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
61enum 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
87enum 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
100enum work_cancel_flags {
101 WORK_CANCEL_DELAYED = 1 << 0, /* canceling a delayed_work */
102 WORK_CANCEL_DISABLE = 1 << 1, /* canceling to disable */
103};
104
105enum 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
184struct 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#ifdef CONFIG_PREEMPT_RT
226 spinlock_t cb_lock; /* BH worker cancel lock */
227#endif
228 /*
229 * Destruction of pool is RCU protected to allow dereferences
230 * from get_work_pool().
231 */
232 struct rcu_head rcu;
233};
234
235/*
236 * Per-pool_workqueue statistics. These can be monitored using
237 * tools/workqueue/wq_monitor.py.
238 */
239enum pool_workqueue_stats {
240 PWQ_STAT_STARTED, /* work items started execution */
241 PWQ_STAT_COMPLETED, /* work items completed execution */
242 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
243 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
244 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
245 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
246 PWQ_STAT_MAYDAY, /* maydays to rescuer */
247 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
248
249 PWQ_NR_STATS,
250};
251
252/*
253 * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT
254 * of work_struct->data are used for flags and the remaining high bits
255 * point to the pwq; thus, pwqs need to be aligned at two's power of the
256 * number of flag bits.
257 */
258struct pool_workqueue {
259 struct worker_pool *pool; /* I: the associated pool */
260 struct workqueue_struct *wq; /* I: the owning workqueue */
261 int work_color; /* L: current color */
262 int flush_color; /* L: flushing color */
263 int refcnt; /* L: reference count */
264 int nr_in_flight[WORK_NR_COLORS];
265 /* L: nr of in_flight works */
266 bool plugged; /* L: execution suspended */
267
268 /*
269 * nr_active management and WORK_STRUCT_INACTIVE:
270 *
271 * When pwq->nr_active >= max_active, new work item is queued to
272 * pwq->inactive_works instead of pool->worklist and marked with
273 * WORK_STRUCT_INACTIVE.
274 *
275 * All work items marked with WORK_STRUCT_INACTIVE do not participate in
276 * nr_active and all work items in pwq->inactive_works are marked with
277 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are
278 * in pwq->inactive_works. Some of them are ready to run in
279 * pool->worklist or worker->scheduled. Those work itmes are only struct
280 * wq_barrier which is used for flush_work() and should not participate
281 * in nr_active. For non-barrier work item, it is marked with
282 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
283 */
284 int nr_active; /* L: nr of active works */
285 struct list_head inactive_works; /* L: inactive works */
286 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */
287 struct list_head pwqs_node; /* WR: node on wq->pwqs */
288 struct list_head mayday_node; /* MD: node on wq->maydays */
289
290 u64 stats[PWQ_NR_STATS];
291
292 /*
293 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
294 * and pwq_release_workfn() for details. pool_workqueue itself is also
295 * RCU protected so that the first pwq can be determined without
296 * grabbing wq->mutex.
297 */
298 struct kthread_work release_work;
299 struct rcu_head rcu;
300} __aligned(1 << WORK_STRUCT_PWQ_SHIFT);
301
302/*
303 * Structure used to wait for workqueue flush.
304 */
305struct wq_flusher {
306 struct list_head list; /* WQ: list of flushers */
307 int flush_color; /* WQ: flush color waiting for */
308 struct completion done; /* flush completion */
309};
310
311struct wq_device;
312
313/*
314 * Unlike in a per-cpu workqueue where max_active limits its concurrency level
315 * on each CPU, in an unbound workqueue, max_active applies to the whole system.
316 * As sharing a single nr_active across multiple sockets can be very expensive,
317 * the counting and enforcement is per NUMA node.
318 *
319 * The following struct is used to enforce per-node max_active. When a pwq wants
320 * to start executing a work item, it should increment ->nr using
321 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over
322 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish
323 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in
324 * round-robin order.
325 */
326struct wq_node_nr_active {
327 int max; /* per-node max_active */
328 atomic_t nr; /* per-node nr_active */
329 raw_spinlock_t lock; /* nests inside pool locks */
330 struct list_head pending_pwqs; /* LN: pwqs with inactive works */
331};
332
333/*
334 * The externally visible workqueue. It relays the issued work items to
335 * the appropriate worker_pool through its pool_workqueues.
336 */
337struct workqueue_struct {
338 struct list_head pwqs; /* WR: all pwqs of this wq */
339 struct list_head list; /* PR: list of all workqueues */
340
341 struct mutex mutex; /* protects this wq */
342 int work_color; /* WQ: current work color */
343 int flush_color; /* WQ: current flush color */
344 atomic_t nr_pwqs_to_flush; /* flush in progress */
345 struct wq_flusher *first_flusher; /* WQ: first flusher */
346 struct list_head flusher_queue; /* WQ: flush waiters */
347 struct list_head flusher_overflow; /* WQ: flush overflow list */
348
349 struct list_head maydays; /* MD: pwqs requesting rescue */
350 struct worker *rescuer; /* MD: rescue worker */
351
352 int nr_drainers; /* WQ: drain in progress */
353
354 /* See alloc_workqueue() function comment for info on min/max_active */
355 int max_active; /* WO: max active works */
356 int min_active; /* WO: min active works */
357 int saved_max_active; /* WQ: saved max_active */
358 int saved_min_active; /* WQ: saved min_active */
359
360 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
361 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */
362
363#ifdef CONFIG_SYSFS
364 struct wq_device *wq_dev; /* I: for sysfs interface */
365#endif
366#ifdef CONFIG_LOCKDEP
367 char *lock_name;
368 struct lock_class_key key;
369 struct lockdep_map __lockdep_map;
370 struct lockdep_map *lockdep_map;
371#endif
372 char name[WQ_NAME_LEN]; /* I: workqueue name */
373
374 /*
375 * Destruction of workqueue_struct is RCU protected to allow walking
376 * the workqueues list without grabbing wq_pool_mutex.
377 * This is used to dump all workqueues from sysrq.
378 */
379 struct rcu_head rcu;
380
381 /* hot fields used during command issue, aligned to cacheline */
382 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
383 struct pool_workqueue __rcu * __percpu *cpu_pwq; /* I: per-cpu pwqs */
384 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */
385};
386
387/*
388 * Each pod type describes how CPUs should be grouped for unbound workqueues.
389 * See the comment above workqueue_attrs->affn_scope.
390 */
391struct wq_pod_type {
392 int nr_pods; /* number of pods */
393 cpumask_var_t *pod_cpus; /* pod -> cpus */
394 int *pod_node; /* pod -> node */
395 int *cpu_pod; /* cpu -> pod */
396};
397
398struct work_offq_data {
399 u32 pool_id;
400 u32 disable;
401 u32 flags;
402};
403
404static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
405 [WQ_AFFN_DFL] = "default",
406 [WQ_AFFN_CPU] = "cpu",
407 [WQ_AFFN_SMT] = "smt",
408 [WQ_AFFN_CACHE] = "cache",
409 [WQ_AFFN_NUMA] = "numa",
410 [WQ_AFFN_SYSTEM] = "system",
411};
412
413/*
414 * Per-cpu work items which run for longer than the following threshold are
415 * automatically considered CPU intensive and excluded from concurrency
416 * management to prevent them from noticeably delaying other per-cpu work items.
417 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
418 * The actual value is initialized in wq_cpu_intensive_thresh_init().
419 */
420static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
421module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
422#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
423static unsigned int wq_cpu_intensive_warning_thresh = 4;
424module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644);
425#endif
426
427/* see the comment above the definition of WQ_POWER_EFFICIENT */
428static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
429module_param_named(power_efficient, wq_power_efficient, bool, 0444);
430
431static bool wq_online; /* can kworkers be created yet? */
432static bool wq_topo_initialized __read_mostly = false;
433
434static struct kmem_cache *pwq_cache;
435
436static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
437static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
438
439/* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
440static struct workqueue_attrs *unbound_wq_update_pwq_attrs_buf;
441
442static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
443static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
444static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
445/* wait for manager to go away */
446static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
447
448static LIST_HEAD(workqueues); /* PR: list of all workqueues */
449static bool workqueue_freezing; /* PL: have wqs started freezing? */
450
451/* PL: mirror the cpu_online_mask excluding the CPU in the midst of hotplugging */
452static cpumask_var_t wq_online_cpumask;
453
454/* PL&A: allowable cpus for unbound wqs and work items */
455static cpumask_var_t wq_unbound_cpumask;
456
457/* PL: user requested unbound cpumask via sysfs */
458static cpumask_var_t wq_requested_unbound_cpumask;
459
460/* PL: isolated cpumask to be excluded from unbound cpumask */
461static cpumask_var_t wq_isolated_cpumask;
462
463/* for further constrain wq_unbound_cpumask by cmdline parameter*/
464static struct cpumask wq_cmdline_cpumask __initdata;
465
466/* CPU where unbound work was last round robin scheduled from this CPU */
467static DEFINE_PER_CPU(int, wq_rr_cpu_last);
468
469/*
470 * Local execution of unbound work items is no longer guaranteed. The
471 * following always forces round-robin CPU selection on unbound work items
472 * to uncover usages which depend on it.
473 */
474#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
475static bool wq_debug_force_rr_cpu = true;
476#else
477static bool wq_debug_force_rr_cpu = false;
478#endif
479module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
480
481/* to raise softirq for the BH worker pools on other CPUs */
482static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS], bh_pool_irq_works);
483
484/* the BH worker pools */
485static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], bh_worker_pools);
486
487/* the per-cpu worker pools */
488static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
489
490static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
491
492/* PL: hash of all unbound pools keyed by pool->attrs */
493static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
494
495/* I: attributes used when instantiating standard unbound pools on demand */
496static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
497
498/* I: attributes used when instantiating ordered pools on demand */
499static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
500
501/*
502 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
503 * process context while holding a pool lock. Bounce to a dedicated kthread
504 * worker to avoid A-A deadlocks.
505 */
506static struct kthread_worker *pwq_release_worker __ro_after_init;
507
508struct workqueue_struct *system_wq __ro_after_init;
509EXPORT_SYMBOL(system_wq);
510struct workqueue_struct *system_percpu_wq __ro_after_init;
511EXPORT_SYMBOL(system_percpu_wq);
512struct workqueue_struct *system_highpri_wq __ro_after_init;
513EXPORT_SYMBOL_GPL(system_highpri_wq);
514struct workqueue_struct *system_long_wq __ro_after_init;
515EXPORT_SYMBOL_GPL(system_long_wq);
516struct workqueue_struct *system_unbound_wq __ro_after_init;
517EXPORT_SYMBOL_GPL(system_unbound_wq);
518struct workqueue_struct *system_dfl_wq __ro_after_init;
519EXPORT_SYMBOL_GPL(system_dfl_wq);
520struct workqueue_struct *system_freezable_wq __ro_after_init;
521EXPORT_SYMBOL_GPL(system_freezable_wq);
522struct workqueue_struct *system_power_efficient_wq __ro_after_init;
523EXPORT_SYMBOL_GPL(system_power_efficient_wq);
524struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
525EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
526struct workqueue_struct *system_bh_wq;
527EXPORT_SYMBOL_GPL(system_bh_wq);
528struct workqueue_struct *system_bh_highpri_wq;
529EXPORT_SYMBOL_GPL(system_bh_highpri_wq);
530
531static int worker_thread(void *__worker);
532static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
533static void show_pwq(struct pool_workqueue *pwq);
534static void show_one_worker_pool(struct worker_pool *pool);
535
536#define CREATE_TRACE_POINTS
537#include <trace/events/workqueue.h>
538
539#define assert_rcu_or_pool_mutex() \
540 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
541 !lockdep_is_held(&wq_pool_mutex), \
542 "RCU or wq_pool_mutex should be held")
543
544#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
545 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
546 !lockdep_is_held(&wq->mutex) && \
547 !lockdep_is_held(&wq_pool_mutex), \
548 "RCU, wq->mutex or wq_pool_mutex should be held")
549
550#define for_each_bh_worker_pool(pool, cpu) \
551 for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \
552 (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
553 (pool)++)
554
555#define for_each_cpu_worker_pool(pool, cpu) \
556 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
557 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
558 (pool)++)
559
560/**
561 * for_each_pool - iterate through all worker_pools in the system
562 * @pool: iteration cursor
563 * @pi: integer used for iteration
564 *
565 * This must be called either with wq_pool_mutex held or RCU read
566 * locked. If the pool needs to be used beyond the locking in effect, the
567 * caller is responsible for guaranteeing that the pool stays online.
568 *
569 * The if/else clause exists only for the lockdep assertion and can be
570 * ignored.
571 */
572#define for_each_pool(pool, pi) \
573 idr_for_each_entry(&worker_pool_idr, pool, pi) \
574 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
575 else
576
577/**
578 * for_each_pool_worker - iterate through all workers of a worker_pool
579 * @worker: iteration cursor
580 * @pool: worker_pool to iterate workers of
581 *
582 * This must be called with wq_pool_attach_mutex.
583 *
584 * The if/else clause exists only for the lockdep assertion and can be
585 * ignored.
586 */
587#define for_each_pool_worker(worker, pool) \
588 list_for_each_entry((worker), &(pool)->workers, node) \
589 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
590 else
591
592/**
593 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
594 * @pwq: iteration cursor
595 * @wq: the target workqueue
596 *
597 * This must be called either with wq->mutex held or RCU read locked.
598 * If the pwq needs to be used beyond the locking in effect, the caller is
599 * responsible for guaranteeing that the pwq stays online.
600 *
601 * The if/else clause exists only for the lockdep assertion and can be
602 * ignored.
603 */
604#define for_each_pwq(pwq, wq) \
605 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
606 lockdep_is_held(&(wq->mutex)))
607
608#ifdef CONFIG_DEBUG_OBJECTS_WORK
609
610static const struct debug_obj_descr work_debug_descr;
611
612static void *work_debug_hint(void *addr)
613{
614 return ((struct work_struct *) addr)->func;
615}
616
617static bool work_is_static_object(void *addr)
618{
619 struct work_struct *work = addr;
620
621 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
622}
623
624/*
625 * fixup_init is called when:
626 * - an active object is initialized
627 */
628static bool work_fixup_init(void *addr, enum debug_obj_state state)
629{
630 struct work_struct *work = addr;
631
632 switch (state) {
633 case ODEBUG_STATE_ACTIVE:
634 cancel_work_sync(work);
635 debug_object_init(work, &work_debug_descr);
636 return true;
637 default:
638 return false;
639 }
640}
641
642/*
643 * fixup_free is called when:
644 * - an active object is freed
645 */
646static bool work_fixup_free(void *addr, enum debug_obj_state state)
647{
648 struct work_struct *work = addr;
649
650 switch (state) {
651 case ODEBUG_STATE_ACTIVE:
652 cancel_work_sync(work);
653 debug_object_free(work, &work_debug_descr);
654 return true;
655 default:
656 return false;
657 }
658}
659
660static const struct debug_obj_descr work_debug_descr = {
661 .name = "work_struct",
662 .debug_hint = work_debug_hint,
663 .is_static_object = work_is_static_object,
664 .fixup_init = work_fixup_init,
665 .fixup_free = work_fixup_free,
666};
667
668static inline void debug_work_activate(struct work_struct *work)
669{
670 debug_object_activate(work, &work_debug_descr);
671}
672
673static inline void debug_work_deactivate(struct work_struct *work)
674{
675 debug_object_deactivate(work, &work_debug_descr);
676}
677
678void __init_work(struct work_struct *work, int onstack)
679{
680 if (onstack)
681 debug_object_init_on_stack(work, &work_debug_descr);
682 else
683 debug_object_init(work, &work_debug_descr);
684}
685EXPORT_SYMBOL_GPL(__init_work);
686
687void destroy_work_on_stack(struct work_struct *work)
688{
689 debug_object_free(work, &work_debug_descr);
690}
691EXPORT_SYMBOL_GPL(destroy_work_on_stack);
692
693void destroy_delayed_work_on_stack(struct delayed_work *work)
694{
695 timer_destroy_on_stack(&work->timer);
696 debug_object_free(&work->work, &work_debug_descr);
697}
698EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
699
700#else
701static inline void debug_work_activate(struct work_struct *work) { }
702static inline void debug_work_deactivate(struct work_struct *work) { }
703#endif
704
705/**
706 * worker_pool_assign_id - allocate ID and assign it to @pool
707 * @pool: the pool pointer of interest
708 *
709 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
710 * successfully, -errno on failure.
711 */
712static int worker_pool_assign_id(struct worker_pool *pool)
713{
714 int ret;
715
716 lockdep_assert_held(&wq_pool_mutex);
717
718 ret = idr_alloc(&worker_pool_idr, ptr: pool, start: 0, WORK_OFFQ_POOL_NONE,
719 GFP_KERNEL);
720 if (ret >= 0) {
721 pool->id = ret;
722 return 0;
723 }
724 return ret;
725}
726
727static struct pool_workqueue __rcu **
728unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
729{
730 if (cpu >= 0)
731 return per_cpu_ptr(wq->cpu_pwq, cpu);
732 else
733 return &wq->dfl_pwq;
734}
735
736/* @cpu < 0 for dfl_pwq */
737static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu)
738{
739 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
740 lockdep_is_held(&wq_pool_mutex) ||
741 lockdep_is_held(&wq->mutex));
742}
743
744/**
745 * unbound_effective_cpumask - effective cpumask of an unbound workqueue
746 * @wq: workqueue of interest
747 *
748 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which
749 * is masked with wq_unbound_cpumask to determine the effective cpumask. The
750 * default pwq is always mapped to the pool with the current effective cpumask.
751 */
752static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq)
753{
754 return unbound_pwq(wq, cpu: -1)->pool->attrs->__pod_cpumask;
755}
756
757static unsigned int work_color_to_flags(int color)
758{
759 return color << WORK_STRUCT_COLOR_SHIFT;
760}
761
762static int get_work_color(unsigned long work_data)
763{
764 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
765 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
766}
767
768static int work_next_color(int color)
769{
770 return (color + 1) % WORK_NR_COLORS;
771}
772
773static unsigned long pool_offq_flags(struct worker_pool *pool)
774{
775 return (pool->flags & POOL_BH) ? WORK_OFFQ_BH : 0;
776}
777
778/*
779 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
780 * contain the pointer to the queued pwq. Once execution starts, the flag
781 * is cleared and the high bits contain OFFQ flags and pool ID.
782 *
783 * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling()
784 * can be used to set the pwq, pool or clear work->data. These functions should
785 * only be called while the work is owned - ie. while the PENDING bit is set.
786 *
787 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
788 * corresponding to a work. Pool is available once the work has been
789 * queued anywhere after initialization until it is sync canceled. pwq is
790 * available only while the work item is queued.
791 */
792static inline void set_work_data(struct work_struct *work, unsigned long data)
793{
794 WARN_ON_ONCE(!work_pending(work));
795 atomic_long_set(v: &work->data, i: data | work_static(work));
796}
797
798static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
799 unsigned long flags)
800{
801 set_work_data(work, data: (unsigned long)pwq | WORK_STRUCT_PENDING |
802 WORK_STRUCT_PWQ | flags);
803}
804
805static void set_work_pool_and_keep_pending(struct work_struct *work,
806 int pool_id, unsigned long flags)
807{
808 set_work_data(work, data: ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
809 WORK_STRUCT_PENDING | flags);
810}
811
812static void set_work_pool_and_clear_pending(struct work_struct *work,
813 int pool_id, unsigned long flags)
814{
815 /*
816 * The following wmb is paired with the implied mb in
817 * test_and_set_bit(PENDING) and ensures all updates to @work made
818 * here are visible to and precede any updates by the next PENDING
819 * owner.
820 */
821 smp_wmb();
822 set_work_data(work, data: ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
823 flags);
824 /*
825 * The following mb guarantees that previous clear of a PENDING bit
826 * will not be reordered with any speculative LOADS or STORES from
827 * work->current_func, which is executed afterwards. This possible
828 * reordering can lead to a missed execution on attempt to queue
829 * the same @work. E.g. consider this case:
830 *
831 * CPU#0 CPU#1
832 * ---------------------------- --------------------------------
833 *
834 * 1 STORE event_indicated
835 * 2 queue_work_on() {
836 * 3 test_and_set_bit(PENDING)
837 * 4 } set_..._and_clear_pending() {
838 * 5 set_work_data() # clear bit
839 * 6 smp_mb()
840 * 7 work->current_func() {
841 * 8 LOAD event_indicated
842 * }
843 *
844 * Without an explicit full barrier speculative LOAD on line 8 can
845 * be executed before CPU#0 does STORE on line 1. If that happens,
846 * CPU#0 observes the PENDING bit is still set and new execution of
847 * a @work is not queued in a hope, that CPU#1 will eventually
848 * finish the queued @work. Meanwhile CPU#1 does not see
849 * event_indicated is set, because speculative LOAD was executed
850 * before actual STORE.
851 */
852 smp_mb();
853}
854
855static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
856{
857 return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK);
858}
859
860static struct pool_workqueue *get_work_pwq(struct work_struct *work)
861{
862 unsigned long data = atomic_long_read(v: &work->data);
863
864 if (data & WORK_STRUCT_PWQ)
865 return work_struct_pwq(data);
866 else
867 return NULL;
868}
869
870/**
871 * get_work_pool - return the worker_pool a given work was associated with
872 * @work: the work item of interest
873 *
874 * Pools are created and destroyed under wq_pool_mutex, and allows read
875 * access under RCU read lock. As such, this function should be
876 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
877 *
878 * All fields of the returned pool are accessible as long as the above
879 * mentioned locking is in effect. If the returned pool needs to be used
880 * beyond the critical section, the caller is responsible for ensuring the
881 * returned pool is and stays online.
882 *
883 * Return: The worker_pool @work was last associated with. %NULL if none.
884 */
885static struct worker_pool *get_work_pool(struct work_struct *work)
886{
887 unsigned long data = atomic_long_read(v: &work->data);
888 int pool_id;
889
890 assert_rcu_or_pool_mutex();
891
892 if (data & WORK_STRUCT_PWQ)
893 return work_struct_pwq(data)->pool;
894
895 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
896 if (pool_id == WORK_OFFQ_POOL_NONE)
897 return NULL;
898
899 return idr_find(&worker_pool_idr, id: pool_id);
900}
901
902static unsigned long shift_and_mask(unsigned long v, u32 shift, u32 bits)
903{
904 return (v >> shift) & ((1U << bits) - 1);
905}
906
907static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data)
908{
909 WARN_ON_ONCE(data & WORK_STRUCT_PWQ);
910
911 offqd->pool_id = shift_and_mask(v: data, shift: WORK_OFFQ_POOL_SHIFT,
912 bits: WORK_OFFQ_POOL_BITS);
913 offqd->disable = shift_and_mask(v: data, shift: WORK_OFFQ_DISABLE_SHIFT,
914 bits: WORK_OFFQ_DISABLE_BITS);
915 offqd->flags = data & WORK_OFFQ_FLAG_MASK;
916}
917
918static unsigned long work_offqd_pack_flags(struct work_offq_data *offqd)
919{
920 return ((unsigned long)offqd->disable << WORK_OFFQ_DISABLE_SHIFT) |
921 ((unsigned long)offqd->flags);
922}
923
924/*
925 * Policy functions. These define the policies on how the global worker
926 * pools are managed. Unless noted otherwise, these functions assume that
927 * they're being called with pool->lock held.
928 */
929
930/*
931 * Need to wake up a worker? Called from anything but currently
932 * running workers.
933 *
934 * Note that, because unbound workers never contribute to nr_running, this
935 * function will always return %true for unbound pools as long as the
936 * worklist isn't empty.
937 */
938static bool need_more_worker(struct worker_pool *pool)
939{
940 return !list_empty(head: &pool->worklist) && !pool->nr_running;
941}
942
943/* Can I start working? Called from busy but !running workers. */
944static bool may_start_working(struct worker_pool *pool)
945{
946 return pool->nr_idle;
947}
948
949/* Do I need to keep working? Called from currently running workers. */
950static bool keep_working(struct worker_pool *pool)
951{
952 return !list_empty(head: &pool->worklist) && (pool->nr_running <= 1);
953}
954
955/* Do we need a new worker? Called from manager. */
956static bool need_to_create_worker(struct worker_pool *pool)
957{
958 return need_more_worker(pool) && !may_start_working(pool);
959}
960
961/* Do we have too many workers and should some go away? */
962static bool too_many_workers(struct worker_pool *pool)
963{
964 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
965 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
966 int nr_busy = pool->nr_workers - nr_idle;
967
968 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
969}
970
971/**
972 * worker_set_flags - set worker flags and adjust nr_running accordingly
973 * @worker: self
974 * @flags: flags to set
975 *
976 * Set @flags in @worker->flags and adjust nr_running accordingly.
977 */
978static inline void worker_set_flags(struct worker *worker, unsigned int flags)
979{
980 struct worker_pool *pool = worker->pool;
981
982 lockdep_assert_held(&pool->lock);
983
984 /* If transitioning into NOT_RUNNING, adjust nr_running. */
985 if ((flags & WORKER_NOT_RUNNING) &&
986 !(worker->flags & WORKER_NOT_RUNNING)) {
987 pool->nr_running--;
988 }
989
990 worker->flags |= flags;
991}
992
993/**
994 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
995 * @worker: self
996 * @flags: flags to clear
997 *
998 * Clear @flags in @worker->flags and adjust nr_running accordingly.
999 */
1000static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
1001{
1002 struct worker_pool *pool = worker->pool;
1003 unsigned int oflags = worker->flags;
1004
1005 lockdep_assert_held(&pool->lock);
1006
1007 worker->flags &= ~flags;
1008
1009 /*
1010 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1011 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1012 * of multiple flags, not a single flag.
1013 */
1014 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1015 if (!(worker->flags & WORKER_NOT_RUNNING))
1016 pool->nr_running++;
1017}
1018
1019/* Return the first idle worker. Called with pool->lock held. */
1020static struct worker *first_idle_worker(struct worker_pool *pool)
1021{
1022 if (unlikely(list_empty(&pool->idle_list)))
1023 return NULL;
1024
1025 return list_first_entry(&pool->idle_list, struct worker, entry);
1026}
1027
1028/**
1029 * worker_enter_idle - enter idle state
1030 * @worker: worker which is entering idle state
1031 *
1032 * @worker is entering idle state. Update stats and idle timer if
1033 * necessary.
1034 *
1035 * LOCKING:
1036 * raw_spin_lock_irq(pool->lock).
1037 */
1038static void worker_enter_idle(struct worker *worker)
1039{
1040 struct worker_pool *pool = worker->pool;
1041
1042 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1043 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1044 (worker->hentry.next || worker->hentry.pprev)))
1045 return;
1046
1047 /* can't use worker_set_flags(), also called from create_worker() */
1048 worker->flags |= WORKER_IDLE;
1049 pool->nr_idle++;
1050 worker->last_active = jiffies;
1051
1052 /* idle_list is LIFO */
1053 list_add(new: &worker->entry, head: &pool->idle_list);
1054
1055 if (too_many_workers(pool) && !timer_pending(timer: &pool->idle_timer))
1056 mod_timer(timer: &pool->idle_timer, expires: jiffies + IDLE_WORKER_TIMEOUT);
1057
1058 /* Sanity check nr_running. */
1059 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1060}
1061
1062/**
1063 * worker_leave_idle - leave idle state
1064 * @worker: worker which is leaving idle state
1065 *
1066 * @worker is leaving idle state. Update stats.
1067 *
1068 * LOCKING:
1069 * raw_spin_lock_irq(pool->lock).
1070 */
1071static void worker_leave_idle(struct worker *worker)
1072{
1073 struct worker_pool *pool = worker->pool;
1074
1075 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1076 return;
1077 worker_clr_flags(worker, flags: WORKER_IDLE);
1078 pool->nr_idle--;
1079 list_del_init(entry: &worker->entry);
1080}
1081
1082/**
1083 * find_worker_executing_work - find worker which is executing a work
1084 * @pool: pool of interest
1085 * @work: work to find worker for
1086 *
1087 * Find a worker which is executing @work on @pool by searching
1088 * @pool->busy_hash which is keyed by the address of @work. For a worker
1089 * to match, its current execution should match the address of @work and
1090 * its work function. This is to avoid unwanted dependency between
1091 * unrelated work executions through a work item being recycled while still
1092 * being executed.
1093 *
1094 * This is a bit tricky. A work item may be freed once its execution
1095 * starts and nothing prevents the freed area from being recycled for
1096 * another work item. If the same work item address ends up being reused
1097 * before the original execution finishes, workqueue will identify the
1098 * recycled work item as currently executing and make it wait until the
1099 * current execution finishes, introducing an unwanted dependency.
1100 *
1101 * This function checks the work item address and work function to avoid
1102 * false positives. Note that this isn't complete as one may construct a
1103 * work function which can introduce dependency onto itself through a
1104 * recycled work item. Well, if somebody wants to shoot oneself in the
1105 * foot that badly, there's only so much we can do, and if such deadlock
1106 * actually occurs, it should be easy to locate the culprit work function.
1107 *
1108 * CONTEXT:
1109 * raw_spin_lock_irq(pool->lock).
1110 *
1111 * Return:
1112 * Pointer to worker which is executing @work if found, %NULL
1113 * otherwise.
1114 */
1115static struct worker *find_worker_executing_work(struct worker_pool *pool,
1116 struct work_struct *work)
1117{
1118 struct worker *worker;
1119
1120 hash_for_each_possible(pool->busy_hash, worker, hentry,
1121 (unsigned long)work)
1122 if (worker->current_work == work &&
1123 worker->current_func == work->func)
1124 return worker;
1125
1126 return NULL;
1127}
1128
1129/**
1130 * move_linked_works - move linked works to a list
1131 * @work: start of series of works to be scheduled
1132 * @head: target list to append @work to
1133 * @nextp: out parameter for nested worklist walking
1134 *
1135 * Schedule linked works starting from @work to @head. Work series to be
1136 * scheduled starts at @work and includes any consecutive work with
1137 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1138 * @nextp.
1139 *
1140 * CONTEXT:
1141 * raw_spin_lock_irq(pool->lock).
1142 */
1143static void move_linked_works(struct work_struct *work, struct list_head *head,
1144 struct work_struct **nextp)
1145{
1146 struct work_struct *n;
1147
1148 /*
1149 * Linked worklist will always end before the end of the list,
1150 * use NULL for list head.
1151 */
1152 list_for_each_entry_safe_from(work, n, NULL, entry) {
1153 list_move_tail(list: &work->entry, head);
1154 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1155 break;
1156 }
1157
1158 /*
1159 * If we're already inside safe list traversal and have moved
1160 * multiple works to the scheduled queue, the next position
1161 * needs to be updated.
1162 */
1163 if (nextp)
1164 *nextp = n;
1165}
1166
1167/**
1168 * assign_work - assign a work item and its linked work items to a worker
1169 * @work: work to assign
1170 * @worker: worker to assign to
1171 * @nextp: out parameter for nested worklist walking
1172 *
1173 * Assign @work and its linked work items to @worker. If @work is already being
1174 * executed by another worker in the same pool, it'll be punted there.
1175 *
1176 * If @nextp is not NULL, it's updated to point to the next work of the last
1177 * scheduled work. This allows assign_work() to be nested inside
1178 * list_for_each_entry_safe().
1179 *
1180 * Returns %true if @work was successfully assigned to @worker. %false if @work
1181 * was punted to another worker already executing it.
1182 */
1183static bool assign_work(struct work_struct *work, struct worker *worker,
1184 struct work_struct **nextp)
1185{
1186 struct worker_pool *pool = worker->pool;
1187 struct worker *collision;
1188
1189 lockdep_assert_held(&pool->lock);
1190
1191 /*
1192 * A single work shouldn't be executed concurrently by multiple workers.
1193 * __queue_work() ensures that @work doesn't jump to a different pool
1194 * while still running in the previous pool. Here, we should ensure that
1195 * @work is not executed concurrently by multiple workers from the same
1196 * pool. Check whether anyone is already processing the work. If so,
1197 * defer the work to the currently executing one.
1198 */
1199 collision = find_worker_executing_work(pool, work);
1200 if (unlikely(collision)) {
1201 move_linked_works(work, head: &collision->scheduled, nextp);
1202 return false;
1203 }
1204
1205 move_linked_works(work, head: &worker->scheduled, nextp);
1206 return true;
1207}
1208
1209static struct irq_work *bh_pool_irq_work(struct worker_pool *pool)
1210{
1211 int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0;
1212
1213 return &per_cpu(bh_pool_irq_works, pool->cpu)[high];
1214}
1215
1216static void kick_bh_pool(struct worker_pool *pool)
1217{
1218#ifdef CONFIG_SMP
1219 /* see drain_dead_softirq_workfn() for BH_DRAINING */
1220 if (unlikely(pool->cpu != smp_processor_id() &&
1221 !(pool->flags & POOL_BH_DRAINING))) {
1222 irq_work_queue_on(work: bh_pool_irq_work(pool), cpu: pool->cpu);
1223 return;
1224 }
1225#endif
1226 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
1227 raise_softirq_irqoff(nr: HI_SOFTIRQ);
1228 else
1229 raise_softirq_irqoff(nr: TASKLET_SOFTIRQ);
1230}
1231
1232/**
1233 * kick_pool - wake up an idle worker if necessary
1234 * @pool: pool to kick
1235 *
1236 * @pool may have pending work items. Wake up worker if necessary. Returns
1237 * whether a worker was woken up.
1238 */
1239static bool kick_pool(struct worker_pool *pool)
1240{
1241 struct worker *worker = first_idle_worker(pool);
1242 struct task_struct *p;
1243
1244 lockdep_assert_held(&pool->lock);
1245
1246 if (!need_more_worker(pool) || !worker)
1247 return false;
1248
1249 if (pool->flags & POOL_BH) {
1250 kick_bh_pool(pool);
1251 return true;
1252 }
1253
1254 p = worker->task;
1255
1256#ifdef CONFIG_SMP
1257 /*
1258 * Idle @worker is about to execute @work and waking up provides an
1259 * opportunity to migrate @worker at a lower cost by setting the task's
1260 * wake_cpu field. Let's see if we want to move @worker to improve
1261 * execution locality.
1262 *
1263 * We're waking the worker that went idle the latest and there's some
1264 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1265 * so, setting the wake_cpu won't do anything. As this is a best-effort
1266 * optimization and the race window is narrow, let's leave as-is for
1267 * now. If this becomes pronounced, we can skip over workers which are
1268 * still on cpu when picking an idle worker.
1269 *
1270 * If @pool has non-strict affinity, @worker might have ended up outside
1271 * its affinity scope. Repatriate.
1272 */
1273 if (!pool->attrs->affn_strict &&
1274 !cpumask_test_cpu(cpu: p->wake_cpu, cpumask: pool->attrs->__pod_cpumask)) {
1275 struct work_struct *work = list_first_entry(&pool->worklist,
1276 struct work_struct, entry);
1277 int wake_cpu = cpumask_any_and_distribute(src1p: pool->attrs->__pod_cpumask,
1278 cpu_online_mask);
1279 if (wake_cpu < nr_cpu_ids) {
1280 p->wake_cpu = wake_cpu;
1281 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1282 }
1283 }
1284#endif
1285 wake_up_process(tsk: p);
1286 return true;
1287}
1288
1289#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1290
1291/*
1292 * Concurrency-managed per-cpu work items that hog CPU for longer than
1293 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1294 * which prevents them from stalling other concurrency-managed work items. If a
1295 * work function keeps triggering this mechanism, it's likely that the work item
1296 * should be using an unbound workqueue instead.
1297 *
1298 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1299 * and report them so that they can be examined and converted to use unbound
1300 * workqueues as appropriate. To avoid flooding the console, each violating work
1301 * function is tracked and reported with exponential backoff.
1302 */
1303#define WCI_MAX_ENTS 128
1304
1305struct wci_ent {
1306 work_func_t func;
1307 atomic64_t cnt;
1308 struct hlist_node hash_node;
1309};
1310
1311static struct wci_ent wci_ents[WCI_MAX_ENTS];
1312static int wci_nr_ents;
1313static DEFINE_RAW_SPINLOCK(wci_lock);
1314static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1315
1316static struct wci_ent *wci_find_ent(work_func_t func)
1317{
1318 struct wci_ent *ent;
1319
1320 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1321 (unsigned long)func) {
1322 if (ent->func == func)
1323 return ent;
1324 }
1325 return NULL;
1326}
1327
1328static void wq_cpu_intensive_report(work_func_t func)
1329{
1330 struct wci_ent *ent;
1331
1332restart:
1333 ent = wci_find_ent(func);
1334 if (ent) {
1335 u64 cnt;
1336
1337 /*
1338 * Start reporting from the warning_thresh and back off
1339 * exponentially.
1340 */
1341 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1342 if (wq_cpu_intensive_warning_thresh &&
1343 cnt >= wq_cpu_intensive_warning_thresh &&
1344 is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh))
1345 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1346 ent->func, wq_cpu_intensive_thresh_us,
1347 atomic64_read(&ent->cnt));
1348 return;
1349 }
1350
1351 /*
1352 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1353 * is exhausted, something went really wrong and we probably made enough
1354 * noise already.
1355 */
1356 if (wci_nr_ents >= WCI_MAX_ENTS)
1357 return;
1358
1359 raw_spin_lock(&wci_lock);
1360
1361 if (wci_nr_ents >= WCI_MAX_ENTS) {
1362 raw_spin_unlock(&wci_lock);
1363 return;
1364 }
1365
1366 if (wci_find_ent(func)) {
1367 raw_spin_unlock(&wci_lock);
1368 goto restart;
1369 }
1370
1371 ent = &wci_ents[wci_nr_ents++];
1372 ent->func = func;
1373 atomic64_set(&ent->cnt, 0);
1374 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1375
1376 raw_spin_unlock(&wci_lock);
1377
1378 goto restart;
1379}
1380
1381#else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1382static void wq_cpu_intensive_report(work_func_t func) {}
1383#endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1384
1385/**
1386 * wq_worker_running - a worker is running again
1387 * @task: task waking up
1388 *
1389 * This function is called when a worker returns from schedule()
1390 */
1391void wq_worker_running(struct task_struct *task)
1392{
1393 struct worker *worker = kthread_data(k: task);
1394
1395 if (!READ_ONCE(worker->sleeping))
1396 return;
1397
1398 /*
1399 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1400 * and the nr_running increment below, we may ruin the nr_running reset
1401 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1402 * pool. Protect against such race.
1403 */
1404 preempt_disable();
1405 if (!(worker->flags & WORKER_NOT_RUNNING))
1406 worker->pool->nr_running++;
1407 preempt_enable();
1408
1409 /*
1410 * CPU intensive auto-detection cares about how long a work item hogged
1411 * CPU without sleeping. Reset the starting timestamp on wakeup.
1412 */
1413 worker->current_at = worker->task->se.sum_exec_runtime;
1414
1415 WRITE_ONCE(worker->sleeping, 0);
1416}
1417
1418/**
1419 * wq_worker_sleeping - a worker is going to sleep
1420 * @task: task going to sleep
1421 *
1422 * This function is called from schedule() when a busy worker is
1423 * going to sleep.
1424 */
1425void wq_worker_sleeping(struct task_struct *task)
1426{
1427 struct worker *worker = kthread_data(k: task);
1428 struct worker_pool *pool;
1429
1430 /*
1431 * Rescuers, which may not have all the fields set up like normal
1432 * workers, also reach here, let's not access anything before
1433 * checking NOT_RUNNING.
1434 */
1435 if (worker->flags & WORKER_NOT_RUNNING)
1436 return;
1437
1438 pool = worker->pool;
1439
1440 /* Return if preempted before wq_worker_running() was reached */
1441 if (READ_ONCE(worker->sleeping))
1442 return;
1443
1444 WRITE_ONCE(worker->sleeping, 1);
1445 raw_spin_lock_irq(&pool->lock);
1446
1447 /*
1448 * Recheck in case unbind_workers() preempted us. We don't
1449 * want to decrement nr_running after the worker is unbound
1450 * and nr_running has been reset.
1451 */
1452 if (worker->flags & WORKER_NOT_RUNNING) {
1453 raw_spin_unlock_irq(&pool->lock);
1454 return;
1455 }
1456
1457 pool->nr_running--;
1458 if (kick_pool(pool))
1459 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1460
1461 raw_spin_unlock_irq(&pool->lock);
1462}
1463
1464/**
1465 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1466 * @task: task currently running
1467 *
1468 * Called from sched_tick(). We're in the IRQ context and the current
1469 * worker's fields which follow the 'K' locking rule can be accessed safely.
1470 */
1471void wq_worker_tick(struct task_struct *task)
1472{
1473 struct worker *worker = kthread_data(k: task);
1474 struct pool_workqueue *pwq = worker->current_pwq;
1475 struct worker_pool *pool = worker->pool;
1476
1477 if (!pwq)
1478 return;
1479
1480 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1481
1482 if (!wq_cpu_intensive_thresh_us)
1483 return;
1484
1485 /*
1486 * If the current worker is concurrency managed and hogged the CPU for
1487 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1488 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1489 *
1490 * Set @worker->sleeping means that @worker is in the process of
1491 * switching out voluntarily and won't be contributing to
1492 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1493 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1494 * double decrements. The task is releasing the CPU anyway. Let's skip.
1495 * We probably want to make this prettier in the future.
1496 */
1497 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1498 worker->task->se.sum_exec_runtime - worker->current_at <
1499 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1500 return;
1501
1502 raw_spin_lock(&pool->lock);
1503
1504 worker_set_flags(worker, flags: WORKER_CPU_INTENSIVE);
1505 wq_cpu_intensive_report(func: worker->current_func);
1506 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1507
1508 if (kick_pool(pool))
1509 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1510
1511 raw_spin_unlock(&pool->lock);
1512}
1513
1514/**
1515 * wq_worker_last_func - retrieve worker's last work function
1516 * @task: Task to retrieve last work function of.
1517 *
1518 * Determine the last function a worker executed. This is called from
1519 * the scheduler to get a worker's last known identity.
1520 *
1521 * CONTEXT:
1522 * raw_spin_lock_irq(rq->lock)
1523 *
1524 * This function is called during schedule() when a kworker is going
1525 * to sleep. It's used by psi to identify aggregation workers during
1526 * dequeuing, to allow periodic aggregation to shut-off when that
1527 * worker is the last task in the system or cgroup to go to sleep.
1528 *
1529 * As this function doesn't involve any workqueue-related locking, it
1530 * only returns stable values when called from inside the scheduler's
1531 * queuing and dequeuing paths, when @task, which must be a kworker,
1532 * is guaranteed to not be processing any works.
1533 *
1534 * Return:
1535 * The last work function %current executed as a worker, NULL if it
1536 * hasn't executed any work yet.
1537 */
1538work_func_t wq_worker_last_func(struct task_struct *task)
1539{
1540 struct worker *worker = kthread_data(k: task);
1541
1542 return worker->last_func;
1543}
1544
1545/**
1546 * wq_node_nr_active - Determine wq_node_nr_active to use
1547 * @wq: workqueue of interest
1548 * @node: NUMA node, can be %NUMA_NO_NODE
1549 *
1550 * Determine wq_node_nr_active to use for @wq on @node. Returns:
1551 *
1552 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active.
1553 *
1554 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1555 *
1556 * - Otherwise, node_nr_active[@node].
1557 */
1558static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq,
1559 int node)
1560{
1561 if (!(wq->flags & WQ_UNBOUND))
1562 return NULL;
1563
1564 if (node == NUMA_NO_NODE)
1565 node = nr_node_ids;
1566
1567 return wq->node_nr_active[node];
1568}
1569
1570/**
1571 * wq_update_node_max_active - Update per-node max_actives to use
1572 * @wq: workqueue to update
1573 * @off_cpu: CPU that's going down, -1 if a CPU is not going down
1574 *
1575 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1576 * distributed among nodes according to the proportions of numbers of online
1577 * cpus. The result is always between @wq->min_active and max_active.
1578 */
1579static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1580{
1581 struct cpumask *effective = unbound_effective_cpumask(wq);
1582 int min_active = READ_ONCE(wq->min_active);
1583 int max_active = READ_ONCE(wq->max_active);
1584 int total_cpus, node;
1585
1586 lockdep_assert_held(&wq->mutex);
1587
1588 if (!wq_topo_initialized)
1589 return;
1590
1591 if (off_cpu >= 0 && !cpumask_test_cpu(cpu: off_cpu, cpumask: effective))
1592 off_cpu = -1;
1593
1594 total_cpus = cpumask_weight_and(srcp1: effective, cpu_online_mask);
1595 if (off_cpu >= 0)
1596 total_cpus--;
1597
1598 /* If all CPUs of the wq get offline, use the default values */
1599 if (unlikely(!total_cpus)) {
1600 for_each_node(node)
1601 wq_node_nr_active(wq, node)->max = min_active;
1602
1603 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1604 return;
1605 }
1606
1607 for_each_node(node) {
1608 int node_cpus;
1609
1610 node_cpus = cpumask_weight_and(srcp1: effective, srcp2: cpumask_of_node(node));
1611 if (off_cpu >= 0 && cpu_to_node(cpu: off_cpu) == node)
1612 node_cpus--;
1613
1614 wq_node_nr_active(wq, node)->max =
1615 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1616 min_active, max_active);
1617 }
1618
1619 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1620}
1621
1622/**
1623 * get_pwq - get an extra reference on the specified pool_workqueue
1624 * @pwq: pool_workqueue to get
1625 *
1626 * Obtain an extra reference on @pwq. The caller should guarantee that
1627 * @pwq has positive refcnt and be holding the matching pool->lock.
1628 */
1629static void get_pwq(struct pool_workqueue *pwq)
1630{
1631 lockdep_assert_held(&pwq->pool->lock);
1632 WARN_ON_ONCE(pwq->refcnt <= 0);
1633 pwq->refcnt++;
1634}
1635
1636/**
1637 * put_pwq - put a pool_workqueue reference
1638 * @pwq: pool_workqueue to put
1639 *
1640 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1641 * destruction. The caller should be holding the matching pool->lock.
1642 */
1643static void put_pwq(struct pool_workqueue *pwq)
1644{
1645 lockdep_assert_held(&pwq->pool->lock);
1646 if (likely(--pwq->refcnt))
1647 return;
1648 /*
1649 * @pwq can't be released under pool->lock, bounce to a dedicated
1650 * kthread_worker to avoid A-A deadlocks.
1651 */
1652 kthread_queue_work(worker: pwq_release_worker, work: &pwq->release_work);
1653}
1654
1655/**
1656 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1657 * @pwq: pool_workqueue to put (can be %NULL)
1658 *
1659 * put_pwq() with locking. This function also allows %NULL @pwq.
1660 */
1661static void put_pwq_unlocked(struct pool_workqueue *pwq)
1662{
1663 if (pwq) {
1664 /*
1665 * As both pwqs and pools are RCU protected, the
1666 * following lock operations are safe.
1667 */
1668 raw_spin_lock_irq(&pwq->pool->lock);
1669 put_pwq(pwq);
1670 raw_spin_unlock_irq(&pwq->pool->lock);
1671 }
1672}
1673
1674static bool pwq_is_empty(struct pool_workqueue *pwq)
1675{
1676 return !pwq->nr_active && list_empty(head: &pwq->inactive_works);
1677}
1678
1679static void __pwq_activate_work(struct pool_workqueue *pwq,
1680 struct work_struct *work)
1681{
1682 unsigned long *wdb = work_data_bits(work);
1683
1684 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1685 trace_workqueue_activate_work(work);
1686 if (list_empty(head: &pwq->pool->worklist))
1687 pwq->pool->watchdog_ts = jiffies;
1688 move_linked_works(work, head: &pwq->pool->worklist, NULL);
1689 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1690}
1691
1692static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1693{
1694 int max = READ_ONCE(nna->max);
1695 int old = atomic_read(v: &nna->nr);
1696
1697 do {
1698 if (old >= max)
1699 return false;
1700 } while (!atomic_try_cmpxchg_relaxed(v: &nna->nr, old: &old, new: old + 1));
1701
1702 return true;
1703}
1704
1705/**
1706 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1707 * @pwq: pool_workqueue of interest
1708 * @fill: max_active may have increased, try to increase concurrency level
1709 *
1710 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1711 * successfully obtained. %false otherwise.
1712 */
1713static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1714{
1715 struct workqueue_struct *wq = pwq->wq;
1716 struct worker_pool *pool = pwq->pool;
1717 struct wq_node_nr_active *nna = wq_node_nr_active(wq, node: pool->node);
1718 bool obtained = false;
1719
1720 lockdep_assert_held(&pool->lock);
1721
1722 if (!nna) {
1723 /* BH or per-cpu workqueue, pwq->nr_active is sufficient */
1724 obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1725 goto out;
1726 }
1727
1728 if (unlikely(pwq->plugged))
1729 return false;
1730
1731 /*
1732 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1733 * already waiting on $nna, pwq_dec_nr_active() will maintain the
1734 * concurrency level. Don't jump the line.
1735 *
1736 * We need to ignore the pending test after max_active has increased as
1737 * pwq_dec_nr_active() can only maintain the concurrency level but not
1738 * increase it. This is indicated by @fill.
1739 */
1740 if (!list_empty(head: &pwq->pending_node) && likely(!fill))
1741 goto out;
1742
1743 obtained = tryinc_node_nr_active(nna);
1744 if (obtained)
1745 goto out;
1746
1747 /*
1748 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1749 * and try again. The smp_mb() is paired with the implied memory barrier
1750 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1751 * we see the decremented $nna->nr or they see non-empty
1752 * $nna->pending_pwqs.
1753 */
1754 raw_spin_lock(&nna->lock);
1755
1756 if (list_empty(head: &pwq->pending_node))
1757 list_add_tail(new: &pwq->pending_node, head: &nna->pending_pwqs);
1758 else if (likely(!fill))
1759 goto out_unlock;
1760
1761 smp_mb();
1762
1763 obtained = tryinc_node_nr_active(nna);
1764
1765 /*
1766 * If @fill, @pwq might have already been pending. Being spuriously
1767 * pending in cold paths doesn't affect anything. Let's leave it be.
1768 */
1769 if (obtained && likely(!fill))
1770 list_del_init(entry: &pwq->pending_node);
1771
1772out_unlock:
1773 raw_spin_unlock(&nna->lock);
1774out:
1775 if (obtained)
1776 pwq->nr_active++;
1777 return obtained;
1778}
1779
1780/**
1781 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1782 * @pwq: pool_workqueue of interest
1783 * @fill: max_active may have increased, try to increase concurrency level
1784 *
1785 * Activate the first inactive work item of @pwq if available and allowed by
1786 * max_active limit.
1787 *
1788 * Returns %true if an inactive work item has been activated. %false if no
1789 * inactive work item is found or max_active limit is reached.
1790 */
1791static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1792{
1793 struct work_struct *work =
1794 list_first_entry_or_null(&pwq->inactive_works,
1795 struct work_struct, entry);
1796
1797 if (work && pwq_tryinc_nr_active(pwq, fill)) {
1798 __pwq_activate_work(pwq, work);
1799 return true;
1800 } else {
1801 return false;
1802 }
1803}
1804
1805/**
1806 * unplug_oldest_pwq - unplug the oldest pool_workqueue
1807 * @wq: workqueue_struct where its oldest pwq is to be unplugged
1808 *
1809 * This function should only be called for ordered workqueues where only the
1810 * oldest pwq is unplugged, the others are plugged to suspend execution to
1811 * ensure proper work item ordering::
1812 *
1813 * dfl_pwq --------------+ [P] - plugged
1814 * |
1815 * v
1816 * pwqs -> A -> B [P] -> C [P] (newest)
1817 * | | |
1818 * 1 3 5
1819 * | | |
1820 * 2 4 6
1821 *
1822 * When the oldest pwq is drained and removed, this function should be called
1823 * to unplug the next oldest one to start its work item execution. Note that
1824 * pwq's are linked into wq->pwqs with the oldest first, so the first one in
1825 * the list is the oldest.
1826 */
1827static void unplug_oldest_pwq(struct workqueue_struct *wq)
1828{
1829 struct pool_workqueue *pwq;
1830
1831 lockdep_assert_held(&wq->mutex);
1832
1833 /* Caller should make sure that pwqs isn't empty before calling */
1834 pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue,
1835 pwqs_node);
1836 raw_spin_lock_irq(&pwq->pool->lock);
1837 if (pwq->plugged) {
1838 pwq->plugged = false;
1839 if (pwq_activate_first_inactive(pwq, fill: true))
1840 kick_pool(pool: pwq->pool);
1841 }
1842 raw_spin_unlock_irq(&pwq->pool->lock);
1843}
1844
1845/**
1846 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1847 * @nna: wq_node_nr_active to activate a pending pwq for
1848 * @caller_pool: worker_pool the caller is locking
1849 *
1850 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1851 * @caller_pool may be unlocked and relocked to lock other worker_pools.
1852 */
1853static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1854 struct worker_pool *caller_pool)
1855{
1856 struct worker_pool *locked_pool = caller_pool;
1857 struct pool_workqueue *pwq;
1858 struct work_struct *work;
1859
1860 lockdep_assert_held(&caller_pool->lock);
1861
1862 raw_spin_lock(&nna->lock);
1863retry:
1864 pwq = list_first_entry_or_null(&nna->pending_pwqs,
1865 struct pool_workqueue, pending_node);
1866 if (!pwq)
1867 goto out_unlock;
1868
1869 /*
1870 * If @pwq is for a different pool than @locked_pool, we need to lock
1871 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1872 * / lock dance. For that, we also need to release @nna->lock as it's
1873 * nested inside pool locks.
1874 */
1875 if (pwq->pool != locked_pool) {
1876 raw_spin_unlock(&locked_pool->lock);
1877 locked_pool = pwq->pool;
1878 if (!raw_spin_trylock(&locked_pool->lock)) {
1879 raw_spin_unlock(&nna->lock);
1880 raw_spin_lock(&locked_pool->lock);
1881 raw_spin_lock(&nna->lock);
1882 goto retry;
1883 }
1884 }
1885
1886 /*
1887 * $pwq may not have any inactive work items due to e.g. cancellations.
1888 * Drop it from pending_pwqs and see if there's another one.
1889 */
1890 work = list_first_entry_or_null(&pwq->inactive_works,
1891 struct work_struct, entry);
1892 if (!work) {
1893 list_del_init(entry: &pwq->pending_node);
1894 goto retry;
1895 }
1896
1897 /*
1898 * Acquire an nr_active count and activate the inactive work item. If
1899 * $pwq still has inactive work items, rotate it to the end of the
1900 * pending_pwqs so that we round-robin through them. This means that
1901 * inactive work items are not activated in queueing order which is fine
1902 * given that there has never been any ordering across different pwqs.
1903 */
1904 if (likely(tryinc_node_nr_active(nna))) {
1905 pwq->nr_active++;
1906 __pwq_activate_work(pwq, work);
1907
1908 if (list_empty(head: &pwq->inactive_works))
1909 list_del_init(entry: &pwq->pending_node);
1910 else
1911 list_move_tail(list: &pwq->pending_node, head: &nna->pending_pwqs);
1912
1913 /* if activating a foreign pool, make sure it's running */
1914 if (pwq->pool != caller_pool)
1915 kick_pool(pool: pwq->pool);
1916 }
1917
1918out_unlock:
1919 raw_spin_unlock(&nna->lock);
1920 if (locked_pool != caller_pool) {
1921 raw_spin_unlock(&locked_pool->lock);
1922 raw_spin_lock(&caller_pool->lock);
1923 }
1924}
1925
1926/**
1927 * pwq_dec_nr_active - Retire an active count
1928 * @pwq: pool_workqueue of interest
1929 *
1930 * Decrement @pwq's nr_active and try to activate the first inactive work item.
1931 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
1932 */
1933static void pwq_dec_nr_active(struct pool_workqueue *pwq)
1934{
1935 struct worker_pool *pool = pwq->pool;
1936 struct wq_node_nr_active *nna = wq_node_nr_active(wq: pwq->wq, node: pool->node);
1937
1938 lockdep_assert_held(&pool->lock);
1939
1940 /*
1941 * @pwq->nr_active should be decremented for both percpu and unbound
1942 * workqueues.
1943 */
1944 pwq->nr_active--;
1945
1946 /*
1947 * For a percpu workqueue, it's simple. Just need to kick the first
1948 * inactive work item on @pwq itself.
1949 */
1950 if (!nna) {
1951 pwq_activate_first_inactive(pwq, fill: false);
1952 return;
1953 }
1954
1955 /*
1956 * If @pwq is for an unbound workqueue, it's more complicated because
1957 * multiple pwqs and pools may be sharing the nr_active count. When a
1958 * pwq needs to wait for an nr_active count, it puts itself on
1959 * $nna->pending_pwqs. The following atomic_dec_return()'s implied
1960 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
1961 * guarantee that either we see non-empty pending_pwqs or they see
1962 * decremented $nna->nr.
1963 *
1964 * $nna->max may change as CPUs come online/offline and @pwq->wq's
1965 * max_active gets updated. However, it is guaranteed to be equal to or
1966 * larger than @pwq->wq->min_active which is above zero unless freezing.
1967 * This maintains the forward progress guarantee.
1968 */
1969 if (atomic_dec_return(v: &nna->nr) >= READ_ONCE(nna->max))
1970 return;
1971
1972 if (!list_empty(head: &nna->pending_pwqs))
1973 node_activate_pending_pwq(nna, caller_pool: pool);
1974}
1975
1976/**
1977 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1978 * @pwq: pwq of interest
1979 * @work_data: work_data of work which left the queue
1980 *
1981 * A work either has completed or is removed from pending queue,
1982 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1983 *
1984 * NOTE:
1985 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock
1986 * and thus should be called after all other state updates for the in-flight
1987 * work item is complete.
1988 *
1989 * CONTEXT:
1990 * raw_spin_lock_irq(pool->lock).
1991 */
1992static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1993{
1994 int color = get_work_color(work_data);
1995
1996 if (!(work_data & WORK_STRUCT_INACTIVE))
1997 pwq_dec_nr_active(pwq);
1998
1999 pwq->nr_in_flight[color]--;
2000
2001 /* is flush in progress and are we at the flushing tip? */
2002 if (likely(pwq->flush_color != color))
2003 goto out_put;
2004
2005 /* are there still in-flight works? */
2006 if (pwq->nr_in_flight[color])
2007 goto out_put;
2008
2009 /* this pwq is done, clear flush_color */
2010 pwq->flush_color = -1;
2011
2012 /*
2013 * If this was the last pwq, wake up the first flusher. It
2014 * will handle the rest.
2015 */
2016 if (atomic_dec_and_test(v: &pwq->wq->nr_pwqs_to_flush))
2017 complete(&pwq->wq->first_flusher->done);
2018out_put:
2019 put_pwq(pwq);
2020}
2021
2022/**
2023 * try_to_grab_pending - steal work item from worklist and disable irq
2024 * @work: work item to steal
2025 * @cflags: %WORK_CANCEL_ flags
2026 * @irq_flags: place to store irq state
2027 *
2028 * Try to grab PENDING bit of @work. This function can handle @work in any
2029 * stable state - idle, on timer or on worklist.
2030 *
2031 * Return:
2032 *
2033 * ======== ================================================================
2034 * 1 if @work was pending and we successfully stole PENDING
2035 * 0 if @work was idle and we claimed PENDING
2036 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
2037 * ======== ================================================================
2038 *
2039 * Note:
2040 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2041 * interrupted while holding PENDING and @work off queue, irq must be
2042 * disabled on entry. This, combined with delayed_work->timer being
2043 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2044 *
2045 * On successful return, >= 0, irq is disabled and the caller is
2046 * responsible for releasing it using local_irq_restore(*@irq_flags).
2047 *
2048 * This function is safe to call from any context including IRQ handler.
2049 */
2050static int try_to_grab_pending(struct work_struct *work, u32 cflags,
2051 unsigned long *irq_flags)
2052{
2053 struct worker_pool *pool;
2054 struct pool_workqueue *pwq;
2055
2056 local_irq_save(*irq_flags);
2057
2058 /* try to steal the timer if it exists */
2059 if (cflags & WORK_CANCEL_DELAYED) {
2060 struct delayed_work *dwork = to_delayed_work(work);
2061
2062 /*
2063 * dwork->timer is irqsafe. If timer_delete() fails, it's
2064 * guaranteed that the timer is not queued anywhere and not
2065 * running on the local CPU.
2066 */
2067 if (likely(timer_delete(&dwork->timer)))
2068 return 1;
2069 }
2070
2071 /* try to claim PENDING the normal way */
2072 if (!test_and_set_bit(nr: WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2073 return 0;
2074
2075 rcu_read_lock();
2076 /*
2077 * The queueing is in progress, or it is already queued. Try to
2078 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2079 */
2080 pool = get_work_pool(work);
2081 if (!pool)
2082 goto fail;
2083
2084 raw_spin_lock(&pool->lock);
2085 /*
2086 * work->data is guaranteed to point to pwq only while the work
2087 * item is queued on pwq->wq, and both updating work->data to point
2088 * to pwq on queueing and to pool on dequeueing are done under
2089 * pwq->pool->lock. This in turn guarantees that, if work->data
2090 * points to pwq which is associated with a locked pool, the work
2091 * item is currently queued on that pool.
2092 */
2093 pwq = get_work_pwq(work);
2094 if (pwq && pwq->pool == pool) {
2095 unsigned long work_data = *work_data_bits(work);
2096
2097 debug_work_deactivate(work);
2098
2099 /*
2100 * A cancelable inactive work item must be in the
2101 * pwq->inactive_works since a queued barrier can't be
2102 * canceled (see the comments in insert_wq_barrier()).
2103 *
2104 * An inactive work item cannot be deleted directly because
2105 * it might have linked barrier work items which, if left
2106 * on the inactive_works list, will confuse pwq->nr_active
2107 * management later on and cause stall. Move the linked
2108 * barrier work items to the worklist when deleting the grabbed
2109 * item. Also keep WORK_STRUCT_INACTIVE in work_data, so that
2110 * it doesn't participate in nr_active management in later
2111 * pwq_dec_nr_in_flight().
2112 */
2113 if (work_data & WORK_STRUCT_INACTIVE)
2114 move_linked_works(work, head: &pwq->pool->worklist, NULL);
2115
2116 list_del_init(entry: &work->entry);
2117
2118 /*
2119 * work->data points to pwq iff queued. Let's point to pool. As
2120 * this destroys work->data needed by the next step, stash it.
2121 */
2122 set_work_pool_and_keep_pending(work, pool_id: pool->id,
2123 flags: pool_offq_flags(pool));
2124
2125 /* must be the last step, see the function comment */
2126 pwq_dec_nr_in_flight(pwq, work_data);
2127
2128 raw_spin_unlock(&pool->lock);
2129 rcu_read_unlock();
2130 return 1;
2131 }
2132 raw_spin_unlock(&pool->lock);
2133fail:
2134 rcu_read_unlock();
2135 local_irq_restore(*irq_flags);
2136 return -EAGAIN;
2137}
2138
2139/**
2140 * work_grab_pending - steal work item from worklist and disable irq
2141 * @work: work item to steal
2142 * @cflags: %WORK_CANCEL_ flags
2143 * @irq_flags: place to store IRQ state
2144 *
2145 * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer
2146 * or on worklist.
2147 *
2148 * Can be called from any context. IRQ is disabled on return with IRQ state
2149 * stored in *@irq_flags. The caller is responsible for re-enabling it using
2150 * local_irq_restore().
2151 *
2152 * Returns %true if @work was pending. %false if idle.
2153 */
2154static bool work_grab_pending(struct work_struct *work, u32 cflags,
2155 unsigned long *irq_flags)
2156{
2157 int ret;
2158
2159 while (true) {
2160 ret = try_to_grab_pending(work, cflags, irq_flags);
2161 if (ret >= 0)
2162 return ret;
2163 cpu_relax();
2164 }
2165}
2166
2167/**
2168 * insert_work - insert a work into a pool
2169 * @pwq: pwq @work belongs to
2170 * @work: work to insert
2171 * @head: insertion point
2172 * @extra_flags: extra WORK_STRUCT_* flags to set
2173 *
2174 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2175 * work_struct flags.
2176 *
2177 * CONTEXT:
2178 * raw_spin_lock_irq(pool->lock).
2179 */
2180static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2181 struct list_head *head, unsigned int extra_flags)
2182{
2183 debug_work_activate(work);
2184
2185 /* record the work call stack in order to print it in KASAN reports */
2186 kasan_record_aux_stack(ptr: work);
2187
2188 /* we own @work, set data and link */
2189 set_work_pwq(work, pwq, flags: extra_flags);
2190 list_add_tail(new: &work->entry, head);
2191 get_pwq(pwq);
2192}
2193
2194/*
2195 * Test whether @work is being queued from another work executing on the
2196 * same workqueue.
2197 */
2198static bool is_chained_work(struct workqueue_struct *wq)
2199{
2200 struct worker *worker;
2201
2202 worker = current_wq_worker();
2203 /*
2204 * Return %true iff I'm a worker executing a work item on @wq. If
2205 * I'm @worker, it's safe to dereference it without locking.
2206 */
2207 return worker && worker->current_pwq->wq == wq;
2208}
2209
2210/*
2211 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2212 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2213 * avoid perturbing sensitive tasks.
2214 */
2215static int wq_select_unbound_cpu(int cpu)
2216{
2217 int new_cpu;
2218
2219 if (likely(!wq_debug_force_rr_cpu)) {
2220 if (cpumask_test_cpu(cpu, cpumask: wq_unbound_cpumask))
2221 return cpu;
2222 } else {
2223 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2224 }
2225
2226 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2227 new_cpu = cpumask_next_and_wrap(n: new_cpu, src1p: wq_unbound_cpumask, cpu_online_mask);
2228 if (unlikely(new_cpu >= nr_cpu_ids))
2229 return cpu;
2230 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2231
2232 return new_cpu;
2233}
2234
2235static 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_ONCE(!is_chained_work(wq), "workqueue: cannot queue %ps on wq %s\n",
2258 work->func, wq->name))) {
2259 return;
2260 }
2261 rcu_read_lock();
2262retry:
2263 /* pwq which will be used unless @work is executing elsewhere */
2264 if (req_cpu == WORK_CPU_UNBOUND) {
2265 if (wq->flags & WQ_UNBOUND)
2266 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2267 else
2268 cpu = raw_smp_processor_id();
2269 }
2270
2271 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2272 pool = pwq->pool;
2273
2274 /*
2275 * If @work was previously on a different pool, it might still be
2276 * running there, in which case the work needs to be queued on that
2277 * pool to guarantee non-reentrancy.
2278 *
2279 * For ordered workqueue, work items must be queued on the newest pwq
2280 * for accurate order management. Guaranteed order also guarantees
2281 * non-reentrancy. See the comments above unplug_oldest_pwq().
2282 */
2283 last_pool = get_work_pool(work);
2284 if (last_pool && last_pool != pool && !(wq->flags & __WQ_ORDERED)) {
2285 struct worker *worker;
2286
2287 raw_spin_lock(&last_pool->lock);
2288
2289 worker = find_worker_executing_work(pool: last_pool, work);
2290
2291 if (worker && worker->current_pwq->wq == wq) {
2292 pwq = worker->current_pwq;
2293 pool = pwq->pool;
2294 WARN_ON_ONCE(pool != last_pool);
2295 } else {
2296 /* meh... not running there, queue here */
2297 raw_spin_unlock(&last_pool->lock);
2298 raw_spin_lock(&pool->lock);
2299 }
2300 } else {
2301 raw_spin_lock(&pool->lock);
2302 }
2303
2304 /*
2305 * pwq is determined and locked. For unbound pools, we could have raced
2306 * with pwq release and it could already be dead. If its refcnt is zero,
2307 * repeat pwq selection. Note that unbound pwqs never die without
2308 * another pwq replacing it in cpu_pwq or while work items are executing
2309 * on it, so the retrying is guaranteed to make forward-progress.
2310 */
2311 if (unlikely(!pwq->refcnt)) {
2312 if (wq->flags & WQ_UNBOUND) {
2313 raw_spin_unlock(&pool->lock);
2314 cpu_relax();
2315 goto retry;
2316 }
2317 /* oops */
2318 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2319 wq->name, cpu);
2320 }
2321
2322 /* pwq determined, queue */
2323 trace_workqueue_queue_work(req_cpu, pwq, work);
2324
2325 if (WARN_ON(!list_empty(&work->entry)))
2326 goto out;
2327
2328 pwq->nr_in_flight[pwq->work_color]++;
2329 work_flags = work_color_to_flags(color: pwq->work_color);
2330
2331 /*
2332 * Limit the number of concurrently active work items to max_active.
2333 * @work must also queue behind existing inactive work items to maintain
2334 * ordering when max_active changes. See wq_adjust_max_active().
2335 */
2336 if (list_empty(head: &pwq->inactive_works) && pwq_tryinc_nr_active(pwq, fill: false)) {
2337 if (list_empty(head: &pool->worklist))
2338 pool->watchdog_ts = jiffies;
2339
2340 trace_workqueue_activate_work(work);
2341 insert_work(pwq, work, head: &pool->worklist, extra_flags: work_flags);
2342 kick_pool(pool);
2343 } else {
2344 work_flags |= WORK_STRUCT_INACTIVE;
2345 insert_work(pwq, work, head: &pwq->inactive_works, extra_flags: work_flags);
2346 }
2347
2348out:
2349 raw_spin_unlock(&pool->lock);
2350 rcu_read_unlock();
2351}
2352
2353static bool clear_pending_if_disabled(struct work_struct *work)
2354{
2355 unsigned long data = *work_data_bits(work);
2356 struct work_offq_data offqd;
2357
2358 if (likely((data & WORK_STRUCT_PWQ) ||
2359 !(data & WORK_OFFQ_DISABLE_MASK)))
2360 return false;
2361
2362 work_offqd_unpack(offqd: &offqd, data);
2363 set_work_pool_and_clear_pending(work, pool_id: offqd.pool_id,
2364 flags: work_offqd_pack_flags(offqd: &offqd));
2365 return true;
2366}
2367
2368/**
2369 * queue_work_on - queue work on specific cpu
2370 * @cpu: CPU number to execute work on
2371 * @wq: workqueue to use
2372 * @work: work to queue
2373 *
2374 * We queue the work to a specific CPU, the caller must ensure it
2375 * can't go away. Callers that fail to ensure that the specified
2376 * CPU cannot go away will execute on a randomly chosen CPU.
2377 * But note well that callers specifying a CPU that never has been
2378 * online will get a splat.
2379 *
2380 * Return: %false if @work was already on a queue, %true otherwise.
2381 */
2382bool queue_work_on(int cpu, struct workqueue_struct *wq,
2383 struct work_struct *work)
2384{
2385 bool ret = false;
2386 unsigned long irq_flags;
2387
2388 local_irq_save(irq_flags);
2389
2390 if (!test_and_set_bit(nr: WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2391 !clear_pending_if_disabled(work)) {
2392 __queue_work(cpu, wq, work);
2393 ret = true;
2394 }
2395
2396 local_irq_restore(irq_flags);
2397 return ret;
2398}
2399EXPORT_SYMBOL(queue_work_on);
2400
2401/**
2402 * select_numa_node_cpu - Select a CPU based on NUMA node
2403 * @node: NUMA node ID that we want to select a CPU from
2404 *
2405 * This function will attempt to find a "random" cpu available on a given
2406 * node. If there are no CPUs available on the given node it will return
2407 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2408 * available CPU if we need to schedule this work.
2409 */
2410static int select_numa_node_cpu(int node)
2411{
2412 int cpu;
2413
2414 /* Delay binding to CPU if node is not valid or online */
2415 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2416 return WORK_CPU_UNBOUND;
2417
2418 /* Use local node/cpu if we are already there */
2419 cpu = raw_smp_processor_id();
2420 if (node == cpu_to_node(cpu))
2421 return cpu;
2422
2423 /* Use "random" otherwise know as "first" online CPU of node */
2424 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2425
2426 /* If CPU is valid return that, otherwise just defer */
2427 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2428}
2429
2430/**
2431 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2432 * @node: NUMA node that we are targeting the work for
2433 * @wq: workqueue to use
2434 * @work: work to queue
2435 *
2436 * We queue the work to a "random" CPU within a given NUMA node. The basic
2437 * idea here is to provide a way to somehow associate work with a given
2438 * NUMA node.
2439 *
2440 * This function will only make a best effort attempt at getting this onto
2441 * the right NUMA node. If no node is requested or the requested node is
2442 * offline then we just fall back to standard queue_work behavior.
2443 *
2444 * Currently the "random" CPU ends up being the first available CPU in the
2445 * intersection of cpu_online_mask and the cpumask of the node, unless we
2446 * are running on the node. In that case we just use the current CPU.
2447 *
2448 * Return: %false if @work was already on a queue, %true otherwise.
2449 */
2450bool queue_work_node(int node, struct workqueue_struct *wq,
2451 struct work_struct *work)
2452{
2453 unsigned long irq_flags;
2454 bool ret = false;
2455
2456 /*
2457 * This current implementation is specific to unbound workqueues.
2458 * Specifically we only return the first available CPU for a given
2459 * node instead of cycling through individual CPUs within the node.
2460 *
2461 * If this is used with a per-cpu workqueue then the logic in
2462 * workqueue_select_cpu_near would need to be updated to allow for
2463 * some round robin type logic.
2464 */
2465 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2466
2467 local_irq_save(irq_flags);
2468
2469 if (!test_and_set_bit(nr: WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2470 !clear_pending_if_disabled(work)) {
2471 int cpu = select_numa_node_cpu(node);
2472
2473 __queue_work(cpu, wq, work);
2474 ret = true;
2475 }
2476
2477 local_irq_restore(irq_flags);
2478 return ret;
2479}
2480EXPORT_SYMBOL_GPL(queue_work_node);
2481
2482void delayed_work_timer_fn(struct timer_list *t)
2483{
2484 struct delayed_work *dwork = timer_container_of(dwork, t, timer);
2485
2486 /* should have been called from irqsafe timer with irq already off */
2487 __queue_work(cpu: dwork->cpu, wq: dwork->wq, work: &dwork->work);
2488}
2489EXPORT_SYMBOL(delayed_work_timer_fn);
2490
2491static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2492 struct delayed_work *dwork, unsigned long delay)
2493{
2494 struct timer_list *timer = &dwork->timer;
2495 struct work_struct *work = &dwork->work;
2496
2497 WARN_ON_ONCE(!wq);
2498 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2499 WARN_ON_ONCE(timer_pending(timer));
2500 WARN_ON_ONCE(!list_empty(&work->entry));
2501
2502 /*
2503 * If @delay is 0, queue @dwork->work immediately. This is for
2504 * both optimization and correctness. The earliest @timer can
2505 * expire is on the closest next tick and delayed_work users depend
2506 * on that there's no such delay when @delay is 0.
2507 */
2508 if (!delay) {
2509 __queue_work(cpu, wq, work: &dwork->work);
2510 return;
2511 }
2512
2513 WARN_ON_ONCE(cpu != WORK_CPU_UNBOUND && !cpu_online(cpu));
2514 dwork->wq = wq;
2515 dwork->cpu = cpu;
2516 timer->expires = jiffies + delay;
2517
2518 if (housekeeping_enabled(type: HK_TYPE_TIMER)) {
2519 /* If the current cpu is a housekeeping cpu, use it. */
2520 cpu = smp_processor_id();
2521 if (!housekeeping_test_cpu(cpu, type: HK_TYPE_TIMER))
2522 cpu = housekeeping_any_cpu(type: HK_TYPE_TIMER);
2523 add_timer_on(timer, cpu);
2524 } else {
2525 if (likely(cpu == WORK_CPU_UNBOUND))
2526 add_timer_global(timer);
2527 else
2528 add_timer_on(timer, cpu);
2529 }
2530}
2531
2532/**
2533 * queue_delayed_work_on - queue work on specific CPU after delay
2534 * @cpu: CPU number to execute work on
2535 * @wq: workqueue to use
2536 * @dwork: work to queue
2537 * @delay: number of jiffies to wait before queueing
2538 *
2539 * We queue the delayed_work to a specific CPU, for non-zero delays the
2540 * caller must ensure it is online and can't go away. Callers that fail
2541 * to ensure this, may get @dwork->timer queued to an offlined CPU and
2542 * this will prevent queueing of @dwork->work unless the offlined CPU
2543 * becomes online again.
2544 *
2545 * Return: %false if @work was already on a queue, %true otherwise. If
2546 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2547 * execution.
2548 */
2549bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2550 struct delayed_work *dwork, unsigned long delay)
2551{
2552 struct work_struct *work = &dwork->work;
2553 bool ret = false;
2554 unsigned long irq_flags;
2555
2556 /* read the comment in __queue_work() */
2557 local_irq_save(irq_flags);
2558
2559 if (!test_and_set_bit(nr: WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2560 !clear_pending_if_disabled(work)) {
2561 __queue_delayed_work(cpu, wq, dwork, delay);
2562 ret = true;
2563 }
2564
2565 local_irq_restore(irq_flags);
2566 return ret;
2567}
2568EXPORT_SYMBOL(queue_delayed_work_on);
2569
2570/**
2571 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2572 * @cpu: CPU number to execute work on
2573 * @wq: workqueue to use
2574 * @dwork: work to queue
2575 * @delay: number of jiffies to wait before queueing
2576 *
2577 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2578 * modify @dwork's timer so that it expires after @delay. If @delay is
2579 * zero, @work is guaranteed to be scheduled immediately regardless of its
2580 * current state.
2581 *
2582 * Return: %false if @dwork was idle and queued, %true if @dwork was
2583 * pending and its timer was modified.
2584 *
2585 * This function is safe to call from any context including IRQ handler.
2586 * See try_to_grab_pending() for details.
2587 */
2588bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2589 struct delayed_work *dwork, unsigned long delay)
2590{
2591 unsigned long irq_flags;
2592 bool ret;
2593
2594 ret = work_grab_pending(work: &dwork->work, cflags: WORK_CANCEL_DELAYED, irq_flags: &irq_flags);
2595
2596 if (!clear_pending_if_disabled(work: &dwork->work))
2597 __queue_delayed_work(cpu, wq, dwork, delay);
2598
2599 local_irq_restore(irq_flags);
2600 return ret;
2601}
2602EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2603
2604static void rcu_work_rcufn(struct rcu_head *rcu)
2605{
2606 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2607
2608 /* read the comment in __queue_work() */
2609 local_irq_disable();
2610 __queue_work(cpu: WORK_CPU_UNBOUND, wq: rwork->wq, work: &rwork->work);
2611 local_irq_enable();
2612}
2613
2614/**
2615 * queue_rcu_work - queue work after a RCU grace period
2616 * @wq: workqueue to use
2617 * @rwork: work to queue
2618 *
2619 * Return: %false if @rwork was already pending, %true otherwise. Note
2620 * that a full RCU grace period is guaranteed only after a %true return.
2621 * While @rwork is guaranteed to be executed after a %false return, the
2622 * execution may happen before a full RCU grace period has passed.
2623 */
2624bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2625{
2626 struct work_struct *work = &rwork->work;
2627
2628 /*
2629 * rcu_work can't be canceled or disabled. Warn if the user reached
2630 * inside @rwork and disabled the inner work.
2631 */
2632 if (!test_and_set_bit(nr: WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2633 !WARN_ON_ONCE(clear_pending_if_disabled(work))) {
2634 rwork->wq = wq;
2635 call_rcu_hurry(head: &rwork->rcu, func: rcu_work_rcufn);
2636 return true;
2637 }
2638
2639 return false;
2640}
2641EXPORT_SYMBOL(queue_rcu_work);
2642
2643static struct worker *alloc_worker(int node)
2644{
2645 struct worker *worker;
2646
2647 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2648 if (worker) {
2649 INIT_LIST_HEAD(list: &worker->entry);
2650 INIT_LIST_HEAD(list: &worker->scheduled);
2651 INIT_LIST_HEAD(list: &worker->node);
2652 /* on creation a worker is in !idle && prep state */
2653 worker->flags = WORKER_PREP;
2654 }
2655 return worker;
2656}
2657
2658static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2659{
2660 if (pool->cpu < 0 && pool->attrs->affn_strict)
2661 return pool->attrs->__pod_cpumask;
2662 else
2663 return pool->attrs->cpumask;
2664}
2665
2666/**
2667 * worker_attach_to_pool() - attach a worker to a pool
2668 * @worker: worker to be attached
2669 * @pool: the target pool
2670 *
2671 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2672 * cpu-binding of @worker are kept coordinated with the pool across
2673 * cpu-[un]hotplugs.
2674 */
2675static void worker_attach_to_pool(struct worker *worker,
2676 struct worker_pool *pool)
2677{
2678 mutex_lock(lock: &wq_pool_attach_mutex);
2679
2680 /*
2681 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable
2682 * across this function. See the comments above the flag definition for
2683 * details. BH workers are, while per-CPU, always DISASSOCIATED.
2684 */
2685 if (pool->flags & POOL_DISASSOCIATED) {
2686 worker->flags |= WORKER_UNBOUND;
2687 } else {
2688 WARN_ON_ONCE(pool->flags & POOL_BH);
2689 kthread_set_per_cpu(k: worker->task, cpu: pool->cpu);
2690 }
2691
2692 if (worker->rescue_wq)
2693 set_cpus_allowed_ptr(p: worker->task, new_mask: pool_allowed_cpus(pool));
2694
2695 list_add_tail(new: &worker->node, head: &pool->workers);
2696 worker->pool = pool;
2697
2698 mutex_unlock(lock: &wq_pool_attach_mutex);
2699}
2700
2701static void unbind_worker(struct worker *worker)
2702{
2703 lockdep_assert_held(&wq_pool_attach_mutex);
2704
2705 kthread_set_per_cpu(k: worker->task, cpu: -1);
2706 if (cpumask_intersects(src1p: wq_unbound_cpumask, cpu_active_mask))
2707 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2708 else
2709 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2710}
2711
2712
2713static void detach_worker(struct worker *worker)
2714{
2715 lockdep_assert_held(&wq_pool_attach_mutex);
2716
2717 unbind_worker(worker);
2718 list_del(entry: &worker->node);
2719}
2720
2721/**
2722 * worker_detach_from_pool() - detach a worker from its pool
2723 * @worker: worker which is attached to its pool
2724 *
2725 * Undo the attaching which had been done in worker_attach_to_pool(). The
2726 * caller worker shouldn't access to the pool after detached except it has
2727 * other reference to the pool.
2728 */
2729static void worker_detach_from_pool(struct worker *worker)
2730{
2731 struct worker_pool *pool = worker->pool;
2732
2733 /* there is one permanent BH worker per CPU which should never detach */
2734 WARN_ON_ONCE(pool->flags & POOL_BH);
2735
2736 mutex_lock(lock: &wq_pool_attach_mutex);
2737 detach_worker(worker);
2738 worker->pool = NULL;
2739 mutex_unlock(lock: &wq_pool_attach_mutex);
2740
2741 /* clear leftover flags without pool->lock after it is detached */
2742 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2743}
2744
2745static int format_worker_id(char *buf, size_t size, struct worker *worker,
2746 struct worker_pool *pool)
2747{
2748 if (worker->rescue_wq)
2749 return scnprintf(buf, size, fmt: "kworker/R-%s",
2750 worker->rescue_wq->name);
2751
2752 if (pool) {
2753 if (pool->cpu >= 0)
2754 return scnprintf(buf, size, fmt: "kworker/%d:%d%s",
2755 pool->cpu, worker->id,
2756 pool->attrs->nice < 0 ? "H" : "");
2757 else
2758 return scnprintf(buf, size, fmt: "kworker/u%d:%d",
2759 pool->id, worker->id);
2760 } else {
2761 return scnprintf(buf, size, fmt: "kworker/dying");
2762 }
2763}
2764
2765/**
2766 * create_worker - create a new workqueue worker
2767 * @pool: pool the new worker will belong to
2768 *
2769 * Create and start a new worker which is attached to @pool.
2770 *
2771 * CONTEXT:
2772 * Might sleep. Does GFP_KERNEL allocations.
2773 *
2774 * Return:
2775 * Pointer to the newly created worker.
2776 */
2777static struct worker *create_worker(struct worker_pool *pool)
2778{
2779 struct worker *worker;
2780 int id;
2781
2782 /* ID is needed to determine kthread name */
2783 id = ida_alloc(ida: &pool->worker_ida, GFP_KERNEL);
2784 if (id < 0) {
2785 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2786 ERR_PTR(id));
2787 return NULL;
2788 }
2789
2790 worker = alloc_worker(node: pool->node);
2791 if (!worker) {
2792 pr_err_once("workqueue: Failed to allocate a worker\n");
2793 goto fail;
2794 }
2795
2796 worker->id = id;
2797
2798 if (!(pool->flags & POOL_BH)) {
2799 char id_buf[WORKER_ID_LEN];
2800
2801 format_worker_id(buf: id_buf, size: sizeof(id_buf), worker, pool);
2802 worker->task = kthread_create_on_node(threadfn: worker_thread, data: worker,
2803 node: pool->node, namefmt: "%s", id_buf);
2804 if (IS_ERR(ptr: worker->task)) {
2805 if (PTR_ERR(ptr: worker->task) == -EINTR) {
2806 pr_err("workqueue: Interrupted when creating a worker thread \"%s\"\n",
2807 id_buf);
2808 } else {
2809 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2810 worker->task);
2811 }
2812 goto fail;
2813 }
2814
2815 set_user_nice(p: worker->task, nice: pool->attrs->nice);
2816 kthread_bind_mask(k: worker->task, mask: pool_allowed_cpus(pool));
2817 }
2818
2819 /* successful, attach the worker to the pool */
2820 worker_attach_to_pool(worker, pool);
2821
2822 /* start the newly created worker */
2823 raw_spin_lock_irq(&pool->lock);
2824
2825 worker->pool->nr_workers++;
2826 worker_enter_idle(worker);
2827
2828 /*
2829 * @worker is waiting on a completion in kthread() and will trigger hung
2830 * check if not woken up soon. As kick_pool() is noop if @pool is empty,
2831 * wake it up explicitly.
2832 */
2833 if (worker->task)
2834 wake_up_process(tsk: worker->task);
2835
2836 raw_spin_unlock_irq(&pool->lock);
2837
2838 return worker;
2839
2840fail:
2841 ida_free(&pool->worker_ida, id);
2842 kfree(objp: worker);
2843 return NULL;
2844}
2845
2846static void detach_dying_workers(struct list_head *cull_list)
2847{
2848 struct worker *worker;
2849
2850 list_for_each_entry(worker, cull_list, entry)
2851 detach_worker(worker);
2852}
2853
2854static void reap_dying_workers(struct list_head *cull_list)
2855{
2856 struct worker *worker, *tmp;
2857
2858 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2859 list_del_init(entry: &worker->entry);
2860 kthread_stop_put(k: worker->task);
2861 kfree(objp: worker);
2862 }
2863}
2864
2865/**
2866 * set_worker_dying - Tag a worker for destruction
2867 * @worker: worker to be destroyed
2868 * @list: transfer worker away from its pool->idle_list and into list
2869 *
2870 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2871 * should be idle.
2872 *
2873 * CONTEXT:
2874 * raw_spin_lock_irq(pool->lock).
2875 */
2876static void set_worker_dying(struct worker *worker, struct list_head *list)
2877{
2878 struct worker_pool *pool = worker->pool;
2879
2880 lockdep_assert_held(&pool->lock);
2881 lockdep_assert_held(&wq_pool_attach_mutex);
2882
2883 /* sanity check frenzy */
2884 if (WARN_ON(worker->current_work) ||
2885 WARN_ON(!list_empty(&worker->scheduled)) ||
2886 WARN_ON(!(worker->flags & WORKER_IDLE)))
2887 return;
2888
2889 pool->nr_workers--;
2890 pool->nr_idle--;
2891
2892 worker->flags |= WORKER_DIE;
2893
2894 list_move(list: &worker->entry, head: list);
2895
2896 /* get an extra task struct reference for later kthread_stop_put() */
2897 get_task_struct(t: worker->task);
2898}
2899
2900/**
2901 * idle_worker_timeout - check if some idle workers can now be deleted.
2902 * @t: The pool's idle_timer that just expired
2903 *
2904 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2905 * worker_leave_idle(), as a worker flicking between idle and active while its
2906 * pool is at the too_many_workers() tipping point would cause too much timer
2907 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2908 * it expire and re-evaluate things from there.
2909 */
2910static void idle_worker_timeout(struct timer_list *t)
2911{
2912 struct worker_pool *pool = timer_container_of(pool, t, idle_timer);
2913 bool do_cull = false;
2914
2915 if (work_pending(&pool->idle_cull_work))
2916 return;
2917
2918 raw_spin_lock_irq(&pool->lock);
2919
2920 if (too_many_workers(pool)) {
2921 struct worker *worker;
2922 unsigned long expires;
2923
2924 /* idle_list is kept in LIFO order, check the last one */
2925 worker = list_last_entry(&pool->idle_list, struct worker, entry);
2926 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2927 do_cull = !time_before(jiffies, expires);
2928
2929 if (!do_cull)
2930 mod_timer(timer: &pool->idle_timer, expires);
2931 }
2932 raw_spin_unlock_irq(&pool->lock);
2933
2934 if (do_cull)
2935 queue_work(wq: system_dfl_wq, work: &pool->idle_cull_work);
2936}
2937
2938/**
2939 * idle_cull_fn - cull workers that have been idle for too long.
2940 * @work: the pool's work for handling these idle workers
2941 *
2942 * This goes through a pool's idle workers and gets rid of those that have been
2943 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2944 *
2945 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2946 * culled, so this also resets worker affinity. This requires a sleepable
2947 * context, hence the split between timer callback and work item.
2948 */
2949static void idle_cull_fn(struct work_struct *work)
2950{
2951 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2952 LIST_HEAD(cull_list);
2953
2954 /*
2955 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2956 * cannot proceed beyong set_pf_worker() in its self-destruct path.
2957 * This is required as a previously-preempted worker could run after
2958 * set_worker_dying() has happened but before detach_dying_workers() did.
2959 */
2960 mutex_lock(lock: &wq_pool_attach_mutex);
2961 raw_spin_lock_irq(&pool->lock);
2962
2963 while (too_many_workers(pool)) {
2964 struct worker *worker;
2965 unsigned long expires;
2966
2967 worker = list_last_entry(&pool->idle_list, struct worker, entry);
2968 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2969
2970 if (time_before(jiffies, expires)) {
2971 mod_timer(timer: &pool->idle_timer, expires);
2972 break;
2973 }
2974
2975 set_worker_dying(worker, list: &cull_list);
2976 }
2977
2978 raw_spin_unlock_irq(&pool->lock);
2979 detach_dying_workers(cull_list: &cull_list);
2980 mutex_unlock(lock: &wq_pool_attach_mutex);
2981
2982 reap_dying_workers(cull_list: &cull_list);
2983}
2984
2985static void send_mayday(struct work_struct *work)
2986{
2987 struct pool_workqueue *pwq = get_work_pwq(work);
2988 struct workqueue_struct *wq = pwq->wq;
2989
2990 lockdep_assert_held(&wq_mayday_lock);
2991
2992 if (!wq->rescuer)
2993 return;
2994
2995 /* mayday mayday mayday */
2996 if (list_empty(head: &pwq->mayday_node)) {
2997 /*
2998 * If @pwq is for an unbound wq, its base ref may be put at
2999 * any time due to an attribute change. Pin @pwq until the
3000 * rescuer is done with it.
3001 */
3002 get_pwq(pwq);
3003 list_add_tail(new: &pwq->mayday_node, head: &wq->maydays);
3004 wake_up_process(tsk: wq->rescuer->task);
3005 pwq->stats[PWQ_STAT_MAYDAY]++;
3006 }
3007}
3008
3009static void pool_mayday_timeout(struct timer_list *t)
3010{
3011 struct worker_pool *pool = timer_container_of(pool, t, mayday_timer);
3012 struct work_struct *work;
3013
3014 raw_spin_lock_irq(&pool->lock);
3015 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
3016
3017 if (need_to_create_worker(pool)) {
3018 /*
3019 * We've been trying to create a new worker but
3020 * haven't been successful. We might be hitting an
3021 * allocation deadlock. Send distress signals to
3022 * rescuers.
3023 */
3024 list_for_each_entry(work, &pool->worklist, entry)
3025 send_mayday(work);
3026 }
3027
3028 raw_spin_unlock(&wq_mayday_lock);
3029 raw_spin_unlock_irq(&pool->lock);
3030
3031 mod_timer(timer: &pool->mayday_timer, expires: jiffies + MAYDAY_INTERVAL);
3032}
3033
3034/**
3035 * maybe_create_worker - create a new worker if necessary
3036 * @pool: pool to create a new worker for
3037 *
3038 * Create a new worker for @pool if necessary. @pool is guaranteed to
3039 * have at least one idle worker on return from this function. If
3040 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3041 * sent to all rescuers with works scheduled on @pool to resolve
3042 * possible allocation deadlock.
3043 *
3044 * On return, need_to_create_worker() is guaranteed to be %false and
3045 * may_start_working() %true.
3046 *
3047 * LOCKING:
3048 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3049 * multiple times. Does GFP_KERNEL allocations. Called only from
3050 * manager.
3051 */
3052static void maybe_create_worker(struct worker_pool *pool)
3053__releases(&pool->lock)
3054__acquires(&pool->lock)
3055{
3056restart:
3057 raw_spin_unlock_irq(&pool->lock);
3058
3059 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3060 mod_timer(timer: &pool->mayday_timer, expires: jiffies + MAYDAY_INITIAL_TIMEOUT);
3061
3062 while (true) {
3063 if (create_worker(pool) || !need_to_create_worker(pool))
3064 break;
3065
3066 schedule_timeout_interruptible(timeout: CREATE_COOLDOWN);
3067
3068 if (!need_to_create_worker(pool))
3069 break;
3070 }
3071
3072 timer_delete_sync(timer: &pool->mayday_timer);
3073 raw_spin_lock_irq(&pool->lock);
3074 /*
3075 * This is necessary even after a new worker was just successfully
3076 * created as @pool->lock was dropped and the new worker might have
3077 * already become busy.
3078 */
3079 if (need_to_create_worker(pool))
3080 goto restart;
3081}
3082
3083#ifdef CONFIG_PREEMPT_RT
3084static void worker_lock_callback(struct worker_pool *pool)
3085{
3086 spin_lock(&pool->cb_lock);
3087}
3088
3089static void worker_unlock_callback(struct worker_pool *pool)
3090{
3091 spin_unlock(&pool->cb_lock);
3092}
3093
3094static void workqueue_callback_cancel_wait_running(struct worker_pool *pool)
3095{
3096 spin_lock(&pool->cb_lock);
3097 spin_unlock(&pool->cb_lock);
3098}
3099
3100#else
3101
3102static void worker_lock_callback(struct worker_pool *pool) { }
3103static void worker_unlock_callback(struct worker_pool *pool) { }
3104static void workqueue_callback_cancel_wait_running(struct worker_pool *pool) { }
3105
3106#endif
3107
3108/**
3109 * manage_workers - manage worker pool
3110 * @worker: self
3111 *
3112 * Assume the manager role and manage the worker pool @worker belongs
3113 * to. At any given time, there can be only zero or one manager per
3114 * pool. The exclusion is handled automatically by this function.
3115 *
3116 * The caller can safely start processing works on false return. On
3117 * true return, it's guaranteed that need_to_create_worker() is false
3118 * and may_start_working() is true.
3119 *
3120 * CONTEXT:
3121 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3122 * multiple times. Does GFP_KERNEL allocations.
3123 *
3124 * Return:
3125 * %false if the pool doesn't need management and the caller can safely
3126 * start processing works, %true if management function was performed and
3127 * the conditions that the caller verified before calling the function may
3128 * no longer be true.
3129 */
3130static bool manage_workers(struct worker *worker)
3131{
3132 struct worker_pool *pool = worker->pool;
3133
3134 if (pool->flags & POOL_MANAGER_ACTIVE)
3135 return false;
3136
3137 pool->flags |= POOL_MANAGER_ACTIVE;
3138 pool->manager = worker;
3139
3140 maybe_create_worker(pool);
3141
3142 pool->manager = NULL;
3143 pool->flags &= ~POOL_MANAGER_ACTIVE;
3144 rcuwait_wake_up(w: &manager_wait);
3145 return true;
3146}
3147
3148/**
3149 * process_one_work - process single work
3150 * @worker: self
3151 * @work: work to process
3152 *
3153 * Process @work. This function contains all the logics necessary to
3154 * process a single work including synchronization against and
3155 * interaction with other workers on the same cpu, queueing and
3156 * flushing. As long as context requirement is met, any worker can
3157 * call this function to process a work.
3158 *
3159 * CONTEXT:
3160 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3161 */
3162static void process_one_work(struct worker *worker, struct work_struct *work)
3163__releases(&pool->lock)
3164__acquires(&pool->lock)
3165{
3166 struct pool_workqueue *pwq = get_work_pwq(work);
3167 struct worker_pool *pool = worker->pool;
3168 unsigned long work_data;
3169 int lockdep_start_depth, rcu_start_depth;
3170 bool bh_draining = pool->flags & POOL_BH_DRAINING;
3171#ifdef CONFIG_LOCKDEP
3172 /*
3173 * It is permissible to free the struct work_struct from
3174 * inside the function that is called from it, this we need to
3175 * take into account for lockdep too. To avoid bogus "held
3176 * lock freed" warnings as well as problems when looking into
3177 * work->lockdep_map, make a copy and use that here.
3178 */
3179 struct lockdep_map lockdep_map;
3180
3181 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3182#endif
3183 /* ensure we're on the correct CPU */
3184 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3185 raw_smp_processor_id() != pool->cpu);
3186
3187 /* claim and dequeue */
3188 debug_work_deactivate(work);
3189 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3190 worker->current_work = work;
3191 worker->current_func = work->func;
3192 worker->current_pwq = pwq;
3193 if (worker->task)
3194 worker->current_at = worker->task->se.sum_exec_runtime;
3195 work_data = *work_data_bits(work);
3196 worker->current_color = get_work_color(work_data);
3197
3198 /*
3199 * Record wq name for cmdline and debug reporting, may get
3200 * overridden through set_worker_desc().
3201 */
3202 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3203
3204 list_del_init(entry: &work->entry);
3205
3206 /*
3207 * CPU intensive works don't participate in concurrency management.
3208 * They're the scheduler's responsibility. This takes @worker out
3209 * of concurrency management and the next code block will chain
3210 * execution of the pending work items.
3211 */
3212 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3213 worker_set_flags(worker, flags: WORKER_CPU_INTENSIVE);
3214
3215 /*
3216 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3217 * since nr_running would always be >= 1 at this point. This is used to
3218 * chain execution of the pending work items for WORKER_NOT_RUNNING
3219 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3220 */
3221 kick_pool(pool);
3222
3223 /*
3224 * Record the last pool and clear PENDING which should be the last
3225 * update to @work. Also, do this inside @pool->lock so that
3226 * PENDING and queued state changes happen together while IRQ is
3227 * disabled.
3228 */
3229 set_work_pool_and_clear_pending(work, pool_id: pool->id, flags: pool_offq_flags(pool));
3230
3231 pwq->stats[PWQ_STAT_STARTED]++;
3232 raw_spin_unlock_irq(&pool->lock);
3233
3234 rcu_start_depth = rcu_preempt_depth();
3235 lockdep_start_depth = lockdep_depth(current);
3236 /* see drain_dead_softirq_workfn() */
3237 if (!bh_draining)
3238 lock_map_acquire(pwq->wq->lockdep_map);
3239 lock_map_acquire(&lockdep_map);
3240 /*
3241 * Strictly speaking we should mark the invariant state without holding
3242 * any locks, that is, before these two lock_map_acquire()'s.
3243 *
3244 * However, that would result in:
3245 *
3246 * A(W1)
3247 * WFC(C)
3248 * A(W1)
3249 * C(C)
3250 *
3251 * Which would create W1->C->W1 dependencies, even though there is no
3252 * actual deadlock possible. There are two solutions, using a
3253 * read-recursive acquire on the work(queue) 'locks', but this will then
3254 * hit the lockdep limitation on recursive locks, or simply discard
3255 * these locks.
3256 *
3257 * AFAICT there is no possible deadlock scenario between the
3258 * flush_work() and complete() primitives (except for single-threaded
3259 * workqueues), so hiding them isn't a problem.
3260 */
3261 lockdep_invariant_state(force: true);
3262 trace_workqueue_execute_start(work);
3263 worker->current_func(work);
3264 /*
3265 * While we must be careful to not use "work" after this, the trace
3266 * point will only record its address.
3267 */
3268 trace_workqueue_execute_end(work, function: worker->current_func);
3269
3270 lock_map_release(&lockdep_map);
3271 if (!bh_draining)
3272 lock_map_release(pwq->wq->lockdep_map);
3273
3274 if (unlikely((worker->task && in_atomic()) ||
3275 lockdep_depth(current) != lockdep_start_depth ||
3276 rcu_preempt_depth() != rcu_start_depth)) {
3277 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n"
3278 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n",
3279 current->comm, task_pid_nr(current), preempt_count(),
3280 lockdep_start_depth, lockdep_depth(current),
3281 rcu_start_depth, rcu_preempt_depth(),
3282 worker->current_func);
3283 debug_show_held_locks(current);
3284 dump_stack();
3285 }
3286
3287 /*
3288 * The following prevents a kworker from hogging CPU on !PREEMPTION
3289 * kernels, where a requeueing work item waiting for something to
3290 * happen could deadlock with stop_machine as such work item could
3291 * indefinitely requeue itself while all other CPUs are trapped in
3292 * stop_machine. At the same time, report a quiescent RCU state so
3293 * the same condition doesn't freeze RCU.
3294 */
3295 if (worker->task)
3296 cond_resched();
3297
3298 raw_spin_lock_irq(&pool->lock);
3299
3300 pwq->stats[PWQ_STAT_COMPLETED]++;
3301
3302 /*
3303 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3304 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3305 * wq_cpu_intensive_thresh_us. Clear it.
3306 */
3307 worker_clr_flags(worker, flags: WORKER_CPU_INTENSIVE);
3308
3309 /* tag the worker for identification in schedule() */
3310 worker->last_func = worker->current_func;
3311
3312 /* we're done with it, release */
3313 hash_del(node: &worker->hentry);
3314 worker->current_work = NULL;
3315 worker->current_func = NULL;
3316 worker->current_pwq = NULL;
3317 worker->current_color = INT_MAX;
3318
3319 /* must be the last step, see the function comment */
3320 pwq_dec_nr_in_flight(pwq, work_data);
3321}
3322
3323/**
3324 * process_scheduled_works - process scheduled works
3325 * @worker: self
3326 *
3327 * Process all scheduled works. Please note that the scheduled list
3328 * may change while processing a work, so this function repeatedly
3329 * fetches a work from the top and executes it.
3330 *
3331 * CONTEXT:
3332 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3333 * multiple times.
3334 */
3335static void process_scheduled_works(struct worker *worker)
3336{
3337 struct work_struct *work;
3338 bool first = true;
3339
3340 while ((work = list_first_entry_or_null(&worker->scheduled,
3341 struct work_struct, entry))) {
3342 if (first) {
3343 worker->pool->watchdog_ts = jiffies;
3344 first = false;
3345 }
3346 process_one_work(worker, work);
3347 }
3348}
3349
3350static void set_pf_worker(bool val)
3351{
3352 mutex_lock(lock: &wq_pool_attach_mutex);
3353 if (val)
3354 current->flags |= PF_WQ_WORKER;
3355 else
3356 current->flags &= ~PF_WQ_WORKER;
3357 mutex_unlock(lock: &wq_pool_attach_mutex);
3358}
3359
3360/**
3361 * worker_thread - the worker thread function
3362 * @__worker: self
3363 *
3364 * The worker thread function. All workers belong to a worker_pool -
3365 * either a per-cpu one or dynamic unbound one. These workers process all
3366 * work items regardless of their specific target workqueue. The only
3367 * exception is work items which belong to workqueues with a rescuer which
3368 * will be explained in rescuer_thread().
3369 *
3370 * Return: 0
3371 */
3372static int worker_thread(void *__worker)
3373{
3374 struct worker *worker = __worker;
3375 struct worker_pool *pool = worker->pool;
3376
3377 /* tell the scheduler that this is a workqueue worker */
3378 set_pf_worker(true);
3379woke_up:
3380 raw_spin_lock_irq(&pool->lock);
3381
3382 /* am I supposed to die? */
3383 if (unlikely(worker->flags & WORKER_DIE)) {
3384 raw_spin_unlock_irq(&pool->lock);
3385 set_pf_worker(false);
3386 /*
3387 * The worker is dead and PF_WQ_WORKER is cleared, worker->pool
3388 * shouldn't be accessed, reset it to NULL in case otherwise.
3389 */
3390 worker->pool = NULL;
3391 ida_free(&pool->worker_ida, id: worker->id);
3392 return 0;
3393 }
3394
3395 worker_leave_idle(worker);
3396recheck:
3397 /* no more worker necessary? */
3398 if (!need_more_worker(pool))
3399 goto sleep;
3400
3401 /* do we need to manage? */
3402 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3403 goto recheck;
3404
3405 /*
3406 * ->scheduled list can only be filled while a worker is
3407 * preparing to process a work or actually processing it.
3408 * Make sure nobody diddled with it while I was sleeping.
3409 */
3410 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3411
3412 /*
3413 * Finish PREP stage. We're guaranteed to have at least one idle
3414 * worker or that someone else has already assumed the manager
3415 * role. This is where @worker starts participating in concurrency
3416 * management if applicable and concurrency management is restored
3417 * after being rebound. See rebind_workers() for details.
3418 */
3419 worker_clr_flags(worker, flags: WORKER_PREP | WORKER_REBOUND);
3420
3421 do {
3422 struct work_struct *work =
3423 list_first_entry(&pool->worklist,
3424 struct work_struct, entry);
3425
3426 if (assign_work(work, worker, NULL))
3427 process_scheduled_works(worker);
3428 } while (keep_working(pool));
3429
3430 worker_set_flags(worker, flags: WORKER_PREP);
3431sleep:
3432 /*
3433 * pool->lock is held and there's no work to process and no need to
3434 * manage, sleep. Workers are woken up only while holding
3435 * pool->lock or from local cpu, so setting the current state
3436 * before releasing pool->lock is enough to prevent losing any
3437 * event.
3438 */
3439 worker_enter_idle(worker);
3440 __set_current_state(TASK_IDLE);
3441 raw_spin_unlock_irq(&pool->lock);
3442 schedule();
3443 goto woke_up;
3444}
3445
3446/**
3447 * rescuer_thread - the rescuer thread function
3448 * @__rescuer: self
3449 *
3450 * Workqueue rescuer thread function. There's one rescuer for each
3451 * workqueue which has WQ_MEM_RECLAIM set.
3452 *
3453 * Regular work processing on a pool may block trying to create a new
3454 * worker which uses GFP_KERNEL allocation which has slight chance of
3455 * developing into deadlock if some works currently on the same queue
3456 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3457 * the problem rescuer solves.
3458 *
3459 * When such condition is possible, the pool summons rescuers of all
3460 * workqueues which have works queued on the pool and let them process
3461 * those works so that forward progress can be guaranteed.
3462 *
3463 * This should happen rarely.
3464 *
3465 * Return: 0
3466 */
3467static int rescuer_thread(void *__rescuer)
3468{
3469 struct worker *rescuer = __rescuer;
3470 struct workqueue_struct *wq = rescuer->rescue_wq;
3471 bool should_stop;
3472
3473 set_user_nice(current, nice: RESCUER_NICE_LEVEL);
3474
3475 /*
3476 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3477 * doesn't participate in concurrency management.
3478 */
3479 set_pf_worker(true);
3480repeat:
3481 set_current_state(TASK_IDLE);
3482
3483 /*
3484 * By the time the rescuer is requested to stop, the workqueue
3485 * shouldn't have any work pending, but @wq->maydays may still have
3486 * pwq(s) queued. This can happen by non-rescuer workers consuming
3487 * all the work items before the rescuer got to them. Go through
3488 * @wq->maydays processing before acting on should_stop so that the
3489 * list is always empty on exit.
3490 */
3491 should_stop = kthread_should_stop();
3492
3493 /* see whether any pwq is asking for help */
3494 raw_spin_lock_irq(&wq_mayday_lock);
3495
3496 while (!list_empty(head: &wq->maydays)) {
3497 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3498 struct pool_workqueue, mayday_node);
3499 struct worker_pool *pool = pwq->pool;
3500 struct work_struct *work, *n;
3501
3502 __set_current_state(TASK_RUNNING);
3503 list_del_init(entry: &pwq->mayday_node);
3504
3505 raw_spin_unlock_irq(&wq_mayday_lock);
3506
3507 worker_attach_to_pool(worker: rescuer, pool);
3508
3509 raw_spin_lock_irq(&pool->lock);
3510
3511 /*
3512 * Slurp in all works issued via this workqueue and
3513 * process'em.
3514 */
3515 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3516 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3517 if (get_work_pwq(work) == pwq &&
3518 assign_work(work, worker: rescuer, nextp: &n))
3519 pwq->stats[PWQ_STAT_RESCUED]++;
3520 }
3521
3522 if (!list_empty(head: &rescuer->scheduled)) {
3523 process_scheduled_works(worker: rescuer);
3524
3525 /*
3526 * The above execution of rescued work items could
3527 * have created more to rescue through
3528 * pwq_activate_first_inactive() or chained
3529 * queueing. Let's put @pwq back on mayday list so
3530 * that such back-to-back work items, which may be
3531 * being used to relieve memory pressure, don't
3532 * incur MAYDAY_INTERVAL delay inbetween.
3533 */
3534 if (pwq->nr_active && need_to_create_worker(pool)) {
3535 raw_spin_lock(&wq_mayday_lock);
3536 /*
3537 * Queue iff we aren't racing destruction
3538 * and somebody else hasn't queued it already.
3539 */
3540 if (wq->rescuer && list_empty(head: &pwq->mayday_node)) {
3541 get_pwq(pwq);
3542 list_add_tail(new: &pwq->mayday_node, head: &wq->maydays);
3543 }
3544 raw_spin_unlock(&wq_mayday_lock);
3545 }
3546 }
3547
3548 /*
3549 * Leave this pool. Notify regular workers; otherwise, we end up
3550 * with 0 concurrency and stalling the execution.
3551 */
3552 kick_pool(pool);
3553
3554 raw_spin_unlock_irq(&pool->lock);
3555
3556 worker_detach_from_pool(worker: rescuer);
3557
3558 /*
3559 * Put the reference grabbed by send_mayday(). @pool might
3560 * go away any time after it.
3561 */
3562 put_pwq_unlocked(pwq);
3563
3564 raw_spin_lock_irq(&wq_mayday_lock);
3565 }
3566
3567 raw_spin_unlock_irq(&wq_mayday_lock);
3568
3569 if (should_stop) {
3570 __set_current_state(TASK_RUNNING);
3571 set_pf_worker(false);
3572 return 0;
3573 }
3574
3575 /* rescuers should never participate in concurrency management */
3576 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3577 schedule();
3578 goto repeat;
3579}
3580
3581static void bh_worker(struct worker *worker)
3582{
3583 struct worker_pool *pool = worker->pool;
3584 int nr_restarts = BH_WORKER_RESTARTS;
3585 unsigned long end = jiffies + BH_WORKER_JIFFIES;
3586
3587 worker_lock_callback(pool);
3588 raw_spin_lock_irq(&pool->lock);
3589 worker_leave_idle(worker);
3590
3591 /*
3592 * This function follows the structure of worker_thread(). See there for
3593 * explanations on each step.
3594 */
3595 if (!need_more_worker(pool))
3596 goto done;
3597
3598 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3599 worker_clr_flags(worker, flags: WORKER_PREP | WORKER_REBOUND);
3600
3601 do {
3602 struct work_struct *work =
3603 list_first_entry(&pool->worklist,
3604 struct work_struct, entry);
3605
3606 if (assign_work(work, worker, NULL))
3607 process_scheduled_works(worker);
3608 } while (keep_working(pool) &&
3609 --nr_restarts && time_before(jiffies, end));
3610
3611 worker_set_flags(worker, flags: WORKER_PREP);
3612done:
3613 worker_enter_idle(worker);
3614 kick_pool(pool);
3615 raw_spin_unlock_irq(&pool->lock);
3616 worker_unlock_callback(pool);
3617}
3618
3619/*
3620 * TODO: Convert all tasklet users to workqueue and use softirq directly.
3621 *
3622 * This is currently called from tasklet[_hi]action() and thus is also called
3623 * whenever there are tasklets to run. Let's do an early exit if there's nothing
3624 * queued. Once conversion from tasklet is complete, the need_more_worker() test
3625 * can be dropped.
3626 *
3627 * After full conversion, we'll add worker->softirq_action, directly use the
3628 * softirq action and obtain the worker pointer from the softirq_action pointer.
3629 */
3630void workqueue_softirq_action(bool highpri)
3631{
3632 struct worker_pool *pool =
3633 &per_cpu(bh_worker_pools, smp_processor_id())[highpri];
3634 if (need_more_worker(pool))
3635 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3636}
3637
3638struct wq_drain_dead_softirq_work {
3639 struct work_struct work;
3640 struct worker_pool *pool;
3641 struct completion done;
3642};
3643
3644static void drain_dead_softirq_workfn(struct work_struct *work)
3645{
3646 struct wq_drain_dead_softirq_work *dead_work =
3647 container_of(work, struct wq_drain_dead_softirq_work, work);
3648 struct worker_pool *pool = dead_work->pool;
3649 bool repeat;
3650
3651 /*
3652 * @pool's CPU is dead and we want to execute its still pending work
3653 * items from this BH work item which is running on a different CPU. As
3654 * its CPU is dead, @pool can't be kicked and, as work execution path
3655 * will be nested, a lockdep annotation needs to be suppressed. Mark
3656 * @pool with %POOL_BH_DRAINING for the special treatments.
3657 */
3658 raw_spin_lock_irq(&pool->lock);
3659 pool->flags |= POOL_BH_DRAINING;
3660 raw_spin_unlock_irq(&pool->lock);
3661
3662 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3663
3664 raw_spin_lock_irq(&pool->lock);
3665 pool->flags &= ~POOL_BH_DRAINING;
3666 repeat = need_more_worker(pool);
3667 raw_spin_unlock_irq(&pool->lock);
3668
3669 /*
3670 * bh_worker() might hit consecutive execution limit and bail. If there
3671 * still are pending work items, reschedule self and return so that we
3672 * don't hog this CPU's BH.
3673 */
3674 if (repeat) {
3675 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3676 queue_work(wq: system_bh_highpri_wq, work);
3677 else
3678 queue_work(wq: system_bh_wq, work);
3679 } else {
3680 complete(&dead_work->done);
3681 }
3682}
3683
3684/*
3685 * @cpu is dead. Drain the remaining BH work items on the current CPU. It's
3686 * possible to allocate dead_work per CPU and avoid flushing. However, then we
3687 * have to worry about draining overlapping with CPU coming back online or
3688 * nesting (one CPU's dead_work queued on another CPU which is also dead and so
3689 * on). Let's keep it simple and drain them synchronously. These are BH work
3690 * items which shouldn't be requeued on the same pool. Shouldn't take long.
3691 */
3692void workqueue_softirq_dead(unsigned int cpu)
3693{
3694 int i;
3695
3696 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
3697 struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i];
3698 struct wq_drain_dead_softirq_work dead_work;
3699
3700 if (!need_more_worker(pool))
3701 continue;
3702
3703 INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn);
3704 dead_work.pool = pool;
3705 init_completion(x: &dead_work.done);
3706
3707 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3708 queue_work(wq: system_bh_highpri_wq, work: &dead_work.work);
3709 else
3710 queue_work(wq: system_bh_wq, work: &dead_work.work);
3711
3712 wait_for_completion(&dead_work.done);
3713 destroy_work_on_stack(work: &dead_work.work);
3714 }
3715}
3716
3717/**
3718 * check_flush_dependency - check for flush dependency sanity
3719 * @target_wq: workqueue being flushed
3720 * @target_work: work item being flushed (NULL for workqueue flushes)
3721 * @from_cancel: are we called from the work cancel path
3722 *
3723 * %current is trying to flush the whole @target_wq or @target_work on it.
3724 * If this is not the cancel path (which implies work being flushed is either
3725 * already running, or will not be at all), check if @target_wq doesn't have
3726 * %WQ_MEM_RECLAIM and verify that %current is not reclaiming memory or running
3727 * on a workqueue which doesn't have %WQ_MEM_RECLAIM as that can break forward-
3728 * progress guarantee leading to a deadlock.
3729 */
3730static void check_flush_dependency(struct workqueue_struct *target_wq,
3731 struct work_struct *target_work,
3732 bool from_cancel)
3733{
3734 work_func_t target_func;
3735 struct worker *worker;
3736
3737 if (from_cancel || target_wq->flags & WQ_MEM_RECLAIM)
3738 return;
3739
3740 worker = current_wq_worker();
3741 target_func = target_work ? target_work->func : NULL;
3742
3743 WARN_ONCE(current->flags & PF_MEMALLOC,
3744 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3745 current->pid, current->comm, target_wq->name, target_func);
3746 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3747 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3748 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3749 worker->current_pwq->wq->name, worker->current_func,
3750 target_wq->name, target_func);
3751}
3752
3753struct wq_barrier {
3754 struct work_struct work;
3755 struct completion done;
3756 struct task_struct *task; /* purely informational */
3757};
3758
3759static void wq_barrier_func(struct work_struct *work)
3760{
3761 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3762 complete(&barr->done);
3763}
3764
3765/**
3766 * insert_wq_barrier - insert a barrier work
3767 * @pwq: pwq to insert barrier into
3768 * @barr: wq_barrier to insert
3769 * @target: target work to attach @barr to
3770 * @worker: worker currently executing @target, NULL if @target is not executing
3771 *
3772 * @barr is linked to @target such that @barr is completed only after
3773 * @target finishes execution. Please note that the ordering
3774 * guarantee is observed only with respect to @target and on the local
3775 * cpu.
3776 *
3777 * Currently, a queued barrier can't be canceled. This is because
3778 * try_to_grab_pending() can't determine whether the work to be
3779 * grabbed is at the head of the queue and thus can't clear LINKED
3780 * flag of the previous work while there must be a valid next work
3781 * after a work with LINKED flag set.
3782 *
3783 * Note that when @worker is non-NULL, @target may be modified
3784 * underneath us, so we can't reliably determine pwq from @target.
3785 *
3786 * CONTEXT:
3787 * raw_spin_lock_irq(pool->lock).
3788 */
3789static void insert_wq_barrier(struct pool_workqueue *pwq,
3790 struct wq_barrier *barr,
3791 struct work_struct *target, struct worker *worker)
3792{
3793 static __maybe_unused struct lock_class_key bh_key, thr_key;
3794 unsigned int work_flags = 0;
3795 unsigned int work_color;
3796 struct list_head *head;
3797
3798 /*
3799 * debugobject calls are safe here even with pool->lock locked
3800 * as we know for sure that this will not trigger any of the
3801 * checks and call back into the fixup functions where we
3802 * might deadlock.
3803 *
3804 * BH and threaded workqueues need separate lockdep keys to avoid
3805 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W}
3806 * usage".
3807 */
3808 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func,
3809 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key);
3810 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3811
3812 init_completion_map(&barr->done, &target->lockdep_map);
3813
3814 barr->task = current;
3815
3816 /* The barrier work item does not participate in nr_active. */
3817 work_flags |= WORK_STRUCT_INACTIVE;
3818
3819 /*
3820 * If @target is currently being executed, schedule the
3821 * barrier to the worker; otherwise, put it after @target.
3822 */
3823 if (worker) {
3824 head = worker->scheduled.next;
3825 work_color = worker->current_color;
3826 } else {
3827 unsigned long *bits = work_data_bits(target);
3828
3829 head = target->entry.next;
3830 /* there can already be other linked works, inherit and set */
3831 work_flags |= *bits & WORK_STRUCT_LINKED;
3832 work_color = get_work_color(work_data: *bits);
3833 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3834 }
3835
3836 pwq->nr_in_flight[work_color]++;
3837 work_flags |= work_color_to_flags(color: work_color);
3838
3839 insert_work(pwq, work: &barr->work, head, extra_flags: work_flags);
3840}
3841
3842/**
3843 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3844 * @wq: workqueue being flushed
3845 * @flush_color: new flush color, < 0 for no-op
3846 * @work_color: new work color, < 0 for no-op
3847 *
3848 * Prepare pwqs for workqueue flushing.
3849 *
3850 * If @flush_color is non-negative, flush_color on all pwqs should be
3851 * -1. If no pwq has in-flight commands at the specified color, all
3852 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3853 * has in flight commands, its pwq->flush_color is set to
3854 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3855 * wakeup logic is armed and %true is returned.
3856 *
3857 * The caller should have initialized @wq->first_flusher prior to
3858 * calling this function with non-negative @flush_color. If
3859 * @flush_color is negative, no flush color update is done and %false
3860 * is returned.
3861 *
3862 * If @work_color is non-negative, all pwqs should have the same
3863 * work_color which is previous to @work_color and all will be
3864 * advanced to @work_color.
3865 *
3866 * CONTEXT:
3867 * mutex_lock(wq->mutex).
3868 *
3869 * Return:
3870 * %true if @flush_color >= 0 and there's something to flush. %false
3871 * otherwise.
3872 */
3873static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3874 int flush_color, int work_color)
3875{
3876 bool wait = false;
3877 struct pool_workqueue *pwq;
3878 struct worker_pool *current_pool = NULL;
3879
3880 if (flush_color >= 0) {
3881 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3882 atomic_set(v: &wq->nr_pwqs_to_flush, i: 1);
3883 }
3884
3885 /*
3886 * For unbound workqueue, pwqs will map to only a few pools.
3887 * Most of the time, pwqs within the same pool will be linked
3888 * sequentially to wq->pwqs by cpu index. So in the majority
3889 * of pwq iters, the pool is the same, only doing lock/unlock
3890 * if the pool has changed. This can largely reduce expensive
3891 * lock operations.
3892 */
3893 for_each_pwq(pwq, wq) {
3894 if (current_pool != pwq->pool) {
3895 if (likely(current_pool))
3896 raw_spin_unlock_irq(&current_pool->lock);
3897 current_pool = pwq->pool;
3898 raw_spin_lock_irq(&current_pool->lock);
3899 }
3900
3901 if (flush_color >= 0) {
3902 WARN_ON_ONCE(pwq->flush_color != -1);
3903
3904 if (pwq->nr_in_flight[flush_color]) {
3905 pwq->flush_color = flush_color;
3906 atomic_inc(v: &wq->nr_pwqs_to_flush);
3907 wait = true;
3908 }
3909 }
3910
3911 if (work_color >= 0) {
3912 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3913 pwq->work_color = work_color;
3914 }
3915
3916 }
3917
3918 if (current_pool)
3919 raw_spin_unlock_irq(&current_pool->lock);
3920
3921 if (flush_color >= 0 && atomic_dec_and_test(v: &wq->nr_pwqs_to_flush))
3922 complete(&wq->first_flusher->done);
3923
3924 return wait;
3925}
3926
3927static void touch_wq_lockdep_map(struct workqueue_struct *wq)
3928{
3929#ifdef CONFIG_LOCKDEP
3930 if (unlikely(!wq->lockdep_map))
3931 return;
3932
3933 if (wq->flags & WQ_BH)
3934 local_bh_disable();
3935
3936 lock_map_acquire(wq->lockdep_map);
3937 lock_map_release(wq->lockdep_map);
3938
3939 if (wq->flags & WQ_BH)
3940 local_bh_enable();
3941#endif
3942}
3943
3944static void touch_work_lockdep_map(struct work_struct *work,
3945 struct workqueue_struct *wq)
3946{
3947#ifdef CONFIG_LOCKDEP
3948 if (wq->flags & WQ_BH)
3949 local_bh_disable();
3950
3951 lock_map_acquire(&work->lockdep_map);
3952 lock_map_release(&work->lockdep_map);
3953
3954 if (wq->flags & WQ_BH)
3955 local_bh_enable();
3956#endif
3957}
3958
3959/**
3960 * __flush_workqueue - ensure that any scheduled work has run to completion.
3961 * @wq: workqueue to flush
3962 *
3963 * This function sleeps until all work items which were queued on entry
3964 * have finished execution, but it is not livelocked by new incoming ones.
3965 */
3966void __flush_workqueue(struct workqueue_struct *wq)
3967{
3968 struct wq_flusher this_flusher = {
3969 .list = LIST_HEAD_INIT(this_flusher.list),
3970 .flush_color = -1,
3971 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, (*wq->lockdep_map)),
3972 };
3973 int next_color;
3974
3975 if (WARN_ON(!wq_online))
3976 return;
3977
3978 touch_wq_lockdep_map(wq);
3979
3980 mutex_lock(lock: &wq->mutex);
3981
3982 /*
3983 * Start-to-wait phase
3984 */
3985 next_color = work_next_color(color: wq->work_color);
3986
3987 if (next_color != wq->flush_color) {
3988 /*
3989 * Color space is not full. The current work_color
3990 * becomes our flush_color and work_color is advanced
3991 * by one.
3992 */
3993 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3994 this_flusher.flush_color = wq->work_color;
3995 wq->work_color = next_color;
3996
3997 if (!wq->first_flusher) {
3998 /* no flush in progress, become the first flusher */
3999 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
4000
4001 wq->first_flusher = &this_flusher;
4002
4003 if (!flush_workqueue_prep_pwqs(wq, flush_color: wq->flush_color,
4004 work_color: wq->work_color)) {
4005 /* nothing to flush, done */
4006 wq->flush_color = next_color;
4007 wq->first_flusher = NULL;
4008 goto out_unlock;
4009 }
4010 } else {
4011 /* wait in queue */
4012 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
4013 list_add_tail(new: &this_flusher.list, head: &wq->flusher_queue);
4014 flush_workqueue_prep_pwqs(wq, flush_color: -1, work_color: wq->work_color);
4015 }
4016 } else {
4017 /*
4018 * Oops, color space is full, wait on overflow queue.
4019 * The next flush completion will assign us
4020 * flush_color and transfer to flusher_queue.
4021 */
4022 list_add_tail(new: &this_flusher.list, head: &wq->flusher_overflow);
4023 }
4024
4025 check_flush_dependency(target_wq: wq, NULL, from_cancel: false);
4026
4027 mutex_unlock(lock: &wq->mutex);
4028
4029 wait_for_completion(&this_flusher.done);
4030
4031 /*
4032 * Wake-up-and-cascade phase
4033 *
4034 * First flushers are responsible for cascading flushes and
4035 * handling overflow. Non-first flushers can simply return.
4036 */
4037 if (READ_ONCE(wq->first_flusher) != &this_flusher)
4038 return;
4039
4040 mutex_lock(lock: &wq->mutex);
4041
4042 /* we might have raced, check again with mutex held */
4043 if (wq->first_flusher != &this_flusher)
4044 goto out_unlock;
4045
4046 WRITE_ONCE(wq->first_flusher, NULL);
4047
4048 WARN_ON_ONCE(!list_empty(&this_flusher.list));
4049 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
4050
4051 while (true) {
4052 struct wq_flusher *next, *tmp;
4053
4054 /* complete all the flushers sharing the current flush color */
4055 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
4056 if (next->flush_color != wq->flush_color)
4057 break;
4058 list_del_init(entry: &next->list);
4059 complete(&next->done);
4060 }
4061
4062 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4063 wq->flush_color != work_next_color(wq->work_color));
4064
4065 /* this flush_color is finished, advance by one */
4066 wq->flush_color = work_next_color(color: wq->flush_color);
4067
4068 /* one color has been freed, handle overflow queue */
4069 if (!list_empty(head: &wq->flusher_overflow)) {
4070 /*
4071 * Assign the same color to all overflowed
4072 * flushers, advance work_color and append to
4073 * flusher_queue. This is the start-to-wait
4074 * phase for these overflowed flushers.
4075 */
4076 list_for_each_entry(tmp, &wq->flusher_overflow, list)
4077 tmp->flush_color = wq->work_color;
4078
4079 wq->work_color = work_next_color(color: wq->work_color);
4080
4081 list_splice_tail_init(list: &wq->flusher_overflow,
4082 head: &wq->flusher_queue);
4083 flush_workqueue_prep_pwqs(wq, flush_color: -1, work_color: wq->work_color);
4084 }
4085
4086 if (list_empty(head: &wq->flusher_queue)) {
4087 WARN_ON_ONCE(wq->flush_color != wq->work_color);
4088 break;
4089 }
4090
4091 /*
4092 * Need to flush more colors. Make the next flusher
4093 * the new first flusher and arm pwqs.
4094 */
4095 WARN_ON_ONCE(wq->flush_color == wq->work_color);
4096 WARN_ON_ONCE(wq->flush_color != next->flush_color);
4097
4098 list_del_init(entry: &next->list);
4099 wq->first_flusher = next;
4100
4101 if (flush_workqueue_prep_pwqs(wq, flush_color: wq->flush_color, work_color: -1))
4102 break;
4103
4104 /*
4105 * Meh... this color is already done, clear first
4106 * flusher and repeat cascading.
4107 */
4108 wq->first_flusher = NULL;
4109 }
4110
4111out_unlock:
4112 mutex_unlock(lock: &wq->mutex);
4113}
4114EXPORT_SYMBOL(__flush_workqueue);
4115
4116/**
4117 * drain_workqueue - drain a workqueue
4118 * @wq: workqueue to drain
4119 *
4120 * Wait until the workqueue becomes empty. While draining is in progress,
4121 * only chain queueing is allowed. IOW, only currently pending or running
4122 * work items on @wq can queue further work items on it. @wq is flushed
4123 * repeatedly until it becomes empty. The number of flushing is determined
4124 * by the depth of chaining and should be relatively short. Whine if it
4125 * takes too long.
4126 */
4127void drain_workqueue(struct workqueue_struct *wq)
4128{
4129 unsigned int flush_cnt = 0;
4130 struct pool_workqueue *pwq;
4131
4132 /*
4133 * __queue_work() needs to test whether there are drainers, is much
4134 * hotter than drain_workqueue() and already looks at @wq->flags.
4135 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4136 */
4137 mutex_lock(lock: &wq->mutex);
4138 if (!wq->nr_drainers++)
4139 wq->flags |= __WQ_DRAINING;
4140 mutex_unlock(lock: &wq->mutex);
4141reflush:
4142 __flush_workqueue(wq);
4143
4144 mutex_lock(lock: &wq->mutex);
4145
4146 for_each_pwq(pwq, wq) {
4147 bool drained;
4148
4149 raw_spin_lock_irq(&pwq->pool->lock);
4150 drained = pwq_is_empty(pwq);
4151 raw_spin_unlock_irq(&pwq->pool->lock);
4152
4153 if (drained)
4154 continue;
4155
4156 if (++flush_cnt == 10 ||
4157 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4158 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
4159 wq->name, __func__, flush_cnt);
4160
4161 mutex_unlock(lock: &wq->mutex);
4162 goto reflush;
4163 }
4164
4165 if (!--wq->nr_drainers)
4166 wq->flags &= ~__WQ_DRAINING;
4167 mutex_unlock(lock: &wq->mutex);
4168}
4169EXPORT_SYMBOL_GPL(drain_workqueue);
4170
4171static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
4172 bool from_cancel)
4173{
4174 struct worker *worker = NULL;
4175 struct worker_pool *pool;
4176 struct pool_workqueue *pwq;
4177 struct workqueue_struct *wq;
4178
4179 rcu_read_lock();
4180 pool = get_work_pool(work);
4181 if (!pool) {
4182 rcu_read_unlock();
4183 return false;
4184 }
4185
4186 raw_spin_lock_irq(&pool->lock);
4187 /* see the comment in try_to_grab_pending() with the same code */
4188 pwq = get_work_pwq(work);
4189 if (pwq) {
4190 if (unlikely(pwq->pool != pool))
4191 goto already_gone;
4192 } else {
4193 worker = find_worker_executing_work(pool, work);
4194 if (!worker)
4195 goto already_gone;
4196 pwq = worker->current_pwq;
4197 }
4198
4199 wq = pwq->wq;
4200 check_flush_dependency(target_wq: wq, target_work: work, from_cancel);
4201
4202 insert_wq_barrier(pwq, barr, target: work, worker);
4203 raw_spin_unlock_irq(&pool->lock);
4204
4205 touch_work_lockdep_map(work, wq);
4206
4207 /*
4208 * Force a lock recursion deadlock when using flush_work() inside a
4209 * single-threaded or rescuer equipped workqueue.
4210 *
4211 * For single threaded workqueues the deadlock happens when the work
4212 * is after the work issuing the flush_work(). For rescuer equipped
4213 * workqueues the deadlock happens when the rescuer stalls, blocking
4214 * forward progress.
4215 */
4216 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer))
4217 touch_wq_lockdep_map(wq);
4218
4219 rcu_read_unlock();
4220 return true;
4221already_gone:
4222 raw_spin_unlock_irq(&pool->lock);
4223 rcu_read_unlock();
4224 return false;
4225}
4226
4227static bool __flush_work(struct work_struct *work, bool from_cancel)
4228{
4229 struct wq_barrier barr;
4230
4231 if (WARN_ON(!wq_online))
4232 return false;
4233
4234 if (WARN_ON(!work->func))
4235 return false;
4236
4237 if (!start_flush_work(work, barr: &barr, from_cancel))
4238 return false;
4239
4240 /*
4241 * start_flush_work() returned %true. If @from_cancel is set, we know
4242 * that @work must have been executing during start_flush_work() and
4243 * can't currently be queued. Its data must contain OFFQ bits. If @work
4244 * was queued on a BH workqueue, we also know that it was running in the
4245 * BH context and thus can be busy-waited.
4246 */
4247 if (from_cancel) {
4248 unsigned long data = *work_data_bits(work);
4249
4250 if (!WARN_ON_ONCE(data & WORK_STRUCT_PWQ) &&
4251 (data & WORK_OFFQ_BH)) {
4252 /*
4253 * On RT, prevent a live lock when %current preempted
4254 * soft interrupt processing by blocking on lock which
4255 * is owned by the thread invoking the callback.
4256 */
4257 while (!try_wait_for_completion(x: &barr.done)) {
4258 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
4259 struct worker_pool *pool;
4260
4261 guard(rcu)();
4262 pool = get_work_pool(work);
4263 if (pool)
4264 workqueue_callback_cancel_wait_running(pool);
4265 } else {
4266 cpu_relax();
4267 }
4268 }
4269 goto out_destroy;
4270 }
4271 }
4272
4273 wait_for_completion(&barr.done);
4274
4275out_destroy:
4276 destroy_work_on_stack(work: &barr.work);
4277 return true;
4278}
4279
4280/**
4281 * flush_work - wait for a work to finish executing the last queueing instance
4282 * @work: the work to flush
4283 *
4284 * Wait until @work has finished execution. @work is guaranteed to be idle
4285 * on return if it hasn't been requeued since flush started.
4286 *
4287 * Return:
4288 * %true if flush_work() waited for the work to finish execution,
4289 * %false if it was already idle.
4290 */
4291bool flush_work(struct work_struct *work)
4292{
4293 might_sleep();
4294 return __flush_work(work, from_cancel: false);
4295}
4296EXPORT_SYMBOL_GPL(flush_work);
4297
4298/**
4299 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4300 * @dwork: the delayed work to flush
4301 *
4302 * Delayed timer is cancelled and the pending work is queued for
4303 * immediate execution. Like flush_work(), this function only
4304 * considers the last queueing instance of @dwork.
4305 *
4306 * Return:
4307 * %true if flush_work() waited for the work to finish execution,
4308 * %false if it was already idle.
4309 */
4310bool flush_delayed_work(struct delayed_work *dwork)
4311{
4312 local_irq_disable();
4313 if (timer_delete_sync(timer: &dwork->timer))
4314 __queue_work(cpu: dwork->cpu, wq: dwork->wq, work: &dwork->work);
4315 local_irq_enable();
4316 return flush_work(&dwork->work);
4317}
4318EXPORT_SYMBOL(flush_delayed_work);
4319
4320/**
4321 * flush_rcu_work - wait for a rwork to finish executing the last queueing
4322 * @rwork: the rcu work to flush
4323 *
4324 * Return:
4325 * %true if flush_rcu_work() waited for the work to finish execution,
4326 * %false if it was already idle.
4327 */
4328bool flush_rcu_work(struct rcu_work *rwork)
4329{
4330 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4331 rcu_barrier();
4332 flush_work(&rwork->work);
4333 return true;
4334 } else {
4335 return flush_work(&rwork->work);
4336 }
4337}
4338EXPORT_SYMBOL(flush_rcu_work);
4339
4340static void work_offqd_disable(struct work_offq_data *offqd)
4341{
4342 const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1;
4343
4344 if (likely(offqd->disable < max))
4345 offqd->disable++;
4346 else
4347 WARN_ONCE(true, "workqueue: work disable count overflowed\n");
4348}
4349
4350static void work_offqd_enable(struct work_offq_data *offqd)
4351{
4352 if (likely(offqd->disable > 0))
4353 offqd->disable--;
4354 else
4355 WARN_ONCE(true, "workqueue: work disable count underflowed\n");
4356}
4357
4358static bool __cancel_work(struct work_struct *work, u32 cflags)
4359{
4360 struct work_offq_data offqd;
4361 unsigned long irq_flags;
4362 int ret;
4363
4364 ret = work_grab_pending(work, cflags, irq_flags: &irq_flags);
4365
4366 work_offqd_unpack(offqd: &offqd, data: *work_data_bits(work));
4367
4368 if (cflags & WORK_CANCEL_DISABLE)
4369 work_offqd_disable(offqd: &offqd);
4370
4371 set_work_pool_and_clear_pending(work, pool_id: offqd.pool_id,
4372 flags: work_offqd_pack_flags(offqd: &offqd));
4373 local_irq_restore(irq_flags);
4374 return ret;
4375}
4376
4377static bool __cancel_work_sync(struct work_struct *work, u32 cflags)
4378{
4379 bool ret;
4380
4381 ret = __cancel_work(work, cflags: cflags | WORK_CANCEL_DISABLE);
4382
4383 if (*work_data_bits(work) & WORK_OFFQ_BH)
4384 WARN_ON_ONCE(in_hardirq());
4385 else
4386 might_sleep();
4387
4388 /*
4389 * Skip __flush_work() during early boot when we know that @work isn't
4390 * executing. This allows canceling during early boot.
4391 */
4392 if (wq_online)
4393 __flush_work(work, from_cancel: true);
4394
4395 if (!(cflags & WORK_CANCEL_DISABLE))
4396 enable_work(work);
4397
4398 return ret;
4399}
4400
4401/*
4402 * See cancel_delayed_work()
4403 */
4404bool cancel_work(struct work_struct *work)
4405{
4406 return __cancel_work(work, cflags: 0);
4407}
4408EXPORT_SYMBOL(cancel_work);
4409
4410/**
4411 * cancel_work_sync - cancel a work and wait for it to finish
4412 * @work: the work to cancel
4413 *
4414 * Cancel @work and wait for its execution to finish. This function can be used
4415 * even if the work re-queues itself or migrates to another workqueue. On return
4416 * from this function, @work is guaranteed to be not pending or executing on any
4417 * CPU as long as there aren't racing enqueues.
4418 *
4419 * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's.
4420 * Use cancel_delayed_work_sync() instead.
4421 *
4422 * Must be called from a sleepable context if @work was last queued on a non-BH
4423 * workqueue. Can also be called from non-hardirq atomic contexts including BH
4424 * if @work was last queued on a BH workqueue.
4425 *
4426 * Returns %true if @work was pending, %false otherwise.
4427 */
4428bool cancel_work_sync(struct work_struct *work)
4429{
4430 return __cancel_work_sync(work, cflags: 0);
4431}
4432EXPORT_SYMBOL_GPL(cancel_work_sync);
4433
4434/**
4435 * cancel_delayed_work - cancel a delayed work
4436 * @dwork: delayed_work to cancel
4437 *
4438 * Kill off a pending delayed_work.
4439 *
4440 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4441 * pending.
4442 *
4443 * Note:
4444 * The work callback function may still be running on return, unless
4445 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4446 * use cancel_delayed_work_sync() to wait on it.
4447 *
4448 * This function is safe to call from any context including IRQ handler.
4449 */
4450bool cancel_delayed_work(struct delayed_work *dwork)
4451{
4452 return __cancel_work(work: &dwork->work, cflags: WORK_CANCEL_DELAYED);
4453}
4454EXPORT_SYMBOL(cancel_delayed_work);
4455
4456/**
4457 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4458 * @dwork: the delayed work cancel
4459 *
4460 * This is cancel_work_sync() for delayed works.
4461 *
4462 * Return:
4463 * %true if @dwork was pending, %false otherwise.
4464 */
4465bool cancel_delayed_work_sync(struct delayed_work *dwork)
4466{
4467 return __cancel_work_sync(work: &dwork->work, cflags: WORK_CANCEL_DELAYED);
4468}
4469EXPORT_SYMBOL(cancel_delayed_work_sync);
4470
4471/**
4472 * disable_work - Disable and cancel a work item
4473 * @work: work item to disable
4474 *
4475 * Disable @work by incrementing its disable count and cancel it if currently
4476 * pending. As long as the disable count is non-zero, any attempt to queue @work
4477 * will fail and return %false. The maximum supported disable depth is 2 to the
4478 * power of %WORK_OFFQ_DISABLE_BITS, currently 65536.
4479 *
4480 * Can be called from any context. Returns %true if @work was pending, %false
4481 * otherwise.
4482 */
4483bool disable_work(struct work_struct *work)
4484{
4485 return __cancel_work(work, cflags: WORK_CANCEL_DISABLE);
4486}
4487EXPORT_SYMBOL_GPL(disable_work);
4488
4489/**
4490 * disable_work_sync - Disable, cancel and drain a work item
4491 * @work: work item to disable
4492 *
4493 * Similar to disable_work() but also wait for @work to finish if currently
4494 * executing.
4495 *
4496 * Must be called from a sleepable context if @work was last queued on a non-BH
4497 * workqueue. Can also be called from non-hardirq atomic contexts including BH
4498 * if @work was last queued on a BH workqueue.
4499 *
4500 * Returns %true if @work was pending, %false otherwise.
4501 */
4502bool disable_work_sync(struct work_struct *work)
4503{
4504 return __cancel_work_sync(work, cflags: WORK_CANCEL_DISABLE);
4505}
4506EXPORT_SYMBOL_GPL(disable_work_sync);
4507
4508/**
4509 * enable_work - Enable a work item
4510 * @work: work item to enable
4511 *
4512 * Undo disable_work[_sync]() by decrementing @work's disable count. @work can
4513 * only be queued if its disable count is 0.
4514 *
4515 * Can be called from any context. Returns %true if the disable count reached 0.
4516 * Otherwise, %false.
4517 */
4518bool enable_work(struct work_struct *work)
4519{
4520 struct work_offq_data offqd;
4521 unsigned long irq_flags;
4522
4523 work_grab_pending(work, cflags: 0, irq_flags: &irq_flags);
4524
4525 work_offqd_unpack(offqd: &offqd, data: *work_data_bits(work));
4526 work_offqd_enable(offqd: &offqd);
4527 set_work_pool_and_clear_pending(work, pool_id: offqd.pool_id,
4528 flags: work_offqd_pack_flags(offqd: &offqd));
4529 local_irq_restore(irq_flags);
4530
4531 return !offqd.disable;
4532}
4533EXPORT_SYMBOL_GPL(enable_work);
4534
4535/**
4536 * disable_delayed_work - Disable and cancel a delayed work item
4537 * @dwork: delayed work item to disable
4538 *
4539 * disable_work() for delayed work items.
4540 */
4541bool disable_delayed_work(struct delayed_work *dwork)
4542{
4543 return __cancel_work(work: &dwork->work,
4544 cflags: WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4545}
4546EXPORT_SYMBOL_GPL(disable_delayed_work);
4547
4548/**
4549 * disable_delayed_work_sync - Disable, cancel and drain a delayed work item
4550 * @dwork: delayed work item to disable
4551 *
4552 * disable_work_sync() for delayed work items.
4553 */
4554bool disable_delayed_work_sync(struct delayed_work *dwork)
4555{
4556 return __cancel_work_sync(work: &dwork->work,
4557 cflags: WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4558}
4559EXPORT_SYMBOL_GPL(disable_delayed_work_sync);
4560
4561/**
4562 * enable_delayed_work - Enable a delayed work item
4563 * @dwork: delayed work item to enable
4564 *
4565 * enable_work() for delayed work items.
4566 */
4567bool enable_delayed_work(struct delayed_work *dwork)
4568{
4569 return enable_work(&dwork->work);
4570}
4571EXPORT_SYMBOL_GPL(enable_delayed_work);
4572
4573/**
4574 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4575 * @func: the function to call
4576 *
4577 * schedule_on_each_cpu() executes @func on each online CPU using the
4578 * system workqueue and blocks until all CPUs have completed.
4579 * schedule_on_each_cpu() is very slow.
4580 *
4581 * Return:
4582 * 0 on success, -errno on failure.
4583 */
4584int schedule_on_each_cpu(work_func_t func)
4585{
4586 int cpu;
4587 struct work_struct __percpu *works;
4588
4589 works = alloc_percpu(struct work_struct);
4590 if (!works)
4591 return -ENOMEM;
4592
4593 cpus_read_lock();
4594
4595 for_each_online_cpu(cpu) {
4596 struct work_struct *work = per_cpu_ptr(works, cpu);
4597
4598 INIT_WORK(work, func);
4599 schedule_work_on(cpu, work);
4600 }
4601
4602 for_each_online_cpu(cpu)
4603 flush_work(per_cpu_ptr(works, cpu));
4604
4605 cpus_read_unlock();
4606 free_percpu(pdata: works);
4607 return 0;
4608}
4609
4610/**
4611 * execute_in_process_context - reliably execute the routine with user context
4612 * @fn: the function to execute
4613 * @ew: guaranteed storage for the execute work structure (must
4614 * be available when the work executes)
4615 *
4616 * Executes the function immediately if process context is available,
4617 * otherwise schedules the function for delayed execution.
4618 *
4619 * Return: 0 - function was executed
4620 * 1 - function was scheduled for execution
4621 */
4622int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4623{
4624 if (!in_interrupt()) {
4625 fn(&ew->work);
4626 return 0;
4627 }
4628
4629 INIT_WORK(&ew->work, fn);
4630 schedule_work(work: &ew->work);
4631
4632 return 1;
4633}
4634EXPORT_SYMBOL_GPL(execute_in_process_context);
4635
4636/**
4637 * free_workqueue_attrs - free a workqueue_attrs
4638 * @attrs: workqueue_attrs to free
4639 *
4640 * Undo alloc_workqueue_attrs().
4641 */
4642void free_workqueue_attrs(struct workqueue_attrs *attrs)
4643{
4644 if (attrs) {
4645 free_cpumask_var(mask: attrs->cpumask);
4646 free_cpumask_var(mask: attrs->__pod_cpumask);
4647 kfree(objp: attrs);
4648 }
4649}
4650
4651/**
4652 * alloc_workqueue_attrs - allocate a workqueue_attrs
4653 *
4654 * Allocate a new workqueue_attrs, initialize with default settings and
4655 * return it.
4656 *
4657 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4658 */
4659struct workqueue_attrs *alloc_workqueue_attrs_noprof(void)
4660{
4661 struct workqueue_attrs *attrs;
4662
4663 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4664 if (!attrs)
4665 goto fail;
4666 if (!alloc_cpumask_var(mask: &attrs->cpumask, GFP_KERNEL))
4667 goto fail;
4668 if (!alloc_cpumask_var(mask: &attrs->__pod_cpumask, GFP_KERNEL))
4669 goto fail;
4670
4671 cpumask_copy(dstp: attrs->cpumask, cpu_possible_mask);
4672 attrs->affn_scope = WQ_AFFN_DFL;
4673 return attrs;
4674fail:
4675 free_workqueue_attrs(attrs);
4676 return NULL;
4677}
4678
4679static void copy_workqueue_attrs(struct workqueue_attrs *to,
4680 const struct workqueue_attrs *from)
4681{
4682 to->nice = from->nice;
4683 cpumask_copy(dstp: to->cpumask, srcp: from->cpumask);
4684 cpumask_copy(dstp: to->__pod_cpumask, srcp: from->__pod_cpumask);
4685 to->affn_strict = from->affn_strict;
4686
4687 /*
4688 * Unlike hash and equality test, copying shouldn't ignore wq-only
4689 * fields as copying is used for both pool and wq attrs. Instead,
4690 * get_unbound_pool() explicitly clears the fields.
4691 */
4692 to->affn_scope = from->affn_scope;
4693 to->ordered = from->ordered;
4694}
4695
4696/*
4697 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4698 * comments in 'struct workqueue_attrs' definition.
4699 */
4700static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4701{
4702 attrs->affn_scope = WQ_AFFN_NR_TYPES;
4703 attrs->ordered = false;
4704 if (attrs->affn_strict)
4705 cpumask_copy(dstp: attrs->cpumask, cpu_possible_mask);
4706}
4707
4708/* hash value of the content of @attr */
4709static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4710{
4711 u32 hash = 0;
4712
4713 hash = jhash_1word(a: attrs->nice, initval: hash);
4714 hash = jhash_1word(a: attrs->affn_strict, initval: hash);
4715 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4716 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), initval: hash);
4717 if (!attrs->affn_strict)
4718 hash = jhash(cpumask_bits(attrs->cpumask),
4719 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), initval: hash);
4720 return hash;
4721}
4722
4723/* content equality test */
4724static bool wqattrs_equal(const struct workqueue_attrs *a,
4725 const struct workqueue_attrs *b)
4726{
4727 if (a->nice != b->nice)
4728 return false;
4729 if (a->affn_strict != b->affn_strict)
4730 return false;
4731 if (!cpumask_equal(src1p: a->__pod_cpumask, src2p: b->__pod_cpumask))
4732 return false;
4733 if (!a->affn_strict && !cpumask_equal(src1p: a->cpumask, src2p: b->cpumask))
4734 return false;
4735 return true;
4736}
4737
4738/* Update @attrs with actually available CPUs */
4739static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4740 const cpumask_t *unbound_cpumask)
4741{
4742 /*
4743 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4744 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4745 * @unbound_cpumask.
4746 */
4747 cpumask_and(dstp: attrs->cpumask, src1p: attrs->cpumask, src2p: unbound_cpumask);
4748 if (unlikely(cpumask_empty(attrs->cpumask)))
4749 cpumask_copy(dstp: attrs->cpumask, srcp: unbound_cpumask);
4750}
4751
4752/* find wq_pod_type to use for @attrs */
4753static const struct wq_pod_type *
4754wqattrs_pod_type(const struct workqueue_attrs *attrs)
4755{
4756 enum wq_affn_scope scope;
4757 struct wq_pod_type *pt;
4758
4759 /* to synchronize access to wq_affn_dfl */
4760 lockdep_assert_held(&wq_pool_mutex);
4761
4762 if (attrs->affn_scope == WQ_AFFN_DFL)
4763 scope = wq_affn_dfl;
4764 else
4765 scope = attrs->affn_scope;
4766
4767 pt = &wq_pod_types[scope];
4768
4769 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4770 likely(pt->nr_pods))
4771 return pt;
4772
4773 /*
4774 * Before workqueue_init_topology(), only SYSTEM is available which is
4775 * initialized in workqueue_init_early().
4776 */
4777 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4778 BUG_ON(!pt->nr_pods);
4779 return pt;
4780}
4781
4782/**
4783 * init_worker_pool - initialize a newly zalloc'd worker_pool
4784 * @pool: worker_pool to initialize
4785 *
4786 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4787 *
4788 * Return: 0 on success, -errno on failure. Even on failure, all fields
4789 * inside @pool proper are initialized and put_unbound_pool() can be called
4790 * on @pool safely to release it.
4791 */
4792static int init_worker_pool(struct worker_pool *pool)
4793{
4794 raw_spin_lock_init(&pool->lock);
4795 pool->id = -1;
4796 pool->cpu = -1;
4797 pool->node = NUMA_NO_NODE;
4798 pool->flags |= POOL_DISASSOCIATED;
4799 pool->watchdog_ts = jiffies;
4800 INIT_LIST_HEAD(list: &pool->worklist);
4801 INIT_LIST_HEAD(list: &pool->idle_list);
4802 hash_init(pool->busy_hash);
4803
4804 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4805 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4806
4807 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4808
4809 INIT_LIST_HEAD(list: &pool->workers);
4810
4811 ida_init(ida: &pool->worker_ida);
4812 INIT_HLIST_NODE(h: &pool->hash_node);
4813 pool->refcnt = 1;
4814#ifdef CONFIG_PREEMPT_RT
4815 spin_lock_init(&pool->cb_lock);
4816#endif
4817
4818 /* shouldn't fail above this point */
4819 pool->attrs = alloc_workqueue_attrs();
4820 if (!pool->attrs)
4821 return -ENOMEM;
4822
4823 wqattrs_clear_for_pool(attrs: pool->attrs);
4824
4825 return 0;
4826}
4827
4828#ifdef CONFIG_LOCKDEP
4829static void wq_init_lockdep(struct workqueue_struct *wq)
4830{
4831 char *lock_name;
4832
4833 lockdep_register_key(&wq->key);
4834 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4835 if (!lock_name)
4836 lock_name = wq->name;
4837
4838 wq->lock_name = lock_name;
4839 wq->lockdep_map = &wq->__lockdep_map;
4840 lockdep_init_map(wq->lockdep_map, lock_name, &wq->key, 0);
4841}
4842
4843static void wq_unregister_lockdep(struct workqueue_struct *wq)
4844{
4845 if (wq->lockdep_map != &wq->__lockdep_map)
4846 return;
4847
4848 lockdep_unregister_key(&wq->key);
4849}
4850
4851static void wq_free_lockdep(struct workqueue_struct *wq)
4852{
4853 if (wq->lockdep_map != &wq->__lockdep_map)
4854 return;
4855
4856 if (wq->lock_name != wq->name)
4857 kfree(wq->lock_name);
4858}
4859#else
4860static void wq_init_lockdep(struct workqueue_struct *wq)
4861{
4862}
4863
4864static void wq_unregister_lockdep(struct workqueue_struct *wq)
4865{
4866}
4867
4868static void wq_free_lockdep(struct workqueue_struct *wq)
4869{
4870}
4871#endif
4872
4873static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
4874{
4875 int node;
4876
4877 for_each_node(node) {
4878 kfree(objp: nna_ar[node]);
4879 nna_ar[node] = NULL;
4880 }
4881
4882 kfree(objp: nna_ar[nr_node_ids]);
4883 nna_ar[nr_node_ids] = NULL;
4884}
4885
4886static void init_node_nr_active(struct wq_node_nr_active *nna)
4887{
4888 nna->max = WQ_DFL_MIN_ACTIVE;
4889 atomic_set(v: &nna->nr, i: 0);
4890 raw_spin_lock_init(&nna->lock);
4891 INIT_LIST_HEAD(list: &nna->pending_pwqs);
4892}
4893
4894/*
4895 * Each node's nr_active counter will be accessed mostly from its own node and
4896 * should be allocated in the node.
4897 */
4898static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
4899{
4900 struct wq_node_nr_active *nna;
4901 int node;
4902
4903 for_each_node(node) {
4904 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
4905 if (!nna)
4906 goto err_free;
4907 init_node_nr_active(nna);
4908 nna_ar[node] = nna;
4909 }
4910
4911 /* [nr_node_ids] is used as the fallback */
4912 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
4913 if (!nna)
4914 goto err_free;
4915 init_node_nr_active(nna);
4916 nna_ar[nr_node_ids] = nna;
4917
4918 return 0;
4919
4920err_free:
4921 free_node_nr_active(nna_ar);
4922 return -ENOMEM;
4923}
4924
4925static void rcu_free_wq(struct rcu_head *rcu)
4926{
4927 struct workqueue_struct *wq =
4928 container_of(rcu, struct workqueue_struct, rcu);
4929
4930 if (wq->flags & WQ_UNBOUND)
4931 free_node_nr_active(nna_ar: wq->node_nr_active);
4932
4933 wq_free_lockdep(wq);
4934 free_percpu(pdata: wq->cpu_pwq);
4935 free_workqueue_attrs(attrs: wq->unbound_attrs);
4936 kfree(objp: wq);
4937}
4938
4939static void rcu_free_pool(struct rcu_head *rcu)
4940{
4941 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4942
4943 ida_destroy(ida: &pool->worker_ida);
4944 free_workqueue_attrs(attrs: pool->attrs);
4945 kfree(objp: pool);
4946}
4947
4948/**
4949 * put_unbound_pool - put a worker_pool
4950 * @pool: worker_pool to put
4951 *
4952 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4953 * safe manner. get_unbound_pool() calls this function on its failure path
4954 * and this function should be able to release pools which went through,
4955 * successfully or not, init_worker_pool().
4956 *
4957 * Should be called with wq_pool_mutex held.
4958 */
4959static void put_unbound_pool(struct worker_pool *pool)
4960{
4961 struct worker *worker;
4962 LIST_HEAD(cull_list);
4963
4964 lockdep_assert_held(&wq_pool_mutex);
4965
4966 if (--pool->refcnt)
4967 return;
4968
4969 /* sanity checks */
4970 if (WARN_ON(!(pool->cpu < 0)) ||
4971 WARN_ON(!list_empty(&pool->worklist)))
4972 return;
4973
4974 /* release id and unhash */
4975 if (pool->id >= 0)
4976 idr_remove(&worker_pool_idr, id: pool->id);
4977 hash_del(node: &pool->hash_node);
4978
4979 /*
4980 * Become the manager and destroy all workers. This prevents
4981 * @pool's workers from blocking on attach_mutex. We're the last
4982 * manager and @pool gets freed with the flag set.
4983 *
4984 * Having a concurrent manager is quite unlikely to happen as we can
4985 * only get here with
4986 * pwq->refcnt == pool->refcnt == 0
4987 * which implies no work queued to the pool, which implies no worker can
4988 * become the manager. However a worker could have taken the role of
4989 * manager before the refcnts dropped to 0, since maybe_create_worker()
4990 * drops pool->lock
4991 */
4992 while (true) {
4993 rcuwait_wait_event(&manager_wait,
4994 !(pool->flags & POOL_MANAGER_ACTIVE),
4995 TASK_UNINTERRUPTIBLE);
4996
4997 mutex_lock(lock: &wq_pool_attach_mutex);
4998 raw_spin_lock_irq(&pool->lock);
4999 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
5000 pool->flags |= POOL_MANAGER_ACTIVE;
5001 break;
5002 }
5003 raw_spin_unlock_irq(&pool->lock);
5004 mutex_unlock(lock: &wq_pool_attach_mutex);
5005 }
5006
5007 while ((worker = first_idle_worker(pool)))
5008 set_worker_dying(worker, list: &cull_list);
5009 WARN_ON(pool->nr_workers || pool->nr_idle);
5010 raw_spin_unlock_irq(&pool->lock);
5011
5012 detach_dying_workers(cull_list: &cull_list);
5013
5014 mutex_unlock(lock: &wq_pool_attach_mutex);
5015
5016 reap_dying_workers(cull_list: &cull_list);
5017
5018 /* shut down the timers */
5019 timer_delete_sync(timer: &pool->idle_timer);
5020 cancel_work_sync(&pool->idle_cull_work);
5021 timer_delete_sync(timer: &pool->mayday_timer);
5022
5023 /* RCU protected to allow dereferences from get_work_pool() */
5024 call_rcu(head: &pool->rcu, func: rcu_free_pool);
5025}
5026
5027/**
5028 * get_unbound_pool - get a worker_pool with the specified attributes
5029 * @attrs: the attributes of the worker_pool to get
5030 *
5031 * Obtain a worker_pool which has the same attributes as @attrs, bump the
5032 * reference count and return it. If there already is a matching
5033 * worker_pool, it will be used; otherwise, this function attempts to
5034 * create a new one.
5035 *
5036 * Should be called with wq_pool_mutex held.
5037 *
5038 * Return: On success, a worker_pool with the same attributes as @attrs.
5039 * On failure, %NULL.
5040 */
5041static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
5042{
5043 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
5044 u32 hash = wqattrs_hash(attrs);
5045 struct worker_pool *pool;
5046 int pod, node = NUMA_NO_NODE;
5047
5048 lockdep_assert_held(&wq_pool_mutex);
5049
5050 /* do we already have a matching pool? */
5051 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
5052 if (wqattrs_equal(a: pool->attrs, b: attrs)) {
5053 pool->refcnt++;
5054 return pool;
5055 }
5056 }
5057
5058 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
5059 for (pod = 0; pod < pt->nr_pods; pod++) {
5060 if (cpumask_subset(src1p: attrs->__pod_cpumask, src2p: pt->pod_cpus[pod])) {
5061 node = pt->pod_node[pod];
5062 break;
5063 }
5064 }
5065
5066 /* nope, create a new one */
5067 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
5068 if (!pool || init_worker_pool(pool) < 0)
5069 goto fail;
5070
5071 pool->node = node;
5072 copy_workqueue_attrs(to: pool->attrs, from: attrs);
5073 wqattrs_clear_for_pool(attrs: pool->attrs);
5074
5075 if (worker_pool_assign_id(pool) < 0)
5076 goto fail;
5077
5078 /* create and start the initial worker */
5079 if (wq_online && !create_worker(pool))
5080 goto fail;
5081
5082 /* install */
5083 hash_add(unbound_pool_hash, &pool->hash_node, hash);
5084
5085 return pool;
5086fail:
5087 if (pool)
5088 put_unbound_pool(pool);
5089 return NULL;
5090}
5091
5092/*
5093 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
5094 * refcnt and needs to be destroyed.
5095 */
5096static void pwq_release_workfn(struct kthread_work *work)
5097{
5098 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
5099 release_work);
5100 struct workqueue_struct *wq = pwq->wq;
5101 struct worker_pool *pool = pwq->pool;
5102 bool is_last = false;
5103
5104 /*
5105 * When @pwq is not linked, it doesn't hold any reference to the
5106 * @wq, and @wq is invalid to access.
5107 */
5108 if (!list_empty(head: &pwq->pwqs_node)) {
5109 mutex_lock(lock: &wq->mutex);
5110 list_del_rcu(entry: &pwq->pwqs_node);
5111 is_last = list_empty(head: &wq->pwqs);
5112
5113 /*
5114 * For ordered workqueue with a plugged dfl_pwq, restart it now.
5115 */
5116 if (!is_last && (wq->flags & __WQ_ORDERED))
5117 unplug_oldest_pwq(wq);
5118
5119 mutex_unlock(lock: &wq->mutex);
5120 }
5121
5122 if (wq->flags & WQ_UNBOUND) {
5123 mutex_lock(lock: &wq_pool_mutex);
5124 put_unbound_pool(pool);
5125 mutex_unlock(lock: &wq_pool_mutex);
5126 }
5127
5128 if (!list_empty(head: &pwq->pending_node)) {
5129 struct wq_node_nr_active *nna =
5130 wq_node_nr_active(wq: pwq->wq, node: pwq->pool->node);
5131
5132 raw_spin_lock_irq(&nna->lock);
5133 list_del_init(entry: &pwq->pending_node);
5134 raw_spin_unlock_irq(&nna->lock);
5135 }
5136
5137 kfree_rcu(pwq, rcu);
5138
5139 /*
5140 * If we're the last pwq going away, @wq is already dead and no one
5141 * is gonna access it anymore. Schedule RCU free.
5142 */
5143 if (is_last) {
5144 wq_unregister_lockdep(wq);
5145 call_rcu(head: &wq->rcu, func: rcu_free_wq);
5146 }
5147}
5148
5149/* initialize newly allocated @pwq which is associated with @wq and @pool */
5150static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
5151 struct worker_pool *pool)
5152{
5153 BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK);
5154
5155 memset(s: pwq, c: 0, n: sizeof(*pwq));
5156
5157 pwq->pool = pool;
5158 pwq->wq = wq;
5159 pwq->flush_color = -1;
5160 pwq->refcnt = 1;
5161 INIT_LIST_HEAD(list: &pwq->inactive_works);
5162 INIT_LIST_HEAD(list: &pwq->pending_node);
5163 INIT_LIST_HEAD(list: &pwq->pwqs_node);
5164 INIT_LIST_HEAD(list: &pwq->mayday_node);
5165 kthread_init_work(&pwq->release_work, pwq_release_workfn);
5166}
5167
5168/* sync @pwq with the current state of its associated wq and link it */
5169static void link_pwq(struct pool_workqueue *pwq)
5170{
5171 struct workqueue_struct *wq = pwq->wq;
5172
5173 lockdep_assert_held(&wq->mutex);
5174
5175 /* may be called multiple times, ignore if already linked */
5176 if (!list_empty(head: &pwq->pwqs_node))
5177 return;
5178
5179 /* set the matching work_color */
5180 pwq->work_color = wq->work_color;
5181
5182 /* link in @pwq */
5183 list_add_tail_rcu(new: &pwq->pwqs_node, head: &wq->pwqs);
5184}
5185
5186/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5187static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
5188 const struct workqueue_attrs *attrs)
5189{
5190 struct worker_pool *pool;
5191 struct pool_workqueue *pwq;
5192
5193 lockdep_assert_held(&wq_pool_mutex);
5194
5195 pool = get_unbound_pool(attrs);
5196 if (!pool)
5197 return NULL;
5198
5199 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5200 if (!pwq) {
5201 put_unbound_pool(pool);
5202 return NULL;
5203 }
5204
5205 init_pwq(pwq, wq, pool);
5206 return pwq;
5207}
5208
5209static void apply_wqattrs_lock(void)
5210{
5211 mutex_lock(lock: &wq_pool_mutex);
5212}
5213
5214static void apply_wqattrs_unlock(void)
5215{
5216 mutex_unlock(lock: &wq_pool_mutex);
5217}
5218
5219/**
5220 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
5221 * @attrs: the wq_attrs of the default pwq of the target workqueue
5222 * @cpu: the target CPU
5223 *
5224 * Calculate the cpumask a workqueue with @attrs should use on @pod.
5225 * The result is stored in @attrs->__pod_cpumask.
5226 *
5227 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
5228 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
5229 * intersection of the possible CPUs of @pod and @attrs->cpumask.
5230 *
5231 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
5232 */
5233static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu)
5234{
5235 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5236 int pod = pt->cpu_pod[cpu];
5237
5238 /* calculate possible CPUs in @pod that @attrs wants */
5239 cpumask_and(dstp: attrs->__pod_cpumask, src1p: pt->pod_cpus[pod], src2p: attrs->cpumask);
5240 /* does @pod have any online CPUs @attrs wants? */
5241 if (!cpumask_intersects(src1p: attrs->__pod_cpumask, src2p: wq_online_cpumask)) {
5242 cpumask_copy(dstp: attrs->__pod_cpumask, srcp: attrs->cpumask);
5243 return;
5244 }
5245}
5246
5247/* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
5248static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
5249 int cpu, struct pool_workqueue *pwq)
5250{
5251 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
5252 struct pool_workqueue *old_pwq;
5253
5254 lockdep_assert_held(&wq_pool_mutex);
5255 lockdep_assert_held(&wq->mutex);
5256
5257 /* link_pwq() can handle duplicate calls */
5258 link_pwq(pwq);
5259
5260 old_pwq = rcu_access_pointer(*slot);
5261 rcu_assign_pointer(*slot, pwq);
5262 return old_pwq;
5263}
5264
5265/* context to store the prepared attrs & pwqs before applying */
5266struct apply_wqattrs_ctx {
5267 struct workqueue_struct *wq; /* target workqueue */
5268 struct workqueue_attrs *attrs; /* attrs to apply */
5269 struct list_head list; /* queued for batching commit */
5270 struct pool_workqueue *dfl_pwq;
5271 struct pool_workqueue *pwq_tbl[];
5272};
5273
5274/* free the resources after success or abort */
5275static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5276{
5277 if (ctx) {
5278 int cpu;
5279
5280 for_each_possible_cpu(cpu)
5281 put_pwq_unlocked(pwq: ctx->pwq_tbl[cpu]);
5282 put_pwq_unlocked(pwq: ctx->dfl_pwq);
5283
5284 free_workqueue_attrs(attrs: ctx->attrs);
5285
5286 kfree(objp: ctx);
5287 }
5288}
5289
5290/* allocate the attrs and pwqs for later installation */
5291static struct apply_wqattrs_ctx *
5292apply_wqattrs_prepare(struct workqueue_struct *wq,
5293 const struct workqueue_attrs *attrs,
5294 const cpumask_var_t unbound_cpumask)
5295{
5296 struct apply_wqattrs_ctx *ctx;
5297 struct workqueue_attrs *new_attrs;
5298 int cpu;
5299
5300 lockdep_assert_held(&wq_pool_mutex);
5301
5302 if (WARN_ON(attrs->affn_scope < 0 ||
5303 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
5304 return ERR_PTR(error: -EINVAL);
5305
5306 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
5307
5308 new_attrs = alloc_workqueue_attrs();
5309 if (!ctx || !new_attrs)
5310 goto out_free;
5311
5312 /*
5313 * If something goes wrong during CPU up/down, we'll fall back to
5314 * the default pwq covering whole @attrs->cpumask. Always create
5315 * it even if we don't use it immediately.
5316 */
5317 copy_workqueue_attrs(to: new_attrs, from: attrs);
5318 wqattrs_actualize_cpumask(attrs: new_attrs, unbound_cpumask);
5319 cpumask_copy(dstp: new_attrs->__pod_cpumask, srcp: new_attrs->cpumask);
5320 ctx->dfl_pwq = alloc_unbound_pwq(wq, attrs: new_attrs);
5321 if (!ctx->dfl_pwq)
5322 goto out_free;
5323
5324 for_each_possible_cpu(cpu) {
5325 if (new_attrs->ordered) {
5326 ctx->dfl_pwq->refcnt++;
5327 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
5328 } else {
5329 wq_calc_pod_cpumask(attrs: new_attrs, cpu);
5330 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, attrs: new_attrs);
5331 if (!ctx->pwq_tbl[cpu])
5332 goto out_free;
5333 }
5334 }
5335
5336 /* save the user configured attrs and sanitize it. */
5337 copy_workqueue_attrs(to: new_attrs, from: attrs);
5338 cpumask_and(dstp: new_attrs->cpumask, src1p: new_attrs->cpumask, cpu_possible_mask);
5339 cpumask_copy(dstp: new_attrs->__pod_cpumask, srcp: new_attrs->cpumask);
5340 ctx->attrs = new_attrs;
5341
5342 /*
5343 * For initialized ordered workqueues, there should only be one pwq
5344 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution
5345 * of newly queued work items until execution of older work items in
5346 * the old pwq's have completed.
5347 */
5348 if ((wq->flags & __WQ_ORDERED) && !list_empty(head: &wq->pwqs))
5349 ctx->dfl_pwq->plugged = true;
5350
5351 ctx->wq = wq;
5352 return ctx;
5353
5354out_free:
5355 free_workqueue_attrs(attrs: new_attrs);
5356 apply_wqattrs_cleanup(ctx);
5357 return ERR_PTR(error: -ENOMEM);
5358}
5359
5360/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
5361static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5362{
5363 int cpu;
5364
5365 /* all pwqs have been created successfully, let's install'em */
5366 mutex_lock(lock: &ctx->wq->mutex);
5367
5368 copy_workqueue_attrs(to: ctx->wq->unbound_attrs, from: ctx->attrs);
5369
5370 /* save the previous pwqs and install the new ones */
5371 for_each_possible_cpu(cpu)
5372 ctx->pwq_tbl[cpu] = install_unbound_pwq(wq: ctx->wq, cpu,
5373 pwq: ctx->pwq_tbl[cpu]);
5374 ctx->dfl_pwq = install_unbound_pwq(wq: ctx->wq, cpu: -1, pwq: ctx->dfl_pwq);
5375
5376 /* update node_nr_active->max */
5377 wq_update_node_max_active(wq: ctx->wq, off_cpu: -1);
5378
5379 /* rescuer needs to respect wq cpumask changes */
5380 if (ctx->wq->rescuer)
5381 set_cpus_allowed_ptr(p: ctx->wq->rescuer->task,
5382 new_mask: unbound_effective_cpumask(wq: ctx->wq));
5383
5384 mutex_unlock(lock: &ctx->wq->mutex);
5385}
5386
5387static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5388 const struct workqueue_attrs *attrs)
5389{
5390 struct apply_wqattrs_ctx *ctx;
5391
5392 /* only unbound workqueues can change attributes */
5393 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5394 return -EINVAL;
5395
5396 ctx = apply_wqattrs_prepare(wq, attrs, unbound_cpumask: wq_unbound_cpumask);
5397 if (IS_ERR(ptr: ctx))
5398 return PTR_ERR(ptr: ctx);
5399
5400 /* the ctx has been prepared successfully, let's commit it */
5401 apply_wqattrs_commit(ctx);
5402 apply_wqattrs_cleanup(ctx);
5403
5404 return 0;
5405}
5406
5407/**
5408 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5409 * @wq: the target workqueue
5410 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5411 *
5412 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
5413 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
5414 * work items are affine to the pod it was issued on. Older pwqs are released as
5415 * in-flight work items finish. Note that a work item which repeatedly requeues
5416 * itself back-to-back will stay on its current pwq.
5417 *
5418 * Performs GFP_KERNEL allocations.
5419 *
5420 * Return: 0 on success and -errno on failure.
5421 */
5422int apply_workqueue_attrs(struct workqueue_struct *wq,
5423 const struct workqueue_attrs *attrs)
5424{
5425 int ret;
5426
5427 mutex_lock(lock: &wq_pool_mutex);
5428 ret = apply_workqueue_attrs_locked(wq, attrs);
5429 mutex_unlock(lock: &wq_pool_mutex);
5430
5431 return ret;
5432}
5433
5434/**
5435 * unbound_wq_update_pwq - update a pwq slot for CPU hot[un]plug
5436 * @wq: the target workqueue
5437 * @cpu: the CPU to update the pwq slot for
5438 *
5439 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5440 * %CPU_DOWN_FAILED. @cpu is in the same pod of the CPU being hot[un]plugged.
5441 *
5442 *
5443 * If pod affinity can't be adjusted due to memory allocation failure, it falls
5444 * back to @wq->dfl_pwq which may not be optimal but is always correct.
5445 *
5446 * Note that when the last allowed CPU of a pod goes offline for a workqueue
5447 * with a cpumask spanning multiple pods, the workers which were already
5448 * executing the work items for the workqueue will lose their CPU affinity and
5449 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
5450 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
5451 * responsibility to flush the work item from CPU_DOWN_PREPARE.
5452 */
5453static void unbound_wq_update_pwq(struct workqueue_struct *wq, int cpu)
5454{
5455 struct pool_workqueue *old_pwq = NULL, *pwq;
5456 struct workqueue_attrs *target_attrs;
5457
5458 lockdep_assert_held(&wq_pool_mutex);
5459
5460 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
5461 return;
5462
5463 /*
5464 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5465 * Let's use a preallocated one. The following buf is protected by
5466 * CPU hotplug exclusion.
5467 */
5468 target_attrs = unbound_wq_update_pwq_attrs_buf;
5469
5470 copy_workqueue_attrs(to: target_attrs, from: wq->unbound_attrs);
5471 wqattrs_actualize_cpumask(attrs: target_attrs, unbound_cpumask: wq_unbound_cpumask);
5472
5473 /* nothing to do if the target cpumask matches the current pwq */
5474 wq_calc_pod_cpumask(attrs: target_attrs, cpu);
5475 if (wqattrs_equal(a: target_attrs, b: unbound_pwq(wq, cpu)->pool->attrs))
5476 return;
5477
5478 /* create a new pwq */
5479 pwq = alloc_unbound_pwq(wq, attrs: target_attrs);
5480 if (!pwq) {
5481 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
5482 wq->name);
5483 goto use_dfl_pwq;
5484 }
5485
5486 /* Install the new pwq. */
5487 mutex_lock(lock: &wq->mutex);
5488 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5489 goto out_unlock;
5490
5491use_dfl_pwq:
5492 mutex_lock(lock: &wq->mutex);
5493 pwq = unbound_pwq(wq, cpu: -1);
5494 raw_spin_lock_irq(&pwq->pool->lock);
5495 get_pwq(pwq);
5496 raw_spin_unlock_irq(&pwq->pool->lock);
5497 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5498out_unlock:
5499 mutex_unlock(lock: &wq->mutex);
5500 put_pwq_unlocked(pwq: old_pwq);
5501}
5502
5503static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5504{
5505 bool highpri = wq->flags & WQ_HIGHPRI;
5506 int cpu, ret;
5507
5508 lockdep_assert_held(&wq_pool_mutex);
5509
5510 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
5511 if (!wq->cpu_pwq)
5512 goto enomem;
5513
5514 if (!(wq->flags & WQ_UNBOUND)) {
5515 struct worker_pool __percpu *pools;
5516
5517 if (wq->flags & WQ_BH)
5518 pools = bh_worker_pools;
5519 else
5520 pools = cpu_worker_pools;
5521
5522 for_each_possible_cpu(cpu) {
5523 struct pool_workqueue **pwq_p;
5524 struct worker_pool *pool;
5525
5526 pool = &(per_cpu_ptr(pools, cpu)[highpri]);
5527 pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu);
5528
5529 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5530 pool->node);
5531 if (!*pwq_p)
5532 goto enomem;
5533
5534 init_pwq(pwq: *pwq_p, wq, pool);
5535
5536 mutex_lock(lock: &wq->mutex);
5537 link_pwq(pwq: *pwq_p);
5538 mutex_unlock(lock: &wq->mutex);
5539 }
5540 return 0;
5541 }
5542
5543 if (wq->flags & __WQ_ORDERED) {
5544 struct pool_workqueue *dfl_pwq;
5545
5546 ret = apply_workqueue_attrs_locked(wq, attrs: ordered_wq_attrs[highpri]);
5547 /* there should only be single pwq for ordering guarantee */
5548 dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5549 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node ||
5550 wq->pwqs.prev != &dfl_pwq->pwqs_node),
5551 "ordering guarantee broken for workqueue %s\n", wq->name);
5552 } else {
5553 ret = apply_workqueue_attrs_locked(wq, attrs: unbound_std_wq_attrs[highpri]);
5554 }
5555
5556 return ret;
5557
5558enomem:
5559 if (wq->cpu_pwq) {
5560 for_each_possible_cpu(cpu) {
5561 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5562
5563 if (pwq)
5564 kmem_cache_free(s: pwq_cache, objp: pwq);
5565 }
5566 free_percpu(pdata: wq->cpu_pwq);
5567 wq->cpu_pwq = NULL;
5568 }
5569 return -ENOMEM;
5570}
5571
5572static int wq_clamp_max_active(int max_active, unsigned int flags,
5573 const char *name)
5574{
5575 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5576 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5577 max_active, name, 1, WQ_MAX_ACTIVE);
5578
5579 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5580}
5581
5582/*
5583 * Workqueues which may be used during memory reclaim should have a rescuer
5584 * to guarantee forward progress.
5585 */
5586static int init_rescuer(struct workqueue_struct *wq)
5587{
5588 struct worker *rescuer;
5589 char id_buf[WORKER_ID_LEN];
5590 int ret;
5591
5592 lockdep_assert_held(&wq_pool_mutex);
5593
5594 if (!(wq->flags & WQ_MEM_RECLAIM))
5595 return 0;
5596
5597 rescuer = alloc_worker(NUMA_NO_NODE);
5598 if (!rescuer) {
5599 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5600 wq->name);
5601 return -ENOMEM;
5602 }
5603
5604 rescuer->rescue_wq = wq;
5605 format_worker_id(buf: id_buf, size: sizeof(id_buf), worker: rescuer, NULL);
5606
5607 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", id_buf);
5608 if (IS_ERR(ptr: rescuer->task)) {
5609 ret = PTR_ERR(ptr: rescuer->task);
5610 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5611 wq->name, ERR_PTR(ret));
5612 kfree(objp: rescuer);
5613 return ret;
5614 }
5615
5616 wq->rescuer = rescuer;
5617 if (wq->flags & WQ_UNBOUND)
5618 kthread_bind_mask(k: rescuer->task, mask: unbound_effective_cpumask(wq));
5619 else
5620 kthread_bind_mask(k: rescuer->task, cpu_possible_mask);
5621 wake_up_process(tsk: rescuer->task);
5622
5623 return 0;
5624}
5625
5626/**
5627 * wq_adjust_max_active - update a wq's max_active to the current setting
5628 * @wq: target workqueue
5629 *
5630 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5631 * activate inactive work items accordingly. If @wq is freezing, clear
5632 * @wq->max_active to zero.
5633 */
5634static void wq_adjust_max_active(struct workqueue_struct *wq)
5635{
5636 bool activated;
5637 int new_max, new_min;
5638
5639 lockdep_assert_held(&wq->mutex);
5640
5641 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5642 new_max = 0;
5643 new_min = 0;
5644 } else {
5645 new_max = wq->saved_max_active;
5646 new_min = wq->saved_min_active;
5647 }
5648
5649 if (wq->max_active == new_max && wq->min_active == new_min)
5650 return;
5651
5652 /*
5653 * Update @wq->max/min_active and then kick inactive work items if more
5654 * active work items are allowed. This doesn't break work item ordering
5655 * because new work items are always queued behind existing inactive
5656 * work items if there are any.
5657 */
5658 WRITE_ONCE(wq->max_active, new_max);
5659 WRITE_ONCE(wq->min_active, new_min);
5660
5661 if (wq->flags & WQ_UNBOUND)
5662 wq_update_node_max_active(wq, off_cpu: -1);
5663
5664 if (new_max == 0)
5665 return;
5666
5667 /*
5668 * Round-robin through pwq's activating the first inactive work item
5669 * until max_active is filled.
5670 */
5671 do {
5672 struct pool_workqueue *pwq;
5673
5674 activated = false;
5675 for_each_pwq(pwq, wq) {
5676 unsigned long irq_flags;
5677
5678 /* can be called during early boot w/ irq disabled */
5679 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
5680 if (pwq_activate_first_inactive(pwq, fill: true)) {
5681 activated = true;
5682 kick_pool(pool: pwq->pool);
5683 }
5684 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
5685 }
5686 } while (activated);
5687}
5688
5689__printf(1, 0)
5690static struct workqueue_struct *__alloc_workqueue(const char *fmt,
5691 unsigned int flags,
5692 int max_active, va_list args)
5693{
5694 struct workqueue_struct *wq;
5695 size_t wq_size;
5696 int name_len;
5697
5698 if (flags & WQ_BH) {
5699 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS))
5700 return NULL;
5701 if (WARN_ON_ONCE(max_active))
5702 return NULL;
5703 }
5704
5705 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5706 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5707 flags |= WQ_UNBOUND;
5708
5709 /* allocate wq and format name */
5710 if (flags & WQ_UNBOUND)
5711 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5712 else
5713 wq_size = sizeof(*wq);
5714
5715 wq = kzalloc_noprof(size: wq_size, GFP_KERNEL);
5716 if (!wq)
5717 return NULL;
5718
5719 if (flags & WQ_UNBOUND) {
5720 wq->unbound_attrs = alloc_workqueue_attrs_noprof();
5721 if (!wq->unbound_attrs)
5722 goto err_free_wq;
5723 }
5724
5725 name_len = vsnprintf(buf: wq->name, size: sizeof(wq->name), fmt, args);
5726
5727 if (name_len >= WQ_NAME_LEN)
5728 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5729 wq->name);
5730
5731 if (flags & WQ_BH) {
5732 /*
5733 * BH workqueues always share a single execution context per CPU
5734 * and don't impose any max_active limit.
5735 */
5736 max_active = INT_MAX;
5737 } else {
5738 max_active = max_active ?: WQ_DFL_ACTIVE;
5739 max_active = wq_clamp_max_active(max_active, flags, name: wq->name);
5740 }
5741
5742 /* init wq */
5743 wq->flags = flags;
5744 wq->max_active = max_active;
5745 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5746 wq->saved_max_active = wq->max_active;
5747 wq->saved_min_active = wq->min_active;
5748 mutex_init(&wq->mutex);
5749 atomic_set(v: &wq->nr_pwqs_to_flush, i: 0);
5750 INIT_LIST_HEAD(list: &wq->pwqs);
5751 INIT_LIST_HEAD(list: &wq->flusher_queue);
5752 INIT_LIST_HEAD(list: &wq->flusher_overflow);
5753 INIT_LIST_HEAD(list: &wq->maydays);
5754
5755 INIT_LIST_HEAD(list: &wq->list);
5756
5757 if (flags & WQ_UNBOUND) {
5758 if (alloc_node_nr_active(nna_ar: wq->node_nr_active) < 0)
5759 goto err_free_wq;
5760 }
5761
5762 /*
5763 * wq_pool_mutex protects the workqueues list, allocations of PWQs,
5764 * and the global freeze state.
5765 */
5766 apply_wqattrs_lock();
5767
5768 if (alloc_and_link_pwqs(wq) < 0)
5769 goto err_unlock_free_node_nr_active;
5770
5771 mutex_lock(lock: &wq->mutex);
5772 wq_adjust_max_active(wq);
5773 mutex_unlock(lock: &wq->mutex);
5774
5775 list_add_tail_rcu(new: &wq->list, head: &workqueues);
5776
5777 if (wq_online && init_rescuer(wq) < 0)
5778 goto err_unlock_destroy;
5779
5780 apply_wqattrs_unlock();
5781
5782 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5783 goto err_destroy;
5784
5785 return wq;
5786
5787err_unlock_free_node_nr_active:
5788 apply_wqattrs_unlock();
5789 /*
5790 * Failed alloc_and_link_pwqs() may leave pending pwq->release_work,
5791 * flushing the pwq_release_worker ensures that the pwq_release_workfn()
5792 * completes before calling kfree(wq).
5793 */
5794 if (wq->flags & WQ_UNBOUND) {
5795 kthread_flush_worker(worker: pwq_release_worker);
5796 free_node_nr_active(nna_ar: wq->node_nr_active);
5797 }
5798err_free_wq:
5799 free_workqueue_attrs(attrs: wq->unbound_attrs);
5800 kfree(objp: wq);
5801 return NULL;
5802err_unlock_destroy:
5803 apply_wqattrs_unlock();
5804err_destroy:
5805 destroy_workqueue(wq);
5806 return NULL;
5807}
5808
5809__printf(1, 4)
5810struct workqueue_struct *alloc_workqueue_noprof(const char *fmt,
5811 unsigned int flags,
5812 int max_active, ...)
5813{
5814 struct workqueue_struct *wq;
5815 va_list args;
5816
5817 va_start(args, max_active);
5818 wq = __alloc_workqueue(fmt, flags, max_active, args);
5819 va_end(args);
5820 if (!wq)
5821 return NULL;
5822
5823 wq_init_lockdep(wq);
5824
5825 return wq;
5826}
5827EXPORT_SYMBOL_GPL(alloc_workqueue_noprof);
5828
5829#ifdef CONFIG_LOCKDEP
5830__printf(1, 5)
5831struct workqueue_struct *
5832alloc_workqueue_lockdep_map(const char *fmt, unsigned int flags,
5833 int max_active, struct lockdep_map *lockdep_map, ...)
5834{
5835 struct workqueue_struct *wq;
5836 va_list args;
5837
5838 va_start(args, lockdep_map);
5839 wq = __alloc_workqueue(fmt, flags, max_active, args);
5840 va_end(args);
5841 if (!wq)
5842 return NULL;
5843
5844 wq->lockdep_map = lockdep_map;
5845
5846 return wq;
5847}
5848EXPORT_SYMBOL_GPL(alloc_workqueue_lockdep_map);
5849#endif
5850
5851static bool pwq_busy(struct pool_workqueue *pwq)
5852{
5853 int i;
5854
5855 for (i = 0; i < WORK_NR_COLORS; i++)
5856 if (pwq->nr_in_flight[i])
5857 return true;
5858
5859 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
5860 return true;
5861 if (!pwq_is_empty(pwq))
5862 return true;
5863
5864 return false;
5865}
5866
5867/**
5868 * destroy_workqueue - safely terminate a workqueue
5869 * @wq: target workqueue
5870 *
5871 * Safely destroy a workqueue. All work currently pending will be done first.
5872 *
5873 * This function does NOT guarantee that non-pending work that has been
5874 * submitted with queue_delayed_work() and similar functions will be done
5875 * before destroying the workqueue. The fundamental problem is that, currently,
5876 * the workqueue has no way of accessing non-pending delayed_work. delayed_work
5877 * is only linked on the timer-side. All delayed_work must, therefore, be
5878 * canceled before calling this function.
5879 *
5880 * TODO: It would be better if the problem described above wouldn't exist and
5881 * destroy_workqueue() would cleanly cancel all pending and non-pending
5882 * delayed_work.
5883 */
5884void destroy_workqueue(struct workqueue_struct *wq)
5885{
5886 struct pool_workqueue *pwq;
5887 int cpu;
5888
5889 /*
5890 * Remove it from sysfs first so that sanity check failure doesn't
5891 * lead to sysfs name conflicts.
5892 */
5893 workqueue_sysfs_unregister(wq);
5894
5895 /* mark the workqueue destruction is in progress */
5896 mutex_lock(lock: &wq->mutex);
5897 wq->flags |= __WQ_DESTROYING;
5898 mutex_unlock(lock: &wq->mutex);
5899
5900 /* drain it before proceeding with destruction */
5901 drain_workqueue(wq);
5902
5903 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5904 if (wq->rescuer) {
5905 struct worker *rescuer = wq->rescuer;
5906
5907 /* this prevents new queueing */
5908 raw_spin_lock_irq(&wq_mayday_lock);
5909 wq->rescuer = NULL;
5910 raw_spin_unlock_irq(&wq_mayday_lock);
5911
5912 /* rescuer will empty maydays list before exiting */
5913 kthread_stop(k: rescuer->task);
5914 kfree(objp: rescuer);
5915 }
5916
5917 /*
5918 * Sanity checks - grab all the locks so that we wait for all
5919 * in-flight operations which may do put_pwq().
5920 */
5921 mutex_lock(lock: &wq_pool_mutex);
5922 mutex_lock(lock: &wq->mutex);
5923 for_each_pwq(pwq, wq) {
5924 raw_spin_lock_irq(&pwq->pool->lock);
5925 if (WARN_ON(pwq_busy(pwq))) {
5926 pr_warn("%s: %s has the following busy pwq\n",
5927 __func__, wq->name);
5928 show_pwq(pwq);
5929 raw_spin_unlock_irq(&pwq->pool->lock);
5930 mutex_unlock(lock: &wq->mutex);
5931 mutex_unlock(lock: &wq_pool_mutex);
5932 show_one_workqueue(wq);
5933 return;
5934 }
5935 raw_spin_unlock_irq(&pwq->pool->lock);
5936 }
5937 mutex_unlock(lock: &wq->mutex);
5938
5939 /*
5940 * wq list is used to freeze wq, remove from list after
5941 * flushing is complete in case freeze races us.
5942 */
5943 list_del_rcu(entry: &wq->list);
5944 mutex_unlock(lock: &wq_pool_mutex);
5945
5946 /*
5947 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5948 * to put the base refs. @wq will be auto-destroyed from the last
5949 * pwq_put. RCU read lock prevents @wq from going away from under us.
5950 */
5951 rcu_read_lock();
5952
5953 for_each_possible_cpu(cpu) {
5954 put_pwq_unlocked(pwq: unbound_pwq(wq, cpu));
5955 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
5956 }
5957
5958 put_pwq_unlocked(pwq: unbound_pwq(wq, cpu: -1));
5959 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
5960
5961 rcu_read_unlock();
5962}
5963EXPORT_SYMBOL_GPL(destroy_workqueue);
5964
5965/**
5966 * workqueue_set_max_active - adjust max_active of a workqueue
5967 * @wq: target workqueue
5968 * @max_active: new max_active value.
5969 *
5970 * Set max_active of @wq to @max_active. See the alloc_workqueue() function
5971 * comment.
5972 *
5973 * CONTEXT:
5974 * Don't call from IRQ context.
5975 */
5976void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5977{
5978 /* max_active doesn't mean anything for BH workqueues */
5979 if (WARN_ON(wq->flags & WQ_BH))
5980 return;
5981 /* disallow meddling with max_active for ordered workqueues */
5982 if (WARN_ON(wq->flags & __WQ_ORDERED))
5983 return;
5984
5985 max_active = wq_clamp_max_active(max_active, flags: wq->flags, name: wq->name);
5986
5987 mutex_lock(lock: &wq->mutex);
5988
5989 wq->saved_max_active = max_active;
5990 if (wq->flags & WQ_UNBOUND)
5991 wq->saved_min_active = min(wq->saved_min_active, max_active);
5992
5993 wq_adjust_max_active(wq);
5994
5995 mutex_unlock(lock: &wq->mutex);
5996}
5997EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5998
5999/**
6000 * workqueue_set_min_active - adjust min_active of an unbound workqueue
6001 * @wq: target unbound workqueue
6002 * @min_active: new min_active value
6003 *
6004 * Set min_active of an unbound workqueue. Unlike other types of workqueues, an
6005 * unbound workqueue is not guaranteed to be able to process max_active
6006 * interdependent work items. Instead, an unbound workqueue is guaranteed to be
6007 * able to process min_active number of interdependent work items which is
6008 * %WQ_DFL_MIN_ACTIVE by default.
6009 *
6010 * Use this function to adjust the min_active value between 0 and the current
6011 * max_active.
6012 */
6013void workqueue_set_min_active(struct workqueue_struct *wq, int min_active)
6014{
6015 /* min_active is only meaningful for non-ordered unbound workqueues */
6016 if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) !=
6017 WQ_UNBOUND))
6018 return;
6019
6020 mutex_lock(lock: &wq->mutex);
6021 wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active);
6022 wq_adjust_max_active(wq);
6023 mutex_unlock(lock: &wq->mutex);
6024}
6025
6026/**
6027 * current_work - retrieve %current task's work struct
6028 *
6029 * Determine if %current task is a workqueue worker and what it's working on.
6030 * Useful to find out the context that the %current task is running in.
6031 *
6032 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
6033 */
6034struct work_struct *current_work(void)
6035{
6036 struct worker *worker = current_wq_worker();
6037
6038 return worker ? worker->current_work : NULL;
6039}
6040EXPORT_SYMBOL(current_work);
6041
6042/**
6043 * current_is_workqueue_rescuer - is %current workqueue rescuer?
6044 *
6045 * Determine whether %current is a workqueue rescuer. Can be used from
6046 * work functions to determine whether it's being run off the rescuer task.
6047 *
6048 * Return: %true if %current is a workqueue rescuer. %false otherwise.
6049 */
6050bool current_is_workqueue_rescuer(void)
6051{
6052 struct worker *worker = current_wq_worker();
6053
6054 return worker && worker->rescue_wq;
6055}
6056
6057/**
6058 * workqueue_congested - test whether a workqueue is congested
6059 * @cpu: CPU in question
6060 * @wq: target workqueue
6061 *
6062 * Test whether @wq's cpu workqueue for @cpu is congested. There is
6063 * no synchronization around this function and the test result is
6064 * unreliable and only useful as advisory hints or for debugging.
6065 *
6066 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
6067 *
6068 * With the exception of ordered workqueues, all workqueues have per-cpu
6069 * pool_workqueues, each with its own congested state. A workqueue being
6070 * congested on one CPU doesn't mean that the workqueue is contested on any
6071 * other CPUs.
6072 *
6073 * Return:
6074 * %true if congested, %false otherwise.
6075 */
6076bool workqueue_congested(int cpu, struct workqueue_struct *wq)
6077{
6078 struct pool_workqueue *pwq;
6079 bool ret;
6080
6081 preempt_disable();
6082
6083 if (cpu == WORK_CPU_UNBOUND)
6084 cpu = smp_processor_id();
6085
6086 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
6087 ret = !list_empty(head: &pwq->inactive_works);
6088
6089 preempt_enable();
6090
6091 return ret;
6092}
6093EXPORT_SYMBOL_GPL(workqueue_congested);
6094
6095/**
6096 * work_busy - test whether a work is currently pending or running
6097 * @work: the work to be tested
6098 *
6099 * Test whether @work is currently pending or running. There is no
6100 * synchronization around this function and the test result is
6101 * unreliable and only useful as advisory hints or for debugging.
6102 *
6103 * Return:
6104 * OR'd bitmask of WORK_BUSY_* bits.
6105 */
6106unsigned int work_busy(struct work_struct *work)
6107{
6108 struct worker_pool *pool;
6109 unsigned long irq_flags;
6110 unsigned int ret = 0;
6111
6112 if (work_pending(work))
6113 ret |= WORK_BUSY_PENDING;
6114
6115 rcu_read_lock();
6116 pool = get_work_pool(work);
6117 if (pool) {
6118 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6119 if (find_worker_executing_work(pool, work))
6120 ret |= WORK_BUSY_RUNNING;
6121 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6122 }
6123 rcu_read_unlock();
6124
6125 return ret;
6126}
6127EXPORT_SYMBOL_GPL(work_busy);
6128
6129/**
6130 * set_worker_desc - set description for the current work item
6131 * @fmt: printf-style format string
6132 * @...: arguments for the format string
6133 *
6134 * This function can be called by a running work function to describe what
6135 * the work item is about. If the worker task gets dumped, this
6136 * information will be printed out together to help debugging. The
6137 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
6138 */
6139void set_worker_desc(const char *fmt, ...)
6140{
6141 struct worker *worker = current_wq_worker();
6142 va_list args;
6143
6144 if (worker) {
6145 va_start(args, fmt);
6146 vsnprintf(buf: worker->desc, size: sizeof(worker->desc), fmt, args);
6147 va_end(args);
6148 }
6149}
6150EXPORT_SYMBOL_GPL(set_worker_desc);
6151
6152/**
6153 * print_worker_info - print out worker information and description
6154 * @log_lvl: the log level to use when printing
6155 * @task: target task
6156 *
6157 * If @task is a worker and currently executing a work item, print out the
6158 * name of the workqueue being serviced and worker description set with
6159 * set_worker_desc() by the currently executing work item.
6160 *
6161 * This function can be safely called on any task as long as the
6162 * task_struct itself is accessible. While safe, this function isn't
6163 * synchronized and may print out mixups or garbages of limited length.
6164 */
6165void print_worker_info(const char *log_lvl, struct task_struct *task)
6166{
6167 work_func_t *fn = NULL;
6168 char name[WQ_NAME_LEN] = { };
6169 char desc[WORKER_DESC_LEN] = { };
6170 struct pool_workqueue *pwq = NULL;
6171 struct workqueue_struct *wq = NULL;
6172 struct worker *worker;
6173
6174 if (!(task->flags & PF_WQ_WORKER))
6175 return;
6176
6177 /*
6178 * This function is called without any synchronization and @task
6179 * could be in any state. Be careful with dereferences.
6180 */
6181 worker = kthread_probe_data(k: task);
6182
6183 /*
6184 * Carefully copy the associated workqueue's workfn, name and desc.
6185 * Keep the original last '\0' in case the original is garbage.
6186 */
6187 copy_from_kernel_nofault(dst: &fn, src: &worker->current_func, size: sizeof(fn));
6188 copy_from_kernel_nofault(dst: &pwq, src: &worker->current_pwq, size: sizeof(pwq));
6189 copy_from_kernel_nofault(dst: &wq, src: &pwq->wq, size: sizeof(wq));
6190 copy_from_kernel_nofault(dst: name, src: wq->name, size: sizeof(name) - 1);
6191 copy_from_kernel_nofault(dst: desc, src: worker->desc, size: sizeof(desc) - 1);
6192
6193 if (fn || name[0] || desc[0]) {
6194 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
6195 if (strcmp(name, desc))
6196 pr_cont(" (%s)", desc);
6197 pr_cont("\n");
6198 }
6199}
6200
6201static void pr_cont_pool_info(struct worker_pool *pool)
6202{
6203 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
6204 if (pool->node != NUMA_NO_NODE)
6205 pr_cont(" node=%d", pool->node);
6206 pr_cont(" flags=0x%x", pool->flags);
6207 if (pool->flags & POOL_BH)
6208 pr_cont(" bh%s",
6209 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6210 else
6211 pr_cont(" nice=%d", pool->attrs->nice);
6212}
6213
6214static void pr_cont_worker_id(struct worker *worker)
6215{
6216 struct worker_pool *pool = worker->pool;
6217
6218 if (pool->flags & WQ_BH)
6219 pr_cont("bh%s",
6220 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6221 else
6222 pr_cont("%d%s", task_pid_nr(worker->task),
6223 worker->rescue_wq ? "(RESCUER)" : "");
6224}
6225
6226struct pr_cont_work_struct {
6227 bool comma;
6228 work_func_t func;
6229 long ctr;
6230};
6231
6232static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
6233{
6234 if (!pcwsp->ctr)
6235 goto out_record;
6236 if (func == pcwsp->func) {
6237 pcwsp->ctr++;
6238 return;
6239 }
6240 if (pcwsp->ctr == 1)
6241 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
6242 else
6243 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
6244 pcwsp->ctr = 0;
6245out_record:
6246 if ((long)func == -1L)
6247 return;
6248 pcwsp->comma = comma;
6249 pcwsp->func = func;
6250 pcwsp->ctr = 1;
6251}
6252
6253static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
6254{
6255 if (work->func == wq_barrier_func) {
6256 struct wq_barrier *barr;
6257
6258 barr = container_of(work, struct wq_barrier, work);
6259
6260 pr_cont_work_flush(comma, func: (work_func_t)-1, pcwsp);
6261 pr_cont("%s BAR(%d)", comma ? "," : "",
6262 task_pid_nr(barr->task));
6263 } else {
6264 if (!comma)
6265 pr_cont_work_flush(comma, func: (work_func_t)-1, pcwsp);
6266 pr_cont_work_flush(comma, func: work->func, pcwsp);
6267 }
6268}
6269
6270static void show_pwq(struct pool_workqueue *pwq)
6271{
6272 struct pr_cont_work_struct pcws = { .ctr = 0, };
6273 struct worker_pool *pool = pwq->pool;
6274 struct work_struct *work;
6275 struct worker *worker;
6276 bool has_in_flight = false, has_pending = false;
6277 int bkt;
6278
6279 pr_info(" pwq %d:", pool->id);
6280 pr_cont_pool_info(pool);
6281
6282 pr_cont(" active=%d refcnt=%d%s\n",
6283 pwq->nr_active, pwq->refcnt,
6284 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6285
6286 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6287 if (worker->current_pwq == pwq) {
6288 has_in_flight = true;
6289 break;
6290 }
6291 }
6292 if (has_in_flight) {
6293 bool comma = false;
6294
6295 pr_info(" in-flight:");
6296 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6297 if (worker->current_pwq != pwq)
6298 continue;
6299
6300 pr_cont(" %s", comma ? "," : "");
6301 pr_cont_worker_id(worker);
6302 pr_cont(":%ps", worker->current_func);
6303 list_for_each_entry(work, &worker->scheduled, entry)
6304 pr_cont_work(comma: false, work, pcwsp: &pcws);
6305 pr_cont_work_flush(comma, func: (work_func_t)-1L, pcwsp: &pcws);
6306 comma = true;
6307 }
6308 pr_cont("\n");
6309 }
6310
6311 list_for_each_entry(work, &pool->worklist, entry) {
6312 if (get_work_pwq(work) == pwq) {
6313 has_pending = true;
6314 break;
6315 }
6316 }
6317 if (has_pending) {
6318 bool comma = false;
6319
6320 pr_info(" pending:");
6321 list_for_each_entry(work, &pool->worklist, entry) {
6322 if (get_work_pwq(work) != pwq)
6323 continue;
6324
6325 pr_cont_work(comma, work, pcwsp: &pcws);
6326 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6327 }
6328 pr_cont_work_flush(comma, func: (work_func_t)-1L, pcwsp: &pcws);
6329 pr_cont("\n");
6330 }
6331
6332 if (!list_empty(head: &pwq->inactive_works)) {
6333 bool comma = false;
6334
6335 pr_info(" inactive:");
6336 list_for_each_entry(work, &pwq->inactive_works, entry) {
6337 pr_cont_work(comma, work, pcwsp: &pcws);
6338 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6339 }
6340 pr_cont_work_flush(comma, func: (work_func_t)-1L, pcwsp: &pcws);
6341 pr_cont("\n");
6342 }
6343}
6344
6345/**
6346 * show_one_workqueue - dump state of specified workqueue
6347 * @wq: workqueue whose state will be printed
6348 */
6349void show_one_workqueue(struct workqueue_struct *wq)
6350{
6351 struct pool_workqueue *pwq;
6352 bool idle = true;
6353 unsigned long irq_flags;
6354
6355 for_each_pwq(pwq, wq) {
6356 if (!pwq_is_empty(pwq)) {
6357 idle = false;
6358 break;
6359 }
6360 }
6361 if (idle) /* Nothing to print for idle workqueue */
6362 return;
6363
6364 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6365
6366 for_each_pwq(pwq, wq) {
6367 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
6368 if (!pwq_is_empty(pwq)) {
6369 /*
6370 * Defer printing to avoid deadlocks in console
6371 * drivers that queue work while holding locks
6372 * also taken in their write paths.
6373 */
6374 printk_deferred_enter();
6375 show_pwq(pwq);
6376 printk_deferred_exit();
6377 }
6378 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
6379 /*
6380 * We could be printing a lot from atomic context, e.g.
6381 * sysrq-t -> show_all_workqueues(). Avoid triggering
6382 * hard lockup.
6383 */
6384 touch_nmi_watchdog();
6385 }
6386
6387}
6388
6389/**
6390 * show_one_worker_pool - dump state of specified worker pool
6391 * @pool: worker pool whose state will be printed
6392 */
6393static void show_one_worker_pool(struct worker_pool *pool)
6394{
6395 struct worker *worker;
6396 bool first = true;
6397 unsigned long irq_flags;
6398 unsigned long hung = 0;
6399
6400 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6401 if (pool->nr_workers == pool->nr_idle)
6402 goto next_pool;
6403
6404 /* How long the first pending work is waiting for a worker. */
6405 if (!list_empty(head: &pool->worklist))
6406 hung = jiffies_to_msecs(j: jiffies - pool->watchdog_ts) / 1000;
6407
6408 /*
6409 * Defer printing to avoid deadlocks in console drivers that
6410 * queue work while holding locks also taken in their write
6411 * paths.
6412 */
6413 printk_deferred_enter();
6414 pr_info("pool %d:", pool->id);
6415 pr_cont_pool_info(pool);
6416 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
6417 if (pool->manager)
6418 pr_cont(" manager: %d",
6419 task_pid_nr(pool->manager->task));
6420 list_for_each_entry(worker, &pool->idle_list, entry) {
6421 pr_cont(" %s", first ? "idle: " : "");
6422 pr_cont_worker_id(worker);
6423 first = false;
6424 }
6425 pr_cont("\n");
6426 printk_deferred_exit();
6427next_pool:
6428 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6429 /*
6430 * We could be printing a lot from atomic context, e.g.
6431 * sysrq-t -> show_all_workqueues(). Avoid triggering
6432 * hard lockup.
6433 */
6434 touch_nmi_watchdog();
6435
6436}
6437
6438/**
6439 * show_all_workqueues - dump workqueue state
6440 *
6441 * Called from a sysrq handler and prints out all busy workqueues and pools.
6442 */
6443void show_all_workqueues(void)
6444{
6445 struct workqueue_struct *wq;
6446 struct worker_pool *pool;
6447 int pi;
6448
6449 rcu_read_lock();
6450
6451 pr_info("Showing busy workqueues and worker pools:\n");
6452
6453 list_for_each_entry_rcu(wq, &workqueues, list)
6454 show_one_workqueue(wq);
6455
6456 for_each_pool(pool, pi)
6457 show_one_worker_pool(pool);
6458
6459 rcu_read_unlock();
6460}
6461
6462/**
6463 * show_freezable_workqueues - dump freezable workqueue state
6464 *
6465 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
6466 * still busy.
6467 */
6468void show_freezable_workqueues(void)
6469{
6470 struct workqueue_struct *wq;
6471
6472 rcu_read_lock();
6473
6474 pr_info("Showing freezable workqueues that are still busy:\n");
6475
6476 list_for_each_entry_rcu(wq, &workqueues, list) {
6477 if (!(wq->flags & WQ_FREEZABLE))
6478 continue;
6479 show_one_workqueue(wq);
6480 }
6481
6482 rcu_read_unlock();
6483}
6484
6485/* used to show worker information through /proc/PID/{comm,stat,status} */
6486void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
6487{
6488 /* stabilize PF_WQ_WORKER and worker pool association */
6489 mutex_lock(lock: &wq_pool_attach_mutex);
6490
6491 if (task->flags & PF_WQ_WORKER) {
6492 struct worker *worker = kthread_data(k: task);
6493 struct worker_pool *pool = worker->pool;
6494 int off;
6495
6496 off = format_worker_id(buf, size, worker, pool);
6497
6498 if (pool) {
6499 raw_spin_lock_irq(&pool->lock);
6500 /*
6501 * ->desc tracks information (wq name or
6502 * set_worker_desc()) for the latest execution. If
6503 * current, prepend '+', otherwise '-'.
6504 */
6505 if (worker->desc[0] != '\0') {
6506 if (worker->current_work)
6507 scnprintf(buf: buf + off, size: size - off, fmt: "+%s",
6508 worker->desc);
6509 else
6510 scnprintf(buf: buf + off, size: size - off, fmt: "-%s",
6511 worker->desc);
6512 }
6513 raw_spin_unlock_irq(&pool->lock);
6514 }
6515 } else {
6516 strscpy(buf, task->comm, size);
6517 }
6518
6519 mutex_unlock(lock: &wq_pool_attach_mutex);
6520}
6521
6522#ifdef CONFIG_SMP
6523
6524/*
6525 * CPU hotplug.
6526 *
6527 * There are two challenges in supporting CPU hotplug. Firstly, there
6528 * are a lot of assumptions on strong associations among work, pwq and
6529 * pool which make migrating pending and scheduled works very
6530 * difficult to implement without impacting hot paths. Secondly,
6531 * worker pools serve mix of short, long and very long running works making
6532 * blocked draining impractical.
6533 *
6534 * This is solved by allowing the pools to be disassociated from the CPU
6535 * running as an unbound one and allowing it to be reattached later if the
6536 * cpu comes back online.
6537 */
6538
6539static void unbind_workers(int cpu)
6540{
6541 struct worker_pool *pool;
6542 struct worker *worker;
6543
6544 for_each_cpu_worker_pool(pool, cpu) {
6545 mutex_lock(lock: &wq_pool_attach_mutex);
6546 raw_spin_lock_irq(&pool->lock);
6547
6548 /*
6549 * We've blocked all attach/detach operations. Make all workers
6550 * unbound and set DISASSOCIATED. Before this, all workers
6551 * must be on the cpu. After this, they may become diasporas.
6552 * And the preemption disabled section in their sched callbacks
6553 * are guaranteed to see WORKER_UNBOUND since the code here
6554 * is on the same cpu.
6555 */
6556 for_each_pool_worker(worker, pool)
6557 worker->flags |= WORKER_UNBOUND;
6558
6559 pool->flags |= POOL_DISASSOCIATED;
6560
6561 /*
6562 * The handling of nr_running in sched callbacks are disabled
6563 * now. Zap nr_running. After this, nr_running stays zero and
6564 * need_more_worker() and keep_working() are always true as
6565 * long as the worklist is not empty. This pool now behaves as
6566 * an unbound (in terms of concurrency management) pool which
6567 * are served by workers tied to the pool.
6568 */
6569 pool->nr_running = 0;
6570
6571 /*
6572 * With concurrency management just turned off, a busy
6573 * worker blocking could lead to lengthy stalls. Kick off
6574 * unbound chain execution of currently pending work items.
6575 */
6576 kick_pool(pool);
6577
6578 raw_spin_unlock_irq(&pool->lock);
6579
6580 for_each_pool_worker(worker, pool)
6581 unbind_worker(worker);
6582
6583 mutex_unlock(lock: &wq_pool_attach_mutex);
6584 }
6585}
6586
6587/**
6588 * rebind_workers - rebind all workers of a pool to the associated CPU
6589 * @pool: pool of interest
6590 *
6591 * @pool->cpu is coming online. Rebind all workers to the CPU.
6592 */
6593static void rebind_workers(struct worker_pool *pool)
6594{
6595 struct worker *worker;
6596
6597 lockdep_assert_held(&wq_pool_attach_mutex);
6598
6599 /*
6600 * Restore CPU affinity of all workers. As all idle workers should
6601 * be on the run-queue of the associated CPU before any local
6602 * wake-ups for concurrency management happen, restore CPU affinity
6603 * of all workers first and then clear UNBOUND. As we're called
6604 * from CPU_ONLINE, the following shouldn't fail.
6605 */
6606 for_each_pool_worker(worker, pool) {
6607 kthread_set_per_cpu(k: worker->task, cpu: pool->cpu);
6608 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6609 pool_allowed_cpus(pool)) < 0);
6610 }
6611
6612 raw_spin_lock_irq(&pool->lock);
6613
6614 pool->flags &= ~POOL_DISASSOCIATED;
6615
6616 for_each_pool_worker(worker, pool) {
6617 unsigned int worker_flags = worker->flags;
6618
6619 /*
6620 * We want to clear UNBOUND but can't directly call
6621 * worker_clr_flags() or adjust nr_running. Atomically
6622 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6623 * @worker will clear REBOUND using worker_clr_flags() when
6624 * it initiates the next execution cycle thus restoring
6625 * concurrency management. Note that when or whether
6626 * @worker clears REBOUND doesn't affect correctness.
6627 *
6628 * WRITE_ONCE() is necessary because @worker->flags may be
6629 * tested without holding any lock in
6630 * wq_worker_running(). Without it, NOT_RUNNING test may
6631 * fail incorrectly leading to premature concurrency
6632 * management operations.
6633 */
6634 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6635 worker_flags |= WORKER_REBOUND;
6636 worker_flags &= ~WORKER_UNBOUND;
6637 WRITE_ONCE(worker->flags, worker_flags);
6638 }
6639
6640 raw_spin_unlock_irq(&pool->lock);
6641}
6642
6643/**
6644 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6645 * @pool: unbound pool of interest
6646 * @cpu: the CPU which is coming up
6647 *
6648 * An unbound pool may end up with a cpumask which doesn't have any online
6649 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6650 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6651 * online CPU before, cpus_allowed of all its workers should be restored.
6652 */
6653static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6654{
6655 static cpumask_t cpumask;
6656 struct worker *worker;
6657
6658 lockdep_assert_held(&wq_pool_attach_mutex);
6659
6660 /* is @cpu allowed for @pool? */
6661 if (!cpumask_test_cpu(cpu, cpumask: pool->attrs->cpumask))
6662 return;
6663
6664 cpumask_and(dstp: &cpumask, src1p: pool->attrs->cpumask, cpu_online_mask);
6665
6666 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6667 for_each_pool_worker(worker, pool)
6668 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6669}
6670
6671int workqueue_prepare_cpu(unsigned int cpu)
6672{
6673 struct worker_pool *pool;
6674
6675 for_each_cpu_worker_pool(pool, cpu) {
6676 if (pool->nr_workers)
6677 continue;
6678 if (!create_worker(pool))
6679 return -ENOMEM;
6680 }
6681 return 0;
6682}
6683
6684int workqueue_online_cpu(unsigned int cpu)
6685{
6686 struct worker_pool *pool;
6687 struct workqueue_struct *wq;
6688 int pi;
6689
6690 mutex_lock(lock: &wq_pool_mutex);
6691
6692 cpumask_set_cpu(cpu, dstp: wq_online_cpumask);
6693
6694 for_each_pool(pool, pi) {
6695 /* BH pools aren't affected by hotplug */
6696 if (pool->flags & POOL_BH)
6697 continue;
6698
6699 mutex_lock(lock: &wq_pool_attach_mutex);
6700 if (pool->cpu == cpu)
6701 rebind_workers(pool);
6702 else if (pool->cpu < 0)
6703 restore_unbound_workers_cpumask(pool, cpu);
6704 mutex_unlock(lock: &wq_pool_attach_mutex);
6705 }
6706
6707 /* update pod affinity of unbound workqueues */
6708 list_for_each_entry(wq, &workqueues, list) {
6709 struct workqueue_attrs *attrs = wq->unbound_attrs;
6710
6711 if (attrs) {
6712 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6713 int tcpu;
6714
6715 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6716 unbound_wq_update_pwq(wq, cpu: tcpu);
6717
6718 mutex_lock(lock: &wq->mutex);
6719 wq_update_node_max_active(wq, off_cpu: -1);
6720 mutex_unlock(lock: &wq->mutex);
6721 }
6722 }
6723
6724 mutex_unlock(lock: &wq_pool_mutex);
6725 return 0;
6726}
6727
6728int workqueue_offline_cpu(unsigned int cpu)
6729{
6730 struct workqueue_struct *wq;
6731
6732 /* unbinding per-cpu workers should happen on the local CPU */
6733 if (WARN_ON(cpu != smp_processor_id()))
6734 return -1;
6735
6736 unbind_workers(cpu);
6737
6738 /* update pod affinity of unbound workqueues */
6739 mutex_lock(lock: &wq_pool_mutex);
6740
6741 cpumask_clear_cpu(cpu, dstp: wq_online_cpumask);
6742
6743 list_for_each_entry(wq, &workqueues, list) {
6744 struct workqueue_attrs *attrs = wq->unbound_attrs;
6745
6746 if (attrs) {
6747 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6748 int tcpu;
6749
6750 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6751 unbound_wq_update_pwq(wq, cpu: tcpu);
6752
6753 mutex_lock(lock: &wq->mutex);
6754 wq_update_node_max_active(wq, off_cpu: cpu);
6755 mutex_unlock(lock: &wq->mutex);
6756 }
6757 }
6758 mutex_unlock(lock: &wq_pool_mutex);
6759
6760 return 0;
6761}
6762
6763struct work_for_cpu {
6764 struct work_struct work;
6765 long (*fn)(void *);
6766 void *arg;
6767 long ret;
6768};
6769
6770static void work_for_cpu_fn(struct work_struct *work)
6771{
6772 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6773
6774 wfc->ret = wfc->fn(wfc->arg);
6775}
6776
6777/**
6778 * work_on_cpu_key - run a function in thread context on a particular cpu
6779 * @cpu: the cpu to run on
6780 * @fn: the function to run
6781 * @arg: the function arg
6782 * @key: The lock class key for lock debugging purposes
6783 *
6784 * It is up to the caller to ensure that the cpu doesn't go offline.
6785 * The caller must not hold any locks which would prevent @fn from completing.
6786 *
6787 * Return: The value @fn returns.
6788 */
6789long work_on_cpu_key(int cpu, long (*fn)(void *),
6790 void *arg, struct lock_class_key *key)
6791{
6792 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6793
6794 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6795 schedule_work_on(cpu, work: &wfc.work);
6796 flush_work(&wfc.work);
6797 destroy_work_on_stack(work: &wfc.work);
6798 return wfc.ret;
6799}
6800EXPORT_SYMBOL_GPL(work_on_cpu_key);
6801#endif /* CONFIG_SMP */
6802
6803#ifdef CONFIG_FREEZER
6804
6805/**
6806 * freeze_workqueues_begin - begin freezing workqueues
6807 *
6808 * Start freezing workqueues. After this function returns, all freezable
6809 * workqueues will queue new works to their inactive_works list instead of
6810 * pool->worklist.
6811 *
6812 * CONTEXT:
6813 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6814 */
6815void freeze_workqueues_begin(void)
6816{
6817 struct workqueue_struct *wq;
6818
6819 mutex_lock(lock: &wq_pool_mutex);
6820
6821 WARN_ON_ONCE(workqueue_freezing);
6822 workqueue_freezing = true;
6823
6824 list_for_each_entry(wq, &workqueues, list) {
6825 mutex_lock(lock: &wq->mutex);
6826 wq_adjust_max_active(wq);
6827 mutex_unlock(lock: &wq->mutex);
6828 }
6829
6830 mutex_unlock(lock: &wq_pool_mutex);
6831}
6832
6833/**
6834 * freeze_workqueues_busy - are freezable workqueues still busy?
6835 *
6836 * Check whether freezing is complete. This function must be called
6837 * between freeze_workqueues_begin() and thaw_workqueues().
6838 *
6839 * CONTEXT:
6840 * Grabs and releases wq_pool_mutex.
6841 *
6842 * Return:
6843 * %true if some freezable workqueues are still busy. %false if freezing
6844 * is complete.
6845 */
6846bool freeze_workqueues_busy(void)
6847{
6848 bool busy = false;
6849 struct workqueue_struct *wq;
6850 struct pool_workqueue *pwq;
6851
6852 mutex_lock(lock: &wq_pool_mutex);
6853
6854 WARN_ON_ONCE(!workqueue_freezing);
6855
6856 list_for_each_entry(wq, &workqueues, list) {
6857 if (!(wq->flags & WQ_FREEZABLE))
6858 continue;
6859 /*
6860 * nr_active is monotonically decreasing. It's safe
6861 * to peek without lock.
6862 */
6863 rcu_read_lock();
6864 for_each_pwq(pwq, wq) {
6865 WARN_ON_ONCE(pwq->nr_active < 0);
6866 if (pwq->nr_active) {
6867 busy = true;
6868 rcu_read_unlock();
6869 goto out_unlock;
6870 }
6871 }
6872 rcu_read_unlock();
6873 }
6874out_unlock:
6875 mutex_unlock(lock: &wq_pool_mutex);
6876 return busy;
6877}
6878
6879/**
6880 * thaw_workqueues - thaw workqueues
6881 *
6882 * Thaw workqueues. Normal queueing is restored and all collected
6883 * frozen works are transferred to their respective pool worklists.
6884 *
6885 * CONTEXT:
6886 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6887 */
6888void thaw_workqueues(void)
6889{
6890 struct workqueue_struct *wq;
6891
6892 mutex_lock(lock: &wq_pool_mutex);
6893
6894 if (!workqueue_freezing)
6895 goto out_unlock;
6896
6897 workqueue_freezing = false;
6898
6899 /* restore max_active and repopulate worklist */
6900 list_for_each_entry(wq, &workqueues, list) {
6901 mutex_lock(lock: &wq->mutex);
6902 wq_adjust_max_active(wq);
6903 mutex_unlock(lock: &wq->mutex);
6904 }
6905
6906out_unlock:
6907 mutex_unlock(lock: &wq_pool_mutex);
6908}
6909#endif /* CONFIG_FREEZER */
6910
6911static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6912{
6913 LIST_HEAD(ctxs);
6914 int ret = 0;
6915 struct workqueue_struct *wq;
6916 struct apply_wqattrs_ctx *ctx, *n;
6917
6918 lockdep_assert_held(&wq_pool_mutex);
6919
6920 list_for_each_entry(wq, &workqueues, list) {
6921 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING))
6922 continue;
6923
6924 ctx = apply_wqattrs_prepare(wq, attrs: wq->unbound_attrs, unbound_cpumask);
6925 if (IS_ERR(ptr: ctx)) {
6926 ret = PTR_ERR(ptr: ctx);
6927 break;
6928 }
6929
6930 list_add_tail(new: &ctx->list, head: &ctxs);
6931 }
6932
6933 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6934 if (!ret)
6935 apply_wqattrs_commit(ctx);
6936 apply_wqattrs_cleanup(ctx);
6937 }
6938
6939 if (!ret) {
6940 mutex_lock(lock: &wq_pool_attach_mutex);
6941 cpumask_copy(dstp: wq_unbound_cpumask, srcp: unbound_cpumask);
6942 mutex_unlock(lock: &wq_pool_attach_mutex);
6943 }
6944 return ret;
6945}
6946
6947/**
6948 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
6949 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
6950 *
6951 * This function can be called from cpuset code to provide a set of isolated
6952 * CPUs that should be excluded from wq_unbound_cpumask.
6953 */
6954int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
6955{
6956 cpumask_var_t cpumask;
6957 int ret = 0;
6958
6959 if (!zalloc_cpumask_var(mask: &cpumask, GFP_KERNEL))
6960 return -ENOMEM;
6961
6962 mutex_lock(lock: &wq_pool_mutex);
6963
6964 /*
6965 * If the operation fails, it will fall back to
6966 * wq_requested_unbound_cpumask which is initially set to
6967 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
6968 * by any subsequent write to workqueue/cpumask sysfs file.
6969 */
6970 if (!cpumask_andnot(dstp: cpumask, src1p: wq_requested_unbound_cpumask, src2p: exclude_cpumask))
6971 cpumask_copy(dstp: cpumask, srcp: wq_requested_unbound_cpumask);
6972 if (!cpumask_equal(src1p: cpumask, src2p: wq_unbound_cpumask))
6973 ret = workqueue_apply_unbound_cpumask(unbound_cpumask: cpumask);
6974
6975 /* Save the current isolated cpumask & export it via sysfs */
6976 if (!ret)
6977 cpumask_copy(dstp: wq_isolated_cpumask, srcp: exclude_cpumask);
6978
6979 mutex_unlock(lock: &wq_pool_mutex);
6980 free_cpumask_var(mask: cpumask);
6981 return ret;
6982}
6983
6984static int parse_affn_scope(const char *val)
6985{
6986 int i;
6987
6988 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6989 if (!strncasecmp(s1: val, s2: wq_affn_names[i], n: strlen(wq_affn_names[i])))
6990 return i;
6991 }
6992 return -EINVAL;
6993}
6994
6995static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
6996{
6997 struct workqueue_struct *wq;
6998 int affn, cpu;
6999
7000 affn = parse_affn_scope(val);
7001 if (affn < 0)
7002 return affn;
7003 if (affn == WQ_AFFN_DFL)
7004 return -EINVAL;
7005
7006 cpus_read_lock();
7007 mutex_lock(lock: &wq_pool_mutex);
7008
7009 wq_affn_dfl = affn;
7010
7011 list_for_each_entry(wq, &workqueues, list) {
7012 for_each_online_cpu(cpu)
7013 unbound_wq_update_pwq(wq, cpu);
7014 }
7015
7016 mutex_unlock(lock: &wq_pool_mutex);
7017 cpus_read_unlock();
7018
7019 return 0;
7020}
7021
7022static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
7023{
7024 return scnprintf(buf: buffer, PAGE_SIZE, fmt: "%s\n", wq_affn_names[wq_affn_dfl]);
7025}
7026
7027static const struct kernel_param_ops wq_affn_dfl_ops = {
7028 .set = wq_affn_dfl_set,
7029 .get = wq_affn_dfl_get,
7030};
7031
7032module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
7033
7034#ifdef CONFIG_SYSFS
7035/*
7036 * Workqueues with WQ_SYSFS flag set is visible to userland via
7037 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
7038 * following attributes.
7039 *
7040 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
7041 * max_active RW int : maximum number of in-flight work items
7042 *
7043 * Unbound workqueues have the following extra attributes.
7044 *
7045 * nice RW int : nice value of the workers
7046 * cpumask RW mask : bitmask of allowed CPUs for the workers
7047 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
7048 * affinity_strict RW bool : worker CPU affinity is strict
7049 */
7050struct wq_device {
7051 struct workqueue_struct *wq;
7052 struct device dev;
7053};
7054
7055static struct workqueue_struct *dev_to_wq(struct device *dev)
7056{
7057 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7058
7059 return wq_dev->wq;
7060}
7061
7062static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
7063 char *buf)
7064{
7065 struct workqueue_struct *wq = dev_to_wq(dev);
7066
7067 return scnprintf(buf, PAGE_SIZE, fmt: "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
7068}
7069static DEVICE_ATTR_RO(per_cpu);
7070
7071static ssize_t max_active_show(struct device *dev,
7072 struct device_attribute *attr, char *buf)
7073{
7074 struct workqueue_struct *wq = dev_to_wq(dev);
7075
7076 return scnprintf(buf, PAGE_SIZE, fmt: "%d\n", wq->saved_max_active);
7077}
7078
7079static ssize_t max_active_store(struct device *dev,
7080 struct device_attribute *attr, const char *buf,
7081 size_t count)
7082{
7083 struct workqueue_struct *wq = dev_to_wq(dev);
7084 int val;
7085
7086 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
7087 return -EINVAL;
7088
7089 workqueue_set_max_active(wq, val);
7090 return count;
7091}
7092static DEVICE_ATTR_RW(max_active);
7093
7094static struct attribute *wq_sysfs_attrs[] = {
7095 &dev_attr_per_cpu.attr,
7096 &dev_attr_max_active.attr,
7097 NULL,
7098};
7099ATTRIBUTE_GROUPS(wq_sysfs);
7100
7101static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
7102 char *buf)
7103{
7104 struct workqueue_struct *wq = dev_to_wq(dev);
7105 int written;
7106
7107 mutex_lock(lock: &wq->mutex);
7108 written = scnprintf(buf, PAGE_SIZE, fmt: "%d\n", wq->unbound_attrs->nice);
7109 mutex_unlock(lock: &wq->mutex);
7110
7111 return written;
7112}
7113
7114/* prepare workqueue_attrs for sysfs store operations */
7115static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
7116{
7117 struct workqueue_attrs *attrs;
7118
7119 lockdep_assert_held(&wq_pool_mutex);
7120
7121 attrs = alloc_workqueue_attrs();
7122 if (!attrs)
7123 return NULL;
7124
7125 copy_workqueue_attrs(to: attrs, from: wq->unbound_attrs);
7126 return attrs;
7127}
7128
7129static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
7130 const char *buf, size_t count)
7131{
7132 struct workqueue_struct *wq = dev_to_wq(dev);
7133 struct workqueue_attrs *attrs;
7134 int ret = -ENOMEM;
7135
7136 apply_wqattrs_lock();
7137
7138 attrs = wq_sysfs_prep_attrs(wq);
7139 if (!attrs)
7140 goto out_unlock;
7141
7142 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
7143 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
7144 ret = apply_workqueue_attrs_locked(wq, attrs);
7145 else
7146 ret = -EINVAL;
7147
7148out_unlock:
7149 apply_wqattrs_unlock();
7150 free_workqueue_attrs(attrs);
7151 return ret ?: count;
7152}
7153
7154static ssize_t wq_cpumask_show(struct device *dev,
7155 struct device_attribute *attr, char *buf)
7156{
7157 struct workqueue_struct *wq = dev_to_wq(dev);
7158 int written;
7159
7160 mutex_lock(lock: &wq->mutex);
7161 written = scnprintf(buf, PAGE_SIZE, fmt: "%*pb\n",
7162 cpumask_pr_args(wq->unbound_attrs->cpumask));
7163 mutex_unlock(lock: &wq->mutex);
7164 return written;
7165}
7166
7167static ssize_t wq_cpumask_store(struct device *dev,
7168 struct device_attribute *attr,
7169 const char *buf, size_t count)
7170{
7171 struct workqueue_struct *wq = dev_to_wq(dev);
7172 struct workqueue_attrs *attrs;
7173 int ret = -ENOMEM;
7174
7175 apply_wqattrs_lock();
7176
7177 attrs = wq_sysfs_prep_attrs(wq);
7178 if (!attrs)
7179 goto out_unlock;
7180
7181 ret = cpumask_parse(buf, dstp: attrs->cpumask);
7182 if (!ret)
7183 ret = apply_workqueue_attrs_locked(wq, attrs);
7184
7185out_unlock:
7186 apply_wqattrs_unlock();
7187 free_workqueue_attrs(attrs);
7188 return ret ?: count;
7189}
7190
7191static ssize_t wq_affn_scope_show(struct device *dev,
7192 struct device_attribute *attr, char *buf)
7193{
7194 struct workqueue_struct *wq = dev_to_wq(dev);
7195 int written;
7196
7197 mutex_lock(lock: &wq->mutex);
7198 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
7199 written = scnprintf(buf, PAGE_SIZE, fmt: "%s (%s)\n",
7200 wq_affn_names[WQ_AFFN_DFL],
7201 wq_affn_names[wq_affn_dfl]);
7202 else
7203 written = scnprintf(buf, PAGE_SIZE, fmt: "%s\n",
7204 wq_affn_names[wq->unbound_attrs->affn_scope]);
7205 mutex_unlock(lock: &wq->mutex);
7206
7207 return written;
7208}
7209
7210static ssize_t wq_affn_scope_store(struct device *dev,
7211 struct device_attribute *attr,
7212 const char *buf, size_t count)
7213{
7214 struct workqueue_struct *wq = dev_to_wq(dev);
7215 struct workqueue_attrs *attrs;
7216 int affn, ret = -ENOMEM;
7217
7218 affn = parse_affn_scope(val: buf);
7219 if (affn < 0)
7220 return affn;
7221
7222 apply_wqattrs_lock();
7223 attrs = wq_sysfs_prep_attrs(wq);
7224 if (attrs) {
7225 attrs->affn_scope = affn;
7226 ret = apply_workqueue_attrs_locked(wq, attrs);
7227 }
7228 apply_wqattrs_unlock();
7229 free_workqueue_attrs(attrs);
7230 return ret ?: count;
7231}
7232
7233static ssize_t wq_affinity_strict_show(struct device *dev,
7234 struct device_attribute *attr, char *buf)
7235{
7236 struct workqueue_struct *wq = dev_to_wq(dev);
7237
7238 return scnprintf(buf, PAGE_SIZE, fmt: "%d\n",
7239 wq->unbound_attrs->affn_strict);
7240}
7241
7242static ssize_t wq_affinity_strict_store(struct device *dev,
7243 struct device_attribute *attr,
7244 const char *buf, size_t count)
7245{
7246 struct workqueue_struct *wq = dev_to_wq(dev);
7247 struct workqueue_attrs *attrs;
7248 int v, ret = -ENOMEM;
7249
7250 if (sscanf(buf, "%d", &v) != 1)
7251 return -EINVAL;
7252
7253 apply_wqattrs_lock();
7254 attrs = wq_sysfs_prep_attrs(wq);
7255 if (attrs) {
7256 attrs->affn_strict = (bool)v;
7257 ret = apply_workqueue_attrs_locked(wq, attrs);
7258 }
7259 apply_wqattrs_unlock();
7260 free_workqueue_attrs(attrs);
7261 return ret ?: count;
7262}
7263
7264static struct device_attribute wq_sysfs_unbound_attrs[] = {
7265 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7266 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7267 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
7268 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
7269 __ATTR_NULL,
7270};
7271
7272static const struct bus_type wq_subsys = {
7273 .name = "workqueue",
7274 .dev_groups = wq_sysfs_groups,
7275};
7276
7277/**
7278 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
7279 * @cpumask: the cpumask to set
7280 *
7281 * The low-level workqueues cpumask is a global cpumask that limits
7282 * the affinity of all unbound workqueues. This function check the @cpumask
7283 * and apply it to all unbound workqueues and updates all pwqs of them.
7284 *
7285 * Return: 0 - Success
7286 * -EINVAL - Invalid @cpumask
7287 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
7288 */
7289static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
7290{
7291 int ret = -EINVAL;
7292
7293 /*
7294 * Not excluding isolated cpus on purpose.
7295 * If the user wishes to include them, we allow that.
7296 */
7297 cpumask_and(dstp: cpumask, src1p: cpumask, cpu_possible_mask);
7298 if (!cpumask_empty(srcp: cpumask)) {
7299 ret = 0;
7300 apply_wqattrs_lock();
7301 if (!cpumask_equal(src1p: cpumask, src2p: wq_unbound_cpumask))
7302 ret = workqueue_apply_unbound_cpumask(unbound_cpumask: cpumask);
7303 if (!ret)
7304 cpumask_copy(dstp: wq_requested_unbound_cpumask, srcp: cpumask);
7305 apply_wqattrs_unlock();
7306 }
7307
7308 return ret;
7309}
7310
7311static ssize_t __wq_cpumask_show(struct device *dev,
7312 struct device_attribute *attr, char *buf, cpumask_var_t mask)
7313{
7314 int written;
7315
7316 mutex_lock(lock: &wq_pool_mutex);
7317 written = scnprintf(buf, PAGE_SIZE, fmt: "%*pb\n", cpumask_pr_args(mask));
7318 mutex_unlock(lock: &wq_pool_mutex);
7319
7320 return written;
7321}
7322
7323static ssize_t cpumask_requested_show(struct device *dev,
7324 struct device_attribute *attr, char *buf)
7325{
7326 return __wq_cpumask_show(dev, attr, buf, mask: wq_requested_unbound_cpumask);
7327}
7328static DEVICE_ATTR_RO(cpumask_requested);
7329
7330static ssize_t cpumask_isolated_show(struct device *dev,
7331 struct device_attribute *attr, char *buf)
7332{
7333 return __wq_cpumask_show(dev, attr, buf, mask: wq_isolated_cpumask);
7334}
7335static DEVICE_ATTR_RO(cpumask_isolated);
7336
7337static ssize_t cpumask_show(struct device *dev,
7338 struct device_attribute *attr, char *buf)
7339{
7340 return __wq_cpumask_show(dev, attr, buf, mask: wq_unbound_cpumask);
7341}
7342
7343static ssize_t cpumask_store(struct device *dev,
7344 struct device_attribute *attr, const char *buf, size_t count)
7345{
7346 cpumask_var_t cpumask;
7347 int ret;
7348
7349 if (!zalloc_cpumask_var(mask: &cpumask, GFP_KERNEL))
7350 return -ENOMEM;
7351
7352 ret = cpumask_parse(buf, dstp: cpumask);
7353 if (!ret)
7354 ret = workqueue_set_unbound_cpumask(cpumask);
7355
7356 free_cpumask_var(mask: cpumask);
7357 return ret ? ret : count;
7358}
7359static DEVICE_ATTR_RW(cpumask);
7360
7361static struct attribute *wq_sysfs_cpumask_attrs[] = {
7362 &dev_attr_cpumask.attr,
7363 &dev_attr_cpumask_requested.attr,
7364 &dev_attr_cpumask_isolated.attr,
7365 NULL,
7366};
7367ATTRIBUTE_GROUPS(wq_sysfs_cpumask);
7368
7369static int __init wq_sysfs_init(void)
7370{
7371 return subsys_virtual_register(subsys: &wq_subsys, groups: wq_sysfs_cpumask_groups);
7372}
7373core_initcall(wq_sysfs_init);
7374
7375static void wq_device_release(struct device *dev)
7376{
7377 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7378
7379 kfree(objp: wq_dev);
7380}
7381
7382/**
7383 * workqueue_sysfs_register - make a workqueue visible in sysfs
7384 * @wq: the workqueue to register
7385 *
7386 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
7387 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7388 * which is the preferred method.
7389 *
7390 * Workqueue user should use this function directly iff it wants to apply
7391 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7392 * apply_workqueue_attrs() may race against userland updating the
7393 * attributes.
7394 *
7395 * Return: 0 on success, -errno on failure.
7396 */
7397int workqueue_sysfs_register(struct workqueue_struct *wq)
7398{
7399 struct wq_device *wq_dev;
7400 int ret;
7401
7402 /*
7403 * Adjusting max_active breaks ordering guarantee. Disallow exposing
7404 * ordered workqueues.
7405 */
7406 if (WARN_ON(wq->flags & __WQ_ORDERED))
7407 return -EINVAL;
7408
7409 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7410 if (!wq_dev)
7411 return -ENOMEM;
7412
7413 wq_dev->wq = wq;
7414 wq_dev->dev.bus = &wq_subsys;
7415 wq_dev->dev.release = wq_device_release;
7416 dev_set_name(dev: &wq_dev->dev, name: "%s", wq->name);
7417
7418 /*
7419 * unbound_attrs are created separately. Suppress uevent until
7420 * everything is ready.
7421 */
7422 dev_set_uevent_suppress(dev: &wq_dev->dev, val: true);
7423
7424 ret = device_register(dev: &wq_dev->dev);
7425 if (ret) {
7426 put_device(dev: &wq_dev->dev);
7427 wq->wq_dev = NULL;
7428 return ret;
7429 }
7430
7431 if (wq->flags & WQ_UNBOUND) {
7432 struct device_attribute *attr;
7433
7434 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7435 ret = device_create_file(device: &wq_dev->dev, entry: attr);
7436 if (ret) {
7437 device_unregister(dev: &wq_dev->dev);
7438 wq->wq_dev = NULL;
7439 return ret;
7440 }
7441 }
7442 }
7443
7444 dev_set_uevent_suppress(dev: &wq_dev->dev, val: false);
7445 kobject_uevent(kobj: &wq_dev->dev.kobj, action: KOBJ_ADD);
7446 return 0;
7447}
7448
7449/**
7450 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7451 * @wq: the workqueue to unregister
7452 *
7453 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7454 */
7455static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7456{
7457 struct wq_device *wq_dev = wq->wq_dev;
7458
7459 if (!wq->wq_dev)
7460 return;
7461
7462 wq->wq_dev = NULL;
7463 device_unregister(dev: &wq_dev->dev);
7464}
7465#else /* CONFIG_SYSFS */
7466static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7467#endif /* CONFIG_SYSFS */
7468
7469/*
7470 * Workqueue watchdog.
7471 *
7472 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7473 * flush dependency, a concurrency managed work item which stays RUNNING
7474 * indefinitely. Workqueue stalls can be very difficult to debug as the
7475 * usual warning mechanisms don't trigger and internal workqueue state is
7476 * largely opaque.
7477 *
7478 * Workqueue watchdog monitors all worker pools periodically and dumps
7479 * state if some pools failed to make forward progress for a while where
7480 * forward progress is defined as the first item on ->worklist changing.
7481 *
7482 * This mechanism is controlled through the kernel parameter
7483 * "workqueue.watchdog_thresh" which can be updated at runtime through the
7484 * corresponding sysfs parameter file.
7485 */
7486#ifdef CONFIG_WQ_WATCHDOG
7487
7488static unsigned long wq_watchdog_thresh = 30;
7489static struct timer_list wq_watchdog_timer;
7490
7491static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7492static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7493
7494static unsigned int wq_panic_on_stall;
7495module_param_named(panic_on_stall, wq_panic_on_stall, uint, 0644);
7496
7497/*
7498 * Show workers that might prevent the processing of pending work items.
7499 * The only candidates are CPU-bound workers in the running state.
7500 * Pending work items should be handled by another idle worker
7501 * in all other situations.
7502 */
7503static void show_cpu_pool_hog(struct worker_pool *pool)
7504{
7505 struct worker *worker;
7506 unsigned long irq_flags;
7507 int bkt;
7508
7509 raw_spin_lock_irqsave(&pool->lock, irq_flags);
7510
7511 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
7512 if (task_is_running(worker->task)) {
7513 /*
7514 * Defer printing to avoid deadlocks in console
7515 * drivers that queue work while holding locks
7516 * also taken in their write paths.
7517 */
7518 printk_deferred_enter();
7519
7520 pr_info("pool %d:\n", pool->id);
7521 sched_show_task(worker->task);
7522
7523 printk_deferred_exit();
7524 }
7525 }
7526
7527 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
7528}
7529
7530static void show_cpu_pools_hogs(void)
7531{
7532 struct worker_pool *pool;
7533 int pi;
7534
7535 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
7536
7537 rcu_read_lock();
7538
7539 for_each_pool(pool, pi) {
7540 if (pool->cpu_stall)
7541 show_cpu_pool_hog(pool);
7542
7543 }
7544
7545 rcu_read_unlock();
7546}
7547
7548static void panic_on_wq_watchdog(void)
7549{
7550 static unsigned int wq_stall;
7551
7552 if (wq_panic_on_stall) {
7553 wq_stall++;
7554 BUG_ON(wq_stall >= wq_panic_on_stall);
7555 }
7556}
7557
7558static void wq_watchdog_reset_touched(void)
7559{
7560 int cpu;
7561
7562 wq_watchdog_touched = jiffies;
7563 for_each_possible_cpu(cpu)
7564 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7565}
7566
7567static void wq_watchdog_timer_fn(struct timer_list *unused)
7568{
7569 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7570 bool lockup_detected = false;
7571 bool cpu_pool_stall = false;
7572 unsigned long now = jiffies;
7573 struct worker_pool *pool;
7574 int pi;
7575
7576 if (!thresh)
7577 return;
7578
7579 for_each_pool(pool, pi) {
7580 unsigned long pool_ts, touched, ts;
7581
7582 pool->cpu_stall = false;
7583 if (list_empty(&pool->worklist))
7584 continue;
7585
7586 /*
7587 * If a virtual machine is stopped by the host it can look to
7588 * the watchdog like a stall.
7589 */
7590 kvm_check_and_clear_guest_paused();
7591
7592 /* get the latest of pool and touched timestamps */
7593 if (pool->cpu >= 0)
7594 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7595 else
7596 touched = READ_ONCE(wq_watchdog_touched);
7597 pool_ts = READ_ONCE(pool->watchdog_ts);
7598
7599 if (time_after(pool_ts, touched))
7600 ts = pool_ts;
7601 else
7602 ts = touched;
7603
7604 /* did we stall? */
7605 if (time_after(now, ts + thresh)) {
7606 lockup_detected = true;
7607 if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) {
7608 pool->cpu_stall = true;
7609 cpu_pool_stall = true;
7610 }
7611 pr_emerg("BUG: workqueue lockup - pool");
7612 pr_cont_pool_info(pool);
7613 pr_cont(" stuck for %us!\n",
7614 jiffies_to_msecs(now - pool_ts) / 1000);
7615 }
7616
7617
7618 }
7619
7620 if (lockup_detected)
7621 show_all_workqueues();
7622
7623 if (cpu_pool_stall)
7624 show_cpu_pools_hogs();
7625
7626 if (lockup_detected)
7627 panic_on_wq_watchdog();
7628
7629 wq_watchdog_reset_touched();
7630 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7631}
7632
7633notrace void wq_watchdog_touch(int cpu)
7634{
7635 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7636 unsigned long touch_ts = READ_ONCE(wq_watchdog_touched);
7637 unsigned long now = jiffies;
7638
7639 if (cpu >= 0)
7640 per_cpu(wq_watchdog_touched_cpu, cpu) = now;
7641 else
7642 WARN_ONCE(1, "%s should be called with valid CPU", __func__);
7643
7644 /* Don't unnecessarily store to global cacheline */
7645 if (time_after(now, touch_ts + thresh / 4))
7646 WRITE_ONCE(wq_watchdog_touched, jiffies);
7647}
7648
7649static void wq_watchdog_set_thresh(unsigned long thresh)
7650{
7651 wq_watchdog_thresh = 0;
7652 timer_delete_sync(&wq_watchdog_timer);
7653
7654 if (thresh) {
7655 wq_watchdog_thresh = thresh;
7656 wq_watchdog_reset_touched();
7657 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7658 }
7659}
7660
7661static int wq_watchdog_param_set_thresh(const char *val,
7662 const struct kernel_param *kp)
7663{
7664 unsigned long thresh;
7665 int ret;
7666
7667 ret = kstrtoul(val, 0, &thresh);
7668 if (ret)
7669 return ret;
7670
7671 if (system_percpu_wq)
7672 wq_watchdog_set_thresh(thresh);
7673 else
7674 wq_watchdog_thresh = thresh;
7675
7676 return 0;
7677}
7678
7679static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7680 .set = wq_watchdog_param_set_thresh,
7681 .get = param_get_ulong,
7682};
7683
7684module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7685 0644);
7686
7687static void wq_watchdog_init(void)
7688{
7689 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7690 wq_watchdog_set_thresh(wq_watchdog_thresh);
7691}
7692
7693#else /* CONFIG_WQ_WATCHDOG */
7694
7695static inline void wq_watchdog_init(void) { }
7696
7697#endif /* CONFIG_WQ_WATCHDOG */
7698
7699static void bh_pool_kick_normal(struct irq_work *irq_work)
7700{
7701 raise_softirq_irqoff(nr: TASKLET_SOFTIRQ);
7702}
7703
7704static void bh_pool_kick_highpri(struct irq_work *irq_work)
7705{
7706 raise_softirq_irqoff(nr: HI_SOFTIRQ);
7707}
7708
7709static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
7710{
7711 if (!cpumask_intersects(src1p: wq_unbound_cpumask, src2p: mask)) {
7712 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
7713 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7714 return;
7715 }
7716
7717 cpumask_and(dstp: wq_unbound_cpumask, src1p: wq_unbound_cpumask, src2p: mask);
7718}
7719
7720static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice)
7721{
7722 BUG_ON(init_worker_pool(pool));
7723 pool->cpu = cpu;
7724 cpumask_copy(dstp: pool->attrs->cpumask, cpumask_of(cpu));
7725 cpumask_copy(dstp: pool->attrs->__pod_cpumask, cpumask_of(cpu));
7726 pool->attrs->nice = nice;
7727 pool->attrs->affn_strict = true;
7728 pool->node = cpu_to_node(cpu);
7729
7730 /* alloc pool ID */
7731 mutex_lock(lock: &wq_pool_mutex);
7732 BUG_ON(worker_pool_assign_id(pool));
7733 mutex_unlock(lock: &wq_pool_mutex);
7734}
7735
7736/**
7737 * workqueue_init_early - early init for workqueue subsystem
7738 *
7739 * This is the first step of three-staged workqueue subsystem initialization and
7740 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7741 * up. It sets up all the data structures and system workqueues and allows early
7742 * boot code to create workqueues and queue/cancel work items. Actual work item
7743 * execution starts only after kthreads can be created and scheduled right
7744 * before early initcalls.
7745 */
7746void __init workqueue_init_early(void)
7747{
7748 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7749 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7750 void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal,
7751 bh_pool_kick_highpri };
7752 int i, cpu;
7753
7754 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7755
7756 BUG_ON(!alloc_cpumask_var(&wq_online_cpumask, GFP_KERNEL));
7757 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7758 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
7759 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
7760
7761 cpumask_copy(dstp: wq_online_cpumask, cpu_online_mask);
7762 cpumask_copy(dstp: wq_unbound_cpumask, cpu_possible_mask);
7763 restrict_unbound_cpumask(name: "HK_TYPE_WQ", mask: housekeeping_cpumask(type: HK_TYPE_WQ));
7764 restrict_unbound_cpumask(name: "HK_TYPE_DOMAIN", mask: housekeeping_cpumask(type: HK_TYPE_DOMAIN));
7765 if (!cpumask_empty(srcp: &wq_cmdline_cpumask))
7766 restrict_unbound_cpumask(name: "workqueue.unbound_cpus", mask: &wq_cmdline_cpumask);
7767
7768 cpumask_copy(dstp: wq_requested_unbound_cpumask, srcp: wq_unbound_cpumask);
7769 cpumask_andnot(dstp: wq_isolated_cpumask, cpu_possible_mask,
7770 src2p: housekeeping_cpumask(type: HK_TYPE_DOMAIN));
7771 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7772
7773 unbound_wq_update_pwq_attrs_buf = alloc_workqueue_attrs();
7774 BUG_ON(!unbound_wq_update_pwq_attrs_buf);
7775
7776 /*
7777 * If nohz_full is enabled, set power efficient workqueue as unbound.
7778 * This allows workqueue items to be moved to HK CPUs.
7779 */
7780 if (housekeeping_enabled(type: HK_TYPE_TICK))
7781 wq_power_efficient = true;
7782
7783 /* initialize WQ_AFFN_SYSTEM pods */
7784 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7785 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7786 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7787 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
7788
7789 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7790
7791 pt->nr_pods = 1;
7792 cpumask_copy(dstp: pt->pod_cpus[0], cpu_possible_mask);
7793 pt->pod_node[0] = NUMA_NO_NODE;
7794 pt->cpu_pod[0] = 0;
7795
7796 /* initialize BH and CPU pools */
7797 for_each_possible_cpu(cpu) {
7798 struct worker_pool *pool;
7799
7800 i = 0;
7801 for_each_bh_worker_pool(pool, cpu) {
7802 init_cpu_worker_pool(pool, cpu, nice: std_nice[i]);
7803 pool->flags |= POOL_BH;
7804 init_irq_work(work: bh_pool_irq_work(pool), func: irq_work_fns[i]);
7805 i++;
7806 }
7807
7808 i = 0;
7809 for_each_cpu_worker_pool(pool, cpu)
7810 init_cpu_worker_pool(pool, cpu, nice: std_nice[i++]);
7811 }
7812
7813 /* create default unbound and ordered wq attrs */
7814 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7815 struct workqueue_attrs *attrs;
7816
7817 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7818 attrs->nice = std_nice[i];
7819 unbound_std_wq_attrs[i] = attrs;
7820
7821 /*
7822 * An ordered wq should have only one pwq as ordering is
7823 * guaranteed by max_active which is enforced by pwqs.
7824 */
7825 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7826 attrs->nice = std_nice[i];
7827 attrs->ordered = true;
7828 ordered_wq_attrs[i] = attrs;
7829 }
7830
7831 system_wq = alloc_workqueue("events", WQ_PERCPU, 0);
7832 system_percpu_wq = alloc_workqueue("events", WQ_PERCPU, 0);
7833 system_highpri_wq = alloc_workqueue("events_highpri",
7834 WQ_HIGHPRI | WQ_PERCPU, 0);
7835 system_long_wq = alloc_workqueue("events_long", WQ_PERCPU, 0);
7836 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, WQ_MAX_ACTIVE);
7837 system_dfl_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, WQ_MAX_ACTIVE);
7838 system_freezable_wq = alloc_workqueue("events_freezable",
7839 WQ_FREEZABLE | WQ_PERCPU, 0);
7840 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7841 WQ_POWER_EFFICIENT | WQ_PERCPU, 0);
7842 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
7843 WQ_FREEZABLE | WQ_POWER_EFFICIENT | WQ_PERCPU, 0);
7844 system_bh_wq = alloc_workqueue("events_bh", WQ_BH | WQ_PERCPU, 0);
7845 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri",
7846 WQ_BH | WQ_HIGHPRI | WQ_PERCPU, 0);
7847 BUG_ON(!system_wq || !system_percpu_wq|| !system_highpri_wq || !system_long_wq ||
7848 !system_unbound_wq || !system_freezable_wq || !system_dfl_wq ||
7849 !system_power_efficient_wq ||
7850 !system_freezable_power_efficient_wq ||
7851 !system_bh_wq || !system_bh_highpri_wq);
7852}
7853
7854static void __init wq_cpu_intensive_thresh_init(void)
7855{
7856 unsigned long thresh;
7857 unsigned long bogo;
7858
7859 pwq_release_worker = kthread_run_worker(0, "pool_workqueue_release");
7860 BUG_ON(IS_ERR(pwq_release_worker));
7861
7862 /* if the user set it to a specific value, keep it */
7863 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7864 return;
7865
7866 /*
7867 * The default of 10ms is derived from the fact that most modern (as of
7868 * 2023) processors can do a lot in 10ms and that it's just below what
7869 * most consider human-perceivable. However, the kernel also runs on a
7870 * lot slower CPUs including microcontrollers where the threshold is way
7871 * too low.
7872 *
7873 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7874 * This is by no means accurate but it doesn't have to be. The mechanism
7875 * is still useful even when the threshold is fully scaled up. Also, as
7876 * the reports would usually be applicable to everyone, some machines
7877 * operating on longer thresholds won't significantly diminish their
7878 * usefulness.
7879 */
7880 thresh = 10 * USEC_PER_MSEC;
7881
7882 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
7883 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7884 if (bogo < 4000)
7885 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7886
7887 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7888 loops_per_jiffy, bogo, thresh);
7889
7890 wq_cpu_intensive_thresh_us = thresh;
7891}
7892
7893/**
7894 * workqueue_init - bring workqueue subsystem fully online
7895 *
7896 * This is the second step of three-staged workqueue subsystem initialization
7897 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
7898 * been created and work items queued on them, but there are no kworkers
7899 * executing the work items yet. Populate the worker pools with the initial
7900 * workers and enable future kworker creations.
7901 */
7902void __init workqueue_init(void)
7903{
7904 struct workqueue_struct *wq;
7905 struct worker_pool *pool;
7906 int cpu, bkt;
7907
7908 wq_cpu_intensive_thresh_init();
7909
7910 mutex_lock(lock: &wq_pool_mutex);
7911
7912 /*
7913 * Per-cpu pools created earlier could be missing node hint. Fix them
7914 * up. Also, create a rescuer for workqueues that requested it.
7915 */
7916 for_each_possible_cpu(cpu) {
7917 for_each_bh_worker_pool(pool, cpu)
7918 pool->node = cpu_to_node(cpu);
7919 for_each_cpu_worker_pool(pool, cpu)
7920 pool->node = cpu_to_node(cpu);
7921 }
7922
7923 list_for_each_entry(wq, &workqueues, list) {
7924 WARN(init_rescuer(wq),
7925 "workqueue: failed to create early rescuer for %s",
7926 wq->name);
7927 }
7928
7929 mutex_unlock(lock: &wq_pool_mutex);
7930
7931 /*
7932 * Create the initial workers. A BH pool has one pseudo worker that
7933 * represents the shared BH execution context and thus doesn't get
7934 * affected by hotplug events. Create the BH pseudo workers for all
7935 * possible CPUs here.
7936 */
7937 for_each_possible_cpu(cpu)
7938 for_each_bh_worker_pool(pool, cpu)
7939 BUG_ON(!create_worker(pool));
7940
7941 for_each_online_cpu(cpu) {
7942 for_each_cpu_worker_pool(pool, cpu) {
7943 pool->flags &= ~POOL_DISASSOCIATED;
7944 BUG_ON(!create_worker(pool));
7945 }
7946 }
7947
7948 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7949 BUG_ON(!create_worker(pool));
7950
7951 wq_online = true;
7952 wq_watchdog_init();
7953}
7954
7955/*
7956 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7957 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7958 * and consecutive pod ID. The rest of @pt is initialized accordingly.
7959 */
7960static void __init init_pod_type(struct wq_pod_type *pt,
7961 bool (*cpus_share_pod)(int, int))
7962{
7963 int cur, pre, cpu, pod;
7964
7965 pt->nr_pods = 0;
7966
7967 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
7968 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7969 BUG_ON(!pt->cpu_pod);
7970
7971 for_each_possible_cpu(cur) {
7972 for_each_possible_cpu(pre) {
7973 if (pre >= cur) {
7974 pt->cpu_pod[cur] = pt->nr_pods++;
7975 break;
7976 }
7977 if (cpus_share_pod(cur, pre)) {
7978 pt->cpu_pod[cur] = pt->cpu_pod[pre];
7979 break;
7980 }
7981 }
7982 }
7983
7984 /* init the rest to match @pt->cpu_pod[] */
7985 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7986 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7987 BUG_ON(!pt->pod_cpus || !pt->pod_node);
7988
7989 for (pod = 0; pod < pt->nr_pods; pod++)
7990 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7991
7992 for_each_possible_cpu(cpu) {
7993 cpumask_set_cpu(cpu, dstp: pt->pod_cpus[pt->cpu_pod[cpu]]);
7994 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7995 }
7996}
7997
7998static bool __init cpus_dont_share(int cpu0, int cpu1)
7999{
8000 return false;
8001}
8002
8003static bool __init cpus_share_smt(int cpu0, int cpu1)
8004{
8005#ifdef CONFIG_SCHED_SMT
8006 return cpumask_test_cpu(cpu: cpu0, cpumask: cpu_smt_mask(cpu: cpu1));
8007#else
8008 return false;
8009#endif
8010}
8011
8012static bool __init cpus_share_numa(int cpu0, int cpu1)
8013{
8014 return cpu_to_node(cpu: cpu0) == cpu_to_node(cpu: cpu1);
8015}
8016
8017/**
8018 * workqueue_init_topology - initialize CPU pods for unbound workqueues
8019 *
8020 * This is the third step of three-staged workqueue subsystem initialization and
8021 * invoked after SMP and topology information are fully initialized. It
8022 * initializes the unbound CPU pods accordingly.
8023 */
8024void __init workqueue_init_topology(void)
8025{
8026 struct workqueue_struct *wq;
8027 int cpu;
8028
8029 init_pod_type(pt: &wq_pod_types[WQ_AFFN_CPU], cpus_share_pod: cpus_dont_share);
8030 init_pod_type(pt: &wq_pod_types[WQ_AFFN_SMT], cpus_share_pod: cpus_share_smt);
8031 init_pod_type(pt: &wq_pod_types[WQ_AFFN_CACHE], cpus_share_pod: cpus_share_cache);
8032 init_pod_type(pt: &wq_pod_types[WQ_AFFN_NUMA], cpus_share_pod: cpus_share_numa);
8033
8034 wq_topo_initialized = true;
8035
8036 mutex_lock(lock: &wq_pool_mutex);
8037
8038 /*
8039 * Workqueues allocated earlier would have all CPUs sharing the default
8040 * worker pool. Explicitly call unbound_wq_update_pwq() on all workqueue
8041 * and CPU combinations to apply per-pod sharing.
8042 */
8043 list_for_each_entry(wq, &workqueues, list) {
8044 for_each_online_cpu(cpu)
8045 unbound_wq_update_pwq(wq, cpu);
8046 if (wq->flags & WQ_UNBOUND) {
8047 mutex_lock(lock: &wq->mutex);
8048 wq_update_node_max_active(wq, off_cpu: -1);
8049 mutex_unlock(lock: &wq->mutex);
8050 }
8051 }
8052
8053 mutex_unlock(lock: &wq_pool_mutex);
8054}
8055
8056void __warn_flushing_systemwide_wq(void)
8057{
8058 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
8059 dump_stack();
8060}
8061EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
8062
8063static int __init workqueue_unbound_cpus_setup(char *str)
8064{
8065 if (cpulist_parse(buf: str, dstp: &wq_cmdline_cpumask) < 0) {
8066 cpumask_clear(dstp: &wq_cmdline_cpumask);
8067 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
8068 }
8069
8070 return 1;
8071}
8072__setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);
8073