projects/linux·blob master · refs

8644 lines · 245.6 KiB

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 <[email protected]>
9 * Andrew Morton
10 * Kai Petzke <[email protected]>
11 * Theodore Ts'o <[email protected]>
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 <[email protected]>
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/device/devres.h>
45#include <linux/lockdep.h>
46#include <linux/idr.h>
47#include <linux/jhash.h>
48#include <linux/hashtable.h>
49#include <linux/rculist.h>
50#include <linux/nodemask.h>
51#include <linux/moduleparam.h>
52#include <linux/uaccess.h>
53#include <linux/sched/isolation.h>
54#include <linux/sched/debug.h>
55#include <linux/nmi.h>
56#include <linux/kvm_para.h>
57#include <linux/delay.h>
58#include <linux/irq_work.h>
59
60#include "workqueue_internal.h"
61
62enum worker_pool_flags {
63 /*
64 * worker_pool flags
65 *
66 * A bound pool is either associated or disassociated with its CPU.
67 * While associated (!DISASSOCIATED), all workers are bound to the
68 * CPU and none has %WORKER_UNBOUND set and concurrency management
69 * is in effect.
70 *
71 * While DISASSOCIATED, the cpu may be offline and all workers have
72 * %WORKER_UNBOUND set and concurrency management disabled, and may
73 * be executing on any CPU. The pool behaves as an unbound one.
74 *
75 * Note that DISASSOCIATED should be flipped only while holding
76 * wq_pool_attach_mutex to avoid changing binding state while
77 * worker_attach_to_pool() is in progress.
78 *
79 * As there can only be one concurrent BH execution context per CPU, a
80 * BH pool is per-CPU and always DISASSOCIATED.
81 */
82 POOL_BH = 1 << 0, /* is a BH pool */
83 POOL_MANAGER_ACTIVE = 1 << 1, /* being managed */
84 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
85 POOL_BH_DRAINING = 1 << 3, /* draining after CPU offline */
86};
87
88enum worker_flags {
89 /* worker flags */
90 WORKER_DIE = 1 << 1, /* die die die */
91 WORKER_IDLE = 1 << 2, /* is idle */
92 WORKER_PREP = 1 << 3, /* preparing to run works */
93 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
94 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
95 WORKER_REBOUND = 1 << 8, /* worker was rebound */
96
97 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
98 WORKER_UNBOUND | WORKER_REBOUND,
99};
100
101enum work_cancel_flags {
102 WORK_CANCEL_DELAYED = 1 << 0, /* canceling a delayed_work */
103 WORK_CANCEL_DISABLE = 1 << 1, /* canceling to disable */
104};
105
106enum wq_internal_consts {
107 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
108
109 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
110 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
111
112 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
113 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
114
115 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
116 /* call for help after 10ms
117 (min two ticks) */
118 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
119 CREATE_COOLDOWN = HZ, /* time to breath after fail */
120
121 RESCUER_BATCH = 16, /* process items per turn */
122
123 /*
124 * Rescue workers are used only on emergencies and shared by
125 * all cpus. Give MIN_NICE.
126 */
127 RESCUER_NICE_LEVEL = MIN_NICE,
128 HIGHPRI_NICE_LEVEL = MIN_NICE,
129
130 WQ_NAME_LEN = 32,
131 WORKER_ID_LEN = 10 + WQ_NAME_LEN, /* "kworker/R-" + WQ_NAME_LEN */
132};
133
134/* Layout of shards within one LLC pod */
135struct llc_shard_layout {
136 int nr_large_shards; /* number of large shards (cores_per_shard + 1) */
137 int cores_per_shard; /* base number of cores per default shard */
138 int nr_shards; /* total number of shards */
139 /* nr_default shards = (nr_shards - nr_large_shards) */
140};
141
142/*
143 * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and
144 * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because
145 * msecs_to_jiffies() can't be an initializer.
146 */
147#define BH_WORKER_JIFFIES msecs_to_jiffies(2)
148#define BH_WORKER_RESTARTS 10
149
150/*
151 * Structure fields follow one of the following exclusion rules.
152 *
153 * I: Modifiable by initialization/destruction paths and read-only for
154 * everyone else.
155 *
156 * P: Preemption protected. Disabling preemption is enough and should
157 * only be modified and accessed from the local cpu.
158 *
159 * L: pool->lock protected. Access with pool->lock held.
160 *
161 * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for
162 * reads.
163 *
164 * K: Only modified by worker while holding pool->lock. Can be safely read by
165 * self, while holding pool->lock or from IRQ context if %current is the
166 * kworker.
167 *
168 * S: Only modified by worker self.
169 *
170 * A: wq_pool_attach_mutex protected.
171 *
172 * PL: wq_pool_mutex protected.
173 *
174 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
175 *
176 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
177 *
178 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
179 * RCU for reads.
180 *
181 * WQ: wq->mutex protected.
182 *
183 * WR: wq->mutex protected for writes. RCU protected for reads.
184 *
185 * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read
186 * with READ_ONCE() without locking.
187 *
188 * MD: wq_mayday_lock protected.
189 *
190 * WD: Used internally by the watchdog.
191 */
192
193/* struct worker is defined in workqueue_internal.h */
194
195struct worker_pool {
196 raw_spinlock_t lock; /* the pool lock */
197 int cpu; /* I: the associated cpu */
198 int node; /* I: the associated node ID */
199 int id; /* I: pool ID */
200 unsigned int flags; /* L: flags */
201
202 unsigned long last_progress_ts; /* L: last forward progress timestamp */
203 bool cpu_stall; /* WD: stalled cpu bound pool */
204
205 /*
206 * The counter is incremented in a process context on the associated CPU
207 * w/ preemption disabled, and decremented or reset in the same context
208 * but w/ pool->lock held. The readers grab pool->lock and are
209 * guaranteed to see if the counter reached zero.
210 */
211 int nr_running;
212
213 struct list_head worklist; /* L: list of pending works */
214
215 int nr_workers; /* L: total number of workers */
216 int nr_idle; /* L: currently idle workers */
217
218 struct list_head idle_list; /* L: list of idle workers */
219 struct timer_list idle_timer; /* L: worker idle timeout */
220 struct work_struct idle_cull_work; /* L: worker idle cleanup */
221
222 struct timer_list mayday_timer; /* L: SOS timer for workers */
223
224 /* a workers is either on busy_hash or idle_list, or the manager */
225 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
226 /* L: hash of busy workers */
227
228 struct worker *manager; /* L: purely informational */
229 /* L: last worker woken by kick_pool() */
230 struct worker *last_woken_worker;
231 struct list_head workers; /* A: attached workers */
232
233 struct ida worker_ida; /* worker IDs for task name */
234
235 struct workqueue_attrs *attrs; /* I: worker attributes */
236 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
237 int refcnt; /* PL: refcnt for unbound pools */
238#ifdef CONFIG_PREEMPT_RT
239 spinlock_t cb_lock; /* BH worker cancel lock */
240#endif
241 /*
242 * Destruction of pool is RCU protected to allow dereferences
243 * from get_work_pool().
244 */
245 struct rcu_head rcu;
246};
247
248/*
249 * Per-pool_workqueue statistics. These can be monitored using
250 * tools/workqueue/wq_monitor.py.
251 */
252enum pool_workqueue_stats {
253 PWQ_STAT_STARTED, /* work items started execution */
254 PWQ_STAT_COMPLETED, /* work items completed execution */
255 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
256 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
257 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
258 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
259 PWQ_STAT_MAYDAY, /* maydays to rescuer */
260 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
261
262 PWQ_NR_STATS,
263};
264
265/*
266 * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT
267 * of work_struct->data are used for flags and the remaining high bits
268 * point to the pwq; thus, pwqs need to be aligned at two's power of the
269 * number of flag bits.
270 */
271struct pool_workqueue {
272 struct worker_pool *pool; /* I: the associated pool */
273 struct workqueue_struct *wq; /* I: the owning workqueue */
274 int work_color; /* L: current color */
275 int flush_color; /* L: flushing color */
276 int refcnt; /* L: reference count */
277 int nr_in_flight[WORK_NR_COLORS];
278 /* L: nr of in_flight works */
279 bool plugged; /* L: execution suspended */
280
281 /*
282 * nr_active management and WORK_STRUCT_INACTIVE:
283 *
284 * When pwq->nr_active >= max_active, new work item is queued to
285 * pwq->inactive_works instead of pool->worklist and marked with
286 * WORK_STRUCT_INACTIVE.
287 *
288 * All work items marked with WORK_STRUCT_INACTIVE do not participate in
289 * nr_active and all work items in pwq->inactive_works are marked with
290 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are
291 * in pwq->inactive_works. Some of them are ready to run in
292 * pool->worklist or worker->scheduled. Those work itmes are only struct
293 * wq_barrier which is used for flush_work() and should not participate
294 * in nr_active. For non-barrier work item, it is marked with
295 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
296 */
297 int nr_active; /* L: nr of active works */
298 struct list_head inactive_works; /* L: inactive works */
299 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */
300 struct list_head pwqs_node; /* WR: node on wq->pwqs */
301 struct list_head mayday_node; /* MD: node on wq->maydays */
302 struct work_struct mayday_cursor; /* L: cursor on pool->worklist */
303
304 u64 stats[PWQ_NR_STATS];
305
306 /*
307 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
308 * and pwq_release_workfn() for details. pool_workqueue itself is also
309 * RCU protected so that the first pwq can be determined without
310 * grabbing wq->mutex.
311 */
312 struct kthread_work release_work;
313 struct rcu_head rcu;
314} __aligned(1 << WORK_STRUCT_PWQ_SHIFT);
315
316/*
317 * Structure used to wait for workqueue flush.
318 */
319struct wq_flusher {
320 struct list_head list; /* WQ: list of flushers */
321 int flush_color; /* WQ: flush color waiting for */
322 struct completion done; /* flush completion */
323};
324
325struct wq_device;
326
327/*
328 * Unlike in a per-cpu workqueue where max_active limits its concurrency level
329 * on each CPU, in an unbound workqueue, max_active applies to the whole system.
330 * As sharing a single nr_active across multiple sockets can be very expensive,
331 * the counting and enforcement is per NUMA node.
332 *
333 * The following struct is used to enforce per-node max_active. When a pwq wants
334 * to start executing a work item, it should increment ->nr using
335 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over
336 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish
337 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in
338 * round-robin order.
339 */
340struct wq_node_nr_active {
341 int max; /* per-node max_active */
342 atomic_t nr; /* per-node nr_active */
343 raw_spinlock_t lock; /* nests inside pool locks */
344 struct list_head pending_pwqs; /* LN: pwqs with inactive works */
345};
346
347/*
348 * The externally visible workqueue. It relays the issued work items to
349 * the appropriate worker_pool through its pool_workqueues.
350 */
351struct workqueue_struct {
352 struct list_head pwqs; /* WR: all pwqs of this wq */
353 struct list_head list; /* PR: list of all workqueues */
354
355 struct mutex mutex; /* protects this wq */
356 int work_color; /* WQ: current work color */
357 int flush_color; /* WQ: current flush color */
358 atomic_t nr_pwqs_to_flush; /* flush in progress */
359 struct wq_flusher *first_flusher; /* WQ: first flusher */
360 struct list_head flusher_queue; /* WQ: flush waiters */
361 struct list_head flusher_overflow; /* WQ: flush overflow list */
362
363 struct list_head maydays; /* MD: pwqs requesting rescue */
364 struct worker *rescuer; /* MD: rescue worker */
365
366 int nr_drainers; /* WQ: drain in progress */
367
368 /* See alloc_workqueue() function comment for info on min/max_active */
369 int max_active; /* WO: max active works */
370 int min_active; /* WO: min active works */
371 int saved_max_active; /* WQ: saved max_active */
372 int saved_min_active; /* WQ: saved min_active */
373
374 struct workqueue_attrs *attrs; /* PW: workqueue attributes */
375 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */
376
377#ifdef CONFIG_SYSFS
378 struct wq_device *wq_dev; /* I: for sysfs interface */
379#endif
380#ifdef CONFIG_LOCKDEP
381 char *lock_name;
382 struct lock_class_key key;
383 struct lockdep_map __lockdep_map;
384 struct lockdep_map *lockdep_map;
385#endif
386 char name[WQ_NAME_LEN]; /* I: workqueue name */
387
388 /*
389 * Destruction of workqueue_struct is RCU protected to allow walking
390 * the workqueues list without grabbing wq_pool_mutex.
391 * This is used to dump all workqueues from sysrq.
392 */
393 struct rcu_head rcu;
394
395 /* hot fields used during command issue, aligned to cacheline */
396 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
397 struct pool_workqueue __rcu * __percpu *cpu_pwq; /* I: per-cpu pwqs */
398 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */
399};
400
401/*
402 * Each pod type describes how CPUs should be grouped for unbound workqueues.
403 * See the comment above workqueue_attrs->affn_scope.
404 */
405struct wq_pod_type {
406 int nr_pods; /* number of pods */
407 cpumask_var_t *pod_cpus; /* pod -> cpus */
408 int *pod_node; /* pod -> node */
409 int *cpu_pod; /* cpu -> pod */
410};
411
412struct work_offq_data {
413 u32 pool_id;
414 u32 disable;
415 u32 flags;
416};
417
418static const char * const wq_affn_names[WQ_AFFN_NR_TYPES] = {
419 [WQ_AFFN_DFL] = "default",
420 [WQ_AFFN_CPU] = "cpu",
421 [WQ_AFFN_SMT] = "smt",
422 [WQ_AFFN_CACHE] = "cache",
423 [WQ_AFFN_CACHE_SHARD] = "cache_shard",
424 [WQ_AFFN_NUMA] = "numa",
425 [WQ_AFFN_SYSTEM] = "system",
426};
427
428/*
429 * Per-cpu work items which run for longer than the following threshold are
430 * automatically considered CPU intensive and excluded from concurrency
431 * management to prevent them from noticeably delaying other per-cpu work items.
432 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
433 * The actual value is initialized in wq_cpu_intensive_thresh_init().
434 */
435static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
436module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
437#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
438static unsigned int wq_cpu_intensive_warning_thresh = 4;
439module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644);
440#endif
441
442/* see the comment above the definition of WQ_POWER_EFFICIENT */
443static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
444module_param_named(power_efficient, wq_power_efficient, bool, 0444);
445
446static unsigned int wq_cache_shard_size = 8;
447module_param_named(cache_shard_size, wq_cache_shard_size, uint, 0444);
448
449static bool wq_online; /* can kworkers be created yet? */
450static bool wq_topo_initialized __read_mostly = false;
451
452static struct kmem_cache *pwq_cache;
453
454static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
455static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE_SHARD;
456
457/* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
458static struct workqueue_attrs *unbound_wq_update_pwq_attrs_buf;
459
460static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
461static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
462static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
463/* wait for manager to go away */
464static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
465
466static LIST_HEAD(workqueues); /* PR: list of all workqueues */
467static bool workqueue_freezing; /* PL: have wqs started freezing? */
468
469/* PL: mirror the cpu_online_mask excluding the CPU in the midst of hotplugging */
470static cpumask_var_t wq_online_cpumask;
471
472/* PL&A: allowable cpus for unbound wqs and work items */
473static cpumask_var_t wq_unbound_cpumask;
474
475/* PL: user requested unbound cpumask via sysfs */
476static cpumask_var_t wq_requested_unbound_cpumask;
477
478/* PL: isolated cpumask to be excluded from unbound cpumask */
479static cpumask_var_t wq_isolated_cpumask;
480
481/* for further constrain wq_unbound_cpumask by cmdline parameter*/
482static struct cpumask wq_cmdline_cpumask __initdata;
483
484/* CPU where unbound work was last round robin scheduled from this CPU */
485static DEFINE_PER_CPU(int, wq_rr_cpu_last);
486
487/*
488 * Local execution of unbound work items is no longer guaranteed. The
489 * following always forces round-robin CPU selection on unbound work items
490 * to uncover usages which depend on it.
491 */
492#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
493static bool wq_debug_force_rr_cpu = true;
494#else
495static bool wq_debug_force_rr_cpu = false;
496#endif
497module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
498
499/* to raise softirq for the BH worker pools on other CPUs */
500static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS], bh_pool_irq_works);
501
502/* the BH worker pools */
503static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], bh_worker_pools);
504
505/* the per-cpu worker pools */
506static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
507
508static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
509
510/* PL: hash of all unbound pools keyed by pool->attrs */
511static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
512
513/* I: attributes used when instantiating standard unbound pools on demand */
514static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
515
516/* I: attributes used when instantiating ordered pools on demand */
517static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
518
519/*
520 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
521 * process context while holding a pool lock. Bounce to a dedicated kthread
522 * worker to avoid A-A deadlocks.
523 */
524static struct kthread_worker *pwq_release_worker __ro_after_init;
525
526struct workqueue_struct *system_wq __ro_after_init;
527EXPORT_SYMBOL(system_wq);
528struct workqueue_struct *system_percpu_wq __ro_after_init;
529EXPORT_SYMBOL(system_percpu_wq);
530struct workqueue_struct *system_highpri_wq __ro_after_init;
531EXPORT_SYMBOL_GPL(system_highpri_wq);
532struct workqueue_struct *system_long_wq __ro_after_init;
533EXPORT_SYMBOL_GPL(system_long_wq);
534struct workqueue_struct *system_unbound_wq __ro_after_init;
535EXPORT_SYMBOL_GPL(system_unbound_wq);
536struct workqueue_struct *system_dfl_wq __ro_after_init;
537EXPORT_SYMBOL_GPL(system_dfl_wq);
538struct workqueue_struct *system_freezable_wq __ro_after_init;
539EXPORT_SYMBOL_GPL(system_freezable_wq);
540struct workqueue_struct *system_power_efficient_wq __ro_after_init;
541EXPORT_SYMBOL_GPL(system_power_efficient_wq);
542struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
543EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
544struct workqueue_struct *system_bh_wq;
545EXPORT_SYMBOL_GPL(system_bh_wq);
546struct workqueue_struct *system_bh_highpri_wq;
547EXPORT_SYMBOL_GPL(system_bh_highpri_wq);
548struct workqueue_struct *system_dfl_long_wq __ro_after_init;
549EXPORT_SYMBOL_GPL(system_dfl_long_wq);
550
551static int worker_thread(void *__worker);
552static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
553static void show_pwq(struct pool_workqueue *pwq);
554static void show_one_worker_pool(struct worker_pool *pool);
555
556#define CREATE_TRACE_POINTS
557#include <trace/events/workqueue.h>
558
559#define assert_rcu_or_pool_mutex() \
560 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
561 !lockdep_is_held(&wq_pool_mutex), \
562 "RCU or wq_pool_mutex should be held")
563
564#define for_each_bh_worker_pool(pool, cpu) \
565 for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \
566 (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
567 (pool)++)
568
569#define for_each_cpu_worker_pool(pool, cpu) \
570 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
571 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
572 (pool)++)
573
574/**
575 * for_each_pool - iterate through all worker_pools in the system
576 * @pool: iteration cursor
577 * @pi: integer used for iteration
578 *
579 * This must be called either with wq_pool_mutex held or RCU read
580 * locked. If the pool needs to be used beyond the locking in effect, the
581 * caller is responsible for guaranteeing that the pool stays online.
582 *
583 * The if/else clause exists only for the lockdep assertion and can be
584 * ignored.
585 */
586#define for_each_pool(pool, pi) \
587 idr_for_each_entry(&worker_pool_idr, pool, pi) \
588 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
589 else
590
591/**
592 * for_each_pool_worker - iterate through all workers of a worker_pool
593 * @worker: iteration cursor
594 * @pool: worker_pool to iterate workers of
595 *
596 * This must be called with wq_pool_attach_mutex.
597 *
598 * The if/else clause exists only for the lockdep assertion and can be
599 * ignored.
600 */
601#define for_each_pool_worker(worker, pool) \
602 list_for_each_entry((worker), &(pool)->workers, node) \
603 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
604 else
605
606/**
607 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
608 * @pwq: iteration cursor
609 * @wq: the target workqueue
610 *
611 * This must be called either with wq->mutex held or RCU read locked.
612 * If the pwq needs to be used beyond the locking in effect, the caller is
613 * responsible for guaranteeing that the pwq stays online.
614 *
615 * The if/else clause exists only for the lockdep assertion and can be
616 * ignored.
617 */
618#define for_each_pwq(pwq, wq) \
619 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
620 lockdep_is_held(&(wq->mutex)))
621
622#ifdef CONFIG_DEBUG_OBJECTS_WORK
623
624static const struct debug_obj_descr work_debug_descr;
625
626static void *work_debug_hint(void *addr)
627{
628 return ((struct work_struct *) addr)->func;
629}
630
631static bool work_is_static_object(void *addr)
632{
633 struct work_struct *work = addr;
634
635 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
636}
637
638/*
639 * fixup_init is called when:
640 * - an active object is initialized
641 */
642static bool work_fixup_init(void *addr, enum debug_obj_state state)
643{
644 struct work_struct *work = addr;
645
646 switch (state) {
647 case ODEBUG_STATE_ACTIVE:
648 cancel_work_sync(work);
649 debug_object_init(work, &work_debug_descr);
650 return true;
651 default:
652 return false;
653 }
654}
655
656/*
657 * fixup_free is called when:
658 * - an active object is freed
659 */
660static bool work_fixup_free(void *addr, enum debug_obj_state state)
661{
662 struct work_struct *work = addr;
663
664 switch (state) {
665 case ODEBUG_STATE_ACTIVE:
666 cancel_work_sync(work);
667 debug_object_free(work, &work_debug_descr);
668 return true;
669 default:
670 return false;
671 }
672}
673
674static const struct debug_obj_descr work_debug_descr = {
675 .name = "work_struct",
676 .debug_hint = work_debug_hint,
677 .is_static_object = work_is_static_object,
678 .fixup_init = work_fixup_init,
679 .fixup_free = work_fixup_free,
680};
681
682static inline void debug_work_activate(struct work_struct *work)
683{
684 debug_object_activate(work, &work_debug_descr);
685}
686
687static inline void debug_work_deactivate(struct work_struct *work)
688{
689 debug_object_deactivate(work, &work_debug_descr);
690}
691
692void __init_work(struct work_struct *work, int onstack)
693{
694 if (onstack)
695 debug_object_init_on_stack(work, &work_debug_descr);
696 else
697 debug_object_init(work, &work_debug_descr);
698}
699EXPORT_SYMBOL_GPL(__init_work);
700
701void destroy_work_on_stack(struct work_struct *work)
702{
703 debug_object_free(work, &work_debug_descr);
704}
705EXPORT_SYMBOL_GPL(destroy_work_on_stack);
706
707void destroy_delayed_work_on_stack(struct delayed_work *work)
708{
709 timer_destroy_on_stack(&work->timer);
710 debug_object_free(&work->work, &work_debug_descr);
711}
712EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
713
714#else
715static inline void debug_work_activate(struct work_struct *work) { }
716static inline void debug_work_deactivate(struct work_struct *work) { }
717#endif
718
719/**
720 * worker_pool_assign_id - allocate ID and assign it to @pool
721 * @pool: the pool pointer of interest
722 *
723 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
724 * successfully, -errno on failure.
725 */
726static int worker_pool_assign_id(struct worker_pool *pool)
727{
728 int ret;
729
730 lockdep_assert_held(&wq_pool_mutex);
731
732 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
733 GFP_KERNEL);
734 if (ret >= 0) {
735 pool->id = ret;
736 return 0;
737 }
738 return ret;
739}
740
741static struct pool_workqueue __rcu **
742unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
743{
744 if (cpu >= 0)
745 return per_cpu_ptr(wq->cpu_pwq, cpu);
746 else
747 return &wq->dfl_pwq;
748}
749
750/* @cpu < 0 for dfl_pwq */
751static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu)
752{
753 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
754 lockdep_is_held(&wq_pool_mutex) ||
755 lockdep_is_held(&wq->mutex));
756}
757
758/**
759 * unbound_effective_cpumask - effective cpumask of an unbound workqueue
760 * @wq: workqueue of interest
761 *
762 * @wq->attrs->cpumask contains the cpumask requested by the user which
763 * is masked with wq_unbound_cpumask to determine the effective cpumask. The
764 * default pwq is always mapped to the pool with the current effective cpumask.
765 */
766static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq)
767{
768 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask;
769}
770
771static unsigned int work_color_to_flags(int color)
772{
773 return color << WORK_STRUCT_COLOR_SHIFT;
774}
775
776static int get_work_color(unsigned long work_data)
777{
778 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
779 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
780}
781
782static int work_next_color(int color)
783{
784 return (color + 1) % WORK_NR_COLORS;
785}
786
787static unsigned long pool_offq_flags(struct worker_pool *pool)
788{
789 return (pool->flags & POOL_BH) ? WORK_OFFQ_BH : 0;
790}
791
792/*
793 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
794 * contain the pointer to the queued pwq. Once execution starts, the flag
795 * is cleared and the high bits contain OFFQ flags and pool ID.
796 *
797 * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling()
798 * can be used to set the pwq, pool or clear work->data. These functions should
799 * only be called while the work is owned - ie. while the PENDING bit is set.
800 *
801 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
802 * corresponding to a work. Pool is available once the work has been
803 * queued anywhere after initialization until it is sync canceled. pwq is
804 * available only while the work item is queued.
805 */
806static inline void set_work_data(struct work_struct *work, unsigned long data)
807{
808 WARN_ON_ONCE(!work_pending(work));
809 atomic_long_set(&work->data, data | work_static(work));
810}
811
812static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
813 unsigned long flags)
814{
815 set_work_data(work, (unsigned long)pwq | WORK_STRUCT_PENDING |
816 WORK_STRUCT_PWQ | flags);
817}
818
819static void set_work_pool_and_keep_pending(struct work_struct *work,
820 int pool_id, unsigned long flags)
821{
822 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
823 WORK_STRUCT_PENDING | flags);
824}
825
826static void set_work_pool_and_clear_pending(struct work_struct *work,
827 int pool_id, unsigned long flags)
828{
829 /*
830 * The following wmb is paired with the implied mb in
831 * test_and_set_bit(PENDING) and ensures all updates to @work made
832 * here are visible to and precede any updates by the next PENDING
833 * owner.
834 */
835 smp_wmb();
836 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
837 flags);
838 /*
839 * The following mb guarantees that previous clear of a PENDING bit
840 * will not be reordered with any speculative LOADS or STORES from
841 * work->current_func, which is executed afterwards. This possible
842 * reordering can lead to a missed execution on attempt to queue
843 * the same @work. E.g. consider this case:
844 *
845 * CPU#0 CPU#1
846 * ---------------------------- --------------------------------
847 *
848 * 1 STORE event_indicated
849 * 2 queue_work_on() {
850 * 3 test_and_set_bit(PENDING)
851 * 4 } set_..._and_clear_pending() {
852 * 5 set_work_data() # clear bit
853 * 6 smp_mb()
854 * 7 work->current_func() {
855 * 8 LOAD event_indicated
856 * }
857 *
858 * Without an explicit full barrier speculative LOAD on line 8 can
859 * be executed before CPU#0 does STORE on line 1. If that happens,
860 * CPU#0 observes the PENDING bit is still set and new execution of
861 * a @work is not queued in a hope, that CPU#1 will eventually
862 * finish the queued @work. Meanwhile CPU#1 does not see
863 * event_indicated is set, because speculative LOAD was executed
864 * before actual STORE.
865 */
866 smp_mb();
867}
868
869static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
870{
871 return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK);
872}
873
874static struct pool_workqueue *get_work_pwq(struct work_struct *work)
875{
876 unsigned long data = atomic_long_read(&work->data);
877
878 if (data & WORK_STRUCT_PWQ)
879 return work_struct_pwq(data);
880 else
881 return NULL;
882}
883
884/**
885 * get_work_pool - return the worker_pool a given work was associated with
886 * @work: the work item of interest
887 *
888 * Pools are created and destroyed under wq_pool_mutex, and allows read
889 * access under RCU read lock. As such, this function should be
890 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
891 *
892 * All fields of the returned pool are accessible as long as the above
893 * mentioned locking is in effect. If the returned pool needs to be used
894 * beyond the critical section, the caller is responsible for ensuring the
895 * returned pool is and stays online.
896 *
897 * Return: The worker_pool @work was last associated with. %NULL if none.
898 */
899static struct worker_pool *get_work_pool(struct work_struct *work)
900{
901 unsigned long data = atomic_long_read(&work->data);
902 int pool_id;
903
904 assert_rcu_or_pool_mutex();
905
906 if (data & WORK_STRUCT_PWQ)
907 return work_struct_pwq(data)->pool;
908
909 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
910 if (pool_id == WORK_OFFQ_POOL_NONE)
911 return NULL;
912
913 return idr_find(&worker_pool_idr, pool_id);
914}
915
916static unsigned long shift_and_mask(unsigned long v, u32 shift, u32 bits)
917{
918 return (v >> shift) & ((1U << bits) - 1);
919}
920
921static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data)
922{
923 WARN_ON_ONCE(data & WORK_STRUCT_PWQ);
924
925 offqd->pool_id = shift_and_mask(data, WORK_OFFQ_POOL_SHIFT,
926 WORK_OFFQ_POOL_BITS);
927 offqd->disable = shift_and_mask(data, WORK_OFFQ_DISABLE_SHIFT,
928 WORK_OFFQ_DISABLE_BITS);
929 offqd->flags = data & WORK_OFFQ_FLAG_MASK;
930}
931
932static unsigned long work_offqd_pack_flags(struct work_offq_data *offqd)
933{
934 return ((unsigned long)offqd->disable << WORK_OFFQ_DISABLE_SHIFT) |
935 ((unsigned long)offqd->flags);
936}
937
938/*
939 * Policy functions. These define the policies on how the global worker
940 * pools are managed. Unless noted otherwise, these functions assume that
941 * they're being called with pool->lock held.
942 */
943
944/*
945 * Need to wake up a worker? Called from anything but currently
946 * running workers.
947 *
948 * Note that, because unbound workers never contribute to nr_running, this
949 * function will always return %true for unbound pools as long as the
950 * worklist isn't empty.
951 */
952static bool need_more_worker(struct worker_pool *pool)
953{
954 return !list_empty(&pool->worklist) && !pool->nr_running;
955}
956
957/* Can I start working? Called from busy but !running workers. */
958static bool may_start_working(struct worker_pool *pool)
959{
960 return pool->nr_idle;
961}
962
963/* Do I need to keep working? Called from currently running workers. */
964static bool keep_working(struct worker_pool *pool)
965{
966 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
967}
968
969/* Do we need a new worker? Called from manager. */
970static bool need_to_create_worker(struct worker_pool *pool)
971{
972 return need_more_worker(pool) && !may_start_working(pool);
973}
974
975/* Do we have too many workers and should some go away? */
976static bool too_many_workers(struct worker_pool *pool)
977{
978 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
979 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
980 int nr_busy = pool->nr_workers - nr_idle;
981
982 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
983}
984
985/**
986 * worker_set_flags - set worker flags and adjust nr_running accordingly
987 * @worker: self
988 * @flags: flags to set
989 *
990 * Set @flags in @worker->flags and adjust nr_running accordingly.
991 */
992static inline void worker_set_flags(struct worker *worker, unsigned int flags)
993{
994 struct worker_pool *pool = worker->pool;
995
996 lockdep_assert_held(&pool->lock);
997
998 /* If transitioning into NOT_RUNNING, adjust nr_running. */
999 if ((flags & WORKER_NOT_RUNNING) &&
1000 !(worker->flags & WORKER_NOT_RUNNING)) {
1001 pool->nr_running--;
1002 }
1003
1004 worker->flags |= flags;
1005}
1006
1007/**
1008 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
1009 * @worker: self
1010 * @flags: flags to clear
1011 *
1012 * Clear @flags in @worker->flags and adjust nr_running accordingly.
1013 */
1014static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
1015{
1016 struct worker_pool *pool = worker->pool;
1017 unsigned int oflags = worker->flags;
1018
1019 lockdep_assert_held(&pool->lock);
1020
1021 worker->flags &= ~flags;
1022
1023 /*
1024 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1025 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1026 * of multiple flags, not a single flag.
1027 */
1028 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1029 if (!(worker->flags & WORKER_NOT_RUNNING))
1030 pool->nr_running++;
1031}
1032
1033/* Return the first idle worker. Called with pool->lock held. */
1034static struct worker *first_idle_worker(struct worker_pool *pool)
1035{
1036 if (unlikely(list_empty(&pool->idle_list)))
1037 return NULL;
1038
1039 return list_first_entry(&pool->idle_list, struct worker, entry);
1040}
1041
1042/**
1043 * worker_enter_idle - enter idle state
1044 * @worker: worker which is entering idle state
1045 *
1046 * @worker is entering idle state. Update stats and idle timer if
1047 * necessary.
1048 *
1049 * LOCKING:
1050 * raw_spin_lock_irq(pool->lock).
1051 */
1052static void worker_enter_idle(struct worker *worker)
1053{
1054 struct worker_pool *pool = worker->pool;
1055
1056 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1057 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1058 (worker->hentry.next || worker->hentry.pprev)))
1059 return;
1060
1061 /* can't use worker_set_flags(), also called from create_worker() */
1062 worker->flags |= WORKER_IDLE;
1063 pool->nr_idle++;
1064 worker->last_active = jiffies;
1065
1066 /* idle_list is LIFO */
1067 list_add(&worker->entry, &pool->idle_list);
1068
1069 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1070 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1071
1072 /* Sanity check nr_running. */
1073 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1074}
1075
1076/**
1077 * worker_leave_idle - leave idle state
1078 * @worker: worker which is leaving idle state
1079 *
1080 * @worker is leaving idle state. Update stats.
1081 *
1082 * LOCKING:
1083 * raw_spin_lock_irq(pool->lock).
1084 */
1085static void worker_leave_idle(struct worker *worker)
1086{
1087 struct worker_pool *pool = worker->pool;
1088
1089 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1090 return;
1091 worker_clr_flags(worker, WORKER_IDLE);
1092 pool->nr_idle--;
1093 list_del_init(&worker->entry);
1094}
1095
1096/**
1097 * find_worker_executing_work - find worker which is executing a work
1098 * @pool: pool of interest
1099 * @work: work to find worker for
1100 *
1101 * Find a worker which is executing @work on @pool by searching
1102 * @pool->busy_hash which is keyed by the address of @work. For a worker
1103 * to match, its current execution should match the address of @work and
1104 * its work function. This is to avoid unwanted dependency between
1105 * unrelated work executions through a work item being recycled while still
1106 * being executed.
1107 *
1108 * This is a bit tricky. A work item may be freed once its execution
1109 * starts and nothing prevents the freed area from being recycled for
1110 * another work item. If the same work item address ends up being reused
1111 * before the original execution finishes, workqueue will identify the
1112 * recycled work item as currently executing and make it wait until the
1113 * current execution finishes, introducing an unwanted dependency.
1114 *
1115 * This function checks the work item address and work function to avoid
1116 * false positives. Note that this isn't complete as one may construct a
1117 * work function which can introduce dependency onto itself through a
1118 * recycled work item. Well, if somebody wants to shoot oneself in the
1119 * foot that badly, there's only so much we can do, and if such deadlock
1120 * actually occurs, it should be easy to locate the culprit work function.
1121 *
1122 * CONTEXT:
1123 * raw_spin_lock_irq(pool->lock).
1124 *
1125 * Return:
1126 * Pointer to worker which is executing @work if found, %NULL
1127 * otherwise.
1128 */
1129static struct worker *find_worker_executing_work(struct worker_pool *pool,
1130 struct work_struct *work)
1131{
1132 struct worker *worker;
1133
1134 hash_for_each_possible(pool->busy_hash, worker, hentry,
1135 (unsigned long)work)
1136 if (worker->current_work == work &&
1137 worker->current_func == work->func)
1138 return worker;
1139
1140 return NULL;
1141}
1142
1143static void mayday_cursor_func(struct work_struct *work)
1144{
1145 /* should not be processed, only for marking position */
1146 BUG();
1147}
1148
1149/**
1150 * move_linked_works - move linked works to a list
1151 * @work: start of series of works to be scheduled
1152 * @head: target list to append @work to
1153 * @nextp: out parameter for nested worklist walking
1154 *
1155 * Schedule linked works starting from @work to @head. Work series to be
1156 * scheduled starts at @work and includes any consecutive work with
1157 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1158 * @nextp.
1159 *
1160 * CONTEXT:
1161 * raw_spin_lock_irq(pool->lock).
1162 */
1163static void move_linked_works(struct work_struct *work, struct list_head *head,
1164 struct work_struct **nextp)
1165{
1166 struct work_struct *n;
1167
1168 /*
1169 * Linked worklist will always end before the end of the list,
1170 * use NULL for list head.
1171 */
1172 list_for_each_entry_safe_from(work, n, NULL, entry) {
1173 list_move_tail(&work->entry, head);
1174 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1175 break;
1176 }
1177
1178 /*
1179 * If we're already inside safe list traversal and have moved
1180 * multiple works to the scheduled queue, the next position
1181 * needs to be updated.
1182 */
1183 if (nextp)
1184 *nextp = n;
1185}
1186
1187/**
1188 * assign_work - assign a work item and its linked work items to a worker
1189 * @work: work to assign
1190 * @worker: worker to assign to
1191 * @nextp: out parameter for nested worklist walking
1192 *
1193 * Assign @work and its linked work items to @worker. If @work is already being
1194 * executed by another worker in the same pool, it'll be punted there.
1195 *
1196 * If @nextp is not NULL, it's updated to point to the next work of the last
1197 * scheduled work. This allows assign_work() to be nested inside
1198 * list_for_each_entry_safe().
1199 *
1200 * Returns %true if @work was successfully assigned to @worker. %false if @work
1201 * was punted to another worker already executing it.
1202 */
1203static bool assign_work(struct work_struct *work, struct worker *worker,
1204 struct work_struct **nextp)
1205{
1206 struct worker_pool *pool = worker->pool;
1207 struct worker *collision;
1208
1209 lockdep_assert_held(&pool->lock);
1210
1211 /* The cursor work should not be processed */
1212 if (unlikely(work->func == mayday_cursor_func)) {
1213 /* only worker_thread() can possibly take this branch */
1214 WARN_ON_ONCE(worker->rescue_wq);
1215 if (nextp)
1216 *nextp = list_next_entry(work, entry);
1217 list_del_init(&work->entry);
1218 return false;
1219 }
1220
1221 /*
1222 * A single work shouldn't be executed concurrently by multiple workers.
1223 * __queue_work() ensures that @work doesn't jump to a different pool
1224 * while still running in the previous pool. Here, we should ensure that
1225 * @work is not executed concurrently by multiple workers from the same
1226 * pool. Check whether anyone is already processing the work. If so,
1227 * defer the work to the currently executing one.
1228 */
1229 collision = find_worker_executing_work(pool, work);
1230 if (unlikely(collision)) {
1231 move_linked_works(work, &collision->scheduled, nextp);
1232 return false;
1233 }
1234
1235 move_linked_works(work, &worker->scheduled, nextp);
1236 return true;
1237}
1238
1239static struct irq_work *bh_pool_irq_work(struct worker_pool *pool)
1240{
1241 int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0;
1242
1243 return &per_cpu(bh_pool_irq_works, pool->cpu)[high];
1244}
1245
1246static void kick_bh_pool(struct worker_pool *pool)
1247{
1248#ifdef CONFIG_SMP
1249 /* see drain_dead_softirq_workfn() for BH_DRAINING */
1250 if (unlikely(pool->cpu != smp_processor_id() &&
1251 !(pool->flags & POOL_BH_DRAINING))) {
1252 irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu);
1253 return;
1254 }
1255#endif
1256 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
1257 raise_softirq_irqoff(HI_SOFTIRQ);
1258 else
1259 raise_softirq_irqoff(TASKLET_SOFTIRQ);
1260}
1261
1262/**
1263 * kick_pool_pick - select an idle worker to kick, deferring the wakeup
1264 * @pool: pool to kick
1265 * @wakep: out-param, set to the task to wake after pool->lock is dropped
1266 *
1267 * Like kick_pool() but, for a regular (non-BH) pool, returns the picked
1268 * worker's task via @wakep instead of waking it, so the caller can issue the
1269 * wakeup after dropping pool->lock (the wakeup takes rq->lock). Worker
1270 * selection, wake_cpu setup and the BH kick still happen under the lock.
1271 * Returns whether a worker was selected or kicked.
1272 *
1273 * Must be called with @pool->lock held.
1274 */
1275static bool kick_pool_pick(struct worker_pool *pool, struct task_struct **wakep)
1276{
1277 struct worker *worker = first_idle_worker(pool);
1278 struct task_struct *p;
1279
1280 lockdep_assert_held(&pool->lock);
1281
1282 *wakep = NULL;
1283
1284 if (!need_more_worker(pool) || !worker)
1285 return false;
1286
1287 if (pool->flags & POOL_BH) {
1288 kick_bh_pool(pool);
1289 return true;
1290 }
1291
1292 p = worker->task;
1293
1294#ifdef CONFIG_SMP
1295 /*
1296 * Idle @worker is about to execute @work and waking up provides an
1297 * opportunity to migrate @worker at a lower cost by setting the task's
1298 * wake_cpu field. Let's see if we want to move @worker to improve
1299 * execution locality.
1300 *
1301 * We're waking the worker that went idle the latest and there's some
1302 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1303 * so, setting the wake_cpu won't do anything. As this is a best-effort
1304 * optimization and the race window is narrow, let's leave as-is for
1305 * now. If this becomes pronounced, we can skip over workers which are
1306 * still on cpu when picking an idle worker.
1307 *
1308 * If @pool has non-strict affinity, @worker might have ended up outside
1309 * its affinity scope. Repatriate.
1310 */
1311 if (!pool->attrs->affn_strict &&
1312 !cpumask_test_cpu(READ_ONCE(p->wake_cpu),
1313 pool->attrs->__pod_cpumask)) {
1314 struct work_struct *work = list_first_entry(&pool->worklist,
1315 struct work_struct, entry);
1316 int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask,
1317 cpu_online_mask);
1318 if (wake_cpu < nr_cpu_ids) {
1319 WRITE_ONCE(p->wake_cpu, wake_cpu);
1320 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1321 }
1322 }
1323#endif
1324 /* Track the last idle worker woken, used for stall diagnostics. */
1325 pool->last_woken_worker = worker;
1326
1327 *wakep = p;
1328 return true;
1329}
1330
1331/**
1332 * kick_pool - wake up an idle worker if necessary
1333 * @pool: pool to kick
1334 *
1335 * @pool may have pending work items. Wake up worker if necessary. Returns
1336 * whether a worker was woken up.
1337 */
1338static bool kick_pool(struct worker_pool *pool)
1339{
1340 struct task_struct *p;
1341 bool kicked = kick_pool_pick(pool, &p);
1342
1343 if (p)
1344 wake_up_process(p);
1345 return kicked;
1346}
1347
1348#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1349
1350/*
1351 * Concurrency-managed per-cpu work items that hog CPU for longer than
1352 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1353 * which prevents them from stalling other concurrency-managed work items. If a
1354 * work function keeps triggering this mechanism, it's likely that the work item
1355 * should be using an unbound workqueue instead.
1356 *
1357 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1358 * and report them so that they can be examined and converted to use unbound
1359 * workqueues as appropriate. To avoid flooding the console, each violating work
1360 * function is tracked and reported with exponential backoff.
1361 */
1362#define WCI_MAX_ENTS 128
1363
1364struct wci_ent {
1365 work_func_t func;
1366 atomic64_t cnt;
1367 struct hlist_node hash_node;
1368};
1369
1370static struct wci_ent wci_ents[WCI_MAX_ENTS];
1371static int wci_nr_ents;
1372static DEFINE_RAW_SPINLOCK(wci_lock);
1373static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1374
1375static struct wci_ent *wci_find_ent(work_func_t func)
1376{
1377 struct wci_ent *ent;
1378
1379 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1380 (unsigned long)func) {
1381 if (ent->func == func)
1382 return ent;
1383 }
1384 return NULL;
1385}
1386
1387static void wq_cpu_intensive_report(work_func_t func)
1388{
1389 struct wci_ent *ent;
1390
1391restart:
1392 ent = wci_find_ent(func);
1393 if (ent) {
1394 u64 cnt;
1395
1396 /*
1397 * Start reporting from the warning_thresh and back off
1398 * exponentially.
1399 */
1400 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1401 if (wq_cpu_intensive_warning_thresh &&
1402 cnt >= wq_cpu_intensive_warning_thresh &&
1403 is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh))
1404 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1405 ent->func, wq_cpu_intensive_thresh_us,
1406 atomic64_read(&ent->cnt));
1407 return;
1408 }
1409
1410 /*
1411 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1412 * is exhausted, something went really wrong and we probably made enough
1413 * noise already.
1414 */
1415 if (wci_nr_ents >= WCI_MAX_ENTS)
1416 return;
1417
1418 raw_spin_lock(&wci_lock);
1419
1420 if (wci_nr_ents >= WCI_MAX_ENTS) {
1421 raw_spin_unlock(&wci_lock);
1422 return;
1423 }
1424
1425 if (wci_find_ent(func)) {
1426 raw_spin_unlock(&wci_lock);
1427 goto restart;
1428 }
1429
1430 ent = &wci_ents[wci_nr_ents++];
1431 ent->func = func;
1432 atomic64_set(&ent->cnt, 0);
1433 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1434
1435 raw_spin_unlock(&wci_lock);
1436
1437 goto restart;
1438}
1439
1440#else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1441static void wq_cpu_intensive_report(work_func_t func) {}
1442#endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1443
1444/**
1445 * wq_worker_running - a worker is running again
1446 * @task: task waking up
1447 *
1448 * This function is called when a worker returns from schedule()
1449 */
1450void wq_worker_running(struct task_struct *task)
1451{
1452 struct worker *worker = kthread_data(task);
1453
1454 if (!READ_ONCE(worker->sleeping))
1455 return;
1456
1457 /*
1458 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1459 * and the nr_running increment below, we may ruin the nr_running reset
1460 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1461 * pool. Protect against such race.
1462 */
1463 preempt_disable();
1464 if (!(worker->flags & WORKER_NOT_RUNNING))
1465 worker->pool->nr_running++;
1466 preempt_enable();
1467
1468 /*
1469 * CPU intensive auto-detection cares about how long a work item hogged
1470 * CPU without sleeping. Reset the starting timestamp on wakeup.
1471 */
1472 worker->current_at = READ_ONCE(worker->task->se.sum_exec_runtime);
1473
1474 WRITE_ONCE(worker->sleeping, 0);
1475}
1476
1477/**
1478 * wq_worker_sleeping - a worker is going to sleep
1479 * @task: task going to sleep
1480 *
1481 * This function is called from schedule() when a busy worker is
1482 * going to sleep.
1483 */
1484void wq_worker_sleeping(struct task_struct *task)
1485{
1486 struct worker *worker = kthread_data(task);
1487 struct worker_pool *pool;
1488
1489 /*
1490 * Rescuers, which may not have all the fields set up like normal
1491 * workers, also reach here, let's not access anything before
1492 * checking NOT_RUNNING.
1493 */
1494 if (worker->flags & WORKER_NOT_RUNNING)
1495 return;
1496
1497 pool = worker->pool;
1498
1499 /* Return if preempted before wq_worker_running() was reached */
1500 if (READ_ONCE(worker->sleeping))
1501 return;
1502
1503 WRITE_ONCE(worker->sleeping, 1);
1504 raw_spin_lock_irq(&pool->lock);
1505
1506 /*
1507 * Recheck in case unbind_workers() preempted us. We don't
1508 * want to decrement nr_running after the worker is unbound
1509 * and nr_running has been reset.
1510 */
1511 if (worker->flags & WORKER_NOT_RUNNING) {
1512 raw_spin_unlock_irq(&pool->lock);
1513 return;
1514 }
1515
1516 pool->nr_running--;
1517 if (kick_pool(pool))
1518 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1519
1520 raw_spin_unlock_irq(&pool->lock);
1521}
1522
1523/**
1524 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1525 * @task: task currently running
1526 *
1527 * Called from sched_tick(). We're in the IRQ context and the current
1528 * worker's fields which follow the 'K' locking rule can be accessed safely.
1529 */
1530void wq_worker_tick(struct task_struct *task)
1531{
1532 struct worker *worker = kthread_data(task);
1533 struct pool_workqueue *pwq = worker->current_pwq;
1534 struct worker_pool *pool = worker->pool;
1535
1536 if (!pwq)
1537 return;
1538
1539 /*
1540 * @pwq is shared across CPUs for unbound wqs and this advisory stat is
1541 * bumped outside pool->lock, so the update is intentionally racy.
1542 */
1543 data_race(pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC);
1544
1545 if (!wq_cpu_intensive_thresh_us)
1546 return;
1547
1548 /*
1549 * If the current worker is concurrency managed and hogged the CPU for
1550 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1551 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1552 *
1553 * Set @worker->sleeping means that @worker is in the process of
1554 * switching out voluntarily and won't be contributing to
1555 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1556 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1557 * double decrements. The task is releasing the CPU anyway. Let's skip.
1558 * We probably want to make this prettier in the future.
1559 */
1560 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1561 READ_ONCE(worker->task->se.sum_exec_runtime) - worker->current_at <
1562 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1563 return;
1564
1565 raw_spin_lock(&pool->lock);
1566
1567 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1568 wq_cpu_intensive_report(worker->current_func);
1569 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1570
1571 if (kick_pool(pool))
1572 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1573
1574 raw_spin_unlock(&pool->lock);
1575}
1576
1577/**
1578 * wq_worker_last_func - retrieve worker's last work function
1579 * @task: Task to retrieve last work function of.
1580 *
1581 * Determine the last function a worker executed. This is called from
1582 * the scheduler to get a worker's last known identity.
1583 *
1584 * CONTEXT:
1585 * raw_spin_lock_irq(rq->lock)
1586 *
1587 * This function is called during schedule() when a kworker is going
1588 * to sleep. It's used by psi to identify aggregation workers during
1589 * dequeuing, to allow periodic aggregation to shut-off when that
1590 * worker is the last task in the system or cgroup to go to sleep.
1591 *
1592 * As this function doesn't involve any workqueue-related locking, it
1593 * only returns stable values when called from inside the scheduler's
1594 * queuing and dequeuing paths, when @task, which must be a kworker,
1595 * is guaranteed to not be processing any works.
1596 *
1597 * Return:
1598 * The last work function %current executed as a worker, NULL if it
1599 * hasn't executed any work yet.
1600 */
1601work_func_t wq_worker_last_func(struct task_struct *task)
1602{
1603 struct worker *worker = kthread_data(task);
1604
1605 return worker->last_func;
1606}
1607
1608/* True if @pool is a static per-cpu pool rather than an unbound one. */
1609static bool is_percpu_pool(struct worker_pool *pool)
1610{
1611 return pool->cpu >= 0;
1612}
1613
1614/**
1615 * wq_node_nr_active - Determine wq_node_nr_active to use
1616 * @wq: workqueue of interest
1617 * @node: NUMA node, can be %NUMA_NO_NODE
1618 *
1619 * Determine wq_node_nr_active to use for @wq on @node. @wq must be unbound.
1620 * Returns:
1621 *
1622 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1623 *
1624 * - Otherwise, node_nr_active[@node].
1625 */
1626static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq,
1627 int node)
1628{
1629 BUG_ON(!(wq->flags & WQ_UNBOUND));
1630
1631 if (node == NUMA_NO_NODE)
1632 node = nr_node_ids;
1633
1634 return wq->node_nr_active[node];
1635}
1636
1637/**
1638 * wq_update_node_max_active - Update per-node max_actives to use
1639 * @wq: workqueue to update
1640 * @off_cpu: CPU that's going down, -1 if a CPU is not going down
1641 *
1642 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1643 * distributed among nodes according to the proportions of numbers of online
1644 * cpus. The result is always between @wq->min_active and max_active.
1645 */
1646static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1647{
1648 struct cpumask *effective = unbound_effective_cpumask(wq);
1649 int min_active = READ_ONCE(wq->min_active);
1650 int max_active = READ_ONCE(wq->max_active);
1651 int total_cpus, node;
1652
1653 lockdep_assert_held(&wq->mutex);
1654
1655 if (!wq_topo_initialized)
1656 return;
1657
1658 if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective))
1659 off_cpu = -1;
1660
1661 total_cpus = cpumask_weight_and(effective, cpu_online_mask);
1662 if (off_cpu >= 0)
1663 total_cpus--;
1664
1665 /* If all CPUs of the wq get offline, use the default values */
1666 if (unlikely(!total_cpus)) {
1667 for_each_node(node)
1668 wq_node_nr_active(wq, node)->max = min_active;
1669
1670 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1671 return;
1672 }
1673
1674 for_each_node(node) {
1675 int node_cpus;
1676
1677 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node));
1678 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node)
1679 node_cpus--;
1680
1681 wq_node_nr_active(wq, node)->max =
1682 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1683 min_active, max_active);
1684 }
1685
1686 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1687}
1688
1689/**
1690 * get_pwq - get an extra reference on the specified pool_workqueue
1691 * @pwq: pool_workqueue to get
1692 *
1693 * Obtain an extra reference on @pwq. The caller should guarantee that
1694 * @pwq has positive refcnt and be holding the matching pool->lock.
1695 */
1696static void get_pwq(struct pool_workqueue *pwq)
1697{
1698 lockdep_assert_held(&pwq->pool->lock);
1699 WARN_ON_ONCE(pwq->refcnt <= 0);
1700 pwq->refcnt++;
1701}
1702
1703/**
1704 * put_pwq - put a pool_workqueue reference
1705 * @pwq: pool_workqueue to put
1706 *
1707 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1708 * destruction. The caller should be holding the matching pool->lock.
1709 */
1710static void put_pwq(struct pool_workqueue *pwq)
1711{
1712 lockdep_assert_held(&pwq->pool->lock);
1713 if (likely(--pwq->refcnt))
1714 return;
1715 /*
1716 * @pwq can't be released under pool->lock, bounce to a dedicated
1717 * kthread_worker to avoid A-A deadlocks.
1718 */
1719 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1720}
1721
1722/**
1723 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1724 * @pwq: pool_workqueue to put (can be %NULL)
1725 *
1726 * put_pwq() with locking. This function also allows %NULL @pwq.
1727 */
1728static void put_pwq_unlocked(struct pool_workqueue *pwq)
1729{
1730 if (pwq) {
1731 /*
1732 * As both pwqs and pools are RCU protected, the
1733 * following lock operations are safe.
1734 */
1735 raw_spin_lock_irq(&pwq->pool->lock);
1736 put_pwq(pwq);
1737 raw_spin_unlock_irq(&pwq->pool->lock);
1738 }
1739}
1740
1741static bool pwq_is_empty(struct pool_workqueue *pwq)
1742{
1743 return !pwq->nr_active && list_empty(&pwq->inactive_works);
1744}
1745
1746static void __pwq_activate_work(struct pool_workqueue *pwq,
1747 struct work_struct *work)
1748{
1749 unsigned long *wdb = work_data_bits(work);
1750
1751 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1752 trace_workqueue_activate_work(work);
1753 if (list_empty(&pwq->pool->worklist))
1754 pwq->pool->last_progress_ts = jiffies;
1755 move_linked_works(work, &pwq->pool->worklist, NULL);
1756 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1757}
1758
1759static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1760{
1761 int max = READ_ONCE(nna->max);
1762 int old = atomic_read(&nna->nr);
1763
1764 do {
1765 if (old >= max)
1766 return false;
1767 } while (!atomic_try_cmpxchg_relaxed(&nna->nr, &old, old + 1));
1768
1769 return true;
1770}
1771
1772/**
1773 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1774 * @pwq: pool_workqueue of interest
1775 * @fill: max_active may have increased, try to increase concurrency level
1776 *
1777 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1778 * successfully obtained. %false otherwise.
1779 */
1780static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1781{
1782 struct workqueue_struct *wq = pwq->wq;
1783 struct worker_pool *pool = pwq->pool;
1784 struct wq_node_nr_active *nna;
1785 bool obtained = false;
1786
1787 lockdep_assert_held(&pool->lock);
1788
1789 /*
1790 * A concurrency-managed per-cpu pool accounts nr_active per pwq, so
1791 * pwq->nr_active against wq->max_active is sufficient.
1792 */
1793 if (is_percpu_pool(pool)) {
1794 obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1795 goto out;
1796 }
1797
1798 if (unlikely(pwq->plugged))
1799 return false;
1800
1801 nna = wq_node_nr_active(wq, pool->node);
1802
1803 /*
1804 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1805 * already waiting on $nna, pwq_dec_nr_active() will maintain the
1806 * concurrency level. Don't jump the line.
1807 *
1808 * We need to ignore the pending test after max_active has increased as
1809 * pwq_dec_nr_active() can only maintain the concurrency level but not
1810 * increase it. This is indicated by @fill.
1811 */
1812 if (!list_empty(&pwq->pending_node) && likely(!fill))
1813 goto out;
1814
1815 obtained = tryinc_node_nr_active(nna);
1816 if (obtained)
1817 goto out;
1818
1819 /*
1820 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1821 * and try again. The smp_mb() is paired with the implied memory barrier
1822 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1823 * we see the decremented $nna->nr or they see non-empty
1824 * $nna->pending_pwqs.
1825 */
1826 raw_spin_lock(&nna->lock);
1827
1828 if (list_empty(&pwq->pending_node))
1829 list_add_tail(&pwq->pending_node, &nna->pending_pwqs);
1830 else if (likely(!fill))
1831 goto out_unlock;
1832
1833 smp_mb();
1834
1835 obtained = tryinc_node_nr_active(nna);
1836
1837 /*
1838 * If @fill, @pwq might have already been pending. Being spuriously
1839 * pending in cold paths doesn't affect anything. Let's leave it be.
1840 */
1841 if (obtained && likely(!fill))
1842 list_del_init(&pwq->pending_node);
1843
1844out_unlock:
1845 raw_spin_unlock(&nna->lock);
1846out:
1847 if (obtained)
1848 pwq->nr_active++;
1849 return obtained;
1850}
1851
1852/**
1853 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1854 * @pwq: pool_workqueue of interest
1855 * @fill: max_active may have increased, try to increase concurrency level
1856 *
1857 * Activate the first inactive work item of @pwq if available and allowed by
1858 * max_active limit.
1859 *
1860 * Returns %true if an inactive work item has been activated. %false if no
1861 * inactive work item is found or max_active limit is reached.
1862 */
1863static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1864{
1865 struct work_struct *work =
1866 list_first_entry_or_null(&pwq->inactive_works,
1867 struct work_struct, entry);
1868
1869 if (work && pwq_tryinc_nr_active(pwq, fill)) {
1870 __pwq_activate_work(pwq, work);
1871 return true;
1872 } else {
1873 return false;
1874 }
1875}
1876
1877/**
1878 * unplug_oldest_pwq - unplug the oldest pool_workqueue
1879 * @wq: workqueue_struct where its oldest pwq is to be unplugged
1880 *
1881 * This function should only be called for ordered workqueues where only the
1882 * oldest pwq is unplugged, the others are plugged to suspend execution to
1883 * ensure proper work item ordering::
1884 *
1885 * dfl_pwq --------------+ [P] - plugged
1886 * |
1887 * v
1888 * pwqs -> A -> B [P] -> C [P] (newest)
1889 * | | |
1890 * 1 3 5
1891 * | | |
1892 * 2 4 6
1893 *
1894 * When the oldest pwq is drained and removed, this function should be called
1895 * to unplug the next oldest one to start its work item execution. Note that
1896 * pwq's are linked into wq->pwqs with the oldest first, so the first one in
1897 * the list is the oldest.
1898 */
1899static void unplug_oldest_pwq(struct workqueue_struct *wq)
1900{
1901 struct pool_workqueue *pwq;
1902
1903 lockdep_assert_held(&wq->mutex);
1904
1905 /* Caller should make sure that pwqs isn't empty before calling */
1906 pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue,
1907 pwqs_node);
1908 raw_spin_lock_irq(&pwq->pool->lock);
1909 if (pwq->plugged) {
1910 pwq->plugged = false;
1911 if (pwq_activate_first_inactive(pwq, true)) {
1912 /*
1913 * While plugged, queueing skips activation which
1914 * includes bumping the nr_active count and adding the
1915 * pwq to nna->pending_pwqs if the count can't be
1916 * obtained. We need to restore both for the pwq being
1917 * unplugged. The first call activates the first
1918 * inactive work item and the second, if there are more
1919 * inactive, puts the pwq on pending_pwqs.
1920 */
1921 pwq_activate_first_inactive(pwq, false);
1922
1923 kick_pool(pwq->pool);
1924 }
1925 }
1926 raw_spin_unlock_irq(&pwq->pool->lock);
1927}
1928
1929/**
1930 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1931 * @nna: wq_node_nr_active to activate a pending pwq for
1932 * @caller_pool: worker_pool the caller is locking
1933 *
1934 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1935 * @caller_pool may be unlocked and relocked to lock other worker_pools.
1936 */
1937static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1938 struct worker_pool *caller_pool)
1939{
1940 struct worker_pool *locked_pool = caller_pool;
1941 struct pool_workqueue *pwq;
1942 struct work_struct *work;
1943
1944 lockdep_assert_held(&caller_pool->lock);
1945
1946 raw_spin_lock(&nna->lock);
1947retry:
1948 pwq = list_first_entry_or_null(&nna->pending_pwqs,
1949 struct pool_workqueue, pending_node);
1950 if (!pwq)
1951 goto out_unlock;
1952
1953 /*
1954 * If @pwq is for a different pool than @locked_pool, we need to lock
1955 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1956 * / lock dance. For that, we also need to release @nna->lock as it's
1957 * nested inside pool locks.
1958 */
1959 if (pwq->pool != locked_pool) {
1960 raw_spin_unlock(&locked_pool->lock);
1961 locked_pool = pwq->pool;
1962 if (!raw_spin_trylock(&locked_pool->lock)) {
1963 raw_spin_unlock(&nna->lock);
1964 raw_spin_lock(&locked_pool->lock);
1965 raw_spin_lock(&nna->lock);
1966 goto retry;
1967 }
1968 }
1969
1970 /*
1971 * $pwq may not have any inactive work items due to e.g. cancellations.
1972 * Drop it from pending_pwqs and see if there's another one.
1973 */
1974 work = list_first_entry_or_null(&pwq->inactive_works,
1975 struct work_struct, entry);
1976 if (!work) {
1977 list_del_init(&pwq->pending_node);
1978 goto retry;
1979 }
1980
1981 /*
1982 * Acquire an nr_active count and activate the inactive work item. If
1983 * $pwq still has inactive work items, rotate it to the end of the
1984 * pending_pwqs so that we round-robin through them. This means that
1985 * inactive work items are not activated in queueing order which is fine
1986 * given that there has never been any ordering across different pwqs.
1987 */
1988 if (likely(tryinc_node_nr_active(nna))) {
1989 pwq->nr_active++;
1990 __pwq_activate_work(pwq, work);
1991
1992 if (list_empty(&pwq->inactive_works))
1993 list_del_init(&pwq->pending_node);
1994 else
1995 list_move_tail(&pwq->pending_node, &nna->pending_pwqs);
1996
1997 /* if activating a foreign pool, make sure it's running */
1998 if (pwq->pool != caller_pool)
1999 kick_pool(pwq->pool);
2000 }
2001
2002out_unlock:
2003 raw_spin_unlock(&nna->lock);
2004 if (locked_pool != caller_pool) {
2005 raw_spin_unlock(&locked_pool->lock);
2006 raw_spin_lock(&caller_pool->lock);
2007 }
2008}
2009
2010/**
2011 * pwq_dec_nr_active - Retire an active count
2012 * @pwq: pool_workqueue of interest
2013 *
2014 * Decrement @pwq's nr_active and try to activate the first inactive work item.
2015 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
2016 */
2017static void pwq_dec_nr_active(struct pool_workqueue *pwq)
2018{
2019 struct worker_pool *pool = pwq->pool;
2020 struct wq_node_nr_active *nna;
2021
2022 lockdep_assert_held(&pool->lock);
2023
2024 /*
2025 * @pwq->nr_active should be decremented for both percpu and unbound
2026 * workqueues.
2027 */
2028 pwq->nr_active--;
2029
2030 /*
2031 * A concurrency-managed per-cpu pool only needs to kick the first
2032 * inactive work item on @pwq itself.
2033 */
2034 if (is_percpu_pool(pool)) {
2035 pwq_activate_first_inactive(pwq, false);
2036 return;
2037 }
2038
2039 nna = wq_node_nr_active(pwq->wq, pool->node);
2040
2041 /*
2042 * If @pwq is for an unbound workqueue, it's more complicated because
2043 * multiple pwqs and pools may be sharing the nr_active count. When a
2044 * pwq needs to wait for an nr_active count, it puts itself on
2045 * $nna->pending_pwqs. The following atomic_dec_return()'s implied
2046 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
2047 * guarantee that either we see non-empty pending_pwqs or they see
2048 * decremented $nna->nr.
2049 *
2050 * $nna->max may change as CPUs come online/offline and @pwq->wq's
2051 * max_active gets updated. However, it is guaranteed to be equal to or
2052 * larger than @pwq->wq->min_active which is above zero unless freezing.
2053 * This maintains the forward progress guarantee.
2054 */
2055 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max))
2056 return;
2057
2058 if (!list_empty(&nna->pending_pwqs))
2059 node_activate_pending_pwq(nna, pool);
2060}
2061
2062/**
2063 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
2064 * @pwq: pwq of interest
2065 * @work_data: work_data of work which left the queue
2066 *
2067 * A work either has completed or is removed from pending queue,
2068 * decrement nr_in_flight of its pwq and handle workqueue flushing.
2069 *
2070 * NOTE:
2071 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock
2072 * and thus should be called after all other state updates for the in-flight
2073 * work item is complete.
2074 *
2075 * CONTEXT:
2076 * raw_spin_lock_irq(pool->lock).
2077 */
2078static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
2079{
2080 int color = get_work_color(work_data);
2081
2082 if (!(work_data & WORK_STRUCT_INACTIVE))
2083 pwq_dec_nr_active(pwq);
2084
2085 pwq->nr_in_flight[color]--;
2086
2087 /* is flush in progress and are we at the flushing tip? */
2088 if (likely(pwq->flush_color != color))
2089 goto out_put;
2090
2091 /* are there still in-flight works? */
2092 if (pwq->nr_in_flight[color])
2093 goto out_put;
2094
2095 /* this pwq is done, clear flush_color */
2096 pwq->flush_color = -1;
2097
2098 /*
2099 * If this was the last pwq, wake up the first flusher. It
2100 * will handle the rest.
2101 */
2102 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2103 complete(&pwq->wq->first_flusher->done);
2104out_put:
2105 put_pwq(pwq);
2106}
2107
2108/**
2109 * try_to_grab_pending - steal work item from worklist and disable irq
2110 * @work: work item to steal
2111 * @cflags: %WORK_CANCEL_ flags
2112 * @irq_flags: place to store irq state
2113 *
2114 * Try to grab PENDING bit of @work. This function can handle @work in any
2115 * stable state - idle, on timer or on worklist.
2116 *
2117 * Return:
2118 *
2119 * ======== ================================================================
2120 * 1 if @work was pending and we successfully stole PENDING
2121 * 0 if @work was idle and we claimed PENDING
2122 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
2123 * ======== ================================================================
2124 *
2125 * Note:
2126 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2127 * interrupted while holding PENDING and @work off queue, irq must be
2128 * disabled on entry. This, combined with delayed_work->timer being
2129 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2130 *
2131 * On successful return, >= 0, irq is disabled and the caller is
2132 * responsible for releasing it using local_irq_restore(*@irq_flags).
2133 *
2134 * This function is safe to call from any context including IRQ handler.
2135 */
2136static int try_to_grab_pending(struct work_struct *work, u32 cflags,
2137 unsigned long *irq_flags)
2138{
2139 struct worker_pool *pool;
2140 struct pool_workqueue *pwq;
2141
2142 local_irq_save(*irq_flags);
2143
2144 /* try to steal the timer if it exists */
2145 if (cflags & WORK_CANCEL_DELAYED) {
2146 struct delayed_work *dwork = to_delayed_work(work);
2147
2148 /*
2149 * dwork->timer is irqsafe. If timer_delete() fails, it's
2150 * guaranteed that the timer is not queued anywhere and not
2151 * running on the local CPU.
2152 */
2153 if (likely(timer_delete(&dwork->timer)))
2154 return 1;
2155 }
2156
2157 /* try to claim PENDING the normal way */
2158 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2159 return 0;
2160
2161 rcu_read_lock();
2162 /*
2163 * The queueing is in progress, or it is already queued. Try to
2164 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2165 */
2166 pool = get_work_pool(work);
2167 if (!pool)
2168 goto fail;
2169
2170 raw_spin_lock(&pool->lock);
2171 /*
2172 * work->data is guaranteed to point to pwq only while the work
2173 * item is queued on pwq->wq, and both updating work->data to point
2174 * to pwq on queueing and to pool on dequeueing are done under
2175 * pwq->pool->lock. This in turn guarantees that, if work->data
2176 * points to pwq which is associated with a locked pool, the work
2177 * item is currently queued on that pool.
2178 */
2179 pwq = get_work_pwq(work);
2180 if (pwq && pwq->pool == pool) {
2181 unsigned long work_data = *work_data_bits(work);
2182
2183 debug_work_deactivate(work);
2184
2185 /*
2186 * A cancelable inactive work item must be in the
2187 * pwq->inactive_works since a queued barrier can't be
2188 * canceled (see the comments in insert_wq_barrier()).
2189 *
2190 * An inactive work item cannot be deleted directly because
2191 * it might have linked barrier work items which, if left
2192 * on the inactive_works list, will confuse pwq->nr_active
2193 * management later on and cause stall. Move the linked
2194 * barrier work items to the worklist when deleting the grabbed
2195 * item. Also keep WORK_STRUCT_INACTIVE in work_data, so that
2196 * it doesn't participate in nr_active management in later
2197 * pwq_dec_nr_in_flight().
2198 */
2199 if (work_data & WORK_STRUCT_INACTIVE)
2200 move_linked_works(work, &pwq->pool->worklist, NULL);
2201
2202 list_del_init(&work->entry);
2203
2204 /*
2205 * work->data points to pwq iff queued. Let's point to pool. As
2206 * this destroys work->data needed by the next step, stash it.
2207 */
2208 set_work_pool_and_keep_pending(work, pool->id,
2209 pool_offq_flags(pool));
2210
2211 /* must be the last step, see the function comment */
2212 pwq_dec_nr_in_flight(pwq, work_data);
2213
2214 raw_spin_unlock(&pool->lock);
2215 rcu_read_unlock();
2216 return 1;
2217 }
2218 raw_spin_unlock(&pool->lock);
2219fail:
2220 rcu_read_unlock();
2221 local_irq_restore(*irq_flags);
2222 return -EAGAIN;
2223}
2224
2225/**
2226 * work_grab_pending - steal work item from worklist and disable irq
2227 * @work: work item to steal
2228 * @cflags: %WORK_CANCEL_ flags
2229 * @irq_flags: place to store IRQ state
2230 *
2231 * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer
2232 * or on worklist.
2233 *
2234 * Can be called from any context. IRQ is disabled on return with IRQ state
2235 * stored in *@irq_flags. The caller is responsible for re-enabling it using
2236 * local_irq_restore().
2237 *
2238 * Returns %true if @work was pending. %false if idle.
2239 */
2240static bool work_grab_pending(struct work_struct *work, u32 cflags,
2241 unsigned long *irq_flags)
2242{
2243 int ret;
2244
2245 while (true) {
2246 ret = try_to_grab_pending(work, cflags, irq_flags);
2247 if (ret >= 0)
2248 return ret;
2249 cpu_relax();
2250 }
2251}
2252
2253/**
2254 * insert_work - insert a work into a pool
2255 * @pwq: pwq @work belongs to
2256 * @work: work to insert
2257 * @head: insertion point
2258 * @extra_flags: extra WORK_STRUCT_* flags to set
2259 *
2260 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2261 * work_struct flags.
2262 *
2263 * CONTEXT:
2264 * raw_spin_lock_irq(pool->lock).
2265 */
2266static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2267 struct list_head *head, unsigned int extra_flags)
2268{
2269 debug_work_activate(work);
2270
2271 /* record the work call stack in order to print it in KASAN reports */
2272 kasan_record_aux_stack(work);
2273
2274 /* we own @work, set data and link */
2275 set_work_pwq(work, pwq, extra_flags);
2276 list_add_tail(&work->entry, head);
2277 get_pwq(pwq);
2278}
2279
2280/*
2281 * Test whether @work is being queued from another work executing on the
2282 * same workqueue.
2283 */
2284static bool is_chained_work(struct workqueue_struct *wq)
2285{
2286 struct worker *worker;
2287
2288 worker = current_wq_worker();
2289 /*
2290 * Return %true iff I'm a worker executing a work item on @wq. If
2291 * I'm @worker, it's safe to dereference it without locking.
2292 */
2293 return worker && worker->current_pwq && worker->current_pwq->wq == wq;
2294}
2295
2296/*
2297 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2298 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2299 * avoid perturbing sensitive tasks.
2300 */
2301static int wq_select_unbound_cpu(int cpu)
2302{
2303 int new_cpu;
2304
2305 if (likely(!wq_debug_force_rr_cpu)) {
2306 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2307 return cpu;
2308 } else {
2309 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2310 }
2311
2312 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2313 new_cpu = cpumask_next_and_wrap(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2314 if (unlikely(new_cpu >= nr_cpu_ids))
2315 return cpu;
2316 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2317
2318 return new_cpu;
2319}
2320
2321static void __queue_work(int cpu, struct workqueue_struct *wq,
2322 struct work_struct *work)
2323{
2324 struct pool_workqueue *pwq;
2325 struct worker_pool *last_pool, *pool;
2326 struct task_struct *wake_task = NULL;
2327 unsigned int work_flags;
2328 unsigned int req_cpu = cpu;
2329
2330 /*
2331 * NOTE: Check whether the used workqueue is deprecated and warn
2332 */
2333 if (unlikely(wq->flags & __WQ_DEPRECATED))
2334 pr_warn_once("workqueue: work func %ps enqueued on deprecated workqueue. "
2335 "Use system_{percpu|dfl}_wq instead.\n",
2336 work->func);
2337
2338 /*
2339 * While a work item is PENDING && off queue, a task trying to
2340 * steal the PENDING will busy-loop waiting for it to either get
2341 * queued or lose PENDING. Grabbing PENDING and queueing should
2342 * happen with IRQ disabled.
2343 */
2344 lockdep_assert_irqs_disabled();
2345
2346 /*
2347 * For a draining wq, only works from the same workqueue are
2348 * allowed. The __WQ_DESTROYING helps to spot the issue that
2349 * queues a new work item to a wq after destroy_workqueue(wq).
2350 */
2351 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
2352 WARN_ONCE(!is_chained_work(wq), "workqueue: cannot queue %ps on wq %s\n",
2353 work->func, wq->name))) {
2354 struct work_offq_data offqd;
2355
2356 /*
2357 * State on entry: PENDING is set, work is off-queue (no
2358 * insert_work() has run).
2359 *
2360 * Returning without clearing PENDING would leave the work
2361 * in a weird state (PENDING=1, PWQ=0, entry empty)
2362 */
2363 work_offqd_unpack(&offqd, *work_data_bits(work));
2364 set_work_pool_and_clear_pending(work, offqd.pool_id,
2365 work_offqd_pack_flags(&offqd));
2366 return;
2367 }
2368 rcu_read_lock();
2369retry:
2370 /* pwq which will be used unless @work is executing elsewhere */
2371 if (req_cpu == WORK_CPU_UNBOUND) {
2372 if (wq->flags & WQ_UNBOUND)
2373 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2374 else
2375 cpu = raw_smp_processor_id();
2376 }
2377
2378 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2379 pool = pwq->pool;
2380
2381 /*
2382 * If @work was previously on a different pool, it might still be
2383 * running there, in which case the work needs to be queued on that
2384 * pool to guarantee non-reentrancy.
2385 *
2386 * For ordered workqueue, work items must be queued on the newest pwq
2387 * for accurate order management. Guaranteed order also guarantees
2388 * non-reentrancy. See the comments above unplug_oldest_pwq().
2389 */
2390 last_pool = get_work_pool(work);
2391 if (last_pool && last_pool != pool && !(wq->flags & __WQ_ORDERED)) {
2392 struct worker *worker;
2393
2394 raw_spin_lock(&last_pool->lock);
2395
2396 worker = find_worker_executing_work(last_pool, work);
2397
2398 if (worker && worker->current_pwq->wq == wq) {
2399 pwq = worker->current_pwq;
2400 pool = pwq->pool;
2401 WARN_ON_ONCE(pool != last_pool);
2402 } else {
2403 /* meh... not running there, queue here */
2404 raw_spin_unlock(&last_pool->lock);
2405 raw_spin_lock(&pool->lock);
2406 }
2407 } else {
2408 raw_spin_lock(&pool->lock);
2409 }
2410
2411 /*
2412 * pwq is determined and locked. For unbound pools, we could have raced
2413 * with pwq release and it could already be dead. If its refcnt is zero,
2414 * repeat pwq selection. Note that unbound pwqs never die without
2415 * another pwq replacing it in cpu_pwq or while work items are executing
2416 * on it, so the retrying is guaranteed to make forward-progress.
2417 */
2418 if (unlikely(!pwq->refcnt)) {
2419 if (wq->flags & WQ_UNBOUND) {
2420 raw_spin_unlock(&pool->lock);
2421 cpu_relax();
2422 goto retry;
2423 }
2424 /* oops */
2425 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2426 wq->name, cpu);
2427 }
2428
2429 /* pwq determined, queue */
2430 trace_workqueue_queue_work(req_cpu, pwq, work);
2431
2432 if (WARN_ON(!list_empty(&work->entry)))
2433 goto out;
2434
2435 pwq->nr_in_flight[pwq->work_color]++;
2436 work_flags = work_color_to_flags(pwq->work_color);
2437
2438 /*
2439 * Limit the number of concurrently active work items to max_active.
2440 * @work must also queue behind existing inactive work items to maintain
2441 * ordering when max_active changes. See wq_adjust_max_active().
2442 */
2443 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) {
2444 if (list_empty(&pool->worklist))
2445 pool->last_progress_ts = jiffies;
2446
2447 trace_workqueue_activate_work(work);
2448 insert_work(pwq, work, &pool->worklist, work_flags);
2449 kick_pool_pick(pool, &wake_task);
2450 } else {
2451 work_flags |= WORK_STRUCT_INACTIVE;
2452 insert_work(pwq, work, &pwq->inactive_works, work_flags);
2453 }
2454
2455out:
2456 raw_spin_unlock(&pool->lock);
2457 if (wake_task)
2458 wake_up_process(wake_task);
2459 rcu_read_unlock();
2460}
2461
2462static bool clear_pending_if_disabled(struct work_struct *work)
2463{
2464 unsigned long data = *work_data_bits(work);
2465 struct work_offq_data offqd;
2466
2467 if (likely((data & WORK_STRUCT_PWQ) ||
2468 !(data & WORK_OFFQ_DISABLE_MASK)))
2469 return false;
2470
2471 work_offqd_unpack(&offqd, data);
2472 set_work_pool_and_clear_pending(work, offqd.pool_id,
2473 work_offqd_pack_flags(&offqd));
2474 return true;
2475}
2476
2477/**
2478 * queue_work_on - queue work on specific cpu
2479 * @cpu: CPU number to execute work on
2480 * @wq: workqueue to use
2481 * @work: work to queue
2482 *
2483 * We queue the work to a specific CPU, the caller must ensure it
2484 * can't go away. Callers that fail to ensure that the specified
2485 * CPU cannot go away will execute on a randomly chosen CPU.
2486 * But note well that callers specifying a CPU that never has been
2487 * online will get a splat.
2488 *
2489 * Return: %false if @work was already on a queue, %true otherwise.
2490 */
2491bool queue_work_on(int cpu, struct workqueue_struct *wq,
2492 struct work_struct *work)
2493{
2494 bool ret = false;
2495 unsigned long irq_flags;
2496
2497 local_irq_save(irq_flags);
2498
2499 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2500 !clear_pending_if_disabled(work)) {
2501 __queue_work(cpu, wq, work);
2502 ret = true;
2503 }
2504
2505 local_irq_restore(irq_flags);
2506 return ret;
2507}
2508EXPORT_SYMBOL(queue_work_on);
2509
2510/**
2511 * select_numa_node_cpu - Select a CPU based on NUMA node
2512 * @node: NUMA node ID that we want to select a CPU from
2513 *
2514 * This function will attempt to find a "random" cpu available on a given
2515 * node. If there are no CPUs available on the given node it will return
2516 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2517 * available CPU if we need to schedule this work.
2518 */
2519static int select_numa_node_cpu(int node)
2520{
2521 int cpu;
2522
2523 /* Delay binding to CPU if node is not valid or online */
2524 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2525 return WORK_CPU_UNBOUND;
2526
2527 /* Use local node/cpu if we are already there */
2528 cpu = raw_smp_processor_id();
2529 if (node == cpu_to_node(cpu))
2530 return cpu;
2531
2532 /* Use "random" otherwise know as "first" online CPU of node */
2533 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2534
2535 /* If CPU is valid return that, otherwise just defer */
2536 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2537}
2538
2539/**
2540 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2541 * @node: NUMA node that we are targeting the work for
2542 * @wq: workqueue to use
2543 * @work: work to queue
2544 *
2545 * We queue the work to a "random" CPU within a given NUMA node. The basic
2546 * idea here is to provide a way to somehow associate work with a given
2547 * NUMA node.
2548 *
2549 * This function will only make a best effort attempt at getting this onto
2550 * the right NUMA node. If no node is requested or the requested node is
2551 * offline then we just fall back to standard queue_work behavior.
2552 *
2553 * Currently the "random" CPU ends up being the first available CPU in the
2554 * intersection of cpu_online_mask and the cpumask of the node, unless we
2555 * are running on the node. In that case we just use the current CPU.
2556 *
2557 * Return: %false if @work was already on a queue, %true otherwise.
2558 */
2559bool queue_work_node(int node, struct workqueue_struct *wq,
2560 struct work_struct *work)
2561{
2562 unsigned long irq_flags;
2563 bool ret = false;
2564
2565 /*
2566 * This current implementation is specific to unbound workqueues.
2567 * Specifically we only return the first available CPU for a given
2568 * node instead of cycling through individual CPUs within the node.
2569 *
2570 * If this is used with a per-cpu workqueue then the logic in
2571 * workqueue_select_cpu_near would need to be updated to allow for
2572 * some round robin type logic.
2573 */
2574 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2575
2576 local_irq_save(irq_flags);
2577
2578 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2579 !clear_pending_if_disabled(work)) {
2580 int cpu = select_numa_node_cpu(node);
2581
2582 __queue_work(cpu, wq, work);
2583 ret = true;
2584 }
2585
2586 local_irq_restore(irq_flags);
2587 return ret;
2588}
2589EXPORT_SYMBOL_GPL(queue_work_node);
2590
2591void delayed_work_timer_fn(struct timer_list *t)
2592{
2593 struct delayed_work *dwork = timer_container_of(dwork, t, timer);
2594
2595 /* should have been called from irqsafe timer with irq already off */
2596 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2597}
2598EXPORT_SYMBOL(delayed_work_timer_fn);
2599
2600static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2601 struct delayed_work *dwork, unsigned long delay)
2602{
2603 struct timer_list *timer = &dwork->timer;
2604 struct work_struct *work = &dwork->work;
2605
2606 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2607 WARN_ON_ONCE(timer_pending(timer));
2608 WARN_ON_ONCE(!list_empty(&work->entry));
2609
2610 /*
2611 * If @delay is 0, queue @dwork->work immediately. This is for
2612 * both optimization and correctness. The earliest @timer can
2613 * expire is on the closest next tick and delayed_work users depend
2614 * on that there's no such delay when @delay is 0.
2615 */
2616 if (!delay) {
2617 __queue_work(cpu, wq, &dwork->work);
2618 return;
2619 }
2620
2621 WARN_ON_ONCE(cpu != WORK_CPU_UNBOUND && !cpu_online(cpu));
2622 dwork->wq = wq;
2623 dwork->cpu = cpu;
2624 timer->expires = jiffies + delay;
2625
2626 if (housekeeping_enabled(HK_TYPE_TIMER)) {
2627 /* If the current cpu is a housekeeping cpu, use it. */
2628 cpu = smp_processor_id();
2629 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER))
2630 cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
2631 add_timer_on(timer, cpu);
2632 } else {
2633 if (likely(cpu == WORK_CPU_UNBOUND))
2634 add_timer_global(timer);
2635 else
2636 add_timer_on(timer, cpu);
2637 }
2638}
2639
2640/**
2641 * queue_delayed_work_on - queue work on specific CPU after delay
2642 * @cpu: CPU number to execute work on
2643 * @wq: workqueue to use
2644 * @dwork: work to queue
2645 * @delay: number of jiffies to wait before queueing
2646 *
2647 * We queue the delayed_work to a specific CPU, for non-zero delays the
2648 * caller must ensure it is online and can't go away. Callers that fail
2649 * to ensure this, may get @dwork->timer queued to an offlined CPU and
2650 * this will prevent queueing of @dwork->work unless the offlined CPU
2651 * becomes online again.
2652 *
2653 * Return: %false if @work was already on a queue, %true otherwise. If
2654 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2655 * execution.
2656 */
2657bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2658 struct delayed_work *dwork, unsigned long delay)
2659{
2660 struct work_struct *work = &dwork->work;
2661 bool ret = false;
2662 unsigned long irq_flags;
2663
2664 /* read the comment in __queue_work() */
2665 local_irq_save(irq_flags);
2666
2667 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2668 !clear_pending_if_disabled(work)) {
2669 __queue_delayed_work(cpu, wq, dwork, delay);
2670 ret = true;
2671 }
2672
2673 local_irq_restore(irq_flags);
2674 return ret;
2675}
2676EXPORT_SYMBOL(queue_delayed_work_on);
2677
2678/**
2679 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2680 * @cpu: CPU number to execute work on
2681 * @wq: workqueue to use
2682 * @dwork: work to queue
2683 * @delay: number of jiffies to wait before queueing
2684 *
2685 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2686 * modify @dwork's timer so that it expires after @delay. If @delay is
2687 * zero, @work is guaranteed to be scheduled immediately regardless of its
2688 * current state.
2689 *
2690 * Return: %false if @dwork was idle and queued, %true if @dwork was
2691 * pending and its timer was modified.
2692 *
2693 * This function is safe to call from any context including IRQ handler.
2694 * See try_to_grab_pending() for details.
2695 */
2696bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2697 struct delayed_work *dwork, unsigned long delay)
2698{
2699 unsigned long irq_flags;
2700 bool ret;
2701
2702 ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags);
2703
2704 if (!clear_pending_if_disabled(&dwork->work))
2705 __queue_delayed_work(cpu, wq, dwork, delay);
2706
2707 local_irq_restore(irq_flags);
2708 return ret;
2709}
2710EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2711
2712static void rcu_work_rcufn(struct rcu_head *rcu)
2713{
2714 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2715
2716 /* read the comment in __queue_work() */
2717 local_irq_disable();
2718 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2719 local_irq_enable();
2720}
2721
2722/**
2723 * queue_rcu_work - queue work after a RCU grace period
2724 * @wq: workqueue to use
2725 * @rwork: work to queue
2726 *
2727 * Return: %false if @rwork was already pending, %true otherwise. Note
2728 * that a full RCU grace period is guaranteed only after a %true return.
2729 * While @rwork is guaranteed to be executed after a %false return, the
2730 * execution may happen before a full RCU grace period has passed.
2731 */
2732bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2733{
2734 struct work_struct *work = &rwork->work;
2735
2736 /*
2737 * rcu_work can't be canceled or disabled. Warn if the user reached
2738 * inside @rwork and disabled the inner work.
2739 */
2740 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2741 !WARN_ON_ONCE(clear_pending_if_disabled(work))) {
2742 rwork->wq = wq;
2743 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2744 return true;
2745 }
2746
2747 return false;
2748}
2749EXPORT_SYMBOL(queue_rcu_work);
2750
2751static struct worker *alloc_worker(int node)
2752{
2753 struct worker *worker;
2754
2755 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2756 if (worker) {
2757 INIT_LIST_HEAD(&worker->entry);
2758 INIT_LIST_HEAD(&worker->scheduled);
2759 INIT_LIST_HEAD(&worker->node);
2760 /* on creation a worker is in !idle && prep state */
2761 worker->flags = WORKER_PREP;
2762 }
2763 return worker;
2764}
2765
2766static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2767{
2768 if (!is_percpu_pool(pool) && pool->attrs->affn_strict)
2769 return pool->attrs->__pod_cpumask;
2770 else
2771 return pool->attrs->cpumask;
2772}
2773
2774/**
2775 * worker_attach_to_pool() - attach a worker to a pool
2776 * @worker: worker to be attached
2777 * @pool: the target pool
2778 *
2779 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2780 * cpu-binding of @worker are kept coordinated with the pool across
2781 * cpu-[un]hotplugs.
2782 */
2783static void worker_attach_to_pool(struct worker *worker,
2784 struct worker_pool *pool)
2785{
2786 mutex_lock(&wq_pool_attach_mutex);
2787
2788 /*
2789 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable
2790 * across this function. See the comments above the flag definition for
2791 * details. BH workers are, while per-CPU, always DISASSOCIATED.
2792 */
2793 if (pool->flags & POOL_DISASSOCIATED) {
2794 worker->flags |= WORKER_UNBOUND;
2795 } else {
2796 WARN_ON_ONCE(pool->flags & POOL_BH);
2797 kthread_set_per_cpu(worker->task, pool->cpu);
2798 }
2799
2800 if (worker->rescue_wq)
2801 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2802
2803 list_add_tail(&worker->node, &pool->workers);
2804 worker->pool = pool;
2805
2806 mutex_unlock(&wq_pool_attach_mutex);
2807}
2808
2809static void unbind_worker(struct worker *worker)
2810{
2811 lockdep_assert_held(&wq_pool_attach_mutex);
2812
2813 kthread_set_per_cpu(worker->task, -1);
2814 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2815 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2816 else
2817 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2818}
2819
2820
2821static void detach_worker(struct worker *worker)
2822{
2823 lockdep_assert_held(&wq_pool_attach_mutex);
2824
2825 unbind_worker(worker);
2826 list_del(&worker->node);
2827}
2828
2829/**
2830 * worker_detach_from_pool() - detach a worker from its pool
2831 * @worker: worker which is attached to its pool
2832 *
2833 * Undo the attaching which had been done in worker_attach_to_pool(). The
2834 * caller worker shouldn't access to the pool after detached except it has
2835 * other reference to the pool.
2836 */
2837static void worker_detach_from_pool(struct worker *worker)
2838{
2839 struct worker_pool *pool = worker->pool;
2840
2841 /* there is one permanent BH worker per CPU which should never detach */
2842 WARN_ON_ONCE(pool->flags & POOL_BH);
2843
2844 mutex_lock(&wq_pool_attach_mutex);
2845 detach_worker(worker);
2846 worker->pool = NULL;
2847 mutex_unlock(&wq_pool_attach_mutex);
2848
2849 /* clear leftover flags without pool->lock after it is detached */
2850 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2851}
2852
2853static int format_worker_id(char *buf, size_t size, struct worker *worker,
2854 struct worker_pool *pool)
2855{
2856 if (worker->rescue_wq)
2857 return scnprintf(buf, size, "kworker/R-%s",
2858 worker->rescue_wq->name);
2859
2860 if (pool) {
2861 if (pool->cpu >= 0)
2862 return scnprintf(buf, size, "kworker/%d:%d%s",
2863 pool->cpu, worker->id,
2864 pool->attrs->nice < 0 ? "H" : "");
2865 else
2866 return scnprintf(buf, size, "kworker/u%d:%d",
2867 pool->id, worker->id);
2868 } else {
2869 return scnprintf(buf, size, "kworker/dying");
2870 }
2871}
2872
2873/**
2874 * create_worker - create a new workqueue worker
2875 * @pool: pool the new worker will belong to
2876 *
2877 * Create and start a new worker which is attached to @pool.
2878 *
2879 * CONTEXT:
2880 * Might sleep. Does GFP_KERNEL allocations.
2881 *
2882 * Return:
2883 * Pointer to the newly created worker.
2884 */
2885static struct worker *create_worker(struct worker_pool *pool)
2886{
2887 struct worker *worker;
2888 int id;
2889
2890 /* ID is needed to determine kthread name */
2891 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2892 if (id < 0) {
2893 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2894 ERR_PTR(id));
2895 return NULL;
2896 }
2897
2898 worker = alloc_worker(pool->node);
2899 if (!worker) {
2900 pr_err_once("workqueue: Failed to allocate a worker\n");
2901 goto fail;
2902 }
2903
2904 worker->id = id;
2905
2906 if (!(pool->flags & POOL_BH)) {
2907 char id_buf[WORKER_ID_LEN];
2908
2909 format_worker_id(id_buf, sizeof(id_buf), worker, pool);
2910 worker->task = kthread_create_on_node(worker_thread, worker,
2911 pool->node, "%s", id_buf);
2912 if (IS_ERR(worker->task)) {
2913 if (PTR_ERR(worker->task) == -EINTR) {
2914 pr_err("workqueue: Interrupted when creating a worker thread \"%s\"\n",
2915 id_buf);
2916 } else {
2917 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2918 worker->task);
2919 }
2920 goto fail;
2921 }
2922
2923 set_user_nice(worker->task, pool->attrs->nice);
2924 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2925 }
2926
2927 /* successful, attach the worker to the pool */
2928 worker_attach_to_pool(worker, pool);
2929
2930 /* start the newly created worker */
2931 raw_spin_lock_irq(&pool->lock);
2932
2933 worker->pool->nr_workers++;
2934 worker_enter_idle(worker);
2935
2936 /*
2937 * @worker is waiting on a completion in kthread() and will trigger hung
2938 * check if not woken up soon. As kick_pool() is noop if @pool is empty,
2939 * wake it up explicitly.
2940 */
2941 if (worker->task)
2942 wake_up_process(worker->task);
2943
2944 raw_spin_unlock_irq(&pool->lock);
2945
2946 return worker;
2947
2948fail:
2949 ida_free(&pool->worker_ida, id);
2950 kfree(worker);
2951 return NULL;
2952}
2953
2954static void detach_dying_workers(struct list_head *cull_list)
2955{
2956 struct worker *worker;
2957
2958 list_for_each_entry(worker, cull_list, entry)
2959 detach_worker(worker);
2960}
2961
2962static void reap_dying_workers(struct list_head *cull_list)
2963{
2964 struct worker *worker, *tmp;
2965
2966 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2967 list_del_init(&worker->entry);
2968 kthread_stop_put(worker->task);
2969 kfree(worker);
2970 }
2971}
2972
2973/**
2974 * set_worker_dying - Tag a worker for destruction
2975 * @worker: worker to be destroyed
2976 * @list: transfer worker away from its pool->idle_list and into list
2977 *
2978 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2979 * should be idle.
2980 *
2981 * CONTEXT:
2982 * raw_spin_lock_irq(pool->lock).
2983 */
2984static void set_worker_dying(struct worker *worker, struct list_head *list)
2985{
2986 struct worker_pool *pool = worker->pool;
2987
2988 lockdep_assert_held(&pool->lock);
2989 lockdep_assert_held(&wq_pool_attach_mutex);
2990
2991 /* sanity check frenzy */
2992 if (WARN_ON(worker->current_work) ||
2993 WARN_ON(!list_empty(&worker->scheduled)) ||
2994 WARN_ON(!(worker->flags & WORKER_IDLE)))
2995 return;
2996
2997 pool->nr_workers--;
2998 pool->nr_idle--;
2999
3000 /*
3001 * Clear last_woken_worker if it points to this worker, so that
3002 * show_cpu_pool_busy_workers() cannot dereference a freed worker.
3003 */
3004 if (pool->last_woken_worker == worker)
3005 pool->last_woken_worker = NULL;
3006
3007 worker->flags |= WORKER_DIE;
3008
3009 list_move(&worker->entry, list);
3010
3011 /* get an extra task struct reference for later kthread_stop_put() */
3012 get_task_struct(worker->task);
3013}
3014
3015/**
3016 * idle_worker_timeout - check if some idle workers can now be deleted.
3017 * @t: The pool's idle_timer that just expired
3018 *
3019 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
3020 * worker_leave_idle(), as a worker flicking between idle and active while its
3021 * pool is at the too_many_workers() tipping point would cause too much timer
3022 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
3023 * it expire and re-evaluate things from there.
3024 */
3025static void idle_worker_timeout(struct timer_list *t)
3026{
3027 struct worker_pool *pool = timer_container_of(pool, t, idle_timer);
3028 bool do_cull = false;
3029
3030 if (work_pending(&pool->idle_cull_work))
3031 return;
3032
3033 raw_spin_lock_irq(&pool->lock);
3034
3035 if (too_many_workers(pool)) {
3036 struct worker *worker;
3037 unsigned long expires;
3038
3039 /* idle_list is kept in LIFO order, check the last one */
3040 worker = list_last_entry(&pool->idle_list, struct worker, entry);
3041 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
3042 do_cull = !time_before(jiffies, expires);
3043
3044 if (!do_cull)
3045 mod_timer(&pool->idle_timer, expires);
3046 }
3047 raw_spin_unlock_irq(&pool->lock);
3048
3049 if (do_cull)
3050 queue_work(system_dfl_wq, &pool->idle_cull_work);
3051}
3052
3053/**
3054 * idle_cull_fn - cull workers that have been idle for too long.
3055 * @work: the pool's work for handling these idle workers
3056 *
3057 * This goes through a pool's idle workers and gets rid of those that have been
3058 * idle for at least IDLE_WORKER_TIMEOUT seconds.
3059 *
3060 * We don't want to disturb isolated CPUs because of a pcpu kworker being
3061 * culled, so this also resets worker affinity. This requires a sleepable
3062 * context, hence the split between timer callback and work item.
3063 */
3064static void idle_cull_fn(struct work_struct *work)
3065{
3066 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
3067 LIST_HEAD(cull_list);
3068
3069 /*
3070 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
3071 * cannot proceed beyong set_pf_worker() in its self-destruct path.
3072 * This is required as a previously-preempted worker could run after
3073 * set_worker_dying() has happened but before detach_dying_workers() did.
3074 */
3075 mutex_lock(&wq_pool_attach_mutex);
3076 raw_spin_lock_irq(&pool->lock);
3077
3078 while (too_many_workers(pool)) {
3079 struct worker *worker;
3080 unsigned long expires;
3081
3082 worker = list_last_entry(&pool->idle_list, struct worker, entry);
3083 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
3084
3085 if (time_before(jiffies, expires)) {
3086 mod_timer(&pool->idle_timer, expires);
3087 break;
3088 }
3089
3090 set_worker_dying(worker, &cull_list);
3091 }
3092
3093 raw_spin_unlock_irq(&pool->lock);
3094 detach_dying_workers(&cull_list);
3095 mutex_unlock(&wq_pool_attach_mutex);
3096
3097 reap_dying_workers(&cull_list);
3098}
3099
3100static void send_mayday(struct pool_workqueue *pwq)
3101{
3102 struct workqueue_struct *wq = pwq->wq;
3103
3104 lockdep_assert_held(&wq_mayday_lock);
3105
3106 if (!wq->rescuer)
3107 return;
3108
3109 /* mayday mayday mayday */
3110 if (list_empty(&pwq->mayday_node)) {
3111 /*
3112 * If @pwq is for an unbound wq, its base ref may be put at
3113 * any time due to an attribute change. Pin @pwq until the
3114 * rescuer is done with it.
3115 */
3116 get_pwq(pwq);
3117 list_add_tail(&pwq->mayday_node, &wq->maydays);
3118 wake_up_process(wq->rescuer->task);
3119 pwq->stats[PWQ_STAT_MAYDAY]++;
3120 }
3121}
3122
3123static void pool_mayday_timeout(struct timer_list *t)
3124{
3125 struct worker_pool *pool = timer_container_of(pool, t, mayday_timer);
3126 struct work_struct *work;
3127
3128 raw_spin_lock_irq(&pool->lock);
3129 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
3130
3131 if (need_to_create_worker(pool)) {
3132 /*
3133 * We've been trying to create a new worker but
3134 * haven't been successful. We might be hitting an
3135 * allocation deadlock. Send distress signals to
3136 * rescuers.
3137 */
3138 list_for_each_entry(work, &pool->worklist, entry)
3139 send_mayday(get_work_pwq(work));
3140 }
3141
3142 raw_spin_unlock(&wq_mayday_lock);
3143 raw_spin_unlock_irq(&pool->lock);
3144
3145 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3146}
3147
3148/**
3149 * maybe_create_worker - create a new worker if necessary
3150 * @pool: pool to create a new worker for
3151 *
3152 * Create a new worker for @pool if necessary. @pool is guaranteed to
3153 * have at least one idle worker on return from this function. If
3154 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3155 * sent to all rescuers with works scheduled on @pool to resolve
3156 * possible allocation deadlock.
3157 *
3158 * On return, need_to_create_worker() is guaranteed to be %false and
3159 * may_start_working() %true.
3160 *
3161 * LOCKING:
3162 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3163 * multiple times. Does GFP_KERNEL allocations. Called only from
3164 * manager.
3165 */
3166static void maybe_create_worker(struct worker_pool *pool)
3167__releases(&pool->lock)
3168__acquires(&pool->lock)
3169{
3170restart:
3171 raw_spin_unlock_irq(&pool->lock);
3172
3173 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3174 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3175
3176 while (true) {
3177 if (create_worker(pool) || !need_to_create_worker(pool))
3178 break;
3179
3180 schedule_timeout_interruptible(CREATE_COOLDOWN);
3181
3182 if (!need_to_create_worker(pool))
3183 break;
3184 }
3185
3186 timer_delete_sync(&pool->mayday_timer);
3187 raw_spin_lock_irq(&pool->lock);
3188 /*
3189 * This is necessary even after a new worker was just successfully
3190 * created as @pool->lock was dropped and the new worker might have
3191 * already become busy.
3192 */
3193 if (need_to_create_worker(pool))
3194 goto restart;
3195}
3196
3197#ifdef CONFIG_PREEMPT_RT
3198static void worker_lock_callback(struct worker_pool *pool)
3199{
3200 /*
3201 * SINGLE_DEPTH_NESTING is for a dead pool's bh_worker() running from
3202 * drain_dead_softirq_workfn() inside a live pool's bh_worker(). The
3203 * unlocked read is stable: the flag is only set while @pool's CPU is
3204 * dead, inside a serialized hotplug operation. data_race() as the value
3205 * only affects the lockdep annotation and the read can be elided when
3206 * lockdep is disabled.
3207 */
3208 spin_lock_nested(&pool->cb_lock,
3209 data_race(pool->flags) & POOL_BH_DRAINING ? SINGLE_DEPTH_NESTING : 0);
3210}
3211
3212static void worker_unlock_callback(struct worker_pool *pool)
3213{
3214 spin_unlock(&pool->cb_lock);
3215}
3216
3217static void workqueue_callback_cancel_wait_running(struct worker_pool *pool)
3218{
3219 spin_lock(&pool->cb_lock);
3220 spin_unlock(&pool->cb_lock);
3221}
3222
3223#else
3224
3225static void worker_lock_callback(struct worker_pool *pool) { }
3226static void worker_unlock_callback(struct worker_pool *pool) { }
3227static void workqueue_callback_cancel_wait_running(struct worker_pool *pool) { }
3228
3229#endif
3230
3231/**
3232 * manage_workers - manage worker pool
3233 * @worker: self
3234 *
3235 * Assume the manager role and manage the worker pool @worker belongs
3236 * to. At any given time, there can be only zero or one manager per
3237 * pool. The exclusion is handled automatically by this function.
3238 *
3239 * The caller can safely start processing works on false return. On
3240 * true return, it's guaranteed that need_to_create_worker() is false
3241 * and may_start_working() is true.
3242 *
3243 * CONTEXT:
3244 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3245 * multiple times. Does GFP_KERNEL allocations.
3246 *
3247 * Return:
3248 * %false if the pool doesn't need management and the caller can safely
3249 * start processing works, %true if management function was performed and
3250 * the conditions that the caller verified before calling the function may
3251 * no longer be true.
3252 */
3253static bool manage_workers(struct worker *worker)
3254{
3255 struct worker_pool *pool = worker->pool;
3256
3257 if (pool->flags & POOL_MANAGER_ACTIVE)
3258 return false;
3259
3260 pool->flags |= POOL_MANAGER_ACTIVE;
3261 pool->manager = worker;
3262
3263 maybe_create_worker(pool);
3264
3265 pool->manager = NULL;
3266 pool->flags &= ~POOL_MANAGER_ACTIVE;
3267 rcuwait_wake_up(&manager_wait);
3268 return true;
3269}
3270
3271/**
3272 * process_one_work - process single work
3273 * @worker: self
3274 * @work: work to process
3275 *
3276 * Process @work. This function contains all the logics necessary to
3277 * process a single work including synchronization against and
3278 * interaction with other workers on the same cpu, queueing and
3279 * flushing. As long as context requirement is met, any worker can
3280 * call this function to process a work.
3281 *
3282 * CONTEXT:
3283 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3284 */
3285static void process_one_work(struct worker *worker, struct work_struct *work)
3286__releases(&pool->lock)
3287__acquires(&pool->lock)
3288{
3289 struct pool_workqueue *pwq = get_work_pwq(work);
3290 struct worker_pool *pool = worker->pool;
3291 struct task_struct *wake_task = NULL;
3292 unsigned long work_data;
3293 int lockdep_start_depth, rcu_start_depth;
3294 bool bh_draining = pool->flags & POOL_BH_DRAINING;
3295#ifdef CONFIG_LOCKDEP
3296 /*
3297 * It is permissible to free the struct work_struct from
3298 * inside the function that is called from it, this we need to
3299 * take into account for lockdep too. To avoid bogus "held
3300 * lock freed" warnings as well as problems when looking into
3301 * work->lockdep_map, make a copy and use that here.
3302 */
3303 struct lockdep_map lockdep_map;
3304
3305 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3306#endif
3307 /* ensure we're on the correct CPU */
3308 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3309 raw_smp_processor_id() != pool->cpu);
3310
3311 /* claim and dequeue */
3312 debug_work_deactivate(work);
3313 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3314 worker->current_work = work;
3315 worker->current_func = work->func;
3316 worker->current_pwq = pwq;
3317 if (worker->task)
3318 worker->current_at = READ_ONCE(worker->task->se.sum_exec_runtime);
3319 worker->current_start = jiffies;
3320 work_data = *work_data_bits(work);
3321 worker->current_color = get_work_color(work_data);
3322
3323 /*
3324 * Record wq name for cmdline and debug reporting, may get
3325 * overridden through set_worker_desc().
3326 */
3327 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3328
3329 list_del_init(&work->entry);
3330
3331 /*
3332 * CPU intensive works don't participate in concurrency management.
3333 * They're the scheduler's responsibility. This takes @worker out
3334 * of concurrency management and the next code block will chain
3335 * execution of the pending work items.
3336 */
3337 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3338 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3339
3340 /*
3341 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3342 * since nr_running would always be >= 1 at this point. This is used to
3343 * chain execution of the pending work items for WORKER_NOT_RUNNING
3344 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3345 *
3346 * Select the worker under pool->lock; the wakeup is deferred until
3347 * after the lock is dropped, guarded by the rcu_read_lock() below.
3348 */
3349 kick_pool_pick(pool, &wake_task);
3350
3351 /*
3352 * Record the last pool and clear PENDING which should be the last
3353 * update to @work. Also, do this inside @pool->lock so that
3354 * PENDING and queued state changes happen together while IRQ is
3355 * disabled.
3356 */
3357 set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool));
3358
3359 pwq->stats[PWQ_STAT_STARTED]++;
3360
3361 rcu_read_lock();
3362 raw_spin_unlock_irq(&pool->lock);
3363 if (wake_task)
3364 wake_up_process(wake_task);
3365 rcu_read_unlock();
3366
3367 rcu_start_depth = rcu_preempt_depth();
3368 lockdep_start_depth = lockdep_depth(current);
3369 /* see drain_dead_softirq_workfn() */
3370 if (!bh_draining)
3371 lock_map_acquire(pwq->wq->lockdep_map);
3372 lock_map_acquire(&lockdep_map);
3373 /*
3374 * Strictly speaking we should mark the invariant state without holding
3375 * any locks, that is, before these two lock_map_acquire()'s.
3376 *
3377 * However, that would result in:
3378 *
3379 * A(W1)
3380 * WFC(C)
3381 * A(W1)
3382 * C(C)
3383 *
3384 * Which would create W1->C->W1 dependencies, even though there is no
3385 * actual deadlock possible. There are two solutions, using a
3386 * read-recursive acquire on the work(queue) 'locks', but this will then
3387 * hit the lockdep limitation on recursive locks, or simply discard
3388 * these locks.
3389 *
3390 * AFAICT there is no possible deadlock scenario between the
3391 * flush_work() and complete() primitives (except for single-threaded
3392 * workqueues), so hiding them isn't a problem.
3393 */
3394 lockdep_invariant_state(true);
3395 trace_workqueue_execute_start(work);
3396 worker->current_func(work);
3397 /*
3398 * While we must be careful to not use "work" after this, the trace
3399 * point will only record its address.
3400 */
3401 trace_workqueue_execute_end(work, worker->current_func);
3402
3403 lock_map_release(&lockdep_map);
3404 if (!bh_draining)
3405 lock_map_release(pwq->wq->lockdep_map);
3406
3407 if (unlikely((worker->task && in_atomic()) ||
3408 lockdep_depth(current) != lockdep_start_depth ||
3409 rcu_preempt_depth() != rcu_start_depth)) {
3410 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n"
3411 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n",
3412 current->comm, task_pid_nr(current), preempt_count(),
3413 lockdep_start_depth, lockdep_depth(current),
3414 rcu_start_depth, rcu_preempt_depth(),
3415 worker->current_func);
3416 debug_show_held_locks(current);
3417 dump_stack();
3418 }
3419
3420 /*
3421 * The following prevents a kworker from hogging CPU on !PREEMPTION
3422 * kernels, where a requeueing work item waiting for something to
3423 * happen could deadlock with stop_machine as such work item could
3424 * indefinitely requeue itself while all other CPUs are trapped in
3425 * stop_machine. At the same time, report a quiescent RCU state so
3426 * the same condition doesn't freeze RCU.
3427 */
3428 if (worker->task)
3429 cond_resched();
3430
3431 raw_spin_lock_irq(&pool->lock);
3432
3433 pwq->stats[PWQ_STAT_COMPLETED]++;
3434
3435 /*
3436 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3437 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3438 * wq_cpu_intensive_thresh_us. Clear it.
3439 */
3440 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3441
3442 /* tag the worker for identification in schedule() */
3443 worker->last_func = worker->current_func;
3444
3445 /* we're done with it, release */
3446 hash_del(&worker->hentry);
3447 worker->current_work = NULL;
3448 worker->current_func = NULL;
3449 worker->current_pwq = NULL;
3450 worker->current_color = INT_MAX;
3451
3452 /* must be the last step, see the function comment */
3453 pwq_dec_nr_in_flight(pwq, work_data);
3454}
3455
3456/**
3457 * process_scheduled_works - process scheduled works
3458 * @worker: self
3459 *
3460 * Process all scheduled works. Please note that the scheduled list
3461 * may change while processing a work, so this function repeatedly
3462 * fetches a work from the top and executes it.
3463 *
3464 * CONTEXT:
3465 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3466 * multiple times.
3467 */
3468static void process_scheduled_works(struct worker *worker)
3469{
3470 struct work_struct *work;
3471 bool first = true;
3472
3473 while ((work = list_first_entry_or_null(&worker->scheduled,
3474 struct work_struct, entry))) {
3475 if (first) {
3476 worker->pool->last_progress_ts = jiffies;
3477 first = false;
3478 }
3479 process_one_work(worker, work);
3480 }
3481}
3482
3483static void set_pf_worker(bool val)
3484{
3485 mutex_lock(&wq_pool_attach_mutex);
3486 if (val)
3487 current->flags |= PF_WQ_WORKER;
3488 else
3489 current->flags &= ~PF_WQ_WORKER;
3490 mutex_unlock(&wq_pool_attach_mutex);
3491}
3492
3493/**
3494 * worker_thread - the worker thread function
3495 * @__worker: self
3496 *
3497 * The worker thread function. All workers belong to a worker_pool -
3498 * either a per-cpu one or dynamic unbound one. These workers process all
3499 * work items regardless of their specific target workqueue. The only
3500 * exception is work items which belong to workqueues with a rescuer which
3501 * will be explained in rescuer_thread().
3502 *
3503 * Return: 0
3504 */
3505static int worker_thread(void *__worker)
3506{
3507 struct worker *worker = __worker;
3508 struct worker_pool *pool = worker->pool;
3509
3510 /* tell the scheduler that this is a workqueue worker */
3511 set_pf_worker(true);
3512woke_up:
3513 raw_spin_lock_irq(&pool->lock);
3514
3515 /* am I supposed to die? */
3516 if (unlikely(worker->flags & WORKER_DIE)) {
3517 raw_spin_unlock_irq(&pool->lock);
3518 set_pf_worker(false);
3519 /*
3520 * The worker is dead and PF_WQ_WORKER is cleared, worker->pool
3521 * shouldn't be accessed, reset it to NULL in case otherwise.
3522 */
3523 worker->pool = NULL;
3524 ida_free(&pool->worker_ida, worker->id);
3525 return 0;
3526 }
3527
3528 worker_leave_idle(worker);
3529recheck:
3530 /* no more worker necessary? */
3531 if (!need_more_worker(pool))
3532 goto sleep;
3533
3534 /* do we need to manage? */
3535 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3536 goto recheck;
3537
3538 /*
3539 * ->scheduled list can only be filled while a worker is
3540 * preparing to process a work or actually processing it.
3541 * Make sure nobody diddled with it while I was sleeping.
3542 */
3543 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3544
3545 /*
3546 * Finish PREP stage. We're guaranteed to have at least one idle
3547 * worker or that someone else has already assumed the manager
3548 * role. This is where @worker starts participating in concurrency
3549 * management if applicable and concurrency management is restored
3550 * after being rebound. See rebind_workers() for details.
3551 */
3552 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3553
3554 do {
3555 struct work_struct *work =
3556 list_first_entry(&pool->worklist,
3557 struct work_struct, entry);
3558
3559 if (assign_work(work, worker, NULL))
3560 process_scheduled_works(worker);
3561 } while (keep_working(pool));
3562
3563 worker_set_flags(worker, WORKER_PREP);
3564sleep:
3565 /*
3566 * pool->lock is held and there's no work to process and no need to
3567 * manage, sleep. Workers are woken up only while holding
3568 * pool->lock or from local cpu, so setting the current state
3569 * before releasing pool->lock is enough to prevent losing any
3570 * event.
3571 */
3572 worker_enter_idle(worker);
3573 __set_current_state(TASK_IDLE);
3574 raw_spin_unlock_irq(&pool->lock);
3575 schedule();
3576 goto woke_up;
3577}
3578
3579static bool assign_rescuer_work(struct pool_workqueue *pwq, struct worker *rescuer)
3580{
3581 struct worker_pool *pool = pwq->pool;
3582 struct work_struct *cursor = &pwq->mayday_cursor;
3583 struct work_struct *work, *n;
3584
3585 /* have work items to rescue? */
3586 if (!pwq->nr_active)
3587 return false;
3588
3589 /* need rescue? */
3590 if (!need_to_create_worker(pool)) {
3591 /*
3592 * The pool has idle workers and doesn't need the rescuer, so it
3593 * could simply return false here.
3594 *
3595 * However, the memory pressure might not be fully relieved.
3596 * In PERCPU pool with concurrency enabled, having idle workers
3597 * does not necessarily mean memory pressure is gone; it may
3598 * simply mean regular workers have woken up, completed their
3599 * work, and gone idle again due to concurrency limits.
3600 *
3601 * In this case, those working workers may later sleep again,
3602 * the pool may run out of idle workers, and it will have to
3603 * allocate new ones and wait for the timer to send mayday,
3604 * causing unnecessary delay - especially if memory pressure
3605 * was never resolved throughout.
3606 *
3607 * Do more work if memory pressure is still on to reduce
3608 * relapse, using (pool->flags & POOL_MANAGER_ACTIVE), though
3609 * not precisely, unless there are other PWQs needing help.
3610 */
3611 if (!(pool->flags & POOL_MANAGER_ACTIVE) ||
3612 !list_empty(&pwq->wq->maydays))
3613 return false;
3614 }
3615
3616 /* search from the start or cursor if available */
3617 if (list_empty(&cursor->entry))
3618 work = list_first_entry(&pool->worklist, struct work_struct, entry);
3619 else
3620 work = list_next_entry(cursor, entry);
3621
3622 /* find the next work item to rescue */
3623 list_for_each_entry_safe_from(work, n, &pool->worklist, entry) {
3624 if (get_work_pwq(work) == pwq && assign_work(work, rescuer, &n)) {
3625 pwq->stats[PWQ_STAT_RESCUED]++;
3626 /* put the cursor for next search */
3627 list_move_tail(&cursor->entry, &n->entry);
3628 return true;
3629 }
3630 }
3631
3632 return false;
3633}
3634
3635/**
3636 * rescuer_thread - the rescuer thread function
3637 * @__rescuer: self
3638 *
3639 * Workqueue rescuer thread function. There's one rescuer for each
3640 * workqueue which has WQ_MEM_RECLAIM set.
3641 *
3642 * Regular work processing on a pool may block trying to create a new
3643 * worker which uses GFP_KERNEL allocation which has slight chance of
3644 * developing into deadlock if some works currently on the same queue
3645 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3646 * the problem rescuer solves.
3647 *
3648 * When such condition is possible, the pool summons rescuers of all
3649 * workqueues which have works queued on the pool and let them process
3650 * those works so that forward progress can be guaranteed.
3651 *
3652 * This should happen rarely.
3653 *
3654 * Return: 0
3655 */
3656static int rescuer_thread(void *__rescuer)
3657{
3658 struct worker *rescuer = __rescuer;
3659 struct workqueue_struct *wq = rescuer->rescue_wq;
3660 bool should_stop;
3661
3662 set_user_nice(current, RESCUER_NICE_LEVEL);
3663
3664 /*
3665 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3666 * doesn't participate in concurrency management.
3667 */
3668 set_pf_worker(true);
3669repeat:
3670 set_current_state(TASK_IDLE);
3671
3672 /*
3673 * By the time the rescuer is requested to stop, the workqueue
3674 * shouldn't have any work pending, but @wq->maydays may still have
3675 * pwq(s) queued. This can happen by non-rescuer workers consuming
3676 * all the work items before the rescuer got to them. Go through
3677 * @wq->maydays processing before acting on should_stop so that the
3678 * list is always empty on exit.
3679 */
3680 should_stop = kthread_should_stop();
3681
3682 /* see whether any pwq is asking for help */
3683 raw_spin_lock_irq(&wq_mayday_lock);
3684
3685 while (!list_empty(&wq->maydays)) {
3686 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3687 struct pool_workqueue, mayday_node);
3688 struct worker_pool *pool = pwq->pool;
3689 unsigned int count = 0;
3690
3691 __set_current_state(TASK_RUNNING);
3692 list_del_init(&pwq->mayday_node);
3693
3694 raw_spin_unlock_irq(&wq_mayday_lock);
3695
3696 worker_attach_to_pool(rescuer, pool);
3697
3698 raw_spin_lock_irq(&pool->lock);
3699
3700 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3701
3702 while (assign_rescuer_work(pwq, rescuer)) {
3703 process_scheduled_works(rescuer);
3704
3705 /*
3706 * If the per-turn work item limit is reached and other
3707 * PWQs are in mayday, requeue mayday for this PWQ and
3708 * let the rescuer handle the other PWQs first.
3709 */
3710 if (++count > RESCUER_BATCH && !list_empty(&pwq->wq->maydays) &&
3711 pwq->nr_active && need_to_create_worker(pool)) {
3712 raw_spin_lock(&wq_mayday_lock);
3713 send_mayday(pwq);
3714 raw_spin_unlock(&wq_mayday_lock);
3715 break;
3716 }
3717 }
3718
3719 /* The cursor can not be left behind without the rescuer watching it. */
3720 if (!list_empty(&pwq->mayday_cursor.entry) && list_empty(&pwq->mayday_node))
3721 list_del_init(&pwq->mayday_cursor.entry);
3722
3723 /*
3724 * Leave this pool. Notify regular workers; otherwise, we end up
3725 * with 0 concurrency and stalling the execution.
3726 */
3727 kick_pool(pool);
3728
3729 raw_spin_unlock_irq(&pool->lock);
3730
3731 worker_detach_from_pool(rescuer);
3732
3733 /*
3734 * Put the reference grabbed by send_mayday(). @pool might
3735 * go away any time after it.
3736 */
3737 put_pwq_unlocked(pwq);
3738
3739 raw_spin_lock_irq(&wq_mayday_lock);
3740 }
3741
3742 raw_spin_unlock_irq(&wq_mayday_lock);
3743
3744 if (should_stop) {
3745 __set_current_state(TASK_RUNNING);
3746 set_pf_worker(false);
3747 return 0;
3748 }
3749
3750 /* rescuers should never participate in concurrency management */
3751 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3752 schedule();
3753 goto repeat;
3754}
3755
3756static void bh_worker(struct worker *worker)
3757{
3758 struct worker_pool *pool = worker->pool;
3759 int nr_restarts = BH_WORKER_RESTARTS;
3760 unsigned long end = jiffies + BH_WORKER_JIFFIES;
3761
3762 worker_lock_callback(pool);
3763 raw_spin_lock_irq(&pool->lock);
3764 worker_leave_idle(worker);
3765
3766 /*
3767 * This function follows the structure of worker_thread(). See there for
3768 * explanations on each step.
3769 */
3770 if (!need_more_worker(pool))
3771 goto done;
3772
3773 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3774 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3775
3776 do {
3777 struct work_struct *work =
3778 list_first_entry(&pool->worklist,
3779 struct work_struct, entry);
3780
3781 if (assign_work(work, worker, NULL))
3782 process_scheduled_works(worker);
3783 } while (keep_working(pool) &&
3784 --nr_restarts && time_before(jiffies, end));
3785
3786 worker_set_flags(worker, WORKER_PREP);
3787done:
3788 worker_enter_idle(worker);
3789 kick_pool(pool);
3790 raw_spin_unlock_irq(&pool->lock);
3791 worker_unlock_callback(pool);
3792}
3793
3794/*
3795 * TODO: Convert all tasklet users to workqueue and use softirq directly.
3796 *
3797 * This is currently called from tasklet[_hi]action() and thus is also called
3798 * whenever there are tasklets to run. Let's do an early exit if there's nothing
3799 * queued. Once conversion from tasklet is complete, the need_more_worker() test
3800 * can be dropped.
3801 *
3802 * After full conversion, we'll add worker->softirq_action, directly use the
3803 * softirq action and obtain the worker pointer from the softirq_action pointer.
3804 */
3805void workqueue_softirq_action(bool highpri)
3806{
3807 struct worker_pool *pool =
3808 &per_cpu(bh_worker_pools, smp_processor_id())[highpri];
3809 if (need_more_worker(pool))
3810 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3811}
3812
3813struct wq_drain_dead_softirq_work {
3814 struct work_struct work;
3815 struct worker_pool *pool;
3816 struct completion done;
3817};
3818
3819static void drain_dead_softirq_workfn(struct work_struct *work)
3820{
3821 struct wq_drain_dead_softirq_work *dead_work =
3822 container_of(work, struct wq_drain_dead_softirq_work, work);
3823 struct worker_pool *pool = dead_work->pool;
3824 bool repeat;
3825
3826 /*
3827 * @pool's CPU is dead and we want to execute its still pending work
3828 * items from this BH work item which is running on a different CPU. As
3829 * its CPU is dead, @pool can't be kicked and, as work execution path
3830 * will be nested, a lockdep annotation needs to be suppressed. Mark
3831 * @pool with %POOL_BH_DRAINING for the special treatments.
3832 */
3833 raw_spin_lock_irq(&pool->lock);
3834 pool->flags |= POOL_BH_DRAINING;
3835 raw_spin_unlock_irq(&pool->lock);
3836
3837 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3838
3839 raw_spin_lock_irq(&pool->lock);
3840 pool->flags &= ~POOL_BH_DRAINING;
3841 repeat = need_more_worker(pool);
3842 raw_spin_unlock_irq(&pool->lock);
3843
3844 /*
3845 * bh_worker() might hit consecutive execution limit and bail. If there
3846 * still are pending work items, reschedule self and return so that we
3847 * don't hog this CPU's BH.
3848 */
3849 if (repeat) {
3850 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3851 queue_work(system_bh_highpri_wq, work);
3852 else
3853 queue_work(system_bh_wq, work);
3854 } else {
3855 complete(&dead_work->done);
3856 }
3857}
3858
3859/*
3860 * @cpu is dead. Drain the remaining BH work items on the current CPU. It's
3861 * possible to allocate dead_work per CPU and avoid flushing. However, then we
3862 * have to worry about draining overlapping with CPU coming back online or
3863 * nesting (one CPU's dead_work queued on another CPU which is also dead and so
3864 * on). Let's keep it simple and drain them synchronously. These are BH work
3865 * items which shouldn't be requeued on the same pool. Shouldn't take long.
3866 */
3867void workqueue_softirq_dead(unsigned int cpu)
3868{
3869 int i;
3870
3871 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
3872 struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i];
3873 struct wq_drain_dead_softirq_work dead_work;
3874
3875 if (!need_more_worker(pool))
3876 continue;
3877
3878 INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn);
3879 dead_work.pool = pool;
3880 init_completion(&dead_work.done);
3881
3882 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3883 queue_work(system_bh_highpri_wq, &dead_work.work);
3884 else
3885 queue_work(system_bh_wq, &dead_work.work);
3886
3887 wait_for_completion(&dead_work.done);
3888 destroy_work_on_stack(&dead_work.work);
3889 }
3890}
3891
3892/**
3893 * check_flush_dependency - check for flush dependency sanity
3894 * @target_wq: workqueue being flushed
3895 * @target_work: work item being flushed (NULL for workqueue flushes)
3896 * @from_cancel: are we called from the work cancel path
3897 *
3898 * %current is trying to flush the whole @target_wq or @target_work on it.
3899 * If this is not the cancel path (which implies work being flushed is either
3900 * already running, or will not be at all), check if @target_wq doesn't have
3901 * %WQ_MEM_RECLAIM and verify that %current is not reclaiming memory or running
3902 * on a workqueue which doesn't have %WQ_MEM_RECLAIM as that can break forward-
3903 * progress guarantee leading to a deadlock.
3904 */
3905static void check_flush_dependency(struct workqueue_struct *target_wq,
3906 struct work_struct *target_work,
3907 bool from_cancel)
3908{
3909 work_func_t target_func;
3910 struct worker *worker;
3911
3912 if (from_cancel || target_wq->flags & WQ_MEM_RECLAIM)
3913 return;
3914
3915 worker = current_wq_worker();
3916 target_func = target_work ? target_work->func : NULL;
3917
3918 WARN_ONCE(current->flags & PF_MEMALLOC,
3919 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3920 current->pid, current->comm, target_wq->name, target_func);
3921 WARN_ONCE(worker && worker->current_pwq && ((worker->current_pwq->wq->flags &
3922 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3923 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3924 worker->current_pwq->wq->name, worker->current_func,
3925 target_wq->name, target_func);
3926}
3927
3928struct wq_barrier {
3929 struct work_struct work;
3930 struct completion done;
3931 struct task_struct *task; /* purely informational */
3932};
3933
3934static void wq_barrier_func(struct work_struct *work)
3935{
3936 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3937 complete(&barr->done);
3938}
3939
3940/**
3941 * insert_wq_barrier - insert a barrier work
3942 * @pwq: pwq to insert barrier into
3943 * @barr: wq_barrier to insert
3944 * @target: target work to attach @barr to
3945 * @worker: worker currently executing @target, NULL if @target is not executing
3946 *
3947 * @barr is linked to @target such that @barr is completed only after
3948 * @target finishes execution. Please note that the ordering
3949 * guarantee is observed only with respect to @target and on the local
3950 * cpu.
3951 *
3952 * Currently, a queued barrier can't be canceled. This is because
3953 * try_to_grab_pending() can't determine whether the work to be
3954 * grabbed is at the head of the queue and thus can't clear LINKED
3955 * flag of the previous work while there must be a valid next work
3956 * after a work with LINKED flag set.
3957 *
3958 * Note that when @worker is non-NULL, @target may be modified
3959 * underneath us, so we can't reliably determine pwq from @target.
3960 *
3961 * CONTEXT:
3962 * raw_spin_lock_irq(pool->lock).
3963 */
3964static void insert_wq_barrier(struct pool_workqueue *pwq,
3965 struct wq_barrier *barr,
3966 struct work_struct *target, struct worker *worker)
3967{
3968 static __maybe_unused struct lock_class_key bh_key, thr_key;
3969 unsigned int work_flags = 0;
3970 unsigned int work_color;
3971 struct list_head *head;
3972
3973 /*
3974 * debugobject calls are safe here even with pool->lock locked
3975 * as we know for sure that this will not trigger any of the
3976 * checks and call back into the fixup functions where we
3977 * might deadlock.
3978 *
3979 * BH and threaded workqueues need separate lockdep keys to avoid
3980 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W}
3981 * usage".
3982 */
3983 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func,
3984 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key);
3985 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3986
3987 init_completion_map(&barr->done, &target->lockdep_map);
3988
3989 barr->task = current;
3990
3991 /* The barrier work item does not participate in nr_active. */
3992 work_flags |= WORK_STRUCT_INACTIVE;
3993
3994 /*
3995 * If @target is currently being executed, schedule the
3996 * barrier to the worker; otherwise, put it after @target.
3997 */
3998 if (worker) {
3999 head = worker->scheduled.next;
4000 work_color = worker->current_color;
4001 } else {
4002 unsigned long *bits = work_data_bits(target);
4003
4004 head = target->entry.next;
4005 /* there can already be other linked works, inherit and set */
4006 work_flags |= *bits & WORK_STRUCT_LINKED;
4007 work_color = get_work_color(*bits);
4008 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
4009 }
4010
4011 pwq->nr_in_flight[work_color]++;
4012 work_flags |= work_color_to_flags(work_color);
4013
4014 insert_work(pwq, &barr->work, head, work_flags);
4015}
4016
4017/**
4018 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
4019 * @wq: workqueue being flushed
4020 * @flush_color: new flush color, < 0 for no-op
4021 * @work_color: new work color, < 0 for no-op
4022 *
4023 * Prepare pwqs for workqueue flushing.
4024 *
4025 * If @flush_color is non-negative, flush_color on all pwqs should be
4026 * -1. If no pwq has in-flight commands at the specified color, all
4027 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
4028 * has in flight commands, its pwq->flush_color is set to
4029 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
4030 * wakeup logic is armed and %true is returned.
4031 *
4032 * The caller should have initialized @wq->first_flusher prior to
4033 * calling this function with non-negative @flush_color. If
4034 * @flush_color is negative, no flush color update is done and %false
4035 * is returned.
4036 *
4037 * If @work_color is non-negative, all pwqs should have the same
4038 * work_color which is previous to @work_color and all will be
4039 * advanced to @work_color.
4040 *
4041 * CONTEXT:
4042 * mutex_lock(wq->mutex).
4043 *
4044 * Return:
4045 * %true if @flush_color >= 0 and there's something to flush. %false
4046 * otherwise.
4047 */
4048static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
4049 int flush_color, int work_color)
4050{
4051 bool wait = false;
4052 struct pool_workqueue *pwq;
4053 struct worker_pool *current_pool = NULL;
4054
4055 if (flush_color >= 0) {
4056 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
4057 atomic_set(&wq->nr_pwqs_to_flush, 1);
4058 }
4059
4060 /*
4061 * For unbound workqueue, pwqs will map to only a few pools.
4062 * Most of the time, pwqs within the same pool will be linked
4063 * sequentially to wq->pwqs by cpu index. So in the majority
4064 * of pwq iters, the pool is the same, only doing lock/unlock
4065 * if the pool has changed. This can largely reduce expensive
4066 * lock operations.
4067 */
4068 for_each_pwq(pwq, wq) {
4069 if (current_pool != pwq->pool) {
4070 if (likely(current_pool))
4071 raw_spin_unlock_irq(&current_pool->lock);
4072 current_pool = pwq->pool;
4073 raw_spin_lock_irq(&current_pool->lock);
4074 }
4075
4076 if (flush_color >= 0) {
4077 WARN_ON_ONCE(pwq->flush_color != -1);
4078
4079 if (pwq->nr_in_flight[flush_color]) {
4080 pwq->flush_color = flush_color;
4081 atomic_inc(&wq->nr_pwqs_to_flush);
4082 wait = true;
4083 }
4084 }
4085
4086 if (work_color >= 0) {
4087 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
4088 pwq->work_color = work_color;
4089 }
4090
4091 }
4092
4093 if (current_pool)
4094 raw_spin_unlock_irq(&current_pool->lock);
4095
4096 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
4097 complete(&wq->first_flusher->done);
4098
4099 return wait;
4100}
4101
4102static void touch_wq_lockdep_map(struct workqueue_struct *wq)
4103{
4104#ifdef CONFIG_LOCKDEP
4105 if (unlikely(!wq->lockdep_map))
4106 return;
4107
4108 if (wq->flags & WQ_BH)
4109 local_bh_disable();
4110
4111 lock_map_acquire(wq->lockdep_map);
4112 lock_map_release(wq->lockdep_map);
4113
4114 if (wq->flags & WQ_BH)
4115 local_bh_enable();
4116#endif
4117}
4118
4119static void touch_work_lockdep_map(struct work_struct *work,
4120 struct workqueue_struct *wq)
4121{
4122#ifdef CONFIG_LOCKDEP
4123 if (wq->flags & WQ_BH)
4124 local_bh_disable();
4125
4126 lock_map_acquire(&work->lockdep_map);
4127 lock_map_release(&work->lockdep_map);
4128
4129 if (wq->flags & WQ_BH)
4130 local_bh_enable();
4131#endif
4132}
4133
4134/**
4135 * __flush_workqueue - ensure that any scheduled work has run to completion.
4136 * @wq: workqueue to flush
4137 *
4138 * This function sleeps until all work items which were queued on entry
4139 * have finished execution, but it is not livelocked by new incoming ones.
4140 */
4141void __flush_workqueue(struct workqueue_struct *wq)
4142{
4143 struct wq_flusher this_flusher = {
4144 .list = LIST_HEAD_INIT(this_flusher.list),
4145 .flush_color = -1,
4146 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, (*wq->lockdep_map)),
4147 };
4148 int next_color;
4149
4150 if (WARN_ON(!wq_online))
4151 return;
4152
4153 touch_wq_lockdep_map(wq);
4154
4155 mutex_lock(&wq->mutex);
4156
4157 /*
4158 * Start-to-wait phase
4159 */
4160 next_color = work_next_color(wq->work_color);
4161
4162 if (next_color != wq->flush_color) {
4163 /*
4164 * Color space is not full. The current work_color
4165 * becomes our flush_color and work_color is advanced
4166 * by one.
4167 */
4168 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
4169 this_flusher.flush_color = wq->work_color;
4170 wq->work_color = next_color;
4171
4172 if (!wq->first_flusher) {
4173 /* no flush in progress, become the first flusher */
4174 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
4175
4176 wq->first_flusher = &this_flusher;
4177
4178 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
4179 wq->work_color)) {
4180 /* nothing to flush, done */
4181 wq->flush_color = next_color;
4182 wq->first_flusher = NULL;
4183 goto out_unlock;
4184 }
4185 } else {
4186 /* wait in queue */
4187 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
4188 list_add_tail(&this_flusher.list, &wq->flusher_queue);
4189 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4190 }
4191 } else {
4192 /*
4193 * Oops, color space is full, wait on overflow queue.
4194 * The next flush completion will assign us
4195 * flush_color and transfer to flusher_queue.
4196 */
4197 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
4198 }
4199
4200 check_flush_dependency(wq, NULL, false);
4201
4202 mutex_unlock(&wq->mutex);
4203
4204 wait_for_completion(&this_flusher.done);
4205
4206 /*
4207 * Wake-up-and-cascade phase
4208 *
4209 * First flushers are responsible for cascading flushes and
4210 * handling overflow. Non-first flushers can simply return.
4211 */
4212 if (READ_ONCE(wq->first_flusher) != &this_flusher)
4213 return;
4214
4215 mutex_lock(&wq->mutex);
4216
4217 /* we might have raced, check again with mutex held */
4218 if (wq->first_flusher != &this_flusher)
4219 goto out_unlock;
4220
4221 WRITE_ONCE(wq->first_flusher, NULL);
4222
4223 WARN_ON_ONCE(!list_empty(&this_flusher.list));
4224 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
4225
4226 while (true) {
4227 struct wq_flusher *next, *tmp;
4228
4229 /* complete all the flushers sharing the current flush color */
4230 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
4231 if (next->flush_color != wq->flush_color)
4232 break;
4233 list_del_init(&next->list);
4234 complete(&next->done);
4235 }
4236
4237 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4238 wq->flush_color != work_next_color(wq->work_color));
4239
4240 /* this flush_color is finished, advance by one */
4241 wq->flush_color = work_next_color(wq->flush_color);
4242
4243 /* one color has been freed, handle overflow queue */
4244 if (!list_empty(&wq->flusher_overflow)) {
4245 /*
4246 * Assign the same color to all overflowed
4247 * flushers, advance work_color and append to
4248 * flusher_queue. This is the start-to-wait
4249 * phase for these overflowed flushers.
4250 */
4251 list_for_each_entry(tmp, &wq->flusher_overflow, list)
4252 tmp->flush_color = wq->work_color;
4253
4254 wq->work_color = work_next_color(wq->work_color);
4255
4256 list_splice_tail_init(&wq->flusher_overflow,
4257 &wq->flusher_queue);
4258 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4259 }
4260
4261 if (list_empty(&wq->flusher_queue)) {
4262 WARN_ON_ONCE(wq->flush_color != wq->work_color);
4263 break;
4264 }
4265
4266 /*
4267 * Need to flush more colors. Make the next flusher
4268 * the new first flusher and arm pwqs.
4269 */
4270 WARN_ON_ONCE(wq->flush_color == wq->work_color);
4271 WARN_ON_ONCE(wq->flush_color != next->flush_color);
4272
4273 list_del_init(&next->list);
4274 wq->first_flusher = next;
4275
4276 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4277 break;
4278
4279 /*
4280 * Meh... this color is already done, clear first
4281 * flusher and repeat cascading.
4282 */
4283 wq->first_flusher = NULL;
4284 }
4285
4286out_unlock:
4287 mutex_unlock(&wq->mutex);
4288}
4289EXPORT_SYMBOL(__flush_workqueue);
4290
4291/**
4292 * drain_workqueue - drain a workqueue
4293 * @wq: workqueue to drain
4294 *
4295 * Wait until the workqueue becomes empty. While draining is in progress,
4296 * only chain queueing is allowed. IOW, only currently pending or running
4297 * work items on @wq can queue further work items on it. @wq is flushed
4298 * repeatedly until it becomes empty. The number of flushing is determined
4299 * by the depth of chaining and should be relatively short. Whine if it
4300 * takes too long.
4301 */
4302void drain_workqueue(struct workqueue_struct *wq)
4303{
4304 unsigned int flush_cnt = 0;
4305 struct pool_workqueue *pwq;
4306
4307 /*
4308 * __queue_work() needs to test whether there are drainers, is much
4309 * hotter than drain_workqueue() and already looks at @wq->flags.
4310 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4311 */
4312 mutex_lock(&wq->mutex);
4313 if (!wq->nr_drainers++)
4314 wq->flags |= __WQ_DRAINING;
4315 mutex_unlock(&wq->mutex);
4316reflush:
4317 __flush_workqueue(wq);
4318
4319 mutex_lock(&wq->mutex);
4320
4321 for_each_pwq(pwq, wq) {
4322 bool drained;
4323
4324 raw_spin_lock_irq(&pwq->pool->lock);
4325 drained = pwq_is_empty(pwq);
4326 raw_spin_unlock_irq(&pwq->pool->lock);
4327
4328 if (drained)
4329 continue;
4330
4331 if (++flush_cnt == 10 ||
4332 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4333 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
4334 wq->name, __func__, flush_cnt);
4335
4336 mutex_unlock(&wq->mutex);
4337 goto reflush;
4338 }
4339
4340 if (!--wq->nr_drainers)
4341 wq->flags &= ~__WQ_DRAINING;
4342 mutex_unlock(&wq->mutex);
4343}
4344EXPORT_SYMBOL_GPL(drain_workqueue);
4345
4346static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
4347 bool from_cancel)
4348{
4349 struct worker *worker = NULL;
4350 struct worker_pool *pool;
4351 struct pool_workqueue *pwq;
4352 struct workqueue_struct *wq;
4353
4354 rcu_read_lock();
4355 pool = get_work_pool(work);
4356 if (!pool) {
4357 rcu_read_unlock();
4358 return false;
4359 }
4360
4361 raw_spin_lock_irq(&pool->lock);
4362 /* see the comment in try_to_grab_pending() with the same code */
4363 pwq = get_work_pwq(work);
4364 if (pwq) {
4365 if (unlikely(pwq->pool != pool))
4366 goto already_gone;
4367 } else {
4368 worker = find_worker_executing_work(pool, work);
4369 if (!worker)
4370 goto already_gone;
4371 pwq = worker->current_pwq;
4372 }
4373
4374 wq = pwq->wq;
4375 check_flush_dependency(wq, work, from_cancel);
4376
4377 insert_wq_barrier(pwq, barr, work, worker);
4378 raw_spin_unlock_irq(&pool->lock);
4379
4380 touch_work_lockdep_map(work, wq);
4381
4382 /*
4383 * Force a lock recursion deadlock when using flush_work() inside a
4384 * single-threaded or rescuer equipped workqueue.
4385 *
4386 * For single threaded workqueues the deadlock happens when the work
4387 * is after the work issuing the flush_work(). For rescuer equipped
4388 * workqueues the deadlock happens when the rescuer stalls, blocking
4389 * forward progress.
4390 */
4391 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer))
4392 touch_wq_lockdep_map(wq);
4393
4394 rcu_read_unlock();
4395 return true;
4396already_gone:
4397 raw_spin_unlock_irq(&pool->lock);
4398 rcu_read_unlock();
4399 return false;
4400}
4401
4402static bool __flush_work(struct work_struct *work, bool from_cancel)
4403{
4404 struct wq_barrier barr;
4405
4406 if (WARN_ON(!wq_online))
4407 return false;
4408
4409 if (WARN_ON(!work->func))
4410 return false;
4411
4412 if (!start_flush_work(work, &barr, from_cancel))
4413 return false;
4414
4415 /*
4416 * start_flush_work() returned %true. If @from_cancel is set, we know
4417 * that @work must have been executing during start_flush_work() and
4418 * can't currently be queued. Its data must contain OFFQ bits. If @work
4419 * was queued on a BH workqueue, we also know that it was running in the
4420 * BH context and thus can be busy-waited.
4421 */
4422 if (from_cancel) {
4423 unsigned long data = *work_data_bits(work);
4424
4425 if (!WARN_ON_ONCE(data & WORK_STRUCT_PWQ) &&
4426 (data & WORK_OFFQ_BH)) {
4427 /*
4428 * On RT, prevent a live lock when %current preempted
4429 * soft interrupt processing by blocking on lock which
4430 * is owned by the thread invoking the callback.
4431 */
4432 while (!try_wait_for_completion(&barr.done)) {
4433 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
4434 struct worker_pool *pool;
4435
4436 guard(rcu)();
4437 pool = get_work_pool(work);
4438 if (pool)
4439 workqueue_callback_cancel_wait_running(pool);
4440 } else {
4441 cpu_relax();
4442 }
4443 }
4444 goto out_destroy;
4445 }
4446 }
4447
4448 wait_for_completion(&barr.done);
4449
4450out_destroy:
4451 destroy_work_on_stack(&barr.work);
4452 return true;
4453}
4454
4455/**
4456 * flush_work - wait for a work to finish executing the last queueing instance
4457 * @work: the work to flush
4458 *
4459 * Wait until @work has finished execution. @work is guaranteed to be idle
4460 * on return if it hasn't been requeued since flush started.
4461 *
4462 * Return:
4463 * %true if flush_work() waited for the work to finish execution,
4464 * %false if it was already idle.
4465 */
4466bool flush_work(struct work_struct *work)
4467{
4468 might_sleep();
4469 return __flush_work(work, false);
4470}
4471EXPORT_SYMBOL_GPL(flush_work);
4472
4473/**
4474 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4475 * @dwork: the delayed work to flush
4476 *
4477 * Delayed timer is cancelled and the pending work is queued for
4478 * immediate execution. Like flush_work(), this function only
4479 * considers the last queueing instance of @dwork.
4480 *
4481 * Return:
4482 * %true if flush_work() waited for the work to finish execution,
4483 * %false if it was already idle.
4484 */
4485bool flush_delayed_work(struct delayed_work *dwork)
4486{
4487 local_irq_disable();
4488 if (timer_delete_sync(&dwork->timer))
4489 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
4490 local_irq_enable();
4491 return flush_work(&dwork->work);
4492}
4493EXPORT_SYMBOL(flush_delayed_work);
4494
4495/**
4496 * flush_rcu_work - wait for a rwork to finish executing the last queueing
4497 * @rwork: the rcu work to flush
4498 *
4499 * Return:
4500 * %true if flush_rcu_work() waited for the work to finish execution,
4501 * %false if it was already idle.
4502 */
4503bool flush_rcu_work(struct rcu_work *rwork)
4504{
4505 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4506 rcu_barrier();
4507 flush_work(&rwork->work);
4508 return true;
4509 } else {
4510 return flush_work(&rwork->work);
4511 }
4512}
4513EXPORT_SYMBOL(flush_rcu_work);
4514
4515static void work_offqd_disable(struct work_offq_data *offqd)
4516{
4517 const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1;
4518
4519 if (likely(offqd->disable < max))
4520 offqd->disable++;
4521 else
4522 WARN_ONCE(true, "workqueue: work disable count overflowed\n");
4523}
4524
4525static void work_offqd_enable(struct work_offq_data *offqd)
4526{
4527 if (likely(offqd->disable > 0))
4528 offqd->disable--;
4529 else
4530 WARN_ONCE(true, "workqueue: work disable count underflowed\n");
4531}
4532
4533static bool __cancel_work(struct work_struct *work, u32 cflags)
4534{
4535 struct work_offq_data offqd;
4536 unsigned long irq_flags;
4537 int ret;
4538
4539 ret = work_grab_pending(work, cflags, &irq_flags);
4540
4541 work_offqd_unpack(&offqd, *work_data_bits(work));
4542
4543 if (cflags & WORK_CANCEL_DISABLE)
4544 work_offqd_disable(&offqd);
4545
4546 set_work_pool_and_clear_pending(work, offqd.pool_id,
4547 work_offqd_pack_flags(&offqd));
4548 local_irq_restore(irq_flags);
4549 return ret;
4550}
4551
4552static bool __cancel_work_sync(struct work_struct *work, u32 cflags)
4553{
4554 bool ret;
4555
4556 ret = __cancel_work(work, cflags | WORK_CANCEL_DISABLE);
4557
4558 if (*work_data_bits(work) & WORK_OFFQ_BH)
4559 WARN_ON_ONCE(in_hardirq());
4560 else
4561 might_sleep();
4562
4563 /*
4564 * Skip __flush_work() during early boot when we know that @work isn't
4565 * executing. This allows canceling during early boot.
4566 */
4567 if (wq_online)
4568 __flush_work(work, true);
4569
4570 if (!(cflags & WORK_CANCEL_DISABLE))
4571 enable_work(work);
4572
4573 return ret;
4574}
4575
4576/*
4577 * See cancel_delayed_work()
4578 */
4579bool cancel_work(struct work_struct *work)
4580{
4581 return __cancel_work(work, 0);
4582}
4583EXPORT_SYMBOL(cancel_work);
4584
4585/**
4586 * cancel_work_sync - cancel a work and wait for it to finish
4587 * @work: the work to cancel
4588 *
4589 * Cancel @work and wait for its execution to finish. This function can be used
4590 * even if the work re-queues itself or migrates to another workqueue. On return
4591 * from this function, @work is guaranteed to be not pending or executing on any
4592 * CPU as long as there aren't racing enqueues.
4593 *
4594 * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's.
4595 * Use cancel_delayed_work_sync() instead.
4596 *
4597 * Must be called from a sleepable context if @work was last queued on a non-BH
4598 * workqueue. Can also be called from non-hardirq atomic contexts including BH
4599 * if @work was last queued on a BH workqueue.
4600 *
4601 * Returns %true if @work was pending, %false otherwise.
4602 */
4603bool cancel_work_sync(struct work_struct *work)
4604{
4605 return __cancel_work_sync(work, 0);
4606}
4607EXPORT_SYMBOL_GPL(cancel_work_sync);
4608
4609/**
4610 * cancel_delayed_work - cancel a delayed work
4611 * @dwork: delayed_work to cancel
4612 *
4613 * Kill off a pending delayed_work.
4614 *
4615 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4616 * pending.
4617 *
4618 * Note:
4619 * The work callback function may still be running on return, unless
4620 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4621 * use cancel_delayed_work_sync() to wait on it.
4622 *
4623 * This function is safe to call from any context including IRQ handler.
4624 */
4625bool cancel_delayed_work(struct delayed_work *dwork)
4626{
4627 return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED);
4628}
4629EXPORT_SYMBOL(cancel_delayed_work);
4630
4631/**
4632 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4633 * @dwork: the delayed work cancel
4634 *
4635 * This is cancel_work_sync() for delayed works.
4636 *
4637 * Return:
4638 * %true if @dwork was pending, %false otherwise.
4639 */
4640bool cancel_delayed_work_sync(struct delayed_work *dwork)
4641{
4642 return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED);
4643}
4644EXPORT_SYMBOL(cancel_delayed_work_sync);
4645
4646/**
4647 * disable_work - Disable and cancel a work item
4648 * @work: work item to disable
4649 *
4650 * Disable @work by incrementing its disable count and cancel it if currently
4651 * pending. As long as the disable count is non-zero, any attempt to queue @work
4652 * will fail and return %false. The maximum supported disable depth is 2 to the
4653 * power of %WORK_OFFQ_DISABLE_BITS, currently 65536.
4654 *
4655 * Can be called from any context. Returns %true if @work was pending, %false
4656 * otherwise.
4657 */
4658bool disable_work(struct work_struct *work)
4659{
4660 return __cancel_work(work, WORK_CANCEL_DISABLE);
4661}
4662EXPORT_SYMBOL_GPL(disable_work);
4663
4664/**
4665 * disable_work_sync - Disable, cancel and drain a work item
4666 * @work: work item to disable
4667 *
4668 * Similar to disable_work() but also wait for @work to finish if currently
4669 * executing.
4670 *
4671 * Must be called from a sleepable context if @work was last queued on a non-BH
4672 * workqueue. Can also be called from non-hardirq atomic contexts including BH
4673 * if @work was last queued on a BH workqueue.
4674 *
4675 * Returns %true if @work was pending, %false otherwise.
4676 */
4677bool disable_work_sync(struct work_struct *work)
4678{
4679 return __cancel_work_sync(work, WORK_CANCEL_DISABLE);
4680}
4681EXPORT_SYMBOL_GPL(disable_work_sync);
4682
4683/**
4684 * enable_work - Enable a work item
4685 * @work: work item to enable
4686 *
4687 * Undo disable_work[_sync]() by decrementing @work's disable count. @work can
4688 * only be queued if its disable count is 0.
4689 *
4690 * Can be called from any context. Returns %true if the disable count reached 0.
4691 * Otherwise, %false.
4692 */
4693bool enable_work(struct work_struct *work)
4694{
4695 struct work_offq_data offqd;
4696 unsigned long irq_flags;
4697
4698 work_grab_pending(work, 0, &irq_flags);
4699
4700 work_offqd_unpack(&offqd, *work_data_bits(work));
4701 work_offqd_enable(&offqd);
4702 set_work_pool_and_clear_pending(work, offqd.pool_id,
4703 work_offqd_pack_flags(&offqd));
4704 local_irq_restore(irq_flags);
4705
4706 return !offqd.disable;
4707}
4708EXPORT_SYMBOL_GPL(enable_work);
4709
4710/**
4711 * disable_delayed_work - Disable and cancel a delayed work item
4712 * @dwork: delayed work item to disable
4713 *
4714 * disable_work() for delayed work items.
4715 */
4716bool disable_delayed_work(struct delayed_work *dwork)
4717{
4718 return __cancel_work(&dwork->work,
4719 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4720}
4721EXPORT_SYMBOL_GPL(disable_delayed_work);
4722
4723/**
4724 * disable_delayed_work_sync - Disable, cancel and drain a delayed work item
4725 * @dwork: delayed work item to disable
4726 *
4727 * disable_work_sync() for delayed work items.
4728 */
4729bool disable_delayed_work_sync(struct delayed_work *dwork)
4730{
4731 return __cancel_work_sync(&dwork->work,
4732 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4733}
4734EXPORT_SYMBOL_GPL(disable_delayed_work_sync);
4735
4736/**
4737 * enable_delayed_work - Enable a delayed work item
4738 * @dwork: delayed work item to enable
4739 *
4740 * enable_work() for delayed work items.
4741 */
4742bool enable_delayed_work(struct delayed_work *dwork)
4743{
4744 return enable_work(&dwork->work);
4745}
4746EXPORT_SYMBOL_GPL(enable_delayed_work);
4747
4748/**
4749 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4750 * @func: the function to call
4751 *
4752 * schedule_on_each_cpu() executes @func on each online CPU using the
4753 * system workqueue and blocks until all CPUs have completed.
4754 * schedule_on_each_cpu() is very slow.
4755 *
4756 * Return:
4757 * 0 on success, -errno on failure.
4758 */
4759int schedule_on_each_cpu(work_func_t func)
4760{
4761 int cpu;
4762 struct work_struct __percpu *works;
4763
4764 works = alloc_percpu(struct work_struct);
4765 if (!works)
4766 return -ENOMEM;
4767
4768 cpus_read_lock();
4769
4770 for_each_online_cpu(cpu) {
4771 struct work_struct *work = per_cpu_ptr(works, cpu);
4772
4773 INIT_WORK(work, func);
4774 schedule_work_on(cpu, work);
4775 }
4776
4777 for_each_online_cpu(cpu)
4778 flush_work(per_cpu_ptr(works, cpu));
4779
4780 cpus_read_unlock();
4781 free_percpu(works);
4782 return 0;
4783}
4784
4785/**
4786 * execute_in_process_context - reliably execute the routine with user context
4787 * @fn: the function to execute
4788 * @ew: guaranteed storage for the execute work structure (must
4789 * be available when the work executes)
4790 *
4791 * Executes the function immediately if process context is available,
4792 * otherwise schedules the function for delayed execution.
4793 *
4794 * Return: 0 - function was executed
4795 * 1 - function was scheduled for execution
4796 */
4797int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4798{
4799 if (!in_interrupt()) {
4800 fn(&ew->work);
4801 return 0;
4802 }
4803
4804 INIT_WORK(&ew->work, fn);
4805 schedule_work(&ew->work);
4806
4807 return 1;
4808}
4809EXPORT_SYMBOL_GPL(execute_in_process_context);
4810
4811/**
4812 * free_workqueue_attrs - free a workqueue_attrs
4813 * @attrs: workqueue_attrs to free
4814 *
4815 * Undo alloc_workqueue_attrs().
4816 */
4817void free_workqueue_attrs(struct workqueue_attrs *attrs)
4818{
4819 if (attrs) {
4820 free_cpumask_var(attrs->cpumask);
4821 free_cpumask_var(attrs->__pod_cpumask);
4822 kfree(attrs);
4823 }
4824}
4825
4826/**
4827 * alloc_workqueue_attrs - allocate a workqueue_attrs
4828 *
4829 * Allocate a new workqueue_attrs, initialize with default settings and
4830 * return it.
4831 *
4832 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4833 */
4834struct workqueue_attrs *alloc_workqueue_attrs_noprof(void)
4835{
4836 struct workqueue_attrs *attrs;
4837
4838 attrs = kzalloc_obj(*attrs);
4839 if (!attrs)
4840 goto fail;
4841 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4842 goto fail;
4843 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4844 goto fail;
4845
4846 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4847 attrs->affn_scope = WQ_AFFN_DFL;
4848 return attrs;
4849fail:
4850 free_workqueue_attrs(attrs);
4851 return NULL;
4852}
4853
4854static void copy_workqueue_attrs(struct workqueue_attrs *to,
4855 const struct workqueue_attrs *from)
4856{
4857 to->nice = from->nice;
4858 cpumask_copy(to->cpumask, from->cpumask);
4859 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4860 to->affn_strict = from->affn_strict;
4861
4862 /*
4863 * Unlike hash and equality test, copying shouldn't ignore wq-only
4864 * fields as copying is used for both pool and wq attrs. Instead,
4865 * get_unbound_pool() explicitly clears the fields.
4866 */
4867 to->affn_scope = from->affn_scope;
4868 to->ordered = from->ordered;
4869}
4870
4871/*
4872 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4873 * comments in 'struct workqueue_attrs' definition.
4874 */
4875static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4876{
4877 attrs->affn_scope = WQ_AFFN_NR_TYPES;
4878 attrs->ordered = false;
4879 if (attrs->affn_strict)
4880 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4881}
4882
4883/* hash value of the content of @attr */
4884static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4885{
4886 u32 hash = 0;
4887
4888 hash = jhash_1word(attrs->nice, hash);
4889 hash = jhash_1word(attrs->affn_strict, hash);
4890 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4891 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4892 if (!attrs->affn_strict)
4893 hash = jhash(cpumask_bits(attrs->cpumask),
4894 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4895 return hash;
4896}
4897
4898/* content equality test */
4899static bool wqattrs_equal(const struct workqueue_attrs *a,
4900 const struct workqueue_attrs *b)
4901{
4902 if (a->nice != b->nice)
4903 return false;
4904 if (a->affn_strict != b->affn_strict)
4905 return false;
4906 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4907 return false;
4908 if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask))
4909 return false;
4910 return true;
4911}
4912
4913/* Update @attrs with actually available CPUs */
4914static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4915 const cpumask_t *unbound_cpumask)
4916{
4917 /*
4918 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4919 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4920 * @unbound_cpumask.
4921 */
4922 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4923 if (unlikely(cpumask_empty(attrs->cpumask)))
4924 cpumask_copy(attrs->cpumask, unbound_cpumask);
4925}
4926
4927/* find wq_pod_type to use for @attrs */
4928static const struct wq_pod_type *
4929wqattrs_pod_type(const struct workqueue_attrs *attrs)
4930{
4931 enum wq_affn_scope scope;
4932 struct wq_pod_type *pt;
4933
4934 /* to synchronize access to wq_affn_dfl */
4935 lockdep_assert_held(&wq_pool_mutex);
4936
4937 if (attrs->affn_scope == WQ_AFFN_DFL)
4938 scope = wq_affn_dfl;
4939 else
4940 scope = attrs->affn_scope;
4941
4942 pt = &wq_pod_types[scope];
4943
4944 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4945 likely(pt->nr_pods))
4946 return pt;
4947
4948 /*
4949 * Before workqueue_init_topology(), only SYSTEM is available which is
4950 * initialized in workqueue_init_early().
4951 */
4952 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4953 BUG_ON(!pt->nr_pods);
4954 return pt;
4955}
4956
4957/**
4958 * init_worker_pool - initialize a newly zalloc'd worker_pool
4959 * @pool: worker_pool to initialize
4960 *
4961 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4962 *
4963 * Return: 0 on success, -errno on failure. Even on failure, all fields
4964 * inside @pool proper are initialized and put_unbound_pool() can be called
4965 * on @pool safely to release it.
4966 */
4967static int init_worker_pool(struct worker_pool *pool)
4968{
4969 raw_spin_lock_init(&pool->lock);
4970 pool->id = -1;
4971 pool->cpu = -1;
4972 pool->node = NUMA_NO_NODE;
4973 pool->flags |= POOL_DISASSOCIATED;
4974 pool->last_progress_ts = jiffies;
4975 INIT_LIST_HEAD(&pool->worklist);
4976 INIT_LIST_HEAD(&pool->idle_list);
4977 hash_init(pool->busy_hash);
4978
4979 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4980 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4981
4982 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4983
4984 INIT_LIST_HEAD(&pool->workers);
4985
4986 ida_init(&pool->worker_ida);
4987 INIT_HLIST_NODE(&pool->hash_node);
4988 pool->refcnt = 1;
4989#ifdef CONFIG_PREEMPT_RT
4990 spin_lock_init(&pool->cb_lock);
4991#endif
4992
4993 /* shouldn't fail above this point */
4994 pool->attrs = alloc_workqueue_attrs();
4995 if (!pool->attrs)
4996 return -ENOMEM;
4997
4998 wqattrs_clear_for_pool(pool->attrs);
4999
5000 return 0;
5001}
5002
5003#ifdef CONFIG_LOCKDEP
5004static void wq_init_lockdep(struct workqueue_struct *wq)
5005{
5006 char *lock_name;
5007
5008 lockdep_register_key(&wq->key);
5009 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
5010 if (!lock_name)
5011 lock_name = wq->name;
5012
5013 wq->lock_name = lock_name;
5014 wq->lockdep_map = &wq->__lockdep_map;
5015 lockdep_init_map(wq->lockdep_map, lock_name, &wq->key, 0);
5016}
5017
5018static void wq_unregister_lockdep(struct workqueue_struct *wq)
5019{
5020 if (wq->lockdep_map != &wq->__lockdep_map)
5021 return;
5022
5023 lockdep_unregister_key(&wq->key);
5024}
5025
5026static void wq_free_lockdep(struct workqueue_struct *wq)
5027{
5028 if (wq->lockdep_map != &wq->__lockdep_map)
5029 return;
5030
5031 if (wq->lock_name != wq->name)
5032 kfree(wq->lock_name);
5033}
5034#else
5035static void wq_init_lockdep(struct workqueue_struct *wq)
5036{
5037}
5038
5039static void wq_unregister_lockdep(struct workqueue_struct *wq)
5040{
5041}
5042
5043static void wq_free_lockdep(struct workqueue_struct *wq)
5044{
5045}
5046#endif
5047
5048static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
5049{
5050 int node;
5051
5052 for_each_node(node) {
5053 kfree(nna_ar[node]);
5054 nna_ar[node] = NULL;
5055 }
5056
5057 kfree(nna_ar[nr_node_ids]);
5058 nna_ar[nr_node_ids] = NULL;
5059}
5060
5061static void init_node_nr_active(struct wq_node_nr_active *nna)
5062{
5063 nna->max = WQ_DFL_MIN_ACTIVE;
5064 atomic_set(&nna->nr, 0);
5065 raw_spin_lock_init(&nna->lock);
5066 INIT_LIST_HEAD(&nna->pending_pwqs);
5067}
5068
5069/*
5070 * Each node's nr_active counter will be accessed mostly from its own node and
5071 * should be allocated in the node.
5072 */
5073static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
5074{
5075 struct wq_node_nr_active *nna;
5076 int node;
5077
5078 for_each_node(node) {
5079 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
5080 if (!nna)
5081 goto err_free;
5082 init_node_nr_active(nna);
5083 nna_ar[node] = nna;
5084 }
5085
5086 /* [nr_node_ids] is used as the fallback */
5087 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
5088 if (!nna)
5089 goto err_free;
5090 init_node_nr_active(nna);
5091 nna_ar[nr_node_ids] = nna;
5092
5093 return 0;
5094
5095err_free:
5096 free_node_nr_active(nna_ar);
5097 return -ENOMEM;
5098}
5099
5100static void rcu_free_wq(struct rcu_head *rcu)
5101{
5102 struct workqueue_struct *wq =
5103 container_of(rcu, struct workqueue_struct, rcu);
5104
5105 if (wq->flags & WQ_UNBOUND)
5106 free_node_nr_active(wq->node_nr_active);
5107
5108 wq_free_lockdep(wq);
5109 free_percpu(wq->cpu_pwq);
5110 free_workqueue_attrs(wq->attrs);
5111 kfree(wq);
5112}
5113
5114static void rcu_free_pool(struct rcu_head *rcu)
5115{
5116 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
5117
5118 ida_destroy(&pool->worker_ida);
5119 free_workqueue_attrs(pool->attrs);
5120 kfree(pool);
5121}
5122
5123/**
5124 * put_unbound_pool - put a worker_pool
5125 * @pool: worker_pool to put
5126 *
5127 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
5128 * safe manner. get_unbound_pool() calls this function on its failure path
5129 * and this function should be able to release pools which went through,
5130 * successfully or not, init_worker_pool().
5131 *
5132 * Should be called with wq_pool_mutex held.
5133 */
5134static void put_unbound_pool(struct worker_pool *pool)
5135{
5136 struct worker *worker;
5137 LIST_HEAD(cull_list);
5138
5139 lockdep_assert_held(&wq_pool_mutex);
5140
5141 if (--pool->refcnt)
5142 return;
5143
5144 /* sanity checks */
5145 if (WARN_ON(is_percpu_pool(pool)) ||
5146 WARN_ON(!list_empty(&pool->worklist)))
5147 return;
5148
5149 /* release id and unhash */
5150 if (pool->id >= 0)
5151 idr_remove(&worker_pool_idr, pool->id);
5152 hash_del(&pool->hash_node);
5153
5154 /*
5155 * Become the manager and destroy all workers. This prevents
5156 * @pool's workers from blocking on attach_mutex. We're the last
5157 * manager and @pool gets freed with the flag set.
5158 *
5159 * Having a concurrent manager is quite unlikely to happen as we can
5160 * only get here with
5161 * pwq->refcnt == pool->refcnt == 0
5162 * which implies no work queued to the pool, which implies no worker can
5163 * become the manager. However a worker could have taken the role of
5164 * manager before the refcnts dropped to 0, since maybe_create_worker()
5165 * drops pool->lock
5166 */
5167 while (true) {
5168 rcuwait_wait_event(&manager_wait,
5169 !(pool->flags & POOL_MANAGER_ACTIVE),
5170 TASK_UNINTERRUPTIBLE);
5171
5172 mutex_lock(&wq_pool_attach_mutex);
5173 raw_spin_lock_irq(&pool->lock);
5174 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
5175 pool->flags |= POOL_MANAGER_ACTIVE;
5176 break;
5177 }
5178 raw_spin_unlock_irq(&pool->lock);
5179 mutex_unlock(&wq_pool_attach_mutex);
5180 }
5181
5182 while ((worker = first_idle_worker(pool)))
5183 set_worker_dying(worker, &cull_list);
5184 WARN_ON(pool->nr_workers || pool->nr_idle);
5185 raw_spin_unlock_irq(&pool->lock);
5186
5187 detach_dying_workers(&cull_list);
5188
5189 mutex_unlock(&wq_pool_attach_mutex);
5190
5191 reap_dying_workers(&cull_list);
5192
5193 /* shut down the timers */
5194 timer_delete_sync(&pool->idle_timer);
5195 cancel_work_sync(&pool->idle_cull_work);
5196 timer_delete_sync(&pool->mayday_timer);
5197
5198 /* RCU protected to allow dereferences from get_work_pool() */
5199 call_rcu(&pool->rcu, rcu_free_pool);
5200}
5201
5202/**
5203 * get_unbound_pool - get a worker_pool with the specified attributes
5204 * @attrs: the attributes of the worker_pool to get
5205 *
5206 * Obtain a worker_pool which has the same attributes as @attrs, bump the
5207 * reference count and return it. If there already is a matching
5208 * worker_pool, it will be used; otherwise, this function attempts to
5209 * create a new one.
5210 *
5211 * Should be called with wq_pool_mutex held.
5212 *
5213 * Return: On success, a worker_pool with the same attributes as @attrs.
5214 * On failure, %NULL.
5215 */
5216static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
5217{
5218 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
5219 u32 hash = wqattrs_hash(attrs);
5220 struct worker_pool *pool;
5221 int pod, node = NUMA_NO_NODE;
5222
5223 lockdep_assert_held(&wq_pool_mutex);
5224
5225 /* do we already have a matching pool? */
5226 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
5227 if (wqattrs_equal(pool->attrs, attrs)) {
5228 pool->refcnt++;
5229 return pool;
5230 }
5231 }
5232
5233 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
5234 for (pod = 0; pod < pt->nr_pods; pod++) {
5235 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
5236 node = pt->pod_node[pod];
5237 break;
5238 }
5239 }
5240
5241 /* nope, create a new one */
5242 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
5243 if (!pool || init_worker_pool(pool) < 0)
5244 goto fail;
5245
5246 pool->node = node;
5247 copy_workqueue_attrs(pool->attrs, attrs);
5248 wqattrs_clear_for_pool(pool->attrs);
5249
5250 if (worker_pool_assign_id(pool) < 0)
5251 goto fail;
5252
5253 /* create and start the initial worker */
5254 if (wq_online && !create_worker(pool))
5255 goto fail;
5256
5257 /* install */
5258 hash_add(unbound_pool_hash, &pool->hash_node, hash);
5259
5260 return pool;
5261fail:
5262 if (pool)
5263 put_unbound_pool(pool);
5264 return NULL;
5265}
5266
5267/*
5268 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
5269 * refcnt and needs to be destroyed.
5270 */
5271static void pwq_release_workfn(struct kthread_work *work)
5272{
5273 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
5274 release_work);
5275 struct workqueue_struct *wq = pwq->wq;
5276 struct worker_pool *pool = pwq->pool;
5277 bool is_last = false;
5278
5279 /*
5280 * When @pwq is not linked, it doesn't hold any reference to the
5281 * @wq, and @wq is invalid to access.
5282 */
5283 if (!list_empty(&pwq->pwqs_node)) {
5284 mutex_lock(&wq->mutex);
5285 list_del_rcu(&pwq->pwqs_node);
5286 is_last = list_empty(&wq->pwqs);
5287
5288 /*
5289 * For ordered workqueue with a plugged dfl_pwq, restart it now.
5290 */
5291 if (!is_last && (wq->flags & __WQ_ORDERED))
5292 unplug_oldest_pwq(wq);
5293
5294 mutex_unlock(&wq->mutex);
5295 }
5296
5297 if (!list_empty(&pwq->pending_node)) {
5298 struct wq_node_nr_active *nna =
5299 wq_node_nr_active(pwq->wq, pwq->pool->node);
5300
5301 raw_spin_lock_irq(&nna->lock);
5302 list_del_init(&pwq->pending_node);
5303 raw_spin_unlock_irq(&nna->lock);
5304 }
5305
5306 if (!is_percpu_pool(pool)) {
5307 mutex_lock(&wq_pool_mutex);
5308 put_unbound_pool(pool);
5309 mutex_unlock(&wq_pool_mutex);
5310 }
5311
5312 kfree_rcu(pwq, rcu);
5313
5314 /*
5315 * If we're the last pwq going away, @wq is already dead and no one
5316 * is gonna access it anymore. Schedule RCU free.
5317 */
5318 if (is_last) {
5319 wq_unregister_lockdep(wq);
5320 call_rcu(&wq->rcu, rcu_free_wq);
5321 }
5322}
5323
5324/* initialize newly allocated @pwq which is associated with @wq and @pool */
5325static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
5326 struct worker_pool *pool)
5327{
5328 BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK);
5329
5330 memset(pwq, 0, sizeof(*pwq));
5331
5332 pwq->pool = pool;
5333 pwq->wq = wq;
5334 pwq->flush_color = -1;
5335 pwq->refcnt = 1;
5336 INIT_LIST_HEAD(&pwq->inactive_works);
5337 INIT_LIST_HEAD(&pwq->pending_node);
5338 INIT_LIST_HEAD(&pwq->pwqs_node);
5339 INIT_LIST_HEAD(&pwq->mayday_node);
5340 kthread_init_work(&pwq->release_work, pwq_release_workfn);
5341
5342 /*
5343 * Set the dummy cursor work with valid function and get_work_pwq().
5344 *
5345 * The cursor work should only be in the pwq->pool->worklist, and
5346 * should not be treated as a processable work item.
5347 *
5348 * WORK_STRUCT_PENDING and WORK_STRUCT_INACTIVE just make it less
5349 * surprise for kernel debugging tools and reviewers.
5350 */
5351 INIT_WORK(&pwq->mayday_cursor, mayday_cursor_func);
5352 atomic_long_set(&pwq->mayday_cursor.data, (unsigned long)pwq |
5353 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | WORK_STRUCT_INACTIVE);
5354}
5355
5356/* sync @pwq with the current state of its associated wq and link it */
5357static void link_pwq(struct pool_workqueue *pwq)
5358{
5359 struct workqueue_struct *wq = pwq->wq;
5360
5361 lockdep_assert_held(&wq->mutex);
5362
5363 /* may be called multiple times, ignore if already linked */
5364 if (!list_empty(&pwq->pwqs_node))
5365 return;
5366
5367 /* set the matching work_color */
5368 pwq->work_color = wq->work_color;
5369
5370 /* link in @pwq */
5371 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs);
5372}
5373
5374/* Return the static per-cpu worker_pool that backs @wq on @cpu. */
5375static struct worker_pool *get_percpu_pool(struct workqueue_struct *wq, int cpu)
5376{
5377 struct worker_pool __percpu *pools;
5378 bool highpri = wq->flags & WQ_HIGHPRI;
5379
5380 if (wq->flags & WQ_BH)
5381 pools = bh_worker_pools;
5382 else
5383 pools = cpu_worker_pools;
5384
5385 return &per_cpu_ptr(pools, cpu)[highpri];
5386}
5387
5388/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5389static struct pool_workqueue *alloc_pwq(struct workqueue_struct *wq,
5390 const struct workqueue_attrs *attrs)
5391{
5392 struct worker_pool *pool;
5393 struct pool_workqueue *pwq;
5394
5395 lockdep_assert_held(&wq_pool_mutex);
5396
5397 pool = get_unbound_pool(attrs);
5398 if (!pool)
5399 return NULL;
5400
5401 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5402 if (!pwq) {
5403 put_unbound_pool(pool);
5404 return NULL;
5405 }
5406
5407 init_pwq(pwq, wq, pool);
5408 return pwq;
5409}
5410
5411/**
5412 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
5413 * @attrs: the wq_attrs of the default pwq of the target workqueue
5414 * @cpu: the target CPU
5415 *
5416 * Calculate the cpumask a workqueue with @attrs should use on @pod.
5417 * The result is stored in @attrs->__pod_cpumask.
5418 *
5419 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
5420 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
5421 * intersection of the possible CPUs of @pod and @attrs->cpumask.
5422 *
5423 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
5424 */
5425static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu)
5426{
5427 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5428 int pod = pt->cpu_pod[cpu];
5429
5430 /* calculate possible CPUs in @pod that @attrs wants */
5431 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
5432 /* does @pod have any online CPUs @attrs wants? */
5433 if (!cpumask_intersects(attrs->__pod_cpumask, wq_online_cpumask)) {
5434 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
5435 return;
5436 }
5437}
5438
5439/* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
5440static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
5441 int cpu, struct pool_workqueue *pwq)
5442{
5443 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
5444 struct pool_workqueue *old_pwq;
5445
5446 lockdep_assert_held(&wq_pool_mutex);
5447 lockdep_assert_held(&wq->mutex);
5448
5449 /* link_pwq() can handle duplicate calls */
5450 link_pwq(pwq);
5451
5452 old_pwq = rcu_access_pointer(*slot);
5453 rcu_assign_pointer(*slot, pwq);
5454 return old_pwq;
5455}
5456
5457/* context to store the prepared attrs & pwqs before applying */
5458struct apply_wqattrs_ctx {
5459 struct workqueue_struct *wq; /* target workqueue */
5460 struct workqueue_attrs *attrs; /* attrs to apply */
5461 struct list_head list; /* queued for batching commit */
5462 struct pool_workqueue *dfl_pwq;
5463 struct pool_workqueue *pwq_tbl[];
5464};
5465
5466/* free the resources after success or abort */
5467static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5468{
5469 if (ctx) {
5470 int cpu;
5471
5472 for_each_possible_cpu(cpu)
5473 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
5474 put_pwq_unlocked(ctx->dfl_pwq);
5475
5476 free_workqueue_attrs(ctx->attrs);
5477
5478 kfree(ctx);
5479 }
5480}
5481
5482/* allocate the attrs and pwqs for later installation */
5483static struct apply_wqattrs_ctx *
5484apply_wqattrs_prepare(struct workqueue_struct *wq,
5485 const struct workqueue_attrs *attrs,
5486 const cpumask_var_t unbound_cpumask)
5487{
5488 struct apply_wqattrs_ctx *ctx;
5489 struct workqueue_attrs *new_attrs;
5490 int cpu;
5491
5492 lockdep_assert_held(&wq_pool_mutex);
5493
5494 if (WARN_ON(attrs->affn_scope < 0 ||
5495 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
5496 return ERR_PTR(-EINVAL);
5497
5498 ctx = kzalloc_flex(*ctx, pwq_tbl, nr_cpu_ids);
5499
5500 new_attrs = alloc_workqueue_attrs();
5501 if (!ctx || !new_attrs)
5502 goto out_free;
5503
5504 /*
5505 * If something goes wrong during CPU up/down, we'll fall back to
5506 * the default pwq covering whole @attrs->cpumask. Always create
5507 * it even if we don't use it immediately.
5508 */
5509 copy_workqueue_attrs(new_attrs, attrs);
5510 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
5511 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5512 ctx->dfl_pwq = alloc_pwq(wq, new_attrs);
5513 if (!ctx->dfl_pwq)
5514 goto out_free;
5515
5516 for_each_possible_cpu(cpu) {
5517 if (new_attrs->ordered) {
5518 ctx->dfl_pwq->refcnt++;
5519 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
5520 } else {
5521 wq_calc_pod_cpumask(new_attrs, cpu);
5522 ctx->pwq_tbl[cpu] = alloc_pwq(wq, new_attrs);
5523 if (!ctx->pwq_tbl[cpu])
5524 goto out_free;
5525 }
5526 }
5527
5528 /* save the user configured attrs and sanitize it. */
5529 copy_workqueue_attrs(new_attrs, attrs);
5530 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
5531 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5532 ctx->attrs = new_attrs;
5533
5534 /*
5535 * For initialized ordered workqueues, there should only be one pwq
5536 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution
5537 * of newly queued work items until execution of older work items in
5538 * the old pwq's have completed.
5539 */
5540 if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))
5541 ctx->dfl_pwq->plugged = true;
5542
5543 ctx->wq = wq;
5544 return ctx;
5545
5546out_free:
5547 free_workqueue_attrs(new_attrs);
5548 apply_wqattrs_cleanup(ctx);
5549 return ERR_PTR(-ENOMEM);
5550}
5551
5552/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
5553static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5554{
5555 int cpu;
5556
5557 /* all pwqs have been created successfully, let's install'em */
5558 mutex_lock(&ctx->wq->mutex);
5559
5560 copy_workqueue_attrs(ctx->wq->attrs, ctx->attrs);
5561
5562 /* save the previous pwqs and install the new ones */
5563 for_each_possible_cpu(cpu)
5564 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
5565 ctx->pwq_tbl[cpu]);
5566 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq);
5567
5568 /* update node_nr_active->max, which only unbound workqueues have */
5569 if (ctx->wq->flags & WQ_UNBOUND)
5570 wq_update_node_max_active(ctx->wq, -1);
5571
5572 mutex_unlock(&ctx->wq->mutex);
5573}
5574
5575static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5576 const struct workqueue_attrs *attrs)
5577{
5578 struct apply_wqattrs_ctx *ctx;
5579
5580 /* only unbound workqueues can change attributes */
5581 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5582 return -EINVAL;
5583
5584 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
5585 if (IS_ERR(ctx))
5586 return PTR_ERR(ctx);
5587
5588 /* the ctx has been prepared successfully, let's commit it */
5589 apply_wqattrs_commit(ctx);
5590 apply_wqattrs_cleanup(ctx);
5591
5592 return 0;
5593}
5594
5595/**
5596 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5597 * @wq: the target workqueue
5598 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5599 *
5600 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
5601 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
5602 * work items are affine to the pod it was issued on. Older pwqs are released as
5603 * in-flight work items finish. Note that a work item which repeatedly requeues
5604 * itself back-to-back will stay on its current pwq.
5605 *
5606 * Performs GFP_KERNEL allocations.
5607 *
5608 * Return: 0 on success and -errno on failure.
5609 */
5610int apply_workqueue_attrs(struct workqueue_struct *wq,
5611 const struct workqueue_attrs *attrs)
5612{
5613 int ret;
5614
5615 mutex_lock(&wq_pool_mutex);
5616 ret = apply_workqueue_attrs_locked(wq, attrs);
5617 mutex_unlock(&wq_pool_mutex);
5618
5619 return ret;
5620}
5621
5622/**
5623 * unbound_wq_update_pwq - update a pwq slot for CPU hot[un]plug
5624 * @wq: the target workqueue
5625 * @cpu: the CPU to update the pwq slot for
5626 *
5627 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5628 * %CPU_DOWN_FAILED. @cpu is in the same pod of the CPU being hot[un]plugged.
5629 *
5630 *
5631 * If pod affinity can't be adjusted due to memory allocation failure, it falls
5632 * back to @wq->dfl_pwq which may not be optimal but is always correct.
5633 *
5634 * Note that when the last allowed CPU of a pod goes offline for a workqueue
5635 * with a cpumask spanning multiple pods, the workers which were already
5636 * executing the work items for the workqueue will lose their CPU affinity and
5637 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
5638 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
5639 * responsibility to flush the work item from CPU_DOWN_PREPARE.
5640 */
5641static void unbound_wq_update_pwq(struct workqueue_struct *wq, int cpu)
5642{
5643 struct pool_workqueue *old_pwq = NULL, *pwq;
5644 struct workqueue_attrs *target_attrs;
5645
5646 lockdep_assert_held(&wq_pool_mutex);
5647
5648 if (!(wq->flags & WQ_UNBOUND) || wq->attrs->ordered)
5649 return;
5650
5651 /*
5652 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5653 * Let's use a preallocated one. The following buf is protected by
5654 * CPU hotplug exclusion.
5655 */
5656 target_attrs = unbound_wq_update_pwq_attrs_buf;
5657
5658 copy_workqueue_attrs(target_attrs, wq->attrs);
5659 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
5660
5661 /* nothing to do if the target cpumask matches the current pwq */
5662 wq_calc_pod_cpumask(target_attrs, cpu);
5663 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs))
5664 return;
5665
5666 /* create a new pwq */
5667 pwq = alloc_pwq(wq, target_attrs);
5668 if (!pwq) {
5669 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
5670 wq->name);
5671 goto use_dfl_pwq;
5672 }
5673
5674 /* Install the new pwq. */
5675 mutex_lock(&wq->mutex);
5676 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5677 goto out_unlock;
5678
5679use_dfl_pwq:
5680 mutex_lock(&wq->mutex);
5681 pwq = unbound_pwq(wq, -1);
5682 raw_spin_lock_irq(&pwq->pool->lock);
5683 get_pwq(pwq);
5684 raw_spin_unlock_irq(&pwq->pool->lock);
5685 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5686out_unlock:
5687 mutex_unlock(&wq->mutex);
5688 put_pwq_unlocked(old_pwq);
5689}
5690
5691static int alloc_and_link_percpu_pwqs(struct workqueue_struct *wq)
5692{
5693 struct pool_workqueue *pwq;
5694 int cpu;
5695
5696 for_each_possible_cpu(cpu) {
5697 struct worker_pool *pool = get_percpu_pool(wq, cpu);
5698
5699 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5700 if (!pwq)
5701 return -ENOMEM;
5702
5703 init_pwq(pwq, wq, pool);
5704
5705 mutex_lock(&wq->mutex);
5706 link_pwq(pwq);
5707 mutex_unlock(&wq->mutex);
5708
5709 rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq);
5710 }
5711
5712 return 0;
5713}
5714
5715static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5716{
5717 bool highpri = wq->flags & WQ_HIGHPRI;
5718 int cpu, ret;
5719
5720 lockdep_assert_held(&wq_pool_mutex);
5721
5722 wq->cpu_pwq = alloc_percpu(struct pool_workqueue __rcu *);
5723 if (!wq->cpu_pwq)
5724 goto enomem;
5725
5726 if (!(wq->flags & WQ_UNBOUND)) {
5727 ret = alloc_and_link_percpu_pwqs(wq);
5728 } else if (wq->flags & __WQ_ORDERED) {
5729 struct pool_workqueue *dfl_pwq;
5730
5731 ret = apply_workqueue_attrs_locked(wq, ordered_wq_attrs[highpri]);
5732 /* there should only be single pwq for ordering guarantee */
5733 dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5734 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node ||
5735 wq->pwqs.prev != &dfl_pwq->pwqs_node),
5736 "ordering guarantee broken for workqueue %s\n", wq->name);
5737 } else {
5738 ret = apply_workqueue_attrs_locked(wq, unbound_std_wq_attrs[highpri]);
5739 }
5740
5741 if (ret)
5742 goto enomem;
5743 return 0;
5744
5745enomem:
5746 if (wq->cpu_pwq) {
5747 for_each_possible_cpu(cpu) {
5748 struct pool_workqueue __rcu **slot;
5749 struct pool_workqueue *pwq;
5750
5751 slot = per_cpu_ptr(wq->cpu_pwq, cpu);
5752 pwq = rcu_access_pointer(*slot);
5753 if (pwq) {
5754 /*
5755 * Unlink pwq from wq->pwqs since link_pwq()
5756 * may have already added it. wq->mutex is not
5757 * needed as the wq has not been published yet.
5758 */
5759 if (!list_empty(&pwq->pwqs_node))
5760 list_del_rcu(&pwq->pwqs_node);
5761 kmem_cache_free(pwq_cache, pwq);
5762 }
5763 }
5764 free_percpu(wq->cpu_pwq);
5765 wq->cpu_pwq = NULL;
5766 }
5767 return -ENOMEM;
5768}
5769
5770static int wq_clamp_max_active(int max_active, unsigned int flags,
5771 const char *name)
5772{
5773 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5774 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5775 max_active, name, 1, WQ_MAX_ACTIVE);
5776
5777 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5778}
5779
5780/*
5781 * Workqueues which may be used during memory reclaim should have a rescuer
5782 * to guarantee forward progress.
5783 */
5784static int init_rescuer(struct workqueue_struct *wq)
5785{
5786 struct worker *rescuer;
5787 char id_buf[WORKER_ID_LEN];
5788 int ret;
5789
5790 lockdep_assert_held(&wq_pool_mutex);
5791
5792 if (!(wq->flags & WQ_MEM_RECLAIM))
5793 return 0;
5794
5795 rescuer = alloc_worker(NUMA_NO_NODE);
5796 if (!rescuer) {
5797 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5798 wq->name);
5799 return -ENOMEM;
5800 }
5801
5802 rescuer->rescue_wq = wq;
5803 format_worker_id(id_buf, sizeof(id_buf), rescuer, NULL);
5804
5805 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", id_buf);
5806 if (IS_ERR(rescuer->task)) {
5807 ret = PTR_ERR(rescuer->task);
5808 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5809 wq->name, ERR_PTR(ret));
5810 kfree(rescuer);
5811 return ret;
5812 }
5813
5814 wq->rescuer = rescuer;
5815
5816 /* initial cpumask is consistent with the detached rescuer and unbind_worker() */
5817 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
5818 kthread_bind_mask(rescuer->task, wq_unbound_cpumask);
5819 else
5820 kthread_bind_mask(rescuer->task, cpu_possible_mask);
5821
5822 wake_up_process(rescuer->task);
5823
5824 return 0;
5825}
5826
5827/**
5828 * wq_adjust_max_active - update a wq's max_active to the current setting
5829 * @wq: target workqueue
5830 *
5831 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5832 * activate inactive work items accordingly. If @wq is freezing, clear
5833 * @wq->max_active to zero.
5834 */
5835static void wq_adjust_max_active(struct workqueue_struct *wq)
5836{
5837 bool activated;
5838 int new_max, new_min;
5839
5840 lockdep_assert_held(&wq->mutex);
5841
5842 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5843 new_max = 0;
5844 new_min = 0;
5845 } else {
5846 new_max = wq->saved_max_active;
5847 new_min = wq->saved_min_active;
5848 }
5849
5850 if (wq->max_active == new_max && wq->min_active == new_min)
5851 return;
5852
5853 /*
5854 * Update @wq->max/min_active and then kick inactive work items if more
5855 * active work items are allowed. This doesn't break work item ordering
5856 * because new work items are always queued behind existing inactive
5857 * work items if there are any.
5858 */
5859 WRITE_ONCE(wq->max_active, new_max);
5860 WRITE_ONCE(wq->min_active, new_min);
5861
5862 if (wq->flags & WQ_UNBOUND)
5863 wq_update_node_max_active(wq, -1);
5864
5865 if (new_max == 0)
5866 return;
5867
5868 /*
5869 * Round-robin through pwq's activating the first inactive work item
5870 * until max_active is filled.
5871 */
5872 do {
5873 struct pool_workqueue *pwq;
5874
5875 activated = false;
5876 for_each_pwq(pwq, wq) {
5877 unsigned long irq_flags;
5878
5879 /* can be called during early boot w/ irq disabled */
5880 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
5881 if (pwq_activate_first_inactive(pwq, true)) {
5882 activated = true;
5883 kick_pool(pwq->pool);
5884 }
5885 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
5886 }
5887 } while (activated);
5888}
5889
5890__printf(1, 0)
5891static struct workqueue_struct *__alloc_workqueue(const char *fmt,
5892 unsigned int flags,
5893 int max_active, va_list args)
5894{
5895 struct workqueue_struct *wq;
5896 size_t wq_size;
5897 int name_len;
5898
5899 if (flags & WQ_BH) {
5900 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS))
5901 return NULL;
5902 if (WARN_ON_ONCE(max_active))
5903 return NULL;
5904 }
5905
5906 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5907 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5908 flags = (flags & ~WQ_PERCPU) | WQ_UNBOUND;
5909
5910 /* allocate wq and format name */
5911 if (flags & WQ_UNBOUND)
5912 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5913 else
5914 wq_size = sizeof(*wq);
5915
5916 wq = kzalloc_noprof(wq_size, GFP_KERNEL);
5917 if (!wq)
5918 return NULL;
5919
5920 wq->attrs = alloc_workqueue_attrs_noprof();
5921 if (!wq->attrs)
5922 goto err_free_wq;
5923
5924 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5925
5926 if (name_len >= WQ_NAME_LEN)
5927 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5928 wq->name);
5929
5930 /*
5931 * One among WQ_PERCPU and WQ_UNBOUND must be set, but not both.
5932 * - If neither is set, default to WQ_PERCPU
5933 * - If both are set, default to WQ_UNBOUND
5934 *
5935 * This code can be removed after workqueue are unbound by default
5936 */
5937 if (unlikely(!(flags & (WQ_UNBOUND | WQ_PERCPU)))) {
5938 WARN_ONCE(1, "workqueue: %s is using neither WQ_PERCPU or WQ_UNBOUND. "
5939 "Setting WQ_PERCPU.\n", wq->name);
5940 flags |= WQ_PERCPU;
5941 } else if (unlikely((flags & WQ_PERCPU) && (flags & WQ_UNBOUND))) {
5942 WARN_ONCE(1, "workqueue: %s uses both WQ_PERCPU and WQ_UNBOUND. "
5943 "Dropped WQ_PERCPU, keeping WQ_UNBOUND.\n", wq->name);
5944 flags &= ~WQ_PERCPU;
5945 }
5946
5947 if (flags & WQ_BH) {
5948 /*
5949 * BH workqueues always share a single execution context per CPU
5950 * and don't impose any max_active limit.
5951 */
5952 max_active = INT_MAX;
5953 } else {
5954 max_active = max_active ?: WQ_DFL_ACTIVE;
5955 max_active = wq_clamp_max_active(max_active, flags, wq->name);
5956 }
5957
5958 /* init wq */
5959 wq->flags = flags;
5960 wq->max_active = max_active;
5961 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5962 wq->saved_max_active = wq->max_active;
5963 wq->saved_min_active = wq->min_active;
5964 mutex_init(&wq->mutex);
5965 atomic_set(&wq->nr_pwqs_to_flush, 0);
5966 INIT_LIST_HEAD(&wq->pwqs);
5967 INIT_LIST_HEAD(&wq->flusher_queue);
5968 INIT_LIST_HEAD(&wq->flusher_overflow);
5969 INIT_LIST_HEAD(&wq->maydays);
5970
5971 INIT_LIST_HEAD(&wq->list);
5972
5973 if (flags & WQ_UNBOUND) {
5974 if (alloc_node_nr_active(wq->node_nr_active) < 0)
5975 goto err_free_wq;
5976 }
5977
5978 /*
5979 * wq_pool_mutex protects the workqueues list, allocations of PWQs,
5980 * and the global freeze state.
5981 */
5982 mutex_lock(&wq_pool_mutex);
5983
5984 if (alloc_and_link_pwqs(wq) < 0)
5985 goto err_unlock_free_node_nr_active;
5986
5987 mutex_lock(&wq->mutex);
5988 wq_adjust_max_active(wq);
5989 mutex_unlock(&wq->mutex);
5990
5991 list_add_tail_rcu(&wq->list, &workqueues);
5992
5993 if (wq_online && init_rescuer(wq) < 0)
5994 goto err_unlock_destroy;
5995
5996 mutex_unlock(&wq_pool_mutex);
5997
5998 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5999 goto err_destroy;
6000
6001 return wq;
6002
6003err_unlock_free_node_nr_active:
6004 mutex_unlock(&wq_pool_mutex);
6005 /*
6006 * Failed alloc_and_link_pwqs() may leave pending pwq->release_work,
6007 * flushing the pwq_release_worker ensures that the pwq_release_workfn()
6008 * completes before calling kfree(wq).
6009 */
6010 if (wq->flags & WQ_UNBOUND) {
6011 kthread_flush_worker(pwq_release_worker);
6012 free_node_nr_active(wq->node_nr_active);
6013 }
6014err_free_wq:
6015 free_workqueue_attrs(wq->attrs);
6016 kfree(wq);
6017 return NULL;
6018err_unlock_destroy:
6019 mutex_unlock(&wq_pool_mutex);
6020err_destroy:
6021 destroy_workqueue(wq);
6022 return NULL;
6023}
6024
6025__printf(1, 0)
6026static struct workqueue_struct *alloc_workqueue_va(const char *fmt,
6027 unsigned int flags,
6028 int max_active,
6029 va_list args)
6030{
6031 struct workqueue_struct *wq;
6032
6033 wq = __alloc_workqueue(fmt, flags, max_active, args);
6034 if (wq)
6035 wq_init_lockdep(wq);
6036
6037 return wq;
6038}
6039
6040__printf(1, 4)
6041struct workqueue_struct *alloc_workqueue_noprof(const char *fmt,
6042 unsigned int flags,
6043 int max_active, ...)
6044{
6045 struct workqueue_struct *wq;
6046 va_list args;
6047
6048 va_start(args, max_active);
6049 wq = alloc_workqueue_va(fmt, flags, max_active, args);
6050 va_end(args);
6051
6052 return wq;
6053}
6054EXPORT_SYMBOL_GPL(alloc_workqueue_noprof);
6055
6056static void devm_workqueue_release(void *res)
6057{
6058 destroy_workqueue(res);
6059}
6060
6061__printf(2, 5) struct workqueue_struct *
6062devm_alloc_workqueue_noprof(struct device *dev, const char *fmt,
6063 unsigned int flags, int max_active, ...)
6064{
6065 struct workqueue_struct *wq;
6066 va_list args;
6067 int ret;
6068
6069 va_start(args, max_active);
6070 wq = alloc_workqueue_va(fmt, flags, max_active, args);
6071 va_end(args);
6072 if (!wq)
6073 return NULL;
6074
6075 ret = devm_add_action_or_reset(dev, devm_workqueue_release, wq);
6076 if (ret)
6077 return NULL;
6078
6079 return wq;
6080}
6081EXPORT_SYMBOL_GPL(devm_alloc_workqueue_noprof);
6082
6083#ifdef CONFIG_LOCKDEP
6084__printf(1, 5)
6085struct workqueue_struct *
6086alloc_workqueue_lockdep_map(const char *fmt, unsigned int flags,
6087 int max_active, struct lockdep_map *lockdep_map, ...)
6088{
6089 struct workqueue_struct *wq;
6090 va_list args;
6091
6092 va_start(args, lockdep_map);
6093 wq = __alloc_workqueue(fmt, flags, max_active, args);
6094 va_end(args);
6095 if (!wq)
6096 return NULL;
6097
6098 wq->lockdep_map = lockdep_map;
6099
6100 return wq;
6101}
6102EXPORT_SYMBOL_GPL(alloc_workqueue_lockdep_map);
6103#endif
6104
6105static bool pwq_busy(struct pool_workqueue *pwq)
6106{
6107 int i;
6108
6109 for (i = 0; i < WORK_NR_COLORS; i++)
6110 if (pwq->nr_in_flight[i])
6111 return true;
6112
6113 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
6114 return true;
6115 if (!pwq_is_empty(pwq))
6116 return true;
6117
6118 return false;
6119}
6120
6121/**
6122 * destroy_workqueue - safely terminate a workqueue
6123 * @wq: target workqueue
6124 *
6125 * Safely destroy a workqueue. All work currently pending will be done first.
6126 *
6127 * This function does NOT guarantee that non-pending work that has been
6128 * submitted with queue_delayed_work() and similar functions will be done
6129 * before destroying the workqueue. The fundamental problem is that, currently,
6130 * the workqueue has no way of accessing non-pending delayed_work. delayed_work
6131 * is only linked on the timer-side. All delayed_work must, therefore, be
6132 * canceled before calling this function.
6133 *
6134 * TODO: It would be better if the problem described above wouldn't exist and
6135 * destroy_workqueue() would cleanly cancel all pending and non-pending
6136 * delayed_work.
6137 */
6138void destroy_workqueue(struct workqueue_struct *wq)
6139{
6140 struct pool_workqueue *pwq;
6141 int cpu;
6142
6143 /*
6144 * Remove it from sysfs first so that sanity check failure doesn't
6145 * lead to sysfs name conflicts.
6146 */
6147 workqueue_sysfs_unregister(wq);
6148
6149 /* mark the workqueue destruction is in progress */
6150 mutex_lock(&wq->mutex);
6151 wq->flags |= __WQ_DESTROYING;
6152 mutex_unlock(&wq->mutex);
6153
6154 /* drain it before proceeding with destruction */
6155 drain_workqueue(wq);
6156
6157 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
6158 if (wq->rescuer) {
6159 /* rescuer will empty maydays list before exiting */
6160 kthread_stop(wq->rescuer->task);
6161 kfree(wq->rescuer);
6162 wq->rescuer = NULL;
6163 }
6164
6165 /*
6166 * Sanity checks - grab all the locks so that we wait for all
6167 * in-flight operations which may do put_pwq().
6168 */
6169 mutex_lock(&wq_pool_mutex);
6170 mutex_lock(&wq->mutex);
6171 for_each_pwq(pwq, wq) {
6172 raw_spin_lock_irq(&pwq->pool->lock);
6173 if (WARN_ON(pwq_busy(pwq))) {
6174 pr_warn("%s: %s has the following busy pwq\n",
6175 __func__, wq->name);
6176 show_pwq(pwq);
6177 raw_spin_unlock_irq(&pwq->pool->lock);
6178 mutex_unlock(&wq->mutex);
6179 mutex_unlock(&wq_pool_mutex);
6180 show_one_workqueue(wq);
6181 return;
6182 }
6183 raw_spin_unlock_irq(&pwq->pool->lock);
6184 }
6185 mutex_unlock(&wq->mutex);
6186
6187 /*
6188 * wq list is used to freeze wq, remove from list after
6189 * flushing is complete in case freeze races us.
6190 */
6191 list_del_rcu(&wq->list);
6192 mutex_unlock(&wq_pool_mutex);
6193
6194 /*
6195 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
6196 * to put the base refs. @wq will be auto-destroyed from the last
6197 * pwq_put. RCU read lock prevents @wq from going away from under us.
6198 */
6199 rcu_read_lock();
6200
6201 for_each_possible_cpu(cpu) {
6202 put_pwq_unlocked(unbound_pwq(wq, cpu));
6203 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
6204 }
6205
6206 put_pwq_unlocked(unbound_pwq(wq, -1));
6207 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
6208
6209 rcu_read_unlock();
6210}
6211EXPORT_SYMBOL_GPL(destroy_workqueue);
6212
6213/**
6214 * workqueue_set_max_active - adjust max_active of a workqueue
6215 * @wq: target workqueue
6216 * @max_active: new max_active value.
6217 *
6218 * Set max_active of @wq to @max_active. See the alloc_workqueue() function
6219 * comment.
6220 *
6221 * CONTEXT:
6222 * Don't call from IRQ context.
6223 */
6224void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
6225{
6226 /* max_active doesn't mean anything for BH workqueues */
6227 if (WARN_ON(wq->flags & WQ_BH))
6228 return;
6229 /* disallow meddling with max_active for ordered workqueues */
6230 if (WARN_ON(wq->flags & __WQ_ORDERED))
6231 return;
6232
6233 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
6234
6235 mutex_lock(&wq->mutex);
6236
6237 wq->saved_max_active = max_active;
6238 if (wq->flags & WQ_UNBOUND)
6239 wq->saved_min_active = min(wq->saved_min_active, max_active);
6240
6241 wq_adjust_max_active(wq);
6242
6243 mutex_unlock(&wq->mutex);
6244}
6245EXPORT_SYMBOL_GPL(workqueue_set_max_active);
6246
6247/**
6248 * workqueue_set_min_active - adjust min_active of an unbound workqueue
6249 * @wq: target unbound workqueue
6250 * @min_active: new min_active value
6251 *
6252 * Set min_active of an unbound workqueue. Unlike other types of workqueues, an
6253 * unbound workqueue is not guaranteed to be able to process max_active
6254 * interdependent work items. Instead, an unbound workqueue is guaranteed to be
6255 * able to process min_active number of interdependent work items which is
6256 * %WQ_DFL_MIN_ACTIVE by default.
6257 *
6258 * Use this function to adjust the min_active value between 0 and the current
6259 * max_active.
6260 */
6261void workqueue_set_min_active(struct workqueue_struct *wq, int min_active)
6262{
6263 /* min_active is only meaningful for non-ordered unbound workqueues */
6264 if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) !=
6265 WQ_UNBOUND))
6266 return;
6267
6268 mutex_lock(&wq->mutex);
6269 wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active);
6270 wq_adjust_max_active(wq);
6271 mutex_unlock(&wq->mutex);
6272}
6273
6274/**
6275 * current_work - retrieve %current task's work struct
6276 *
6277 * Determine if %current task is a workqueue worker and what it's working on.
6278 * Useful to find out the context that the %current task is running in.
6279 *
6280 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
6281 */
6282struct work_struct *current_work(void)
6283{
6284 struct worker *worker = current_wq_worker();
6285
6286 return worker ? worker->current_work : NULL;
6287}
6288EXPORT_SYMBOL(current_work);
6289
6290/**
6291 * current_is_workqueue_rescuer - is %current workqueue rescuer?
6292 *
6293 * Determine whether %current is a workqueue rescuer. Can be used from
6294 * work functions to determine whether it's being run off the rescuer task.
6295 *
6296 * Return: %true if %current is a workqueue rescuer. %false otherwise.
6297 */
6298bool current_is_workqueue_rescuer(void)
6299{
6300 struct worker *worker = current_wq_worker();
6301
6302 return worker && worker->rescue_wq;
6303}
6304
6305/**
6306 * current_is_workqueue_mem_reclaim - is %current a %WQ_MEM_RECLAIM worker?
6307 *
6308 * Determine whether %current is a workqueue worker executing on a workqueue
6309 * created with %WQ_MEM_RECLAIM. This mirrors the condition that
6310 * check_flush_dependency() warns on: flushing (or otherwise waiting on) a
6311 * !WQ_MEM_RECLAIM workqueue from such a context breaks the forward-progress
6312 * guarantee and can deadlock. Callers that may recurse into such a flush --
6313 * e.g. NFS LOCALIO submitting into a stacked filesystem that flushes its own
6314 * !WQ_MEM_RECLAIM workqueue -- can use this to decide whether they must defer
6315 * the work to a !WQ_MEM_RECLAIM workqueue rather than run it inline.
6316 *
6317 * Return: %true if %current is a %WQ_MEM_RECLAIM worker. %false otherwise.
6318 */
6319bool current_is_workqueue_mem_reclaim(void)
6320{
6321 struct worker *worker = current_wq_worker();
6322
6323 return worker && worker->current_pwq &&
6324 ((worker->current_pwq->wq->flags &
6325 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM);
6326}
6327EXPORT_SYMBOL_GPL(current_is_workqueue_mem_reclaim);
6328
6329/**
6330 * workqueue_congested - test whether a workqueue is congested
6331 * @cpu: CPU in question
6332 * @wq: target workqueue
6333 *
6334 * Test whether @wq's cpu workqueue for @cpu is congested. There is
6335 * no synchronization around this function and the test result is
6336 * unreliable and only useful as advisory hints or for debugging.
6337 *
6338 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
6339 *
6340 * With the exception of ordered workqueues, all workqueues have per-cpu
6341 * pool_workqueues, each with its own congested state. A workqueue being
6342 * congested on one CPU doesn't mean that the workqueue is contested on any
6343 * other CPUs.
6344 *
6345 * Return:
6346 * %true if congested, %false otherwise.
6347 */
6348bool workqueue_congested(int cpu, struct workqueue_struct *wq)
6349{
6350 struct pool_workqueue *pwq;
6351 bool ret;
6352
6353 preempt_disable();
6354
6355 if (cpu == WORK_CPU_UNBOUND)
6356 cpu = smp_processor_id();
6357
6358 pwq = rcu_dereference_sched(*per_cpu_ptr(wq->cpu_pwq, cpu));
6359 ret = !list_empty(&pwq->inactive_works);
6360
6361 preempt_enable();
6362
6363 return ret;
6364}
6365EXPORT_SYMBOL_GPL(workqueue_congested);
6366
6367/**
6368 * work_busy - test whether a work is currently pending or running
6369 * @work: the work to be tested
6370 *
6371 * Test whether @work is currently pending or running. There is no
6372 * synchronization around this function and the test result is
6373 * unreliable and only useful as advisory hints or for debugging.
6374 *
6375 * Return:
6376 * OR'd bitmask of WORK_BUSY_* bits.
6377 */
6378unsigned int work_busy(struct work_struct *work)
6379{
6380 struct worker_pool *pool;
6381 unsigned long irq_flags;
6382 unsigned int ret = 0;
6383
6384 if (work_pending(work))
6385 ret |= WORK_BUSY_PENDING;
6386
6387 rcu_read_lock();
6388 pool = get_work_pool(work);
6389 if (pool) {
6390 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6391 if (find_worker_executing_work(pool, work))
6392 ret |= WORK_BUSY_RUNNING;
6393 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6394 }
6395 rcu_read_unlock();
6396
6397 return ret;
6398}
6399EXPORT_SYMBOL_GPL(work_busy);
6400
6401/**
6402 * set_worker_desc - set description for the current work item
6403 * @fmt: printf-style format string
6404 * @...: arguments for the format string
6405 *
6406 * This function can be called by a running work function to describe what
6407 * the work item is about. If the worker task gets dumped, this
6408 * information will be printed out together to help debugging. The
6409 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
6410 */
6411void set_worker_desc(const char *fmt, ...)
6412{
6413 struct worker *worker = current_wq_worker();
6414 va_list args;
6415
6416 if (worker) {
6417 va_start(args, fmt);
6418 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
6419 va_end(args);
6420 }
6421}
6422EXPORT_SYMBOL_GPL(set_worker_desc);
6423
6424/**
6425 * print_worker_info - print out worker information and description
6426 * @log_lvl: the log level to use when printing
6427 * @task: target task
6428 *
6429 * If @task is a worker and currently executing a work item, print out the
6430 * name of the workqueue being serviced and worker description set with
6431 * set_worker_desc() by the currently executing work item.
6432 *
6433 * This function can be safely called on any task as long as the
6434 * task_struct itself is accessible. While safe, this function isn't
6435 * synchronized and may print out mixups or garbages of limited length.
6436 */
6437void print_worker_info(const char *log_lvl, struct task_struct *task)
6438{
6439 work_func_t fn = NULL;
6440 char name[WQ_NAME_LEN] = { };
6441 char desc[WORKER_DESC_LEN] = { };
6442 struct pool_workqueue *pwq = NULL;
6443 struct workqueue_struct *wq = NULL;
6444 struct worker *worker;
6445
6446 if (!(task->flags & PF_WQ_WORKER))
6447 return;
6448
6449 /*
6450 * This function is called without any synchronization and @task
6451 * could be in any state. Be careful with dereferences.
6452 */
6453 worker = kthread_probe_data(task);
6454
6455 /*
6456 * Carefully copy the associated workqueue's workfn, name and desc.
6457 * Keep the original last '\0' in case the original is garbage.
6458 */
6459 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
6460 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
6461 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
6462 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
6463 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
6464
6465 if (fn || name[0] || desc[0]) {
6466 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
6467 if (strcmp(name, desc))
6468 pr_cont(" (%s)", desc);
6469 pr_cont("\n");
6470 }
6471}
6472
6473static void pr_cont_pool_info(struct worker_pool *pool)
6474{
6475 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
6476 if (pool->node != NUMA_NO_NODE)
6477 pr_cont(" node=%d", pool->node);
6478 pr_cont(" flags=0x%x", pool->flags);
6479 if (pool->flags & POOL_BH)
6480 pr_cont(" bh%s",
6481 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6482 else
6483 pr_cont(" nice=%d", pool->attrs->nice);
6484}
6485
6486static void pr_cont_worker_id(struct worker *worker)
6487{
6488 struct worker_pool *pool = worker->pool;
6489
6490 if (pool->flags & POOL_BH)
6491 pr_cont("bh%s",
6492 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6493 else
6494 pr_cont("%d%s", task_pid_nr(worker->task),
6495 worker->rescue_wq ? "(RESCUER)" : "");
6496}
6497
6498struct pr_cont_work_struct {
6499 bool comma;
6500 work_func_t func;
6501 long ctr;
6502};
6503
6504static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
6505{
6506 if (!pcwsp->ctr)
6507 goto out_record;
6508 if (func == pcwsp->func) {
6509 pcwsp->ctr++;
6510 return;
6511 }
6512 if (pcwsp->ctr == 1)
6513 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
6514 else
6515 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
6516 pcwsp->ctr = 0;
6517out_record:
6518 if ((long)func == -1L)
6519 return;
6520 pcwsp->comma = comma;
6521 pcwsp->func = func;
6522 pcwsp->ctr = 1;
6523}
6524
6525static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
6526{
6527 if (work->func == wq_barrier_func) {
6528 struct wq_barrier *barr;
6529
6530 barr = container_of(work, struct wq_barrier, work);
6531
6532 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6533 pr_cont("%s BAR(%d)", comma ? "," : "",
6534 task_pid_nr(barr->task));
6535 } else {
6536 if (!comma)
6537 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6538 pr_cont_work_flush(comma, work->func, pcwsp);
6539 }
6540}
6541
6542static void show_pwq(struct pool_workqueue *pwq)
6543{
6544 struct pr_cont_work_struct pcws = { .ctr = 0, };
6545 struct worker_pool *pool = pwq->pool;
6546 struct work_struct *work;
6547 struct worker *worker;
6548 bool has_in_flight = false, has_pending = false;
6549 int bkt;
6550
6551 pr_info(" pwq %d:", pool->id);
6552 pr_cont_pool_info(pool);
6553
6554 pr_cont(" active=%d refcnt=%d%s\n",
6555 pwq->nr_active, pwq->refcnt,
6556 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6557
6558 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6559 if (worker->current_pwq == pwq) {
6560 has_in_flight = true;
6561 break;
6562 }
6563 }
6564 if (has_in_flight) {
6565 bool comma = false;
6566
6567 pr_info(" in-flight:");
6568 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6569 if (worker->current_pwq != pwq)
6570 continue;
6571
6572 pr_cont(" %s", comma ? "," : "");
6573 pr_cont_worker_id(worker);
6574 pr_cont(":%ps", worker->current_func);
6575 pr_cont(" for %us",
6576 jiffies_to_msecs(jiffies - worker->current_start) / 1000);
6577 list_for_each_entry(work, &worker->scheduled, entry)
6578 pr_cont_work(false, work, &pcws);
6579 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6580 comma = true;
6581 }
6582 pr_cont("\n");
6583 }
6584
6585 list_for_each_entry(work, &pool->worklist, entry) {
6586 if (get_work_pwq(work) == pwq) {
6587 has_pending = true;
6588 break;
6589 }
6590 }
6591 if (has_pending) {
6592 bool comma = false;
6593
6594 pr_info(" pending:");
6595 list_for_each_entry(work, &pool->worklist, entry) {
6596 if (get_work_pwq(work) != pwq)
6597 continue;
6598
6599 pr_cont_work(comma, work, &pcws);
6600 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6601 }
6602 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6603 pr_cont("\n");
6604 }
6605
6606 if (!list_empty(&pwq->inactive_works)) {
6607 bool comma = false;
6608
6609 pr_info(" inactive:");
6610 list_for_each_entry(work, &pwq->inactive_works, entry) {
6611 pr_cont_work(comma, work, &pcws);
6612 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6613 }
6614 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6615 pr_cont("\n");
6616 }
6617}
6618
6619/**
6620 * show_one_workqueue - dump state of specified workqueue
6621 * @wq: workqueue whose state will be printed
6622 */
6623void show_one_workqueue(struct workqueue_struct *wq)
6624{
6625 struct pool_workqueue *pwq;
6626 bool idle = true;
6627 unsigned long irq_flags;
6628
6629 for_each_pwq(pwq, wq) {
6630 if (!pwq_is_empty(pwq)) {
6631 idle = false;
6632 break;
6633 }
6634 }
6635 if (idle) /* Nothing to print for idle workqueue */
6636 return;
6637
6638 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6639
6640 for_each_pwq(pwq, wq) {
6641 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
6642 if (!pwq_is_empty(pwq)) {
6643 /*
6644 * Defer printing to avoid deadlocks in console
6645 * drivers that queue work while holding locks
6646 * also taken in their write paths.
6647 */
6648 printk_deferred_enter();
6649 show_pwq(pwq);
6650 printk_deferred_exit();
6651 }
6652 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
6653 /*
6654 * We could be printing a lot from atomic context, e.g.
6655 * sysrq-t -> show_all_workqueues(). Avoid triggering
6656 * hard lockup.
6657 */
6658 touch_nmi_watchdog();
6659 }
6660
6661}
6662
6663/**
6664 * show_one_worker_pool - dump state of specified worker pool
6665 * @pool: worker pool whose state will be printed
6666 */
6667static void show_one_worker_pool(struct worker_pool *pool)
6668{
6669 struct worker *worker;
6670 bool first = true;
6671 unsigned long irq_flags;
6672 unsigned long hung = 0;
6673
6674 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6675 if (pool->nr_workers == pool->nr_idle)
6676 goto next_pool;
6677
6678 /* How long the first pending work is waiting for a worker. */
6679 if (!list_empty(&pool->worklist))
6680 hung = jiffies_to_msecs(jiffies - pool->last_progress_ts) / 1000;
6681
6682 /*
6683 * Defer printing to avoid deadlocks in console drivers that
6684 * queue work while holding locks also taken in their write
6685 * paths.
6686 */
6687 printk_deferred_enter();
6688 pr_info("pool %d:", pool->id);
6689 pr_cont_pool_info(pool);
6690 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
6691 if (pool->manager)
6692 pr_cont(" manager: %d",
6693 task_pid_nr(pool->manager->task));
6694 list_for_each_entry(worker, &pool->idle_list, entry) {
6695 pr_cont(" %s", first ? "idle: " : "");
6696 pr_cont_worker_id(worker);
6697 first = false;
6698 }
6699 pr_cont("\n");
6700 printk_deferred_exit();
6701next_pool:
6702 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6703 /*
6704 * We could be printing a lot from atomic context, e.g.
6705 * sysrq-t -> show_all_workqueues(). Avoid triggering
6706 * hard lockup.
6707 */
6708 touch_nmi_watchdog();
6709
6710}
6711
6712/**
6713 * show_all_workqueues - dump workqueue state
6714 *
6715 * Called from a sysrq handler and prints out all busy workqueues and pools.
6716 */
6717void show_all_workqueues(void)
6718{
6719 struct workqueue_struct *wq;
6720 struct worker_pool *pool;
6721 int pi;
6722
6723 rcu_read_lock();
6724
6725 pr_info("Showing busy workqueues and worker pools:\n");
6726
6727 list_for_each_entry_rcu(wq, &workqueues, list)
6728 show_one_workqueue(wq);
6729
6730 for_each_pool(pool, pi)
6731 show_one_worker_pool(pool);
6732
6733 rcu_read_unlock();
6734}
6735
6736/**
6737 * show_freezable_workqueues - dump freezable workqueue state
6738 *
6739 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
6740 * still busy.
6741 */
6742void show_freezable_workqueues(void)
6743{
6744 struct workqueue_struct *wq;
6745
6746 rcu_read_lock();
6747
6748 pr_info("Showing freezable workqueues that are still busy:\n");
6749
6750 list_for_each_entry_rcu(wq, &workqueues, list) {
6751 if (!(wq->flags & WQ_FREEZABLE))
6752 continue;
6753 show_one_workqueue(wq);
6754 }
6755
6756 rcu_read_unlock();
6757}
6758
6759/* used to show worker information through /proc/PID/{comm,stat,status} */
6760void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
6761{
6762 /* stabilize PF_WQ_WORKER and worker pool association */
6763 mutex_lock(&wq_pool_attach_mutex);
6764
6765 if (task->flags & PF_WQ_WORKER) {
6766 struct worker *worker = kthread_data(task);
6767 struct worker_pool *pool = worker->pool;
6768 int off;
6769
6770 off = format_worker_id(buf, size, worker, pool);
6771
6772 if (pool) {
6773 raw_spin_lock_irq(&pool->lock);
6774 /*
6775 * ->desc tracks information (wq name or
6776 * set_worker_desc()) for the latest execution. If
6777 * current, prepend '+', otherwise '-'.
6778 */
6779 if (worker->desc[0] != '\0') {
6780 if (worker->current_work)
6781 scnprintf(buf + off, size - off, "+%s",
6782 worker->desc);
6783 else
6784 scnprintf(buf + off, size - off, "-%s",
6785 worker->desc);
6786 }
6787 raw_spin_unlock_irq(&pool->lock);
6788 }
6789 } else {
6790 strscpy(buf, task->comm, size);
6791 }
6792
6793 mutex_unlock(&wq_pool_attach_mutex);
6794}
6795
6796#ifdef CONFIG_SMP
6797
6798/*
6799 * CPU hotplug.
6800 *
6801 * There are two challenges in supporting CPU hotplug. Firstly, there
6802 * are a lot of assumptions on strong associations among work, pwq and
6803 * pool which make migrating pending and scheduled works very
6804 * difficult to implement without impacting hot paths. Secondly,
6805 * worker pools serve mix of short, long and very long running works making
6806 * blocked draining impractical.
6807 *
6808 * This is solved by allowing the pools to be disassociated from the CPU
6809 * running as an unbound one and allowing it to be reattached later if the
6810 * cpu comes back online.
6811 */
6812
6813static void unbind_workers(int cpu)
6814{
6815 struct worker_pool *pool;
6816 struct worker *worker;
6817
6818 for_each_cpu_worker_pool(pool, cpu) {
6819 mutex_lock(&wq_pool_attach_mutex);
6820 raw_spin_lock_irq(&pool->lock);
6821
6822 /*
6823 * We've blocked all attach/detach operations. Make all workers
6824 * unbound and set DISASSOCIATED. Before this, all workers
6825 * must be on the cpu. After this, they may become diasporas.
6826 * And the preemption disabled section in their sched callbacks
6827 * are guaranteed to see WORKER_UNBOUND since the code here
6828 * is on the same cpu.
6829 */
6830 for_each_pool_worker(worker, pool)
6831 worker->flags |= WORKER_UNBOUND;
6832
6833 pool->flags |= POOL_DISASSOCIATED;
6834
6835 /*
6836 * The handling of nr_running in sched callbacks are disabled
6837 * now. Zap nr_running. After this, nr_running stays zero and
6838 * need_more_worker() and keep_working() are always true as
6839 * long as the worklist is not empty. This pool now behaves as
6840 * an unbound (in terms of concurrency management) pool which
6841 * are served by workers tied to the pool.
6842 */
6843 pool->nr_running = 0;
6844
6845 /*
6846 * With concurrency management just turned off, a busy
6847 * worker blocking could lead to lengthy stalls. Kick off
6848 * unbound chain execution of currently pending work items.
6849 */
6850 kick_pool(pool);
6851
6852 raw_spin_unlock_irq(&pool->lock);
6853
6854 for_each_pool_worker(worker, pool)
6855 unbind_worker(worker);
6856
6857 mutex_unlock(&wq_pool_attach_mutex);
6858 }
6859}
6860
6861/**
6862 * rebind_workers - rebind all workers of a pool to the associated CPU
6863 * @pool: pool of interest
6864 *
6865 * @pool->cpu is coming online. Rebind all workers to the CPU.
6866 */
6867static void rebind_workers(struct worker_pool *pool)
6868{
6869 struct worker *worker;
6870
6871 lockdep_assert_held(&wq_pool_attach_mutex);
6872
6873 /*
6874 * Restore CPU affinity of all workers. As all idle workers should
6875 * be on the run-queue of the associated CPU before any local
6876 * wake-ups for concurrency management happen, restore CPU affinity
6877 * of all workers first and then clear UNBOUND. As we're called
6878 * from CPU_ONLINE, the following shouldn't fail.
6879 */
6880 for_each_pool_worker(worker, pool) {
6881 kthread_set_per_cpu(worker->task, pool->cpu);
6882 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6883 pool_allowed_cpus(pool)) < 0);
6884 }
6885
6886 raw_spin_lock_irq(&pool->lock);
6887
6888 pool->flags &= ~POOL_DISASSOCIATED;
6889
6890 for_each_pool_worker(worker, pool) {
6891 unsigned int worker_flags = worker->flags;
6892
6893 /*
6894 * We want to clear UNBOUND but can't directly call
6895 * worker_clr_flags() or adjust nr_running. Atomically
6896 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6897 * @worker will clear REBOUND using worker_clr_flags() when
6898 * it initiates the next execution cycle thus restoring
6899 * concurrency management. Note that when or whether
6900 * @worker clears REBOUND doesn't affect correctness.
6901 *
6902 * WRITE_ONCE() is necessary because @worker->flags may be
6903 * tested without holding any lock in
6904 * wq_worker_running(). Without it, NOT_RUNNING test may
6905 * fail incorrectly leading to premature concurrency
6906 * management operations.
6907 */
6908 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6909 worker_flags |= WORKER_REBOUND;
6910 worker_flags &= ~WORKER_UNBOUND;
6911 WRITE_ONCE(worker->flags, worker_flags);
6912 }
6913
6914 raw_spin_unlock_irq(&pool->lock);
6915}
6916
6917/**
6918 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6919 * @pool: unbound pool of interest
6920 * @cpu: the CPU which is coming up
6921 *
6922 * An unbound pool may end up with a cpumask which doesn't have any online
6923 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6924 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6925 * online CPU before, cpus_allowed of all its workers should be restored.
6926 */
6927static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6928{
6929 static cpumask_t cpumask;
6930 struct worker *worker;
6931
6932 lockdep_assert_held(&wq_pool_attach_mutex);
6933
6934 /* is @cpu allowed for @pool? */
6935 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6936 return;
6937
6938 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6939
6940 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6941 for_each_pool_worker(worker, pool)
6942 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6943}
6944
6945int workqueue_prepare_cpu(unsigned int cpu)
6946{
6947 struct worker_pool *pool;
6948
6949 for_each_cpu_worker_pool(pool, cpu) {
6950 if (pool->nr_workers)
6951 continue;
6952 if (!create_worker(pool))
6953 return -ENOMEM;
6954 }
6955 return 0;
6956}
6957
6958int workqueue_online_cpu(unsigned int cpu)
6959{
6960 struct worker_pool *pool;
6961 struct workqueue_struct *wq;
6962 int pi;
6963
6964 mutex_lock(&wq_pool_mutex);
6965
6966 cpumask_set_cpu(cpu, wq_online_cpumask);
6967
6968 for_each_pool(pool, pi) {
6969 /* BH pools aren't affected by hotplug */
6970 if (pool->flags & POOL_BH)
6971 continue;
6972
6973 mutex_lock(&wq_pool_attach_mutex);
6974 if (pool->cpu == cpu)
6975 rebind_workers(pool);
6976 else if (pool->cpu < 0)
6977 restore_unbound_workers_cpumask(pool, cpu);
6978 mutex_unlock(&wq_pool_attach_mutex);
6979 }
6980
6981 /* update pod affinity of unbound workqueues */
6982 list_for_each_entry(wq, &workqueues, list) {
6983 struct workqueue_attrs *attrs = wq->attrs;
6984
6985 if (wq->flags & WQ_UNBOUND) {
6986 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6987 int tcpu;
6988
6989 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6990 unbound_wq_update_pwq(wq, tcpu);
6991
6992 mutex_lock(&wq->mutex);
6993 wq_update_node_max_active(wq, -1);
6994 mutex_unlock(&wq->mutex);
6995 }
6996 }
6997
6998 mutex_unlock(&wq_pool_mutex);
6999 return 0;
7000}
7001
7002int workqueue_offline_cpu(unsigned int cpu)
7003{
7004 struct workqueue_struct *wq;
7005
7006 /* unbinding per-cpu workers should happen on the local CPU */
7007 if (WARN_ON(cpu != smp_processor_id()))
7008 return -1;
7009
7010 unbind_workers(cpu);
7011
7012 /* update pod affinity of unbound workqueues */
7013 mutex_lock(&wq_pool_mutex);
7014
7015 cpumask_clear_cpu(cpu, wq_online_cpumask);
7016
7017 list_for_each_entry(wq, &workqueues, list) {
7018 struct workqueue_attrs *attrs = wq->attrs;
7019
7020 if (wq->flags & WQ_UNBOUND) {
7021 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
7022 int tcpu;
7023
7024 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
7025 unbound_wq_update_pwq(wq, tcpu);
7026
7027 mutex_lock(&wq->mutex);
7028 wq_update_node_max_active(wq, cpu);
7029 mutex_unlock(&wq->mutex);
7030 }
7031 }
7032 mutex_unlock(&wq_pool_mutex);
7033
7034 return 0;
7035}
7036
7037struct work_for_cpu {
7038 struct work_struct work;
7039 long (*fn)(void *);
7040 void *arg;
7041 long ret;
7042};
7043
7044static void work_for_cpu_fn(struct work_struct *work)
7045{
7046 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
7047
7048 wfc->ret = wfc->fn(wfc->arg);
7049}
7050
7051/**
7052 * work_on_cpu_key - run a function in thread context on a particular cpu
7053 * @cpu: the cpu to run on
7054 * @fn: the function to run
7055 * @arg: the function arg
7056 * @key: The lock class key for lock debugging purposes
7057 *
7058 * It is up to the caller to ensure that the cpu doesn't go offline.
7059 * The caller must not hold any locks which would prevent @fn from completing.
7060 *
7061 * Return: The value @fn returns.
7062 */
7063long work_on_cpu_key(int cpu, long (*fn)(void *),
7064 void *arg, struct lock_class_key *key)
7065{
7066 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
7067
7068 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
7069 schedule_work_on(cpu, &wfc.work);
7070 flush_work(&wfc.work);
7071 destroy_work_on_stack(&wfc.work);
7072 return wfc.ret;
7073}
7074EXPORT_SYMBOL_GPL(work_on_cpu_key);
7075#endif /* CONFIG_SMP */
7076
7077#ifdef CONFIG_FREEZER
7078
7079/**
7080 * freeze_workqueues_begin - begin freezing workqueues
7081 *
7082 * Start freezing workqueues. After this function returns, all freezable
7083 * workqueues will queue new works to their inactive_works list instead of
7084 * pool->worklist.
7085 *
7086 * CONTEXT:
7087 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
7088 */
7089void freeze_workqueues_begin(void)
7090{
7091 struct workqueue_struct *wq;
7092
7093 mutex_lock(&wq_pool_mutex);
7094
7095 WARN_ON_ONCE(workqueue_freezing);
7096 workqueue_freezing = true;
7097
7098 list_for_each_entry(wq, &workqueues, list) {
7099 mutex_lock(&wq->mutex);
7100 wq_adjust_max_active(wq);
7101 mutex_unlock(&wq->mutex);
7102 }
7103
7104 mutex_unlock(&wq_pool_mutex);
7105}
7106
7107/**
7108 * freeze_workqueues_busy - are freezable workqueues still busy?
7109 *
7110 * Check whether freezing is complete. This function must be called
7111 * between freeze_workqueues_begin() and thaw_workqueues().
7112 *
7113 * CONTEXT:
7114 * Grabs and releases wq_pool_mutex.
7115 *
7116 * Return:
7117 * %true if some freezable workqueues are still busy. %false if freezing
7118 * is complete.
7119 */
7120bool freeze_workqueues_busy(void)
7121{
7122 bool busy = false;
7123 struct workqueue_struct *wq;
7124 struct pool_workqueue *pwq;
7125
7126 mutex_lock(&wq_pool_mutex);
7127
7128 WARN_ON_ONCE(!workqueue_freezing);
7129
7130 list_for_each_entry(wq, &workqueues, list) {
7131 if (!(wq->flags & WQ_FREEZABLE))
7132 continue;
7133 /*
7134 * nr_active is monotonically decreasing. It's safe
7135 * to peek without lock.
7136 */
7137 rcu_read_lock();
7138 for_each_pwq(pwq, wq) {
7139 WARN_ON_ONCE(pwq->nr_active < 0);
7140 if (pwq->nr_active) {
7141 busy = true;
7142 rcu_read_unlock();
7143 goto out_unlock;
7144 }
7145 }
7146 rcu_read_unlock();
7147 }
7148out_unlock:
7149 mutex_unlock(&wq_pool_mutex);
7150 return busy;
7151}
7152
7153/**
7154 * thaw_workqueues - thaw workqueues
7155 *
7156 * Thaw workqueues. Normal queueing is restored and all collected
7157 * frozen works are transferred to their respective pool worklists.
7158 *
7159 * CONTEXT:
7160 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
7161 */
7162void thaw_workqueues(void)
7163{
7164 struct workqueue_struct *wq;
7165
7166 mutex_lock(&wq_pool_mutex);
7167
7168 if (!workqueue_freezing)
7169 goto out_unlock;
7170
7171 workqueue_freezing = false;
7172
7173 /* restore max_active and repopulate worklist */
7174 list_for_each_entry(wq, &workqueues, list) {
7175 mutex_lock(&wq->mutex);
7176 wq_adjust_max_active(wq);
7177 mutex_unlock(&wq->mutex);
7178 }
7179
7180out_unlock:
7181 mutex_unlock(&wq_pool_mutex);
7182}
7183#endif /* CONFIG_FREEZER */
7184
7185static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
7186{
7187 LIST_HEAD(ctxs);
7188 int ret = 0;
7189 struct workqueue_struct *wq;
7190 struct apply_wqattrs_ctx *ctx, *n;
7191
7192 lockdep_assert_held(&wq_pool_mutex);
7193
7194 list_for_each_entry(wq, &workqueues, list) {
7195 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING))
7196 continue;
7197
7198 ctx = apply_wqattrs_prepare(wq, wq->attrs, unbound_cpumask);
7199 if (IS_ERR(ctx)) {
7200 ret = PTR_ERR(ctx);
7201 break;
7202 }
7203
7204 list_add_tail(&ctx->list, &ctxs);
7205 }
7206
7207 list_for_each_entry_safe(ctx, n, &ctxs, list) {
7208 if (!ret)
7209 apply_wqattrs_commit(ctx);
7210 apply_wqattrs_cleanup(ctx);
7211 }
7212
7213 if (!ret) {
7214 int cpu;
7215 struct worker_pool *pool;
7216 struct worker *worker;
7217
7218 mutex_lock(&wq_pool_attach_mutex);
7219 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
7220 /* rescuer needs to respect cpumask changes when it is not attached */
7221 list_for_each_entry(wq, &workqueues, list) {
7222 if (wq->rescuer && !wq->rescuer->pool)
7223 unbind_worker(wq->rescuer);
7224 }
7225 /* DISASSOCIATED worker needs to respect wq_unbound_cpumask */
7226 for_each_possible_cpu(cpu) {
7227 for_each_cpu_worker_pool(pool, cpu) {
7228 if (!(pool->flags & POOL_DISASSOCIATED))
7229 continue;
7230 for_each_pool_worker(worker, pool)
7231 unbind_worker(worker);
7232 }
7233 }
7234 mutex_unlock(&wq_pool_attach_mutex);
7235 }
7236 return ret;
7237}
7238
7239/**
7240 * workqueue_unbound_housekeeping_update - Propagate housekeeping cpumask update
7241 * @hk: the new housekeeping cpumask
7242 *
7243 * Update the unbound workqueue cpumask on top of the new housekeeping cpumask such
7244 * that the effective unbound affinity is the intersection of the new housekeeping
7245 * with the requested affinity set via nohz_full=/isolcpus= or sysfs.
7246 *
7247 * Return: 0 on success and -errno on failure.
7248 */
7249int workqueue_unbound_housekeeping_update(const struct cpumask *hk)
7250{
7251 cpumask_var_t cpumask;
7252 int ret = 0;
7253
7254 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7255 return -ENOMEM;
7256
7257 mutex_lock(&wq_pool_mutex);
7258
7259 /*
7260 * If the operation fails, it will fall back to
7261 * wq_requested_unbound_cpumask which is initially set to
7262 * HK_TYPE_DOMAIN house keeping mask and rewritten
7263 * by any subsequent write to workqueue/cpumask sysfs file.
7264 */
7265 if (!cpumask_and(cpumask, wq_requested_unbound_cpumask, hk))
7266 cpumask_copy(cpumask, wq_requested_unbound_cpumask);
7267 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
7268 ret = workqueue_apply_unbound_cpumask(cpumask);
7269
7270 /* Save the current isolated cpumask & export it via sysfs */
7271 if (!ret)
7272 cpumask_andnot(wq_isolated_cpumask, cpu_possible_mask, hk);
7273
7274 mutex_unlock(&wq_pool_mutex);
7275 free_cpumask_var(cpumask);
7276 return ret;
7277}
7278
7279static int parse_affn_scope(const char *val)
7280{
7281 return sysfs_match_string(wq_affn_names, val);
7282}
7283
7284static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
7285{
7286 struct workqueue_struct *wq;
7287 int affn, cpu;
7288
7289 affn = parse_affn_scope(val);
7290 if (affn < 0)
7291 return affn;
7292 if (affn == WQ_AFFN_DFL)
7293 return -EINVAL;
7294
7295 cpus_read_lock();
7296 mutex_lock(&wq_pool_mutex);
7297
7298 wq_affn_dfl = affn;
7299
7300 list_for_each_entry(wq, &workqueues, list) {
7301 for_each_online_cpu(cpu)
7302 unbound_wq_update_pwq(wq, cpu);
7303 }
7304
7305 mutex_unlock(&wq_pool_mutex);
7306 cpus_read_unlock();
7307
7308 return 0;
7309}
7310
7311static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
7312{
7313 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
7314}
7315
7316static const struct kernel_param_ops wq_affn_dfl_ops = {
7317 .set = wq_affn_dfl_set,
7318 .get = wq_affn_dfl_get,
7319};
7320
7321module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
7322
7323#ifdef CONFIG_SYSFS
7324/*
7325 * Workqueues with WQ_SYSFS flag set is visible to userland via
7326 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
7327 * following attributes.
7328 *
7329 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
7330 * max_active RW int : maximum number of in-flight work items
7331 *
7332 * Unbound workqueues have the following extra attributes.
7333 *
7334 * nice RW int : nice value of the workers
7335 * cpumask RW mask : bitmask of allowed CPUs for the workers
7336 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
7337 * affinity_strict RW bool : worker CPU affinity is strict
7338 */
7339struct wq_device {
7340 struct workqueue_struct *wq;
7341 struct device dev;
7342};
7343
7344static struct workqueue_struct *dev_to_wq(struct device *dev)
7345{
7346 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7347
7348 return wq_dev->wq;
7349}
7350
7351static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
7352 char *buf)
7353{
7354 struct workqueue_struct *wq = dev_to_wq(dev);
7355
7356 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
7357}
7358static DEVICE_ATTR_RO(per_cpu);
7359
7360static ssize_t max_active_show(struct device *dev,
7361 struct device_attribute *attr, char *buf)
7362{
7363 struct workqueue_struct *wq = dev_to_wq(dev);
7364
7365 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
7366}
7367
7368static ssize_t max_active_store(struct device *dev,
7369 struct device_attribute *attr, const char *buf,
7370 size_t count)
7371{
7372 struct workqueue_struct *wq = dev_to_wq(dev);
7373 int val;
7374
7375 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
7376 return -EINVAL;
7377
7378 workqueue_set_max_active(wq, val);
7379 return count;
7380}
7381static DEVICE_ATTR_RW(max_active);
7382
7383static struct attribute *wq_sysfs_attrs[] = {
7384 &dev_attr_per_cpu.attr,
7385 &dev_attr_max_active.attr,
7386 NULL,
7387};
7388
7389static umode_t wq_sysfs_is_visible(struct kobject *kobj, struct attribute *a, int n)
7390{
7391 struct device *dev = kobj_to_dev(kobj);
7392 struct workqueue_struct *wq = dev_to_wq(dev);
7393
7394 /*
7395 * Adjusting max_active breaks ordering guarantee. Changing it has no
7396 * effect on BH worker. Limit max_active to RO in such case.
7397 */
7398 if (wq->flags & (WQ_BH | __WQ_ORDERED))
7399 return 0444;
7400 return a->mode;
7401}
7402
7403static const struct attribute_group wq_sysfs_group = {
7404 .is_visible = wq_sysfs_is_visible,
7405 .attrs = wq_sysfs_attrs,
7406};
7407__ATTRIBUTE_GROUPS(wq_sysfs);
7408
7409static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
7410 char *buf)
7411{
7412 struct workqueue_struct *wq = dev_to_wq(dev);
7413 int written;
7414
7415 mutex_lock(&wq->mutex);
7416 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->attrs->nice);
7417 mutex_unlock(&wq->mutex);
7418
7419 return written;
7420}
7421
7422/* prepare workqueue_attrs for sysfs store operations */
7423static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
7424{
7425 struct workqueue_attrs *attrs;
7426
7427 lockdep_assert_held(&wq_pool_mutex);
7428
7429 attrs = alloc_workqueue_attrs();
7430 if (!attrs)
7431 return NULL;
7432
7433 copy_workqueue_attrs(attrs, wq->attrs);
7434 return attrs;
7435}
7436
7437static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
7438 const char *buf, size_t count)
7439{
7440 struct workqueue_struct *wq = dev_to_wq(dev);
7441 struct workqueue_attrs *attrs;
7442 int ret = -ENOMEM;
7443
7444 mutex_lock(&wq_pool_mutex);
7445
7446 attrs = wq_sysfs_prep_attrs(wq);
7447 if (!attrs)
7448 goto out_unlock;
7449
7450 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
7451 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
7452 ret = apply_workqueue_attrs_locked(wq, attrs);
7453 else
7454 ret = -EINVAL;
7455
7456out_unlock:
7457 mutex_unlock(&wq_pool_mutex);
7458 free_workqueue_attrs(attrs);
7459 return ret ?: count;
7460}
7461
7462static ssize_t wq_cpumask_show(struct device *dev,
7463 struct device_attribute *attr, char *buf)
7464{
7465 struct workqueue_struct *wq = dev_to_wq(dev);
7466 int written;
7467
7468 mutex_lock(&wq->mutex);
7469 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
7470 cpumask_pr_args(wq->attrs->cpumask));
7471 mutex_unlock(&wq->mutex);
7472 return written;
7473}
7474
7475static ssize_t wq_cpumask_store(struct device *dev,
7476 struct device_attribute *attr,
7477 const char *buf, size_t count)
7478{
7479 struct workqueue_struct *wq = dev_to_wq(dev);
7480 struct workqueue_attrs *attrs;
7481 int ret = -ENOMEM;
7482
7483 mutex_lock(&wq_pool_mutex);
7484
7485 attrs = wq_sysfs_prep_attrs(wq);
7486 if (!attrs)
7487 goto out_unlock;
7488
7489 ret = cpumask_parse(buf, attrs->cpumask);
7490 if (!ret)
7491 ret = apply_workqueue_attrs_locked(wq, attrs);
7492
7493out_unlock:
7494 mutex_unlock(&wq_pool_mutex);
7495 free_workqueue_attrs(attrs);
7496 return ret ?: count;
7497}
7498
7499static ssize_t wq_affn_scope_show(struct device *dev,
7500 struct device_attribute *attr, char *buf)
7501{
7502 struct workqueue_struct *wq = dev_to_wq(dev);
7503 int written;
7504
7505 mutex_lock(&wq->mutex);
7506 if (wq->attrs->affn_scope == WQ_AFFN_DFL)
7507 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
7508 wq_affn_names[WQ_AFFN_DFL],
7509 wq_affn_names[wq_affn_dfl]);
7510 else
7511 written = scnprintf(buf, PAGE_SIZE, "%s\n",
7512 wq_affn_names[wq->attrs->affn_scope]);
7513 mutex_unlock(&wq->mutex);
7514
7515 return written;
7516}
7517
7518static ssize_t wq_affn_scope_store(struct device *dev,
7519 struct device_attribute *attr,
7520 const char *buf, size_t count)
7521{
7522 struct workqueue_struct *wq = dev_to_wq(dev);
7523 struct workqueue_attrs *attrs;
7524 int affn, ret = -ENOMEM;
7525
7526 affn = parse_affn_scope(buf);
7527 if (affn < 0)
7528 return affn;
7529
7530 mutex_lock(&wq_pool_mutex);
7531 attrs = wq_sysfs_prep_attrs(wq);
7532 if (attrs) {
7533 attrs->affn_scope = affn;
7534 ret = apply_workqueue_attrs_locked(wq, attrs);
7535 }
7536 mutex_unlock(&wq_pool_mutex);
7537 free_workqueue_attrs(attrs);
7538 return ret ?: count;
7539}
7540
7541static ssize_t wq_affinity_strict_show(struct device *dev,
7542 struct device_attribute *attr, char *buf)
7543{
7544 struct workqueue_struct *wq = dev_to_wq(dev);
7545
7546 return scnprintf(buf, PAGE_SIZE, "%d\n",
7547 wq->attrs->affn_strict);
7548}
7549
7550static ssize_t wq_affinity_strict_store(struct device *dev,
7551 struct device_attribute *attr,
7552 const char *buf, size_t count)
7553{
7554 struct workqueue_struct *wq = dev_to_wq(dev);
7555 struct workqueue_attrs *attrs;
7556 int v, ret = -ENOMEM;
7557
7558 if (sscanf(buf, "%d", &v) != 1)
7559 return -EINVAL;
7560
7561 mutex_lock(&wq_pool_mutex);
7562 attrs = wq_sysfs_prep_attrs(wq);
7563 if (attrs) {
7564 attrs->affn_strict = (bool)v;
7565 ret = apply_workqueue_attrs_locked(wq, attrs);
7566 }
7567 mutex_unlock(&wq_pool_mutex);
7568 free_workqueue_attrs(attrs);
7569 return ret ?: count;
7570}
7571
7572static struct device_attribute wq_sysfs_unbound_attrs[] = {
7573 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7574 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7575 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
7576 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
7577 __ATTR_NULL,
7578};
7579
7580static const struct bus_type wq_subsys = {
7581 .name = "workqueue",
7582 .dev_groups = wq_sysfs_groups,
7583};
7584
7585/**
7586 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
7587 * @cpumask: the cpumask to set
7588 *
7589 * The low-level workqueues cpumask is a global cpumask that limits
7590 * the affinity of all unbound workqueues. This function check the @cpumask
7591 * and apply it to all unbound workqueues and updates all pwqs of them.
7592 *
7593 * Return: 0 - Success
7594 * -EINVAL - Invalid @cpumask
7595 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
7596 */
7597static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
7598{
7599 int ret = -EINVAL;
7600
7601 /*
7602 * Not excluding isolated cpus on purpose.
7603 * If the user wishes to include them, we allow that.
7604 */
7605 cpumask_and(cpumask, cpumask, cpu_possible_mask);
7606 if (!cpumask_empty(cpumask)) {
7607 ret = 0;
7608 mutex_lock(&wq_pool_mutex);
7609 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
7610 ret = workqueue_apply_unbound_cpumask(cpumask);
7611 if (!ret)
7612 cpumask_copy(wq_requested_unbound_cpumask, cpumask);
7613 mutex_unlock(&wq_pool_mutex);
7614 }
7615
7616 return ret;
7617}
7618
7619static ssize_t __wq_cpumask_show(struct device *dev,
7620 struct device_attribute *attr, char *buf, cpumask_var_t mask)
7621{
7622 int written;
7623
7624 mutex_lock(&wq_pool_mutex);
7625 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
7626 mutex_unlock(&wq_pool_mutex);
7627
7628 return written;
7629}
7630
7631static ssize_t cpumask_requested_show(struct device *dev,
7632 struct device_attribute *attr, char *buf)
7633{
7634 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
7635}
7636static DEVICE_ATTR_RO(cpumask_requested);
7637
7638static ssize_t cpumask_isolated_show(struct device *dev,
7639 struct device_attribute *attr, char *buf)
7640{
7641 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
7642}
7643static DEVICE_ATTR_RO(cpumask_isolated);
7644
7645static ssize_t cpumask_show(struct device *dev,
7646 struct device_attribute *attr, char *buf)
7647{
7648 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
7649}
7650
7651static ssize_t cpumask_store(struct device *dev,
7652 struct device_attribute *attr, const char *buf, size_t count)
7653{
7654 cpumask_var_t cpumask;
7655 int ret;
7656
7657 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7658 return -ENOMEM;
7659
7660 ret = cpumask_parse(buf, cpumask);
7661 if (!ret)
7662 ret = workqueue_set_unbound_cpumask(cpumask);
7663
7664 free_cpumask_var(cpumask);
7665 return ret ? ret : count;
7666}
7667static DEVICE_ATTR_RW(cpumask);
7668
7669static struct attribute *wq_sysfs_cpumask_attrs[] = {
7670 &dev_attr_cpumask.attr,
7671 &dev_attr_cpumask_requested.attr,
7672 &dev_attr_cpumask_isolated.attr,
7673 NULL,
7674};
7675ATTRIBUTE_GROUPS(wq_sysfs_cpumask);
7676
7677static int __init wq_sysfs_init(void)
7678{
7679 return subsys_virtual_register(&wq_subsys, wq_sysfs_cpumask_groups);
7680}
7681core_initcall(wq_sysfs_init);
7682
7683static void wq_device_release(struct device *dev)
7684{
7685 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7686
7687 kfree(wq_dev);
7688}
7689
7690/**
7691 * workqueue_sysfs_register - make a workqueue visible in sysfs
7692 * @wq: the workqueue to register
7693 *
7694 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
7695 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7696 * which is the preferred method.
7697 *
7698 * Workqueue user should use this function directly iff it wants to apply
7699 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7700 * apply_workqueue_attrs() may race against userland updating the
7701 * attributes.
7702 *
7703 * Return: 0 on success, -errno on failure.
7704 */
7705int workqueue_sysfs_register(struct workqueue_struct *wq)
7706{
7707 struct wq_device *wq_dev;
7708 int ret;
7709
7710 wq->wq_dev = wq_dev = kzalloc_obj(*wq_dev);
7711 if (!wq_dev)
7712 return -ENOMEM;
7713
7714 wq_dev->wq = wq;
7715 wq_dev->dev.bus = &wq_subsys;
7716 wq_dev->dev.release = wq_device_release;
7717 dev_set_name(&wq_dev->dev, "%s", wq->name);
7718
7719 /*
7720 * attrs are created separately. Suppress uevent until
7721 * everything is ready.
7722 */
7723 dev_set_uevent_suppress(&wq_dev->dev, true);
7724
7725 ret = device_register(&wq_dev->dev);
7726 if (ret) {
7727 put_device(&wq_dev->dev);
7728 wq->wq_dev = NULL;
7729 return ret;
7730 }
7731
7732 if (wq->flags & WQ_UNBOUND) {
7733 struct device_attribute *attr;
7734
7735 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7736 ret = device_create_file(&wq_dev->dev, attr);
7737 if (ret) {
7738 device_unregister(&wq_dev->dev);
7739 wq->wq_dev = NULL;
7740 return ret;
7741 }
7742 }
7743 }
7744
7745 dev_set_uevent_suppress(&wq_dev->dev, false);
7746 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7747 return 0;
7748}
7749
7750/**
7751 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7752 * @wq: the workqueue to unregister
7753 *
7754 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7755 */
7756static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7757{
7758 struct wq_device *wq_dev = wq->wq_dev;
7759
7760 if (!wq->wq_dev)
7761 return;
7762
7763 wq->wq_dev = NULL;
7764 device_unregister(&wq_dev->dev);
7765}
7766#else /* CONFIG_SYSFS */
7767static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7768#endif /* CONFIG_SYSFS */
7769
7770/*
7771 * Workqueue watchdog.
7772 *
7773 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7774 * flush dependency, a concurrency managed work item which stays RUNNING
7775 * indefinitely. Workqueue stalls can be very difficult to debug as the
7776 * usual warning mechanisms don't trigger and internal workqueue state is
7777 * largely opaque.
7778 *
7779 * Workqueue watchdog monitors all worker pools periodically and dumps
7780 * state if some pools failed to make forward progress for a while where
7781 * forward progress is defined as the first item on ->worklist changing.
7782 *
7783 * This mechanism is controlled through the kernel parameter
7784 * "workqueue.watchdog_thresh" which can be updated at runtime through the
7785 * corresponding sysfs parameter file.
7786 */
7787#ifdef CONFIG_WQ_WATCHDOG
7788
7789static unsigned long wq_watchdog_thresh = 30;
7790static struct timer_list wq_watchdog_timer;
7791
7792static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7793static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7794
7795static unsigned int wq_panic_on_stall = CONFIG_BOOTPARAM_WQ_STALL_PANIC;
7796module_param_named(panic_on_stall, wq_panic_on_stall, uint, 0644);
7797
7798static unsigned int wq_panic_on_stall_time;
7799module_param_named(panic_on_stall_time, wq_panic_on_stall_time, uint, 0644);
7800MODULE_PARM_DESC(panic_on_stall_time, "Panic if stall exceeds this many seconds (0=disabled)");
7801
7802/*
7803 * Report that a pool has no worker in running state, which is a sign that the
7804 * pool may be stuck. Print pool info. Must be called with pool->lock held and
7805 * inside a printk_deferred_enter/exit region.
7806 */
7807static void show_pool_no_running_worker(struct worker_pool *pool)
7808{
7809 lockdep_assert_held(&pool->lock);
7810
7811 printk_deferred_enter();
7812 pr_info("pool %d: no worker in running state, cpu=%d is %s (nr_workers=%d nr_idle=%d)\n",
7813 pool->id, pool->cpu,
7814 idle_cpu(pool->cpu) ? "idle" : "busy",
7815 pool->nr_workers, pool->nr_idle);
7816 pr_info("The pool might have trouble waking an idle worker.\n");
7817 /*
7818 * last_woken_worker and its task are valid here: set_worker_dying()
7819 * clears it under pool->lock before setting WORKER_DIE, so if
7820 * last_woken_worker is non-NULL the kthread has not yet exited and
7821 * worker->task is still alive.
7822 */
7823 if (pool->last_woken_worker) {
7824 pr_info("Backtrace of last woken worker:\n");
7825 sched_show_task(pool->last_woken_worker->task);
7826 } else {
7827 pr_info("Last woken worker empty\n");
7828 }
7829 printk_deferred_exit();
7830}
7831
7832/*
7833 * Show running workers that might prevent the processing of pending work items.
7834 * If no running worker is found, the pool may be stuck waiting for an idle
7835 * worker to be woken, so report the pool state and the last woken worker.
7836 */
7837static void show_cpu_pool_busy_workers(struct worker_pool *pool)
7838{
7839 bool found_running = false;
7840 struct worker *worker;
7841 unsigned long irq_flags;
7842 int cpu, bkt;
7843
7844 raw_spin_lock_irqsave(&pool->lock, irq_flags);
7845
7846 /* Snapshot cpu inside the lock to safely use it after unlock. */
7847 cpu = pool->cpu;
7848
7849 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
7850 /* Skip workers that are not actively running on the CPU. */
7851 if (!task_is_running(worker->task))
7852 continue;
7853
7854 found_running = true;
7855 /*
7856 * Defer printing to avoid deadlocks in console
7857 * drivers that queue work while holding locks
7858 * also taken in their write paths.
7859 */
7860 printk_deferred_enter();
7861
7862 pr_info("pool %d:\n", pool->id);
7863 sched_show_task(worker->task);
7864
7865 printk_deferred_exit();
7866 }
7867
7868 /*
7869 * If no running worker was found, the pool is likely stuck. Print pool
7870 * state and the backtrace of the last woken worker, which is the prime
7871 * suspect for the stall.
7872 */
7873 if (!found_running)
7874 show_pool_no_running_worker(pool);
7875
7876 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
7877
7878 /*
7879 * Trigger a backtrace on the stalled CPU to capture what it is
7880 * currently executing. Skip an offline CPU, whose NMI is never acked
7881 * and would make the backtrace busy-wait until it times out. Done
7882 * after releasing the lock to avoid issues with NMI delivery.
7883 */
7884 if (!found_running && cpu_online(cpu))
7885 trigger_single_cpu_backtrace(cpu);
7886}
7887
7888static void show_cpu_pools_busy_workers(void)
7889{
7890 struct worker_pool *pool;
7891 int pi;
7892
7893 pr_info("Showing backtraces of busy workers in stalled worker pools:\n");
7894
7895 rcu_read_lock();
7896
7897 for_each_pool(pool, pi) {
7898 if (pool->cpu_stall)
7899 show_cpu_pool_busy_workers(pool);
7900
7901 }
7902
7903 rcu_read_unlock();
7904}
7905
7906/*
7907 * It triggers a panic in two scenarios: when the total number of stalls
7908 * exceeds a threshold, and when a stall lasts longer than
7909 * wq_panic_on_stall_time
7910 */
7911static void panic_on_wq_watchdog(unsigned int stall_time_sec)
7912{
7913 static unsigned int wq_stall;
7914
7915 if (wq_panic_on_stall) {
7916 wq_stall++;
7917 if (wq_stall >= wq_panic_on_stall)
7918 panic("workqueue: %u stall(s) exceeded threshold %u\n",
7919 wq_stall, wq_panic_on_stall);
7920 }
7921
7922 if (wq_panic_on_stall_time && stall_time_sec >= wq_panic_on_stall_time)
7923 panic("workqueue: stall lasted %us, exceeding threshold %us\n",
7924 stall_time_sec, wq_panic_on_stall_time);
7925}
7926
7927static void wq_watchdog_reset_touched(void)
7928{
7929 int cpu;
7930
7931 wq_watchdog_touched = jiffies;
7932 for_each_possible_cpu(cpu)
7933 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7934}
7935
7936static void wq_watchdog_timer_fn(struct timer_list *unused)
7937{
7938 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7939 unsigned int max_stall_time = 0;
7940 bool lockup_detected = false;
7941 bool cpu_pool_stall = false;
7942 unsigned long now = jiffies;
7943 struct worker_pool *pool;
7944 unsigned int stall_time;
7945 int pi;
7946
7947 if (!thresh)
7948 return;
7949
7950 for_each_pool(pool, pi) {
7951 unsigned long pool_ts, touched, ts;
7952
7953 pool->cpu_stall = false;
7954 if (list_empty(&pool->worklist))
7955 continue;
7956
7957 /*
7958 * If a virtual machine is stopped by the host it can look to
7959 * the watchdog like a stall.
7960 */
7961 kvm_check_and_clear_guest_paused();
7962
7963 /* get the latest of pool and touched timestamps */
7964 if (pool->cpu >= 0)
7965 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7966 else
7967 touched = READ_ONCE(wq_watchdog_touched);
7968 pool_ts = READ_ONCE(pool->last_progress_ts);
7969
7970 if (time_after(pool_ts, touched))
7971 ts = pool_ts;
7972 else
7973 ts = touched;
7974
7975 /*
7976 * Did we stall?
7977 *
7978 * Do a lockless check first to do not disturb the system.
7979 *
7980 * Prevent false positives by double checking the timestamp
7981 * under pool->lock. The lock makes sure that the check reads
7982 * an updated pool->last_progress_ts when this CPU saw
7983 * an already updated pool->worklist above. It seems better
7984 * than adding another barrier into __queue_work() which
7985 * is a hotter path.
7986 */
7987 if (time_after(now, ts + thresh)) {
7988 scoped_guard(raw_spinlock_irqsave, &pool->lock) {
7989 pool_ts = pool->last_progress_ts;
7990 if (time_after(pool_ts, touched))
7991 ts = pool_ts;
7992 else
7993 ts = touched;
7994 }
7995 if (!time_after(now, ts + thresh))
7996 continue;
7997
7998 lockup_detected = true;
7999 stall_time = jiffies_to_msecs(now - pool_ts) / 1000;
8000 max_stall_time = max(max_stall_time, stall_time);
8001 if (is_percpu_pool(pool) && !(pool->flags & POOL_BH)) {
8002 pool->cpu_stall = true;
8003 cpu_pool_stall = true;
8004 }
8005 pr_emerg("BUG: workqueue lockup - pool");
8006 pr_cont_pool_info(pool);
8007 pr_cont(" stuck for %us!\n", stall_time);
8008 }
8009 }
8010
8011 if (lockup_detected)
8012 show_all_workqueues();
8013
8014 if (cpu_pool_stall)
8015 show_cpu_pools_busy_workers();
8016
8017 if (lockup_detected)
8018 panic_on_wq_watchdog(max_stall_time);
8019
8020 wq_watchdog_reset_touched();
8021 mod_timer(&wq_watchdog_timer, jiffies + thresh);
8022}
8023
8024notrace void wq_watchdog_touch(int cpu)
8025{
8026 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
8027 unsigned long touch_ts = READ_ONCE(wq_watchdog_touched);
8028 unsigned long now = jiffies;
8029
8030 if (cpu >= 0)
8031 per_cpu(wq_watchdog_touched_cpu, cpu) = now;
8032 else
8033 WARN_ONCE(1, "%s should be called with valid CPU", __func__);
8034
8035 /* Don't unnecessarily store to global cacheline */
8036 if (time_after(now, touch_ts + thresh / 4))
8037 WRITE_ONCE(wq_watchdog_touched, jiffies);
8038}
8039
8040static void wq_watchdog_set_thresh(unsigned long thresh)
8041{
8042 wq_watchdog_thresh = 0;
8043 timer_delete_sync(&wq_watchdog_timer);
8044
8045 if (thresh) {
8046 wq_watchdog_thresh = thresh;
8047 wq_watchdog_reset_touched();
8048 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
8049 }
8050}
8051
8052static int wq_watchdog_param_set_thresh(const char *val,
8053 const struct kernel_param *kp)
8054{
8055 unsigned long thresh;
8056 int ret;
8057
8058 ret = kstrtoul(val, 0, &thresh);
8059 if (ret)
8060 return ret;
8061
8062 if (thresh > MAX_JIFFY_OFFSET / HZ)
8063 return -ERANGE;
8064
8065 if (system_percpu_wq)
8066 wq_watchdog_set_thresh(thresh);
8067 else
8068 wq_watchdog_thresh = thresh;
8069
8070 return 0;
8071}
8072
8073static const struct kernel_param_ops wq_watchdog_thresh_ops = {
8074 .set = wq_watchdog_param_set_thresh,
8075 .get = param_get_ulong,
8076};
8077
8078module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
8079 0644);
8080
8081static void wq_watchdog_init(void)
8082{
8083 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
8084 wq_watchdog_set_thresh(wq_watchdog_thresh);
8085}
8086
8087#else /* CONFIG_WQ_WATCHDOG */
8088
8089static inline void wq_watchdog_init(void) { }
8090
8091#endif /* CONFIG_WQ_WATCHDOG */
8092
8093static void bh_pool_kick_normal(struct irq_work *irq_work)
8094{
8095 raise_softirq(TASKLET_SOFTIRQ);
8096}
8097
8098static void bh_pool_kick_highpri(struct irq_work *irq_work)
8099{
8100 raise_softirq(HI_SOFTIRQ);
8101}
8102
8103static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
8104{
8105 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
8106 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
8107 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
8108 return;
8109 }
8110
8111 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
8112}
8113
8114static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice)
8115{
8116 BUG_ON(init_worker_pool(pool));
8117 pool->cpu = cpu;
8118 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
8119 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
8120 pool->attrs->nice = nice;
8121 pool->attrs->affn_strict = true;
8122 pool->node = cpu_to_node(cpu);
8123
8124 /* alloc pool ID */
8125 mutex_lock(&wq_pool_mutex);
8126 BUG_ON(worker_pool_assign_id(pool));
8127 mutex_unlock(&wq_pool_mutex);
8128}
8129
8130/**
8131 * workqueue_init_early - early init for workqueue subsystem
8132 *
8133 * This is the first step of three-staged workqueue subsystem initialization and
8134 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
8135 * up. It sets up all the data structures and system workqueues and allows early
8136 * boot code to create workqueues and queue/cancel work items. Actual work item
8137 * execution starts only after kthreads can be created and scheduled right
8138 * before early initcalls.
8139 */
8140void __init workqueue_init_early(void)
8141{
8142 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
8143 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
8144 void (*irq_work_fns[NR_STD_WORKER_POOLS])(struct irq_work *) =
8145 { bh_pool_kick_normal, bh_pool_kick_highpri };
8146 int i, cpu;
8147
8148 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
8149
8150 BUG_ON(!alloc_cpumask_var(&wq_online_cpumask, GFP_KERNEL));
8151 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
8152 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
8153 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
8154
8155 cpumask_copy(wq_online_cpumask, cpu_online_mask);
8156 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
8157 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
8158 if (!cpumask_empty(&wq_cmdline_cpumask))
8159 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
8160
8161 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
8162 cpumask_andnot(wq_isolated_cpumask, cpu_possible_mask,
8163 housekeeping_cpumask(HK_TYPE_DOMAIN));
8164 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
8165
8166 unbound_wq_update_pwq_attrs_buf = alloc_workqueue_attrs();
8167 BUG_ON(!unbound_wq_update_pwq_attrs_buf);
8168
8169 /*
8170 * If nohz_full is enabled, set power efficient workqueue as unbound.
8171 * This allows workqueue items to be moved to HK CPUs.
8172 */
8173 if (housekeeping_enabled(HK_TYPE_TICK))
8174 wq_power_efficient = true;
8175
8176 /* initialize WQ_AFFN_SYSTEM pods */
8177 pt->pod_cpus = kzalloc_objs(pt->pod_cpus[0], 1);
8178 pt->pod_node = kzalloc_objs(pt->pod_node[0], 1);
8179 pt->cpu_pod = kzalloc_objs(pt->cpu_pod[0], nr_cpu_ids);
8180 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
8181
8182 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
8183
8184 pt->nr_pods = 1;
8185 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
8186 pt->pod_node[0] = NUMA_NO_NODE;
8187 pt->cpu_pod[0] = 0;
8188
8189 /* initialize BH and CPU pools */
8190 for_each_possible_cpu(cpu) {
8191 struct worker_pool *pool;
8192
8193 i = 0;
8194 for_each_bh_worker_pool(pool, cpu) {
8195 init_cpu_worker_pool(pool, cpu, std_nice[i]);
8196 pool->flags |= POOL_BH;
8197 init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]);
8198 i++;
8199 }
8200
8201 i = 0;
8202 for_each_cpu_worker_pool(pool, cpu)
8203 init_cpu_worker_pool(pool, cpu, std_nice[i++]);
8204 }
8205
8206 /* create default unbound and ordered wq attrs */
8207 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
8208 struct workqueue_attrs *attrs;
8209
8210 BUG_ON(!(attrs = alloc_workqueue_attrs()));
8211 attrs->nice = std_nice[i];
8212 unbound_std_wq_attrs[i] = attrs;
8213
8214 /*
8215 * An ordered wq should have only one pwq as ordering is
8216 * guaranteed by max_active which is enforced by pwqs.
8217 */
8218 BUG_ON(!(attrs = alloc_workqueue_attrs()));
8219 attrs->nice = std_nice[i];
8220 attrs->ordered = true;
8221 ordered_wq_attrs[i] = attrs;
8222 }
8223
8224 system_wq = alloc_workqueue("events", WQ_PERCPU | __WQ_DEPRECATED, 0);
8225 system_percpu_wq = alloc_workqueue("events", WQ_PERCPU, 0);
8226 system_highpri_wq = alloc_workqueue("events_highpri",
8227 WQ_HIGHPRI | WQ_PERCPU, 0);
8228 system_long_wq = alloc_workqueue("events_long", WQ_PERCPU, 0);
8229 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND | __WQ_DEPRECATED, WQ_MAX_ACTIVE);
8230 system_dfl_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, WQ_MAX_ACTIVE);
8231 system_freezable_wq = alloc_workqueue("events_freezable",
8232 WQ_FREEZABLE | WQ_PERCPU, 0);
8233 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
8234 WQ_POWER_EFFICIENT | WQ_PERCPU, 0);
8235 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
8236 WQ_FREEZABLE | WQ_POWER_EFFICIENT | WQ_PERCPU, 0);
8237 system_bh_wq = alloc_workqueue("events_bh", WQ_BH | WQ_PERCPU, 0);
8238 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri",
8239 WQ_BH | WQ_HIGHPRI | WQ_PERCPU, 0);
8240 system_dfl_long_wq = alloc_workqueue("events_dfl_long", WQ_UNBOUND, WQ_MAX_ACTIVE);
8241 BUG_ON(!system_wq || !system_percpu_wq|| !system_highpri_wq || !system_long_wq ||
8242 !system_unbound_wq || !system_freezable_wq || !system_dfl_wq ||
8243 !system_power_efficient_wq ||
8244 !system_freezable_power_efficient_wq ||
8245 !system_bh_wq || !system_bh_highpri_wq || !system_dfl_long_wq);
8246}
8247
8248static void __init wq_cpu_intensive_thresh_init(void)
8249{
8250 unsigned long thresh;
8251 unsigned long bogo;
8252
8253 pwq_release_worker = kthread_run_worker(0, "pool_workqueue_release");
8254 BUG_ON(IS_ERR(pwq_release_worker));
8255
8256 /* if the user set it to a specific value, keep it */
8257 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
8258 return;
8259
8260 /*
8261 * The default of 10ms is derived from the fact that most modern (as of
8262 * 2023) processors can do a lot in 10ms and that it's just below what
8263 * most consider human-perceivable. However, the kernel also runs on a
8264 * lot slower CPUs including microcontrollers where the threshold is way
8265 * too low.
8266 *
8267 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
8268 * This is by no means accurate but it doesn't have to be. The mechanism
8269 * is still useful even when the threshold is fully scaled up. Also, as
8270 * the reports would usually be applicable to everyone, some machines
8271 * operating on longer thresholds won't significantly diminish their
8272 * usefulness.
8273 */
8274 thresh = 10 * USEC_PER_MSEC;
8275
8276 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
8277 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
8278 if (bogo < 4000)
8279 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
8280
8281 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
8282 loops_per_jiffy, bogo, thresh);
8283
8284 wq_cpu_intensive_thresh_us = thresh;
8285}
8286
8287/**
8288 * workqueue_init - bring workqueue subsystem fully online
8289 *
8290 * This is the second step of three-staged workqueue subsystem initialization
8291 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
8292 * been created and work items queued on them, but there are no kworkers
8293 * executing the work items yet. Populate the worker pools with the initial
8294 * workers and enable future kworker creations.
8295 */
8296void __init workqueue_init(void)
8297{
8298 struct workqueue_struct *wq;
8299 struct worker_pool *pool;
8300 int cpu, bkt;
8301
8302 wq_cpu_intensive_thresh_init();
8303
8304 mutex_lock(&wq_pool_mutex);
8305
8306 /*
8307 * Per-cpu pools created earlier could be missing node hint. Fix them
8308 * up. Also, create a rescuer for workqueues that requested it.
8309 */
8310 for_each_possible_cpu(cpu) {
8311 for_each_bh_worker_pool(pool, cpu)
8312 pool->node = cpu_to_node(cpu);
8313 for_each_cpu_worker_pool(pool, cpu)
8314 pool->node = cpu_to_node(cpu);
8315 }
8316
8317 list_for_each_entry(wq, &workqueues, list) {
8318 WARN(init_rescuer(wq),
8319 "workqueue: failed to create early rescuer for %s",
8320 wq->name);
8321 }
8322
8323 mutex_unlock(&wq_pool_mutex);
8324
8325 /*
8326 * Create the initial workers. A BH pool has one pseudo worker that
8327 * represents the shared BH execution context and thus doesn't get
8328 * affected by hotplug events. Create the BH pseudo workers for all
8329 * possible CPUs here.
8330 */
8331 for_each_possible_cpu(cpu)
8332 for_each_bh_worker_pool(pool, cpu)
8333 BUG_ON(!create_worker(pool));
8334
8335 for_each_online_cpu(cpu) {
8336 for_each_cpu_worker_pool(pool, cpu) {
8337 pool->flags &= ~POOL_DISASSOCIATED;
8338 BUG_ON(!create_worker(pool));
8339 }
8340 }
8341
8342 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
8343 BUG_ON(!create_worker(pool));
8344
8345 wq_online = true;
8346 wq_watchdog_init();
8347}
8348
8349/*
8350 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
8351 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
8352 * and consecutive pod ID. The rest of @pt is initialized accordingly.
8353 */
8354static void __init init_pod_type(struct wq_pod_type *pt,
8355 bool (*cpus_share_pod)(int, int))
8356{
8357 int cur, pre, cpu, pod;
8358
8359 pt->nr_pods = 0;
8360
8361 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
8362 pt->cpu_pod = kzalloc_objs(pt->cpu_pod[0], nr_cpu_ids);
8363 BUG_ON(!pt->cpu_pod);
8364
8365 for_each_possible_cpu(cur) {
8366 for_each_possible_cpu(pre) {
8367 if (pre >= cur) {
8368 pt->cpu_pod[cur] = pt->nr_pods++;
8369 break;
8370 }
8371 if (cpus_share_pod(cur, pre)) {
8372 pt->cpu_pod[cur] = pt->cpu_pod[pre];
8373 break;
8374 }
8375 }
8376 }
8377
8378 /* init the rest to match @pt->cpu_pod[] */
8379 pt->pod_cpus = kzalloc_objs(pt->pod_cpus[0], pt->nr_pods);
8380 pt->pod_node = kzalloc_objs(pt->pod_node[0], pt->nr_pods);
8381 BUG_ON(!pt->pod_cpus || !pt->pod_node);
8382
8383 for (pod = 0; pod < pt->nr_pods; pod++)
8384 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
8385
8386 for_each_possible_cpu(cpu) {
8387 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
8388 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
8389 }
8390}
8391
8392static bool __init cpus_dont_share(int cpu0, int cpu1)
8393{
8394 return false;
8395}
8396
8397static bool __init cpus_share_smt(int cpu0, int cpu1)
8398{
8399 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
8400}
8401
8402static bool __init cpus_share_numa(int cpu0, int cpu1)
8403{
8404 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
8405}
8406
8407/* Maps each CPU to its shard index within the LLC pod it belongs to */
8408static int cpu_shard_id[NR_CPUS] __initdata;
8409
8410/**
8411 * llc_count_cores - count distinct cores (SMT groups) within an LLC pod
8412 * @pod_cpus: the cpumask of CPUs in the LLC pod
8413 * @smt_pods: the SMT pod type, used to identify sibling groups
8414 *
8415 * A core is represented by the lowest-numbered CPU in its SMT group. Returns
8416 * the number of distinct cores found in @pod_cpus.
8417 */
8418static int __init llc_count_cores(const struct cpumask *pod_cpus,
8419 struct wq_pod_type *smt_pods)
8420{
8421 const struct cpumask *sibling_cpus;
8422 int nr_cores = 0, c;
8423
8424 /*
8425 * Count distinct cores by only counting the first CPU in each
8426 * SMT sibling group.
8427 */
8428 for_each_cpu(c, pod_cpus) {
8429 sibling_cpus = smt_pods->pod_cpus[smt_pods->cpu_pod[c]];
8430 if (cpumask_first(sibling_cpus) == c)
8431 nr_cores++;
8432 }
8433
8434 return nr_cores;
8435}
8436
8437/*
8438 * llc_shard_size - number of cores in a given shard
8439 *
8440 * Cores are spread as evenly as possible. The first @nr_large_shards shards are
8441 * "large shards" with (cores_per_shard + 1) cores; the rest are "default
8442 * shards" with cores_per_shard cores.
8443 */
8444static int __init llc_shard_size(int shard_id, int cores_per_shard, int nr_large_shards)
8445{
8446 /* The first @nr_large_shards shards are large shards */
8447 if (shard_id < nr_large_shards)
8448 return cores_per_shard + 1;
8449
8450 /* The remaining shards are default shards */
8451 return cores_per_shard;
8452}
8453
8454/*
8455 * llc_calc_shard_layout - compute the shard layout for an LLC pod
8456 * @nr_cores: number of distinct cores in the LLC pod
8457 *
8458 * Chooses the number of shards that keeps average shard size closest to
8459 * wq_cache_shard_size. Returns a struct describing the total number of shards,
8460 * the base size of each, and how many are large shards.
8461 */
8462static struct llc_shard_layout __init llc_calc_shard_layout(int nr_cores)
8463{
8464 struct llc_shard_layout layout;
8465
8466 /* Ensure at least one shard; pick the count closest to the target size */
8467 layout.nr_shards = max(1, DIV_ROUND_CLOSEST(nr_cores, wq_cache_shard_size));
8468 layout.cores_per_shard = nr_cores / layout.nr_shards;
8469 layout.nr_large_shards = nr_cores % layout.nr_shards;
8470
8471 return layout;
8472}
8473
8474/*
8475 * llc_shard_is_full - check whether a shard has reached its core capacity
8476 * @cores_in_shard: number of cores already assigned to this shard
8477 * @shard_id: index of the shard being checked
8478 * @layout: the shard layout computed by llc_calc_shard_layout()
8479 *
8480 * Returns true if @cores_in_shard equals the expected size for @shard_id.
8481 */
8482static bool __init llc_shard_is_full(int cores_in_shard, int shard_id,
8483 const struct llc_shard_layout *layout)
8484{
8485 return cores_in_shard == llc_shard_size(shard_id, layout->cores_per_shard,
8486 layout->nr_large_shards);
8487}
8488
8489/**
8490 * llc_populate_cpu_shard_id - populate cpu_shard_id[] for each CPU in an LLC pod
8491 * @pod_cpus: the cpumask of CPUs in the LLC pod
8492 * @smt_pods: the SMT pod type, used to identify sibling groups
8493 * @nr_cores: number of distinct cores in @pod_cpus (from llc_count_cores())
8494 *
8495 * Walks @pod_cpus in order. At each SMT group leader, advances to the next
8496 * shard once the current shard is full. Results are written to cpu_shard_id[].
8497 */
8498static void __init llc_populate_cpu_shard_id(const struct cpumask *pod_cpus,
8499 struct wq_pod_type *smt_pods,
8500 int nr_cores)
8501{
8502 struct llc_shard_layout layout = llc_calc_shard_layout(nr_cores);
8503 const struct cpumask *sibling_cpus;
8504 /* Count the number of cores in the current shard_id */
8505 int cores_in_shard = 0;
8506 unsigned int leader;
8507 /* This is a cursor for the shards. Go from zero to nr_shards - 1*/
8508 int shard_id = 0;
8509 int c;
8510
8511 /* Iterate at every CPU for a given LLC pod, and assign it a shard */
8512 for_each_cpu(c, pod_cpus) {
8513 sibling_cpus = smt_pods->pod_cpus[smt_pods->cpu_pod[c]];
8514 if (cpumask_first(sibling_cpus) == c) {
8515 /* This is the CPU leader for the siblings */
8516 if (llc_shard_is_full(cores_in_shard, shard_id, &layout)) {
8517 shard_id++;
8518 cores_in_shard = 0;
8519 }
8520 cores_in_shard++;
8521 cpu_shard_id[c] = shard_id;
8522 } else {
8523 /*
8524 * The siblings' shard MUST be the same as the leader.
8525 * never split threads in the same core.
8526 */
8527 leader = cpumask_first(sibling_cpus);
8528
8529 /*
8530 * This check silences a Warray-bounds warning on UP
8531 * configs where NR_CPUS=1 makes cpu_shard_id[]
8532 * a single-element array, and the compiler can't
8533 * prove the index is always 0.
8534 */
8535 if (WARN_ON_ONCE(leader >= nr_cpu_ids))
8536 continue;
8537 cpu_shard_id[c] = cpu_shard_id[leader];
8538 }
8539 }
8540
8541 WARN_ON_ONCE(shard_id != (layout.nr_shards - 1));
8542}
8543
8544/**
8545 * precompute_cache_shard_ids - assign each CPU its shard index within its LLC
8546 *
8547 * Iterates over all LLC pods. For each pod, counts distinct cores then assigns
8548 * shard indices to all CPUs in the pod. Must be called after WQ_AFFN_CACHE and
8549 * WQ_AFFN_SMT have been initialized.
8550 */
8551static void __init precompute_cache_shard_ids(void)
8552{
8553 struct wq_pod_type *llc_pods = &wq_pod_types[WQ_AFFN_CACHE];
8554 struct wq_pod_type *smt_pods = &wq_pod_types[WQ_AFFN_SMT];
8555 const struct cpumask *cpus_sharing_llc;
8556 int nr_cores;
8557 int pod;
8558
8559 if (!wq_cache_shard_size) {
8560 pr_warn("workqueue: cache_shard_size must be > 0, setting to 1\n");
8561 wq_cache_shard_size = 1;
8562 }
8563
8564 for (pod = 0; pod < llc_pods->nr_pods; pod++) {
8565 cpus_sharing_llc = llc_pods->pod_cpus[pod];
8566
8567 /* Number of cores in this given LLC */
8568 nr_cores = llc_count_cores(cpus_sharing_llc, smt_pods);
8569 llc_populate_cpu_shard_id(cpus_sharing_llc, smt_pods, nr_cores);
8570 }
8571}
8572
8573/*
8574 * cpus_share_cache_shard - test whether two CPUs belong to the same cache shard
8575 *
8576 * Two CPUs share a cache shard if they are in the same LLC and have the same
8577 * shard index. Used as the pod affinity callback for WQ_AFFN_CACHE_SHARD.
8578 */
8579static bool __init cpus_share_cache_shard(int cpu0, int cpu1)
8580{
8581 if (!cpus_share_cache(cpu0, cpu1))
8582 return false;
8583
8584 return cpu_shard_id[cpu0] == cpu_shard_id[cpu1];
8585}
8586
8587/**
8588 * workqueue_init_topology - initialize CPU pods for unbound workqueues
8589 *
8590 * This is the third step of three-staged workqueue subsystem initialization and
8591 * invoked after SMP and topology information are fully initialized. It
8592 * initializes the unbound CPU pods accordingly.
8593 */
8594void __init workqueue_init_topology(void)
8595{
8596 struct workqueue_struct *wq;
8597 int cpu;
8598
8599 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
8600 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
8601 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
8602 precompute_cache_shard_ids();
8603 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE_SHARD], cpus_share_cache_shard);
8604 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
8605
8606 wq_topo_initialized = true;
8607
8608 mutex_lock(&wq_pool_mutex);
8609
8610 /*
8611 * Workqueues allocated earlier would have all CPUs sharing the default
8612 * worker pool. Explicitly call unbound_wq_update_pwq() on all workqueue
8613 * and CPU combinations to apply per-pod sharing.
8614 */
8615 list_for_each_entry(wq, &workqueues, list) {
8616 for_each_online_cpu(cpu)
8617 unbound_wq_update_pwq(wq, cpu);
8618 if (wq->flags & WQ_UNBOUND) {
8619 mutex_lock(&wq->mutex);
8620 wq_update_node_max_active(wq, -1);
8621 mutex_unlock(&wq->mutex);
8622 }
8623 }
8624
8625 mutex_unlock(&wq_pool_mutex);
8626}
8627
8628void __warn_flushing_systemwide_wq(void)
8629{
8630 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
8631 dump_stack();
8632}
8633EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
8634
8635static int __init workqueue_unbound_cpus_setup(char *str)
8636{
8637 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
8638 cpumask_clear(&wq_cmdline_cpumask);
8639 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
8640 }
8641
8642 return 1;
8643}
8644__setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);