2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.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/nmi.h>
52 #include <linux/kvm_para.h>
54 #include "workqueue_internal.h"
60 * A bound pool is either associated or disassociated with its CPU.
61 * While associated (!DISASSOCIATED), all workers are bound to the
62 * CPU and none has %WORKER_UNBOUND set and concurrency management
65 * While DISASSOCIATED, the cpu may be offline and all workers have
66 * %WORKER_UNBOUND set and concurrency management disabled, and may
67 * be executing on any CPU. The pool behaves as an unbound one.
69 * Note that DISASSOCIATED should be flipped only while holding
70 * attach_mutex to avoid changing binding state while
71 * worker_attach_to_pool() is in progress.
73 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
74 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
77 WORKER_DIE = 1 << 1, /* die die die */
78 WORKER_IDLE = 1 << 2, /* is idle */
79 WORKER_PREP = 1 << 3, /* preparing to run works */
80 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
81 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
82 WORKER_REBOUND = 1 << 8, /* worker was rebound */
84 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
85 WORKER_UNBOUND | WORKER_REBOUND,
87 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
89 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
90 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
92 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
95 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
96 /* call for help after 10ms
98 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
99 CREATE_COOLDOWN = HZ, /* time to breath after fail */
102 * Rescue workers are used only on emergencies and shared by
103 * all cpus. Give MIN_NICE.
105 RESCUER_NICE_LEVEL = MIN_NICE,
106 HIGHPRI_NICE_LEVEL = MIN_NICE,
112 * Structure fields follow one of the following exclusion rules.
114 * I: Modifiable by initialization/destruction paths and read-only for
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
120 * L: pool->lock protected. Access with pool->lock held.
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
127 * A: pool->attach_mutex protected.
129 * PL: wq_pool_mutex protected.
131 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
133 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
135 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
136 * sched-RCU for reads.
138 * WQ: wq->mutex protected.
140 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
142 * MD: wq_mayday_lock protected.
145 /* struct worker is defined in workqueue_internal.h */
148 spinlock_t lock; /* the pool lock */
149 int cpu; /* I: the associated cpu */
150 int node; /* I: the associated node ID */
151 int id; /* I: pool ID */
152 unsigned int flags; /* X: flags */
154 unsigned long watchdog_ts; /* L: watchdog timestamp */
156 struct list_head worklist; /* L: list of pending works */
157 int nr_workers; /* L: total number of workers */
159 /* nr_idle includes the ones off idle_list for rebinding */
160 int nr_idle; /* L: currently idle ones */
162 struct list_head idle_list; /* X: list of idle workers */
163 struct timer_list idle_timer; /* L: worker idle timeout */
164 struct timer_list mayday_timer; /* L: SOS timer for workers */
166 /* a workers is either on busy_hash or idle_list, or the manager */
167 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
168 /* L: hash of busy workers */
170 /* see manage_workers() for details on the two manager mutexes */
171 struct worker *manager; /* L: purely informational */
172 struct mutex attach_mutex; /* attach/detach exclusion */
173 struct list_head workers; /* A: attached workers */
174 struct completion *detach_completion; /* all workers detached */
176 struct ida worker_ida; /* worker IDs for task name */
178 struct workqueue_attrs *attrs; /* I: worker attributes */
179 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
180 int refcnt; /* PL: refcnt for unbound pools */
183 * The current concurrency level. As it's likely to be accessed
184 * from other CPUs during try_to_wake_up(), put it in a separate
187 atomic_t nr_running ____cacheline_aligned_in_smp;
190 * Destruction of pool is sched-RCU protected to allow dereferences
191 * from get_work_pool().
194 } ____cacheline_aligned_in_smp;
197 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
198 * of work_struct->data are used for flags and the remaining high bits
199 * point to the pwq; thus, pwqs need to be aligned at two's power of the
200 * number of flag bits.
202 struct pool_workqueue {
203 struct worker_pool *pool; /* I: the associated pool */
204 struct workqueue_struct *wq; /* I: the owning workqueue */
205 int work_color; /* L: current color */
206 int flush_color; /* L: flushing color */
207 int refcnt; /* L: reference count */
208 int nr_in_flight[WORK_NR_COLORS];
209 /* L: nr of in_flight works */
210 int nr_active; /* L: nr of active works */
211 int max_active; /* L: max active works */
212 struct list_head delayed_works; /* L: delayed works */
213 struct list_head pwqs_node; /* WR: node on wq->pwqs */
214 struct list_head mayday_node; /* MD: node on wq->maydays */
217 * Release of unbound pwq is punted to system_wq. See put_pwq()
218 * and pwq_unbound_release_workfn() for details. pool_workqueue
219 * itself is also sched-RCU protected so that the first pwq can be
220 * determined without grabbing wq->mutex.
222 struct work_struct unbound_release_work;
224 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
227 * Structure used to wait for workqueue flush.
230 struct list_head list; /* WQ: list of flushers */
231 int flush_color; /* WQ: flush color waiting for */
232 struct completion done; /* flush completion */
238 * The externally visible workqueue. It relays the issued work items to
239 * the appropriate worker_pool through its pool_workqueues.
241 struct workqueue_struct {
242 struct list_head pwqs; /* WR: all pwqs of this wq */
243 struct list_head list; /* PR: list of all workqueues */
245 struct mutex mutex; /* protects this wq */
246 int work_color; /* WQ: current work color */
247 int flush_color; /* WQ: current flush color */
248 atomic_t nr_pwqs_to_flush; /* flush in progress */
249 struct wq_flusher *first_flusher; /* WQ: first flusher */
250 struct list_head flusher_queue; /* WQ: flush waiters */
251 struct list_head flusher_overflow; /* WQ: flush overflow list */
253 struct list_head maydays; /* MD: pwqs requesting rescue */
254 struct worker *rescuer; /* I: rescue worker */
256 int nr_drainers; /* WQ: drain in progress */
257 int saved_max_active; /* WQ: saved pwq max_active */
259 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
260 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
263 struct wq_device *wq_dev; /* I: for sysfs interface */
265 #ifdef CONFIG_LOCKDEP
266 struct lockdep_map lockdep_map;
268 char name[WQ_NAME_LEN]; /* I: workqueue name */
271 * Destruction of workqueue_struct is sched-RCU protected to allow
272 * walking the workqueues list without grabbing wq_pool_mutex.
273 * This is used to dump all workqueues from sysrq.
277 /* hot fields used during command issue, aligned to cacheline */
278 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
279 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
280 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
283 static struct kmem_cache *pwq_cache;
285 static cpumask_var_t *wq_numa_possible_cpumask;
286 /* possible CPUs of each node */
288 static bool wq_disable_numa;
289 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
291 /* see the comment above the definition of WQ_POWER_EFFICIENT */
292 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
293 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
295 bool wq_online; /* can kworkers be created yet? */
297 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
299 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
300 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
302 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
303 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
304 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
306 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
307 static bool workqueue_freezing; /* PL: have wqs started freezing? */
309 /* PL: allowable cpus for unbound wqs and work items */
310 static cpumask_var_t wq_unbound_cpumask;
312 /* CPU where unbound work was last round robin scheduled from this CPU */
313 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
316 * Local execution of unbound work items is no longer guaranteed. The
317 * following always forces round-robin CPU selection on unbound work items
318 * to uncover usages which depend on it.
320 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
321 static bool wq_debug_force_rr_cpu = true;
323 static bool wq_debug_force_rr_cpu = false;
325 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
327 /* the per-cpu worker pools */
328 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
330 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
332 /* PL: hash of all unbound pools keyed by pool->attrs */
333 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
335 /* I: attributes used when instantiating standard unbound pools on demand */
336 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
338 /* I: attributes used when instantiating ordered pools on demand */
339 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
341 struct workqueue_struct *system_wq __read_mostly;
342 EXPORT_SYMBOL(system_wq);
343 struct workqueue_struct *system_highpri_wq __read_mostly;
344 EXPORT_SYMBOL_GPL(system_highpri_wq);
345 struct workqueue_struct *system_long_wq __read_mostly;
346 EXPORT_SYMBOL_GPL(system_long_wq);
347 struct workqueue_struct *system_unbound_wq __read_mostly;
348 EXPORT_SYMBOL_GPL(system_unbound_wq);
349 struct workqueue_struct *system_freezable_wq __read_mostly;
350 EXPORT_SYMBOL_GPL(system_freezable_wq);
351 struct workqueue_struct *system_power_efficient_wq __read_mostly;
352 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
353 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
354 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
356 static int worker_thread(void *__worker);
357 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
359 #define CREATE_TRACE_POINTS
360 #include <trace/events/workqueue.h>
362 #define assert_rcu_or_pool_mutex() \
363 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
364 !lockdep_is_held(&wq_pool_mutex), \
365 "sched RCU or wq_pool_mutex should be held")
367 #define assert_rcu_or_wq_mutex(wq) \
368 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
369 !lockdep_is_held(&wq->mutex), \
370 "sched RCU or wq->mutex should be held")
372 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
373 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
374 !lockdep_is_held(&wq->mutex) && \
375 !lockdep_is_held(&wq_pool_mutex), \
376 "sched RCU, wq->mutex or wq_pool_mutex should be held")
378 #define for_each_cpu_worker_pool(pool, cpu) \
379 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
380 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
384 * for_each_pool - iterate through all worker_pools in the system
385 * @pool: iteration cursor
386 * @pi: integer used for iteration
388 * This must be called either with wq_pool_mutex held or sched RCU read
389 * locked. If the pool needs to be used beyond the locking in effect, the
390 * caller is responsible for guaranteeing that the pool stays online.
392 * The if/else clause exists only for the lockdep assertion and can be
395 #define for_each_pool(pool, pi) \
396 idr_for_each_entry(&worker_pool_idr, pool, pi) \
397 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
401 * for_each_pool_worker - iterate through all workers of a worker_pool
402 * @worker: iteration cursor
403 * @pool: worker_pool to iterate workers of
405 * This must be called with @pool->attach_mutex.
407 * The if/else clause exists only for the lockdep assertion and can be
410 #define for_each_pool_worker(worker, pool) \
411 list_for_each_entry((worker), &(pool)->workers, node) \
412 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
416 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
417 * @pwq: iteration cursor
418 * @wq: the target workqueue
420 * This must be called either with wq->mutex held or sched RCU read locked.
421 * If the pwq needs to be used beyond the locking in effect, the caller is
422 * responsible for guaranteeing that the pwq stays online.
424 * The if/else clause exists only for the lockdep assertion and can be
427 #define for_each_pwq(pwq, wq) \
428 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
429 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
432 #ifdef CONFIG_DEBUG_OBJECTS_WORK
434 static struct debug_obj_descr work_debug_descr;
436 static void *work_debug_hint(void *addr)
438 return ((struct work_struct *) addr)->func;
441 static bool work_is_static_object(void *addr)
443 struct work_struct *work = addr;
445 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
449 * fixup_init is called when:
450 * - an active object is initialized
452 static bool work_fixup_init(void *addr, enum debug_obj_state state)
454 struct work_struct *work = addr;
457 case ODEBUG_STATE_ACTIVE:
458 cancel_work_sync(work);
459 debug_object_init(work, &work_debug_descr);
467 * fixup_free is called when:
468 * - an active object is freed
470 static bool work_fixup_free(void *addr, enum debug_obj_state state)
472 struct work_struct *work = addr;
475 case ODEBUG_STATE_ACTIVE:
476 cancel_work_sync(work);
477 debug_object_free(work, &work_debug_descr);
484 static struct debug_obj_descr work_debug_descr = {
485 .name = "work_struct",
486 .debug_hint = work_debug_hint,
487 .is_static_object = work_is_static_object,
488 .fixup_init = work_fixup_init,
489 .fixup_free = work_fixup_free,
492 static inline void debug_work_activate(struct work_struct *work)
494 debug_object_activate(work, &work_debug_descr);
497 static inline void debug_work_deactivate(struct work_struct *work)
499 debug_object_deactivate(work, &work_debug_descr);
502 void __init_work(struct work_struct *work, int onstack)
505 debug_object_init_on_stack(work, &work_debug_descr);
507 debug_object_init(work, &work_debug_descr);
509 EXPORT_SYMBOL_GPL(__init_work);
511 void destroy_work_on_stack(struct work_struct *work)
513 debug_object_free(work, &work_debug_descr);
515 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
517 void destroy_delayed_work_on_stack(struct delayed_work *work)
519 destroy_timer_on_stack(&work->timer);
520 debug_object_free(&work->work, &work_debug_descr);
522 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
525 static inline void debug_work_activate(struct work_struct *work) { }
526 static inline void debug_work_deactivate(struct work_struct *work) { }
530 * worker_pool_assign_id - allocate ID and assing it to @pool
531 * @pool: the pool pointer of interest
533 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
534 * successfully, -errno on failure.
536 static int worker_pool_assign_id(struct worker_pool *pool)
540 lockdep_assert_held(&wq_pool_mutex);
542 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
552 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
553 * @wq: the target workqueue
556 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
558 * If the pwq needs to be used beyond the locking in effect, the caller is
559 * responsible for guaranteeing that the pwq stays online.
561 * Return: The unbound pool_workqueue for @node.
563 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
566 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
569 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
570 * delayed item is pending. The plan is to keep CPU -> NODE
571 * mapping valid and stable across CPU on/offlines. Once that
572 * happens, this workaround can be removed.
574 if (unlikely(node == NUMA_NO_NODE))
577 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
580 static unsigned int work_color_to_flags(int color)
582 return color << WORK_STRUCT_COLOR_SHIFT;
585 static int get_work_color(struct work_struct *work)
587 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
588 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
591 static int work_next_color(int color)
593 return (color + 1) % WORK_NR_COLORS;
597 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
598 * contain the pointer to the queued pwq. Once execution starts, the flag
599 * is cleared and the high bits contain OFFQ flags and pool ID.
601 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
602 * and clear_work_data() can be used to set the pwq, pool or clear
603 * work->data. These functions should only be called while the work is
604 * owned - ie. while the PENDING bit is set.
606 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
607 * corresponding to a work. Pool is available once the work has been
608 * queued anywhere after initialization until it is sync canceled. pwq is
609 * available only while the work item is queued.
611 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
612 * canceled. While being canceled, a work item may have its PENDING set
613 * but stay off timer and worklist for arbitrarily long and nobody should
614 * try to steal the PENDING bit.
616 static inline void set_work_data(struct work_struct *work, unsigned long data,
619 WARN_ON_ONCE(!work_pending(work));
620 atomic_long_set(&work->data, data | flags | work_static(work));
623 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
624 unsigned long extra_flags)
626 set_work_data(work, (unsigned long)pwq,
627 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
630 static void set_work_pool_and_keep_pending(struct work_struct *work,
633 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
634 WORK_STRUCT_PENDING);
637 static void set_work_pool_and_clear_pending(struct work_struct *work,
641 * The following wmb is paired with the implied mb in
642 * test_and_set_bit(PENDING) and ensures all updates to @work made
643 * here are visible to and precede any updates by the next PENDING
647 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
649 * The following mb guarantees that previous clear of a PENDING bit
650 * will not be reordered with any speculative LOADS or STORES from
651 * work->current_func, which is executed afterwards. This possible
652 * reordering can lead to a missed execution on attempt to qeueue
653 * the same @work. E.g. consider this case:
656 * ---------------------------- --------------------------------
658 * 1 STORE event_indicated
659 * 2 queue_work_on() {
660 * 3 test_and_set_bit(PENDING)
661 * 4 } set_..._and_clear_pending() {
662 * 5 set_work_data() # clear bit
664 * 7 work->current_func() {
665 * 8 LOAD event_indicated
668 * Without an explicit full barrier speculative LOAD on line 8 can
669 * be executed before CPU#0 does STORE on line 1. If that happens,
670 * CPU#0 observes the PENDING bit is still set and new execution of
671 * a @work is not queued in a hope, that CPU#1 will eventually
672 * finish the queued @work. Meanwhile CPU#1 does not see
673 * event_indicated is set, because speculative LOAD was executed
674 * before actual STORE.
679 static void clear_work_data(struct work_struct *work)
681 smp_wmb(); /* see set_work_pool_and_clear_pending() */
682 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
685 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
687 unsigned long data = atomic_long_read(&work->data);
689 if (data & WORK_STRUCT_PWQ)
690 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
696 * get_work_pool - return the worker_pool a given work was associated with
697 * @work: the work item of interest
699 * Pools are created and destroyed under wq_pool_mutex, and allows read
700 * access under sched-RCU read lock. As such, this function should be
701 * called under wq_pool_mutex or with preemption disabled.
703 * All fields of the returned pool are accessible as long as the above
704 * mentioned locking is in effect. If the returned pool needs to be used
705 * beyond the critical section, the caller is responsible for ensuring the
706 * returned pool is and stays online.
708 * Return: The worker_pool @work was last associated with. %NULL if none.
710 static struct worker_pool *get_work_pool(struct work_struct *work)
712 unsigned long data = atomic_long_read(&work->data);
715 assert_rcu_or_pool_mutex();
717 if (data & WORK_STRUCT_PWQ)
718 return ((struct pool_workqueue *)
719 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
721 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
722 if (pool_id == WORK_OFFQ_POOL_NONE)
725 return idr_find(&worker_pool_idr, pool_id);
729 * get_work_pool_id - return the worker pool ID a given work is associated with
730 * @work: the work item of interest
732 * Return: The worker_pool ID @work was last associated with.
733 * %WORK_OFFQ_POOL_NONE if none.
735 static int get_work_pool_id(struct work_struct *work)
737 unsigned long data = atomic_long_read(&work->data);
739 if (data & WORK_STRUCT_PWQ)
740 return ((struct pool_workqueue *)
741 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
743 return data >> WORK_OFFQ_POOL_SHIFT;
746 static void mark_work_canceling(struct work_struct *work)
748 unsigned long pool_id = get_work_pool_id(work);
750 pool_id <<= WORK_OFFQ_POOL_SHIFT;
751 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
754 static bool work_is_canceling(struct work_struct *work)
756 unsigned long data = atomic_long_read(&work->data);
758 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
762 * Policy functions. These define the policies on how the global worker
763 * pools are managed. Unless noted otherwise, these functions assume that
764 * they're being called with pool->lock held.
767 static bool __need_more_worker(struct worker_pool *pool)
769 return !atomic_read(&pool->nr_running);
773 * Need to wake up a worker? Called from anything but currently
776 * Note that, because unbound workers never contribute to nr_running, this
777 * function will always return %true for unbound pools as long as the
778 * worklist isn't empty.
780 static bool need_more_worker(struct worker_pool *pool)
782 return !list_empty(&pool->worklist) && __need_more_worker(pool);
785 /* Can I start working? Called from busy but !running workers. */
786 static bool may_start_working(struct worker_pool *pool)
788 return pool->nr_idle;
791 /* Do I need to keep working? Called from currently running workers. */
792 static bool keep_working(struct worker_pool *pool)
794 return !list_empty(&pool->worklist) &&
795 atomic_read(&pool->nr_running) <= 1;
798 /* Do we need a new worker? Called from manager. */
799 static bool need_to_create_worker(struct worker_pool *pool)
801 return need_more_worker(pool) && !may_start_working(pool);
804 /* Do we have too many workers and should some go away? */
805 static bool too_many_workers(struct worker_pool *pool)
807 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
808 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
809 int nr_busy = pool->nr_workers - nr_idle;
811 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
818 /* Return the first idle worker. Safe with preemption disabled */
819 static struct worker *first_idle_worker(struct worker_pool *pool)
821 if (unlikely(list_empty(&pool->idle_list)))
824 return list_first_entry(&pool->idle_list, struct worker, entry);
828 * wake_up_worker - wake up an idle worker
829 * @pool: worker pool to wake worker from
831 * Wake up the first idle worker of @pool.
834 * spin_lock_irq(pool->lock).
836 static void wake_up_worker(struct worker_pool *pool)
838 struct worker *worker = first_idle_worker(pool);
841 wake_up_process(worker->task);
845 * wq_worker_waking_up - a worker is waking up
846 * @task: task waking up
847 * @cpu: CPU @task is waking up to
849 * This function is called during try_to_wake_up() when a worker is
853 * spin_lock_irq(rq->lock)
855 void wq_worker_waking_up(struct task_struct *task, int cpu)
857 struct worker *worker = kthread_data(task);
859 if (!(worker->flags & WORKER_NOT_RUNNING)) {
860 WARN_ON_ONCE(worker->pool->cpu != cpu);
861 atomic_inc(&worker->pool->nr_running);
866 * wq_worker_sleeping - a worker is going to sleep
867 * @task: task going to sleep
869 * This function is called during schedule() when a busy worker is
870 * going to sleep. Worker on the same cpu can be woken up by
871 * returning pointer to its task.
874 * spin_lock_irq(rq->lock)
877 * Worker task on @cpu to wake up, %NULL if none.
879 struct task_struct *wq_worker_sleeping(struct task_struct *task)
881 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
882 struct worker_pool *pool;
885 * Rescuers, which may not have all the fields set up like normal
886 * workers, also reach here, let's not access anything before
887 * checking NOT_RUNNING.
889 if (worker->flags & WORKER_NOT_RUNNING)
894 /* this can only happen on the local cpu */
895 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
899 * The counterpart of the following dec_and_test, implied mb,
900 * worklist not empty test sequence is in insert_work().
901 * Please read comment there.
903 * NOT_RUNNING is clear. This means that we're bound to and
904 * running on the local cpu w/ rq lock held and preemption
905 * disabled, which in turn means that none else could be
906 * manipulating idle_list, so dereferencing idle_list without pool
909 if (atomic_dec_and_test(&pool->nr_running) &&
910 !list_empty(&pool->worklist))
911 to_wakeup = first_idle_worker(pool);
912 return to_wakeup ? to_wakeup->task : NULL;
916 * worker_set_flags - set worker flags and adjust nr_running accordingly
918 * @flags: flags to set
920 * Set @flags in @worker->flags and adjust nr_running accordingly.
923 * spin_lock_irq(pool->lock)
925 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
927 struct worker_pool *pool = worker->pool;
929 WARN_ON_ONCE(worker->task != current);
931 /* If transitioning into NOT_RUNNING, adjust nr_running. */
932 if ((flags & WORKER_NOT_RUNNING) &&
933 !(worker->flags & WORKER_NOT_RUNNING)) {
934 atomic_dec(&pool->nr_running);
937 worker->flags |= flags;
941 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
943 * @flags: flags to clear
945 * Clear @flags in @worker->flags and adjust nr_running accordingly.
948 * spin_lock_irq(pool->lock)
950 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
952 struct worker_pool *pool = worker->pool;
953 unsigned int oflags = worker->flags;
955 WARN_ON_ONCE(worker->task != current);
957 worker->flags &= ~flags;
960 * If transitioning out of NOT_RUNNING, increment nr_running. Note
961 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
962 * of multiple flags, not a single flag.
964 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
965 if (!(worker->flags & WORKER_NOT_RUNNING))
966 atomic_inc(&pool->nr_running);
970 * find_worker_executing_work - find worker which is executing a work
971 * @pool: pool of interest
972 * @work: work to find worker for
974 * Find a worker which is executing @work on @pool by searching
975 * @pool->busy_hash which is keyed by the address of @work. For a worker
976 * to match, its current execution should match the address of @work and
977 * its work function. This is to avoid unwanted dependency between
978 * unrelated work executions through a work item being recycled while still
981 * This is a bit tricky. A work item may be freed once its execution
982 * starts and nothing prevents the freed area from being recycled for
983 * another work item. If the same work item address ends up being reused
984 * before the original execution finishes, workqueue will identify the
985 * recycled work item as currently executing and make it wait until the
986 * current execution finishes, introducing an unwanted dependency.
988 * This function checks the work item address and work function to avoid
989 * false positives. Note that this isn't complete as one may construct a
990 * work function which can introduce dependency onto itself through a
991 * recycled work item. Well, if somebody wants to shoot oneself in the
992 * foot that badly, there's only so much we can do, and if such deadlock
993 * actually occurs, it should be easy to locate the culprit work function.
996 * spin_lock_irq(pool->lock).
999 * Pointer to worker which is executing @work if found, %NULL
1002 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1003 struct work_struct *work)
1005 struct worker *worker;
1007 hash_for_each_possible(pool->busy_hash, worker, hentry,
1008 (unsigned long)work)
1009 if (worker->current_work == work &&
1010 worker->current_func == work->func)
1017 * move_linked_works - move linked works to a list
1018 * @work: start of series of works to be scheduled
1019 * @head: target list to append @work to
1020 * @nextp: out parameter for nested worklist walking
1022 * Schedule linked works starting from @work to @head. Work series to
1023 * be scheduled starts at @work and includes any consecutive work with
1024 * WORK_STRUCT_LINKED set in its predecessor.
1026 * If @nextp is not NULL, it's updated to point to the next work of
1027 * the last scheduled work. This allows move_linked_works() to be
1028 * nested inside outer list_for_each_entry_safe().
1031 * spin_lock_irq(pool->lock).
1033 static void move_linked_works(struct work_struct *work, struct list_head *head,
1034 struct work_struct **nextp)
1036 struct work_struct *n;
1039 * Linked worklist will always end before the end of the list,
1040 * use NULL for list head.
1042 list_for_each_entry_safe_from(work, n, NULL, entry) {
1043 list_move_tail(&work->entry, head);
1044 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1049 * If we're already inside safe list traversal and have moved
1050 * multiple works to the scheduled queue, the next position
1051 * needs to be updated.
1058 * get_pwq - get an extra reference on the specified pool_workqueue
1059 * @pwq: pool_workqueue to get
1061 * Obtain an extra reference on @pwq. The caller should guarantee that
1062 * @pwq has positive refcnt and be holding the matching pool->lock.
1064 static void get_pwq(struct pool_workqueue *pwq)
1066 lockdep_assert_held(&pwq->pool->lock);
1067 WARN_ON_ONCE(pwq->refcnt <= 0);
1072 * put_pwq - put a pool_workqueue reference
1073 * @pwq: pool_workqueue to put
1075 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1076 * destruction. The caller should be holding the matching pool->lock.
1078 static void put_pwq(struct pool_workqueue *pwq)
1080 lockdep_assert_held(&pwq->pool->lock);
1081 if (likely(--pwq->refcnt))
1083 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1086 * @pwq can't be released under pool->lock, bounce to
1087 * pwq_unbound_release_workfn(). This never recurses on the same
1088 * pool->lock as this path is taken only for unbound workqueues and
1089 * the release work item is scheduled on a per-cpu workqueue. To
1090 * avoid lockdep warning, unbound pool->locks are given lockdep
1091 * subclass of 1 in get_unbound_pool().
1093 schedule_work(&pwq->unbound_release_work);
1097 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1098 * @pwq: pool_workqueue to put (can be %NULL)
1100 * put_pwq() with locking. This function also allows %NULL @pwq.
1102 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1106 * As both pwqs and pools are sched-RCU protected, the
1107 * following lock operations are safe.
1109 spin_lock_irq(&pwq->pool->lock);
1111 spin_unlock_irq(&pwq->pool->lock);
1115 static void pwq_activate_delayed_work(struct work_struct *work)
1117 struct pool_workqueue *pwq = get_work_pwq(work);
1119 trace_workqueue_activate_work(work);
1120 if (list_empty(&pwq->pool->worklist))
1121 pwq->pool->watchdog_ts = jiffies;
1122 move_linked_works(work, &pwq->pool->worklist, NULL);
1123 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1127 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1129 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1130 struct work_struct, entry);
1132 pwq_activate_delayed_work(work);
1136 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1137 * @pwq: pwq of interest
1138 * @color: color of work which left the queue
1140 * A work either has completed or is removed from pending queue,
1141 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1144 * spin_lock_irq(pool->lock).
1146 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1148 /* uncolored work items don't participate in flushing or nr_active */
1149 if (color == WORK_NO_COLOR)
1152 pwq->nr_in_flight[color]--;
1155 if (!list_empty(&pwq->delayed_works)) {
1156 /* one down, submit a delayed one */
1157 if (pwq->nr_active < pwq->max_active)
1158 pwq_activate_first_delayed(pwq);
1161 /* is flush in progress and are we at the flushing tip? */
1162 if (likely(pwq->flush_color != color))
1165 /* are there still in-flight works? */
1166 if (pwq->nr_in_flight[color])
1169 /* this pwq is done, clear flush_color */
1170 pwq->flush_color = -1;
1173 * If this was the last pwq, wake up the first flusher. It
1174 * will handle the rest.
1176 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1177 complete(&pwq->wq->first_flusher->done);
1183 * try_to_grab_pending - steal work item from worklist and disable irq
1184 * @work: work item to steal
1185 * @is_dwork: @work is a delayed_work
1186 * @flags: place to store irq state
1188 * Try to grab PENDING bit of @work. This function can handle @work in any
1189 * stable state - idle, on timer or on worklist.
1192 * 1 if @work was pending and we successfully stole PENDING
1193 * 0 if @work was idle and we claimed PENDING
1194 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1195 * -ENOENT if someone else is canceling @work, this state may persist
1196 * for arbitrarily long
1199 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1200 * interrupted while holding PENDING and @work off queue, irq must be
1201 * disabled on entry. This, combined with delayed_work->timer being
1202 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1204 * On successful return, >= 0, irq is disabled and the caller is
1205 * responsible for releasing it using local_irq_restore(*@flags).
1207 * This function is safe to call from any context including IRQ handler.
1209 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1210 unsigned long *flags)
1212 struct worker_pool *pool;
1213 struct pool_workqueue *pwq;
1215 local_irq_save(*flags);
1217 /* try to steal the timer if it exists */
1219 struct delayed_work *dwork = to_delayed_work(work);
1222 * dwork->timer is irqsafe. If del_timer() fails, it's
1223 * guaranteed that the timer is not queued anywhere and not
1224 * running on the local CPU.
1226 if (likely(del_timer(&dwork->timer)))
1230 /* try to claim PENDING the normal way */
1231 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1235 * The queueing is in progress, or it is already queued. Try to
1236 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1238 pool = get_work_pool(work);
1242 spin_lock(&pool->lock);
1244 * work->data is guaranteed to point to pwq only while the work
1245 * item is queued on pwq->wq, and both updating work->data to point
1246 * to pwq on queueing and to pool on dequeueing are done under
1247 * pwq->pool->lock. This in turn guarantees that, if work->data
1248 * points to pwq which is associated with a locked pool, the work
1249 * item is currently queued on that pool.
1251 pwq = get_work_pwq(work);
1252 if (pwq && pwq->pool == pool) {
1253 debug_work_deactivate(work);
1256 * A delayed work item cannot be grabbed directly because
1257 * it might have linked NO_COLOR work items which, if left
1258 * on the delayed_list, will confuse pwq->nr_active
1259 * management later on and cause stall. Make sure the work
1260 * item is activated before grabbing.
1262 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1263 pwq_activate_delayed_work(work);
1265 list_del_init(&work->entry);
1266 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1268 /* work->data points to pwq iff queued, point to pool */
1269 set_work_pool_and_keep_pending(work, pool->id);
1271 spin_unlock(&pool->lock);
1274 spin_unlock(&pool->lock);
1276 local_irq_restore(*flags);
1277 if (work_is_canceling(work))
1284 * insert_work - insert a work into a pool
1285 * @pwq: pwq @work belongs to
1286 * @work: work to insert
1287 * @head: insertion point
1288 * @extra_flags: extra WORK_STRUCT_* flags to set
1290 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1291 * work_struct flags.
1294 * spin_lock_irq(pool->lock).
1296 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1297 struct list_head *head, unsigned int extra_flags)
1299 struct worker_pool *pool = pwq->pool;
1301 /* we own @work, set data and link */
1302 set_work_pwq(work, pwq, extra_flags);
1303 list_add_tail(&work->entry, head);
1307 * Ensure either wq_worker_sleeping() sees the above
1308 * list_add_tail() or we see zero nr_running to avoid workers lying
1309 * around lazily while there are works to be processed.
1313 if (__need_more_worker(pool))
1314 wake_up_worker(pool);
1318 * Test whether @work is being queued from another work executing on the
1321 static bool is_chained_work(struct workqueue_struct *wq)
1323 struct worker *worker;
1325 worker = current_wq_worker();
1327 * Return %true iff I'm a worker execuing a work item on @wq. If
1328 * I'm @worker, it's safe to dereference it without locking.
1330 return worker && worker->current_pwq->wq == wq;
1334 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1335 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1336 * avoid perturbing sensitive tasks.
1338 static int wq_select_unbound_cpu(int cpu)
1340 static bool printed_dbg_warning;
1343 if (likely(!wq_debug_force_rr_cpu)) {
1344 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1346 } else if (!printed_dbg_warning) {
1347 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1348 printed_dbg_warning = true;
1351 if (cpumask_empty(wq_unbound_cpumask))
1354 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1355 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1356 if (unlikely(new_cpu >= nr_cpu_ids)) {
1357 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1358 if (unlikely(new_cpu >= nr_cpu_ids))
1361 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1366 static void __queue_work(int cpu, struct workqueue_struct *wq,
1367 struct work_struct *work)
1369 struct pool_workqueue *pwq;
1370 struct worker_pool *last_pool;
1371 struct list_head *worklist;
1372 unsigned int work_flags;
1373 unsigned int req_cpu = cpu;
1376 * While a work item is PENDING && off queue, a task trying to
1377 * steal the PENDING will busy-loop waiting for it to either get
1378 * queued or lose PENDING. Grabbing PENDING and queueing should
1379 * happen with IRQ disabled.
1381 WARN_ON_ONCE(!irqs_disabled());
1384 /* if draining, only works from the same workqueue are allowed */
1385 if (unlikely(wq->flags & __WQ_DRAINING) &&
1386 WARN_ON_ONCE(!is_chained_work(wq)))
1389 /* pwq which will be used unless @work is executing elsewhere */
1390 if (wq->flags & WQ_UNBOUND) {
1391 if (req_cpu == WORK_CPU_UNBOUND)
1392 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1393 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1395 if (req_cpu == WORK_CPU_UNBOUND)
1396 cpu = raw_smp_processor_id();
1397 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1401 * If @work was previously on a different pool, it might still be
1402 * running there, in which case the work needs to be queued on that
1403 * pool to guarantee non-reentrancy.
1405 last_pool = get_work_pool(work);
1406 if (last_pool && last_pool != pwq->pool) {
1407 struct worker *worker;
1409 spin_lock(&last_pool->lock);
1411 worker = find_worker_executing_work(last_pool, work);
1413 if (worker && worker->current_pwq->wq == wq) {
1414 pwq = worker->current_pwq;
1416 /* meh... not running there, queue here */
1417 spin_unlock(&last_pool->lock);
1418 spin_lock(&pwq->pool->lock);
1421 spin_lock(&pwq->pool->lock);
1425 * pwq is determined and locked. For unbound pools, we could have
1426 * raced with pwq release and it could already be dead. If its
1427 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1428 * without another pwq replacing it in the numa_pwq_tbl or while
1429 * work items are executing on it, so the retrying is guaranteed to
1430 * make forward-progress.
1432 if (unlikely(!pwq->refcnt)) {
1433 if (wq->flags & WQ_UNBOUND) {
1434 spin_unlock(&pwq->pool->lock);
1439 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1443 /* pwq determined, queue */
1444 trace_workqueue_queue_work(req_cpu, pwq, work);
1446 if (WARN_ON(!list_empty(&work->entry))) {
1447 spin_unlock(&pwq->pool->lock);
1451 pwq->nr_in_flight[pwq->work_color]++;
1452 work_flags = work_color_to_flags(pwq->work_color);
1454 if (likely(pwq->nr_active < pwq->max_active)) {
1455 trace_workqueue_activate_work(work);
1457 worklist = &pwq->pool->worklist;
1458 if (list_empty(worklist))
1459 pwq->pool->watchdog_ts = jiffies;
1461 work_flags |= WORK_STRUCT_DELAYED;
1462 worklist = &pwq->delayed_works;
1465 debug_work_activate(work);
1466 insert_work(pwq, work, worklist, work_flags);
1468 spin_unlock(&pwq->pool->lock);
1472 * queue_work_on - queue work on specific cpu
1473 * @cpu: CPU number to execute work on
1474 * @wq: workqueue to use
1475 * @work: work to queue
1477 * We queue the work to a specific CPU, the caller must ensure it
1480 * Return: %false if @work was already on a queue, %true otherwise.
1482 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1483 struct work_struct *work)
1486 unsigned long flags;
1488 local_irq_save(flags);
1490 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1491 __queue_work(cpu, wq, work);
1495 local_irq_restore(flags);
1498 EXPORT_SYMBOL(queue_work_on);
1500 void delayed_work_timer_fn(unsigned long __data)
1502 struct delayed_work *dwork = (struct delayed_work *)__data;
1504 /* should have been called from irqsafe timer with irq already off */
1505 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1507 EXPORT_SYMBOL(delayed_work_timer_fn);
1509 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1510 struct delayed_work *dwork, unsigned long delay)
1512 struct timer_list *timer = &dwork->timer;
1513 struct work_struct *work = &dwork->work;
1516 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1517 timer->data != (unsigned long)dwork);
1518 WARN_ON_ONCE(timer_pending(timer));
1519 WARN_ON_ONCE(!list_empty(&work->entry));
1522 * If @delay is 0, queue @dwork->work immediately. This is for
1523 * both optimization and correctness. The earliest @timer can
1524 * expire is on the closest next tick and delayed_work users depend
1525 * on that there's no such delay when @delay is 0.
1528 __queue_work(cpu, wq, &dwork->work);
1532 timer_stats_timer_set_start_info(&dwork->timer);
1536 timer->expires = jiffies + delay;
1538 if (unlikely(cpu != WORK_CPU_UNBOUND))
1539 add_timer_on(timer, cpu);
1545 * queue_delayed_work_on - queue work on specific CPU after delay
1546 * @cpu: CPU number to execute work on
1547 * @wq: workqueue to use
1548 * @dwork: work to queue
1549 * @delay: number of jiffies to wait before queueing
1551 * Return: %false if @work was already on a queue, %true otherwise. If
1552 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1555 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1556 struct delayed_work *dwork, unsigned long delay)
1558 struct work_struct *work = &dwork->work;
1560 unsigned long flags;
1562 /* read the comment in __queue_work() */
1563 local_irq_save(flags);
1565 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1566 __queue_delayed_work(cpu, wq, dwork, delay);
1570 local_irq_restore(flags);
1573 EXPORT_SYMBOL(queue_delayed_work_on);
1576 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1577 * @cpu: CPU number to execute work on
1578 * @wq: workqueue to use
1579 * @dwork: work to queue
1580 * @delay: number of jiffies to wait before queueing
1582 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1583 * modify @dwork's timer so that it expires after @delay. If @delay is
1584 * zero, @work is guaranteed to be scheduled immediately regardless of its
1587 * Return: %false if @dwork was idle and queued, %true if @dwork was
1588 * pending and its timer was modified.
1590 * This function is safe to call from any context including IRQ handler.
1591 * See try_to_grab_pending() for details.
1593 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1594 struct delayed_work *dwork, unsigned long delay)
1596 unsigned long flags;
1600 ret = try_to_grab_pending(&dwork->work, true, &flags);
1601 } while (unlikely(ret == -EAGAIN));
1603 if (likely(ret >= 0)) {
1604 __queue_delayed_work(cpu, wq, dwork, delay);
1605 local_irq_restore(flags);
1608 /* -ENOENT from try_to_grab_pending() becomes %true */
1611 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1614 * worker_enter_idle - enter idle state
1615 * @worker: worker which is entering idle state
1617 * @worker is entering idle state. Update stats and idle timer if
1621 * spin_lock_irq(pool->lock).
1623 static void worker_enter_idle(struct worker *worker)
1625 struct worker_pool *pool = worker->pool;
1627 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1628 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1629 (worker->hentry.next || worker->hentry.pprev)))
1632 /* can't use worker_set_flags(), also called from create_worker() */
1633 worker->flags |= WORKER_IDLE;
1635 worker->last_active = jiffies;
1637 /* idle_list is LIFO */
1638 list_add(&worker->entry, &pool->idle_list);
1640 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1641 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1644 * Sanity check nr_running. Because wq_unbind_fn() releases
1645 * pool->lock between setting %WORKER_UNBOUND and zapping
1646 * nr_running, the warning may trigger spuriously. Check iff
1647 * unbind is not in progress.
1649 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1650 pool->nr_workers == pool->nr_idle &&
1651 atomic_read(&pool->nr_running));
1655 * worker_leave_idle - leave idle state
1656 * @worker: worker which is leaving idle state
1658 * @worker is leaving idle state. Update stats.
1661 * spin_lock_irq(pool->lock).
1663 static void worker_leave_idle(struct worker *worker)
1665 struct worker_pool *pool = worker->pool;
1667 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1669 worker_clr_flags(worker, WORKER_IDLE);
1671 list_del_init(&worker->entry);
1674 static struct worker *alloc_worker(int node)
1676 struct worker *worker;
1678 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1680 INIT_LIST_HEAD(&worker->entry);
1681 INIT_LIST_HEAD(&worker->scheduled);
1682 INIT_LIST_HEAD(&worker->node);
1683 /* on creation a worker is in !idle && prep state */
1684 worker->flags = WORKER_PREP;
1690 * worker_attach_to_pool() - attach a worker to a pool
1691 * @worker: worker to be attached
1692 * @pool: the target pool
1694 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1695 * cpu-binding of @worker are kept coordinated with the pool across
1698 static void worker_attach_to_pool(struct worker *worker,
1699 struct worker_pool *pool)
1701 mutex_lock(&pool->attach_mutex);
1704 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1705 * online CPUs. It'll be re-applied when any of the CPUs come up.
1707 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1710 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1711 * stable across this function. See the comments above the
1712 * flag definition for details.
1714 if (pool->flags & POOL_DISASSOCIATED)
1715 worker->flags |= WORKER_UNBOUND;
1717 list_add_tail(&worker->node, &pool->workers);
1719 mutex_unlock(&pool->attach_mutex);
1723 * worker_detach_from_pool() - detach a worker from its pool
1724 * @worker: worker which is attached to its pool
1725 * @pool: the pool @worker is attached to
1727 * Undo the attaching which had been done in worker_attach_to_pool(). The
1728 * caller worker shouldn't access to the pool after detached except it has
1729 * other reference to the pool.
1731 static void worker_detach_from_pool(struct worker *worker,
1732 struct worker_pool *pool)
1734 struct completion *detach_completion = NULL;
1736 mutex_lock(&pool->attach_mutex);
1737 list_del(&worker->node);
1738 if (list_empty(&pool->workers))
1739 detach_completion = pool->detach_completion;
1740 mutex_unlock(&pool->attach_mutex);
1742 /* clear leftover flags without pool->lock after it is detached */
1743 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1745 if (detach_completion)
1746 complete(detach_completion);
1750 * create_worker - create a new workqueue worker
1751 * @pool: pool the new worker will belong to
1753 * Create and start a new worker which is attached to @pool.
1756 * Might sleep. Does GFP_KERNEL allocations.
1759 * Pointer to the newly created worker.
1761 static struct worker *create_worker(struct worker_pool *pool)
1763 struct worker *worker = NULL;
1767 /* ID is needed to determine kthread name */
1768 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1772 worker = alloc_worker(pool->node);
1776 worker->pool = pool;
1780 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1781 pool->attrs->nice < 0 ? "H" : "");
1783 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1785 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1786 "kworker/%s", id_buf);
1787 if (IS_ERR(worker->task))
1790 set_user_nice(worker->task, pool->attrs->nice);
1791 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1793 /* successful, attach the worker to the pool */
1794 worker_attach_to_pool(worker, pool);
1796 /* start the newly created worker */
1797 spin_lock_irq(&pool->lock);
1798 worker->pool->nr_workers++;
1799 worker_enter_idle(worker);
1800 wake_up_process(worker->task);
1801 spin_unlock_irq(&pool->lock);
1807 ida_simple_remove(&pool->worker_ida, id);
1813 * destroy_worker - destroy a workqueue worker
1814 * @worker: worker to be destroyed
1816 * Destroy @worker and adjust @pool stats accordingly. The worker should
1820 * spin_lock_irq(pool->lock).
1822 static void destroy_worker(struct worker *worker)
1824 struct worker_pool *pool = worker->pool;
1826 lockdep_assert_held(&pool->lock);
1828 /* sanity check frenzy */
1829 if (WARN_ON(worker->current_work) ||
1830 WARN_ON(!list_empty(&worker->scheduled)) ||
1831 WARN_ON(!(worker->flags & WORKER_IDLE)))
1837 list_del_init(&worker->entry);
1838 worker->flags |= WORKER_DIE;
1839 wake_up_process(worker->task);
1842 static void idle_worker_timeout(unsigned long __pool)
1844 struct worker_pool *pool = (void *)__pool;
1846 spin_lock_irq(&pool->lock);
1848 while (too_many_workers(pool)) {
1849 struct worker *worker;
1850 unsigned long expires;
1852 /* idle_list is kept in LIFO order, check the last one */
1853 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1854 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1856 if (time_before(jiffies, expires)) {
1857 mod_timer(&pool->idle_timer, expires);
1861 destroy_worker(worker);
1864 spin_unlock_irq(&pool->lock);
1867 static void send_mayday(struct work_struct *work)
1869 struct pool_workqueue *pwq = get_work_pwq(work);
1870 struct workqueue_struct *wq = pwq->wq;
1872 lockdep_assert_held(&wq_mayday_lock);
1877 /* mayday mayday mayday */
1878 if (list_empty(&pwq->mayday_node)) {
1880 * If @pwq is for an unbound wq, its base ref may be put at
1881 * any time due to an attribute change. Pin @pwq until the
1882 * rescuer is done with it.
1885 list_add_tail(&pwq->mayday_node, &wq->maydays);
1886 wake_up_process(wq->rescuer->task);
1890 static void pool_mayday_timeout(unsigned long __pool)
1892 struct worker_pool *pool = (void *)__pool;
1893 struct work_struct *work;
1895 spin_lock_irq(&pool->lock);
1896 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1898 if (need_to_create_worker(pool)) {
1900 * We've been trying to create a new worker but
1901 * haven't been successful. We might be hitting an
1902 * allocation deadlock. Send distress signals to
1905 list_for_each_entry(work, &pool->worklist, entry)
1909 spin_unlock(&wq_mayday_lock);
1910 spin_unlock_irq(&pool->lock);
1912 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1916 * maybe_create_worker - create a new worker if necessary
1917 * @pool: pool to create a new worker for
1919 * Create a new worker for @pool if necessary. @pool is guaranteed to
1920 * have at least one idle worker on return from this function. If
1921 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1922 * sent to all rescuers with works scheduled on @pool to resolve
1923 * possible allocation deadlock.
1925 * On return, need_to_create_worker() is guaranteed to be %false and
1926 * may_start_working() %true.
1929 * spin_lock_irq(pool->lock) which may be released and regrabbed
1930 * multiple times. Does GFP_KERNEL allocations. Called only from
1933 static void maybe_create_worker(struct worker_pool *pool)
1934 __releases(&pool->lock)
1935 __acquires(&pool->lock)
1938 spin_unlock_irq(&pool->lock);
1940 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1941 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1944 if (create_worker(pool) || !need_to_create_worker(pool))
1947 schedule_timeout_interruptible(CREATE_COOLDOWN);
1949 if (!need_to_create_worker(pool))
1953 del_timer_sync(&pool->mayday_timer);
1954 spin_lock_irq(&pool->lock);
1956 * This is necessary even after a new worker was just successfully
1957 * created as @pool->lock was dropped and the new worker might have
1958 * already become busy.
1960 if (need_to_create_worker(pool))
1965 * manage_workers - manage worker pool
1968 * Assume the manager role and manage the worker pool @worker belongs
1969 * to. At any given time, there can be only zero or one manager per
1970 * pool. The exclusion is handled automatically by this function.
1972 * The caller can safely start processing works on false return. On
1973 * true return, it's guaranteed that need_to_create_worker() is false
1974 * and may_start_working() is true.
1977 * spin_lock_irq(pool->lock) which may be released and regrabbed
1978 * multiple times. Does GFP_KERNEL allocations.
1981 * %false if the pool doesn't need management and the caller can safely
1982 * start processing works, %true if management function was performed and
1983 * the conditions that the caller verified before calling the function may
1984 * no longer be true.
1986 static bool manage_workers(struct worker *worker)
1988 struct worker_pool *pool = worker->pool;
1990 if (pool->flags & POOL_MANAGER_ACTIVE)
1993 pool->flags |= POOL_MANAGER_ACTIVE;
1994 pool->manager = worker;
1996 maybe_create_worker(pool);
1998 pool->manager = NULL;
1999 pool->flags &= ~POOL_MANAGER_ACTIVE;
2000 wake_up(&wq_manager_wait);
2005 * process_one_work - process single work
2007 * @work: work to process
2009 * Process @work. This function contains all the logics necessary to
2010 * process a single work including synchronization against and
2011 * interaction with other workers on the same cpu, queueing and
2012 * flushing. As long as context requirement is met, any worker can
2013 * call this function to process a work.
2016 * spin_lock_irq(pool->lock) which is released and regrabbed.
2018 static void process_one_work(struct worker *worker, struct work_struct *work)
2019 __releases(&pool->lock)
2020 __acquires(&pool->lock)
2022 struct pool_workqueue *pwq = get_work_pwq(work);
2023 struct worker_pool *pool = worker->pool;
2024 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2026 struct worker *collision;
2027 #ifdef CONFIG_LOCKDEP
2029 * It is permissible to free the struct work_struct from
2030 * inside the function that is called from it, this we need to
2031 * take into account for lockdep too. To avoid bogus "held
2032 * lock freed" warnings as well as problems when looking into
2033 * work->lockdep_map, make a copy and use that here.
2035 struct lockdep_map lockdep_map;
2037 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2039 /* ensure we're on the correct CPU */
2040 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2041 raw_smp_processor_id() != pool->cpu);
2044 * A single work shouldn't be executed concurrently by
2045 * multiple workers on a single cpu. Check whether anyone is
2046 * already processing the work. If so, defer the work to the
2047 * currently executing one.
2049 collision = find_worker_executing_work(pool, work);
2050 if (unlikely(collision)) {
2051 move_linked_works(work, &collision->scheduled, NULL);
2055 /* claim and dequeue */
2056 debug_work_deactivate(work);
2057 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2058 worker->current_work = work;
2059 worker->current_func = work->func;
2060 worker->current_pwq = pwq;
2061 work_color = get_work_color(work);
2063 list_del_init(&work->entry);
2066 * CPU intensive works don't participate in concurrency management.
2067 * They're the scheduler's responsibility. This takes @worker out
2068 * of concurrency management and the next code block will chain
2069 * execution of the pending work items.
2071 if (unlikely(cpu_intensive))
2072 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2075 * Wake up another worker if necessary. The condition is always
2076 * false for normal per-cpu workers since nr_running would always
2077 * be >= 1 at this point. This is used to chain execution of the
2078 * pending work items for WORKER_NOT_RUNNING workers such as the
2079 * UNBOUND and CPU_INTENSIVE ones.
2081 if (need_more_worker(pool))
2082 wake_up_worker(pool);
2085 * Record the last pool and clear PENDING which should be the last
2086 * update to @work. Also, do this inside @pool->lock so that
2087 * PENDING and queued state changes happen together while IRQ is
2090 set_work_pool_and_clear_pending(work, pool->id);
2092 spin_unlock_irq(&pool->lock);
2094 lock_map_acquire_read(&pwq->wq->lockdep_map);
2095 lock_map_acquire(&lockdep_map);
2096 trace_workqueue_execute_start(work);
2097 worker->current_func(work);
2099 * While we must be careful to not use "work" after this, the trace
2100 * point will only record its address.
2102 trace_workqueue_execute_end(work);
2103 lock_map_release(&lockdep_map);
2104 lock_map_release(&pwq->wq->lockdep_map);
2106 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2107 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2108 " last function: %pf\n",
2109 current->comm, preempt_count(), task_pid_nr(current),
2110 worker->current_func);
2111 debug_show_held_locks(current);
2116 * The following prevents a kworker from hogging CPU on !PREEMPT
2117 * kernels, where a requeueing work item waiting for something to
2118 * happen could deadlock with stop_machine as such work item could
2119 * indefinitely requeue itself while all other CPUs are trapped in
2120 * stop_machine. At the same time, report a quiescent RCU state so
2121 * the same condition doesn't freeze RCU.
2123 cond_resched_rcu_qs();
2125 spin_lock_irq(&pool->lock);
2127 /* clear cpu intensive status */
2128 if (unlikely(cpu_intensive))
2129 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2131 /* we're done with it, release */
2132 hash_del(&worker->hentry);
2133 worker->current_work = NULL;
2134 worker->current_func = NULL;
2135 worker->current_pwq = NULL;
2136 worker->desc_valid = false;
2137 pwq_dec_nr_in_flight(pwq, work_color);
2141 * process_scheduled_works - process scheduled works
2144 * Process all scheduled works. Please note that the scheduled list
2145 * may change while processing a work, so this function repeatedly
2146 * fetches a work from the top and executes it.
2149 * spin_lock_irq(pool->lock) which may be released and regrabbed
2152 static void process_scheduled_works(struct worker *worker)
2154 while (!list_empty(&worker->scheduled)) {
2155 struct work_struct *work = list_first_entry(&worker->scheduled,
2156 struct work_struct, entry);
2157 process_one_work(worker, work);
2162 * worker_thread - the worker thread function
2165 * The worker thread function. All workers belong to a worker_pool -
2166 * either a per-cpu one or dynamic unbound one. These workers process all
2167 * work items regardless of their specific target workqueue. The only
2168 * exception is work items which belong to workqueues with a rescuer which
2169 * will be explained in rescuer_thread().
2173 static int worker_thread(void *__worker)
2175 struct worker *worker = __worker;
2176 struct worker_pool *pool = worker->pool;
2178 /* tell the scheduler that this is a workqueue worker */
2179 worker->task->flags |= PF_WQ_WORKER;
2181 spin_lock_irq(&pool->lock);
2183 /* am I supposed to die? */
2184 if (unlikely(worker->flags & WORKER_DIE)) {
2185 spin_unlock_irq(&pool->lock);
2186 WARN_ON_ONCE(!list_empty(&worker->entry));
2187 worker->task->flags &= ~PF_WQ_WORKER;
2189 set_task_comm(worker->task, "kworker/dying");
2190 ida_simple_remove(&pool->worker_ida, worker->id);
2191 worker_detach_from_pool(worker, pool);
2196 worker_leave_idle(worker);
2198 /* no more worker necessary? */
2199 if (!need_more_worker(pool))
2202 /* do we need to manage? */
2203 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2207 * ->scheduled list can only be filled while a worker is
2208 * preparing to process a work or actually processing it.
2209 * Make sure nobody diddled with it while I was sleeping.
2211 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2214 * Finish PREP stage. We're guaranteed to have at least one idle
2215 * worker or that someone else has already assumed the manager
2216 * role. This is where @worker starts participating in concurrency
2217 * management if applicable and concurrency management is restored
2218 * after being rebound. See rebind_workers() for details.
2220 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2223 struct work_struct *work =
2224 list_first_entry(&pool->worklist,
2225 struct work_struct, entry);
2227 pool->watchdog_ts = jiffies;
2229 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2230 /* optimization path, not strictly necessary */
2231 process_one_work(worker, work);
2232 if (unlikely(!list_empty(&worker->scheduled)))
2233 process_scheduled_works(worker);
2235 move_linked_works(work, &worker->scheduled, NULL);
2236 process_scheduled_works(worker);
2238 } while (keep_working(pool));
2240 worker_set_flags(worker, WORKER_PREP);
2243 * pool->lock is held and there's no work to process and no need to
2244 * manage, sleep. Workers are woken up only while holding
2245 * pool->lock or from local cpu, so setting the current state
2246 * before releasing pool->lock is enough to prevent losing any
2249 worker_enter_idle(worker);
2250 __set_current_state(TASK_INTERRUPTIBLE);
2251 spin_unlock_irq(&pool->lock);
2257 * rescuer_thread - the rescuer thread function
2260 * Workqueue rescuer thread function. There's one rescuer for each
2261 * workqueue which has WQ_MEM_RECLAIM set.
2263 * Regular work processing on a pool may block trying to create a new
2264 * worker which uses GFP_KERNEL allocation which has slight chance of
2265 * developing into deadlock if some works currently on the same queue
2266 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2267 * the problem rescuer solves.
2269 * When such condition is possible, the pool summons rescuers of all
2270 * workqueues which have works queued on the pool and let them process
2271 * those works so that forward progress can be guaranteed.
2273 * This should happen rarely.
2277 static int rescuer_thread(void *__rescuer)
2279 struct worker *rescuer = __rescuer;
2280 struct workqueue_struct *wq = rescuer->rescue_wq;
2281 struct list_head *scheduled = &rescuer->scheduled;
2284 set_user_nice(current, RESCUER_NICE_LEVEL);
2287 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2288 * doesn't participate in concurrency management.
2290 rescuer->task->flags |= PF_WQ_WORKER;
2292 set_current_state(TASK_INTERRUPTIBLE);
2295 * By the time the rescuer is requested to stop, the workqueue
2296 * shouldn't have any work pending, but @wq->maydays may still have
2297 * pwq(s) queued. This can happen by non-rescuer workers consuming
2298 * all the work items before the rescuer got to them. Go through
2299 * @wq->maydays processing before acting on should_stop so that the
2300 * list is always empty on exit.
2302 should_stop = kthread_should_stop();
2304 /* see whether any pwq is asking for help */
2305 spin_lock_irq(&wq_mayday_lock);
2307 while (!list_empty(&wq->maydays)) {
2308 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2309 struct pool_workqueue, mayday_node);
2310 struct worker_pool *pool = pwq->pool;
2311 struct work_struct *work, *n;
2314 __set_current_state(TASK_RUNNING);
2315 list_del_init(&pwq->mayday_node);
2317 spin_unlock_irq(&wq_mayday_lock);
2319 worker_attach_to_pool(rescuer, pool);
2321 spin_lock_irq(&pool->lock);
2322 rescuer->pool = pool;
2325 * Slurp in all works issued via this workqueue and
2328 WARN_ON_ONCE(!list_empty(scheduled));
2329 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2330 if (get_work_pwq(work) == pwq) {
2332 pool->watchdog_ts = jiffies;
2333 move_linked_works(work, scheduled, &n);
2338 if (!list_empty(scheduled)) {
2339 process_scheduled_works(rescuer);
2342 * The above execution of rescued work items could
2343 * have created more to rescue through
2344 * pwq_activate_first_delayed() or chained
2345 * queueing. Let's put @pwq back on mayday list so
2346 * that such back-to-back work items, which may be
2347 * being used to relieve memory pressure, don't
2348 * incur MAYDAY_INTERVAL delay inbetween.
2350 if (need_to_create_worker(pool)) {
2351 spin_lock(&wq_mayday_lock);
2353 * Queue iff we aren't racing destruction
2354 * and somebody else hasn't queued it already.
2356 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2358 list_add_tail(&pwq->mayday_node, &wq->maydays);
2360 spin_unlock(&wq_mayday_lock);
2365 * Put the reference grabbed by send_mayday(). @pool won't
2366 * go away while we're still attached to it.
2371 * Leave this pool. If need_more_worker() is %true, notify a
2372 * regular worker; otherwise, we end up with 0 concurrency
2373 * and stalling the execution.
2375 if (need_more_worker(pool))
2376 wake_up_worker(pool);
2378 rescuer->pool = NULL;
2379 spin_unlock_irq(&pool->lock);
2381 worker_detach_from_pool(rescuer, pool);
2383 spin_lock_irq(&wq_mayday_lock);
2386 spin_unlock_irq(&wq_mayday_lock);
2389 __set_current_state(TASK_RUNNING);
2390 rescuer->task->flags &= ~PF_WQ_WORKER;
2394 /* rescuers should never participate in concurrency management */
2395 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2401 * check_flush_dependency - check for flush dependency sanity
2402 * @target_wq: workqueue being flushed
2403 * @target_work: work item being flushed (NULL for workqueue flushes)
2405 * %current is trying to flush the whole @target_wq or @target_work on it.
2406 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2407 * reclaiming memory or running on a workqueue which doesn't have
2408 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2411 static void check_flush_dependency(struct workqueue_struct *target_wq,
2412 struct work_struct *target_work)
2414 work_func_t target_func = target_work ? target_work->func : NULL;
2415 struct worker *worker;
2417 if (target_wq->flags & WQ_MEM_RECLAIM)
2420 worker = current_wq_worker();
2422 WARN_ONCE(current->flags & PF_MEMALLOC,
2423 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2424 current->pid, current->comm, target_wq->name, target_func);
2425 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2426 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2427 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2428 worker->current_pwq->wq->name, worker->current_func,
2429 target_wq->name, target_func);
2433 struct work_struct work;
2434 struct completion done;
2435 struct task_struct *task; /* purely informational */
2438 static void wq_barrier_func(struct work_struct *work)
2440 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2441 complete(&barr->done);
2445 * insert_wq_barrier - insert a barrier work
2446 * @pwq: pwq to insert barrier into
2447 * @barr: wq_barrier to insert
2448 * @target: target work to attach @barr to
2449 * @worker: worker currently executing @target, NULL if @target is not executing
2451 * @barr is linked to @target such that @barr is completed only after
2452 * @target finishes execution. Please note that the ordering
2453 * guarantee is observed only with respect to @target and on the local
2456 * Currently, a queued barrier can't be canceled. This is because
2457 * try_to_grab_pending() can't determine whether the work to be
2458 * grabbed is at the head of the queue and thus can't clear LINKED
2459 * flag of the previous work while there must be a valid next work
2460 * after a work with LINKED flag set.
2462 * Note that when @worker is non-NULL, @target may be modified
2463 * underneath us, so we can't reliably determine pwq from @target.
2466 * spin_lock_irq(pool->lock).
2468 static void insert_wq_barrier(struct pool_workqueue *pwq,
2469 struct wq_barrier *barr,
2470 struct work_struct *target, struct worker *worker)
2472 struct list_head *head;
2473 unsigned int linked = 0;
2476 * debugobject calls are safe here even with pool->lock locked
2477 * as we know for sure that this will not trigger any of the
2478 * checks and call back into the fixup functions where we
2481 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2482 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2483 init_completion(&barr->done);
2484 barr->task = current;
2487 * If @target is currently being executed, schedule the
2488 * barrier to the worker; otherwise, put it after @target.
2491 head = worker->scheduled.next;
2493 unsigned long *bits = work_data_bits(target);
2495 head = target->entry.next;
2496 /* there can already be other linked works, inherit and set */
2497 linked = *bits & WORK_STRUCT_LINKED;
2498 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2501 debug_work_activate(&barr->work);
2502 insert_work(pwq, &barr->work, head,
2503 work_color_to_flags(WORK_NO_COLOR) | linked);
2507 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2508 * @wq: workqueue being flushed
2509 * @flush_color: new flush color, < 0 for no-op
2510 * @work_color: new work color, < 0 for no-op
2512 * Prepare pwqs for workqueue flushing.
2514 * If @flush_color is non-negative, flush_color on all pwqs should be
2515 * -1. If no pwq has in-flight commands at the specified color, all
2516 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2517 * has in flight commands, its pwq->flush_color is set to
2518 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2519 * wakeup logic is armed and %true is returned.
2521 * The caller should have initialized @wq->first_flusher prior to
2522 * calling this function with non-negative @flush_color. If
2523 * @flush_color is negative, no flush color update is done and %false
2526 * If @work_color is non-negative, all pwqs should have the same
2527 * work_color which is previous to @work_color and all will be
2528 * advanced to @work_color.
2531 * mutex_lock(wq->mutex).
2534 * %true if @flush_color >= 0 and there's something to flush. %false
2537 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2538 int flush_color, int work_color)
2541 struct pool_workqueue *pwq;
2543 if (flush_color >= 0) {
2544 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2545 atomic_set(&wq->nr_pwqs_to_flush, 1);
2548 for_each_pwq(pwq, wq) {
2549 struct worker_pool *pool = pwq->pool;
2551 spin_lock_irq(&pool->lock);
2553 if (flush_color >= 0) {
2554 WARN_ON_ONCE(pwq->flush_color != -1);
2556 if (pwq->nr_in_flight[flush_color]) {
2557 pwq->flush_color = flush_color;
2558 atomic_inc(&wq->nr_pwqs_to_flush);
2563 if (work_color >= 0) {
2564 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2565 pwq->work_color = work_color;
2568 spin_unlock_irq(&pool->lock);
2571 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2572 complete(&wq->first_flusher->done);
2578 * flush_workqueue - ensure that any scheduled work has run to completion.
2579 * @wq: workqueue to flush
2581 * This function sleeps until all work items which were queued on entry
2582 * have finished execution, but it is not livelocked by new incoming ones.
2584 void flush_workqueue(struct workqueue_struct *wq)
2586 struct wq_flusher this_flusher = {
2587 .list = LIST_HEAD_INIT(this_flusher.list),
2589 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2593 if (WARN_ON(!wq_online))
2596 lock_map_acquire(&wq->lockdep_map);
2597 lock_map_release(&wq->lockdep_map);
2599 mutex_lock(&wq->mutex);
2602 * Start-to-wait phase
2604 next_color = work_next_color(wq->work_color);
2606 if (next_color != wq->flush_color) {
2608 * Color space is not full. The current work_color
2609 * becomes our flush_color and work_color is advanced
2612 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2613 this_flusher.flush_color = wq->work_color;
2614 wq->work_color = next_color;
2616 if (!wq->first_flusher) {
2617 /* no flush in progress, become the first flusher */
2618 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2620 wq->first_flusher = &this_flusher;
2622 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2624 /* nothing to flush, done */
2625 wq->flush_color = next_color;
2626 wq->first_flusher = NULL;
2631 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2632 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2633 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2637 * Oops, color space is full, wait on overflow queue.
2638 * The next flush completion will assign us
2639 * flush_color and transfer to flusher_queue.
2641 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2644 check_flush_dependency(wq, NULL);
2646 mutex_unlock(&wq->mutex);
2648 wait_for_completion(&this_flusher.done);
2651 * Wake-up-and-cascade phase
2653 * First flushers are responsible for cascading flushes and
2654 * handling overflow. Non-first flushers can simply return.
2656 if (wq->first_flusher != &this_flusher)
2659 mutex_lock(&wq->mutex);
2661 /* we might have raced, check again with mutex held */
2662 if (wq->first_flusher != &this_flusher)
2665 wq->first_flusher = NULL;
2667 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2668 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2671 struct wq_flusher *next, *tmp;
2673 /* complete all the flushers sharing the current flush color */
2674 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2675 if (next->flush_color != wq->flush_color)
2677 list_del_init(&next->list);
2678 complete(&next->done);
2681 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2682 wq->flush_color != work_next_color(wq->work_color));
2684 /* this flush_color is finished, advance by one */
2685 wq->flush_color = work_next_color(wq->flush_color);
2687 /* one color has been freed, handle overflow queue */
2688 if (!list_empty(&wq->flusher_overflow)) {
2690 * Assign the same color to all overflowed
2691 * flushers, advance work_color and append to
2692 * flusher_queue. This is the start-to-wait
2693 * phase for these overflowed flushers.
2695 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2696 tmp->flush_color = wq->work_color;
2698 wq->work_color = work_next_color(wq->work_color);
2700 list_splice_tail_init(&wq->flusher_overflow,
2701 &wq->flusher_queue);
2702 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2705 if (list_empty(&wq->flusher_queue)) {
2706 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2711 * Need to flush more colors. Make the next flusher
2712 * the new first flusher and arm pwqs.
2714 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2715 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2717 list_del_init(&next->list);
2718 wq->first_flusher = next;
2720 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2724 * Meh... this color is already done, clear first
2725 * flusher and repeat cascading.
2727 wq->first_flusher = NULL;
2731 mutex_unlock(&wq->mutex);
2733 EXPORT_SYMBOL(flush_workqueue);
2736 * drain_workqueue - drain a workqueue
2737 * @wq: workqueue to drain
2739 * Wait until the workqueue becomes empty. While draining is in progress,
2740 * only chain queueing is allowed. IOW, only currently pending or running
2741 * work items on @wq can queue further work items on it. @wq is flushed
2742 * repeatedly until it becomes empty. The number of flushing is determined
2743 * by the depth of chaining and should be relatively short. Whine if it
2746 void drain_workqueue(struct workqueue_struct *wq)
2748 unsigned int flush_cnt = 0;
2749 struct pool_workqueue *pwq;
2752 * __queue_work() needs to test whether there are drainers, is much
2753 * hotter than drain_workqueue() and already looks at @wq->flags.
2754 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2756 mutex_lock(&wq->mutex);
2757 if (!wq->nr_drainers++)
2758 wq->flags |= __WQ_DRAINING;
2759 mutex_unlock(&wq->mutex);
2761 flush_workqueue(wq);
2763 mutex_lock(&wq->mutex);
2765 for_each_pwq(pwq, wq) {
2768 spin_lock_irq(&pwq->pool->lock);
2769 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2770 spin_unlock_irq(&pwq->pool->lock);
2775 if (++flush_cnt == 10 ||
2776 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2777 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2778 wq->name, flush_cnt);
2780 mutex_unlock(&wq->mutex);
2784 if (!--wq->nr_drainers)
2785 wq->flags &= ~__WQ_DRAINING;
2786 mutex_unlock(&wq->mutex);
2788 EXPORT_SYMBOL_GPL(drain_workqueue);
2790 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2792 struct worker *worker = NULL;
2793 struct worker_pool *pool;
2794 struct pool_workqueue *pwq;
2798 local_irq_disable();
2799 pool = get_work_pool(work);
2805 spin_lock(&pool->lock);
2806 /* see the comment in try_to_grab_pending() with the same code */
2807 pwq = get_work_pwq(work);
2809 if (unlikely(pwq->pool != pool))
2812 worker = find_worker_executing_work(pool, work);
2815 pwq = worker->current_pwq;
2818 check_flush_dependency(pwq->wq, work);
2820 insert_wq_barrier(pwq, barr, work, worker);
2821 spin_unlock_irq(&pool->lock);
2824 * If @max_active is 1 or rescuer is in use, flushing another work
2825 * item on the same workqueue may lead to deadlock. Make sure the
2826 * flusher is not running on the same workqueue by verifying write
2829 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2830 lock_map_acquire(&pwq->wq->lockdep_map);
2832 lock_map_acquire_read(&pwq->wq->lockdep_map);
2833 lock_map_release(&pwq->wq->lockdep_map);
2837 spin_unlock_irq(&pool->lock);
2842 * flush_work - wait for a work to finish executing the last queueing instance
2843 * @work: the work to flush
2845 * Wait until @work has finished execution. @work is guaranteed to be idle
2846 * on return if it hasn't been requeued since flush started.
2849 * %true if flush_work() waited for the work to finish execution,
2850 * %false if it was already idle.
2852 bool flush_work(struct work_struct *work)
2854 struct wq_barrier barr;
2856 if (WARN_ON(!wq_online))
2859 lock_map_acquire(&work->lockdep_map);
2860 lock_map_release(&work->lockdep_map);
2862 if (start_flush_work(work, &barr)) {
2863 wait_for_completion(&barr.done);
2864 destroy_work_on_stack(&barr.work);
2870 EXPORT_SYMBOL_GPL(flush_work);
2874 struct work_struct *work;
2877 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2879 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2881 if (cwait->work != key)
2883 return autoremove_wake_function(wait, mode, sync, key);
2886 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2888 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2889 unsigned long flags;
2893 ret = try_to_grab_pending(work, is_dwork, &flags);
2895 * If someone else is already canceling, wait for it to
2896 * finish. flush_work() doesn't work for PREEMPT_NONE
2897 * because we may get scheduled between @work's completion
2898 * and the other canceling task resuming and clearing
2899 * CANCELING - flush_work() will return false immediately
2900 * as @work is no longer busy, try_to_grab_pending() will
2901 * return -ENOENT as @work is still being canceled and the
2902 * other canceling task won't be able to clear CANCELING as
2903 * we're hogging the CPU.
2905 * Let's wait for completion using a waitqueue. As this
2906 * may lead to the thundering herd problem, use a custom
2907 * wake function which matches @work along with exclusive
2910 if (unlikely(ret == -ENOENT)) {
2911 struct cwt_wait cwait;
2913 init_wait(&cwait.wait);
2914 cwait.wait.func = cwt_wakefn;
2917 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2918 TASK_UNINTERRUPTIBLE);
2919 if (work_is_canceling(work))
2921 finish_wait(&cancel_waitq, &cwait.wait);
2923 } while (unlikely(ret < 0));
2925 /* tell other tasks trying to grab @work to back off */
2926 mark_work_canceling(work);
2927 local_irq_restore(flags);
2930 * This allows canceling during early boot. We know that @work
2936 clear_work_data(work);
2939 * Paired with prepare_to_wait() above so that either
2940 * waitqueue_active() is visible here or !work_is_canceling() is
2944 if (waitqueue_active(&cancel_waitq))
2945 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2951 * cancel_work_sync - cancel a work and wait for it to finish
2952 * @work: the work to cancel
2954 * Cancel @work and wait for its execution to finish. This function
2955 * can be used even if the work re-queues itself or migrates to
2956 * another workqueue. On return from this function, @work is
2957 * guaranteed to be not pending or executing on any CPU.
2959 * cancel_work_sync(&delayed_work->work) must not be used for
2960 * delayed_work's. Use cancel_delayed_work_sync() instead.
2962 * The caller must ensure that the workqueue on which @work was last
2963 * queued can't be destroyed before this function returns.
2966 * %true if @work was pending, %false otherwise.
2968 bool cancel_work_sync(struct work_struct *work)
2970 return __cancel_work_timer(work, false);
2972 EXPORT_SYMBOL_GPL(cancel_work_sync);
2975 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2976 * @dwork: the delayed work to flush
2978 * Delayed timer is cancelled and the pending work is queued for
2979 * immediate execution. Like flush_work(), this function only
2980 * considers the last queueing instance of @dwork.
2983 * %true if flush_work() waited for the work to finish execution,
2984 * %false if it was already idle.
2986 bool flush_delayed_work(struct delayed_work *dwork)
2988 local_irq_disable();
2989 if (del_timer_sync(&dwork->timer))
2990 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2992 return flush_work(&dwork->work);
2994 EXPORT_SYMBOL(flush_delayed_work);
2996 static bool __cancel_work(struct work_struct *work, bool is_dwork)
2998 unsigned long flags;
3002 ret = try_to_grab_pending(work, is_dwork, &flags);
3003 } while (unlikely(ret == -EAGAIN));
3005 if (unlikely(ret < 0))
3008 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3009 local_irq_restore(flags);
3014 * See cancel_delayed_work()
3016 bool cancel_work(struct work_struct *work)
3018 return __cancel_work(work, false);
3022 * cancel_delayed_work - cancel a delayed work
3023 * @dwork: delayed_work to cancel
3025 * Kill off a pending delayed_work.
3027 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3031 * The work callback function may still be running on return, unless
3032 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3033 * use cancel_delayed_work_sync() to wait on it.
3035 * This function is safe to call from any context including IRQ handler.
3037 bool cancel_delayed_work(struct delayed_work *dwork)
3039 return __cancel_work(&dwork->work, true);
3041 EXPORT_SYMBOL(cancel_delayed_work);
3044 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3045 * @dwork: the delayed work cancel
3047 * This is cancel_work_sync() for delayed works.
3050 * %true if @dwork was pending, %false otherwise.
3052 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3054 return __cancel_work_timer(&dwork->work, true);
3056 EXPORT_SYMBOL(cancel_delayed_work_sync);
3059 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3060 * @func: the function to call
3062 * schedule_on_each_cpu() executes @func on each online CPU using the
3063 * system workqueue and blocks until all CPUs have completed.
3064 * schedule_on_each_cpu() is very slow.
3067 * 0 on success, -errno on failure.
3069 int schedule_on_each_cpu(work_func_t func)
3072 struct work_struct __percpu *works;
3074 works = alloc_percpu(struct work_struct);
3080 for_each_online_cpu(cpu) {
3081 struct work_struct *work = per_cpu_ptr(works, cpu);
3083 INIT_WORK(work, func);
3084 schedule_work_on(cpu, work);
3087 for_each_online_cpu(cpu)
3088 flush_work(per_cpu_ptr(works, cpu));
3096 * execute_in_process_context - reliably execute the routine with user context
3097 * @fn: the function to execute
3098 * @ew: guaranteed storage for the execute work structure (must
3099 * be available when the work executes)
3101 * Executes the function immediately if process context is available,
3102 * otherwise schedules the function for delayed execution.
3104 * Return: 0 - function was executed
3105 * 1 - function was scheduled for execution
3107 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3109 if (!in_interrupt()) {
3114 INIT_WORK(&ew->work, fn);
3115 schedule_work(&ew->work);
3119 EXPORT_SYMBOL_GPL(execute_in_process_context);
3122 * free_workqueue_attrs - free a workqueue_attrs
3123 * @attrs: workqueue_attrs to free
3125 * Undo alloc_workqueue_attrs().
3127 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3130 free_cpumask_var(attrs->cpumask);
3136 * alloc_workqueue_attrs - allocate a workqueue_attrs
3137 * @gfp_mask: allocation mask to use
3139 * Allocate a new workqueue_attrs, initialize with default settings and
3142 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3144 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3146 struct workqueue_attrs *attrs;
3148 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3151 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3154 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3157 free_workqueue_attrs(attrs);
3161 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3162 const struct workqueue_attrs *from)
3164 to->nice = from->nice;
3165 cpumask_copy(to->cpumask, from->cpumask);
3167 * Unlike hash and equality test, this function doesn't ignore
3168 * ->no_numa as it is used for both pool and wq attrs. Instead,
3169 * get_unbound_pool() explicitly clears ->no_numa after copying.
3171 to->no_numa = from->no_numa;
3174 /* hash value of the content of @attr */
3175 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3179 hash = jhash_1word(attrs->nice, hash);
3180 hash = jhash(cpumask_bits(attrs->cpumask),
3181 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3185 /* content equality test */
3186 static bool wqattrs_equal(const struct workqueue_attrs *a,
3187 const struct workqueue_attrs *b)
3189 if (a->nice != b->nice)
3191 if (!cpumask_equal(a->cpumask, b->cpumask))
3197 * init_worker_pool - initialize a newly zalloc'd worker_pool
3198 * @pool: worker_pool to initialize
3200 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3202 * Return: 0 on success, -errno on failure. Even on failure, all fields
3203 * inside @pool proper are initialized and put_unbound_pool() can be called
3204 * on @pool safely to release it.
3206 static int init_worker_pool(struct worker_pool *pool)
3208 spin_lock_init(&pool->lock);
3211 pool->node = NUMA_NO_NODE;
3212 pool->flags |= POOL_DISASSOCIATED;
3213 pool->watchdog_ts = jiffies;
3214 INIT_LIST_HEAD(&pool->worklist);
3215 INIT_LIST_HEAD(&pool->idle_list);
3216 hash_init(pool->busy_hash);
3218 init_timer_deferrable(&pool->idle_timer);
3219 pool->idle_timer.function = idle_worker_timeout;
3220 pool->idle_timer.data = (unsigned long)pool;
3222 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3223 (unsigned long)pool);
3225 mutex_init(&pool->attach_mutex);
3226 INIT_LIST_HEAD(&pool->workers);
3228 ida_init(&pool->worker_ida);
3229 INIT_HLIST_NODE(&pool->hash_node);
3232 /* shouldn't fail above this point */
3233 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3239 static void rcu_free_wq(struct rcu_head *rcu)
3241 struct workqueue_struct *wq =
3242 container_of(rcu, struct workqueue_struct, rcu);
3244 if (!(wq->flags & WQ_UNBOUND))
3245 free_percpu(wq->cpu_pwqs);
3247 free_workqueue_attrs(wq->unbound_attrs);
3253 static void rcu_free_pool(struct rcu_head *rcu)
3255 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3257 ida_destroy(&pool->worker_ida);
3258 free_workqueue_attrs(pool->attrs);
3263 * put_unbound_pool - put a worker_pool
3264 * @pool: worker_pool to put
3266 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3267 * safe manner. get_unbound_pool() calls this function on its failure path
3268 * and this function should be able to release pools which went through,
3269 * successfully or not, init_worker_pool().
3271 * Should be called with wq_pool_mutex held.
3273 static void put_unbound_pool(struct worker_pool *pool)
3275 DECLARE_COMPLETION_ONSTACK(detach_completion);
3276 struct worker *worker;
3278 lockdep_assert_held(&wq_pool_mutex);
3284 if (WARN_ON(!(pool->cpu < 0)) ||
3285 WARN_ON(!list_empty(&pool->worklist)))
3288 /* release id and unhash */
3290 idr_remove(&worker_pool_idr, pool->id);
3291 hash_del(&pool->hash_node);
3294 * Become the manager and destroy all workers. This prevents
3295 * @pool's workers from blocking on attach_mutex. We're the last
3296 * manager and @pool gets freed with the flag set.
3298 spin_lock_irq(&pool->lock);
3299 wait_event_lock_irq(wq_manager_wait,
3300 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3301 pool->flags |= POOL_MANAGER_ACTIVE;
3303 while ((worker = first_idle_worker(pool)))
3304 destroy_worker(worker);
3305 WARN_ON(pool->nr_workers || pool->nr_idle);
3306 spin_unlock_irq(&pool->lock);
3308 mutex_lock(&pool->attach_mutex);
3309 if (!list_empty(&pool->workers))
3310 pool->detach_completion = &detach_completion;
3311 mutex_unlock(&pool->attach_mutex);
3313 if (pool->detach_completion)
3314 wait_for_completion(pool->detach_completion);
3316 /* shut down the timers */
3317 del_timer_sync(&pool->idle_timer);
3318 del_timer_sync(&pool->mayday_timer);
3320 /* sched-RCU protected to allow dereferences from get_work_pool() */
3321 call_rcu_sched(&pool->rcu, rcu_free_pool);
3325 * get_unbound_pool - get a worker_pool with the specified attributes
3326 * @attrs: the attributes of the worker_pool to get
3328 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3329 * reference count and return it. If there already is a matching
3330 * worker_pool, it will be used; otherwise, this function attempts to
3333 * Should be called with wq_pool_mutex held.
3335 * Return: On success, a worker_pool with the same attributes as @attrs.
3336 * On failure, %NULL.
3338 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3340 u32 hash = wqattrs_hash(attrs);
3341 struct worker_pool *pool;
3343 int target_node = NUMA_NO_NODE;
3345 lockdep_assert_held(&wq_pool_mutex);
3347 /* do we already have a matching pool? */
3348 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3349 if (wqattrs_equal(pool->attrs, attrs)) {
3355 /* if cpumask is contained inside a NUMA node, we belong to that node */
3356 if (wq_numa_enabled) {
3357 for_each_node(node) {
3358 if (cpumask_subset(attrs->cpumask,
3359 wq_numa_possible_cpumask[node])) {
3366 /* nope, create a new one */
3367 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3368 if (!pool || init_worker_pool(pool) < 0)
3371 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3372 copy_workqueue_attrs(pool->attrs, attrs);
3373 pool->node = target_node;
3376 * no_numa isn't a worker_pool attribute, always clear it. See
3377 * 'struct workqueue_attrs' comments for detail.
3379 pool->attrs->no_numa = false;
3381 if (worker_pool_assign_id(pool) < 0)
3384 /* create and start the initial worker */
3385 if (wq_online && !create_worker(pool))
3389 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3394 put_unbound_pool(pool);
3398 static void rcu_free_pwq(struct rcu_head *rcu)
3400 kmem_cache_free(pwq_cache,
3401 container_of(rcu, struct pool_workqueue, rcu));
3405 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3406 * and needs to be destroyed.
3408 static void pwq_unbound_release_workfn(struct work_struct *work)
3410 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3411 unbound_release_work);
3412 struct workqueue_struct *wq = pwq->wq;
3413 struct worker_pool *pool = pwq->pool;
3414 bool is_last = false;
3417 * when @pwq is not linked, it doesn't hold any reference to the
3418 * @wq, and @wq is invalid to access.
3420 if (!list_empty(&pwq->pwqs_node)) {
3421 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3424 mutex_lock(&wq->mutex);
3425 list_del_rcu(&pwq->pwqs_node);
3426 is_last = list_empty(&wq->pwqs);
3427 mutex_unlock(&wq->mutex);
3430 mutex_lock(&wq_pool_mutex);
3431 put_unbound_pool(pool);
3432 mutex_unlock(&wq_pool_mutex);
3434 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3437 * If we're the last pwq going away, @wq is already dead and no one
3438 * is gonna access it anymore. Schedule RCU free.
3441 call_rcu_sched(&wq->rcu, rcu_free_wq);
3445 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3446 * @pwq: target pool_workqueue
3448 * If @pwq isn't freezing, set @pwq->max_active to the associated
3449 * workqueue's saved_max_active and activate delayed work items
3450 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3452 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3454 struct workqueue_struct *wq = pwq->wq;
3455 bool freezable = wq->flags & WQ_FREEZABLE;
3456 unsigned long flags;
3458 /* for @wq->saved_max_active */
3459 lockdep_assert_held(&wq->mutex);
3461 /* fast exit for non-freezable wqs */
3462 if (!freezable && pwq->max_active == wq->saved_max_active)
3465 /* this function can be called during early boot w/ irq disabled */
3466 spin_lock_irqsave(&pwq->pool->lock, flags);
3469 * During [un]freezing, the caller is responsible for ensuring that
3470 * this function is called at least once after @workqueue_freezing
3471 * is updated and visible.
3473 if (!freezable || !workqueue_freezing) {
3476 pwq->max_active = wq->saved_max_active;
3478 while (!list_empty(&pwq->delayed_works) &&
3479 pwq->nr_active < pwq->max_active) {
3480 pwq_activate_first_delayed(pwq);
3485 * Need to kick a worker after thawed or an unbound wq's
3486 * max_active is bumped. In realtime scenarios, always kicking a
3487 * worker will cause interference on the isolated cpu cores, so
3488 * let's kick iff work items were activated.
3491 wake_up_worker(pwq->pool);
3493 pwq->max_active = 0;
3496 spin_unlock_irqrestore(&pwq->pool->lock, flags);
3499 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3500 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3501 struct worker_pool *pool)
3503 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3505 memset(pwq, 0, sizeof(*pwq));
3509 pwq->flush_color = -1;
3511 INIT_LIST_HEAD(&pwq->delayed_works);
3512 INIT_LIST_HEAD(&pwq->pwqs_node);
3513 INIT_LIST_HEAD(&pwq->mayday_node);
3514 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3517 /* sync @pwq with the current state of its associated wq and link it */
3518 static void link_pwq(struct pool_workqueue *pwq)
3520 struct workqueue_struct *wq = pwq->wq;
3522 lockdep_assert_held(&wq->mutex);
3524 /* may be called multiple times, ignore if already linked */
3525 if (!list_empty(&pwq->pwqs_node))
3528 /* set the matching work_color */
3529 pwq->work_color = wq->work_color;
3531 /* sync max_active to the current setting */
3532 pwq_adjust_max_active(pwq);
3535 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3538 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3539 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3540 const struct workqueue_attrs *attrs)
3542 struct worker_pool *pool;
3543 struct pool_workqueue *pwq;
3545 lockdep_assert_held(&wq_pool_mutex);
3547 pool = get_unbound_pool(attrs);
3551 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3553 put_unbound_pool(pool);
3557 init_pwq(pwq, wq, pool);
3562 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3563 * @attrs: the wq_attrs of the default pwq of the target workqueue
3564 * @node: the target NUMA node
3565 * @cpu_going_down: if >= 0, the CPU to consider as offline
3566 * @cpumask: outarg, the resulting cpumask
3568 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3569 * @cpu_going_down is >= 0, that cpu is considered offline during
3570 * calculation. The result is stored in @cpumask.
3572 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3573 * enabled and @node has online CPUs requested by @attrs, the returned
3574 * cpumask is the intersection of the possible CPUs of @node and
3577 * The caller is responsible for ensuring that the cpumask of @node stays
3580 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3583 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3584 int cpu_going_down, cpumask_t *cpumask)
3586 if (!wq_numa_enabled || attrs->no_numa)
3589 /* does @node have any online CPUs @attrs wants? */
3590 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3591 if (cpu_going_down >= 0)
3592 cpumask_clear_cpu(cpu_going_down, cpumask);
3594 if (cpumask_empty(cpumask))
3597 /* yeap, return possible CPUs in @node that @attrs wants */
3598 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3599 return !cpumask_equal(cpumask, attrs->cpumask);
3602 cpumask_copy(cpumask, attrs->cpumask);
3606 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3607 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3609 struct pool_workqueue *pwq)
3611 struct pool_workqueue *old_pwq;
3613 lockdep_assert_held(&wq_pool_mutex);
3614 lockdep_assert_held(&wq->mutex);
3616 /* link_pwq() can handle duplicate calls */
3619 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3620 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3624 /* context to store the prepared attrs & pwqs before applying */
3625 struct apply_wqattrs_ctx {
3626 struct workqueue_struct *wq; /* target workqueue */
3627 struct workqueue_attrs *attrs; /* attrs to apply */
3628 struct list_head list; /* queued for batching commit */
3629 struct pool_workqueue *dfl_pwq;
3630 struct pool_workqueue *pwq_tbl[];
3633 /* free the resources after success or abort */
3634 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3640 put_pwq_unlocked(ctx->pwq_tbl[node]);
3641 put_pwq_unlocked(ctx->dfl_pwq);
3643 free_workqueue_attrs(ctx->attrs);
3649 /* allocate the attrs and pwqs for later installation */
3650 static struct apply_wqattrs_ctx *
3651 apply_wqattrs_prepare(struct workqueue_struct *wq,
3652 const struct workqueue_attrs *attrs)
3654 struct apply_wqattrs_ctx *ctx;
3655 struct workqueue_attrs *new_attrs, *tmp_attrs;
3658 lockdep_assert_held(&wq_pool_mutex);
3660 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3663 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3664 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3665 if (!ctx || !new_attrs || !tmp_attrs)
3669 * Calculate the attrs of the default pwq.
3670 * If the user configured cpumask doesn't overlap with the
3671 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3673 copy_workqueue_attrs(new_attrs, attrs);
3674 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3675 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3676 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3679 * We may create multiple pwqs with differing cpumasks. Make a
3680 * copy of @new_attrs which will be modified and used to obtain
3683 copy_workqueue_attrs(tmp_attrs, new_attrs);
3686 * If something goes wrong during CPU up/down, we'll fall back to
3687 * the default pwq covering whole @attrs->cpumask. Always create
3688 * it even if we don't use it immediately.
3690 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3694 for_each_node(node) {
3695 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3696 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3697 if (!ctx->pwq_tbl[node])
3700 ctx->dfl_pwq->refcnt++;
3701 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3705 /* save the user configured attrs and sanitize it. */
3706 copy_workqueue_attrs(new_attrs, attrs);
3707 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3708 ctx->attrs = new_attrs;
3711 free_workqueue_attrs(tmp_attrs);
3715 free_workqueue_attrs(tmp_attrs);
3716 free_workqueue_attrs(new_attrs);
3717 apply_wqattrs_cleanup(ctx);
3721 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3722 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3726 /* all pwqs have been created successfully, let's install'em */
3727 mutex_lock(&ctx->wq->mutex);
3729 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3731 /* save the previous pwq and install the new one */
3733 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3734 ctx->pwq_tbl[node]);
3736 /* @dfl_pwq might not have been used, ensure it's linked */
3737 link_pwq(ctx->dfl_pwq);
3738 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3740 mutex_unlock(&ctx->wq->mutex);
3743 static void apply_wqattrs_lock(void)
3745 /* CPUs should stay stable across pwq creations and installations */
3747 mutex_lock(&wq_pool_mutex);
3750 static void apply_wqattrs_unlock(void)
3752 mutex_unlock(&wq_pool_mutex);
3756 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3757 const struct workqueue_attrs *attrs)
3759 struct apply_wqattrs_ctx *ctx;
3761 /* only unbound workqueues can change attributes */
3762 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3765 /* creating multiple pwqs breaks ordering guarantee */
3766 if (!list_empty(&wq->pwqs)) {
3767 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3770 wq->flags &= ~__WQ_ORDERED;
3773 ctx = apply_wqattrs_prepare(wq, attrs);
3777 /* the ctx has been prepared successfully, let's commit it */
3778 apply_wqattrs_commit(ctx);
3779 apply_wqattrs_cleanup(ctx);
3785 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3786 * @wq: the target workqueue
3787 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3789 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3790 * machines, this function maps a separate pwq to each NUMA node with
3791 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3792 * NUMA node it was issued on. Older pwqs are released as in-flight work
3793 * items finish. Note that a work item which repeatedly requeues itself
3794 * back-to-back will stay on its current pwq.
3796 * Performs GFP_KERNEL allocations.
3798 * Return: 0 on success and -errno on failure.
3800 int apply_workqueue_attrs(struct workqueue_struct *wq,
3801 const struct workqueue_attrs *attrs)
3805 apply_wqattrs_lock();
3806 ret = apply_workqueue_attrs_locked(wq, attrs);
3807 apply_wqattrs_unlock();
3813 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3814 * @wq: the target workqueue
3815 * @cpu: the CPU coming up or going down
3816 * @online: whether @cpu is coming up or going down
3818 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3819 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3822 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3823 * falls back to @wq->dfl_pwq which may not be optimal but is always
3826 * Note that when the last allowed CPU of a NUMA node goes offline for a
3827 * workqueue with a cpumask spanning multiple nodes, the workers which were
3828 * already executing the work items for the workqueue will lose their CPU
3829 * affinity and may execute on any CPU. This is similar to how per-cpu
3830 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3831 * affinity, it's the user's responsibility to flush the work item from
3834 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3837 int node = cpu_to_node(cpu);
3838 int cpu_off = online ? -1 : cpu;
3839 struct pool_workqueue *old_pwq = NULL, *pwq;
3840 struct workqueue_attrs *target_attrs;
3843 lockdep_assert_held(&wq_pool_mutex);
3845 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3846 wq->unbound_attrs->no_numa)
3850 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3851 * Let's use a preallocated one. The following buf is protected by
3852 * CPU hotplug exclusion.
3854 target_attrs = wq_update_unbound_numa_attrs_buf;
3855 cpumask = target_attrs->cpumask;
3857 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3858 pwq = unbound_pwq_by_node(wq, node);
3861 * Let's determine what needs to be done. If the target cpumask is
3862 * different from the default pwq's, we need to compare it to @pwq's
3863 * and create a new one if they don't match. If the target cpumask
3864 * equals the default pwq's, the default pwq should be used.
3866 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3867 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3873 /* create a new pwq */
3874 pwq = alloc_unbound_pwq(wq, target_attrs);
3876 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3881 /* Install the new pwq. */
3882 mutex_lock(&wq->mutex);
3883 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3887 mutex_lock(&wq->mutex);
3888 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3889 get_pwq(wq->dfl_pwq);
3890 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3891 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3893 mutex_unlock(&wq->mutex);
3894 put_pwq_unlocked(old_pwq);
3897 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3899 bool highpri = wq->flags & WQ_HIGHPRI;
3902 if (!(wq->flags & WQ_UNBOUND)) {
3903 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3907 for_each_possible_cpu(cpu) {
3908 struct pool_workqueue *pwq =
3909 per_cpu_ptr(wq->cpu_pwqs, cpu);
3910 struct worker_pool *cpu_pools =
3911 per_cpu(cpu_worker_pools, cpu);
3913 init_pwq(pwq, wq, &cpu_pools[highpri]);
3915 mutex_lock(&wq->mutex);
3917 mutex_unlock(&wq->mutex);
3920 } else if (wq->flags & __WQ_ORDERED) {
3921 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3922 /* there should only be single pwq for ordering guarantee */
3923 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3924 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3925 "ordering guarantee broken for workqueue %s\n", wq->name);
3928 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3932 static int wq_clamp_max_active(int max_active, unsigned int flags,
3935 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3937 if (max_active < 1 || max_active > lim)
3938 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3939 max_active, name, 1, lim);
3941 return clamp_val(max_active, 1, lim);
3944 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3947 struct lock_class_key *key,
3948 const char *lock_name, ...)
3950 size_t tbl_size = 0;
3952 struct workqueue_struct *wq;
3953 struct pool_workqueue *pwq;
3956 * Unbound && max_active == 1 used to imply ordered, which is no
3957 * longer the case on NUMA machines due to per-node pools. While
3958 * alloc_ordered_workqueue() is the right way to create an ordered
3959 * workqueue, keep the previous behavior to avoid subtle breakages
3962 if ((flags & WQ_UNBOUND) && max_active == 1)
3963 flags |= __WQ_ORDERED;
3965 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3966 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3967 flags |= WQ_UNBOUND;
3969 /* allocate wq and format name */
3970 if (flags & WQ_UNBOUND)
3971 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3973 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3977 if (flags & WQ_UNBOUND) {
3978 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3979 if (!wq->unbound_attrs)
3983 va_start(args, lock_name);
3984 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3987 max_active = max_active ?: WQ_DFL_ACTIVE;
3988 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3992 wq->saved_max_active = max_active;
3993 mutex_init(&wq->mutex);
3994 atomic_set(&wq->nr_pwqs_to_flush, 0);
3995 INIT_LIST_HEAD(&wq->pwqs);
3996 INIT_LIST_HEAD(&wq->flusher_queue);
3997 INIT_LIST_HEAD(&wq->flusher_overflow);
3998 INIT_LIST_HEAD(&wq->maydays);
4000 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4001 INIT_LIST_HEAD(&wq->list);
4003 if (alloc_and_link_pwqs(wq) < 0)
4007 * Workqueues which may be used during memory reclaim should
4008 * have a rescuer to guarantee forward progress.
4010 if (flags & WQ_MEM_RECLAIM) {
4011 struct worker *rescuer;
4013 rescuer = alloc_worker(NUMA_NO_NODE);
4017 rescuer->rescue_wq = wq;
4018 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4020 if (IS_ERR(rescuer->task)) {
4025 wq->rescuer = rescuer;
4026 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4027 wake_up_process(rescuer->task);
4030 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4034 * wq_pool_mutex protects global freeze state and workqueues list.
4035 * Grab it, adjust max_active and add the new @wq to workqueues
4038 mutex_lock(&wq_pool_mutex);
4040 mutex_lock(&wq->mutex);
4041 for_each_pwq(pwq, wq)
4042 pwq_adjust_max_active(pwq);
4043 mutex_unlock(&wq->mutex);
4045 list_add_tail_rcu(&wq->list, &workqueues);
4047 mutex_unlock(&wq_pool_mutex);
4052 free_workqueue_attrs(wq->unbound_attrs);
4056 destroy_workqueue(wq);
4059 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4062 * destroy_workqueue - safely terminate a workqueue
4063 * @wq: target workqueue
4065 * Safely destroy a workqueue. All work currently pending will be done first.
4067 void destroy_workqueue(struct workqueue_struct *wq)
4069 struct pool_workqueue *pwq;
4073 * Remove it from sysfs first so that sanity check failure doesn't
4074 * lead to sysfs name conflicts.
4076 workqueue_sysfs_unregister(wq);
4078 /* drain it before proceeding with destruction */
4079 drain_workqueue(wq);
4081 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4083 struct worker *rescuer = wq->rescuer;
4085 /* this prevents new queueing */
4086 spin_lock_irq(&wq_mayday_lock);
4088 spin_unlock_irq(&wq_mayday_lock);
4090 /* rescuer will empty maydays list before exiting */
4091 kthread_stop(rescuer->task);
4096 mutex_lock(&wq->mutex);
4097 for_each_pwq(pwq, wq) {
4100 for (i = 0; i < WORK_NR_COLORS; i++) {
4101 if (WARN_ON(pwq->nr_in_flight[i])) {
4102 mutex_unlock(&wq->mutex);
4107 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4108 WARN_ON(pwq->nr_active) ||
4109 WARN_ON(!list_empty(&pwq->delayed_works))) {
4110 mutex_unlock(&wq->mutex);
4114 mutex_unlock(&wq->mutex);
4117 * wq list is used to freeze wq, remove from list after
4118 * flushing is complete in case freeze races us.
4120 mutex_lock(&wq_pool_mutex);
4121 list_del_rcu(&wq->list);
4122 mutex_unlock(&wq_pool_mutex);
4124 if (!(wq->flags & WQ_UNBOUND)) {
4126 * The base ref is never dropped on per-cpu pwqs. Directly
4127 * schedule RCU free.
4129 call_rcu_sched(&wq->rcu, rcu_free_wq);
4132 * We're the sole accessor of @wq at this point. Directly
4133 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4134 * @wq will be freed when the last pwq is released.
4136 for_each_node(node) {
4137 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4138 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4139 put_pwq_unlocked(pwq);
4143 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4144 * put. Don't access it afterwards.
4148 put_pwq_unlocked(pwq);
4151 EXPORT_SYMBOL_GPL(destroy_workqueue);
4154 * workqueue_set_max_active - adjust max_active of a workqueue
4155 * @wq: target workqueue
4156 * @max_active: new max_active value.
4158 * Set max_active of @wq to @max_active.
4161 * Don't call from IRQ context.
4163 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4165 struct pool_workqueue *pwq;
4167 /* disallow meddling with max_active for ordered workqueues */
4168 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4171 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4173 mutex_lock(&wq->mutex);
4175 wq->flags &= ~__WQ_ORDERED;
4176 wq->saved_max_active = max_active;
4178 for_each_pwq(pwq, wq)
4179 pwq_adjust_max_active(pwq);
4181 mutex_unlock(&wq->mutex);
4183 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4186 * current_work - retrieve %current task's work struct
4188 * Determine if %current task is a workqueue worker and what it's working on.
4189 * Useful to find out the context that the %current task is running in.
4191 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4193 struct work_struct *current_work(void)
4195 struct worker *worker = current_wq_worker();
4197 return worker ? worker->current_work : NULL;
4199 EXPORT_SYMBOL(current_work);
4202 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4204 * Determine whether %current is a workqueue rescuer. Can be used from
4205 * work functions to determine whether it's being run off the rescuer task.
4207 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4209 bool current_is_workqueue_rescuer(void)
4211 struct worker *worker = current_wq_worker();
4213 return worker && worker->rescue_wq;
4217 * workqueue_congested - test whether a workqueue is congested
4218 * @cpu: CPU in question
4219 * @wq: target workqueue
4221 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4222 * no synchronization around this function and the test result is
4223 * unreliable and only useful as advisory hints or for debugging.
4225 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4226 * Note that both per-cpu and unbound workqueues may be associated with
4227 * multiple pool_workqueues which have separate congested states. A
4228 * workqueue being congested on one CPU doesn't mean the workqueue is also
4229 * contested on other CPUs / NUMA nodes.
4232 * %true if congested, %false otherwise.
4234 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4236 struct pool_workqueue *pwq;
4239 rcu_read_lock_sched();
4241 if (cpu == WORK_CPU_UNBOUND)
4242 cpu = smp_processor_id();
4244 if (!(wq->flags & WQ_UNBOUND))
4245 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4247 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4249 ret = !list_empty(&pwq->delayed_works);
4250 rcu_read_unlock_sched();
4254 EXPORT_SYMBOL_GPL(workqueue_congested);
4257 * work_busy - test whether a work is currently pending or running
4258 * @work: the work to be tested
4260 * Test whether @work is currently pending or running. There is no
4261 * synchronization around this function and the test result is
4262 * unreliable and only useful as advisory hints or for debugging.
4265 * OR'd bitmask of WORK_BUSY_* bits.
4267 unsigned int work_busy(struct work_struct *work)
4269 struct worker_pool *pool;
4270 unsigned long flags;
4271 unsigned int ret = 0;
4273 if (work_pending(work))
4274 ret |= WORK_BUSY_PENDING;
4276 local_irq_save(flags);
4277 pool = get_work_pool(work);
4279 spin_lock(&pool->lock);
4280 if (find_worker_executing_work(pool, work))
4281 ret |= WORK_BUSY_RUNNING;
4282 spin_unlock(&pool->lock);
4284 local_irq_restore(flags);
4288 EXPORT_SYMBOL_GPL(work_busy);
4291 * set_worker_desc - set description for the current work item
4292 * @fmt: printf-style format string
4293 * @...: arguments for the format string
4295 * This function can be called by a running work function to describe what
4296 * the work item is about. If the worker task gets dumped, this
4297 * information will be printed out together to help debugging. The
4298 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4300 void set_worker_desc(const char *fmt, ...)
4302 struct worker *worker = current_wq_worker();
4306 va_start(args, fmt);
4307 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4309 worker->desc_valid = true;
4314 * print_worker_info - print out worker information and description
4315 * @log_lvl: the log level to use when printing
4316 * @task: target task
4318 * If @task is a worker and currently executing a work item, print out the
4319 * name of the workqueue being serviced and worker description set with
4320 * set_worker_desc() by the currently executing work item.
4322 * This function can be safely called on any task as long as the
4323 * task_struct itself is accessible. While safe, this function isn't
4324 * synchronized and may print out mixups or garbages of limited length.
4326 void print_worker_info(const char *log_lvl, struct task_struct *task)
4328 work_func_t *fn = NULL;
4329 char name[WQ_NAME_LEN] = { };
4330 char desc[WORKER_DESC_LEN] = { };
4331 struct pool_workqueue *pwq = NULL;
4332 struct workqueue_struct *wq = NULL;
4333 bool desc_valid = false;
4334 struct worker *worker;
4336 if (!(task->flags & PF_WQ_WORKER))
4340 * This function is called without any synchronization and @task
4341 * could be in any state. Be careful with dereferences.
4343 worker = kthread_probe_data(task);
4346 * Carefully copy the associated workqueue's workfn and name. Keep
4347 * the original last '\0' in case the original contains garbage.
4349 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4350 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4351 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4352 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4354 /* copy worker description */
4355 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4357 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4359 if (fn || name[0] || desc[0]) {
4360 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4362 pr_cont(" (%s)", desc);
4367 static void pr_cont_pool_info(struct worker_pool *pool)
4369 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4370 if (pool->node != NUMA_NO_NODE)
4371 pr_cont(" node=%d", pool->node);
4372 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4375 static void pr_cont_work(bool comma, struct work_struct *work)
4377 if (work->func == wq_barrier_func) {
4378 struct wq_barrier *barr;
4380 barr = container_of(work, struct wq_barrier, work);
4382 pr_cont("%s BAR(%d)", comma ? "," : "",
4383 task_pid_nr(barr->task));
4385 pr_cont("%s %pf", comma ? "," : "", work->func);
4389 static void show_pwq(struct pool_workqueue *pwq)
4391 struct worker_pool *pool = pwq->pool;
4392 struct work_struct *work;
4393 struct worker *worker;
4394 bool has_in_flight = false, has_pending = false;
4397 pr_info(" pwq %d:", pool->id);
4398 pr_cont_pool_info(pool);
4400 pr_cont(" active=%d/%d refcnt=%d%s\n",
4401 pwq->nr_active, pwq->max_active, pwq->refcnt,
4402 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4404 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4405 if (worker->current_pwq == pwq) {
4406 has_in_flight = true;
4410 if (has_in_flight) {
4413 pr_info(" in-flight:");
4414 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4415 if (worker->current_pwq != pwq)
4418 pr_cont("%s %d%s:%pf", comma ? "," : "",
4419 task_pid_nr(worker->task),
4420 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4421 worker->current_func);
4422 list_for_each_entry(work, &worker->scheduled, entry)
4423 pr_cont_work(false, work);
4429 list_for_each_entry(work, &pool->worklist, entry) {
4430 if (get_work_pwq(work) == pwq) {
4438 pr_info(" pending:");
4439 list_for_each_entry(work, &pool->worklist, entry) {
4440 if (get_work_pwq(work) != pwq)
4443 pr_cont_work(comma, work);
4444 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4449 if (!list_empty(&pwq->delayed_works)) {
4452 pr_info(" delayed:");
4453 list_for_each_entry(work, &pwq->delayed_works, entry) {
4454 pr_cont_work(comma, work);
4455 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4462 * show_workqueue_state - dump workqueue state
4464 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4465 * all busy workqueues and pools.
4467 void show_workqueue_state(void)
4469 struct workqueue_struct *wq;
4470 struct worker_pool *pool;
4471 unsigned long flags;
4474 rcu_read_lock_sched();
4476 pr_info("Showing busy workqueues and worker pools:\n");
4478 list_for_each_entry_rcu(wq, &workqueues, list) {
4479 struct pool_workqueue *pwq;
4482 for_each_pwq(pwq, wq) {
4483 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4491 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4493 for_each_pwq(pwq, wq) {
4494 spin_lock_irqsave(&pwq->pool->lock, flags);
4495 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4497 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4499 * We could be printing a lot from atomic context, e.g.
4500 * sysrq-t -> show_workqueue_state(). Avoid triggering
4503 touch_nmi_watchdog();
4507 for_each_pool(pool, pi) {
4508 struct worker *worker;
4511 spin_lock_irqsave(&pool->lock, flags);
4512 if (pool->nr_workers == pool->nr_idle)
4515 pr_info("pool %d:", pool->id);
4516 pr_cont_pool_info(pool);
4517 pr_cont(" hung=%us workers=%d",
4518 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4521 pr_cont(" manager: %d",
4522 task_pid_nr(pool->manager->task));
4523 list_for_each_entry(worker, &pool->idle_list, entry) {
4524 pr_cont(" %s%d", first ? "idle: " : "",
4525 task_pid_nr(worker->task));
4530 spin_unlock_irqrestore(&pool->lock, flags);
4532 * We could be printing a lot from atomic context, e.g.
4533 * sysrq-t -> show_workqueue_state(). Avoid triggering
4536 touch_nmi_watchdog();
4539 rcu_read_unlock_sched();
4545 * There are two challenges in supporting CPU hotplug. Firstly, there
4546 * are a lot of assumptions on strong associations among work, pwq and
4547 * pool which make migrating pending and scheduled works very
4548 * difficult to implement without impacting hot paths. Secondly,
4549 * worker pools serve mix of short, long and very long running works making
4550 * blocked draining impractical.
4552 * This is solved by allowing the pools to be disassociated from the CPU
4553 * running as an unbound one and allowing it to be reattached later if the
4554 * cpu comes back online.
4557 static void wq_unbind_fn(struct work_struct *work)
4559 int cpu = smp_processor_id();
4560 struct worker_pool *pool;
4561 struct worker *worker;
4563 for_each_cpu_worker_pool(pool, cpu) {
4564 mutex_lock(&pool->attach_mutex);
4565 spin_lock_irq(&pool->lock);
4568 * We've blocked all attach/detach operations. Make all workers
4569 * unbound and set DISASSOCIATED. Before this, all workers
4570 * except for the ones which are still executing works from
4571 * before the last CPU down must be on the cpu. After
4572 * this, they may become diasporas.
4574 for_each_pool_worker(worker, pool)
4575 worker->flags |= WORKER_UNBOUND;
4577 pool->flags |= POOL_DISASSOCIATED;
4579 spin_unlock_irq(&pool->lock);
4580 mutex_unlock(&pool->attach_mutex);
4583 * Call schedule() so that we cross rq->lock and thus can
4584 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4585 * This is necessary as scheduler callbacks may be invoked
4591 * Sched callbacks are disabled now. Zap nr_running.
4592 * After this, nr_running stays zero and need_more_worker()
4593 * and keep_working() are always true as long as the
4594 * worklist is not empty. This pool now behaves as an
4595 * unbound (in terms of concurrency management) pool which
4596 * are served by workers tied to the pool.
4598 atomic_set(&pool->nr_running, 0);
4601 * With concurrency management just turned off, a busy
4602 * worker blocking could lead to lengthy stalls. Kick off
4603 * unbound chain execution of currently pending work items.
4605 spin_lock_irq(&pool->lock);
4606 wake_up_worker(pool);
4607 spin_unlock_irq(&pool->lock);
4612 * rebind_workers - rebind all workers of a pool to the associated CPU
4613 * @pool: pool of interest
4615 * @pool->cpu is coming online. Rebind all workers to the CPU.
4617 static void rebind_workers(struct worker_pool *pool)
4619 struct worker *worker;
4621 lockdep_assert_held(&pool->attach_mutex);
4624 * Restore CPU affinity of all workers. As all idle workers should
4625 * be on the run-queue of the associated CPU before any local
4626 * wake-ups for concurrency management happen, restore CPU affinity
4627 * of all workers first and then clear UNBOUND. As we're called
4628 * from CPU_ONLINE, the following shouldn't fail.
4630 for_each_pool_worker(worker, pool)
4631 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4632 pool->attrs->cpumask) < 0);
4634 spin_lock_irq(&pool->lock);
4637 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4638 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4639 * being reworked and this can go away in time.
4641 if (!(pool->flags & POOL_DISASSOCIATED)) {
4642 spin_unlock_irq(&pool->lock);
4646 pool->flags &= ~POOL_DISASSOCIATED;
4648 for_each_pool_worker(worker, pool) {
4649 unsigned int worker_flags = worker->flags;
4652 * A bound idle worker should actually be on the runqueue
4653 * of the associated CPU for local wake-ups targeting it to
4654 * work. Kick all idle workers so that they migrate to the
4655 * associated CPU. Doing this in the same loop as
4656 * replacing UNBOUND with REBOUND is safe as no worker will
4657 * be bound before @pool->lock is released.
4659 if (worker_flags & WORKER_IDLE)
4660 wake_up_process(worker->task);
4663 * We want to clear UNBOUND but can't directly call
4664 * worker_clr_flags() or adjust nr_running. Atomically
4665 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4666 * @worker will clear REBOUND using worker_clr_flags() when
4667 * it initiates the next execution cycle thus restoring
4668 * concurrency management. Note that when or whether
4669 * @worker clears REBOUND doesn't affect correctness.
4671 * ACCESS_ONCE() is necessary because @worker->flags may be
4672 * tested without holding any lock in
4673 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4674 * fail incorrectly leading to premature concurrency
4675 * management operations.
4677 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4678 worker_flags |= WORKER_REBOUND;
4679 worker_flags &= ~WORKER_UNBOUND;
4680 ACCESS_ONCE(worker->flags) = worker_flags;
4683 spin_unlock_irq(&pool->lock);
4687 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4688 * @pool: unbound pool of interest
4689 * @cpu: the CPU which is coming up
4691 * An unbound pool may end up with a cpumask which doesn't have any online
4692 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4693 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4694 * online CPU before, cpus_allowed of all its workers should be restored.
4696 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4698 static cpumask_t cpumask;
4699 struct worker *worker;
4701 lockdep_assert_held(&pool->attach_mutex);
4703 /* is @cpu allowed for @pool? */
4704 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4707 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4709 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4710 for_each_pool_worker(worker, pool)
4711 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4714 int workqueue_prepare_cpu(unsigned int cpu)
4716 struct worker_pool *pool;
4718 for_each_cpu_worker_pool(pool, cpu) {
4719 if (pool->nr_workers)
4721 if (!create_worker(pool))
4727 int workqueue_online_cpu(unsigned int cpu)
4729 struct worker_pool *pool;
4730 struct workqueue_struct *wq;
4733 mutex_lock(&wq_pool_mutex);
4735 for_each_pool(pool, pi) {
4736 mutex_lock(&pool->attach_mutex);
4738 if (pool->cpu == cpu)
4739 rebind_workers(pool);
4740 else if (pool->cpu < 0)
4741 restore_unbound_workers_cpumask(pool, cpu);
4743 mutex_unlock(&pool->attach_mutex);
4746 /* update NUMA affinity of unbound workqueues */
4747 list_for_each_entry(wq, &workqueues, list)
4748 wq_update_unbound_numa(wq, cpu, true);
4750 mutex_unlock(&wq_pool_mutex);
4754 int workqueue_offline_cpu(unsigned int cpu)
4756 struct work_struct unbind_work;
4757 struct workqueue_struct *wq;
4759 /* unbinding per-cpu workers should happen on the local CPU */
4760 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4761 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4763 /* update NUMA affinity of unbound workqueues */
4764 mutex_lock(&wq_pool_mutex);
4765 list_for_each_entry(wq, &workqueues, list)
4766 wq_update_unbound_numa(wq, cpu, false);
4767 mutex_unlock(&wq_pool_mutex);
4769 /* wait for per-cpu unbinding to finish */
4770 flush_work(&unbind_work);
4771 destroy_work_on_stack(&unbind_work);
4777 struct work_for_cpu {
4778 struct work_struct work;
4784 static void work_for_cpu_fn(struct work_struct *work)
4786 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4788 wfc->ret = wfc->fn(wfc->arg);
4792 * work_on_cpu - run a function in thread context on a particular cpu
4793 * @cpu: the cpu to run on
4794 * @fn: the function to run
4795 * @arg: the function arg
4797 * It is up to the caller to ensure that the cpu doesn't go offline.
4798 * The caller must not hold any locks which would prevent @fn from completing.
4800 * Return: The value @fn returns.
4802 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4804 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4806 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4807 schedule_work_on(cpu, &wfc.work);
4808 flush_work(&wfc.work);
4809 destroy_work_on_stack(&wfc.work);
4812 EXPORT_SYMBOL_GPL(work_on_cpu);
4813 #endif /* CONFIG_SMP */
4815 #ifdef CONFIG_FREEZER
4818 * freeze_workqueues_begin - begin freezing workqueues
4820 * Start freezing workqueues. After this function returns, all freezable
4821 * workqueues will queue new works to their delayed_works list instead of
4825 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4827 void freeze_workqueues_begin(void)
4829 struct workqueue_struct *wq;
4830 struct pool_workqueue *pwq;
4832 mutex_lock(&wq_pool_mutex);
4834 WARN_ON_ONCE(workqueue_freezing);
4835 workqueue_freezing = true;
4837 list_for_each_entry(wq, &workqueues, list) {
4838 mutex_lock(&wq->mutex);
4839 for_each_pwq(pwq, wq)
4840 pwq_adjust_max_active(pwq);
4841 mutex_unlock(&wq->mutex);
4844 mutex_unlock(&wq_pool_mutex);
4848 * freeze_workqueues_busy - are freezable workqueues still busy?
4850 * Check whether freezing is complete. This function must be called
4851 * between freeze_workqueues_begin() and thaw_workqueues().
4854 * Grabs and releases wq_pool_mutex.
4857 * %true if some freezable workqueues are still busy. %false if freezing
4860 bool freeze_workqueues_busy(void)
4863 struct workqueue_struct *wq;
4864 struct pool_workqueue *pwq;
4866 mutex_lock(&wq_pool_mutex);
4868 WARN_ON_ONCE(!workqueue_freezing);
4870 list_for_each_entry(wq, &workqueues, list) {
4871 if (!(wq->flags & WQ_FREEZABLE))
4874 * nr_active is monotonically decreasing. It's safe
4875 * to peek without lock.
4877 rcu_read_lock_sched();
4878 for_each_pwq(pwq, wq) {
4879 WARN_ON_ONCE(pwq->nr_active < 0);
4880 if (pwq->nr_active) {
4882 rcu_read_unlock_sched();
4886 rcu_read_unlock_sched();
4889 mutex_unlock(&wq_pool_mutex);
4894 * thaw_workqueues - thaw workqueues
4896 * Thaw workqueues. Normal queueing is restored and all collected
4897 * frozen works are transferred to their respective pool worklists.
4900 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4902 void thaw_workqueues(void)
4904 struct workqueue_struct *wq;
4905 struct pool_workqueue *pwq;
4907 mutex_lock(&wq_pool_mutex);
4909 if (!workqueue_freezing)
4912 workqueue_freezing = false;
4914 /* restore max_active and repopulate worklist */
4915 list_for_each_entry(wq, &workqueues, list) {
4916 mutex_lock(&wq->mutex);
4917 for_each_pwq(pwq, wq)
4918 pwq_adjust_max_active(pwq);
4919 mutex_unlock(&wq->mutex);
4923 mutex_unlock(&wq_pool_mutex);
4925 #endif /* CONFIG_FREEZER */
4927 static int workqueue_apply_unbound_cpumask(void)
4931 struct workqueue_struct *wq;
4932 struct apply_wqattrs_ctx *ctx, *n;
4934 lockdep_assert_held(&wq_pool_mutex);
4936 list_for_each_entry(wq, &workqueues, list) {
4937 if (!(wq->flags & WQ_UNBOUND))
4939 /* creating multiple pwqs breaks ordering guarantee */
4940 if (wq->flags & __WQ_ORDERED)
4943 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4949 list_add_tail(&ctx->list, &ctxs);
4952 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4954 apply_wqattrs_commit(ctx);
4955 apply_wqattrs_cleanup(ctx);
4962 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4963 * @cpumask: the cpumask to set
4965 * The low-level workqueues cpumask is a global cpumask that limits
4966 * the affinity of all unbound workqueues. This function check the @cpumask
4967 * and apply it to all unbound workqueues and updates all pwqs of them.
4969 * Retun: 0 - Success
4970 * -EINVAL - Invalid @cpumask
4971 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4973 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4976 cpumask_var_t saved_cpumask;
4978 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4981 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4982 if (!cpumask_empty(cpumask)) {
4983 apply_wqattrs_lock();
4985 /* save the old wq_unbound_cpumask. */
4986 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4988 /* update wq_unbound_cpumask at first and apply it to wqs. */
4989 cpumask_copy(wq_unbound_cpumask, cpumask);
4990 ret = workqueue_apply_unbound_cpumask();
4992 /* restore the wq_unbound_cpumask when failed. */
4994 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4996 apply_wqattrs_unlock();
4999 free_cpumask_var(saved_cpumask);
5005 * Workqueues with WQ_SYSFS flag set is visible to userland via
5006 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5007 * following attributes.
5009 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5010 * max_active RW int : maximum number of in-flight work items
5012 * Unbound workqueues have the following extra attributes.
5014 * id RO int : the associated pool ID
5015 * nice RW int : nice value of the workers
5016 * cpumask RW mask : bitmask of allowed CPUs for the workers
5019 struct workqueue_struct *wq;
5023 static struct workqueue_struct *dev_to_wq(struct device *dev)
5025 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5030 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5033 struct workqueue_struct *wq = dev_to_wq(dev);
5035 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5037 static DEVICE_ATTR_RO(per_cpu);
5039 static ssize_t max_active_show(struct device *dev,
5040 struct device_attribute *attr, char *buf)
5042 struct workqueue_struct *wq = dev_to_wq(dev);
5044 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5047 static ssize_t max_active_store(struct device *dev,
5048 struct device_attribute *attr, const char *buf,
5051 struct workqueue_struct *wq = dev_to_wq(dev);
5054 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5057 workqueue_set_max_active(wq, val);
5060 static DEVICE_ATTR_RW(max_active);
5062 static struct attribute *wq_sysfs_attrs[] = {
5063 &dev_attr_per_cpu.attr,
5064 &dev_attr_max_active.attr,
5067 ATTRIBUTE_GROUPS(wq_sysfs);
5069 static ssize_t wq_pool_ids_show(struct device *dev,
5070 struct device_attribute *attr, char *buf)
5072 struct workqueue_struct *wq = dev_to_wq(dev);
5073 const char *delim = "";
5074 int node, written = 0;
5076 rcu_read_lock_sched();
5077 for_each_node(node) {
5078 written += scnprintf(buf + written, PAGE_SIZE - written,
5079 "%s%d:%d", delim, node,
5080 unbound_pwq_by_node(wq, node)->pool->id);
5083 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5084 rcu_read_unlock_sched();
5089 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5092 struct workqueue_struct *wq = dev_to_wq(dev);
5095 mutex_lock(&wq->mutex);
5096 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5097 mutex_unlock(&wq->mutex);
5102 /* prepare workqueue_attrs for sysfs store operations */
5103 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5105 struct workqueue_attrs *attrs;
5107 lockdep_assert_held(&wq_pool_mutex);
5109 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5113 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5117 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5118 const char *buf, size_t count)
5120 struct workqueue_struct *wq = dev_to_wq(dev);
5121 struct workqueue_attrs *attrs;
5124 apply_wqattrs_lock();
5126 attrs = wq_sysfs_prep_attrs(wq);
5130 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5131 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5132 ret = apply_workqueue_attrs_locked(wq, attrs);
5137 apply_wqattrs_unlock();
5138 free_workqueue_attrs(attrs);
5139 return ret ?: count;
5142 static ssize_t wq_cpumask_show(struct device *dev,
5143 struct device_attribute *attr, char *buf)
5145 struct workqueue_struct *wq = dev_to_wq(dev);
5148 mutex_lock(&wq->mutex);
5149 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5150 cpumask_pr_args(wq->unbound_attrs->cpumask));
5151 mutex_unlock(&wq->mutex);
5155 static ssize_t wq_cpumask_store(struct device *dev,
5156 struct device_attribute *attr,
5157 const char *buf, size_t count)
5159 struct workqueue_struct *wq = dev_to_wq(dev);
5160 struct workqueue_attrs *attrs;
5163 apply_wqattrs_lock();
5165 attrs = wq_sysfs_prep_attrs(wq);
5169 ret = cpumask_parse(buf, attrs->cpumask);
5171 ret = apply_workqueue_attrs_locked(wq, attrs);
5174 apply_wqattrs_unlock();
5175 free_workqueue_attrs(attrs);
5176 return ret ?: count;
5179 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5182 struct workqueue_struct *wq = dev_to_wq(dev);
5185 mutex_lock(&wq->mutex);
5186 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5187 !wq->unbound_attrs->no_numa);
5188 mutex_unlock(&wq->mutex);
5193 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5194 const char *buf, size_t count)
5196 struct workqueue_struct *wq = dev_to_wq(dev);
5197 struct workqueue_attrs *attrs;
5198 int v, ret = -ENOMEM;
5200 apply_wqattrs_lock();
5202 attrs = wq_sysfs_prep_attrs(wq);
5207 if (sscanf(buf, "%d", &v) == 1) {
5208 attrs->no_numa = !v;
5209 ret = apply_workqueue_attrs_locked(wq, attrs);
5213 apply_wqattrs_unlock();
5214 free_workqueue_attrs(attrs);
5215 return ret ?: count;
5218 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5219 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5220 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5221 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5222 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5226 static struct bus_type wq_subsys = {
5227 .name = "workqueue",
5228 .dev_groups = wq_sysfs_groups,
5231 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5232 struct device_attribute *attr, char *buf)
5236 mutex_lock(&wq_pool_mutex);
5237 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5238 cpumask_pr_args(wq_unbound_cpumask));
5239 mutex_unlock(&wq_pool_mutex);
5244 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5245 struct device_attribute *attr, const char *buf, size_t count)
5247 cpumask_var_t cpumask;
5250 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5253 ret = cpumask_parse(buf, cpumask);
5255 ret = workqueue_set_unbound_cpumask(cpumask);
5257 free_cpumask_var(cpumask);
5258 return ret ? ret : count;
5261 static struct device_attribute wq_sysfs_cpumask_attr =
5262 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5263 wq_unbound_cpumask_store);
5265 static int __init wq_sysfs_init(void)
5269 err = subsys_virtual_register(&wq_subsys, NULL);
5273 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5275 core_initcall(wq_sysfs_init);
5277 static void wq_device_release(struct device *dev)
5279 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5285 * workqueue_sysfs_register - make a workqueue visible in sysfs
5286 * @wq: the workqueue to register
5288 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5289 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5290 * which is the preferred method.
5292 * Workqueue user should use this function directly iff it wants to apply
5293 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5294 * apply_workqueue_attrs() may race against userland updating the
5297 * Return: 0 on success, -errno on failure.
5299 int workqueue_sysfs_register(struct workqueue_struct *wq)
5301 struct wq_device *wq_dev;
5305 * Adjusting max_active or creating new pwqs by applying
5306 * attributes breaks ordering guarantee. Disallow exposing ordered
5309 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5312 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5317 wq_dev->dev.bus = &wq_subsys;
5318 wq_dev->dev.release = wq_device_release;
5319 dev_set_name(&wq_dev->dev, "%s", wq->name);
5322 * unbound_attrs are created separately. Suppress uevent until
5323 * everything is ready.
5325 dev_set_uevent_suppress(&wq_dev->dev, true);
5327 ret = device_register(&wq_dev->dev);
5329 put_device(&wq_dev->dev);
5334 if (wq->flags & WQ_UNBOUND) {
5335 struct device_attribute *attr;
5337 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5338 ret = device_create_file(&wq_dev->dev, attr);
5340 device_unregister(&wq_dev->dev);
5347 dev_set_uevent_suppress(&wq_dev->dev, false);
5348 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5353 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5354 * @wq: the workqueue to unregister
5356 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5358 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5360 struct wq_device *wq_dev = wq->wq_dev;
5366 device_unregister(&wq_dev->dev);
5368 #else /* CONFIG_SYSFS */
5369 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5370 #endif /* CONFIG_SYSFS */
5373 * Workqueue watchdog.
5375 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5376 * flush dependency, a concurrency managed work item which stays RUNNING
5377 * indefinitely. Workqueue stalls can be very difficult to debug as the
5378 * usual warning mechanisms don't trigger and internal workqueue state is
5381 * Workqueue watchdog monitors all worker pools periodically and dumps
5382 * state if some pools failed to make forward progress for a while where
5383 * forward progress is defined as the first item on ->worklist changing.
5385 * This mechanism is controlled through the kernel parameter
5386 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5387 * corresponding sysfs parameter file.
5389 #ifdef CONFIG_WQ_WATCHDOG
5391 static void wq_watchdog_timer_fn(unsigned long data);
5393 static unsigned long wq_watchdog_thresh = 30;
5394 static struct timer_list wq_watchdog_timer =
5395 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5397 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5398 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5400 static void wq_watchdog_reset_touched(void)
5404 wq_watchdog_touched = jiffies;
5405 for_each_possible_cpu(cpu)
5406 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5409 static void wq_watchdog_timer_fn(unsigned long data)
5411 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5412 bool lockup_detected = false;
5413 unsigned long now = jiffies;
5414 struct worker_pool *pool;
5422 for_each_pool(pool, pi) {
5423 unsigned long pool_ts, touched, ts;
5425 if (list_empty(&pool->worklist))
5429 * If a virtual machine is stopped by the host it can look to
5430 * the watchdog like a stall.
5432 kvm_check_and_clear_guest_paused();
5434 /* get the latest of pool and touched timestamps */
5435 pool_ts = READ_ONCE(pool->watchdog_ts);
5436 touched = READ_ONCE(wq_watchdog_touched);
5438 if (time_after(pool_ts, touched))
5443 if (pool->cpu >= 0) {
5444 unsigned long cpu_touched =
5445 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5447 if (time_after(cpu_touched, ts))
5452 if (time_after(now, ts + thresh)) {
5453 lockup_detected = true;
5454 pr_emerg("BUG: workqueue lockup - pool");
5455 pr_cont_pool_info(pool);
5456 pr_cont(" stuck for %us!\n",
5457 jiffies_to_msecs(now - pool_ts) / 1000);
5463 if (lockup_detected)
5464 show_workqueue_state();
5466 wq_watchdog_reset_touched();
5467 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5470 void wq_watchdog_touch(int cpu)
5473 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5475 wq_watchdog_touched = jiffies;
5478 static void wq_watchdog_set_thresh(unsigned long thresh)
5480 wq_watchdog_thresh = 0;
5481 del_timer_sync(&wq_watchdog_timer);
5484 wq_watchdog_thresh = thresh;
5485 wq_watchdog_reset_touched();
5486 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5490 static int wq_watchdog_param_set_thresh(const char *val,
5491 const struct kernel_param *kp)
5493 unsigned long thresh;
5496 ret = kstrtoul(val, 0, &thresh);
5501 wq_watchdog_set_thresh(thresh);
5503 wq_watchdog_thresh = thresh;
5508 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5509 .set = wq_watchdog_param_set_thresh,
5510 .get = param_get_ulong,
5513 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5516 static void wq_watchdog_init(void)
5518 wq_watchdog_set_thresh(wq_watchdog_thresh);
5521 #else /* CONFIG_WQ_WATCHDOG */
5523 static inline void wq_watchdog_init(void) { }
5525 #endif /* CONFIG_WQ_WATCHDOG */
5527 static void __init wq_numa_init(void)
5532 if (num_possible_nodes() <= 1)
5535 if (wq_disable_numa) {
5536 pr_info("workqueue: NUMA affinity support disabled\n");
5540 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5541 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5544 * We want masks of possible CPUs of each node which isn't readily
5545 * available. Build one from cpu_to_node() which should have been
5546 * fully initialized by now.
5548 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5552 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5553 node_online(node) ? node : NUMA_NO_NODE));
5555 for_each_possible_cpu(cpu) {
5556 node = cpu_to_node(cpu);
5557 if (WARN_ON(node == NUMA_NO_NODE)) {
5558 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5559 /* happens iff arch is bonkers, let's just proceed */
5562 cpumask_set_cpu(cpu, tbl[node]);
5565 wq_numa_possible_cpumask = tbl;
5566 wq_numa_enabled = true;
5570 * workqueue_init_early - early init for workqueue subsystem
5572 * This is the first half of two-staged workqueue subsystem initialization
5573 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5574 * idr are up. It sets up all the data structures and system workqueues
5575 * and allows early boot code to create workqueues and queue/cancel work
5576 * items. Actual work item execution starts only after kthreads can be
5577 * created and scheduled right before early initcalls.
5579 int __init workqueue_init_early(void)
5581 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5584 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5586 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5587 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5589 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5593 /* initialize CPU pools */
5594 for_each_possible_cpu(cpu) {
5595 struct worker_pool *pool;
5598 for_each_cpu_worker_pool(pool, cpu) {
5599 BUG_ON(init_worker_pool(pool));
5601 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5602 pool->attrs->nice = std_nice[i++];
5603 pool->node = cpu_to_node(cpu);
5606 mutex_lock(&wq_pool_mutex);
5607 BUG_ON(worker_pool_assign_id(pool));
5608 mutex_unlock(&wq_pool_mutex);
5612 /* create default unbound and ordered wq attrs */
5613 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5614 struct workqueue_attrs *attrs;
5616 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5617 attrs->nice = std_nice[i];
5618 unbound_std_wq_attrs[i] = attrs;
5621 * An ordered wq should have only one pwq as ordering is
5622 * guaranteed by max_active which is enforced by pwqs.
5623 * Turn off NUMA so that dfl_pwq is used for all nodes.
5625 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5626 attrs->nice = std_nice[i];
5627 attrs->no_numa = true;
5628 ordered_wq_attrs[i] = attrs;
5631 system_wq = alloc_workqueue("events", 0, 0);
5632 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5633 system_long_wq = alloc_workqueue("events_long", 0, 0);
5634 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5635 WQ_UNBOUND_MAX_ACTIVE);
5636 system_freezable_wq = alloc_workqueue("events_freezable",
5638 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5639 WQ_POWER_EFFICIENT, 0);
5640 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5641 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5643 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5644 !system_unbound_wq || !system_freezable_wq ||
5645 !system_power_efficient_wq ||
5646 !system_freezable_power_efficient_wq);
5652 * workqueue_init - bring workqueue subsystem fully online
5654 * This is the latter half of two-staged workqueue subsystem initialization
5655 * and invoked as soon as kthreads can be created and scheduled.
5656 * Workqueues have been created and work items queued on them, but there
5657 * are no kworkers executing the work items yet. Populate the worker pools
5658 * with the initial workers and enable future kworker creations.
5660 int __init workqueue_init(void)
5662 struct worker_pool *pool;
5665 /* create the initial workers */
5666 for_each_online_cpu(cpu) {
5667 for_each_cpu_worker_pool(pool, cpu) {
5668 pool->flags &= ~POOL_DISASSOCIATED;
5669 BUG_ON(!create_worker(pool));
5673 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5674 BUG_ON(!create_worker(pool));