdog();
6380 
6381 }
6382 
6383 /**
6384  * show_all_workqueues - dump workqueue state
6385  *
6386  * Called from a sysrq handler and prints out all busy workqueues and pools.
6387  */
6388 void show_all_workqueues(void)
6389 {
6390 	struct workqueue_struct *wq;
6391 	struct worker_pool *pool;
6392 	int pi;
6393 
6394 	rcu_read_lock();
6395 
6396 	pr_info("Showing busy workqueues and worker pools:\n");
6397 
6398 	list_for_each_entry_rcu(wq, &workqueues, list)
6399 		show_one_workqueue(wq);
6400 
6401 	for_each_pool(pool, pi)
6402 		show_one_worker_pool(pool);
6403 
6404 	rcu_read_unlock();
6405 }
6406 
6407 /**
6408  * show_freezable_workqueues - dump freezable workqueue state
6409  *
6410  * Called from try_to_freeze_tasks() and prints out all freezable workqueues
6411  * still busy.
6412  */
6413 void show_freezable_workqueues(void)
6414 {
6415 	struct workqueue_struct *wq;
6416 
6417 	rcu_read_lock();
6418 
6419 	pr_info("Showing freezable workqueues that are still busy:\n");
6420 
6421 	list_for_each_entry_rcu(wq, &workqueues, list) {
6422 		if (!(wq->flags & WQ_FREEZABLE))
6423 			continue;
6424 		show_one_workqueue(wq);
6425 	}
6426 
6427 	rcu_read_unlock();
6428 }
6429 
6430 /* used to show worker information through /proc/PID/{comm,stat,status} */
6431 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
6432 {
6433 	/* stabilize PF_WQ_WORKER and worker pool association */
6434 	mutex_lock(&wq_pool_attach_mutex);
6435 
6436 	if (task->flags & PF_WQ_WORKER) {
6437 		struct worker *worker = kthread_data(task);
6438 		struct worker_pool *pool = worker->pool;
6439 		int off;
6440 
6441 		off = format_worker_id(buf, size, worker, pool);
6442 
6443 		if (pool) {
6444 			raw_spin_lock_irq(&pool->lock);
6445 			/*
6446 			 * ->desc tracks information (wq name or
6447 			 * set_worker_desc()) for the latest execution.  If
6448 			 * current, prepend '+', otherwise '-'.
6449 			 */
6450 			if (worker->desc[0] != '\0') {
6451 				if (worker->current_work)
6452 					scnprintf(buf + off, size - off, "+%s",
6453 						  worker->desc);
6454 				else
6455 					scnprintf(buf + off, size - off, "-%s",
6456 						  worker->desc);
6457 			}
6458 			raw_spin_unlock_irq(&pool->lock);
6459 		}
6460 	} else {
6461 		strscpy(buf, task->comm, size);
6462 	}
6463 
6464 	mutex_unlock(&wq_pool_attach_mutex);
6465 }
6466 
6467 #ifdef CONFIG_SMP
6468 
6469 /*
6470  * CPU hotplug.
6471  *
6472  * There are two challenges in supporting CPU hotplug.  Firstly, there
6473  * are a lot of assumptions on strong associations among work, pwq and
6474  * pool which make migrating pending and scheduled works very
6475  * difficult to implement without impacting hot paths.  Secondly,
6476  * worker pools serve mix of short, long and very long running works making
6477  * blocked draining impractical.
6478  *
6479  * This is solved by allowing the pools to be disassociated from the CPU
6480  * running as an unbound one and allowing it to be reattached later if the
6481  * cpu comes back online.
6482  */
6483 
6484 static void unbind_workers(int cpu)
6485 {
6486 	struct worker_pool *pool;
6487 	struct worker *worker;
6488 
6489 	for_each_cpu_worker_pool(pool, cpu) {
6490 		mutex_lock(&wq_pool_attach_mutex);
6491 		raw_spin_lock_irq(&pool->lock);
6492 
6493 		/*
6494 		 * We've blocked all attach/detach operations. Make all workers
6495 		 * unbound and set DISASSOCIATED.  Before this, all workers
6496 		 * must be on the cpu.  After this, they may become diasporas.
6497 		 * And the preemption disabled section in their sched callbacks
6498 		 * are guaranteed to see WORKER_UNBOUND since the code here
6499 		 * is on the same cpu.
6500 		 */
6501 		for_each_pool_worker(worker, pool)
6502 			worker->flags |= WORKER_UNBOUND;
6503 
6504 		pool->flags |= POOL_DISASSOCIATED;
6505 
6506 		/*
6507 		 * The handling of nr_running in sched callbacks are disabled
6508 		 * now.  Zap nr_running.  After this, nr_running stays zero and
6509 		 * need_more_worker() and keep_working() are always true as
6510 		 * long as the worklist is not empty.  This pool now behaves as
6511 		 * an unbound (in terms of concurrency management) pool which
6512 		 * are served by workers tied to the pool.
6513 		 */
6514 		pool->nr_running = 0;
6515 
6516 		/*
6517 		 * With concurrency management just turned off, a busy
6518 		 * worker blocking could lead to lengthy stalls.  Kick off
6519 		 * unbound chain execution of currently pending work items.
6520 		 */
6521 		kick_pool(pool);
6522 
6523 		raw_spin_unlock_irq(&pool->lock);
6524 
6525 		for_each_pool_worker(worker, pool)
6526 			unbind_worker(worker);
6527 
6528 		mutex_unlock(&wq_pool_attach_mutex);
6529 	}
6530 }
6531 
6532 /**
6533  * rebind_workers - rebind all workers of a pool to the associated CPU
6534  * @pool: pool of interest
6535  *
6536  * @pool->cpu is coming online.  Rebind all workers to the CPU.
6537  */
6538 static void rebind_workers(struct worker_pool *pool)
6539 {
6540 	struct worker *worker;
6541 
6542 	lockdep_assert_held(&wq_pool_attach_mutex);
6543 
6544 	/*
6545 	 * Restore CPU affinity of all workers.  As all idle workers should
6546 	 * be on the run-queue of the associated CPU before any local
6547 	 * wake-ups for concurrency management happen, restore CPU affinity
6548 	 * of all workers first and then clear UNBOUND.  As we're called
6549 	 * from CPU_ONLINE, the following shouldn't fail.
6550 	 */
6551 	for_each_pool_worker(worker, pool) {
6552 		kthread_set_per_cpu(worker->task, pool->cpu);
6553 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6554 						  pool_allowed_cpus(pool)) < 0);
6555 	}
6556 
6557 	raw_spin_lock_irq(&pool->lock);
6558 
6559 	pool->flags &= ~POOL_DISASSOCIATED;
6560 
6561 	for_each_pool_worker(worker, pool) {
6562 		unsigned int worker_flags = worker->flags;
6563 
6564 		/*
6565 		 * We want to clear UNBOUND but can't directly call
6566 		 * worker_clr_flags() or adjust nr_running.  Atomically
6567 		 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6568 		 * @worker will clear REBOUND using worker_clr_flags() when
6569 		 * it initiates the next execution cycle thus restoring
6570 		 * concurrency management.  Note that when or whether
6571 		 * @worker clears REBOUND doesn't affect correctness.
6572 		 *
6573 		 * WRITE_ONCE() is necessary because @worker->flags may be
6574 		 * tested without holding any lock in
6575 		 * wq_worker_running().  Without it, NOT_RUNNING test may
6576 		 * fail incorrectly leading to premature concurrency
6577 		 * management operations.
6578 		 */
6579 		WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6580 		worker_flags |= WORKER_REBOUND;
6581 		worker_flags &= ~WORKER_UNBOUND;
6582 		WRITE_ONCE(worker->flags, worker_flags);
6583 	}
6584 
6585 	raw_spin_unlock_irq(&pool->lock);
6586 }
6587 
6588 /**
6589  * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6590  * @pool: unbound pool of interest
6591  * @cpu: the CPU which is coming up
6592  *
6593  * An unbound pool may end up with a cpumask which doesn't have any online
6594  * CPUs.  When a worker of such pool get scheduled, the scheduler resets
6595  * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
6596  * online CPU before, cpus_allowed of all its workers should be restored.
6597  */
6598 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6599 {
6600 	static cpumask_t cpumask;
6601 	struct worker *worker;
6602 
6603 	lockdep_assert_held(&wq_pool_attach_mutex);
6604 
6605 	/* is @cpu allowed for @pool? */
6606 	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6607 		return;
6608 
6609 	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6610 
6611 	/* as we're called from CPU_ONLINE, the following shouldn't fail */
6612 	for_each_pool_worker(worker, pool)
6613 		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6614 }
6615 
6616 int workqueue_prepare_cpu(unsigned int cpu)
6617 {
6618 	struct worker_pool *pool;
6619 
6620 	for_each_cpu_worker_pool(pool, cpu) {
6621 		if (pool->nr_workers)
6622 			continue;
6623 		if (!create_worker(pool))
6624 			return -ENOMEM;
6625 	}
6626 	return 0;
6627 }
6628 
6629 int workqueue_online_cpu(unsigned int cpu)
6630 {
6631 	struct worker_pool *pool;
6632 	struct workqueue_struct *wq;
6633 	int pi;
6634 
6635 	mutex_lock(&wq_pool_mutex);
6636 
6637 	cpumask_set_cpu(cpu, wq_online_cpumask);
6638 
6639 	for_each_pool(pool, pi) {
6640 		/* BH pools aren't affected by hotplug */
6641 		if (pool->flags & POOL_BH)
6642 			continue;
6643 
6644 		mutex_lock(&wq_pool_attach_mutex);
6645 		if (pool->cpu == cpu)
6646 			rebind_workers(pool);
6647 		else if (pool->cpu < 0)
6648 			restore_unbound_workers_cpumask(pool, cpu);
6649 		mutex_unlock(&wq_pool_attach_mutex);
6650 	}
6651 
6652 	/* update pod affinity of unbound workqueues */
6653 	list_for_each_entry(wq, &workqueues, list) {
6654 		struct workqueue_attrs *attrs = wq->unbound_attrs;
6655 
6656 		if (attrs) {
6657 			const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6658 			int tcpu;
6659 
6660 			for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6661 				unbound_wq_update_pwq(wq, tcpu);
6662 
6663 			mutex_lock(&wq->mutex);
6664 			wq_update_node_max_active(wq, -1);
6665 			mutex_unlock(&wq->mutex);
6666 		}
6667 	}
6668 
6669 	mutex_unlock(&wq_pool_mutex);
6670 	return 0;
6671 }
6672 
6673 int workqueue_offline_cpu(unsigned int cpu)
6674 {
6675 	struct workqueue_struct *wq;
6676 
6677 	/* unbinding per-cpu workers should happen on the local CPU */
6678 	if (WARN_ON(cpu != smp_processor_id()))
6679 		return -1;
6680 
6681 	unbind_workers(cpu);
6682 
6683 	/* update pod affinity of unbound workqueues */
6684 	mutex_lock(&wq_pool_mutex);
6685 
6686 	cpumask_clear_cpu(cpu, wq_online_cpumask);
6687 
6688 	list_for_each_entry(wq, &workqueues, list) {
6689 		struct workqueue_attrs *attrs = wq->unbound_attrs;
6690 
6691 		if (attrs) {
6692 			const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6693 			int tcpu;
6694 
6695 			for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6696 				unbound_wq_update_pwq(wq, tcpu);
6697 
6698 			mutex_lock(&wq->mutex);
6699 			wq_update_node_max_active(wq, cpu);
6700 			mutex_unlock(&wq->mutex);
6701 		}
6702 	}
6703 	mutex_unlock(&wq_pool_mutex);
6704 
6705 	return 0;
6706 }
6707 
6708 struct work_for_cpu {
6709 	struct work_struct work;
6710 	long (*fn)(void *);
6711 	void *arg;
6712 	long ret;
6713 };
6714 
6715 static void work_for_cpu_fn(struct work_struct *work)
6716 {
6717 	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6718 
6719 	wfc->ret = wfc->fn(wfc->arg);
6720 }
6721 
6722 /**
6723  * work_on_cpu_key - run a function in thread context on a particular cpu
6724  * @cpu: the cpu to run on
6725  * @fn: the function to run
6726  * @arg: the function arg
6727  * @key: The lock class key for lock debugging purposes
6728  *
6729  * It is up to the caller to ensure that the cpu doesn't go offline.
6730  * The caller must not hold any locks which would prevent @fn from completing.
6731  *
6732  * Return: The value @fn returns.
6733  */
6734 long work_on_cpu_key(int cpu, long (*fn)(void *),
6735 		     void *arg, struct lock_class_key *key)
6736 {
6737 	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6738 
6739 	INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6740 	schedule_work_on(cpu, &wfc.work);
6741 	flush_work(&wfc.work);
6742 	destroy_work_on_stack(&wfc.work);
6743 	return wfc.ret;
6744 }
6745 EXPORT_SYMBOL_GPL(work_on_cpu_key);
6746 
6747 /**
6748  * work_on_cpu_safe_key - run a function in thread context on a particular cpu
6749  * @cpu: the cpu to run on
6750  * @fn:  the function to run
6751  * @arg: the function argument
6752  * @key: The lock class key for lock debugging purposes
6753  *
6754  * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6755  * any locks which would prevent @fn from completing.
6756  *
6757  * Return: The value @fn returns.
6758  */
6759 long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
6760 			  void *arg, struct lock_class_key *key)
6761 {
6762 	long ret = -ENODEV;
6763 
6764 	cpus_read_lock();
6765 	if (cpu_online(cpu))
6766 		ret = work_on_cpu_key(cpu, fn, arg, key);
6767 	cpus_read_unlock();
6768 	return ret;
6769 }
6770 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
6771 #endif /* CONFIG_SMP */
6772 
6773 #ifdef CONFIG_FREEZER
6774 
6775 /**
6776  * freeze_workqueues_begin - begin freezing workqueues
6777  *
6778  * Start freezing workqueues.  After this function returns, all freezable
6779  * workqueues will queue new works to their inactive_works list instead of
6780  * pool->worklist.
6781  *
6782  * CONTEXT:
6783  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6784  */
6785 void freeze_workqueues_begin(void)
6786 {
6787 	struct workqueue_struct *wq;
6788 
6789 	mutex_lock(&wq_pool_mutex);
6790 
6791 	WARN_ON_ONCE(workqueue_freezing);
6792 	workqueue_freezing = true;
6793 
6794 	list_for_each_entry(wq, &workqueues, list) {
6795 		mutex_lock(&wq->mutex);
6796 		wq_adjust_max_active(wq);
6797 		mutex_unlock(&wq->mutex);
6798 	}
6799 
6800 	mutex_unlock(&wq_pool_mutex);
6801 }
6802 
6803 /**
6804  * freeze_workqueues_busy - are freezable workqueues still busy?
6805  *
6806  * Check whether freezing is complete.  This function must be called
6807  * between freeze_workqueues_begin() and thaw_workqueues().
6808  *
6809  * CONTEXT:
6810  * Grabs and releases wq_pool_mutex.
6811  *
6812  * Return:
6813  * %true if some freezable workqueues are still busy.  %false if freezing
6814  * is complete.
6815  */
6816 bool freeze_workqueues_busy(void)
6817 {
6818 	bool busy = false;
6819 	struct workqueue_struct *wq;
6820 	struct pool_workqueue *pwq;
6821 
6822 	mutex_lock(&wq_pool_mutex);
6823 
6824 	WARN_ON_ONCE(!workqueue_freezing);
6825 
6826 	list_for_each_entry(wq, &workqueues, list) {
6827 		if (!(wq->flags & WQ_FREEZABLE))
6828 			continue;
6829 		/*
6830 		 * nr_active is monotonically decreasing.  It's safe
6831 		 * to peek without lock.
6832 		 */
6833 		rcu_read_lock();
6834 		for_each_pwq(pwq, wq) {
6835 			WARN_ON_ONCE(pwq->nr_active < 0);
6836 			if (pwq->nr_active) {
6837 				busy = true;
6838 				rcu_read_unlock();
6839 				goto out_unlock;
6840 			}
6841 		}
6842 		rcu_read_unlock();
6843 	}
6844 out_unlock:
6845 	mutex_unlock(&wq_pool_mutex);
6846 	return busy;
6847 }
6848 
6849 /**
6850  * thaw_workqueues - thaw workqueues
6851  *
6852  * Thaw workqueues.  Normal queueing is restored and all collected
6853  * frozen works are transferred to their respective pool worklists.
6854  *
6855  * CONTEXT:
6856  * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6857  */
6858 void thaw_workqueues(void)
6859 {
6860 	struct workqueue_struct *wq;
6861 
6862 	mutex_lock(&wq_pool_mutex);
6863 
6864 	if (!workqueue_freezing)
6865 		goto out_unlock;
6866 
6867 	workqueue_freezing = false;
6868 
6869 	/* restore max_active and repopulate worklist */
6870 	list_for_each_entry(wq, &workqueues, list) {
6871 		mutex_lock(&wq->mutex);
6872 		wq_adjust_max_active(wq);
6873 		mutex_unlock(&wq->mutex);
6874 	}
6875 
6876 out_unlock:
6877 	mutex_unlock(&wq_pool_mutex);
6878 }
6879 #endif /* CONFIG_FREEZER */
6880 
6881 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6882 {
6883 	LIST_HEAD(ctxs);
6884 	int ret = 0;
6885 	struct workqueue_struct *wq;
6886 	struct apply_wqattrs_ctx *ctx, *n;
6887 
6888 	lockdep_assert_held(&wq_pool_mutex);
6889 
6890 	list_for_each_entry(wq, &workqueues, list) {
6891 		if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING))
6892 			continue;
6893 
6894 		ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
6895 		if (IS_ERR(ctx)) {
6896 			ret = PTR_ERR(ctx);
6897 			break;
6898 		}
6899 
6900 		list_add_tail(&ctx->list, &ctxs);
6901 	}
6902 
6903 	list_for_each_entry_safe(ctx, n, &ctxs, list) {
6904 		if (!ret)
6905 			apply_wqattrs_commit(ctx);
6906 		apply_wqattrs_cleanup(ctx);
6907 	}
6908 
6909 	if (!ret) {
6910 		mutex_lock(&wq_pool_attach_mutex);
6911 		cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
6912 		mutex_unlock(&wq_pool_attach_mutex);
6913 	}
6914 	return ret;
6915 }
6916 
6917 /**
6918  * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
6919  * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
6920  *
6921  * This function can be called from cpuset code to provide a set of isolated
6922  * CPUs that should be excluded from wq_unbound_cpumask.
6923  */
6924 int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
6925 {
6926 	cpumask_var_t cpumask;
6927 	int ret = 0;
6928 
6929 	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6930 		return -ENOMEM;
6931 
6932 	mutex_lock(&wq_pool_mutex);
6933 
6934 	/*
6935 	 * If the operation fails, it will fall back to
6936 	 * wq_requested_unbound_cpumask which is initially set to
6937 	 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
6938 	 * by any subsequent write to workqueue/cpumask sysfs file.
6939 	 */
6940 	if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
6941 		cpumask_copy(cpumask, wq_requested_unbound_cpumask);
6942 	if (!cpumask_equal(cpumask, wq_unbound_cpumask))
6943 		ret = workqueue_apply_unbound_cpumask(cpumask);
6944 
6945 	/* Save the current isolated cpumask & export it via sysfs */
6946 	if (!ret)
6947 		cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
6948 
6949 	mutex_unlock(&wq_pool_mutex);
6950 	free_cpumask_var(cpumask);
6951 	return ret;
6952 }
6953 
6954 static int parse_affn_scope(const char *val)
6955 {
6956 	int i;
6957 
6958 	for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6959 		if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
6960 			return i;
6961 	}
6962 	return -EINVAL;
6963 }
6964 
6965 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
6966 {
6967 	struct workqueue_struct *wq;
6968 	int affn, cpu;
6969 
6970 	affn = parse_affn_scope(val);
6971 	if (affn < 0)
6972 		return affn;
6973 	if (affn == WQ_AFFN_DFL)
6974 		return -EINVAL;
6975 
6976 	cpus_read_lock();
6977 	mutex_lock(&wq_pool_mutex);
6978 
6979 	wq_affn_dfl = affn;
6980 
6981 	list_for_each_entry(wq, &workqueues, list) {
6982 		for_each_online_cpu(cpu)
6983 			unbound_wq_update_pwq(wq, cpu);
6984 	}
6985 
6986 	mutex_unlock(&wq_pool_mutex);
6987 	cpus_read_unlock();
6988 
6989 	return 0;
6990 }
6991 
6992 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
6993 {
6994 	return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
6995 }
6996 
6997 static const struct kernel_param_ops wq_affn_dfl_ops = {
6998 	.set	= wq_affn_dfl_set,
6999 	.get	= wq_affn_dfl_get,
7000 };
7001 
7002 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
7003 
7004 #ifdef CONFIG_SYSFS
7005 /*
7006  * Workqueues with WQ_SYSFS flag set is visible to userland via
7007  * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
7008  * following attributes.
7009  *
7010  *  per_cpu		RO bool	: whether the workqueue is per-cpu or unbound
7011  *  max_active		RW int	: maximum number of in-flight work items
7012  *
7013  * Unbound workqueues have the following extra attributes.
7014  *
7015  *  nice		RW int	: nice value of the workers
7016  *  cpumask		RW mask	: bitmask of allowed CPUs for the workers
7017  *  affinity_scope	RW str  : worker CPU affinity scope (cache, numa, none)
7018  *  affinity_strict	RW bool : worker CPU affinity is strict
7019  */
7020 struct wq_device {
7021 	struct workqueue_struct		*wq;
7022 	struct device			dev;
7023 };
7024 
7025 static struct workqueue_struct *dev_to_wq(struct device *dev)
7026 {
7027 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7028 
7029 	return wq_dev->wq;
7030 }
7031 
7032 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
7033 			    char *buf)
7034 {
7035 	struct workqueue_struct *wq = dev_to_wq(dev);
7036 
7037 	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
7038 }
7039 static DEVICE_ATTR_RO(per_cpu);
7040 
7041 static ssize_t max_active_show(struct device *dev,
7042 			       struct device_attribute *attr, char *buf)
7043 {
7044 	struct workqueue_struct *wq = dev_to_wq(dev);
7045 
7046 	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
7047 }
7048 
7049 static ssize_t max_active_store(struct device *dev,
7050 				struct device_attribute *attr, const char *buf,
7051 				size_t count)
7052 {
7053 	struct workqueue_struct *wq = dev_to_wq(dev);
7054 	int val;
7055 
7056 	if (sscanf(buf, "%d", &val) != 1 || val <= 0)
7057 		return -EINVAL;
7058 
7059 	workqueue_set_max_active(wq, val);
7060 	return count;
7061 }
7062 static DEVICE_ATTR_RW(max_active);
7063 
7064 static struct attribute *wq_sysfs_attrs[] = {
7065 	&dev_attr_per_cpu.attr,
7066 	&dev_attr_max_active.attr,
7067 	NULL,
7068 };
7069 ATTRIBUTE_GROUPS(wq_sysfs);
7070 
7071 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
7072 			    char *buf)
7073 {
7074 	struct workqueue_struct *wq = dev_to_wq(dev);
7075 	int written;
7076 
7077 	mutex_lock(&wq->mutex);
7078 	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
7079 	mutex_unlock(&wq->mutex);
7080 
7081 	return written;
7082 }
7083 
7084 /* prepare workqueue_attrs for sysfs store operations */
7085 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
7086 {
7087 	struct workqueue_attrs *attrs;
7088 
7089 	lockdep_assert_held(&wq_pool_mutex);
7090 
7091 	attrs = alloc_workqueue_attrs();
7092 	if (!attrs)
7093 		return NULL;
7094 
7095 	copy_workqueue_attrs(attrs, wq->unbound_attrs);
7096 	return attrs;
7097 }
7098 
7099 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
7100 			     const char *buf, size_t count)
7101 {
7102 	struct workqueue_struct *wq = dev_to_wq(dev);
7103 	struct workqueue_attrs *attrs;
7104 	int ret = -ENOMEM;
7105 
7106 	apply_wqattrs_lock();
7107 
7108 	attrs = wq_sysfs_prep_attrs(wq);
7109 	if (!attrs)
7110 		goto out_unlock;
7111 
7112 	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
7113 	    attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
7114 		ret = apply_workqueue_attrs_locked(wq, attrs);
7115 	else
7116 		ret = -EINVAL;
7117 
7118 out_unlock:
7119 	apply_wqattrs_unlock();
7120 	free_workqueue_attrs(attrs);
7121 	return ret ?: count;
7122 }
7123 
7124 static ssize_t wq_cpumask_show(struct device *dev,
7125 			       struct device_attribute *attr, char *buf)
7126 {
7127 	struct workqueue_struct *wq = dev_to_wq(dev);
7128 	int written;
7129 
7130 	mutex_lock(&wq->mutex);
7131 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
7132 			    cpumask_pr_args(wq->unbound_attrs->cpumask));
7133 	mutex_unlock(&wq->mutex);
7134 	return written;
7135 }
7136 
7137 static ssize_t wq_cpumask_store(struct device *dev,
7138 				struct device_attribute *attr,
7139 				const char *buf, size_t count)
7140 {
7141 	struct workqueue_struct *wq = dev_to_wq(dev);
7142 	struct workqueue_attrs *attrs;
7143 	int ret = -ENOMEM;
7144 
7145 	apply_wqattrs_lock();
7146 
7147 	attrs = wq_sysfs_prep_attrs(wq);
7148 	if (!attrs)
7149 		goto out_unlock;
7150 
7151 	ret = cpumask_parse(buf, attrs->cpumask);
7152 	if (!ret)
7153 		ret = apply_workqueue_attrs_locked(wq, attrs);
7154 
7155 out_unlock:
7156 	apply_wqattrs_unlock();
7157 	free_workqueue_attrs(attrs);
7158 	return ret ?: count;
7159 }
7160 
7161 static ssize_t wq_affn_scope_show(struct device *dev,
7162 				  struct device_attribute *attr, char *buf)
7163 {
7164 	struct workqueue_struct *wq = dev_to_wq(dev);
7165 	int written;
7166 
7167 	mutex_lock(&wq->mutex);
7168 	if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
7169 		written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
7170 				    wq_affn_names[WQ_AFFN_DFL],
7171 				    wq_affn_names[wq_affn_dfl]);
7172 	else
7173 		written = scnprintf(buf, PAGE_SIZE, "%s\n",
7174 				    wq_affn_names[wq->unbound_attrs->affn_scope]);
7175 	mutex_unlock(&wq->mutex);
7176 
7177 	return written;
7178 }
7179 
7180 static ssize_t wq_affn_scope_store(struct device *dev,
7181 				   struct device_attribute *attr,
7182 				   const char *buf, size_t count)
7183 {
7184 	struct workqueue_struct *wq = dev_to_wq(dev);
7185 	struct workqueue_attrs *attrs;
7186 	int affn, ret = -ENOMEM;
7187 
7188 	affn = parse_affn_scope(buf);
7189 	if (affn < 0)
7190 		return affn;
7191 
7192 	apply_wqattrs_lock();
7193 	attrs = wq_sysfs_prep_attrs(wq);
7194 	if (attrs) {
7195 		attrs->affn_scope = affn;
7196 		ret = apply_workqueue_attrs_locked(wq, attrs);
7197 	}
7198 	apply_wqattrs_unlock();
7199 	free_workqueue_attrs(attrs);
7200 	return ret ?: count;
7201 }
7202 
7203 static ssize_t wq_affinity_strict_show(struct device *dev,
7204 				       struct device_attribute *attr, char *buf)
7205 {
7206 	struct workqueue_struct *wq = dev_to_wq(dev);
7207 
7208 	return scnprintf(buf, PAGE_SIZE, "%d\n",
7209 			 wq->unbound_attrs->affn_strict);
7210 }
7211 
7212 static ssize_t wq_affinity_strict_store(struct device *dev,
7213 					struct device_attribute *attr,
7214 					const char *buf, size_t count)
7215 {
7216 	struct workqueue_struct *wq = dev_to_wq(dev);
7217 	struct workqueue_attrs *attrs;
7218 	int v, ret = -ENOMEM;
7219 
7220 	if (sscanf(buf, "%d", &v) != 1)
7221 		return -EINVAL;
7222 
7223 	apply_wqattrs_lock();
7224 	attrs = wq_sysfs_prep_attrs(wq);
7225 	if (attrs) {
7226 		attrs->affn_strict = (bool)v;
7227 		ret = apply_workqueue_attrs_locked(wq, attrs);
7228 	}
7229 	apply_wqattrs_unlock();
7230 	free_workqueue_attrs(attrs);
7231 	return ret ?: count;
7232 }
7233 
7234 static struct device_attribute wq_sysfs_unbound_attrs[] = {
7235 	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7236 	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7237 	__ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
7238 	__ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
7239 	__ATTR_NULL,
7240 };
7241 
7242 static const struct bus_type wq_subsys = {
7243 	.name				= "workqueue",
7244 	.dev_groups			= wq_sysfs_groups,
7245 };
7246 
7247 /**
7248  *  workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
7249  *  @cpumask: the cpumask to set
7250  *
7251  *  The low-level workqueues cpumask is a global cpumask that limits
7252  *  the affinity of all unbound workqueues.  This function check the @cpumask
7253  *  and apply it to all unbound workqueues and updates all pwqs of them.
7254  *
7255  *  Return:	0	- Success
7256  *		-EINVAL	- Invalid @cpumask
7257  *		-ENOMEM	- Failed to allocate memory for attrs or pwqs.
7258  */
7259 static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
7260 {
7261 	int ret = -EINVAL;
7262 
7263 	/*
7264 	 * Not excluding isolated cpus on purpose.
7265 	 * If the user wishes to include them, we allow that.
7266 	 */
7267 	cpumask_and(cpumask, cpumask, cpu_possible_mask);
7268 	if (!cpumask_empty(cpumask)) {
7269 		ret = 0;
7270 		apply_wqattrs_lock();
7271 		if (!cpumask_equal(cpumask, wq_unbound_cpumask))
7272 			ret = workqueue_apply_unbound_cpumask(cpumask);
7273 		if (!ret)
7274 			cpumask_copy(wq_requested_unbound_cpumask, cpumask);
7275 		apply_wqattrs_unlock();
7276 	}
7277 
7278 	return ret;
7279 }
7280 
7281 static ssize_t __wq_cpumask_show(struct device *dev,
7282 		struct device_attribute *attr, char *buf, cpumask_var_t mask)
7283 {
7284 	int written;
7285 
7286 	mutex_lock(&wq_pool_mutex);
7287 	written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
7288 	mutex_unlock(&wq_pool_mutex);
7289 
7290 	return written;
7291 }
7292 
7293 static ssize_t cpumask_requested_show(struct device *dev,
7294 		struct device_attribute *attr, char *buf)
7295 {
7296 	return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
7297 }
7298 static DEVICE_ATTR_RO(cpumask_requested);
7299 
7300 static ssize_t cpumask_isolated_show(struct device *dev,
7301 		struct device_attribute *attr, char *buf)
7302 {
7303 	return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
7304 }
7305 static DEVICE_ATTR_RO(cpumask_isolated);
7306 
7307 static ssize_t cpumask_show(struct device *dev,
7308 		struct device_attribute *attr, char *buf)
7309 {
7310 	return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
7311 }
7312 
7313 static ssize_t cpumask_store(struct device *dev,
7314 		struct device_attribute *attr, const char *buf, size_t count)
7315 {
7316 	cpumask_var_t cpumask;
7317 	int ret;
7318 
7319 	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7320 		return -ENOMEM;
7321 
7322 	ret = cpumask_parse(buf, cpumask);
7323 	if (!ret)
7324 		ret = workqueue_set_unbound_cpumask(cpumask);
7325 
7326 	free_cpumask_var(cpumask);
7327 	return ret ? ret : count;
7328 }
7329 static DEVICE_ATTR_RW(cpumask);
7330 
7331 static struct attribute *wq_sysfs_cpumask_attrs[] = {
7332 	&dev_attr_cpumask.attr,
7333 	&dev_attr_cpumask_requested.attr,
7334 	&dev_attr_cpumask_isolated.attr,
7335 	NULL,
7336 };
7337 ATTRIBUTE_GROUPS(wq_sysfs_cpumask);
7338 
7339 static int __init wq_sysfs_init(void)
7340 {
7341 	return subsys_virtual_register(&wq_subsys, wq_sysfs_cpumask_groups);
7342 }
7343 core_initcall(wq_sysfs_init);
7344 
7345 static void wq_device_release(struct device *dev)
7346 {
7347 	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7348 
7349 	kfree(wq_dev);
7350 }
7351 
7352 /**
7353  * workqueue_sysfs_register - make a workqueue visible in sysfs
7354  * @wq: the workqueue to register
7355  *
7356  * Expose @wq in sysfs under /sys/bus/workqueue/devices.
7357  * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7358  * which is the preferred method.
7359  *
7360  * Workqueue user should use this function directly iff it wants to apply
7361  * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7362  * apply_workqueue_attrs() may race against userland updating the
7363  * attributes.
7364  *
7365  * Return: 0 on success, -errno on failure.
7366  */
7367 int workqueue_sysfs_register(struct workqueue_struct *wq)
7368 {
7369 	struct wq_device *wq_dev;
7370 	int ret;
7371 
7372 	/*
7373 	 * Adjusting max_active breaks ordering guarantee.  Disallow exposing
7374 	 * ordered workqueues.
7375 	 */
7376 	if (WARN_ON(wq->flags & __WQ_ORDERED))
7377 		return -EINVAL;
7378 
7379 	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7380 	if (!wq_dev)
7381 		return -ENOMEM;
7382 
7383 	wq_dev->wq = wq;
7384 	wq_dev->dev.bus = &wq_subsys;
7385 	wq_dev->dev.release = wq_device_release;
7386 	dev_set_name(&wq_dev->dev, "%s", wq->name);
7387 
7388 	/*
7389 	 * unbound_attrs are created separately.  Suppress uevent until
7390 	 * everything is ready.
7391 	 */
7392 	dev_set_uevent_suppress(&wq_dev->dev, true);
7393 
7394 	ret = device_register(&wq_dev->dev);
7395 	if (ret) {
7396 		put_device(&wq_dev->dev);
7397 		wq->wq_dev = NULL;
7398 		return ret;
7399 	}
7400 
7401 	if (wq->flags & WQ_UNBOUND) {
7402 		struct device_attribute *attr;
7403 
7404 		for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7405 			ret = device_create_file(&wq_dev->dev, attr);
7406 			if (ret) {
7407 				device_unregister(&wq_dev->dev);
7408 				wq->wq_dev = NULL;
7409 				return ret;
7410 			}
7411 		}
7412 	}
7413 
7414 	dev_set_uevent_suppress(&wq_dev->dev, false);
7415 	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7416 	return 0;
7417 }
7418 
7419 /**
7420  * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7421  * @wq: the workqueue to unregister
7422  *
7423  * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7424  */
7425 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7426 {
7427 	struct wq_device *wq_dev = wq->wq_dev;
7428 
7429 	if (!wq->wq_dev)
7430 		return;
7431 
7432 	wq->wq_dev = NULL;
7433 	device_unregister(&wq_dev->dev);
7434 }
7435 #else	/* CONFIG_SYSFS */
7436 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)	{ }
7437 #endif	/* CONFIG_SYSFS */
7438 
7439 /*
7440  * Workqueue watchdog.
7441  *
7442  * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7443  * flush dependency, a concurrency managed work item which stays RUNNING
7444  * indefinitely.  Workqueue stalls can be very difficult to debug as the
7445  * usual warning mechanisms don't trigger and internal workqueue state is
7446  * largely opaque.
7447  *
7448  * Workqueue watchdog monitors all worker pools periodically and dumps
7449  * state if some pools failed to make forward progress for a while where
7450  * forward progress is defined as the first item on ->worklist changing.
7451  *
7452  * This mechanism is controlled through the kernel parameter
7453  * "workqueue.watchdog_thresh" which can be updated at runtime through the
7454  * corresponding sysfs parameter file.
7455  */
7456 #ifdef CONFIG_WQ_WATCHDOG
7457 
7458 static unsigned long wq_watchdog_thresh = 30;
7459 static struct timer_list wq_watchdog_timer;
7460 
7461 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7462 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7463 
7464 static unsigned int wq_panic_on_stall;
7465 module_param_named(panic_on_stall, wq_panic_on_stall, uint, 0644);
7466 
7467 /*
7468  * Show workers that might prevent the processing of pending work items.
7469  * The only candidates are CPU-bound workers in the running state.
7470  * Pending work items should be handled by another idle worker
7471  * in all other situations.
7472  */
7473 static void show_cpu_pool_hog(struct worker_pool *pool)
7474 {
7475 	struct worker *worker;
7476 	unsigned long irq_flags;
7477 	int bkt;
7478 
7479 	raw_spin_lock_irqsave(&pool->lock, irq_flags);
7480 
7481 	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
7482 		if (task_is_running(worker->task)) {
7483 			/*
7484 			 * Defer printing to avoid deadlocks in console
7485 			 * drivers that queue work while holding locks
7486 			 * also taken in their write paths.
7487 			 */
7488 			printk_deferred_enter();
7489 
7490 			pr_info("pool %d:\n", pool->id);
7491 			sched_show_task(worker->task);
7492 
7493 			printk_deferred_exit();
7494 		}
7495 	}
7496 
7497 	raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
7498 }
7499 
7500 static void show_cpu_pools_hogs(void)
7501 {
7502 	struct worker_pool *pool;
7503 	int pi;
7504 
7505 	pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
7506 
7507 	rcu_read_lock();
7508 
7509 	for_each_pool(pool, pi) {
7510 		if (pool->cpu_stall)
7511 			show_cpu_pool_hog(pool);
7512 
7513 	}
7514 
7515 	rcu_read_unlock();
7516 }
7517 
7518 static void panic_on_wq_watchdog(void)
7519 {
7520 	static unsigned int wq_stall;
7521 
7522 	if (wq_panic_on_stall) {
7523 		wq_stall++;
7524 		BUG_ON(wq_stall >= wq_panic_on_stall);
7525 	}
7526 }
7527 
7528 static void wq_watchdog_reset_touched(void)
7529 {
7530 	int cpu;
7531 
7532 	wq_watchdog_touched = jiffies;
7533 	for_each_possible_cpu(cpu)
7534 		per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7535 }
7536 
7537 static void wq_watchdog_timer_fn(struct timer_list *unused)
7538 {
7539 	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7540 	bool lockup_detected = false;
7541 	bool cpu_pool_stall = false;
7542 	unsigned long now = jiffies;
7543 	struct worker_pool *pool;
7544 	int pi;
7545 
7546 	if (!thresh)
7547 		return;
7548 
7549 	rcu_read_lock();
7550 
7551 	for_each_pool(pool, pi) {
7552 		unsigned long pool_ts, touched, ts;
7553 
7554 		pool->cpu_stall = false;
7555 		if (list_empty(&pool->worklist))
7556 			continue;
7557 
7558 		/*
7559 		 * If a virtual machine is stopped by the host it can look to
7560 		 * the watchdog like a stall.
7561 		 */
7562 		kvm_check_and_clear_guest_paused();
7563 
7564 		/* get the latest of pool and touched timestamps */
7565 		if (pool->cpu >= 0)
7566 			touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7567 		else
7568 			touched = READ_ONCE(wq_watchdog_touched);
7569 		pool_ts = READ_ONCE(pool->watchdog_ts);
7570 
7571 		if (time_after(pool_ts, touched))
7572 			ts = pool_ts;
7573 		else
7574 			ts = touched;
7575 
7576 		/* did we stall? */
7577 		if (time_after(now, ts + thresh)) {
7578 			lockup_detected = true;
7579 			if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) {
7580 				pool->cpu_stall = true;
7581 				cpu_pool_stall = true;
7582 			}
7583 			pr_emerg("BUG: workqueue lockup - pool");
7584 			pr_cont_pool_info(pool);
7585 			pr_cont(" stuck for %us!\n",
7586 				jiffies_to_msecs(now - pool_ts) / 1000);
7587 		}
7588 
7589 
7590 	}
7591 
7592 	rcu_read_unlock();
7593 
7594 	if (lockup_detected)
7595 		show_all_workqueues();
7596 
7597 	if (cpu_pool_stall)
7598 		show_cpu_pools_hogs();
7599 
7600 	if (lockup_detected)
7601 		panic_on_wq_watchdog();
7602 
7603 	wq_watchdog_reset_touched();
7604 	mod_timer(&wq_watchdog_timer, jiffies + thresh);
7605 }
7606 
7607 notrace void wq_watchdog_touch(int cpu)
7608 {
7609 	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7610 	unsigned long touch_ts = READ_ONCE(wq_watchdog_touched);
7611 	unsigned long now = jiffies;
7612 
7613 	if (cpu >= 0)
7614 		per_cpu(wq_watchdog_touched_cpu, cpu) = now;
7615 	else
7616 		WARN_ONCE(1, "%s should be called with valid CPU", __func__);
7617 
7618 	/* Don't unnecessarily store to global cacheline */
7619 	if (time_after(now, touch_ts + thresh / 4))
7620 		WRITE_ONCE(wq_watchdog_touched, jiffies);
7621 }
7622 
7623 static void wq_watchdog_set_thresh(unsigned long thresh)
7624 {
7625 	wq_watchdog_thresh = 0;
7626 	del_timer_sync(&wq_watchdog_timer);
7627 
7628 	if (thresh) {
7629 		wq_watchdog_thresh = thresh;
7630 		wq_watchdog_reset_touched();
7631 		mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7632 	}
7633 }
7634 
7635 static int wq_watchdog_param_set_thresh(const char *val,
7636 					const struct kernel_param *kp)
7637 {
7638 	unsigned long thresh;
7639 	int ret;
7640 
7641 	ret = kstrtoul(val, 0, &thresh);
7642 	if (ret)
7643 		return ret;
7644 
7645 	if (system_wq)
7646 		wq_watchdog_set_thresh(thresh);
7647 	else
7648 		wq_watchdog_thresh = thresh;
7649 
7650 	return 0;
7651 }
7652 
7653 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7654 	.set	= wq_watchdog_param_set_thresh,
7655 	.get	= param_get_ulong,
7656 };
7657 
7658 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7659 		0644);
7660 
7661 static void wq_watchdog_init(void)
7662 {
7663 	timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7664 	wq_watchdog_set_thresh(wq_watchdog_thresh);
7665 }
7666 
7667 #else	/* CONFIG_WQ_WATCHDOG */
7668 
7669 static inline void wq_watchdog_init(void) { }
7670 
7671 #endif	/* CONFIG_WQ_WATCHDOG */
7672 
7673 static void bh_pool_kick_normal(struct irq_work *irq_work)
7674 {
7675 	raise_softirq_irqoff(TASKLET_SOFTIRQ);
7676 }
7677 
7678 static void bh_pool_kick_highpri(struct irq_work *irq_work)
7679 {
7680 	raise_softirq_irqoff(HI_SOFTIRQ);
7681 }
7682 
7683 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
7684 {
7685 	if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
7686 		pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
7687 			cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7688 		return;
7689 	}
7690 
7691 	cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
7692 }
7693 
7694 static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice)
7695 {
7696 	BUG_ON(init_worker_pool(pool));
7697 	pool->cpu = cpu;
7698 	cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7699 	cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
7700 	pool->attrs->nice = nice;
7701 	pool->attrs->affn_strict = true;
7702 	pool->node = cpu_to_node(cpu);
7703 
7704 	/* alloc pool ID */
7705 	mutex_lock(&wq_pool_mutex);
7706 	BUG_ON(worker_pool_assign_id(pool));
7707 	mutex_unlock(&wq_pool_mutex);
7708 }
7709 
7710 /**
7711  * workqueue_init_early - early init for workqueue subsystem
7712  *
7713  * This is the first step of three-staged workqueue subsystem initialization and
7714  * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7715  * up. It sets up all the data structures and system workqueues and allows early
7716  * boot code to create workqueues and queue/cancel work items. Actual work item
7717  * execution starts only after kthreads can be created and scheduled right
7718  * before early initcalls.
7719  */
7720 void __init workqueue_init_early(void)
7721 {
7722 	struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7723 	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7724 	void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal,
7725 						       bh_pool_kick_highpri };
7726 	int i, cpu;
7727 
7728 	BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7729 
7730 	BUG_ON(!alloc_cpumask_var(&wq_online_cpumask, GFP_KERNEL));
7731 	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7732 	BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
7733 	BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
7734 
7735 	cpumask_copy(wq_online_cpumask, cpu_online_mask);
7736 	cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
7737 	restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
7738 	restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
7739 	if (!cpumask_empty(&wq_cmdline_cpumask))
7740 		restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
7741 
7742 	cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
7743 
7744 	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7745 
7746 	unbound_wq_update_pwq_attrs_buf = alloc_workqueue_attrs();
7747 	BUG_ON(!unbound_wq_update_pwq_attrs_buf);
7748 
7749 	/*
7750 	 * If nohz_full is enabled, set power efficient workqueue as unbound.
7751 	 * This allows workqueue items to be moved to HK CPUs.
7752 	 */
7753 	if (housekeeping_enabled(HK_TYPE_TICK))
7754 		wq_power_efficient = true;
7755 
7756 	/* initialize WQ_AFFN_SYSTEM pods */
7757 	pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7758 	pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7759 	pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7760 	BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
7761 
7762 	BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7763 
7764 	pt->nr_pods = 1;
7765 	cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
7766 	pt->pod_node[0] = NUMA_NO_NODE;
7767 	pt->cpu_pod[0] = 0;
7768 
7769 	/* initialize BH and CPU pools */
7770 	for_each_possible_cpu(cpu) {
7771 		struct worker_pool *pool;
7772 
7773 		i = 0;
7774 		for_each_bh_worker_pool(pool, cpu) {
7775 			init_cpu_worker_pool(pool, cpu, std_nice[i]);
7776 			pool->flags |= POOL_BH;
7777 			init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]);
7778 			i++;
7779 		}
7780 
7781 		i = 0;
7782 		for_each_cpu_worker_pool(pool, cpu)
7783 			init_cpu_worker_pool(pool, cpu, std_nice[i++]);
7784 	}
7785 
7786 	/* create default unbound and ordered wq attrs */
7787 	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7788 		struct workqueue_attrs *attrs;
7789 
7790 		BUG_ON(!(attrs = alloc_workqueue_attrs()));
7791 		attrs->nice = std_nice[i];
7792 		unbound_std_wq_attrs[i] = attrs;
7793 
7794 		/*
7795 		 * An ordered wq should have only one pwq as ordering is
7796 		 * guaranteed by max_active which is enforced by pwqs.
7797 		 */
7798 		BUG_ON(!(attrs = alloc_workqueue_attrs()));
7799 		attrs->nice = std_nice[i];
7800 		attrs->ordered = true;
7801 		ordered_wq_attrs[i] = attrs;
7802 	}
7803 
7804 	system_wq = alloc_workqueue("events", 0, 0);
7805 	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7806 	system_long_wq = alloc_workqueue("events_long", 0, 0);
7807 	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7808 					    WQ_MAX_ACTIVE);
7809 	system_freezable_wq = alloc_workqueue("events_freezable",
7810 					      WQ_FREEZABLE, 0);
7811 	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7812 					      WQ_POWER_EFFICIENT, 0);
7813 	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
7814 					      WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7815 					      0);
7816 	system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0);
7817 	system_bh_highpri_wq = alloc_workqueue("events_bh_highpri",
7818 					       WQ_BH | WQ_HIGHPRI, 0);
7819 	BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7820 	       !system_unbound_wq || !system_freezable_wq ||
7821 	       !system_power_efficient_wq ||
7822 	       !system_freezable_power_efficient_wq ||
7823 	       !system_bh_wq || !system_bh_highpri_wq);
7824 }
7825 
7826 static void __init wq_cpu_intensive_thresh_init(void)
7827 {
7828 	unsigned long thresh;
7829 	unsigned long bogo;
7830 
7831 	pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
7832 	BUG_ON(IS_ERR(pwq_release_worker));
7833 
7834 	/* if the user set it to a specific value, keep it */
7835 	if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7836 		return;
7837 
7838 	/*
7839 	 * The default of 10ms is derived from the fact that most modern (as of
7840 	 * 2023) processors can do a lot in 10ms and that it's just below what
7841 	 * most consider human-perceivable. However, the kernel also runs on a
7842 	 * lot slower CPUs including microcontrollers where the threshold is way
7843 	 * too low.
7844 	 *
7845 	 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7846 	 * This is by no means accurate but it doesn't have to be. The mechanism
7847 	 * is still useful even when the threshold is fully scaled up. Also, as
7848 	 * the reports would usually be applicable to everyone, some machines
7849 	 * operating on longer thresholds won't significantly diminish their
7850 	 * usefulness.
7851 	 */
7852 	thresh = 10 * USEC_PER_MSEC;
7853 
7854 	/* see init/calibrate.c for lpj -> BogoMIPS calculation */
7855 	bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7856 	if (bogo < 4000)
7857 		thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7858 
7859 	pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7860 		 loops_per_jiffy, bogo, thresh);
7861 
7862 	wq_cpu_intensive_thresh_us = thresh;
7863 }
7864 
7865 /**
7866  * workqueue_init - bring workqueue subsystem fully online
7867  *
7868  * This is the second step of three-staged workqueue subsystem initialization
7869  * and invoked as soon as kthreads can be created and scheduled. Workqueues have
7870  * been created and work items queued on them, but there are no kworkers
7871  * executing the work items yet. Populate the worker pools with the initial
7872  * workers and enable future kworker creations.
7873  */
7874 void __init workqueue_init(void)
7875 {
7876 	struct workqueue_struct *wq;
7877 	struct worker_pool *pool;
7878 	int cpu, bkt;
7879 
7880 	wq_cpu_intensive_thresh_init();
7881 
7882 	mutex_lock(&wq_pool_mutex);
7883 
7884 	/*
7885 	 * Per-cpu pools created earlier could be missing node hint. Fix them
7886 	 * up. Also, create a rescuer for workqueues that requested it.
7887 	 */
7888 	for_each_possible_cpu(cpu) {
7889 		for_each_bh_worker_pool(pool, cpu)
7890 			pool->node = cpu_to_node(cpu);
7891 		for_each_cpu_worker_pool(pool, cpu)
7892 			pool->node = cpu_to_node(cpu);
7893 	}
7894 
7895 	list_for_each_entry(wq, &workqueues, list) {
7896 		WARN(init_rescuer(wq),
7897 		     "workqueue: failed to create early rescuer for %s",
7898 		     wq->name);
7899 	}
7900 
7901 	mutex_unlock(&wq_pool_mutex);
7902 
7903 	/*
7904 	 * Create the initial workers. A BH pool has one pseudo worker that
7905 	 * represents the shared BH execution context and thus doesn't get
7906 	 * affected by hotplug events. Create the BH pseudo workers for all
7907 	 * possible CPUs here.
7908 	 */
7909 	for_each_possible_cpu(cpu)
7910 		for_each_bh_worker_pool(pool, cpu)
7911 			BUG_ON(!create_worker(pool));
7912 
7913 	for_each_online_cpu(cpu) {
7914 		for_each_cpu_worker_pool(pool, cpu) {
7915 			pool->flags &= ~POOL_DISASSOCIATED;
7916 			BUG_ON(!create_worker(pool));
7917 		}
7918 	}
7919 
7920 	hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7921 		BUG_ON(!create_worker(pool));
7922 
7923 	wq_online = true;
7924 	wq_watchdog_init();
7925 }
7926 
7927 /*
7928  * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7929  * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7930  * and consecutive pod ID. The rest of @pt is initialized accordingly.
7931  */
7932 static void __init init_pod_type(struct wq_pod_type *pt,
7933 				 bool (*cpus_share_pod)(int, int))
7934 {
7935 	int cur, pre, cpu, pod;
7936 
7937 	pt->nr_pods = 0;
7938 
7939 	/* init @pt->cpu_pod[] according to @cpus_share_pod() */
7940 	pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7941 	BUG_ON(!pt->cpu_pod);
7942 
7943 	for_each_possible_cpu(cur) {
7944 		for_each_possible_cpu(pre) {
7945 			if (pre >= cur) {
7946 				pt->cpu_pod[cur] = pt->nr_pods++;
7947 				break;
7948 			}
7949 			if (cpus_share_pod(cur, pre)) {
7950 				pt->cpu_pod[cur] = pt->cpu_pod[pre];
7951 				break;
7952 			}
7953 		}
7954 	}
7955 
7956 	/* init the rest to match @pt->cpu_pod[] */
7957 	pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7958 	pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7959 	BUG_ON(!pt->pod_cpus || !pt->pod_node);
7960 
7961 	for (pod = 0; pod < pt->nr_pods; pod++)
7962 		BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7963 
7964 	for_each_possible_cpu(cpu) {
7965 		cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
7966 		pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7967 	}
7968 }
7969 
7970 static bool __init cpus_dont_share(int cpu0, int cpu1)
7971 {
7972 	return false;
7973 }
7974 
7975 static bool __init cpus_share_smt(int cpu0, int cpu1)
7976 {
7977 #ifdef CONFIG_SCHED_SMT
7978 	return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
7979 #else
7980 	return false;
7981 #endif
7982 }
7983 
7984 static bool __init cpus_share_numa(int cpu0, int cpu1)
7985 {
7986 	return cpu_to_node(cpu0) == cpu_to_node(cpu1);
7987 }
7988 
7989 /**
7990  * workqueue_init_topology - initialize CPU pods for unbound workqueues
7991  *
7992  * This is the third step of three-staged workqueue subsystem initialization and
7993  * invoked after SMP and topology information are fully initialized. It
7994  * initializes the unbound CPU pods accordingly.
7995  */
7996 void __init workqueue_init_topology(void)
7997 {
7998 	struct workqueue_struct *wq;
7999 	int cpu;
8000 
8001 	init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
8002 	init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
8003 	init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
8004 	init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
8005 
8006 	wq_topo_initialized = true;
8007 
8008 	mutex_lock(&wq_pool_mutex);
8009 
8010 	/*
8011 	 * Workqueues allocated earlier would have all CPUs sharing the default
8012 	 * worker pool. Explicitly call unbound_wq_update_pwq() on all workqueue
8013 	 * and CPU combinations to apply per-pod sharing.
8014 	 */
8015 	list_for_each_entry(wq, &workqueues, list) {
8016 		for_each_online_cpu(cpu)
8017 			unbound_wq_update_pwq(wq, cpu);
8018 		if (wq->flags & WQ_UNBOUND) {
8019 			mutex_lock(&wq->mutex);
8020 			wq_update_node_max_active(wq, -1);
8021 			mutex_unlock(&wq->mutex);
8022 		}
8023 	}
8024 
8025 	mutex_unlock(&wq_pool_mutex);
8026 }
8027 
8028 void __warn_flushing_systemwide_wq(void)
8029 {
8030 	pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
8031 	dump_stack();
8032 }
8033 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
8034 
8035 static int __init workqueue_unbound_cpus_setup(char *str)
8036 {
8037 	if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
8038 		cpumask_clear(&wq_cmdline_cpumask);
8039 		pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
8040 	}
8041 
8042 	return 1;
8043 }
8044 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);
8045