1 // SPDX-License-Identifier: GPL-2.0-only
3 * kernel/workqueue.c - generic async execution with shared worker pool
5 * Copyright (C) 2002 Ingo Molnar
7 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org>
10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu>
13 * Made to use alloc_percpu by Christoph Lameter.
15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
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.
25 * Please read Documentation/core-api/workqueue.rst for details.
28 #include <linux/export.h>
29 #include <linux/kernel.h>
30 #include <linux/sched.h>
31 #include <linux/init.h>
32 #include <linux/signal.h>
33 #include <linux/completion.h>
34 #include <linux/workqueue.h>
35 #include <linux/slab.h>
36 #include <linux/cpu.h>
37 #include <linux/notifier.h>
38 #include <linux/kthread.h>
39 #include <linux/hardirq.h>
40 #include <linux/mempolicy.h>
41 #include <linux/freezer.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 #include <linux/sched/isolation.h>
52 #include <linux/sched/debug.h>
53 #include <linux/nmi.h>
54 #include <linux/kvm_para.h>
55 #include <linux/delay.h>
57 #include "workqueue_internal.h"
63 * A bound pool is either associated or disassociated with its CPU.
64 * While associated (!DISASSOCIATED), all workers are bound to the
65 * CPU and none has %WORKER_UNBOUND set and concurrency management
68 * While DISASSOCIATED, the cpu may be offline and all workers have
69 * %WORKER_UNBOUND set and concurrency management disabled, and may
70 * be executing on any CPU. The pool behaves as an unbound one.
72 * Note that DISASSOCIATED should be flipped only while holding
73 * wq_pool_attach_mutex to avoid changing binding state while
74 * worker_attach_to_pool() is in progress.
76 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
77 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
80 WORKER_DIE = 1 << 1, /* die die die */
81 WORKER_IDLE = 1 << 2, /* is idle */
82 WORKER_PREP = 1 << 3, /* preparing to run works */
83 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
84 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
85 WORKER_REBOUND = 1 << 8, /* worker was rebound */
87 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
88 WORKER_UNBOUND | WORKER_REBOUND,
90 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
92 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
93 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
95 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
96 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
98 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
99 /* call for help after 10ms
101 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
102 CREATE_COOLDOWN = HZ, /* time to breath after fail */
105 * Rescue workers are used only on emergencies and shared by
106 * all cpus. Give MIN_NICE.
108 RESCUER_NICE_LEVEL = MIN_NICE,
109 HIGHPRI_NICE_LEVEL = MIN_NICE,
115 * Structure fields follow one of the following exclusion rules.
117 * I: Modifiable by initialization/destruction paths and read-only for
120 * P: Preemption protected. Disabling preemption is enough and should
121 * only be modified and accessed from the local cpu.
123 * L: pool->lock protected. Access with pool->lock held.
125 * K: Only modified by worker while holding pool->lock. Can be safely read by
126 * self, while holding pool->lock or from IRQ context if %current is the
129 * S: Only modified by worker self.
131 * A: wq_pool_attach_mutex protected.
133 * PL: wq_pool_mutex protected.
135 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
137 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
139 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
142 * WQ: wq->mutex protected.
144 * WR: wq->mutex protected for writes. RCU protected for reads.
146 * MD: wq_mayday_lock protected.
148 * WD: Used internally by the watchdog.
151 /* struct worker is defined in workqueue_internal.h */
154 raw_spinlock_t lock; /* the pool lock */
155 int cpu; /* I: the associated cpu */
156 int node; /* I: the associated node ID */
157 int id; /* I: pool ID */
158 unsigned int flags; /* L: flags */
160 unsigned long watchdog_ts; /* L: watchdog timestamp */
161 bool cpu_stall; /* WD: stalled cpu bound pool */
164 * The counter is incremented in a process context on the associated CPU
165 * w/ preemption disabled, and decremented or reset in the same context
166 * but w/ pool->lock held. The readers grab pool->lock and are
167 * guaranteed to see if the counter reached zero.
171 struct list_head worklist; /* L: list of pending works */
173 int nr_workers; /* L: total number of workers */
174 int nr_idle; /* L: currently idle workers */
176 struct list_head idle_list; /* L: list of idle workers */
177 struct timer_list idle_timer; /* L: worker idle timeout */
178 struct work_struct idle_cull_work; /* L: worker idle cleanup */
180 struct timer_list mayday_timer; /* L: SOS timer for workers */
182 /* a workers is either on busy_hash or idle_list, or the manager */
183 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
184 /* L: hash of busy workers */
186 struct worker *manager; /* L: purely informational */
187 struct list_head workers; /* A: attached workers */
188 struct list_head dying_workers; /* A: workers about to die */
189 struct completion *detach_completion; /* all workers detached */
191 struct ida worker_ida; /* worker IDs for task name */
193 struct workqueue_attrs *attrs; /* I: worker attributes */
194 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
195 int refcnt; /* PL: refcnt for unbound pools */
198 * Destruction of pool is RCU protected to allow dereferences
199 * from get_work_pool().
205 * Per-pool_workqueue statistics. These can be monitored using
206 * tools/workqueue/wq_monitor.py.
208 enum pool_workqueue_stats {
209 PWQ_STAT_STARTED, /* work items started execution */
210 PWQ_STAT_COMPLETED, /* work items completed execution */
211 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
212 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
213 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
214 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
215 PWQ_STAT_MAYDAY, /* maydays to rescuer */
216 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
222 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
223 * of work_struct->data are used for flags and the remaining high bits
224 * point to the pwq; thus, pwqs need to be aligned at two's power of the
225 * number of flag bits.
227 struct pool_workqueue {
228 struct worker_pool *pool; /* I: the associated pool */
229 struct workqueue_struct *wq; /* I: the owning workqueue */
230 int work_color; /* L: current color */
231 int flush_color; /* L: flushing color */
232 int refcnt; /* L: reference count */
233 int nr_in_flight[WORK_NR_COLORS];
234 /* L: nr of in_flight works */
237 * nr_active management and WORK_STRUCT_INACTIVE:
239 * When pwq->nr_active >= max_active, new work item is queued to
240 * pwq->inactive_works instead of pool->worklist and marked with
241 * WORK_STRUCT_INACTIVE.
243 * All work items marked with WORK_STRUCT_INACTIVE do not participate
244 * in pwq->nr_active and all work items in pwq->inactive_works are
245 * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE
246 * work items are in pwq->inactive_works. Some of them are ready to
247 * run in pool->worklist or worker->scheduled. Those work itmes are
248 * only struct wq_barrier which is used for flush_work() and should
249 * not participate in pwq->nr_active. For non-barrier work item, it
250 * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
252 int nr_active; /* L: nr of active works */
253 int max_active; /* L: max active works */
254 struct list_head inactive_works; /* L: inactive works */
255 struct list_head pwqs_node; /* WR: node on wq->pwqs */
256 struct list_head mayday_node; /* MD: node on wq->maydays */
258 u64 stats[PWQ_NR_STATS];
261 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
262 * and pwq_release_workfn() for details. pool_workqueue itself is also
263 * RCU protected so that the first pwq can be determined without
264 * grabbing wq->mutex.
266 struct kthread_work release_work;
268 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
271 * Structure used to wait for workqueue flush.
274 struct list_head list; /* WQ: list of flushers */
275 int flush_color; /* WQ: flush color waiting for */
276 struct completion done; /* flush completion */
282 * The externally visible workqueue. It relays the issued work items to
283 * the appropriate worker_pool through its pool_workqueues.
285 struct workqueue_struct {
286 struct list_head pwqs; /* WR: all pwqs of this wq */
287 struct list_head list; /* PR: list of all workqueues */
289 struct mutex mutex; /* protects this wq */
290 int work_color; /* WQ: current work color */
291 int flush_color; /* WQ: current flush color */
292 atomic_t nr_pwqs_to_flush; /* flush in progress */
293 struct wq_flusher *first_flusher; /* WQ: first flusher */
294 struct list_head flusher_queue; /* WQ: flush waiters */
295 struct list_head flusher_overflow; /* WQ: flush overflow list */
297 struct list_head maydays; /* MD: pwqs requesting rescue */
298 struct worker *rescuer; /* MD: rescue worker */
300 int nr_drainers; /* WQ: drain in progress */
301 int saved_max_active; /* WQ: saved pwq max_active */
303 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
304 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
307 struct wq_device *wq_dev; /* I: for sysfs interface */
309 #ifdef CONFIG_LOCKDEP
311 struct lock_class_key key;
312 struct lockdep_map lockdep_map;
314 char name[WQ_NAME_LEN]; /* I: workqueue name */
317 * Destruction of workqueue_struct is RCU protected to allow walking
318 * the workqueues list without grabbing wq_pool_mutex.
319 * This is used to dump all workqueues from sysrq.
323 /* hot fields used during command issue, aligned to cacheline */
324 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
325 struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
328 static struct kmem_cache *pwq_cache;
331 * Each pod type describes how CPUs should be grouped for unbound workqueues.
332 * See the comment above workqueue_attrs->affn_scope.
335 int nr_pods; /* number of pods */
336 cpumask_var_t *pod_cpus; /* pod -> cpus */
337 int *pod_node; /* pod -> node */
338 int *cpu_pod; /* cpu -> pod */
341 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
342 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
344 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
345 [WQ_AFFN_DFL] = "default",
346 [WQ_AFFN_CPU] = "cpu",
347 [WQ_AFFN_SMT] = "smt",
348 [WQ_AFFN_CACHE] = "cache",
349 [WQ_AFFN_NUMA] = "numa",
350 [WQ_AFFN_SYSTEM] = "system",
354 * Per-cpu work items which run for longer than the following threshold are
355 * automatically considered CPU intensive and excluded from concurrency
356 * management to prevent them from noticeably delaying other per-cpu work items.
357 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
358 * The actual value is initialized in wq_cpu_intensive_thresh_init().
360 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
361 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
363 /* see the comment above the definition of WQ_POWER_EFFICIENT */
364 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
365 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
367 static bool wq_online; /* can kworkers be created yet? */
369 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
370 static struct workqueue_attrs *wq_update_pod_attrs_buf;
372 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
373 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
374 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
375 /* wait for manager to go away */
376 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
378 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
379 static bool workqueue_freezing; /* PL: have wqs started freezing? */
381 /* PL&A: allowable cpus for unbound wqs and work items */
382 static cpumask_var_t wq_unbound_cpumask;
384 /* for further constrain wq_unbound_cpumask by cmdline parameter*/
385 static struct cpumask wq_cmdline_cpumask __initdata;
387 /* CPU where unbound work was last round robin scheduled from this CPU */
388 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
391 * Local execution of unbound work items is no longer guaranteed. The
392 * following always forces round-robin CPU selection on unbound work items
393 * to uncover usages which depend on it.
395 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
396 static bool wq_debug_force_rr_cpu = true;
398 static bool wq_debug_force_rr_cpu = false;
400 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
402 /* the per-cpu worker pools */
403 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
405 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
407 /* PL: hash of all unbound pools keyed by pool->attrs */
408 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
410 /* I: attributes used when instantiating standard unbound pools on demand */
411 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
413 /* I: attributes used when instantiating ordered pools on demand */
414 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
417 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
418 * process context while holding a pool lock. Bounce to a dedicated kthread
419 * worker to avoid A-A deadlocks.
421 static struct kthread_worker *pwq_release_worker;
423 struct workqueue_struct *system_wq __read_mostly;
424 EXPORT_SYMBOL(system_wq);
425 struct workqueue_struct *system_highpri_wq __read_mostly;
426 EXPORT_SYMBOL_GPL(system_highpri_wq);
427 struct workqueue_struct *system_long_wq __read_mostly;
428 EXPORT_SYMBOL_GPL(system_long_wq);
429 struct workqueue_struct *system_unbound_wq __read_mostly;
430 EXPORT_SYMBOL_GPL(system_unbound_wq);
431 struct workqueue_struct *system_freezable_wq __read_mostly;
432 EXPORT_SYMBOL_GPL(system_freezable_wq);
433 struct workqueue_struct *system_power_efficient_wq __read_mostly;
434 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
435 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
436 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
438 static int worker_thread(void *__worker);
439 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
440 static void show_pwq(struct pool_workqueue *pwq);
441 static void show_one_worker_pool(struct worker_pool *pool);
443 #define CREATE_TRACE_POINTS
444 #include <trace/events/workqueue.h>
446 #define assert_rcu_or_pool_mutex() \
447 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
448 !lockdep_is_held(&wq_pool_mutex), \
449 "RCU or wq_pool_mutex should be held")
451 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
452 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
453 !lockdep_is_held(&wq->mutex) && \
454 !lockdep_is_held(&wq_pool_mutex), \
455 "RCU, wq->mutex or wq_pool_mutex should be held")
457 #define for_each_cpu_worker_pool(pool, cpu) \
458 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
459 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
463 * for_each_pool - iterate through all worker_pools in the system
464 * @pool: iteration cursor
465 * @pi: integer used for iteration
467 * This must be called either with wq_pool_mutex held or RCU read
468 * locked. If the pool needs to be used beyond the locking in effect, the
469 * caller is responsible for guaranteeing that the pool stays online.
471 * The if/else clause exists only for the lockdep assertion and can be
474 #define for_each_pool(pool, pi) \
475 idr_for_each_entry(&worker_pool_idr, pool, pi) \
476 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
480 * for_each_pool_worker - iterate through all workers of a worker_pool
481 * @worker: iteration cursor
482 * @pool: worker_pool to iterate workers of
484 * This must be called with wq_pool_attach_mutex.
486 * The if/else clause exists only for the lockdep assertion and can be
489 #define for_each_pool_worker(worker, pool) \
490 list_for_each_entry((worker), &(pool)->workers, node) \
491 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
495 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
496 * @pwq: iteration cursor
497 * @wq: the target workqueue
499 * This must be called either with wq->mutex held or RCU read locked.
500 * If the pwq needs to be used beyond the locking in effect, the caller is
501 * responsible for guaranteeing that the pwq stays online.
503 * The if/else clause exists only for the lockdep assertion and can be
506 #define for_each_pwq(pwq, wq) \
507 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
508 lockdep_is_held(&(wq->mutex)))
510 #ifdef CONFIG_DEBUG_OBJECTS_WORK
512 static const struct debug_obj_descr work_debug_descr;
514 static void *work_debug_hint(void *addr)
516 return ((struct work_struct *) addr)->func;
519 static bool work_is_static_object(void *addr)
521 struct work_struct *work = addr;
523 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
527 * fixup_init is called when:
528 * - an active object is initialized
530 static bool work_fixup_init(void *addr, enum debug_obj_state state)
532 struct work_struct *work = addr;
535 case ODEBUG_STATE_ACTIVE:
536 cancel_work_sync(work);
537 debug_object_init(work, &work_debug_descr);
545 * fixup_free is called when:
546 * - an active object is freed
548 static bool work_fixup_free(void *addr, enum debug_obj_state state)
550 struct work_struct *work = addr;
553 case ODEBUG_STATE_ACTIVE:
554 cancel_work_sync(work);
555 debug_object_free(work, &work_debug_descr);
562 static const struct debug_obj_descr work_debug_descr = {
563 .name = "work_struct",
564 .debug_hint = work_debug_hint,
565 .is_static_object = work_is_static_object,
566 .fixup_init = work_fixup_init,
567 .fixup_free = work_fixup_free,
570 static inline void debug_work_activate(struct work_struct *work)
572 debug_object_activate(work, &work_debug_descr);
575 static inline void debug_work_deactivate(struct work_struct *work)
577 debug_object_deactivate(work, &work_debug_descr);
580 void __init_work(struct work_struct *work, int onstack)
583 debug_object_init_on_stack(work, &work_debug_descr);
585 debug_object_init(work, &work_debug_descr);
587 EXPORT_SYMBOL_GPL(__init_work);
589 void destroy_work_on_stack(struct work_struct *work)
591 debug_object_free(work, &work_debug_descr);
593 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
595 void destroy_delayed_work_on_stack(struct delayed_work *work)
597 destroy_timer_on_stack(&work->timer);
598 debug_object_free(&work->work, &work_debug_descr);
600 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
603 static inline void debug_work_activate(struct work_struct *work) { }
604 static inline void debug_work_deactivate(struct work_struct *work) { }
608 * worker_pool_assign_id - allocate ID and assign it to @pool
609 * @pool: the pool pointer of interest
611 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
612 * successfully, -errno on failure.
614 static int worker_pool_assign_id(struct worker_pool *pool)
618 lockdep_assert_held(&wq_pool_mutex);
620 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
629 static unsigned int work_color_to_flags(int color)
631 return color << WORK_STRUCT_COLOR_SHIFT;
634 static int get_work_color(unsigned long work_data)
636 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
637 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
640 static int work_next_color(int color)
642 return (color + 1) % WORK_NR_COLORS;
646 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
647 * contain the pointer to the queued pwq. Once execution starts, the flag
648 * is cleared and the high bits contain OFFQ flags and pool ID.
650 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
651 * and clear_work_data() can be used to set the pwq, pool or clear
652 * work->data. These functions should only be called while the work is
653 * owned - ie. while the PENDING bit is set.
655 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
656 * corresponding to a work. Pool is available once the work has been
657 * queued anywhere after initialization until it is sync canceled. pwq is
658 * available only while the work item is queued.
660 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
661 * canceled. While being canceled, a work item may have its PENDING set
662 * but stay off timer and worklist for arbitrarily long and nobody should
663 * try to steal the PENDING bit.
665 static inline void set_work_data(struct work_struct *work, unsigned long data,
668 WARN_ON_ONCE(!work_pending(work));
669 atomic_long_set(&work->data, data | flags | work_static(work));
672 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
673 unsigned long extra_flags)
675 set_work_data(work, (unsigned long)pwq,
676 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
679 static void set_work_pool_and_keep_pending(struct work_struct *work,
682 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
683 WORK_STRUCT_PENDING);
686 static void set_work_pool_and_clear_pending(struct work_struct *work,
690 * The following wmb is paired with the implied mb in
691 * test_and_set_bit(PENDING) and ensures all updates to @work made
692 * here are visible to and precede any updates by the next PENDING
696 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
698 * The following mb guarantees that previous clear of a PENDING bit
699 * will not be reordered with any speculative LOADS or STORES from
700 * work->current_func, which is executed afterwards. This possible
701 * reordering can lead to a missed execution on attempt to queue
702 * the same @work. E.g. consider this case:
705 * ---------------------------- --------------------------------
707 * 1 STORE event_indicated
708 * 2 queue_work_on() {
709 * 3 test_and_set_bit(PENDING)
710 * 4 } set_..._and_clear_pending() {
711 * 5 set_work_data() # clear bit
713 * 7 work->current_func() {
714 * 8 LOAD event_indicated
717 * Without an explicit full barrier speculative LOAD on line 8 can
718 * be executed before CPU#0 does STORE on line 1. If that happens,
719 * CPU#0 observes the PENDING bit is still set and new execution of
720 * a @work is not queued in a hope, that CPU#1 will eventually
721 * finish the queued @work. Meanwhile CPU#1 does not see
722 * event_indicated is set, because speculative LOAD was executed
723 * before actual STORE.
728 static void clear_work_data(struct work_struct *work)
730 smp_wmb(); /* see set_work_pool_and_clear_pending() */
731 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
734 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
736 return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
739 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
741 unsigned long data = atomic_long_read(&work->data);
743 if (data & WORK_STRUCT_PWQ)
744 return work_struct_pwq(data);
750 * get_work_pool - return the worker_pool a given work was associated with
751 * @work: the work item of interest
753 * Pools are created and destroyed under wq_pool_mutex, and allows read
754 * access under RCU read lock. As such, this function should be
755 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
757 * All fields of the returned pool are accessible as long as the above
758 * mentioned locking is in effect. If the returned pool needs to be used
759 * beyond the critical section, the caller is responsible for ensuring the
760 * returned pool is and stays online.
762 * Return: The worker_pool @work was last associated with. %NULL if none.
764 static struct worker_pool *get_work_pool(struct work_struct *work)
766 unsigned long data = atomic_long_read(&work->data);
769 assert_rcu_or_pool_mutex();
771 if (data & WORK_STRUCT_PWQ)
772 return work_struct_pwq(data)->pool;
774 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
775 if (pool_id == WORK_OFFQ_POOL_NONE)
778 return idr_find(&worker_pool_idr, pool_id);
782 * get_work_pool_id - return the worker pool ID a given work is associated with
783 * @work: the work item of interest
785 * Return: The worker_pool ID @work was last associated with.
786 * %WORK_OFFQ_POOL_NONE if none.
788 static int get_work_pool_id(struct work_struct *work)
790 unsigned long data = atomic_long_read(&work->data);
792 if (data & WORK_STRUCT_PWQ)
793 return work_struct_pwq(data)->pool->id;
795 return data >> WORK_OFFQ_POOL_SHIFT;
798 static void mark_work_canceling(struct work_struct *work)
800 unsigned long pool_id = get_work_pool_id(work);
802 pool_id <<= WORK_OFFQ_POOL_SHIFT;
803 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
806 static bool work_is_canceling(struct work_struct *work)
808 unsigned long data = atomic_long_read(&work->data);
810 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
814 * Policy functions. These define the policies on how the global worker
815 * pools are managed. Unless noted otherwise, these functions assume that
816 * they're being called with pool->lock held.
820 * Need to wake up a worker? Called from anything but currently
823 * Note that, because unbound workers never contribute to nr_running, this
824 * function will always return %true for unbound pools as long as the
825 * worklist isn't empty.
827 static bool need_more_worker(struct worker_pool *pool)
829 return !list_empty(&pool->worklist) && !pool->nr_running;
832 /* Can I start working? Called from busy but !running workers. */
833 static bool may_start_working(struct worker_pool *pool)
835 return pool->nr_idle;
838 /* Do I need to keep working? Called from currently running workers. */
839 static bool keep_working(struct worker_pool *pool)
841 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
844 /* Do we need a new worker? Called from manager. */
845 static bool need_to_create_worker(struct worker_pool *pool)
847 return need_more_worker(pool) && !may_start_working(pool);
850 /* Do we have too many workers and should some go away? */
851 static bool too_many_workers(struct worker_pool *pool)
853 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
854 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
855 int nr_busy = pool->nr_workers - nr_idle;
857 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
861 * worker_set_flags - set worker flags and adjust nr_running accordingly
863 * @flags: flags to set
865 * Set @flags in @worker->flags and adjust nr_running accordingly.
867 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
869 struct worker_pool *pool = worker->pool;
871 lockdep_assert_held(&pool->lock);
873 /* If transitioning into NOT_RUNNING, adjust nr_running. */
874 if ((flags & WORKER_NOT_RUNNING) &&
875 !(worker->flags & WORKER_NOT_RUNNING)) {
879 worker->flags |= flags;
883 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
885 * @flags: flags to clear
887 * Clear @flags in @worker->flags and adjust nr_running accordingly.
889 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
891 struct worker_pool *pool = worker->pool;
892 unsigned int oflags = worker->flags;
894 lockdep_assert_held(&pool->lock);
896 worker->flags &= ~flags;
899 * If transitioning out of NOT_RUNNING, increment nr_running. Note
900 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
901 * of multiple flags, not a single flag.
903 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
904 if (!(worker->flags & WORKER_NOT_RUNNING))
908 /* Return the first idle worker. Called with pool->lock held. */
909 static struct worker *first_idle_worker(struct worker_pool *pool)
911 if (unlikely(list_empty(&pool->idle_list)))
914 return list_first_entry(&pool->idle_list, struct worker, entry);
918 * worker_enter_idle - enter idle state
919 * @worker: worker which is entering idle state
921 * @worker is entering idle state. Update stats and idle timer if
925 * raw_spin_lock_irq(pool->lock).
927 static void worker_enter_idle(struct worker *worker)
929 struct worker_pool *pool = worker->pool;
931 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
932 WARN_ON_ONCE(!list_empty(&worker->entry) &&
933 (worker->hentry.next || worker->hentry.pprev)))
936 /* can't use worker_set_flags(), also called from create_worker() */
937 worker->flags |= WORKER_IDLE;
939 worker->last_active = jiffies;
941 /* idle_list is LIFO */
942 list_add(&worker->entry, &pool->idle_list);
944 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
945 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
947 /* Sanity check nr_running. */
948 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
952 * worker_leave_idle - leave idle state
953 * @worker: worker which is leaving idle state
955 * @worker is leaving idle state. Update stats.
958 * raw_spin_lock_irq(pool->lock).
960 static void worker_leave_idle(struct worker *worker)
962 struct worker_pool *pool = worker->pool;
964 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
966 worker_clr_flags(worker, WORKER_IDLE);
968 list_del_init(&worker->entry);
972 * find_worker_executing_work - find worker which is executing a work
973 * @pool: pool of interest
974 * @work: work to find worker for
976 * Find a worker which is executing @work on @pool by searching
977 * @pool->busy_hash which is keyed by the address of @work. For a worker
978 * to match, its current execution should match the address of @work and
979 * its work function. This is to avoid unwanted dependency between
980 * unrelated work executions through a work item being recycled while still
983 * This is a bit tricky. A work item may be freed once its execution
984 * starts and nothing prevents the freed area from being recycled for
985 * another work item. If the same work item address ends up being reused
986 * before the original execution finishes, workqueue will identify the
987 * recycled work item as currently executing and make it wait until the
988 * current execution finishes, introducing an unwanted dependency.
990 * This function checks the work item address and work function to avoid
991 * false positives. Note that this isn't complete as one may construct a
992 * work function which can introduce dependency onto itself through a
993 * recycled work item. Well, if somebody wants to shoot oneself in the
994 * foot that badly, there's only so much we can do, and if such deadlock
995 * actually occurs, it should be easy to locate the culprit work function.
998 * raw_spin_lock_irq(pool->lock).
1001 * Pointer to worker which is executing @work if found, %NULL
1004 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1005 struct work_struct *work)
1007 struct worker *worker;
1009 hash_for_each_possible(pool->busy_hash, worker, hentry,
1010 (unsigned long)work)
1011 if (worker->current_work == work &&
1012 worker->current_func == work->func)
1019 * move_linked_works - move linked works to a list
1020 * @work: start of series of works to be scheduled
1021 * @head: target list to append @work to
1022 * @nextp: out parameter for nested worklist walking
1024 * Schedule linked works starting from @work to @head. Work series to be
1025 * scheduled starts at @work and includes any consecutive work with
1026 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1030 * raw_spin_lock_irq(pool->lock).
1032 static void move_linked_works(struct work_struct *work, struct list_head *head,
1033 struct work_struct **nextp)
1035 struct work_struct *n;
1038 * Linked worklist will always end before the end of the list,
1039 * use NULL for list head.
1041 list_for_each_entry_safe_from(work, n, NULL, entry) {
1042 list_move_tail(&work->entry, head);
1043 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1048 * If we're already inside safe list traversal and have moved
1049 * multiple works to the scheduled queue, the next position
1050 * needs to be updated.
1057 * assign_work - assign a work item and its linked work items to a worker
1058 * @work: work to assign
1059 * @worker: worker to assign to
1060 * @nextp: out parameter for nested worklist walking
1062 * Assign @work and its linked work items to @worker. If @work is already being
1063 * executed by another worker in the same pool, it'll be punted there.
1065 * If @nextp is not NULL, it's updated to point to the next work of the last
1066 * scheduled work. This allows assign_work() to be nested inside
1067 * list_for_each_entry_safe().
1069 * Returns %true if @work was successfully assigned to @worker. %false if @work
1070 * was punted to another worker already executing it.
1072 static bool assign_work(struct work_struct *work, struct worker *worker,
1073 struct work_struct **nextp)
1075 struct worker_pool *pool = worker->pool;
1076 struct worker *collision;
1078 lockdep_assert_held(&pool->lock);
1081 * A single work shouldn't be executed concurrently by multiple workers.
1082 * __queue_work() ensures that @work doesn't jump to a different pool
1083 * while still running in the previous pool. Here, we should ensure that
1084 * @work is not executed concurrently by multiple workers from the same
1085 * pool. Check whether anyone is already processing the work. If so,
1086 * defer the work to the currently executing one.
1088 collision = find_worker_executing_work(pool, work);
1089 if (unlikely(collision)) {
1090 move_linked_works(work, &collision->scheduled, nextp);
1094 move_linked_works(work, &worker->scheduled, nextp);
1099 * kick_pool - wake up an idle worker if necessary
1100 * @pool: pool to kick
1102 * @pool may have pending work items. Wake up worker if necessary. Returns
1103 * whether a worker was woken up.
1105 static bool kick_pool(struct worker_pool *pool)
1107 struct worker *worker = first_idle_worker(pool);
1108 struct task_struct *p;
1110 lockdep_assert_held(&pool->lock);
1112 if (!need_more_worker(pool) || !worker)
1119 * Idle @worker is about to execute @work and waking up provides an
1120 * opportunity to migrate @worker at a lower cost by setting the task's
1121 * wake_cpu field. Let's see if we want to move @worker to improve
1122 * execution locality.
1124 * We're waking the worker that went idle the latest and there's some
1125 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1126 * so, setting the wake_cpu won't do anything. As this is a best-effort
1127 * optimization and the race window is narrow, let's leave as-is for
1128 * now. If this becomes pronounced, we can skip over workers which are
1129 * still on cpu when picking an idle worker.
1131 * If @pool has non-strict affinity, @worker might have ended up outside
1132 * its affinity scope. Repatriate.
1134 if (!pool->attrs->affn_strict &&
1135 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1136 struct work_struct *work = list_first_entry(&pool->worklist,
1137 struct work_struct, entry);
1138 int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask,
1140 if (wake_cpu < nr_cpu_ids) {
1141 p->wake_cpu = wake_cpu;
1142 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1150 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1153 * Concurrency-managed per-cpu work items that hog CPU for longer than
1154 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1155 * which prevents them from stalling other concurrency-managed work items. If a
1156 * work function keeps triggering this mechanism, it's likely that the work item
1157 * should be using an unbound workqueue instead.
1159 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1160 * and report them so that they can be examined and converted to use unbound
1161 * workqueues as appropriate. To avoid flooding the console, each violating work
1162 * function is tracked and reported with exponential backoff.
1164 #define WCI_MAX_ENTS 128
1169 struct hlist_node hash_node;
1172 static struct wci_ent wci_ents[WCI_MAX_ENTS];
1173 static int wci_nr_ents;
1174 static DEFINE_RAW_SPINLOCK(wci_lock);
1175 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1177 static struct wci_ent *wci_find_ent(work_func_t func)
1179 struct wci_ent *ent;
1181 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1182 (unsigned long)func) {
1183 if (ent->func == func)
1189 static void wq_cpu_intensive_report(work_func_t func)
1191 struct wci_ent *ent;
1194 ent = wci_find_ent(func);
1199 * Start reporting from the fourth time and back off
1202 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1203 if (cnt >= 4 && is_power_of_2(cnt))
1204 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1205 ent->func, wq_cpu_intensive_thresh_us,
1206 atomic64_read(&ent->cnt));
1211 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1212 * is exhausted, something went really wrong and we probably made enough
1215 if (wci_nr_ents >= WCI_MAX_ENTS)
1218 raw_spin_lock(&wci_lock);
1220 if (wci_nr_ents >= WCI_MAX_ENTS) {
1221 raw_spin_unlock(&wci_lock);
1225 if (wci_find_ent(func)) {
1226 raw_spin_unlock(&wci_lock);
1230 ent = &wci_ents[wci_nr_ents++];
1232 atomic64_set(&ent->cnt, 1);
1233 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1235 raw_spin_unlock(&wci_lock);
1238 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1239 static void wq_cpu_intensive_report(work_func_t func) {}
1240 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1243 * wq_worker_running - a worker is running again
1244 * @task: task waking up
1246 * This function is called when a worker returns from schedule()
1248 void wq_worker_running(struct task_struct *task)
1250 struct worker *worker = kthread_data(task);
1252 if (!READ_ONCE(worker->sleeping))
1256 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1257 * and the nr_running increment below, we may ruin the nr_running reset
1258 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1259 * pool. Protect against such race.
1262 if (!(worker->flags & WORKER_NOT_RUNNING))
1263 worker->pool->nr_running++;
1267 * CPU intensive auto-detection cares about how long a work item hogged
1268 * CPU without sleeping. Reset the starting timestamp on wakeup.
1270 worker->current_at = worker->task->se.sum_exec_runtime;
1272 WRITE_ONCE(worker->sleeping, 0);
1276 * wq_worker_sleeping - a worker is going to sleep
1277 * @task: task going to sleep
1279 * This function is called from schedule() when a busy worker is
1282 void wq_worker_sleeping(struct task_struct *task)
1284 struct worker *worker = kthread_data(task);
1285 struct worker_pool *pool;
1288 * Rescuers, which may not have all the fields set up like normal
1289 * workers, also reach here, let's not access anything before
1290 * checking NOT_RUNNING.
1292 if (worker->flags & WORKER_NOT_RUNNING)
1295 pool = worker->pool;
1297 /* Return if preempted before wq_worker_running() was reached */
1298 if (READ_ONCE(worker->sleeping))
1301 WRITE_ONCE(worker->sleeping, 1);
1302 raw_spin_lock_irq(&pool->lock);
1305 * Recheck in case unbind_workers() preempted us. We don't
1306 * want to decrement nr_running after the worker is unbound
1307 * and nr_running has been reset.
1309 if (worker->flags & WORKER_NOT_RUNNING) {
1310 raw_spin_unlock_irq(&pool->lock);
1315 if (kick_pool(pool))
1316 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1318 raw_spin_unlock_irq(&pool->lock);
1322 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1323 * @task: task currently running
1325 * Called from scheduler_tick(). We're in the IRQ context and the current
1326 * worker's fields which follow the 'K' locking rule can be accessed safely.
1328 void wq_worker_tick(struct task_struct *task)
1330 struct worker *worker = kthread_data(task);
1331 struct pool_workqueue *pwq = worker->current_pwq;
1332 struct worker_pool *pool = worker->pool;
1337 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1339 if (!wq_cpu_intensive_thresh_us)
1343 * If the current worker is concurrency managed and hogged the CPU for
1344 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1345 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1347 * Set @worker->sleeping means that @worker is in the process of
1348 * switching out voluntarily and won't be contributing to
1349 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1350 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1351 * double decrements. The task is releasing the CPU anyway. Let's skip.
1352 * We probably want to make this prettier in the future.
1354 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1355 worker->task->se.sum_exec_runtime - worker->current_at <
1356 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1359 raw_spin_lock(&pool->lock);
1361 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1362 wq_cpu_intensive_report(worker->current_func);
1363 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1365 if (kick_pool(pool))
1366 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1368 raw_spin_unlock(&pool->lock);
1372 * wq_worker_last_func - retrieve worker's last work function
1373 * @task: Task to retrieve last work function of.
1375 * Determine the last function a worker executed. This is called from
1376 * the scheduler to get a worker's last known identity.
1379 * raw_spin_lock_irq(rq->lock)
1381 * This function is called during schedule() when a kworker is going
1382 * to sleep. It's used by psi to identify aggregation workers during
1383 * dequeuing, to allow periodic aggregation to shut-off when that
1384 * worker is the last task in the system or cgroup to go to sleep.
1386 * As this function doesn't involve any workqueue-related locking, it
1387 * only returns stable values when called from inside the scheduler's
1388 * queuing and dequeuing paths, when @task, which must be a kworker,
1389 * is guaranteed to not be processing any works.
1392 * The last work function %current executed as a worker, NULL if it
1393 * hasn't executed any work yet.
1395 work_func_t wq_worker_last_func(struct task_struct *task)
1397 struct worker *worker = kthread_data(task);
1399 return worker->last_func;
1403 * get_pwq - get an extra reference on the specified pool_workqueue
1404 * @pwq: pool_workqueue to get
1406 * Obtain an extra reference on @pwq. The caller should guarantee that
1407 * @pwq has positive refcnt and be holding the matching pool->lock.
1409 static void get_pwq(struct pool_workqueue *pwq)
1411 lockdep_assert_held(&pwq->pool->lock);
1412 WARN_ON_ONCE(pwq->refcnt <= 0);
1417 * put_pwq - put a pool_workqueue reference
1418 * @pwq: pool_workqueue to put
1420 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1421 * destruction. The caller should be holding the matching pool->lock.
1423 static void put_pwq(struct pool_workqueue *pwq)
1425 lockdep_assert_held(&pwq->pool->lock);
1426 if (likely(--pwq->refcnt))
1429 * @pwq can't be released under pool->lock, bounce to a dedicated
1430 * kthread_worker to avoid A-A deadlocks.
1432 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1436 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1437 * @pwq: pool_workqueue to put (can be %NULL)
1439 * put_pwq() with locking. This function also allows %NULL @pwq.
1441 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1445 * As both pwqs and pools are RCU protected, the
1446 * following lock operations are safe.
1448 raw_spin_lock_irq(&pwq->pool->lock);
1450 raw_spin_unlock_irq(&pwq->pool->lock);
1454 static void pwq_activate_inactive_work(struct work_struct *work)
1456 struct pool_workqueue *pwq = get_work_pwq(work);
1458 trace_workqueue_activate_work(work);
1459 if (list_empty(&pwq->pool->worklist))
1460 pwq->pool->watchdog_ts = jiffies;
1461 move_linked_works(work, &pwq->pool->worklist, NULL);
1462 __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work));
1466 static void pwq_activate_first_inactive(struct pool_workqueue *pwq)
1468 struct work_struct *work = list_first_entry(&pwq->inactive_works,
1469 struct work_struct, entry);
1471 pwq_activate_inactive_work(work);
1475 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1476 * @pwq: pwq of interest
1477 * @work_data: work_data of work which left the queue
1479 * A work either has completed or is removed from pending queue,
1480 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1483 * raw_spin_lock_irq(pool->lock).
1485 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1487 int color = get_work_color(work_data);
1489 if (!(work_data & WORK_STRUCT_INACTIVE)) {
1491 if (!list_empty(&pwq->inactive_works)) {
1492 /* one down, submit an inactive one */
1493 if (pwq->nr_active < pwq->max_active)
1494 pwq_activate_first_inactive(pwq);
1498 pwq->nr_in_flight[color]--;
1500 /* is flush in progress and are we at the flushing tip? */
1501 if (likely(pwq->flush_color != color))
1504 /* are there still in-flight works? */
1505 if (pwq->nr_in_flight[color])
1508 /* this pwq is done, clear flush_color */
1509 pwq->flush_color = -1;
1512 * If this was the last pwq, wake up the first flusher. It
1513 * will handle the rest.
1515 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1516 complete(&pwq->wq->first_flusher->done);
1522 * try_to_grab_pending - steal work item from worklist and disable irq
1523 * @work: work item to steal
1524 * @is_dwork: @work is a delayed_work
1525 * @flags: place to store irq state
1527 * Try to grab PENDING bit of @work. This function can handle @work in any
1528 * stable state - idle, on timer or on worklist.
1532 * ======== ================================================================
1533 * 1 if @work was pending and we successfully stole PENDING
1534 * 0 if @work was idle and we claimed PENDING
1535 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1536 * -ENOENT if someone else is canceling @work, this state may persist
1537 * for arbitrarily long
1538 * ======== ================================================================
1541 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1542 * interrupted while holding PENDING and @work off queue, irq must be
1543 * disabled on entry. This, combined with delayed_work->timer being
1544 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1546 * On successful return, >= 0, irq is disabled and the caller is
1547 * responsible for releasing it using local_irq_restore(*@flags).
1549 * This function is safe to call from any context including IRQ handler.
1551 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1552 unsigned long *flags)
1554 struct worker_pool *pool;
1555 struct pool_workqueue *pwq;
1557 local_irq_save(*flags);
1559 /* try to steal the timer if it exists */
1561 struct delayed_work *dwork = to_delayed_work(work);
1564 * dwork->timer is irqsafe. If del_timer() fails, it's
1565 * guaranteed that the timer is not queued anywhere and not
1566 * running on the local CPU.
1568 if (likely(del_timer(&dwork->timer)))
1572 /* try to claim PENDING the normal way */
1573 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1578 * The queueing is in progress, or it is already queued. Try to
1579 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1581 pool = get_work_pool(work);
1585 raw_spin_lock(&pool->lock);
1587 * work->data is guaranteed to point to pwq only while the work
1588 * item is queued on pwq->wq, and both updating work->data to point
1589 * to pwq on queueing and to pool on dequeueing are done under
1590 * pwq->pool->lock. This in turn guarantees that, if work->data
1591 * points to pwq which is associated with a locked pool, the work
1592 * item is currently queued on that pool.
1594 pwq = get_work_pwq(work);
1595 if (pwq && pwq->pool == pool) {
1596 debug_work_deactivate(work);
1599 * A cancelable inactive work item must be in the
1600 * pwq->inactive_works since a queued barrier can't be
1601 * canceled (see the comments in insert_wq_barrier()).
1603 * An inactive work item cannot be grabbed directly because
1604 * it might have linked barrier work items which, if left
1605 * on the inactive_works list, will confuse pwq->nr_active
1606 * management later on and cause stall. Make sure the work
1607 * item is activated before grabbing.
1609 if (*work_data_bits(work) & WORK_STRUCT_INACTIVE)
1610 pwq_activate_inactive_work(work);
1612 list_del_init(&work->entry);
1613 pwq_dec_nr_in_flight(pwq, *work_data_bits(work));
1615 /* work->data points to pwq iff queued, point to pool */
1616 set_work_pool_and_keep_pending(work, pool->id);
1618 raw_spin_unlock(&pool->lock);
1622 raw_spin_unlock(&pool->lock);
1625 local_irq_restore(*flags);
1626 if (work_is_canceling(work))
1633 * insert_work - insert a work into a pool
1634 * @pwq: pwq @work belongs to
1635 * @work: work to insert
1636 * @head: insertion point
1637 * @extra_flags: extra WORK_STRUCT_* flags to set
1639 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1640 * work_struct flags.
1643 * raw_spin_lock_irq(pool->lock).
1645 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1646 struct list_head *head, unsigned int extra_flags)
1648 debug_work_activate(work);
1650 /* record the work call stack in order to print it in KASAN reports */
1651 kasan_record_aux_stack_noalloc(work);
1653 /* we own @work, set data and link */
1654 set_work_pwq(work, pwq, extra_flags);
1655 list_add_tail(&work->entry, head);
1660 * Test whether @work is being queued from another work executing on the
1663 static bool is_chained_work(struct workqueue_struct *wq)
1665 struct worker *worker;
1667 worker = current_wq_worker();
1669 * Return %true iff I'm a worker executing a work item on @wq. If
1670 * I'm @worker, it's safe to dereference it without locking.
1672 return worker && worker->current_pwq->wq == wq;
1676 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1677 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1678 * avoid perturbing sensitive tasks.
1680 static int wq_select_unbound_cpu(int cpu)
1684 if (likely(!wq_debug_force_rr_cpu)) {
1685 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1688 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
1691 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1692 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1693 if (unlikely(new_cpu >= nr_cpu_ids)) {
1694 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1695 if (unlikely(new_cpu >= nr_cpu_ids))
1698 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1703 static void __queue_work(int cpu, struct workqueue_struct *wq,
1704 struct work_struct *work)
1706 struct pool_workqueue *pwq;
1707 struct worker_pool *last_pool, *pool;
1708 unsigned int work_flags;
1709 unsigned int req_cpu = cpu;
1712 * While a work item is PENDING && off queue, a task trying to
1713 * steal the PENDING will busy-loop waiting for it to either get
1714 * queued or lose PENDING. Grabbing PENDING and queueing should
1715 * happen with IRQ disabled.
1717 lockdep_assert_irqs_disabled();
1721 * For a draining wq, only works from the same workqueue are
1722 * allowed. The __WQ_DESTROYING helps to spot the issue that
1723 * queues a new work item to a wq after destroy_workqueue(wq).
1725 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
1726 WARN_ON_ONCE(!is_chained_work(wq))))
1730 /* pwq which will be used unless @work is executing elsewhere */
1731 if (req_cpu == WORK_CPU_UNBOUND) {
1732 if (wq->flags & WQ_UNBOUND)
1733 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1735 cpu = raw_smp_processor_id();
1738 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
1742 * If @work was previously on a different pool, it might still be
1743 * running there, in which case the work needs to be queued on that
1744 * pool to guarantee non-reentrancy.
1746 last_pool = get_work_pool(work);
1747 if (last_pool && last_pool != pool) {
1748 struct worker *worker;
1750 raw_spin_lock(&last_pool->lock);
1752 worker = find_worker_executing_work(last_pool, work);
1754 if (worker && worker->current_pwq->wq == wq) {
1755 pwq = worker->current_pwq;
1757 WARN_ON_ONCE(pool != last_pool);
1759 /* meh... not running there, queue here */
1760 raw_spin_unlock(&last_pool->lock);
1761 raw_spin_lock(&pool->lock);
1764 raw_spin_lock(&pool->lock);
1768 * pwq is determined and locked. For unbound pools, we could have raced
1769 * with pwq release and it could already be dead. If its refcnt is zero,
1770 * repeat pwq selection. Note that unbound pwqs never die without
1771 * another pwq replacing it in cpu_pwq or while work items are executing
1772 * on it, so the retrying is guaranteed to make forward-progress.
1774 if (unlikely(!pwq->refcnt)) {
1775 if (wq->flags & WQ_UNBOUND) {
1776 raw_spin_unlock(&pool->lock);
1781 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1785 /* pwq determined, queue */
1786 trace_workqueue_queue_work(req_cpu, pwq, work);
1788 if (WARN_ON(!list_empty(&work->entry)))
1791 pwq->nr_in_flight[pwq->work_color]++;
1792 work_flags = work_color_to_flags(pwq->work_color);
1794 if (likely(pwq->nr_active < pwq->max_active)) {
1795 if (list_empty(&pool->worklist))
1796 pool->watchdog_ts = jiffies;
1798 trace_workqueue_activate_work(work);
1800 insert_work(pwq, work, &pool->worklist, work_flags);
1803 work_flags |= WORK_STRUCT_INACTIVE;
1804 insert_work(pwq, work, &pwq->inactive_works, work_flags);
1808 raw_spin_unlock(&pool->lock);
1813 * queue_work_on - queue work on specific cpu
1814 * @cpu: CPU number to execute work on
1815 * @wq: workqueue to use
1816 * @work: work to queue
1818 * We queue the work to a specific CPU, the caller must ensure it
1819 * can't go away. Callers that fail to ensure that the specified
1820 * CPU cannot go away will execute on a randomly chosen CPU.
1821 * But note well that callers specifying a CPU that never has been
1822 * online will get a splat.
1824 * Return: %false if @work was already on a queue, %true otherwise.
1826 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1827 struct work_struct *work)
1830 unsigned long flags;
1832 local_irq_save(flags);
1834 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1835 __queue_work(cpu, wq, work);
1839 local_irq_restore(flags);
1842 EXPORT_SYMBOL(queue_work_on);
1845 * select_numa_node_cpu - Select a CPU based on NUMA node
1846 * @node: NUMA node ID that we want to select a CPU from
1848 * This function will attempt to find a "random" cpu available on a given
1849 * node. If there are no CPUs available on the given node it will return
1850 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1851 * available CPU if we need to schedule this work.
1853 static int select_numa_node_cpu(int node)
1857 /* Delay binding to CPU if node is not valid or online */
1858 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1859 return WORK_CPU_UNBOUND;
1861 /* Use local node/cpu if we are already there */
1862 cpu = raw_smp_processor_id();
1863 if (node == cpu_to_node(cpu))
1866 /* Use "random" otherwise know as "first" online CPU of node */
1867 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1869 /* If CPU is valid return that, otherwise just defer */
1870 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1874 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1875 * @node: NUMA node that we are targeting the work for
1876 * @wq: workqueue to use
1877 * @work: work to queue
1879 * We queue the work to a "random" CPU within a given NUMA node. The basic
1880 * idea here is to provide a way to somehow associate work with a given
1883 * This function will only make a best effort attempt at getting this onto
1884 * the right NUMA node. If no node is requested or the requested node is
1885 * offline then we just fall back to standard queue_work behavior.
1887 * Currently the "random" CPU ends up being the first available CPU in the
1888 * intersection of cpu_online_mask and the cpumask of the node, unless we
1889 * are running on the node. In that case we just use the current CPU.
1891 * Return: %false if @work was already on a queue, %true otherwise.
1893 bool queue_work_node(int node, struct workqueue_struct *wq,
1894 struct work_struct *work)
1896 unsigned long flags;
1900 * This current implementation is specific to unbound workqueues.
1901 * Specifically we only return the first available CPU for a given
1902 * node instead of cycling through individual CPUs within the node.
1904 * If this is used with a per-cpu workqueue then the logic in
1905 * workqueue_select_cpu_near would need to be updated to allow for
1906 * some round robin type logic.
1908 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1910 local_irq_save(flags);
1912 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1913 int cpu = select_numa_node_cpu(node);
1915 __queue_work(cpu, wq, work);
1919 local_irq_restore(flags);
1922 EXPORT_SYMBOL_GPL(queue_work_node);
1924 void delayed_work_timer_fn(struct timer_list *t)
1926 struct delayed_work *dwork = from_timer(dwork, t, timer);
1928 /* should have been called from irqsafe timer with irq already off */
1929 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1931 EXPORT_SYMBOL(delayed_work_timer_fn);
1933 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1934 struct delayed_work *dwork, unsigned long delay)
1936 struct timer_list *timer = &dwork->timer;
1937 struct work_struct *work = &dwork->work;
1940 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1941 WARN_ON_ONCE(timer_pending(timer));
1942 WARN_ON_ONCE(!list_empty(&work->entry));
1945 * If @delay is 0, queue @dwork->work immediately. This is for
1946 * both optimization and correctness. The earliest @timer can
1947 * expire is on the closest next tick and delayed_work users depend
1948 * on that there's no such delay when @delay is 0.
1951 __queue_work(cpu, wq, &dwork->work);
1957 timer->expires = jiffies + delay;
1959 if (unlikely(cpu != WORK_CPU_UNBOUND))
1960 add_timer_on(timer, cpu);
1966 * queue_delayed_work_on - queue work on specific CPU after delay
1967 * @cpu: CPU number to execute work on
1968 * @wq: workqueue to use
1969 * @dwork: work to queue
1970 * @delay: number of jiffies to wait before queueing
1972 * Return: %false if @work was already on a queue, %true otherwise. If
1973 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1976 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1977 struct delayed_work *dwork, unsigned long delay)
1979 struct work_struct *work = &dwork->work;
1981 unsigned long flags;
1983 /* read the comment in __queue_work() */
1984 local_irq_save(flags);
1986 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1987 __queue_delayed_work(cpu, wq, dwork, delay);
1991 local_irq_restore(flags);
1994 EXPORT_SYMBOL(queue_delayed_work_on);
1997 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1998 * @cpu: CPU number to execute work on
1999 * @wq: workqueue to use
2000 * @dwork: work to queue
2001 * @delay: number of jiffies to wait before queueing
2003 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2004 * modify @dwork's timer so that it expires after @delay. If @delay is
2005 * zero, @work is guaranteed to be scheduled immediately regardless of its
2008 * Return: %false if @dwork was idle and queued, %true if @dwork was
2009 * pending and its timer was modified.
2011 * This function is safe to call from any context including IRQ handler.
2012 * See try_to_grab_pending() for details.
2014 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2015 struct delayed_work *dwork, unsigned long delay)
2017 unsigned long flags;
2021 ret = try_to_grab_pending(&dwork->work, true, &flags);
2022 } while (unlikely(ret == -EAGAIN));
2024 if (likely(ret >= 0)) {
2025 __queue_delayed_work(cpu, wq, dwork, delay);
2026 local_irq_restore(flags);
2029 /* -ENOENT from try_to_grab_pending() becomes %true */
2032 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2034 static void rcu_work_rcufn(struct rcu_head *rcu)
2036 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2038 /* read the comment in __queue_work() */
2039 local_irq_disable();
2040 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2045 * queue_rcu_work - queue work after a RCU grace period
2046 * @wq: workqueue to use
2047 * @rwork: work to queue
2049 * Return: %false if @rwork was already pending, %true otherwise. Note
2050 * that a full RCU grace period is guaranteed only after a %true return.
2051 * While @rwork is guaranteed to be executed after a %false return, the
2052 * execution may happen before a full RCU grace period has passed.
2054 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2056 struct work_struct *work = &rwork->work;
2058 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2060 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2066 EXPORT_SYMBOL(queue_rcu_work);
2068 static struct worker *alloc_worker(int node)
2070 struct worker *worker;
2072 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2074 INIT_LIST_HEAD(&worker->entry);
2075 INIT_LIST_HEAD(&worker->scheduled);
2076 INIT_LIST_HEAD(&worker->node);
2077 /* on creation a worker is in !idle && prep state */
2078 worker->flags = WORKER_PREP;
2083 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2085 if (pool->cpu < 0 && pool->attrs->affn_strict)
2086 return pool->attrs->__pod_cpumask;
2088 return pool->attrs->cpumask;
2092 * worker_attach_to_pool() - attach a worker to a pool
2093 * @worker: worker to be attached
2094 * @pool: the target pool
2096 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2097 * cpu-binding of @worker are kept coordinated with the pool across
2100 static void worker_attach_to_pool(struct worker *worker,
2101 struct worker_pool *pool)
2103 mutex_lock(&wq_pool_attach_mutex);
2106 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
2107 * stable across this function. See the comments above the flag
2108 * definition for details.
2110 if (pool->flags & POOL_DISASSOCIATED)
2111 worker->flags |= WORKER_UNBOUND;
2113 kthread_set_per_cpu(worker->task, pool->cpu);
2115 if (worker->rescue_wq)
2116 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2118 list_add_tail(&worker->node, &pool->workers);
2119 worker->pool = pool;
2121 mutex_unlock(&wq_pool_attach_mutex);
2125 * worker_detach_from_pool() - detach a worker from its pool
2126 * @worker: worker which is attached to its pool
2128 * Undo the attaching which had been done in worker_attach_to_pool(). The
2129 * caller worker shouldn't access to the pool after detached except it has
2130 * other reference to the pool.
2132 static void worker_detach_from_pool(struct worker *worker)
2134 struct worker_pool *pool = worker->pool;
2135 struct completion *detach_completion = NULL;
2137 mutex_lock(&wq_pool_attach_mutex);
2139 kthread_set_per_cpu(worker->task, -1);
2140 list_del(&worker->node);
2141 worker->pool = NULL;
2143 if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
2144 detach_completion = pool->detach_completion;
2145 mutex_unlock(&wq_pool_attach_mutex);
2147 /* clear leftover flags without pool->lock after it is detached */
2148 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2150 if (detach_completion)
2151 complete(detach_completion);
2155 * create_worker - create a new workqueue worker
2156 * @pool: pool the new worker will belong to
2158 * Create and start a new worker which is attached to @pool.
2161 * Might sleep. Does GFP_KERNEL allocations.
2164 * Pointer to the newly created worker.
2166 static struct worker *create_worker(struct worker_pool *pool)
2168 struct worker *worker;
2172 /* ID is needed to determine kthread name */
2173 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2175 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2180 worker = alloc_worker(pool->node);
2182 pr_err_once("workqueue: Failed to allocate a worker\n");
2189 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
2190 pool->attrs->nice < 0 ? "H" : "");
2192 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2194 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
2195 "kworker/%s", id_buf);
2196 if (IS_ERR(worker->task)) {
2197 if (PTR_ERR(worker->task) == -EINTR) {
2198 pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2201 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2207 set_user_nice(worker->task, pool->attrs->nice);
2208 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2210 /* successful, attach the worker to the pool */
2211 worker_attach_to_pool(worker, pool);
2213 /* start the newly created worker */
2214 raw_spin_lock_irq(&pool->lock);
2216 worker->pool->nr_workers++;
2217 worker_enter_idle(worker);
2221 * @worker is waiting on a completion in kthread() and will trigger hung
2222 * check if not woken up soon. As kick_pool() might not have waken it
2223 * up, wake it up explicitly once more.
2225 wake_up_process(worker->task);
2227 raw_spin_unlock_irq(&pool->lock);
2232 ida_free(&pool->worker_ida, id);
2237 static void unbind_worker(struct worker *worker)
2239 lockdep_assert_held(&wq_pool_attach_mutex);
2241 kthread_set_per_cpu(worker->task, -1);
2242 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2243 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2245 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2248 static void wake_dying_workers(struct list_head *cull_list)
2250 struct worker *worker, *tmp;
2252 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2253 list_del_init(&worker->entry);
2254 unbind_worker(worker);
2256 * If the worker was somehow already running, then it had to be
2257 * in pool->idle_list when set_worker_dying() happened or we
2258 * wouldn't have gotten here.
2260 * Thus, the worker must either have observed the WORKER_DIE
2261 * flag, or have set its state to TASK_IDLE. Either way, the
2262 * below will be observed by the worker and is safe to do
2263 * outside of pool->lock.
2265 wake_up_process(worker->task);
2270 * set_worker_dying - Tag a worker for destruction
2271 * @worker: worker to be destroyed
2272 * @list: transfer worker away from its pool->idle_list and into list
2274 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2278 * raw_spin_lock_irq(pool->lock).
2280 static void set_worker_dying(struct worker *worker, struct list_head *list)
2282 struct worker_pool *pool = worker->pool;
2284 lockdep_assert_held(&pool->lock);
2285 lockdep_assert_held(&wq_pool_attach_mutex);
2287 /* sanity check frenzy */
2288 if (WARN_ON(worker->current_work) ||
2289 WARN_ON(!list_empty(&worker->scheduled)) ||
2290 WARN_ON(!(worker->flags & WORKER_IDLE)))
2296 worker->flags |= WORKER_DIE;
2298 list_move(&worker->entry, list);
2299 list_move(&worker->node, &pool->dying_workers);
2303 * idle_worker_timeout - check if some idle workers can now be deleted.
2304 * @t: The pool's idle_timer that just expired
2306 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2307 * worker_leave_idle(), as a worker flicking between idle and active while its
2308 * pool is at the too_many_workers() tipping point would cause too much timer
2309 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2310 * it expire and re-evaluate things from there.
2312 static void idle_worker_timeout(struct timer_list *t)
2314 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2315 bool do_cull = false;
2317 if (work_pending(&pool->idle_cull_work))
2320 raw_spin_lock_irq(&pool->lock);
2322 if (too_many_workers(pool)) {
2323 struct worker *worker;
2324 unsigned long expires;
2326 /* idle_list is kept in LIFO order, check the last one */
2327 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2328 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2329 do_cull = !time_before(jiffies, expires);
2332 mod_timer(&pool->idle_timer, expires);
2334 raw_spin_unlock_irq(&pool->lock);
2337 queue_work(system_unbound_wq, &pool->idle_cull_work);
2341 * idle_cull_fn - cull workers that have been idle for too long.
2342 * @work: the pool's work for handling these idle workers
2344 * This goes through a pool's idle workers and gets rid of those that have been
2345 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2347 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2348 * culled, so this also resets worker affinity. This requires a sleepable
2349 * context, hence the split between timer callback and work item.
2351 static void idle_cull_fn(struct work_struct *work)
2353 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2354 LIST_HEAD(cull_list);
2357 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2358 * cannot proceed beyong worker_detach_from_pool() in its self-destruct
2359 * path. This is required as a previously-preempted worker could run after
2360 * set_worker_dying() has happened but before wake_dying_workers() did.
2362 mutex_lock(&wq_pool_attach_mutex);
2363 raw_spin_lock_irq(&pool->lock);
2365 while (too_many_workers(pool)) {
2366 struct worker *worker;
2367 unsigned long expires;
2369 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2370 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2372 if (time_before(jiffies, expires)) {
2373 mod_timer(&pool->idle_timer, expires);
2377 set_worker_dying(worker, &cull_list);
2380 raw_spin_unlock_irq(&pool->lock);
2381 wake_dying_workers(&cull_list);
2382 mutex_unlock(&wq_pool_attach_mutex);
2385 static void send_mayday(struct work_struct *work)
2387 struct pool_workqueue *pwq = get_work_pwq(work);
2388 struct workqueue_struct *wq = pwq->wq;
2390 lockdep_assert_held(&wq_mayday_lock);
2395 /* mayday mayday mayday */
2396 if (list_empty(&pwq->mayday_node)) {
2398 * If @pwq is for an unbound wq, its base ref may be put at
2399 * any time due to an attribute change. Pin @pwq until the
2400 * rescuer is done with it.
2403 list_add_tail(&pwq->mayday_node, &wq->maydays);
2404 wake_up_process(wq->rescuer->task);
2405 pwq->stats[PWQ_STAT_MAYDAY]++;
2409 static void pool_mayday_timeout(struct timer_list *t)
2411 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2412 struct work_struct *work;
2414 raw_spin_lock_irq(&pool->lock);
2415 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
2417 if (need_to_create_worker(pool)) {
2419 * We've been trying to create a new worker but
2420 * haven't been successful. We might be hitting an
2421 * allocation deadlock. Send distress signals to
2424 list_for_each_entry(work, &pool->worklist, entry)
2428 raw_spin_unlock(&wq_mayday_lock);
2429 raw_spin_unlock_irq(&pool->lock);
2431 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2435 * maybe_create_worker - create a new worker if necessary
2436 * @pool: pool to create a new worker for
2438 * Create a new worker for @pool if necessary. @pool is guaranteed to
2439 * have at least one idle worker on return from this function. If
2440 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2441 * sent to all rescuers with works scheduled on @pool to resolve
2442 * possible allocation deadlock.
2444 * On return, need_to_create_worker() is guaranteed to be %false and
2445 * may_start_working() %true.
2448 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2449 * multiple times. Does GFP_KERNEL allocations. Called only from
2452 static void maybe_create_worker(struct worker_pool *pool)
2453 __releases(&pool->lock)
2454 __acquires(&pool->lock)
2457 raw_spin_unlock_irq(&pool->lock);
2459 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2460 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2463 if (create_worker(pool) || !need_to_create_worker(pool))
2466 schedule_timeout_interruptible(CREATE_COOLDOWN);
2468 if (!need_to_create_worker(pool))
2472 del_timer_sync(&pool->mayday_timer);
2473 raw_spin_lock_irq(&pool->lock);
2475 * This is necessary even after a new worker was just successfully
2476 * created as @pool->lock was dropped and the new worker might have
2477 * already become busy.
2479 if (need_to_create_worker(pool))
2484 * manage_workers - manage worker pool
2487 * Assume the manager role and manage the worker pool @worker belongs
2488 * to. At any given time, there can be only zero or one manager per
2489 * pool. The exclusion is handled automatically by this function.
2491 * The caller can safely start processing works on false return. On
2492 * true return, it's guaranteed that need_to_create_worker() is false
2493 * and may_start_working() is true.
2496 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2497 * multiple times. Does GFP_KERNEL allocations.
2500 * %false if the pool doesn't need management and the caller can safely
2501 * start processing works, %true if management function was performed and
2502 * the conditions that the caller verified before calling the function may
2503 * no longer be true.
2505 static bool manage_workers(struct worker *worker)
2507 struct worker_pool *pool = worker->pool;
2509 if (pool->flags & POOL_MANAGER_ACTIVE)
2512 pool->flags |= POOL_MANAGER_ACTIVE;
2513 pool->manager = worker;
2515 maybe_create_worker(pool);
2517 pool->manager = NULL;
2518 pool->flags &= ~POOL_MANAGER_ACTIVE;
2519 rcuwait_wake_up(&manager_wait);
2524 * process_one_work - process single work
2526 * @work: work to process
2528 * Process @work. This function contains all the logics necessary to
2529 * process a single work including synchronization against and
2530 * interaction with other workers on the same cpu, queueing and
2531 * flushing. As long as context requirement is met, any worker can
2532 * call this function to process a work.
2535 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2537 static void process_one_work(struct worker *worker, struct work_struct *work)
2538 __releases(&pool->lock)
2539 __acquires(&pool->lock)
2541 struct pool_workqueue *pwq = get_work_pwq(work);
2542 struct worker_pool *pool = worker->pool;
2543 unsigned long work_data;
2544 #ifdef CONFIG_LOCKDEP
2546 * It is permissible to free the struct work_struct from
2547 * inside the function that is called from it, this we need to
2548 * take into account for lockdep too. To avoid bogus "held
2549 * lock freed" warnings as well as problems when looking into
2550 * work->lockdep_map, make a copy and use that here.
2552 struct lockdep_map lockdep_map;
2554 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2556 /* ensure we're on the correct CPU */
2557 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2558 raw_smp_processor_id() != pool->cpu);
2560 /* claim and dequeue */
2561 debug_work_deactivate(work);
2562 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2563 worker->current_work = work;
2564 worker->current_func = work->func;
2565 worker->current_pwq = pwq;
2566 worker->current_at = worker->task->se.sum_exec_runtime;
2567 work_data = *work_data_bits(work);
2568 worker->current_color = get_work_color(work_data);
2571 * Record wq name for cmdline and debug reporting, may get
2572 * overridden through set_worker_desc().
2574 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2576 list_del_init(&work->entry);
2579 * CPU intensive works don't participate in concurrency management.
2580 * They're the scheduler's responsibility. This takes @worker out
2581 * of concurrency management and the next code block will chain
2582 * execution of the pending work items.
2584 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
2585 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2588 * Kick @pool if necessary. It's always noop for per-cpu worker pools
2589 * since nr_running would always be >= 1 at this point. This is used to
2590 * chain execution of the pending work items for WORKER_NOT_RUNNING
2591 * workers such as the UNBOUND and CPU_INTENSIVE ones.
2596 * Record the last pool and clear PENDING which should be the last
2597 * update to @work. Also, do this inside @pool->lock so that
2598 * PENDING and queued state changes happen together while IRQ is
2601 set_work_pool_and_clear_pending(work, pool->id);
2603 pwq->stats[PWQ_STAT_STARTED]++;
2604 raw_spin_unlock_irq(&pool->lock);
2606 lock_map_acquire(&pwq->wq->lockdep_map);
2607 lock_map_acquire(&lockdep_map);
2609 * Strictly speaking we should mark the invariant state without holding
2610 * any locks, that is, before these two lock_map_acquire()'s.
2612 * However, that would result in:
2619 * Which would create W1->C->W1 dependencies, even though there is no
2620 * actual deadlock possible. There are two solutions, using a
2621 * read-recursive acquire on the work(queue) 'locks', but this will then
2622 * hit the lockdep limitation on recursive locks, or simply discard
2625 * AFAICT there is no possible deadlock scenario between the
2626 * flush_work() and complete() primitives (except for single-threaded
2627 * workqueues), so hiding them isn't a problem.
2629 lockdep_invariant_state(true);
2630 trace_workqueue_execute_start(work);
2631 worker->current_func(work);
2633 * While we must be careful to not use "work" after this, the trace
2634 * point will only record its address.
2636 trace_workqueue_execute_end(work, worker->current_func);
2637 pwq->stats[PWQ_STAT_COMPLETED]++;
2638 lock_map_release(&lockdep_map);
2639 lock_map_release(&pwq->wq->lockdep_map);
2641 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2642 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2643 " last function: %ps\n",
2644 current->comm, preempt_count(), task_pid_nr(current),
2645 worker->current_func);
2646 debug_show_held_locks(current);
2651 * The following prevents a kworker from hogging CPU on !PREEMPTION
2652 * kernels, where a requeueing work item waiting for something to
2653 * happen could deadlock with stop_machine as such work item could
2654 * indefinitely requeue itself while all other CPUs are trapped in
2655 * stop_machine. At the same time, report a quiescent RCU state so
2656 * the same condition doesn't freeze RCU.
2660 raw_spin_lock_irq(&pool->lock);
2663 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
2664 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
2665 * wq_cpu_intensive_thresh_us. Clear it.
2667 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2669 /* tag the worker for identification in schedule() */
2670 worker->last_func = worker->current_func;
2672 /* we're done with it, release */
2673 hash_del(&worker->hentry);
2674 worker->current_work = NULL;
2675 worker->current_func = NULL;
2676 worker->current_pwq = NULL;
2677 worker->current_color = INT_MAX;
2678 pwq_dec_nr_in_flight(pwq, work_data);
2682 * process_scheduled_works - process scheduled works
2685 * Process all scheduled works. Please note that the scheduled list
2686 * may change while processing a work, so this function repeatedly
2687 * fetches a work from the top and executes it.
2690 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2693 static void process_scheduled_works(struct worker *worker)
2695 struct work_struct *work;
2698 while ((work = list_first_entry_or_null(&worker->scheduled,
2699 struct work_struct, entry))) {
2701 worker->pool->watchdog_ts = jiffies;
2704 process_one_work(worker, work);
2708 static void set_pf_worker(bool val)
2710 mutex_lock(&wq_pool_attach_mutex);
2712 current->flags |= PF_WQ_WORKER;
2714 current->flags &= ~PF_WQ_WORKER;
2715 mutex_unlock(&wq_pool_attach_mutex);
2719 * worker_thread - the worker thread function
2722 * The worker thread function. All workers belong to a worker_pool -
2723 * either a per-cpu one or dynamic unbound one. These workers process all
2724 * work items regardless of their specific target workqueue. The only
2725 * exception is work items which belong to workqueues with a rescuer which
2726 * will be explained in rescuer_thread().
2730 static int worker_thread(void *__worker)
2732 struct worker *worker = __worker;
2733 struct worker_pool *pool = worker->pool;
2735 /* tell the scheduler that this is a workqueue worker */
2736 set_pf_worker(true);
2738 raw_spin_lock_irq(&pool->lock);
2740 /* am I supposed to die? */
2741 if (unlikely(worker->flags & WORKER_DIE)) {
2742 raw_spin_unlock_irq(&pool->lock);
2743 set_pf_worker(false);
2745 set_task_comm(worker->task, "kworker/dying");
2746 ida_free(&pool->worker_ida, worker->id);
2747 worker_detach_from_pool(worker);
2748 WARN_ON_ONCE(!list_empty(&worker->entry));
2753 worker_leave_idle(worker);
2755 /* no more worker necessary? */
2756 if (!need_more_worker(pool))
2759 /* do we need to manage? */
2760 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2764 * ->scheduled list can only be filled while a worker is
2765 * preparing to process a work or actually processing it.
2766 * Make sure nobody diddled with it while I was sleeping.
2768 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2771 * Finish PREP stage. We're guaranteed to have at least one idle
2772 * worker or that someone else has already assumed the manager
2773 * role. This is where @worker starts participating in concurrency
2774 * management if applicable and concurrency management is restored
2775 * after being rebound. See rebind_workers() for details.
2777 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2780 struct work_struct *work =
2781 list_first_entry(&pool->worklist,
2782 struct work_struct, entry);
2784 if (assign_work(work, worker, NULL))
2785 process_scheduled_works(worker);
2786 } while (keep_working(pool));
2788 worker_set_flags(worker, WORKER_PREP);
2791 * pool->lock is held and there's no work to process and no need to
2792 * manage, sleep. Workers are woken up only while holding
2793 * pool->lock or from local cpu, so setting the current state
2794 * before releasing pool->lock is enough to prevent losing any
2797 worker_enter_idle(worker);
2798 __set_current_state(TASK_IDLE);
2799 raw_spin_unlock_irq(&pool->lock);
2805 * rescuer_thread - the rescuer thread function
2808 * Workqueue rescuer thread function. There's one rescuer for each
2809 * workqueue which has WQ_MEM_RECLAIM set.
2811 * Regular work processing on a pool may block trying to create a new
2812 * worker which uses GFP_KERNEL allocation which has slight chance of
2813 * developing into deadlock if some works currently on the same queue
2814 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2815 * the problem rescuer solves.
2817 * When such condition is possible, the pool summons rescuers of all
2818 * workqueues which have works queued on the pool and let them process
2819 * those works so that forward progress can be guaranteed.
2821 * This should happen rarely.
2825 static int rescuer_thread(void *__rescuer)
2827 struct worker *rescuer = __rescuer;
2828 struct workqueue_struct *wq = rescuer->rescue_wq;
2831 set_user_nice(current, RESCUER_NICE_LEVEL);
2834 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2835 * doesn't participate in concurrency management.
2837 set_pf_worker(true);
2839 set_current_state(TASK_IDLE);
2842 * By the time the rescuer is requested to stop, the workqueue
2843 * shouldn't have any work pending, but @wq->maydays may still have
2844 * pwq(s) queued. This can happen by non-rescuer workers consuming
2845 * all the work items before the rescuer got to them. Go through
2846 * @wq->maydays processing before acting on should_stop so that the
2847 * list is always empty on exit.
2849 should_stop = kthread_should_stop();
2851 /* see whether any pwq is asking for help */
2852 raw_spin_lock_irq(&wq_mayday_lock);
2854 while (!list_empty(&wq->maydays)) {
2855 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2856 struct pool_workqueue, mayday_node);
2857 struct worker_pool *pool = pwq->pool;
2858 struct work_struct *work, *n;
2860 __set_current_state(TASK_RUNNING);
2861 list_del_init(&pwq->mayday_node);
2863 raw_spin_unlock_irq(&wq_mayday_lock);
2865 worker_attach_to_pool(rescuer, pool);
2867 raw_spin_lock_irq(&pool->lock);
2870 * Slurp in all works issued via this workqueue and
2873 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2874 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2875 if (get_work_pwq(work) == pwq &&
2876 assign_work(work, rescuer, &n))
2877 pwq->stats[PWQ_STAT_RESCUED]++;
2880 if (!list_empty(&rescuer->scheduled)) {
2881 process_scheduled_works(rescuer);
2884 * The above execution of rescued work items could
2885 * have created more to rescue through
2886 * pwq_activate_first_inactive() or chained
2887 * queueing. Let's put @pwq back on mayday list so
2888 * that such back-to-back work items, which may be
2889 * being used to relieve memory pressure, don't
2890 * incur MAYDAY_INTERVAL delay inbetween.
2892 if (pwq->nr_active && need_to_create_worker(pool)) {
2893 raw_spin_lock(&wq_mayday_lock);
2895 * Queue iff we aren't racing destruction
2896 * and somebody else hasn't queued it already.
2898 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2900 list_add_tail(&pwq->mayday_node, &wq->maydays);
2902 raw_spin_unlock(&wq_mayday_lock);
2907 * Put the reference grabbed by send_mayday(). @pool won't
2908 * go away while we're still attached to it.
2913 * Leave this pool. Notify regular workers; otherwise, we end up
2914 * with 0 concurrency and stalling the execution.
2918 raw_spin_unlock_irq(&pool->lock);
2920 worker_detach_from_pool(rescuer);
2922 raw_spin_lock_irq(&wq_mayday_lock);
2925 raw_spin_unlock_irq(&wq_mayday_lock);
2928 __set_current_state(TASK_RUNNING);
2929 set_pf_worker(false);
2933 /* rescuers should never participate in concurrency management */
2934 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2940 * check_flush_dependency - check for flush dependency sanity
2941 * @target_wq: workqueue being flushed
2942 * @target_work: work item being flushed (NULL for workqueue flushes)
2944 * %current is trying to flush the whole @target_wq or @target_work on it.
2945 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2946 * reclaiming memory or running on a workqueue which doesn't have
2947 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2950 static void check_flush_dependency(struct workqueue_struct *target_wq,
2951 struct work_struct *target_work)
2953 work_func_t target_func = target_work ? target_work->func : NULL;
2954 struct worker *worker;
2956 if (target_wq->flags & WQ_MEM_RECLAIM)
2959 worker = current_wq_worker();
2961 WARN_ONCE(current->flags & PF_MEMALLOC,
2962 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2963 current->pid, current->comm, target_wq->name, target_func);
2964 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2965 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2966 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2967 worker->current_pwq->wq->name, worker->current_func,
2968 target_wq->name, target_func);
2972 struct work_struct work;
2973 struct completion done;
2974 struct task_struct *task; /* purely informational */
2977 static void wq_barrier_func(struct work_struct *work)
2979 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2980 complete(&barr->done);
2984 * insert_wq_barrier - insert a barrier work
2985 * @pwq: pwq to insert barrier into
2986 * @barr: wq_barrier to insert
2987 * @target: target work to attach @barr to
2988 * @worker: worker currently executing @target, NULL if @target is not executing
2990 * @barr is linked to @target such that @barr is completed only after
2991 * @target finishes execution. Please note that the ordering
2992 * guarantee is observed only with respect to @target and on the local
2995 * Currently, a queued barrier can't be canceled. This is because
2996 * try_to_grab_pending() can't determine whether the work to be
2997 * grabbed is at the head of the queue and thus can't clear LINKED
2998 * flag of the previous work while there must be a valid next work
2999 * after a work with LINKED flag set.
3001 * Note that when @worker is non-NULL, @target may be modified
3002 * underneath us, so we can't reliably determine pwq from @target.
3005 * raw_spin_lock_irq(pool->lock).
3007 static void insert_wq_barrier(struct pool_workqueue *pwq,
3008 struct wq_barrier *barr,
3009 struct work_struct *target, struct worker *worker)
3011 unsigned int work_flags = 0;
3012 unsigned int work_color;
3013 struct list_head *head;
3016 * debugobject calls are safe here even with pool->lock locked
3017 * as we know for sure that this will not trigger any of the
3018 * checks and call back into the fixup functions where we
3021 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
3022 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3024 init_completion_map(&barr->done, &target->lockdep_map);
3026 barr->task = current;
3028 /* The barrier work item does not participate in pwq->nr_active. */
3029 work_flags |= WORK_STRUCT_INACTIVE;
3032 * If @target is currently being executed, schedule the
3033 * barrier to the worker; otherwise, put it after @target.
3036 head = worker->scheduled.next;
3037 work_color = worker->current_color;
3039 unsigned long *bits = work_data_bits(target);
3041 head = target->entry.next;
3042 /* there can already be other linked works, inherit and set */
3043 work_flags |= *bits & WORK_STRUCT_LINKED;
3044 work_color = get_work_color(*bits);
3045 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3048 pwq->nr_in_flight[work_color]++;
3049 work_flags |= work_color_to_flags(work_color);
3051 insert_work(pwq, &barr->work, head, work_flags);
3055 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3056 * @wq: workqueue being flushed
3057 * @flush_color: new flush color, < 0 for no-op
3058 * @work_color: new work color, < 0 for no-op
3060 * Prepare pwqs for workqueue flushing.
3062 * If @flush_color is non-negative, flush_color on all pwqs should be
3063 * -1. If no pwq has in-flight commands at the specified color, all
3064 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3065 * has in flight commands, its pwq->flush_color is set to
3066 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3067 * wakeup logic is armed and %true is returned.
3069 * The caller should have initialized @wq->first_flusher prior to
3070 * calling this function with non-negative @flush_color. If
3071 * @flush_color is negative, no flush color update is done and %false
3074 * If @work_color is non-negative, all pwqs should have the same
3075 * work_color which is previous to @work_color and all will be
3076 * advanced to @work_color.
3079 * mutex_lock(wq->mutex).
3082 * %true if @flush_color >= 0 and there's something to flush. %false
3085 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3086 int flush_color, int work_color)
3089 struct pool_workqueue *pwq;
3091 if (flush_color >= 0) {
3092 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3093 atomic_set(&wq->nr_pwqs_to_flush, 1);
3096 for_each_pwq(pwq, wq) {
3097 struct worker_pool *pool = pwq->pool;
3099 raw_spin_lock_irq(&pool->lock);
3101 if (flush_color >= 0) {
3102 WARN_ON_ONCE(pwq->flush_color != -1);
3104 if (pwq->nr_in_flight[flush_color]) {
3105 pwq->flush_color = flush_color;
3106 atomic_inc(&wq->nr_pwqs_to_flush);
3111 if (work_color >= 0) {
3112 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3113 pwq->work_color = work_color;
3116 raw_spin_unlock_irq(&pool->lock);
3119 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3120 complete(&wq->first_flusher->done);
3126 * __flush_workqueue - ensure that any scheduled work has run to completion.
3127 * @wq: workqueue to flush
3129 * This function sleeps until all work items which were queued on entry
3130 * have finished execution, but it is not livelocked by new incoming ones.
3132 void __flush_workqueue(struct workqueue_struct *wq)
3134 struct wq_flusher this_flusher = {
3135 .list = LIST_HEAD_INIT(this_flusher.list),
3137 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3141 if (WARN_ON(!wq_online))
3144 lock_map_acquire(&wq->lockdep_map);
3145 lock_map_release(&wq->lockdep_map);
3147 mutex_lock(&wq->mutex);
3150 * Start-to-wait phase
3152 next_color = work_next_color(wq->work_color);
3154 if (next_color != wq->flush_color) {
3156 * Color space is not full. The current work_color
3157 * becomes our flush_color and work_color is advanced
3160 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3161 this_flusher.flush_color = wq->work_color;
3162 wq->work_color = next_color;
3164 if (!wq->first_flusher) {
3165 /* no flush in progress, become the first flusher */
3166 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3168 wq->first_flusher = &this_flusher;
3170 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3172 /* nothing to flush, done */
3173 wq->flush_color = next_color;
3174 wq->first_flusher = NULL;
3179 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3180 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3181 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3185 * Oops, color space is full, wait on overflow queue.
3186 * The next flush completion will assign us
3187 * flush_color and transfer to flusher_queue.
3189 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3192 check_flush_dependency(wq, NULL);
3194 mutex_unlock(&wq->mutex);
3196 wait_for_completion(&this_flusher.done);
3199 * Wake-up-and-cascade phase
3201 * First flushers are responsible for cascading flushes and
3202 * handling overflow. Non-first flushers can simply return.
3204 if (READ_ONCE(wq->first_flusher) != &this_flusher)
3207 mutex_lock(&wq->mutex);
3209 /* we might have raced, check again with mutex held */
3210 if (wq->first_flusher != &this_flusher)
3213 WRITE_ONCE(wq->first_flusher, NULL);
3215 WARN_ON_ONCE(!list_empty(&this_flusher.list));
3216 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3219 struct wq_flusher *next, *tmp;
3221 /* complete all the flushers sharing the current flush color */
3222 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3223 if (next->flush_color != wq->flush_color)
3225 list_del_init(&next->list);
3226 complete(&next->done);
3229 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
3230 wq->flush_color != work_next_color(wq->work_color));
3232 /* this flush_color is finished, advance by one */
3233 wq->flush_color = work_next_color(wq->flush_color);
3235 /* one color has been freed, handle overflow queue */
3236 if (!list_empty(&wq->flusher_overflow)) {
3238 * Assign the same color to all overflowed
3239 * flushers, advance work_color and append to
3240 * flusher_queue. This is the start-to-wait
3241 * phase for these overflowed flushers.
3243 list_for_each_entry(tmp, &wq->flusher_overflow, list)
3244 tmp->flush_color = wq->work_color;
3246 wq->work_color = work_next_color(wq->work_color);
3248 list_splice_tail_init(&wq->flusher_overflow,
3249 &wq->flusher_queue);
3250 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3253 if (list_empty(&wq->flusher_queue)) {
3254 WARN_ON_ONCE(wq->flush_color != wq->work_color);
3259 * Need to flush more colors. Make the next flusher
3260 * the new first flusher and arm pwqs.
3262 WARN_ON_ONCE(wq->flush_color == wq->work_color);
3263 WARN_ON_ONCE(wq->flush_color != next->flush_color);
3265 list_del_init(&next->list);
3266 wq->first_flusher = next;
3268 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
3272 * Meh... this color is already done, clear first
3273 * flusher and repeat cascading.
3275 wq->first_flusher = NULL;
3279 mutex_unlock(&wq->mutex);
3281 EXPORT_SYMBOL(__flush_workqueue);
3284 * drain_workqueue - drain a workqueue
3285 * @wq: workqueue to drain
3287 * Wait until the workqueue becomes empty. While draining is in progress,
3288 * only chain queueing is allowed. IOW, only currently pending or running
3289 * work items on @wq can queue further work items on it. @wq is flushed
3290 * repeatedly until it becomes empty. The number of flushing is determined
3291 * by the depth of chaining and should be relatively short. Whine if it
3294 void drain_workqueue(struct workqueue_struct *wq)
3296 unsigned int flush_cnt = 0;
3297 struct pool_workqueue *pwq;
3300 * __queue_work() needs to test whether there are drainers, is much
3301 * hotter than drain_workqueue() and already looks at @wq->flags.
3302 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
3304 mutex_lock(&wq->mutex);
3305 if (!wq->nr_drainers++)
3306 wq->flags |= __WQ_DRAINING;
3307 mutex_unlock(&wq->mutex);
3309 __flush_workqueue(wq);
3311 mutex_lock(&wq->mutex);
3313 for_each_pwq(pwq, wq) {
3316 raw_spin_lock_irq(&pwq->pool->lock);
3317 drained = !pwq->nr_active && list_empty(&pwq->inactive_works);
3318 raw_spin_unlock_irq(&pwq->pool->lock);
3323 if (++flush_cnt == 10 ||
3324 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3325 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
3326 wq->name, __func__, flush_cnt);
3328 mutex_unlock(&wq->mutex);
3332 if (!--wq->nr_drainers)
3333 wq->flags &= ~__WQ_DRAINING;
3334 mutex_unlock(&wq->mutex);
3336 EXPORT_SYMBOL_GPL(drain_workqueue);
3338 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
3341 struct worker *worker = NULL;
3342 struct worker_pool *pool;
3343 struct pool_workqueue *pwq;
3348 pool = get_work_pool(work);
3354 raw_spin_lock_irq(&pool->lock);
3355 /* see the comment in try_to_grab_pending() with the same code */
3356 pwq = get_work_pwq(work);
3358 if (unlikely(pwq->pool != pool))
3361 worker = find_worker_executing_work(pool, work);
3364 pwq = worker->current_pwq;
3367 check_flush_dependency(pwq->wq, work);
3369 insert_wq_barrier(pwq, barr, work, worker);
3370 raw_spin_unlock_irq(&pool->lock);
3373 * Force a lock recursion deadlock when using flush_work() inside a
3374 * single-threaded or rescuer equipped workqueue.
3376 * For single threaded workqueues the deadlock happens when the work
3377 * is after the work issuing the flush_work(). For rescuer equipped
3378 * workqueues the deadlock happens when the rescuer stalls, blocking
3382 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3383 lock_map_acquire(&pwq->wq->lockdep_map);
3384 lock_map_release(&pwq->wq->lockdep_map);
3389 raw_spin_unlock_irq(&pool->lock);
3394 static bool __flush_work(struct work_struct *work, bool from_cancel)
3396 struct wq_barrier barr;
3398 if (WARN_ON(!wq_online))
3401 if (WARN_ON(!work->func))
3404 lock_map_acquire(&work->lockdep_map);
3405 lock_map_release(&work->lockdep_map);
3407 if (start_flush_work(work, &barr, from_cancel)) {
3408 wait_for_completion(&barr.done);
3409 destroy_work_on_stack(&barr.work);
3417 * flush_work - wait for a work to finish executing the last queueing instance
3418 * @work: the work to flush
3420 * Wait until @work has finished execution. @work is guaranteed to be idle
3421 * on return if it hasn't been requeued since flush started.
3424 * %true if flush_work() waited for the work to finish execution,
3425 * %false if it was already idle.
3427 bool flush_work(struct work_struct *work)
3429 return __flush_work(work, false);
3431 EXPORT_SYMBOL_GPL(flush_work);
3434 wait_queue_entry_t wait;
3435 struct work_struct *work;
3438 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3440 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3442 if (cwait->work != key)
3444 return autoremove_wake_function(wait, mode, sync, key);
3447 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3449 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3450 unsigned long flags;
3454 ret = try_to_grab_pending(work, is_dwork, &flags);
3456 * If someone else is already canceling, wait for it to
3457 * finish. flush_work() doesn't work for PREEMPT_NONE
3458 * because we may get scheduled between @work's completion
3459 * and the other canceling task resuming and clearing
3460 * CANCELING - flush_work() will return false immediately
3461 * as @work is no longer busy, try_to_grab_pending() will
3462 * return -ENOENT as @work is still being canceled and the
3463 * other canceling task won't be able to clear CANCELING as
3464 * we're hogging the CPU.
3466 * Let's wait for completion using a waitqueue. As this
3467 * may lead to the thundering herd problem, use a custom
3468 * wake function which matches @work along with exclusive
3471 if (unlikely(ret == -ENOENT)) {
3472 struct cwt_wait cwait;
3474 init_wait(&cwait.wait);
3475 cwait.wait.func = cwt_wakefn;
3478 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3479 TASK_UNINTERRUPTIBLE);
3480 if (work_is_canceling(work))
3482 finish_wait(&cancel_waitq, &cwait.wait);
3484 } while (unlikely(ret < 0));
3486 /* tell other tasks trying to grab @work to back off */
3487 mark_work_canceling(work);
3488 local_irq_restore(flags);
3491 * This allows canceling during early boot. We know that @work
3495 __flush_work(work, true);
3497 clear_work_data(work);
3500 * Paired with prepare_to_wait() above so that either
3501 * waitqueue_active() is visible here or !work_is_canceling() is
3505 if (waitqueue_active(&cancel_waitq))
3506 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3512 * cancel_work_sync - cancel a work and wait for it to finish
3513 * @work: the work to cancel
3515 * Cancel @work and wait for its execution to finish. This function
3516 * can be used even if the work re-queues itself or migrates to
3517 * another workqueue. On return from this function, @work is
3518 * guaranteed to be not pending or executing on any CPU.
3520 * cancel_work_sync(&delayed_work->work) must not be used for
3521 * delayed_work's. Use cancel_delayed_work_sync() instead.
3523 * The caller must ensure that the workqueue on which @work was last
3524 * queued can't be destroyed before this function returns.
3527 * %true if @work was pending, %false otherwise.
3529 bool cancel_work_sync(struct work_struct *work)
3531 return __cancel_work_timer(work, false);
3533 EXPORT_SYMBOL_GPL(cancel_work_sync);
3536 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3537 * @dwork: the delayed work to flush
3539 * Delayed timer is cancelled and the pending work is queued for
3540 * immediate execution. Like flush_work(), this function only
3541 * considers the last queueing instance of @dwork.
3544 * %true if flush_work() waited for the work to finish execution,
3545 * %false if it was already idle.
3547 bool flush_delayed_work(struct delayed_work *dwork)
3549 local_irq_disable();
3550 if (del_timer_sync(&dwork->timer))
3551 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3553 return flush_work(&dwork->work);
3555 EXPORT_SYMBOL(flush_delayed_work);
3558 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3559 * @rwork: the rcu work to flush
3562 * %true if flush_rcu_work() waited for the work to finish execution,
3563 * %false if it was already idle.
3565 bool flush_rcu_work(struct rcu_work *rwork)
3567 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3569 flush_work(&rwork->work);
3572 return flush_work(&rwork->work);
3575 EXPORT_SYMBOL(flush_rcu_work);
3577 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3579 unsigned long flags;
3583 ret = try_to_grab_pending(work, is_dwork, &flags);
3584 } while (unlikely(ret == -EAGAIN));
3586 if (unlikely(ret < 0))
3589 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3590 local_irq_restore(flags);
3595 * See cancel_delayed_work()
3597 bool cancel_work(struct work_struct *work)
3599 return __cancel_work(work, false);
3601 EXPORT_SYMBOL(cancel_work);
3604 * cancel_delayed_work - cancel a delayed work
3605 * @dwork: delayed_work to cancel
3607 * Kill off a pending delayed_work.
3609 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3613 * The work callback function may still be running on return, unless
3614 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3615 * use cancel_delayed_work_sync() to wait on it.
3617 * This function is safe to call from any context including IRQ handler.
3619 bool cancel_delayed_work(struct delayed_work *dwork)
3621 return __cancel_work(&dwork->work, true);
3623 EXPORT_SYMBOL(cancel_delayed_work);
3626 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3627 * @dwork: the delayed work cancel
3629 * This is cancel_work_sync() for delayed works.
3632 * %true if @dwork was pending, %false otherwise.
3634 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3636 return __cancel_work_timer(&dwork->work, true);
3638 EXPORT_SYMBOL(cancel_delayed_work_sync);
3641 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3642 * @func: the function to call
3644 * schedule_on_each_cpu() executes @func on each online CPU using the
3645 * system workqueue and blocks until all CPUs have completed.
3646 * schedule_on_each_cpu() is very slow.
3649 * 0 on success, -errno on failure.
3651 int schedule_on_each_cpu(work_func_t func)
3654 struct work_struct __percpu *works;
3656 works = alloc_percpu(struct work_struct);
3662 for_each_online_cpu(cpu) {
3663 struct work_struct *work = per_cpu_ptr(works, cpu);
3665 INIT_WORK(work, func);
3666 schedule_work_on(cpu, work);
3669 for_each_online_cpu(cpu)
3670 flush_work(per_cpu_ptr(works, cpu));
3678 * execute_in_process_context - reliably execute the routine with user context
3679 * @fn: the function to execute
3680 * @ew: guaranteed storage for the execute work structure (must
3681 * be available when the work executes)
3683 * Executes the function immediately if process context is available,
3684 * otherwise schedules the function for delayed execution.
3686 * Return: 0 - function was executed
3687 * 1 - function was scheduled for execution
3689 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3691 if (!in_interrupt()) {
3696 INIT_WORK(&ew->work, fn);
3697 schedule_work(&ew->work);
3701 EXPORT_SYMBOL_GPL(execute_in_process_context);
3704 * free_workqueue_attrs - free a workqueue_attrs
3705 * @attrs: workqueue_attrs to free
3707 * Undo alloc_workqueue_attrs().
3709 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3712 free_cpumask_var(attrs->cpumask);
3713 free_cpumask_var(attrs->__pod_cpumask);
3719 * alloc_workqueue_attrs - allocate a workqueue_attrs
3721 * Allocate a new workqueue_attrs, initialize with default settings and
3724 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3726 struct workqueue_attrs *alloc_workqueue_attrs(void)
3728 struct workqueue_attrs *attrs;
3730 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3733 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3735 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
3738 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3739 attrs->affn_scope = WQ_AFFN_DFL;
3742 free_workqueue_attrs(attrs);
3746 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3747 const struct workqueue_attrs *from)
3749 to->nice = from->nice;
3750 cpumask_copy(to->cpumask, from->cpumask);
3751 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
3752 to->affn_strict = from->affn_strict;
3755 * Unlike hash and equality test, copying shouldn't ignore wq-only
3756 * fields as copying is used for both pool and wq attrs. Instead,
3757 * get_unbound_pool() explicitly clears the fields.
3759 to->affn_scope = from->affn_scope;
3760 to->ordered = from->ordered;
3764 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
3765 * comments in 'struct workqueue_attrs' definition.
3767 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
3769 attrs->affn_scope = WQ_AFFN_NR_TYPES;
3770 attrs->ordered = false;
3773 /* hash value of the content of @attr */
3774 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3778 hash = jhash_1word(attrs->nice, hash);
3779 hash = jhash(cpumask_bits(attrs->cpumask),
3780 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3781 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
3782 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3783 hash = jhash_1word(attrs->affn_strict, hash);
3787 /* content equality test */
3788 static bool wqattrs_equal(const struct workqueue_attrs *a,
3789 const struct workqueue_attrs *b)
3791 if (a->nice != b->nice)
3793 if (!cpumask_equal(a->cpumask, b->cpumask))
3795 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
3797 if (a->affn_strict != b->affn_strict)
3802 /* Update @attrs with actually available CPUs */
3803 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
3804 const cpumask_t *unbound_cpumask)
3807 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
3808 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
3811 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
3812 if (unlikely(cpumask_empty(attrs->cpumask)))
3813 cpumask_copy(attrs->cpumask, unbound_cpumask);
3816 /* find wq_pod_type to use for @attrs */
3817 static const struct wq_pod_type *
3818 wqattrs_pod_type(const struct workqueue_attrs *attrs)
3820 enum wq_affn_scope scope;
3821 struct wq_pod_type *pt;
3823 /* to synchronize access to wq_affn_dfl */
3824 lockdep_assert_held(&wq_pool_mutex);
3826 if (attrs->affn_scope == WQ_AFFN_DFL)
3827 scope = wq_affn_dfl;
3829 scope = attrs->affn_scope;
3831 pt = &wq_pod_types[scope];
3833 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
3834 likely(pt->nr_pods))
3838 * Before workqueue_init_topology(), only SYSTEM is available which is
3839 * initialized in workqueue_init_early().
3841 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
3842 BUG_ON(!pt->nr_pods);
3847 * init_worker_pool - initialize a newly zalloc'd worker_pool
3848 * @pool: worker_pool to initialize
3850 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3852 * Return: 0 on success, -errno on failure. Even on failure, all fields
3853 * inside @pool proper are initialized and put_unbound_pool() can be called
3854 * on @pool safely to release it.
3856 static int init_worker_pool(struct worker_pool *pool)
3858 raw_spin_lock_init(&pool->lock);
3861 pool->node = NUMA_NO_NODE;
3862 pool->flags |= POOL_DISASSOCIATED;
3863 pool->watchdog_ts = jiffies;
3864 INIT_LIST_HEAD(&pool->worklist);
3865 INIT_LIST_HEAD(&pool->idle_list);
3866 hash_init(pool->busy_hash);
3868 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3869 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
3871 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3873 INIT_LIST_HEAD(&pool->workers);
3874 INIT_LIST_HEAD(&pool->dying_workers);
3876 ida_init(&pool->worker_ida);
3877 INIT_HLIST_NODE(&pool->hash_node);
3880 /* shouldn't fail above this point */
3881 pool->attrs = alloc_workqueue_attrs();
3885 wqattrs_clear_for_pool(pool->attrs);
3890 #ifdef CONFIG_LOCKDEP
3891 static void wq_init_lockdep(struct workqueue_struct *wq)
3895 lockdep_register_key(&wq->key);
3896 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3898 lock_name = wq->name;
3900 wq->lock_name = lock_name;
3901 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3904 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3906 lockdep_unregister_key(&wq->key);
3909 static void wq_free_lockdep(struct workqueue_struct *wq)
3911 if (wq->lock_name != wq->name)
3912 kfree(wq->lock_name);
3915 static void wq_init_lockdep(struct workqueue_struct *wq)
3919 static void wq_unregister_lockdep(struct workqueue_struct *wq)
3923 static void wq_free_lockdep(struct workqueue_struct *wq)
3928 static void rcu_free_wq(struct rcu_head *rcu)
3930 struct workqueue_struct *wq =
3931 container_of(rcu, struct workqueue_struct, rcu);
3933 wq_free_lockdep(wq);
3934 free_percpu(wq->cpu_pwq);
3935 free_workqueue_attrs(wq->unbound_attrs);
3939 static void rcu_free_pool(struct rcu_head *rcu)
3941 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3943 ida_destroy(&pool->worker_ida);
3944 free_workqueue_attrs(pool->attrs);
3949 * put_unbound_pool - put a worker_pool
3950 * @pool: worker_pool to put
3952 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3953 * safe manner. get_unbound_pool() calls this function on its failure path
3954 * and this function should be able to release pools which went through,
3955 * successfully or not, init_worker_pool().
3957 * Should be called with wq_pool_mutex held.
3959 static void put_unbound_pool(struct worker_pool *pool)
3961 DECLARE_COMPLETION_ONSTACK(detach_completion);
3962 struct worker *worker;
3963 LIST_HEAD(cull_list);
3965 lockdep_assert_held(&wq_pool_mutex);
3971 if (WARN_ON(!(pool->cpu < 0)) ||
3972 WARN_ON(!list_empty(&pool->worklist)))
3975 /* release id and unhash */
3977 idr_remove(&worker_pool_idr, pool->id);
3978 hash_del(&pool->hash_node);
3981 * Become the manager and destroy all workers. This prevents
3982 * @pool's workers from blocking on attach_mutex. We're the last
3983 * manager and @pool gets freed with the flag set.
3985 * Having a concurrent manager is quite unlikely to happen as we can
3986 * only get here with
3987 * pwq->refcnt == pool->refcnt == 0
3988 * which implies no work queued to the pool, which implies no worker can
3989 * become the manager. However a worker could have taken the role of
3990 * manager before the refcnts dropped to 0, since maybe_create_worker()
3994 rcuwait_wait_event(&manager_wait,
3995 !(pool->flags & POOL_MANAGER_ACTIVE),
3996 TASK_UNINTERRUPTIBLE);
3998 mutex_lock(&wq_pool_attach_mutex);
3999 raw_spin_lock_irq(&pool->lock);
4000 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4001 pool->flags |= POOL_MANAGER_ACTIVE;
4004 raw_spin_unlock_irq(&pool->lock);
4005 mutex_unlock(&wq_pool_attach_mutex);
4008 while ((worker = first_idle_worker(pool)))
4009 set_worker_dying(worker, &cull_list);
4010 WARN_ON(pool->nr_workers || pool->nr_idle);
4011 raw_spin_unlock_irq(&pool->lock);
4013 wake_dying_workers(&cull_list);
4015 if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
4016 pool->detach_completion = &detach_completion;
4017 mutex_unlock(&wq_pool_attach_mutex);
4019 if (pool->detach_completion)
4020 wait_for_completion(pool->detach_completion);
4022 /* shut down the timers */
4023 del_timer_sync(&pool->idle_timer);
4024 cancel_work_sync(&pool->idle_cull_work);
4025 del_timer_sync(&pool->mayday_timer);
4027 /* RCU protected to allow dereferences from get_work_pool() */
4028 call_rcu(&pool->rcu, rcu_free_pool);
4032 * get_unbound_pool - get a worker_pool with the specified attributes
4033 * @attrs: the attributes of the worker_pool to get
4035 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4036 * reference count and return it. If there already is a matching
4037 * worker_pool, it will be used; otherwise, this function attempts to
4040 * Should be called with wq_pool_mutex held.
4042 * Return: On success, a worker_pool with the same attributes as @attrs.
4043 * On failure, %NULL.
4045 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4047 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4048 u32 hash = wqattrs_hash(attrs);
4049 struct worker_pool *pool;
4050 int pod, node = NUMA_NO_NODE;
4052 lockdep_assert_held(&wq_pool_mutex);
4054 /* do we already have a matching pool? */
4055 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4056 if (wqattrs_equal(pool->attrs, attrs)) {
4062 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4063 for (pod = 0; pod < pt->nr_pods; pod++) {
4064 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4065 node = pt->pod_node[pod];
4070 /* nope, create a new one */
4071 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4072 if (!pool || init_worker_pool(pool) < 0)
4076 copy_workqueue_attrs(pool->attrs, attrs);
4077 wqattrs_clear_for_pool(pool->attrs);
4079 if (worker_pool_assign_id(pool) < 0)
4082 /* create and start the initial worker */
4083 if (wq_online && !create_worker(pool))
4087 hash_add(unbound_pool_hash, &pool->hash_node, hash);
4092 put_unbound_pool(pool);
4096 static void rcu_free_pwq(struct rcu_head *rcu)
4098 kmem_cache_free(pwq_cache,
4099 container_of(rcu, struct pool_workqueue, rcu));
4103 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
4104 * refcnt and needs to be destroyed.
4106 static void pwq_release_workfn(struct kthread_work *work)
4108 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
4110 struct workqueue_struct *wq = pwq->wq;
4111 struct worker_pool *pool = pwq->pool;
4112 bool is_last = false;
4115 * When @pwq is not linked, it doesn't hold any reference to the
4116 * @wq, and @wq is invalid to access.
4118 if (!list_empty(&pwq->pwqs_node)) {
4119 mutex_lock(&wq->mutex);
4120 list_del_rcu(&pwq->pwqs_node);
4121 is_last = list_empty(&wq->pwqs);
4122 mutex_unlock(&wq->mutex);
4125 if (wq->flags & WQ_UNBOUND) {
4126 mutex_lock(&wq_pool_mutex);
4127 put_unbound_pool(pool);
4128 mutex_unlock(&wq_pool_mutex);
4131 call_rcu(&pwq->rcu, rcu_free_pwq);
4134 * If we're the last pwq going away, @wq is already dead and no one
4135 * is gonna access it anymore. Schedule RCU free.
4138 wq_unregister_lockdep(wq);
4139 call_rcu(&wq->rcu, rcu_free_wq);
4144 * pwq_adjust_max_active - update a pwq's max_active to the current setting
4145 * @pwq: target pool_workqueue
4147 * If @pwq isn't freezing, set @pwq->max_active to the associated
4148 * workqueue's saved_max_active and activate inactive work items
4149 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
4151 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
4153 struct workqueue_struct *wq = pwq->wq;
4154 bool freezable = wq->flags & WQ_FREEZABLE;
4155 unsigned long flags;
4157 /* for @wq->saved_max_active */
4158 lockdep_assert_held(&wq->mutex);
4160 /* fast exit for non-freezable wqs */
4161 if (!freezable && pwq->max_active == wq->saved_max_active)
4164 /* this function can be called during early boot w/ irq disabled */
4165 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
4168 * During [un]freezing, the caller is responsible for ensuring that
4169 * this function is called at least once after @workqueue_freezing
4170 * is updated and visible.
4172 if (!freezable || !workqueue_freezing) {
4173 pwq->max_active = wq->saved_max_active;
4175 while (!list_empty(&pwq->inactive_works) &&
4176 pwq->nr_active < pwq->max_active)
4177 pwq_activate_first_inactive(pwq);
4179 kick_pool(pwq->pool);
4181 pwq->max_active = 0;
4184 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
4187 /* initialize newly allocated @pwq which is associated with @wq and @pool */
4188 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
4189 struct worker_pool *pool)
4191 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
4193 memset(pwq, 0, sizeof(*pwq));
4197 pwq->flush_color = -1;
4199 INIT_LIST_HEAD(&pwq->inactive_works);
4200 INIT_LIST_HEAD(&pwq->pwqs_node);
4201 INIT_LIST_HEAD(&pwq->mayday_node);
4202 kthread_init_work(&pwq->release_work, pwq_release_workfn);
4205 /* sync @pwq with the current state of its associated wq and link it */
4206 static void link_pwq(struct pool_workqueue *pwq)
4208 struct workqueue_struct *wq = pwq->wq;
4210 lockdep_assert_held(&wq->mutex);
4212 /* may be called multiple times, ignore if already linked */
4213 if (!list_empty(&pwq->pwqs_node))
4216 /* set the matching work_color */
4217 pwq->work_color = wq->work_color;
4219 /* sync max_active to the current setting */
4220 pwq_adjust_max_active(pwq);
4223 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
4226 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
4227 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
4228 const struct workqueue_attrs *attrs)
4230 struct worker_pool *pool;
4231 struct pool_workqueue *pwq;
4233 lockdep_assert_held(&wq_pool_mutex);
4235 pool = get_unbound_pool(attrs);
4239 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
4241 put_unbound_pool(pool);
4245 init_pwq(pwq, wq, pool);
4250 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
4251 * @attrs: the wq_attrs of the default pwq of the target workqueue
4252 * @cpu: the target CPU
4253 * @cpu_going_down: if >= 0, the CPU to consider as offline
4255 * Calculate the cpumask a workqueue with @attrs should use on @pod. If
4256 * @cpu_going_down is >= 0, that cpu is considered offline during calculation.
4257 * The result is stored in @attrs->__pod_cpumask.
4259 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
4260 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
4261 * intersection of the possible CPUs of @pod and @attrs->cpumask.
4263 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
4265 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
4268 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
4269 int pod = pt->cpu_pod[cpu];
4271 /* does @pod have any online CPUs @attrs wants? */
4272 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
4273 cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
4274 if (cpu_going_down >= 0)
4275 cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
4277 if (cpumask_empty(attrs->__pod_cpumask)) {
4278 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
4282 /* yeap, return possible CPUs in @pod that @attrs wants */
4283 cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
4285 if (cpumask_empty(attrs->__pod_cpumask))
4286 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
4287 "possible intersect\n");
4290 /* install @pwq into @wq's cpu_pwq and return the old pwq */
4291 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
4292 int cpu, struct pool_workqueue *pwq)
4294 struct pool_workqueue *old_pwq;
4296 lockdep_assert_held(&wq_pool_mutex);
4297 lockdep_assert_held(&wq->mutex);
4299 /* link_pwq() can handle duplicate calls */
4302 old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
4303 rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq);
4307 /* context to store the prepared attrs & pwqs before applying */
4308 struct apply_wqattrs_ctx {
4309 struct workqueue_struct *wq; /* target workqueue */
4310 struct workqueue_attrs *attrs; /* attrs to apply */
4311 struct list_head list; /* queued for batching commit */
4312 struct pool_workqueue *dfl_pwq;
4313 struct pool_workqueue *pwq_tbl[];
4316 /* free the resources after success or abort */
4317 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
4322 for_each_possible_cpu(cpu)
4323 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
4324 put_pwq_unlocked(ctx->dfl_pwq);
4326 free_workqueue_attrs(ctx->attrs);
4332 /* allocate the attrs and pwqs for later installation */
4333 static struct apply_wqattrs_ctx *
4334 apply_wqattrs_prepare(struct workqueue_struct *wq,
4335 const struct workqueue_attrs *attrs,
4336 const cpumask_var_t unbound_cpumask)
4338 struct apply_wqattrs_ctx *ctx;
4339 struct workqueue_attrs *new_attrs;
4342 lockdep_assert_held(&wq_pool_mutex);
4344 if (WARN_ON(attrs->affn_scope < 0 ||
4345 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
4346 return ERR_PTR(-EINVAL);
4348 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
4350 new_attrs = alloc_workqueue_attrs();
4351 if (!ctx || !new_attrs)
4355 * If something goes wrong during CPU up/down, we'll fall back to
4356 * the default pwq covering whole @attrs->cpumask. Always create
4357 * it even if we don't use it immediately.
4359 copy_workqueue_attrs(new_attrs, attrs);
4360 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
4361 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4362 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
4366 for_each_possible_cpu(cpu) {
4367 if (new_attrs->ordered) {
4368 ctx->dfl_pwq->refcnt++;
4369 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
4371 wq_calc_pod_cpumask(new_attrs, cpu, -1);
4372 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
4373 if (!ctx->pwq_tbl[cpu])
4378 /* save the user configured attrs and sanitize it. */
4379 copy_workqueue_attrs(new_attrs, attrs);
4380 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
4381 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4382 ctx->attrs = new_attrs;
4388 free_workqueue_attrs(new_attrs);
4389 apply_wqattrs_cleanup(ctx);
4390 return ERR_PTR(-ENOMEM);
4393 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
4394 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4398 /* all pwqs have been created successfully, let's install'em */
4399 mutex_lock(&ctx->wq->mutex);
4401 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4403 /* save the previous pwq and install the new one */
4404 for_each_possible_cpu(cpu)
4405 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
4408 /* @dfl_pwq might not have been used, ensure it's linked */
4409 link_pwq(ctx->dfl_pwq);
4410 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
4412 mutex_unlock(&ctx->wq->mutex);
4415 static void apply_wqattrs_lock(void)
4417 /* CPUs should stay stable across pwq creations and installations */
4419 mutex_lock(&wq_pool_mutex);
4422 static void apply_wqattrs_unlock(void)
4424 mutex_unlock(&wq_pool_mutex);
4428 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4429 const struct workqueue_attrs *attrs)
4431 struct apply_wqattrs_ctx *ctx;
4433 /* only unbound workqueues can change attributes */
4434 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4437 /* creating multiple pwqs breaks ordering guarantee */
4438 if (!list_empty(&wq->pwqs)) {
4439 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4442 wq->flags &= ~__WQ_ORDERED;
4445 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
4447 return PTR_ERR(ctx);
4449 /* the ctx has been prepared successfully, let's commit it */
4450 apply_wqattrs_commit(ctx);
4451 apply_wqattrs_cleanup(ctx);
4457 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4458 * @wq: the target workqueue
4459 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4461 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
4462 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
4463 * work items are affine to the pod it was issued on. Older pwqs are released as
4464 * in-flight work items finish. Note that a work item which repeatedly requeues
4465 * itself back-to-back will stay on its current pwq.
4467 * Performs GFP_KERNEL allocations.
4469 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
4471 * Return: 0 on success and -errno on failure.
4473 int apply_workqueue_attrs(struct workqueue_struct *wq,
4474 const struct workqueue_attrs *attrs)
4478 lockdep_assert_cpus_held();
4480 mutex_lock(&wq_pool_mutex);
4481 ret = apply_workqueue_attrs_locked(wq, attrs);
4482 mutex_unlock(&wq_pool_mutex);
4488 * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
4489 * @wq: the target workqueue
4490 * @cpu: the CPU to update pool association for
4491 * @hotplug_cpu: the CPU coming up or going down
4492 * @online: whether @cpu is coming up or going down
4494 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4495 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of
4499 * If pod affinity can't be adjusted due to memory allocation failure, it falls
4500 * back to @wq->dfl_pwq which may not be optimal but is always correct.
4502 * Note that when the last allowed CPU of a pod goes offline for a workqueue
4503 * with a cpumask spanning multiple pods, the workers which were already
4504 * executing the work items for the workqueue will lose their CPU affinity and
4505 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
4506 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
4507 * responsibility to flush the work item from CPU_DOWN_PREPARE.
4509 static void wq_update_pod(struct workqueue_struct *wq, int cpu,
4510 int hotplug_cpu, bool online)
4512 int off_cpu = online ? -1 : hotplug_cpu;
4513 struct pool_workqueue *old_pwq = NULL, *pwq;
4514 struct workqueue_attrs *target_attrs;
4516 lockdep_assert_held(&wq_pool_mutex);
4518 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
4522 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4523 * Let's use a preallocated one. The following buf is protected by
4524 * CPU hotplug exclusion.
4526 target_attrs = wq_update_pod_attrs_buf;
4528 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4529 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
4531 /* nothing to do if the target cpumask matches the current pwq */
4532 wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
4533 pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu),
4534 lockdep_is_held(&wq_pool_mutex));
4535 if (wqattrs_equal(target_attrs, pwq->pool->attrs))
4538 /* create a new pwq */
4539 pwq = alloc_unbound_pwq(wq, target_attrs);
4541 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
4546 /* Install the new pwq. */
4547 mutex_lock(&wq->mutex);
4548 old_pwq = install_unbound_pwq(wq, cpu, pwq);
4552 mutex_lock(&wq->mutex);
4553 raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
4554 get_pwq(wq->dfl_pwq);
4555 raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4556 old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq);
4558 mutex_unlock(&wq->mutex);
4559 put_pwq_unlocked(old_pwq);
4562 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4564 bool highpri = wq->flags & WQ_HIGHPRI;
4567 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
4571 if (!(wq->flags & WQ_UNBOUND)) {
4572 for_each_possible_cpu(cpu) {
4573 struct pool_workqueue **pwq_p =
4574 per_cpu_ptr(wq->cpu_pwq, cpu);
4575 struct worker_pool *pool =
4576 &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]);
4578 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
4583 init_pwq(*pwq_p, wq, pool);
4585 mutex_lock(&wq->mutex);
4587 mutex_unlock(&wq->mutex);
4593 if (wq->flags & __WQ_ORDERED) {
4594 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4595 /* there should only be single pwq for ordering guarantee */
4596 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4597 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4598 "ordering guarantee broken for workqueue %s\n", wq->name);
4600 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4604 /* for unbound pwq, flush the pwq_release_worker ensures that the
4605 * pwq_release_workfn() completes before calling kfree(wq).
4608 kthread_flush_worker(pwq_release_worker);
4614 for_each_possible_cpu(cpu) {
4615 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
4618 kmem_cache_free(pwq_cache, pwq);
4620 free_percpu(wq->cpu_pwq);
4626 static int wq_clamp_max_active(int max_active, unsigned int flags,
4629 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
4630 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4631 max_active, name, 1, WQ_MAX_ACTIVE);
4633 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
4637 * Workqueues which may be used during memory reclaim should have a rescuer
4638 * to guarantee forward progress.
4640 static int init_rescuer(struct workqueue_struct *wq)
4642 struct worker *rescuer;
4645 if (!(wq->flags & WQ_MEM_RECLAIM))
4648 rescuer = alloc_worker(NUMA_NO_NODE);
4650 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
4655 rescuer->rescue_wq = wq;
4656 rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
4657 if (IS_ERR(rescuer->task)) {
4658 ret = PTR_ERR(rescuer->task);
4659 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
4660 wq->name, ERR_PTR(ret));
4665 wq->rescuer = rescuer;
4666 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4667 wake_up_process(rescuer->task);
4673 struct workqueue_struct *alloc_workqueue(const char *fmt,
4675 int max_active, ...)
4678 struct workqueue_struct *wq;
4679 struct pool_workqueue *pwq;
4682 * Unbound && max_active == 1 used to imply ordered, which is no longer
4683 * the case on many machines due to per-pod pools. While
4684 * alloc_ordered_workqueue() is the right way to create an ordered
4685 * workqueue, keep the previous behavior to avoid subtle breakages.
4687 if ((flags & WQ_UNBOUND) && max_active == 1)
4688 flags |= __WQ_ORDERED;
4690 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4691 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4692 flags |= WQ_UNBOUND;
4694 /* allocate wq and format name */
4695 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
4699 if (flags & WQ_UNBOUND) {
4700 wq->unbound_attrs = alloc_workqueue_attrs();
4701 if (!wq->unbound_attrs)
4705 va_start(args, max_active);
4706 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4709 max_active = max_active ?: WQ_DFL_ACTIVE;
4710 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4714 wq->saved_max_active = max_active;
4715 mutex_init(&wq->mutex);
4716 atomic_set(&wq->nr_pwqs_to_flush, 0);
4717 INIT_LIST_HEAD(&wq->pwqs);
4718 INIT_LIST_HEAD(&wq->flusher_queue);
4719 INIT_LIST_HEAD(&wq->flusher_overflow);
4720 INIT_LIST_HEAD(&wq->maydays);
4722 wq_init_lockdep(wq);
4723 INIT_LIST_HEAD(&wq->list);
4725 if (alloc_and_link_pwqs(wq) < 0)
4726 goto err_unreg_lockdep;
4728 if (wq_online && init_rescuer(wq) < 0)
4731 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4735 * wq_pool_mutex protects global freeze state and workqueues list.
4736 * Grab it, adjust max_active and add the new @wq to workqueues
4739 mutex_lock(&wq_pool_mutex);
4741 mutex_lock(&wq->mutex);
4742 for_each_pwq(pwq, wq)
4743 pwq_adjust_max_active(pwq);
4744 mutex_unlock(&wq->mutex);
4746 list_add_tail_rcu(&wq->list, &workqueues);
4748 mutex_unlock(&wq_pool_mutex);
4753 wq_unregister_lockdep(wq);
4754 wq_free_lockdep(wq);
4756 free_workqueue_attrs(wq->unbound_attrs);
4760 destroy_workqueue(wq);
4763 EXPORT_SYMBOL_GPL(alloc_workqueue);
4765 static bool pwq_busy(struct pool_workqueue *pwq)
4769 for (i = 0; i < WORK_NR_COLORS; i++)
4770 if (pwq->nr_in_flight[i])
4773 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4775 if (pwq->nr_active || !list_empty(&pwq->inactive_works))
4782 * destroy_workqueue - safely terminate a workqueue
4783 * @wq: target workqueue
4785 * Safely destroy a workqueue. All work currently pending will be done first.
4787 void destroy_workqueue(struct workqueue_struct *wq)
4789 struct pool_workqueue *pwq;
4793 * Remove it from sysfs first so that sanity check failure doesn't
4794 * lead to sysfs name conflicts.
4796 workqueue_sysfs_unregister(wq);
4798 /* mark the workqueue destruction is in progress */
4799 mutex_lock(&wq->mutex);
4800 wq->flags |= __WQ_DESTROYING;
4801 mutex_unlock(&wq->mutex);
4803 /* drain it before proceeding with destruction */
4804 drain_workqueue(wq);
4806 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4808 struct worker *rescuer = wq->rescuer;
4810 /* this prevents new queueing */
4811 raw_spin_lock_irq(&wq_mayday_lock);
4813 raw_spin_unlock_irq(&wq_mayday_lock);
4815 /* rescuer will empty maydays list before exiting */
4816 kthread_stop(rescuer->task);
4821 * Sanity checks - grab all the locks so that we wait for all
4822 * in-flight operations which may do put_pwq().
4824 mutex_lock(&wq_pool_mutex);
4825 mutex_lock(&wq->mutex);
4826 for_each_pwq(pwq, wq) {
4827 raw_spin_lock_irq(&pwq->pool->lock);
4828 if (WARN_ON(pwq_busy(pwq))) {
4829 pr_warn("%s: %s has the following busy pwq\n",
4830 __func__, wq->name);
4832 raw_spin_unlock_irq(&pwq->pool->lock);
4833 mutex_unlock(&wq->mutex);
4834 mutex_unlock(&wq_pool_mutex);
4835 show_one_workqueue(wq);
4838 raw_spin_unlock_irq(&pwq->pool->lock);
4840 mutex_unlock(&wq->mutex);
4843 * wq list is used to freeze wq, remove from list after
4844 * flushing is complete in case freeze races us.
4846 list_del_rcu(&wq->list);
4847 mutex_unlock(&wq_pool_mutex);
4850 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
4851 * to put the base refs. @wq will be auto-destroyed from the last
4852 * pwq_put. RCU read lock prevents @wq from going away from under us.
4856 for_each_possible_cpu(cpu) {
4857 pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
4858 RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL);
4859 put_pwq_unlocked(pwq);
4862 put_pwq_unlocked(wq->dfl_pwq);
4867 EXPORT_SYMBOL_GPL(destroy_workqueue);
4870 * workqueue_set_max_active - adjust max_active of a workqueue
4871 * @wq: target workqueue
4872 * @max_active: new max_active value.
4874 * Set max_active of @wq to @max_active.
4877 * Don't call from IRQ context.
4879 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4881 struct pool_workqueue *pwq;
4883 /* disallow meddling with max_active for ordered workqueues */
4884 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4887 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4889 mutex_lock(&wq->mutex);
4891 wq->flags &= ~__WQ_ORDERED;
4892 wq->saved_max_active = max_active;
4894 for_each_pwq(pwq, wq)
4895 pwq_adjust_max_active(pwq);
4897 mutex_unlock(&wq->mutex);
4899 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4902 * current_work - retrieve %current task's work struct
4904 * Determine if %current task is a workqueue worker and what it's working on.
4905 * Useful to find out the context that the %current task is running in.
4907 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4909 struct work_struct *current_work(void)
4911 struct worker *worker = current_wq_worker();
4913 return worker ? worker->current_work : NULL;
4915 EXPORT_SYMBOL(current_work);
4918 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4920 * Determine whether %current is a workqueue rescuer. Can be used from
4921 * work functions to determine whether it's being run off the rescuer task.
4923 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4925 bool current_is_workqueue_rescuer(void)
4927 struct worker *worker = current_wq_worker();
4929 return worker && worker->rescue_wq;
4933 * workqueue_congested - test whether a workqueue is congested
4934 * @cpu: CPU in question
4935 * @wq: target workqueue
4937 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4938 * no synchronization around this function and the test result is
4939 * unreliable and only useful as advisory hints or for debugging.
4941 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4943 * With the exception of ordered workqueues, all workqueues have per-cpu
4944 * pool_workqueues, each with its own congested state. A workqueue being
4945 * congested on one CPU doesn't mean that the workqueue is contested on any
4949 * %true if congested, %false otherwise.
4951 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4953 struct pool_workqueue *pwq;
4959 if (cpu == WORK_CPU_UNBOUND)
4960 cpu = smp_processor_id();
4962 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
4963 ret = !list_empty(&pwq->inactive_works);
4970 EXPORT_SYMBOL_GPL(workqueue_congested);
4973 * work_busy - test whether a work is currently pending or running
4974 * @work: the work to be tested
4976 * Test whether @work is currently pending or running. There is no
4977 * synchronization around this function and the test result is
4978 * unreliable and only useful as advisory hints or for debugging.
4981 * OR'd bitmask of WORK_BUSY_* bits.
4983 unsigned int work_busy(struct work_struct *work)
4985 struct worker_pool *pool;
4986 unsigned long flags;
4987 unsigned int ret = 0;
4989 if (work_pending(work))
4990 ret |= WORK_BUSY_PENDING;
4993 pool = get_work_pool(work);
4995 raw_spin_lock_irqsave(&pool->lock, flags);
4996 if (find_worker_executing_work(pool, work))
4997 ret |= WORK_BUSY_RUNNING;
4998 raw_spin_unlock_irqrestore(&pool->lock, flags);
5004 EXPORT_SYMBOL_GPL(work_busy);
5007 * set_worker_desc - set description for the current work item
5008 * @fmt: printf-style format string
5009 * @...: arguments for the format string
5011 * This function can be called by a running work function to describe what
5012 * the work item is about. If the worker task gets dumped, this
5013 * information will be printed out together to help debugging. The
5014 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
5016 void set_worker_desc(const char *fmt, ...)
5018 struct worker *worker = current_wq_worker();
5022 va_start(args, fmt);
5023 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
5027 EXPORT_SYMBOL_GPL(set_worker_desc);
5030 * print_worker_info - print out worker information and description
5031 * @log_lvl: the log level to use when printing
5032 * @task: target task
5034 * If @task is a worker and currently executing a work item, print out the
5035 * name of the workqueue being serviced and worker description set with
5036 * set_worker_desc() by the currently executing work item.
5038 * This function can be safely called on any task as long as the
5039 * task_struct itself is accessible. While safe, this function isn't
5040 * synchronized and may print out mixups or garbages of limited length.
5042 void print_worker_info(const char *log_lvl, struct task_struct *task)
5044 work_func_t *fn = NULL;
5045 char name[WQ_NAME_LEN] = { };
5046 char desc[WORKER_DESC_LEN] = { };
5047 struct pool_workqueue *pwq = NULL;
5048 struct workqueue_struct *wq = NULL;
5049 struct worker *worker;
5051 if (!(task->flags & PF_WQ_WORKER))
5055 * This function is called without any synchronization and @task
5056 * could be in any state. Be careful with dereferences.
5058 worker = kthread_probe_data(task);
5061 * Carefully copy the associated workqueue's workfn, name and desc.
5062 * Keep the original last '\0' in case the original is garbage.
5064 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
5065 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
5066 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
5067 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
5068 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
5070 if (fn || name[0] || desc[0]) {
5071 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
5072 if (strcmp(name, desc))
5073 pr_cont(" (%s)", desc);
5078 static void pr_cont_pool_info(struct worker_pool *pool)
5080 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
5081 if (pool->node != NUMA_NO_NODE)
5082 pr_cont(" node=%d", pool->node);
5083 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
5086 struct pr_cont_work_struct {
5092 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
5096 if (func == pcwsp->func) {
5100 if (pcwsp->ctr == 1)
5101 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
5103 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
5106 if ((long)func == -1L)
5108 pcwsp->comma = comma;
5113 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
5115 if (work->func == wq_barrier_func) {
5116 struct wq_barrier *barr;
5118 barr = container_of(work, struct wq_barrier, work);
5120 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5121 pr_cont("%s BAR(%d)", comma ? "," : "",
5122 task_pid_nr(barr->task));
5125 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5126 pr_cont_work_flush(comma, work->func, pcwsp);
5130 static void show_pwq(struct pool_workqueue *pwq)
5132 struct pr_cont_work_struct pcws = { .ctr = 0, };
5133 struct worker_pool *pool = pwq->pool;
5134 struct work_struct *work;
5135 struct worker *worker;
5136 bool has_in_flight = false, has_pending = false;
5139 pr_info(" pwq %d:", pool->id);
5140 pr_cont_pool_info(pool);
5142 pr_cont(" active=%d/%d refcnt=%d%s\n",
5143 pwq->nr_active, pwq->max_active, pwq->refcnt,
5144 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
5146 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5147 if (worker->current_pwq == pwq) {
5148 has_in_flight = true;
5152 if (has_in_flight) {
5155 pr_info(" in-flight:");
5156 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5157 if (worker->current_pwq != pwq)
5160 pr_cont("%s %d%s:%ps", comma ? "," : "",
5161 task_pid_nr(worker->task),
5162 worker->rescue_wq ? "(RESCUER)" : "",
5163 worker->current_func);
5164 list_for_each_entry(work, &worker->scheduled, entry)
5165 pr_cont_work(false, work, &pcws);
5166 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5172 list_for_each_entry(work, &pool->worklist, entry) {
5173 if (get_work_pwq(work) == pwq) {
5181 pr_info(" pending:");
5182 list_for_each_entry(work, &pool->worklist, entry) {
5183 if (get_work_pwq(work) != pwq)
5186 pr_cont_work(comma, work, &pcws);
5187 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5189 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5193 if (!list_empty(&pwq->inactive_works)) {
5196 pr_info(" inactive:");
5197 list_for_each_entry(work, &pwq->inactive_works, entry) {
5198 pr_cont_work(comma, work, &pcws);
5199 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5201 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5207 * show_one_workqueue - dump state of specified workqueue
5208 * @wq: workqueue whose state will be printed
5210 void show_one_workqueue(struct workqueue_struct *wq)
5212 struct pool_workqueue *pwq;
5214 unsigned long flags;
5216 for_each_pwq(pwq, wq) {
5217 if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
5222 if (idle) /* Nothing to print for idle workqueue */
5225 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
5227 for_each_pwq(pwq, wq) {
5228 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
5229 if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
5231 * Defer printing to avoid deadlocks in console
5232 * drivers that queue work while holding locks
5233 * also taken in their write paths.
5235 printk_deferred_enter();
5237 printk_deferred_exit();
5239 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
5241 * We could be printing a lot from atomic context, e.g.
5242 * sysrq-t -> show_all_workqueues(). Avoid triggering
5245 touch_nmi_watchdog();
5251 * show_one_worker_pool - dump state of specified worker pool
5252 * @pool: worker pool whose state will be printed
5254 static void show_one_worker_pool(struct worker_pool *pool)
5256 struct worker *worker;
5258 unsigned long flags;
5259 unsigned long hung = 0;
5261 raw_spin_lock_irqsave(&pool->lock, flags);
5262 if (pool->nr_workers == pool->nr_idle)
5265 /* How long the first pending work is waiting for a worker. */
5266 if (!list_empty(&pool->worklist))
5267 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
5270 * Defer printing to avoid deadlocks in console drivers that
5271 * queue work while holding locks also taken in their write
5274 printk_deferred_enter();
5275 pr_info("pool %d:", pool->id);
5276 pr_cont_pool_info(pool);
5277 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
5279 pr_cont(" manager: %d",
5280 task_pid_nr(pool->manager->task));
5281 list_for_each_entry(worker, &pool->idle_list, entry) {
5282 pr_cont(" %s%d", first ? "idle: " : "",
5283 task_pid_nr(worker->task));
5287 printk_deferred_exit();
5289 raw_spin_unlock_irqrestore(&pool->lock, flags);
5291 * We could be printing a lot from atomic context, e.g.
5292 * sysrq-t -> show_all_workqueues(). Avoid triggering
5295 touch_nmi_watchdog();
5300 * show_all_workqueues - dump workqueue state
5302 * Called from a sysrq handler and prints out all busy workqueues and pools.
5304 void show_all_workqueues(void)
5306 struct workqueue_struct *wq;
5307 struct worker_pool *pool;
5312 pr_info("Showing busy workqueues and worker pools:\n");
5314 list_for_each_entry_rcu(wq, &workqueues, list)
5315 show_one_workqueue(wq);
5317 for_each_pool(pool, pi)
5318 show_one_worker_pool(pool);
5324 * show_freezable_workqueues - dump freezable workqueue state
5326 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
5329 void show_freezable_workqueues(void)
5331 struct workqueue_struct *wq;
5335 pr_info("Showing freezable workqueues that are still busy:\n");
5337 list_for_each_entry_rcu(wq, &workqueues, list) {
5338 if (!(wq->flags & WQ_FREEZABLE))
5340 show_one_workqueue(wq);
5346 /* used to show worker information through /proc/PID/{comm,stat,status} */
5347 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
5351 /* always show the actual comm */
5352 off = strscpy(buf, task->comm, size);
5356 /* stabilize PF_WQ_WORKER and worker pool association */
5357 mutex_lock(&wq_pool_attach_mutex);
5359 if (task->flags & PF_WQ_WORKER) {
5360 struct worker *worker = kthread_data(task);
5361 struct worker_pool *pool = worker->pool;
5364 raw_spin_lock_irq(&pool->lock);
5366 * ->desc tracks information (wq name or
5367 * set_worker_desc()) for the latest execution. If
5368 * current, prepend '+', otherwise '-'.
5370 if (worker->desc[0] != '\0') {
5371 if (worker->current_work)
5372 scnprintf(buf + off, size - off, "+%s",
5375 scnprintf(buf + off, size - off, "-%s",
5378 raw_spin_unlock_irq(&pool->lock);
5382 mutex_unlock(&wq_pool_attach_mutex);
5390 * There are two challenges in supporting CPU hotplug. Firstly, there
5391 * are a lot of assumptions on strong associations among work, pwq and
5392 * pool which make migrating pending and scheduled works very
5393 * difficult to implement without impacting hot paths. Secondly,
5394 * worker pools serve mix of short, long and very long running works making
5395 * blocked draining impractical.
5397 * This is solved by allowing the pools to be disassociated from the CPU
5398 * running as an unbound one and allowing it to be reattached later if the
5399 * cpu comes back online.
5402 static void unbind_workers(int cpu)
5404 struct worker_pool *pool;
5405 struct worker *worker;
5407 for_each_cpu_worker_pool(pool, cpu) {
5408 mutex_lock(&wq_pool_attach_mutex);
5409 raw_spin_lock_irq(&pool->lock);
5412 * We've blocked all attach/detach operations. Make all workers
5413 * unbound and set DISASSOCIATED. Before this, all workers
5414 * must be on the cpu. After this, they may become diasporas.
5415 * And the preemption disabled section in their sched callbacks
5416 * are guaranteed to see WORKER_UNBOUND since the code here
5417 * is on the same cpu.
5419 for_each_pool_worker(worker, pool)
5420 worker->flags |= WORKER_UNBOUND;
5422 pool->flags |= POOL_DISASSOCIATED;
5425 * The handling of nr_running in sched callbacks are disabled
5426 * now. Zap nr_running. After this, nr_running stays zero and
5427 * need_more_worker() and keep_working() are always true as
5428 * long as the worklist is not empty. This pool now behaves as
5429 * an unbound (in terms of concurrency management) pool which
5430 * are served by workers tied to the pool.
5432 pool->nr_running = 0;
5435 * With concurrency management just turned off, a busy
5436 * worker blocking could lead to lengthy stalls. Kick off
5437 * unbound chain execution of currently pending work items.
5441 raw_spin_unlock_irq(&pool->lock);
5443 for_each_pool_worker(worker, pool)
5444 unbind_worker(worker);
5446 mutex_unlock(&wq_pool_attach_mutex);
5451 * rebind_workers - rebind all workers of a pool to the associated CPU
5452 * @pool: pool of interest
5454 * @pool->cpu is coming online. Rebind all workers to the CPU.
5456 static void rebind_workers(struct worker_pool *pool)
5458 struct worker *worker;
5460 lockdep_assert_held(&wq_pool_attach_mutex);
5463 * Restore CPU affinity of all workers. As all idle workers should
5464 * be on the run-queue of the associated CPU before any local
5465 * wake-ups for concurrency management happen, restore CPU affinity
5466 * of all workers first and then clear UNBOUND. As we're called
5467 * from CPU_ONLINE, the following shouldn't fail.
5469 for_each_pool_worker(worker, pool) {
5470 kthread_set_per_cpu(worker->task, pool->cpu);
5471 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
5472 pool_allowed_cpus(pool)) < 0);
5475 raw_spin_lock_irq(&pool->lock);
5477 pool->flags &= ~POOL_DISASSOCIATED;
5479 for_each_pool_worker(worker, pool) {
5480 unsigned int worker_flags = worker->flags;
5483 * We want to clear UNBOUND but can't directly call
5484 * worker_clr_flags() or adjust nr_running. Atomically
5485 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5486 * @worker will clear REBOUND using worker_clr_flags() when
5487 * it initiates the next execution cycle thus restoring
5488 * concurrency management. Note that when or whether
5489 * @worker clears REBOUND doesn't affect correctness.
5491 * WRITE_ONCE() is necessary because @worker->flags may be
5492 * tested without holding any lock in
5493 * wq_worker_running(). Without it, NOT_RUNNING test may
5494 * fail incorrectly leading to premature concurrency
5495 * management operations.
5497 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
5498 worker_flags |= WORKER_REBOUND;
5499 worker_flags &= ~WORKER_UNBOUND;
5500 WRITE_ONCE(worker->flags, worker_flags);
5503 raw_spin_unlock_irq(&pool->lock);
5507 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5508 * @pool: unbound pool of interest
5509 * @cpu: the CPU which is coming up
5511 * An unbound pool may end up with a cpumask which doesn't have any online
5512 * CPUs. When a worker of such pool get scheduled, the scheduler resets
5513 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5514 * online CPU before, cpus_allowed of all its workers should be restored.
5516 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5518 static cpumask_t cpumask;
5519 struct worker *worker;
5521 lockdep_assert_held(&wq_pool_attach_mutex);
5523 /* is @cpu allowed for @pool? */
5524 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5527 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5529 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5530 for_each_pool_worker(worker, pool)
5531 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5534 int workqueue_prepare_cpu(unsigned int cpu)
5536 struct worker_pool *pool;
5538 for_each_cpu_worker_pool(pool, cpu) {
5539 if (pool->nr_workers)
5541 if (!create_worker(pool))
5547 int workqueue_online_cpu(unsigned int cpu)
5549 struct worker_pool *pool;
5550 struct workqueue_struct *wq;
5553 mutex_lock(&wq_pool_mutex);
5555 for_each_pool(pool, pi) {
5556 mutex_lock(&wq_pool_attach_mutex);
5558 if (pool->cpu == cpu)
5559 rebind_workers(pool);
5560 else if (pool->cpu < 0)
5561 restore_unbound_workers_cpumask(pool, cpu);
5563 mutex_unlock(&wq_pool_attach_mutex);
5566 /* update pod affinity of unbound workqueues */
5567 list_for_each_entry(wq, &workqueues, list) {
5568 struct workqueue_attrs *attrs = wq->unbound_attrs;
5571 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5574 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
5575 wq_update_pod(wq, tcpu, cpu, true);
5579 mutex_unlock(&wq_pool_mutex);
5583 int workqueue_offline_cpu(unsigned int cpu)
5585 struct workqueue_struct *wq;
5587 /* unbinding per-cpu workers should happen on the local CPU */
5588 if (WARN_ON(cpu != smp_processor_id()))
5591 unbind_workers(cpu);
5593 /* update pod affinity of unbound workqueues */
5594 mutex_lock(&wq_pool_mutex);
5595 list_for_each_entry(wq, &workqueues, list) {
5596 struct workqueue_attrs *attrs = wq->unbound_attrs;
5599 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5602 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
5603 wq_update_pod(wq, tcpu, cpu, false);
5606 mutex_unlock(&wq_pool_mutex);
5611 struct work_for_cpu {
5612 struct work_struct work;
5618 static void work_for_cpu_fn(struct work_struct *work)
5620 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5622 wfc->ret = wfc->fn(wfc->arg);
5626 * work_on_cpu_key - run a function in thread context on a particular cpu
5627 * @cpu: the cpu to run on
5628 * @fn: the function to run
5629 * @arg: the function arg
5630 * @key: The lock class key for lock debugging purposes
5632 * It is up to the caller to ensure that the cpu doesn't go offline.
5633 * The caller must not hold any locks which would prevent @fn from completing.
5635 * Return: The value @fn returns.
5637 long work_on_cpu_key(int cpu, long (*fn)(void *),
5638 void *arg, struct lock_class_key *key)
5640 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5642 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
5643 schedule_work_on(cpu, &wfc.work);
5644 flush_work(&wfc.work);
5645 destroy_work_on_stack(&wfc.work);
5648 EXPORT_SYMBOL_GPL(work_on_cpu_key);
5651 * work_on_cpu_safe_key - run a function in thread context on a particular cpu
5652 * @cpu: the cpu to run on
5653 * @fn: the function to run
5654 * @arg: the function argument
5655 * @key: The lock class key for lock debugging purposes
5657 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5658 * any locks which would prevent @fn from completing.
5660 * Return: The value @fn returns.
5662 long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
5663 void *arg, struct lock_class_key *key)
5668 if (cpu_online(cpu))
5669 ret = work_on_cpu_key(cpu, fn, arg, key);
5673 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
5674 #endif /* CONFIG_SMP */
5676 #ifdef CONFIG_FREEZER
5679 * freeze_workqueues_begin - begin freezing workqueues
5681 * Start freezing workqueues. After this function returns, all freezable
5682 * workqueues will queue new works to their inactive_works list instead of
5686 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5688 void freeze_workqueues_begin(void)
5690 struct workqueue_struct *wq;
5691 struct pool_workqueue *pwq;
5693 mutex_lock(&wq_pool_mutex);
5695 WARN_ON_ONCE(workqueue_freezing);
5696 workqueue_freezing = true;
5698 list_for_each_entry(wq, &workqueues, list) {
5699 mutex_lock(&wq->mutex);
5700 for_each_pwq(pwq, wq)
5701 pwq_adjust_max_active(pwq);
5702 mutex_unlock(&wq->mutex);
5705 mutex_unlock(&wq_pool_mutex);
5709 * freeze_workqueues_busy - are freezable workqueues still busy?
5711 * Check whether freezing is complete. This function must be called
5712 * between freeze_workqueues_begin() and thaw_workqueues().
5715 * Grabs and releases wq_pool_mutex.
5718 * %true if some freezable workqueues are still busy. %false if freezing
5721 bool freeze_workqueues_busy(void)
5724 struct workqueue_struct *wq;
5725 struct pool_workqueue *pwq;
5727 mutex_lock(&wq_pool_mutex);
5729 WARN_ON_ONCE(!workqueue_freezing);
5731 list_for_each_entry(wq, &workqueues, list) {
5732 if (!(wq->flags & WQ_FREEZABLE))
5735 * nr_active is monotonically decreasing. It's safe
5736 * to peek without lock.
5739 for_each_pwq(pwq, wq) {
5740 WARN_ON_ONCE(pwq->nr_active < 0);
5741 if (pwq->nr_active) {
5750 mutex_unlock(&wq_pool_mutex);
5755 * thaw_workqueues - thaw workqueues
5757 * Thaw workqueues. Normal queueing is restored and all collected
5758 * frozen works are transferred to their respective pool worklists.
5761 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5763 void thaw_workqueues(void)
5765 struct workqueue_struct *wq;
5766 struct pool_workqueue *pwq;
5768 mutex_lock(&wq_pool_mutex);
5770 if (!workqueue_freezing)
5773 workqueue_freezing = false;
5775 /* restore max_active and repopulate worklist */
5776 list_for_each_entry(wq, &workqueues, list) {
5777 mutex_lock(&wq->mutex);
5778 for_each_pwq(pwq, wq)
5779 pwq_adjust_max_active(pwq);
5780 mutex_unlock(&wq->mutex);
5784 mutex_unlock(&wq_pool_mutex);
5786 #endif /* CONFIG_FREEZER */
5788 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
5792 struct workqueue_struct *wq;
5793 struct apply_wqattrs_ctx *ctx, *n;
5795 lockdep_assert_held(&wq_pool_mutex);
5797 list_for_each_entry(wq, &workqueues, list) {
5798 if (!(wq->flags & WQ_UNBOUND))
5800 /* creating multiple pwqs breaks ordering guarantee */
5801 if (wq->flags & __WQ_ORDERED)
5804 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
5810 list_add_tail(&ctx->list, &ctxs);
5813 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5815 apply_wqattrs_commit(ctx);
5816 apply_wqattrs_cleanup(ctx);
5820 mutex_lock(&wq_pool_attach_mutex);
5821 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
5822 mutex_unlock(&wq_pool_attach_mutex);
5828 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5829 * @cpumask: the cpumask to set
5831 * The low-level workqueues cpumask is a global cpumask that limits
5832 * the affinity of all unbound workqueues. This function check the @cpumask
5833 * and apply it to all unbound workqueues and updates all pwqs of them.
5835 * Return: 0 - Success
5836 * -EINVAL - Invalid @cpumask
5837 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5839 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5844 * Not excluding isolated cpus on purpose.
5845 * If the user wishes to include them, we allow that.
5847 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5848 if (!cpumask_empty(cpumask)) {
5849 apply_wqattrs_lock();
5850 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
5855 ret = workqueue_apply_unbound_cpumask(cpumask);
5858 apply_wqattrs_unlock();
5864 static int parse_affn_scope(const char *val)
5868 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
5869 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
5875 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
5877 struct workqueue_struct *wq;
5880 affn = parse_affn_scope(val);
5883 if (affn == WQ_AFFN_DFL)
5887 mutex_lock(&wq_pool_mutex);
5891 list_for_each_entry(wq, &workqueues, list) {
5892 for_each_online_cpu(cpu) {
5893 wq_update_pod(wq, cpu, cpu, true);
5897 mutex_unlock(&wq_pool_mutex);
5903 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
5905 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
5908 static const struct kernel_param_ops wq_affn_dfl_ops = {
5909 .set = wq_affn_dfl_set,
5910 .get = wq_affn_dfl_get,
5913 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
5917 * Workqueues with WQ_SYSFS flag set is visible to userland via
5918 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5919 * following attributes.
5921 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5922 * max_active RW int : maximum number of in-flight work items
5924 * Unbound workqueues have the following extra attributes.
5926 * nice RW int : nice value of the workers
5927 * cpumask RW mask : bitmask of allowed CPUs for the workers
5928 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
5929 * affinity_strict RW bool : worker CPU affinity is strict
5932 struct workqueue_struct *wq;
5936 static struct workqueue_struct *dev_to_wq(struct device *dev)
5938 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5943 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5946 struct workqueue_struct *wq = dev_to_wq(dev);
5948 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5950 static DEVICE_ATTR_RO(per_cpu);
5952 static ssize_t max_active_show(struct device *dev,
5953 struct device_attribute *attr, char *buf)
5955 struct workqueue_struct *wq = dev_to_wq(dev);
5957 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5960 static ssize_t max_active_store(struct device *dev,
5961 struct device_attribute *attr, const char *buf,
5964 struct workqueue_struct *wq = dev_to_wq(dev);
5967 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5970 workqueue_set_max_active(wq, val);
5973 static DEVICE_ATTR_RW(max_active);
5975 static struct attribute *wq_sysfs_attrs[] = {
5976 &dev_attr_per_cpu.attr,
5977 &dev_attr_max_active.attr,
5980 ATTRIBUTE_GROUPS(wq_sysfs);
5982 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5985 struct workqueue_struct *wq = dev_to_wq(dev);
5988 mutex_lock(&wq->mutex);
5989 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5990 mutex_unlock(&wq->mutex);
5995 /* prepare workqueue_attrs for sysfs store operations */
5996 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5998 struct workqueue_attrs *attrs;
6000 lockdep_assert_held(&wq_pool_mutex);
6002 attrs = alloc_workqueue_attrs();
6006 copy_workqueue_attrs(attrs, wq->unbound_attrs);
6010 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
6011 const char *buf, size_t count)
6013 struct workqueue_struct *wq = dev_to_wq(dev);
6014 struct workqueue_attrs *attrs;
6017 apply_wqattrs_lock();
6019 attrs = wq_sysfs_prep_attrs(wq);
6023 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
6024 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
6025 ret = apply_workqueue_attrs_locked(wq, attrs);
6030 apply_wqattrs_unlock();
6031 free_workqueue_attrs(attrs);
6032 return ret ?: count;
6035 static ssize_t wq_cpumask_show(struct device *dev,
6036 struct device_attribute *attr, char *buf)
6038 struct workqueue_struct *wq = dev_to_wq(dev);
6041 mutex_lock(&wq->mutex);
6042 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6043 cpumask_pr_args(wq->unbound_attrs->cpumask));
6044 mutex_unlock(&wq->mutex);
6048 static ssize_t wq_cpumask_store(struct device *dev,
6049 struct device_attribute *attr,
6050 const char *buf, size_t count)
6052 struct workqueue_struct *wq = dev_to_wq(dev);
6053 struct workqueue_attrs *attrs;
6056 apply_wqattrs_lock();
6058 attrs = wq_sysfs_prep_attrs(wq);
6062 ret = cpumask_parse(buf, attrs->cpumask);
6064 ret = apply_workqueue_attrs_locked(wq, attrs);
6067 apply_wqattrs_unlock();
6068 free_workqueue_attrs(attrs);
6069 return ret ?: count;
6072 static ssize_t wq_affn_scope_show(struct device *dev,
6073 struct device_attribute *attr, char *buf)
6075 struct workqueue_struct *wq = dev_to_wq(dev);
6078 mutex_lock(&wq->mutex);
6079 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
6080 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
6081 wq_affn_names[WQ_AFFN_DFL],
6082 wq_affn_names[wq_affn_dfl]);
6084 written = scnprintf(buf, PAGE_SIZE, "%s\n",
6085 wq_affn_names[wq->unbound_attrs->affn_scope]);
6086 mutex_unlock(&wq->mutex);
6091 static ssize_t wq_affn_scope_store(struct device *dev,
6092 struct device_attribute *attr,
6093 const char *buf, size_t count)
6095 struct workqueue_struct *wq = dev_to_wq(dev);
6096 struct workqueue_attrs *attrs;
6097 int affn, ret = -ENOMEM;
6099 affn = parse_affn_scope(buf);
6103 apply_wqattrs_lock();
6104 attrs = wq_sysfs_prep_attrs(wq);
6106 attrs->affn_scope = affn;
6107 ret = apply_workqueue_attrs_locked(wq, attrs);
6109 apply_wqattrs_unlock();
6110 free_workqueue_attrs(attrs);
6111 return ret ?: count;
6114 static ssize_t wq_affinity_strict_show(struct device *dev,
6115 struct device_attribute *attr, char *buf)
6117 struct workqueue_struct *wq = dev_to_wq(dev);
6119 return scnprintf(buf, PAGE_SIZE, "%d\n",
6120 wq->unbound_attrs->affn_strict);
6123 static ssize_t wq_affinity_strict_store(struct device *dev,
6124 struct device_attribute *attr,
6125 const char *buf, size_t count)
6127 struct workqueue_struct *wq = dev_to_wq(dev);
6128 struct workqueue_attrs *attrs;
6129 int v, ret = -ENOMEM;
6131 if (sscanf(buf, "%d", &v) != 1)
6134 apply_wqattrs_lock();
6135 attrs = wq_sysfs_prep_attrs(wq);
6137 attrs->affn_strict = (bool)v;
6138 ret = apply_workqueue_attrs_locked(wq, attrs);
6140 apply_wqattrs_unlock();
6141 free_workqueue_attrs(attrs);
6142 return ret ?: count;
6145 static struct device_attribute wq_sysfs_unbound_attrs[] = {
6146 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
6147 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
6148 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
6149 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
6153 static struct bus_type wq_subsys = {
6154 .name = "workqueue",
6155 .dev_groups = wq_sysfs_groups,
6158 static ssize_t wq_unbound_cpumask_show(struct device *dev,
6159 struct device_attribute *attr, char *buf)
6163 mutex_lock(&wq_pool_mutex);
6164 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6165 cpumask_pr_args(wq_unbound_cpumask));
6166 mutex_unlock(&wq_pool_mutex);
6171 static ssize_t wq_unbound_cpumask_store(struct device *dev,
6172 struct device_attribute *attr, const char *buf, size_t count)
6174 cpumask_var_t cpumask;
6177 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6180 ret = cpumask_parse(buf, cpumask);
6182 ret = workqueue_set_unbound_cpumask(cpumask);
6184 free_cpumask_var(cpumask);
6185 return ret ? ret : count;
6188 static struct device_attribute wq_sysfs_cpumask_attr =
6189 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
6190 wq_unbound_cpumask_store);
6192 static int __init wq_sysfs_init(void)
6194 struct device *dev_root;
6197 err = subsys_virtual_register(&wq_subsys, NULL);
6201 dev_root = bus_get_dev_root(&wq_subsys);
6203 err = device_create_file(dev_root, &wq_sysfs_cpumask_attr);
6204 put_device(dev_root);
6208 core_initcall(wq_sysfs_init);
6210 static void wq_device_release(struct device *dev)
6212 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6218 * workqueue_sysfs_register - make a workqueue visible in sysfs
6219 * @wq: the workqueue to register
6221 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
6222 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
6223 * which is the preferred method.
6225 * Workqueue user should use this function directly iff it wants to apply
6226 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
6227 * apply_workqueue_attrs() may race against userland updating the
6230 * Return: 0 on success, -errno on failure.
6232 int workqueue_sysfs_register(struct workqueue_struct *wq)
6234 struct wq_device *wq_dev;
6238 * Adjusting max_active or creating new pwqs by applying
6239 * attributes breaks ordering guarantee. Disallow exposing ordered
6242 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
6245 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
6250 wq_dev->dev.bus = &wq_subsys;
6251 wq_dev->dev.release = wq_device_release;
6252 dev_set_name(&wq_dev->dev, "%s", wq->name);
6255 * unbound_attrs are created separately. Suppress uevent until
6256 * everything is ready.
6258 dev_set_uevent_suppress(&wq_dev->dev, true);
6260 ret = device_register(&wq_dev->dev);
6262 put_device(&wq_dev->dev);
6267 if (wq->flags & WQ_UNBOUND) {
6268 struct device_attribute *attr;
6270 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
6271 ret = device_create_file(&wq_dev->dev, attr);
6273 device_unregister(&wq_dev->dev);
6280 dev_set_uevent_suppress(&wq_dev->dev, false);
6281 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
6286 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
6287 * @wq: the workqueue to unregister
6289 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
6291 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
6293 struct wq_device *wq_dev = wq->wq_dev;
6299 device_unregister(&wq_dev->dev);
6301 #else /* CONFIG_SYSFS */
6302 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
6303 #endif /* CONFIG_SYSFS */
6306 * Workqueue watchdog.
6308 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
6309 * flush dependency, a concurrency managed work item which stays RUNNING
6310 * indefinitely. Workqueue stalls can be very difficult to debug as the
6311 * usual warning mechanisms don't trigger and internal workqueue state is
6314 * Workqueue watchdog monitors all worker pools periodically and dumps
6315 * state if some pools failed to make forward progress for a while where
6316 * forward progress is defined as the first item on ->worklist changing.
6318 * This mechanism is controlled through the kernel parameter
6319 * "workqueue.watchdog_thresh" which can be updated at runtime through the
6320 * corresponding sysfs parameter file.
6322 #ifdef CONFIG_WQ_WATCHDOG
6324 static unsigned long wq_watchdog_thresh = 30;
6325 static struct timer_list wq_watchdog_timer;
6327 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
6328 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
6331 * Show workers that might prevent the processing of pending work items.
6332 * The only candidates are CPU-bound workers in the running state.
6333 * Pending work items should be handled by another idle worker
6334 * in all other situations.
6336 static void show_cpu_pool_hog(struct worker_pool *pool)
6338 struct worker *worker;
6339 unsigned long flags;
6342 raw_spin_lock_irqsave(&pool->lock, flags);
6344 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6345 if (task_is_running(worker->task)) {
6347 * Defer printing to avoid deadlocks in console
6348 * drivers that queue work while holding locks
6349 * also taken in their write paths.
6351 printk_deferred_enter();
6353 pr_info("pool %d:\n", pool->id);
6354 sched_show_task(worker->task);
6356 printk_deferred_exit();
6360 raw_spin_unlock_irqrestore(&pool->lock, flags);
6363 static void show_cpu_pools_hogs(void)
6365 struct worker_pool *pool;
6368 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
6372 for_each_pool(pool, pi) {
6373 if (pool->cpu_stall)
6374 show_cpu_pool_hog(pool);
6381 static void wq_watchdog_reset_touched(void)
6385 wq_watchdog_touched = jiffies;
6386 for_each_possible_cpu(cpu)
6387 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
6390 static void wq_watchdog_timer_fn(struct timer_list *unused)
6392 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
6393 bool lockup_detected = false;
6394 bool cpu_pool_stall = false;
6395 unsigned long now = jiffies;
6396 struct worker_pool *pool;
6404 for_each_pool(pool, pi) {
6405 unsigned long pool_ts, touched, ts;
6407 pool->cpu_stall = false;
6408 if (list_empty(&pool->worklist))
6412 * If a virtual machine is stopped by the host it can look to
6413 * the watchdog like a stall.
6415 kvm_check_and_clear_guest_paused();
6417 /* get the latest of pool and touched timestamps */
6419 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
6421 touched = READ_ONCE(wq_watchdog_touched);
6422 pool_ts = READ_ONCE(pool->watchdog_ts);
6424 if (time_after(pool_ts, touched))
6430 if (time_after(now, ts + thresh)) {
6431 lockup_detected = true;
6432 if (pool->cpu >= 0) {
6433 pool->cpu_stall = true;
6434 cpu_pool_stall = true;
6436 pr_emerg("BUG: workqueue lockup - pool");
6437 pr_cont_pool_info(pool);
6438 pr_cont(" stuck for %us!\n",
6439 jiffies_to_msecs(now - pool_ts) / 1000);
6447 if (lockup_detected)
6448 show_all_workqueues();
6451 show_cpu_pools_hogs();
6453 wq_watchdog_reset_touched();
6454 mod_timer(&wq_watchdog_timer, jiffies + thresh);
6457 notrace void wq_watchdog_touch(int cpu)
6460 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
6462 wq_watchdog_touched = jiffies;
6465 static void wq_watchdog_set_thresh(unsigned long thresh)
6467 wq_watchdog_thresh = 0;
6468 del_timer_sync(&wq_watchdog_timer);
6471 wq_watchdog_thresh = thresh;
6472 wq_watchdog_reset_touched();
6473 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
6477 static int wq_watchdog_param_set_thresh(const char *val,
6478 const struct kernel_param *kp)
6480 unsigned long thresh;
6483 ret = kstrtoul(val, 0, &thresh);
6488 wq_watchdog_set_thresh(thresh);
6490 wq_watchdog_thresh = thresh;
6495 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
6496 .set = wq_watchdog_param_set_thresh,
6497 .get = param_get_ulong,
6500 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
6503 static void wq_watchdog_init(void)
6505 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
6506 wq_watchdog_set_thresh(wq_watchdog_thresh);
6509 #else /* CONFIG_WQ_WATCHDOG */
6511 static inline void wq_watchdog_init(void) { }
6513 #endif /* CONFIG_WQ_WATCHDOG */
6515 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
6517 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
6518 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
6519 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
6523 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
6527 * workqueue_init_early - early init for workqueue subsystem
6529 * This is the first step of three-staged workqueue subsystem initialization and
6530 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
6531 * up. It sets up all the data structures and system workqueues and allows early
6532 * boot code to create workqueues and queue/cancel work items. Actual work item
6533 * execution starts only after kthreads can be created and scheduled right
6534 * before early initcalls.
6536 void __init workqueue_init_early(void)
6538 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
6539 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
6542 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
6544 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
6545 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
6546 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
6547 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
6548 if (!cpumask_empty(&wq_cmdline_cpumask))
6549 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
6551 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
6553 wq_update_pod_attrs_buf = alloc_workqueue_attrs();
6554 BUG_ON(!wq_update_pod_attrs_buf);
6556 /* initialize WQ_AFFN_SYSTEM pods */
6557 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
6558 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
6559 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
6560 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
6562 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
6565 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
6566 pt->pod_node[0] = NUMA_NO_NODE;
6569 /* initialize CPU pools */
6570 for_each_possible_cpu(cpu) {
6571 struct worker_pool *pool;
6574 for_each_cpu_worker_pool(pool, cpu) {
6575 BUG_ON(init_worker_pool(pool));
6577 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
6578 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
6579 pool->attrs->nice = std_nice[i++];
6580 pool->attrs->affn_strict = true;
6581 pool->node = cpu_to_node(cpu);
6584 mutex_lock(&wq_pool_mutex);
6585 BUG_ON(worker_pool_assign_id(pool));
6586 mutex_unlock(&wq_pool_mutex);
6590 /* create default unbound and ordered wq attrs */
6591 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
6592 struct workqueue_attrs *attrs;
6594 BUG_ON(!(attrs = alloc_workqueue_attrs()));
6595 attrs->nice = std_nice[i];
6596 unbound_std_wq_attrs[i] = attrs;
6599 * An ordered wq should have only one pwq as ordering is
6600 * guaranteed by max_active which is enforced by pwqs.
6602 BUG_ON(!(attrs = alloc_workqueue_attrs()));
6603 attrs->nice = std_nice[i];
6604 attrs->ordered = true;
6605 ordered_wq_attrs[i] = attrs;
6608 system_wq = alloc_workqueue("events", 0, 0);
6609 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
6610 system_long_wq = alloc_workqueue("events_long", 0, 0);
6611 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
6613 system_freezable_wq = alloc_workqueue("events_freezable",
6615 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
6616 WQ_POWER_EFFICIENT, 0);
6617 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
6618 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
6620 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
6621 !system_unbound_wq || !system_freezable_wq ||
6622 !system_power_efficient_wq ||
6623 !system_freezable_power_efficient_wq);
6626 static void __init wq_cpu_intensive_thresh_init(void)
6628 unsigned long thresh;
6631 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
6632 BUG_ON(IS_ERR(pwq_release_worker));
6634 /* if the user set it to a specific value, keep it */
6635 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
6639 * The default of 10ms is derived from the fact that most modern (as of
6640 * 2023) processors can do a lot in 10ms and that it's just below what
6641 * most consider human-perceivable. However, the kernel also runs on a
6642 * lot slower CPUs including microcontrollers where the threshold is way
6645 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
6646 * This is by no means accurate but it doesn't have to be. The mechanism
6647 * is still useful even when the threshold is fully scaled up. Also, as
6648 * the reports would usually be applicable to everyone, some machines
6649 * operating on longer thresholds won't significantly diminish their
6652 thresh = 10 * USEC_PER_MSEC;
6654 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
6655 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
6657 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
6659 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
6660 loops_per_jiffy, bogo, thresh);
6662 wq_cpu_intensive_thresh_us = thresh;
6666 * workqueue_init - bring workqueue subsystem fully online
6668 * This is the second step of three-staged workqueue subsystem initialization
6669 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
6670 * been created and work items queued on them, but there are no kworkers
6671 * executing the work items yet. Populate the worker pools with the initial
6672 * workers and enable future kworker creations.
6674 void __init workqueue_init(void)
6676 struct workqueue_struct *wq;
6677 struct worker_pool *pool;
6680 wq_cpu_intensive_thresh_init();
6682 mutex_lock(&wq_pool_mutex);
6685 * Per-cpu pools created earlier could be missing node hint. Fix them
6686 * up. Also, create a rescuer for workqueues that requested it.
6688 for_each_possible_cpu(cpu) {
6689 for_each_cpu_worker_pool(pool, cpu) {
6690 pool->node = cpu_to_node(cpu);
6694 list_for_each_entry(wq, &workqueues, list) {
6695 WARN(init_rescuer(wq),
6696 "workqueue: failed to create early rescuer for %s",
6700 mutex_unlock(&wq_pool_mutex);
6702 /* create the initial workers */
6703 for_each_online_cpu(cpu) {
6704 for_each_cpu_worker_pool(pool, cpu) {
6705 pool->flags &= ~POOL_DISASSOCIATED;
6706 BUG_ON(!create_worker(pool));
6710 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6711 BUG_ON(!create_worker(pool));
6718 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
6719 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
6720 * and consecutive pod ID. The rest of @pt is initialized accordingly.
6722 static void __init init_pod_type(struct wq_pod_type *pt,
6723 bool (*cpus_share_pod)(int, int))
6725 int cur, pre, cpu, pod;
6729 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
6730 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
6731 BUG_ON(!pt->cpu_pod);
6733 for_each_possible_cpu(cur) {
6734 for_each_possible_cpu(pre) {
6736 pt->cpu_pod[cur] = pt->nr_pods++;
6739 if (cpus_share_pod(cur, pre)) {
6740 pt->cpu_pod[cur] = pt->cpu_pod[pre];
6746 /* init the rest to match @pt->cpu_pod[] */
6747 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
6748 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
6749 BUG_ON(!pt->pod_cpus || !pt->pod_node);
6751 for (pod = 0; pod < pt->nr_pods; pod++)
6752 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
6754 for_each_possible_cpu(cpu) {
6755 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
6756 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
6760 static bool __init cpus_dont_share(int cpu0, int cpu1)
6765 static bool __init cpus_share_smt(int cpu0, int cpu1)
6767 #ifdef CONFIG_SCHED_SMT
6768 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
6774 static bool __init cpus_share_numa(int cpu0, int cpu1)
6776 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
6780 * workqueue_init_topology - initialize CPU pods for unbound workqueues
6782 * This is the third step of there-staged workqueue subsystem initialization and
6783 * invoked after SMP and topology information are fully initialized. It
6784 * initializes the unbound CPU pods accordingly.
6786 void __init workqueue_init_topology(void)
6788 struct workqueue_struct *wq;
6791 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
6792 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
6793 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
6794 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
6796 mutex_lock(&wq_pool_mutex);
6799 * Workqueues allocated earlier would have all CPUs sharing the default
6800 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
6801 * combinations to apply per-pod sharing.
6803 list_for_each_entry(wq, &workqueues, list) {
6804 for_each_online_cpu(cpu) {
6805 wq_update_pod(wq, cpu, cpu, true);
6809 mutex_unlock(&wq_pool_mutex);
6812 void __warn_flushing_systemwide_wq(void)
6814 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
6817 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
6819 static int __init workqueue_unbound_cpus_setup(char *str)
6821 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
6822 cpumask_clear(&wq_cmdline_cpumask);
6823 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
6828 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);