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/core-api/workqueue.rst 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 static 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 inline struct pool_workqueue *work_struct_pwq(unsigned long data)
687 return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
690 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
692 unsigned long data = atomic_long_read(&work->data);
694 if (data & WORK_STRUCT_PWQ)
695 return work_struct_pwq(data);
701 * get_work_pool - return the worker_pool a given work was associated with
702 * @work: the work item of interest
704 * Pools are created and destroyed under wq_pool_mutex, and allows read
705 * access under sched-RCU read lock. As such, this function should be
706 * called under wq_pool_mutex or with preemption disabled.
708 * All fields of the returned pool are accessible as long as the above
709 * mentioned locking is in effect. If the returned pool needs to be used
710 * beyond the critical section, the caller is responsible for ensuring the
711 * returned pool is and stays online.
713 * Return: The worker_pool @work was last associated with. %NULL if none.
715 static struct worker_pool *get_work_pool(struct work_struct *work)
717 unsigned long data = atomic_long_read(&work->data);
720 assert_rcu_or_pool_mutex();
722 if (data & WORK_STRUCT_PWQ)
723 return work_struct_pwq(data)->pool;
725 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
726 if (pool_id == WORK_OFFQ_POOL_NONE)
729 return idr_find(&worker_pool_idr, pool_id);
733 * get_work_pool_id - return the worker pool ID a given work is associated with
734 * @work: the work item of interest
736 * Return: The worker_pool ID @work was last associated with.
737 * %WORK_OFFQ_POOL_NONE if none.
739 static int get_work_pool_id(struct work_struct *work)
741 unsigned long data = atomic_long_read(&work->data);
743 if (data & WORK_STRUCT_PWQ)
744 return work_struct_pwq(data)->pool->id;
746 return data >> WORK_OFFQ_POOL_SHIFT;
749 static void mark_work_canceling(struct work_struct *work)
751 unsigned long pool_id = get_work_pool_id(work);
753 pool_id <<= WORK_OFFQ_POOL_SHIFT;
754 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
757 static bool work_is_canceling(struct work_struct *work)
759 unsigned long data = atomic_long_read(&work->data);
761 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
765 * Policy functions. These define the policies on how the global worker
766 * pools are managed. Unless noted otherwise, these functions assume that
767 * they're being called with pool->lock held.
770 static bool __need_more_worker(struct worker_pool *pool)
772 return !atomic_read(&pool->nr_running);
776 * Need to wake up a worker? Called from anything but currently
779 * Note that, because unbound workers never contribute to nr_running, this
780 * function will always return %true for unbound pools as long as the
781 * worklist isn't empty.
783 static bool need_more_worker(struct worker_pool *pool)
785 return !list_empty(&pool->worklist) && __need_more_worker(pool);
788 /* Can I start working? Called from busy but !running workers. */
789 static bool may_start_working(struct worker_pool *pool)
791 return pool->nr_idle;
794 /* Do I need to keep working? Called from currently running workers. */
795 static bool keep_working(struct worker_pool *pool)
797 return !list_empty(&pool->worklist) &&
798 atomic_read(&pool->nr_running) <= 1;
801 /* Do we need a new worker? Called from manager. */
802 static bool need_to_create_worker(struct worker_pool *pool)
804 return need_more_worker(pool) && !may_start_working(pool);
807 /* Do we have too many workers and should some go away? */
808 static bool too_many_workers(struct worker_pool *pool)
810 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
811 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
812 int nr_busy = pool->nr_workers - nr_idle;
814 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
821 /* Return the first idle worker. Safe with preemption disabled */
822 static struct worker *first_idle_worker(struct worker_pool *pool)
824 if (unlikely(list_empty(&pool->idle_list)))
827 return list_first_entry(&pool->idle_list, struct worker, entry);
831 * wake_up_worker - wake up an idle worker
832 * @pool: worker pool to wake worker from
834 * Wake up the first idle worker of @pool.
837 * spin_lock_irq(pool->lock).
839 static void wake_up_worker(struct worker_pool *pool)
841 struct worker *worker = first_idle_worker(pool);
844 wake_up_process(worker->task);
848 * wq_worker_waking_up - a worker is waking up
849 * @task: task waking up
850 * @cpu: CPU @task is waking up to
852 * This function is called during try_to_wake_up() when a worker is
856 * spin_lock_irq(rq->lock)
858 void wq_worker_waking_up(struct task_struct *task, int cpu)
860 struct worker *worker = kthread_data(task);
862 if (!(worker->flags & WORKER_NOT_RUNNING)) {
863 WARN_ON_ONCE(worker->pool->cpu != cpu);
864 atomic_inc(&worker->pool->nr_running);
869 * wq_worker_sleeping - a worker is going to sleep
870 * @task: task going to sleep
872 * This function is called during schedule() when a busy worker is
873 * going to sleep. Worker on the same cpu can be woken up by
874 * returning pointer to its task.
877 * spin_lock_irq(rq->lock)
880 * Worker task on @cpu to wake up, %NULL if none.
882 struct task_struct *wq_worker_sleeping(struct task_struct *task)
884 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
885 struct worker_pool *pool;
888 * Rescuers, which may not have all the fields set up like normal
889 * workers, also reach here, let's not access anything before
890 * checking NOT_RUNNING.
892 if (worker->flags & WORKER_NOT_RUNNING)
897 /* this can only happen on the local cpu */
898 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
902 * The counterpart of the following dec_and_test, implied mb,
903 * worklist not empty test sequence is in insert_work().
904 * Please read comment there.
906 * NOT_RUNNING is clear. This means that we're bound to and
907 * running on the local cpu w/ rq lock held and preemption
908 * disabled, which in turn means that none else could be
909 * manipulating idle_list, so dereferencing idle_list without pool
912 if (atomic_dec_and_test(&pool->nr_running) &&
913 !list_empty(&pool->worklist))
914 to_wakeup = first_idle_worker(pool);
915 return to_wakeup ? to_wakeup->task : NULL;
919 * worker_set_flags - set worker flags and adjust nr_running accordingly
921 * @flags: flags to set
923 * Set @flags in @worker->flags and adjust nr_running accordingly.
926 * spin_lock_irq(pool->lock)
928 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
930 struct worker_pool *pool = worker->pool;
932 WARN_ON_ONCE(worker->task != current);
934 /* If transitioning into NOT_RUNNING, adjust nr_running. */
935 if ((flags & WORKER_NOT_RUNNING) &&
936 !(worker->flags & WORKER_NOT_RUNNING)) {
937 atomic_dec(&pool->nr_running);
940 worker->flags |= flags;
944 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
946 * @flags: flags to clear
948 * Clear @flags in @worker->flags and adjust nr_running accordingly.
951 * spin_lock_irq(pool->lock)
953 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
955 struct worker_pool *pool = worker->pool;
956 unsigned int oflags = worker->flags;
958 WARN_ON_ONCE(worker->task != current);
960 worker->flags &= ~flags;
963 * If transitioning out of NOT_RUNNING, increment nr_running. Note
964 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
965 * of multiple flags, not a single flag.
967 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
968 if (!(worker->flags & WORKER_NOT_RUNNING))
969 atomic_inc(&pool->nr_running);
973 * find_worker_executing_work - find worker which is executing a work
974 * @pool: pool of interest
975 * @work: work to find worker for
977 * Find a worker which is executing @work on @pool by searching
978 * @pool->busy_hash which is keyed by the address of @work. For a worker
979 * to match, its current execution should match the address of @work and
980 * its work function. This is to avoid unwanted dependency between
981 * unrelated work executions through a work item being recycled while still
984 * This is a bit tricky. A work item may be freed once its execution
985 * starts and nothing prevents the freed area from being recycled for
986 * another work item. If the same work item address ends up being reused
987 * before the original execution finishes, workqueue will identify the
988 * recycled work item as currently executing and make it wait until the
989 * current execution finishes, introducing an unwanted dependency.
991 * This function checks the work item address and work function to avoid
992 * false positives. Note that this isn't complete as one may construct a
993 * work function which can introduce dependency onto itself through a
994 * recycled work item. Well, if somebody wants to shoot oneself in the
995 * foot that badly, there's only so much we can do, and if such deadlock
996 * actually occurs, it should be easy to locate the culprit work function.
999 * spin_lock_irq(pool->lock).
1002 * Pointer to worker which is executing @work if found, %NULL
1005 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1006 struct work_struct *work)
1008 struct worker *worker;
1010 hash_for_each_possible(pool->busy_hash, worker, hentry,
1011 (unsigned long)work)
1012 if (worker->current_work == work &&
1013 worker->current_func == work->func)
1020 * move_linked_works - move linked works to a list
1021 * @work: start of series of works to be scheduled
1022 * @head: target list to append @work to
1023 * @nextp: out parameter for nested worklist walking
1025 * Schedule linked works starting from @work to @head. Work series to
1026 * be scheduled starts at @work and includes any consecutive work with
1027 * WORK_STRUCT_LINKED set in its predecessor.
1029 * If @nextp is not NULL, it's updated to point to the next work of
1030 * the last scheduled work. This allows move_linked_works() to be
1031 * nested inside outer list_for_each_entry_safe().
1034 * spin_lock_irq(pool->lock).
1036 static void move_linked_works(struct work_struct *work, struct list_head *head,
1037 struct work_struct **nextp)
1039 struct work_struct *n;
1042 * Linked worklist will always end before the end of the list,
1043 * use NULL for list head.
1045 list_for_each_entry_safe_from(work, n, NULL, entry) {
1046 list_move_tail(&work->entry, head);
1047 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1052 * If we're already inside safe list traversal and have moved
1053 * multiple works to the scheduled queue, the next position
1054 * needs to be updated.
1061 * get_pwq - get an extra reference on the specified pool_workqueue
1062 * @pwq: pool_workqueue to get
1064 * Obtain an extra reference on @pwq. The caller should guarantee that
1065 * @pwq has positive refcnt and be holding the matching pool->lock.
1067 static void get_pwq(struct pool_workqueue *pwq)
1069 lockdep_assert_held(&pwq->pool->lock);
1070 WARN_ON_ONCE(pwq->refcnt <= 0);
1075 * put_pwq - put a pool_workqueue reference
1076 * @pwq: pool_workqueue to put
1078 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1079 * destruction. The caller should be holding the matching pool->lock.
1081 static void put_pwq(struct pool_workqueue *pwq)
1083 lockdep_assert_held(&pwq->pool->lock);
1084 if (likely(--pwq->refcnt))
1086 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1089 * @pwq can't be released under pool->lock, bounce to
1090 * pwq_unbound_release_workfn(). This never recurses on the same
1091 * pool->lock as this path is taken only for unbound workqueues and
1092 * the release work item is scheduled on a per-cpu workqueue. To
1093 * avoid lockdep warning, unbound pool->locks are given lockdep
1094 * subclass of 1 in get_unbound_pool().
1096 schedule_work(&pwq->unbound_release_work);
1100 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1101 * @pwq: pool_workqueue to put (can be %NULL)
1103 * put_pwq() with locking. This function also allows %NULL @pwq.
1105 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1109 * As both pwqs and pools are sched-RCU protected, the
1110 * following lock operations are safe.
1112 spin_lock_irq(&pwq->pool->lock);
1114 spin_unlock_irq(&pwq->pool->lock);
1118 static void pwq_activate_delayed_work(struct work_struct *work)
1120 struct pool_workqueue *pwq = get_work_pwq(work);
1122 trace_workqueue_activate_work(work);
1123 if (list_empty(&pwq->pool->worklist))
1124 pwq->pool->watchdog_ts = jiffies;
1125 move_linked_works(work, &pwq->pool->worklist, NULL);
1126 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1130 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1132 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1133 struct work_struct, entry);
1135 pwq_activate_delayed_work(work);
1139 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1140 * @pwq: pwq of interest
1141 * @color: color of work which left the queue
1143 * A work either has completed or is removed from pending queue,
1144 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1147 * spin_lock_irq(pool->lock).
1149 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1151 /* uncolored work items don't participate in flushing or nr_active */
1152 if (color == WORK_NO_COLOR)
1155 pwq->nr_in_flight[color]--;
1158 if (!list_empty(&pwq->delayed_works)) {
1159 /* one down, submit a delayed one */
1160 if (pwq->nr_active < pwq->max_active)
1161 pwq_activate_first_delayed(pwq);
1164 /* is flush in progress and are we at the flushing tip? */
1165 if (likely(pwq->flush_color != color))
1168 /* are there still in-flight works? */
1169 if (pwq->nr_in_flight[color])
1172 /* this pwq is done, clear flush_color */
1173 pwq->flush_color = -1;
1176 * If this was the last pwq, wake up the first flusher. It
1177 * will handle the rest.
1179 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1180 complete(&pwq->wq->first_flusher->done);
1186 * try_to_grab_pending - steal work item from worklist and disable irq
1187 * @work: work item to steal
1188 * @is_dwork: @work is a delayed_work
1189 * @flags: place to store irq state
1191 * Try to grab PENDING bit of @work. This function can handle @work in any
1192 * stable state - idle, on timer or on worklist.
1195 * 1 if @work was pending and we successfully stole PENDING
1196 * 0 if @work was idle and we claimed PENDING
1197 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1198 * -ENOENT if someone else is canceling @work, this state may persist
1199 * for arbitrarily long
1202 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1203 * interrupted while holding PENDING and @work off queue, irq must be
1204 * disabled on entry. This, combined with delayed_work->timer being
1205 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1207 * On successful return, >= 0, irq is disabled and the caller is
1208 * responsible for releasing it using local_irq_restore(*@flags).
1210 * This function is safe to call from any context including IRQ handler.
1212 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1213 unsigned long *flags)
1215 struct worker_pool *pool;
1216 struct pool_workqueue *pwq;
1218 local_irq_save(*flags);
1220 /* try to steal the timer if it exists */
1222 struct delayed_work *dwork = to_delayed_work(work);
1225 * dwork->timer is irqsafe. If del_timer() fails, it's
1226 * guaranteed that the timer is not queued anywhere and not
1227 * running on the local CPU.
1229 if (likely(del_timer(&dwork->timer)))
1233 /* try to claim PENDING the normal way */
1234 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1238 * The queueing is in progress, or it is already queued. Try to
1239 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1241 pool = get_work_pool(work);
1245 spin_lock(&pool->lock);
1247 * work->data is guaranteed to point to pwq only while the work
1248 * item is queued on pwq->wq, and both updating work->data to point
1249 * to pwq on queueing and to pool on dequeueing are done under
1250 * pwq->pool->lock. This in turn guarantees that, if work->data
1251 * points to pwq which is associated with a locked pool, the work
1252 * item is currently queued on that pool.
1254 pwq = get_work_pwq(work);
1255 if (pwq && pwq->pool == pool) {
1256 debug_work_deactivate(work);
1259 * A delayed work item cannot be grabbed directly because
1260 * it might have linked NO_COLOR work items which, if left
1261 * on the delayed_list, will confuse pwq->nr_active
1262 * management later on and cause stall. Make sure the work
1263 * item is activated before grabbing.
1265 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1266 pwq_activate_delayed_work(work);
1268 list_del_init(&work->entry);
1269 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1271 /* work->data points to pwq iff queued, point to pool */
1272 set_work_pool_and_keep_pending(work, pool->id);
1274 spin_unlock(&pool->lock);
1277 spin_unlock(&pool->lock);
1279 local_irq_restore(*flags);
1280 if (work_is_canceling(work))
1287 * insert_work - insert a work into a pool
1288 * @pwq: pwq @work belongs to
1289 * @work: work to insert
1290 * @head: insertion point
1291 * @extra_flags: extra WORK_STRUCT_* flags to set
1293 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1294 * work_struct flags.
1297 * spin_lock_irq(pool->lock).
1299 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1300 struct list_head *head, unsigned int extra_flags)
1302 struct worker_pool *pool = pwq->pool;
1304 /* we own @work, set data and link */
1305 set_work_pwq(work, pwq, extra_flags);
1306 list_add_tail(&work->entry, head);
1310 * Ensure either wq_worker_sleeping() sees the above
1311 * list_add_tail() or we see zero nr_running to avoid workers lying
1312 * around lazily while there are works to be processed.
1316 if (__need_more_worker(pool))
1317 wake_up_worker(pool);
1321 * Test whether @work is being queued from another work executing on the
1324 static bool is_chained_work(struct workqueue_struct *wq)
1326 struct worker *worker;
1328 worker = current_wq_worker();
1330 * Return %true iff I'm a worker execuing a work item on @wq. If
1331 * I'm @worker, it's safe to dereference it without locking.
1333 return worker && worker->current_pwq->wq == wq;
1337 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1338 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1339 * avoid perturbing sensitive tasks.
1341 static int wq_select_unbound_cpu(int cpu)
1343 static bool printed_dbg_warning;
1346 if (likely(!wq_debug_force_rr_cpu)) {
1347 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1349 } else if (!printed_dbg_warning) {
1350 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1351 printed_dbg_warning = true;
1354 if (cpumask_empty(wq_unbound_cpumask))
1357 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1358 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1359 if (unlikely(new_cpu >= nr_cpu_ids)) {
1360 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1361 if (unlikely(new_cpu >= nr_cpu_ids))
1364 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1369 static void __queue_work(int cpu, struct workqueue_struct *wq,
1370 struct work_struct *work)
1372 struct pool_workqueue *pwq;
1373 struct worker_pool *last_pool;
1374 struct list_head *worklist;
1375 unsigned int work_flags;
1376 unsigned int req_cpu = cpu;
1379 * While a work item is PENDING && off queue, a task trying to
1380 * steal the PENDING will busy-loop waiting for it to either get
1381 * queued or lose PENDING. Grabbing PENDING and queueing should
1382 * happen with IRQ disabled.
1384 WARN_ON_ONCE(!irqs_disabled());
1387 /* if draining, only works from the same workqueue are allowed */
1388 if (unlikely(wq->flags & __WQ_DRAINING) &&
1389 WARN_ON_ONCE(!is_chained_work(wq)))
1392 /* pwq which will be used unless @work is executing elsewhere */
1393 if (wq->flags & WQ_UNBOUND) {
1394 if (req_cpu == WORK_CPU_UNBOUND)
1395 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1396 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1398 if (req_cpu == WORK_CPU_UNBOUND)
1399 cpu = raw_smp_processor_id();
1400 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1404 * If @work was previously on a different pool, it might still be
1405 * running there, in which case the work needs to be queued on that
1406 * pool to guarantee non-reentrancy.
1408 last_pool = get_work_pool(work);
1409 if (last_pool && last_pool != pwq->pool) {
1410 struct worker *worker;
1412 spin_lock(&last_pool->lock);
1414 worker = find_worker_executing_work(last_pool, work);
1416 if (worker && worker->current_pwq->wq == wq) {
1417 pwq = worker->current_pwq;
1419 /* meh... not running there, queue here */
1420 spin_unlock(&last_pool->lock);
1421 spin_lock(&pwq->pool->lock);
1424 spin_lock(&pwq->pool->lock);
1428 * pwq is determined and locked. For unbound pools, we could have
1429 * raced with pwq release and it could already be dead. If its
1430 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1431 * without another pwq replacing it in the numa_pwq_tbl or while
1432 * work items are executing on it, so the retrying is guaranteed to
1433 * make forward-progress.
1435 if (unlikely(!pwq->refcnt)) {
1436 if (wq->flags & WQ_UNBOUND) {
1437 spin_unlock(&pwq->pool->lock);
1442 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1446 /* pwq determined, queue */
1447 trace_workqueue_queue_work(req_cpu, pwq, work);
1449 if (WARN_ON(!list_empty(&work->entry))) {
1450 spin_unlock(&pwq->pool->lock);
1454 pwq->nr_in_flight[pwq->work_color]++;
1455 work_flags = work_color_to_flags(pwq->work_color);
1457 if (likely(pwq->nr_active < pwq->max_active)) {
1458 trace_workqueue_activate_work(work);
1460 worklist = &pwq->pool->worklist;
1461 if (list_empty(worklist))
1462 pwq->pool->watchdog_ts = jiffies;
1464 work_flags |= WORK_STRUCT_DELAYED;
1465 worklist = &pwq->delayed_works;
1468 debug_work_activate(work);
1469 insert_work(pwq, work, worklist, work_flags);
1471 spin_unlock(&pwq->pool->lock);
1475 * queue_work_on - queue work on specific cpu
1476 * @cpu: CPU number to execute work on
1477 * @wq: workqueue to use
1478 * @work: work to queue
1480 * We queue the work to a specific CPU, the caller must ensure it
1483 * Return: %false if @work was already on a queue, %true otherwise.
1485 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1486 struct work_struct *work)
1489 unsigned long flags;
1491 local_irq_save(flags);
1493 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1494 __queue_work(cpu, wq, work);
1498 local_irq_restore(flags);
1501 EXPORT_SYMBOL(queue_work_on);
1503 void delayed_work_timer_fn(unsigned long __data)
1505 struct delayed_work *dwork = (struct delayed_work *)__data;
1507 /* should have been called from irqsafe timer with irq already off */
1508 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1510 EXPORT_SYMBOL(delayed_work_timer_fn);
1512 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1513 struct delayed_work *dwork, unsigned long delay)
1515 struct timer_list *timer = &dwork->timer;
1516 struct work_struct *work = &dwork->work;
1519 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1520 timer->data != (unsigned long)dwork);
1521 WARN_ON_ONCE(timer_pending(timer));
1522 WARN_ON_ONCE(!list_empty(&work->entry));
1525 * If @delay is 0, queue @dwork->work immediately. This is for
1526 * both optimization and correctness. The earliest @timer can
1527 * expire is on the closest next tick and delayed_work users depend
1528 * on that there's no such delay when @delay is 0.
1531 __queue_work(cpu, wq, &dwork->work);
1537 timer->expires = jiffies + delay;
1539 if (unlikely(cpu != WORK_CPU_UNBOUND))
1540 add_timer_on(timer, cpu);
1546 * queue_delayed_work_on - queue work on specific CPU after delay
1547 * @cpu: CPU number to execute work on
1548 * @wq: workqueue to use
1549 * @dwork: work to queue
1550 * @delay: number of jiffies to wait before queueing
1552 * Return: %false if @work was already on a queue, %true otherwise. If
1553 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1556 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1557 struct delayed_work *dwork, unsigned long delay)
1559 struct work_struct *work = &dwork->work;
1561 unsigned long flags;
1563 /* read the comment in __queue_work() */
1564 local_irq_save(flags);
1566 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1567 __queue_delayed_work(cpu, wq, dwork, delay);
1571 local_irq_restore(flags);
1574 EXPORT_SYMBOL(queue_delayed_work_on);
1577 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1578 * @cpu: CPU number to execute work on
1579 * @wq: workqueue to use
1580 * @dwork: work to queue
1581 * @delay: number of jiffies to wait before queueing
1583 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1584 * modify @dwork's timer so that it expires after @delay. If @delay is
1585 * zero, @work is guaranteed to be scheduled immediately regardless of its
1588 * Return: %false if @dwork was idle and queued, %true if @dwork was
1589 * pending and its timer was modified.
1591 * This function is safe to call from any context including IRQ handler.
1592 * See try_to_grab_pending() for details.
1594 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1595 struct delayed_work *dwork, unsigned long delay)
1597 unsigned long flags;
1601 ret = try_to_grab_pending(&dwork->work, true, &flags);
1602 } while (unlikely(ret == -EAGAIN));
1604 if (likely(ret >= 0)) {
1605 __queue_delayed_work(cpu, wq, dwork, delay);
1606 local_irq_restore(flags);
1609 /* -ENOENT from try_to_grab_pending() becomes %true */
1612 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1615 * worker_enter_idle - enter idle state
1616 * @worker: worker which is entering idle state
1618 * @worker is entering idle state. Update stats and idle timer if
1622 * spin_lock_irq(pool->lock).
1624 static void worker_enter_idle(struct worker *worker)
1626 struct worker_pool *pool = worker->pool;
1628 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1629 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1630 (worker->hentry.next || worker->hentry.pprev)))
1633 /* can't use worker_set_flags(), also called from create_worker() */
1634 worker->flags |= WORKER_IDLE;
1636 worker->last_active = jiffies;
1638 /* idle_list is LIFO */
1639 list_add(&worker->entry, &pool->idle_list);
1641 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1642 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1645 * Sanity check nr_running. Because wq_unbind_fn() releases
1646 * pool->lock between setting %WORKER_UNBOUND and zapping
1647 * nr_running, the warning may trigger spuriously. Check iff
1648 * unbind is not in progress.
1650 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1651 pool->nr_workers == pool->nr_idle &&
1652 atomic_read(&pool->nr_running));
1656 * worker_leave_idle - leave idle state
1657 * @worker: worker which is leaving idle state
1659 * @worker is leaving idle state. Update stats.
1662 * spin_lock_irq(pool->lock).
1664 static void worker_leave_idle(struct worker *worker)
1666 struct worker_pool *pool = worker->pool;
1668 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1670 worker_clr_flags(worker, WORKER_IDLE);
1672 list_del_init(&worker->entry);
1675 static struct worker *alloc_worker(int node)
1677 struct worker *worker;
1679 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1681 INIT_LIST_HEAD(&worker->entry);
1682 INIT_LIST_HEAD(&worker->scheduled);
1683 INIT_LIST_HEAD(&worker->node);
1684 /* on creation a worker is in !idle && prep state */
1685 worker->flags = WORKER_PREP;
1691 * worker_attach_to_pool() - attach a worker to a pool
1692 * @worker: worker to be attached
1693 * @pool: the target pool
1695 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1696 * cpu-binding of @worker are kept coordinated with the pool across
1699 static void worker_attach_to_pool(struct worker *worker,
1700 struct worker_pool *pool)
1702 mutex_lock(&pool->attach_mutex);
1705 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1706 * online CPUs. It'll be re-applied when any of the CPUs come up.
1708 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1711 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1712 * stable across this function. See the comments above the
1713 * flag definition for details.
1715 if (pool->flags & POOL_DISASSOCIATED)
1716 worker->flags |= WORKER_UNBOUND;
1718 list_add_tail(&worker->node, &pool->workers);
1720 mutex_unlock(&pool->attach_mutex);
1724 * worker_detach_from_pool() - detach a worker from its pool
1725 * @worker: worker which is attached to its pool
1726 * @pool: the pool @worker is attached to
1728 * Undo the attaching which had been done in worker_attach_to_pool(). The
1729 * caller worker shouldn't access to the pool after detached except it has
1730 * other reference to the pool.
1732 static void worker_detach_from_pool(struct worker *worker,
1733 struct worker_pool *pool)
1735 struct completion *detach_completion = NULL;
1737 mutex_lock(&pool->attach_mutex);
1738 list_del(&worker->node);
1739 if (list_empty(&pool->workers))
1740 detach_completion = pool->detach_completion;
1741 mutex_unlock(&pool->attach_mutex);
1743 /* clear leftover flags without pool->lock after it is detached */
1744 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1746 if (detach_completion)
1747 complete(detach_completion);
1751 * create_worker - create a new workqueue worker
1752 * @pool: pool the new worker will belong to
1754 * Create and start a new worker which is attached to @pool.
1757 * Might sleep. Does GFP_KERNEL allocations.
1760 * Pointer to the newly created worker.
1762 static struct worker *create_worker(struct worker_pool *pool)
1764 struct worker *worker = NULL;
1768 /* ID is needed to determine kthread name */
1769 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1773 worker = alloc_worker(pool->node);
1777 worker->pool = pool;
1781 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1782 pool->attrs->nice < 0 ? "H" : "");
1784 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1786 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1787 "kworker/%s", id_buf);
1788 if (IS_ERR(worker->task))
1791 set_user_nice(worker->task, pool->attrs->nice);
1792 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1794 /* successful, attach the worker to the pool */
1795 worker_attach_to_pool(worker, pool);
1797 /* start the newly created worker */
1798 spin_lock_irq(&pool->lock);
1799 worker->pool->nr_workers++;
1800 worker_enter_idle(worker);
1801 wake_up_process(worker->task);
1802 spin_unlock_irq(&pool->lock);
1808 ida_simple_remove(&pool->worker_ida, id);
1814 * destroy_worker - destroy a workqueue worker
1815 * @worker: worker to be destroyed
1817 * Destroy @worker and adjust @pool stats accordingly. The worker should
1821 * spin_lock_irq(pool->lock).
1823 static void destroy_worker(struct worker *worker)
1825 struct worker_pool *pool = worker->pool;
1827 lockdep_assert_held(&pool->lock);
1829 /* sanity check frenzy */
1830 if (WARN_ON(worker->current_work) ||
1831 WARN_ON(!list_empty(&worker->scheduled)) ||
1832 WARN_ON(!(worker->flags & WORKER_IDLE)))
1838 list_del_init(&worker->entry);
1839 worker->flags |= WORKER_DIE;
1840 wake_up_process(worker->task);
1843 static void idle_worker_timeout(unsigned long __pool)
1845 struct worker_pool *pool = (void *)__pool;
1847 spin_lock_irq(&pool->lock);
1849 while (too_many_workers(pool)) {
1850 struct worker *worker;
1851 unsigned long expires;
1853 /* idle_list is kept in LIFO order, check the last one */
1854 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1855 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1857 if (time_before(jiffies, expires)) {
1858 mod_timer(&pool->idle_timer, expires);
1862 destroy_worker(worker);
1865 spin_unlock_irq(&pool->lock);
1868 static void send_mayday(struct work_struct *work)
1870 struct pool_workqueue *pwq = get_work_pwq(work);
1871 struct workqueue_struct *wq = pwq->wq;
1873 lockdep_assert_held(&wq_mayday_lock);
1878 /* mayday mayday mayday */
1879 if (list_empty(&pwq->mayday_node)) {
1881 * If @pwq is for an unbound wq, its base ref may be put at
1882 * any time due to an attribute change. Pin @pwq until the
1883 * rescuer is done with it.
1886 list_add_tail(&pwq->mayday_node, &wq->maydays);
1887 wake_up_process(wq->rescuer->task);
1891 static void pool_mayday_timeout(unsigned long __pool)
1893 struct worker_pool *pool = (void *)__pool;
1894 struct work_struct *work;
1896 spin_lock_irq(&pool->lock);
1897 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1899 if (need_to_create_worker(pool)) {
1901 * We've been trying to create a new worker but
1902 * haven't been successful. We might be hitting an
1903 * allocation deadlock. Send distress signals to
1906 list_for_each_entry(work, &pool->worklist, entry)
1910 spin_unlock(&wq_mayday_lock);
1911 spin_unlock_irq(&pool->lock);
1913 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1917 * maybe_create_worker - create a new worker if necessary
1918 * @pool: pool to create a new worker for
1920 * Create a new worker for @pool if necessary. @pool is guaranteed to
1921 * have at least one idle worker on return from this function. If
1922 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1923 * sent to all rescuers with works scheduled on @pool to resolve
1924 * possible allocation deadlock.
1926 * On return, need_to_create_worker() is guaranteed to be %false and
1927 * may_start_working() %true.
1930 * spin_lock_irq(pool->lock) which may be released and regrabbed
1931 * multiple times. Does GFP_KERNEL allocations. Called only from
1934 static void maybe_create_worker(struct worker_pool *pool)
1935 __releases(&pool->lock)
1936 __acquires(&pool->lock)
1939 spin_unlock_irq(&pool->lock);
1941 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1942 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1945 if (create_worker(pool) || !need_to_create_worker(pool))
1948 schedule_timeout_interruptible(CREATE_COOLDOWN);
1950 if (!need_to_create_worker(pool))
1954 del_timer_sync(&pool->mayday_timer);
1955 spin_lock_irq(&pool->lock);
1957 * This is necessary even after a new worker was just successfully
1958 * created as @pool->lock was dropped and the new worker might have
1959 * already become busy.
1961 if (need_to_create_worker(pool))
1966 * manage_workers - manage worker pool
1969 * Assume the manager role and manage the worker pool @worker belongs
1970 * to. At any given time, there can be only zero or one manager per
1971 * pool. The exclusion is handled automatically by this function.
1973 * The caller can safely start processing works on false return. On
1974 * true return, it's guaranteed that need_to_create_worker() is false
1975 * and may_start_working() is true.
1978 * spin_lock_irq(pool->lock) which may be released and regrabbed
1979 * multiple times. Does GFP_KERNEL allocations.
1982 * %false if the pool doesn't need management and the caller can safely
1983 * start processing works, %true if management function was performed and
1984 * the conditions that the caller verified before calling the function may
1985 * no longer be true.
1987 static bool manage_workers(struct worker *worker)
1989 struct worker_pool *pool = worker->pool;
1991 if (pool->flags & POOL_MANAGER_ACTIVE)
1994 pool->flags |= POOL_MANAGER_ACTIVE;
1995 pool->manager = worker;
1997 maybe_create_worker(pool);
1999 pool->manager = NULL;
2000 pool->flags &= ~POOL_MANAGER_ACTIVE;
2001 wake_up(&wq_manager_wait);
2006 * process_one_work - process single work
2008 * @work: work to process
2010 * Process @work. This function contains all the logics necessary to
2011 * process a single work including synchronization against and
2012 * interaction with other workers on the same cpu, queueing and
2013 * flushing. As long as context requirement is met, any worker can
2014 * call this function to process a work.
2017 * spin_lock_irq(pool->lock) which is released and regrabbed.
2019 static void process_one_work(struct worker *worker, struct work_struct *work)
2020 __releases(&pool->lock)
2021 __acquires(&pool->lock)
2023 struct pool_workqueue *pwq = get_work_pwq(work);
2024 struct worker_pool *pool = worker->pool;
2025 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2027 struct worker *collision;
2028 #ifdef CONFIG_LOCKDEP
2030 * It is permissible to free the struct work_struct from
2031 * inside the function that is called from it, this we need to
2032 * take into account for lockdep too. To avoid bogus "held
2033 * lock freed" warnings as well as problems when looking into
2034 * work->lockdep_map, make a copy and use that here.
2036 struct lockdep_map lockdep_map;
2038 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2040 /* ensure we're on the correct CPU */
2041 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2042 raw_smp_processor_id() != pool->cpu);
2045 * A single work shouldn't be executed concurrently by
2046 * multiple workers on a single cpu. Check whether anyone is
2047 * already processing the work. If so, defer the work to the
2048 * currently executing one.
2050 collision = find_worker_executing_work(pool, work);
2051 if (unlikely(collision)) {
2052 move_linked_works(work, &collision->scheduled, NULL);
2056 /* claim and dequeue */
2057 debug_work_deactivate(work);
2058 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2059 worker->current_work = work;
2060 worker->current_func = work->func;
2061 worker->current_pwq = pwq;
2062 work_color = get_work_color(work);
2064 list_del_init(&work->entry);
2067 * CPU intensive works don't participate in concurrency management.
2068 * They're the scheduler's responsibility. This takes @worker out
2069 * of concurrency management and the next code block will chain
2070 * execution of the pending work items.
2072 if (unlikely(cpu_intensive))
2073 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2076 * Wake up another worker if necessary. The condition is always
2077 * false for normal per-cpu workers since nr_running would always
2078 * be >= 1 at this point. This is used to chain execution of the
2079 * pending work items for WORKER_NOT_RUNNING workers such as the
2080 * UNBOUND and CPU_INTENSIVE ones.
2082 if (need_more_worker(pool))
2083 wake_up_worker(pool);
2086 * Record the last pool and clear PENDING which should be the last
2087 * update to @work. Also, do this inside @pool->lock so that
2088 * PENDING and queued state changes happen together while IRQ is
2091 set_work_pool_and_clear_pending(work, pool->id);
2093 spin_unlock_irq(&pool->lock);
2095 lock_map_acquire(&pwq->wq->lockdep_map);
2096 lock_map_acquire(&lockdep_map);
2098 * Strictly speaking we should mark the invariant state without holding
2099 * any locks, that is, before these two lock_map_acquire()'s.
2101 * However, that would result in:
2108 * Which would create W1->C->W1 dependencies, even though there is no
2109 * actual deadlock possible. There are two solutions, using a
2110 * read-recursive acquire on the work(queue) 'locks', but this will then
2111 * hit the lockdep limitation on recursive locks, or simply discard
2114 * AFAICT there is no possible deadlock scenario between the
2115 * flush_work() and complete() primitives (except for single-threaded
2116 * workqueues), so hiding them isn't a problem.
2118 lockdep_invariant_state(true);
2119 trace_workqueue_execute_start(work);
2120 worker->current_func(work);
2122 * While we must be careful to not use "work" after this, the trace
2123 * point will only record its address.
2125 trace_workqueue_execute_end(work);
2126 lock_map_release(&lockdep_map);
2127 lock_map_release(&pwq->wq->lockdep_map);
2129 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2130 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2131 " last function: %pf\n",
2132 current->comm, preempt_count(), task_pid_nr(current),
2133 worker->current_func);
2134 debug_show_held_locks(current);
2139 * The following prevents a kworker from hogging CPU on !PREEMPT
2140 * kernels, where a requeueing work item waiting for something to
2141 * happen could deadlock with stop_machine as such work item could
2142 * indefinitely requeue itself while all other CPUs are trapped in
2143 * stop_machine. At the same time, report a quiescent RCU state so
2144 * the same condition doesn't freeze RCU.
2146 cond_resched_rcu_qs();
2148 spin_lock_irq(&pool->lock);
2150 /* clear cpu intensive status */
2151 if (unlikely(cpu_intensive))
2152 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2154 /* we're done with it, release */
2155 hash_del(&worker->hentry);
2156 worker->current_work = NULL;
2157 worker->current_func = NULL;
2158 worker->current_pwq = NULL;
2159 worker->desc_valid = false;
2160 pwq_dec_nr_in_flight(pwq, work_color);
2164 * process_scheduled_works - process scheduled works
2167 * Process all scheduled works. Please note that the scheduled list
2168 * may change while processing a work, so this function repeatedly
2169 * fetches a work from the top and executes it.
2172 * spin_lock_irq(pool->lock) which may be released and regrabbed
2175 static void process_scheduled_works(struct worker *worker)
2177 while (!list_empty(&worker->scheduled)) {
2178 struct work_struct *work = list_first_entry(&worker->scheduled,
2179 struct work_struct, entry);
2180 process_one_work(worker, work);
2185 * worker_thread - the worker thread function
2188 * The worker thread function. All workers belong to a worker_pool -
2189 * either a per-cpu one or dynamic unbound one. These workers process all
2190 * work items regardless of their specific target workqueue. The only
2191 * exception is work items which belong to workqueues with a rescuer which
2192 * will be explained in rescuer_thread().
2196 static int worker_thread(void *__worker)
2198 struct worker *worker = __worker;
2199 struct worker_pool *pool = worker->pool;
2201 /* tell the scheduler that this is a workqueue worker */
2202 worker->task->flags |= PF_WQ_WORKER;
2204 spin_lock_irq(&pool->lock);
2206 /* am I supposed to die? */
2207 if (unlikely(worker->flags & WORKER_DIE)) {
2208 spin_unlock_irq(&pool->lock);
2209 WARN_ON_ONCE(!list_empty(&worker->entry));
2210 worker->task->flags &= ~PF_WQ_WORKER;
2212 set_task_comm(worker->task, "kworker/dying");
2213 ida_simple_remove(&pool->worker_ida, worker->id);
2214 worker_detach_from_pool(worker, pool);
2219 worker_leave_idle(worker);
2221 /* no more worker necessary? */
2222 if (!need_more_worker(pool))
2225 /* do we need to manage? */
2226 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2230 * ->scheduled list can only be filled while a worker is
2231 * preparing to process a work or actually processing it.
2232 * Make sure nobody diddled with it while I was sleeping.
2234 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2237 * Finish PREP stage. We're guaranteed to have at least one idle
2238 * worker or that someone else has already assumed the manager
2239 * role. This is where @worker starts participating in concurrency
2240 * management if applicable and concurrency management is restored
2241 * after being rebound. See rebind_workers() for details.
2243 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2246 struct work_struct *work =
2247 list_first_entry(&pool->worklist,
2248 struct work_struct, entry);
2250 pool->watchdog_ts = jiffies;
2252 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2253 /* optimization path, not strictly necessary */
2254 process_one_work(worker, work);
2255 if (unlikely(!list_empty(&worker->scheduled)))
2256 process_scheduled_works(worker);
2258 move_linked_works(work, &worker->scheduled, NULL);
2259 process_scheduled_works(worker);
2261 } while (keep_working(pool));
2263 worker_set_flags(worker, WORKER_PREP);
2266 * pool->lock is held and there's no work to process and no need to
2267 * manage, sleep. Workers are woken up only while holding
2268 * pool->lock or from local cpu, so setting the current state
2269 * before releasing pool->lock is enough to prevent losing any
2272 worker_enter_idle(worker);
2273 __set_current_state(TASK_IDLE);
2274 spin_unlock_irq(&pool->lock);
2280 * rescuer_thread - the rescuer thread function
2283 * Workqueue rescuer thread function. There's one rescuer for each
2284 * workqueue which has WQ_MEM_RECLAIM set.
2286 * Regular work processing on a pool may block trying to create a new
2287 * worker which uses GFP_KERNEL allocation which has slight chance of
2288 * developing into deadlock if some works currently on the same queue
2289 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2290 * the problem rescuer solves.
2292 * When such condition is possible, the pool summons rescuers of all
2293 * workqueues which have works queued on the pool and let them process
2294 * those works so that forward progress can be guaranteed.
2296 * This should happen rarely.
2300 static int rescuer_thread(void *__rescuer)
2302 struct worker *rescuer = __rescuer;
2303 struct workqueue_struct *wq = rescuer->rescue_wq;
2304 struct list_head *scheduled = &rescuer->scheduled;
2307 set_user_nice(current, RESCUER_NICE_LEVEL);
2310 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2311 * doesn't participate in concurrency management.
2313 rescuer->task->flags |= PF_WQ_WORKER;
2315 set_current_state(TASK_IDLE);
2318 * By the time the rescuer is requested to stop, the workqueue
2319 * shouldn't have any work pending, but @wq->maydays may still have
2320 * pwq(s) queued. This can happen by non-rescuer workers consuming
2321 * all the work items before the rescuer got to them. Go through
2322 * @wq->maydays processing before acting on should_stop so that the
2323 * list is always empty on exit.
2325 should_stop = kthread_should_stop();
2327 /* see whether any pwq is asking for help */
2328 spin_lock_irq(&wq_mayday_lock);
2330 while (!list_empty(&wq->maydays)) {
2331 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2332 struct pool_workqueue, mayday_node);
2333 struct worker_pool *pool = pwq->pool;
2334 struct work_struct *work, *n;
2337 __set_current_state(TASK_RUNNING);
2338 list_del_init(&pwq->mayday_node);
2340 spin_unlock_irq(&wq_mayday_lock);
2342 worker_attach_to_pool(rescuer, pool);
2344 spin_lock_irq(&pool->lock);
2345 rescuer->pool = pool;
2348 * Slurp in all works issued via this workqueue and
2351 WARN_ON_ONCE(!list_empty(scheduled));
2352 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2353 if (get_work_pwq(work) == pwq) {
2355 pool->watchdog_ts = jiffies;
2356 move_linked_works(work, scheduled, &n);
2361 if (!list_empty(scheduled)) {
2362 process_scheduled_works(rescuer);
2365 * The above execution of rescued work items could
2366 * have created more to rescue through
2367 * pwq_activate_first_delayed() or chained
2368 * queueing. Let's put @pwq back on mayday list so
2369 * that such back-to-back work items, which may be
2370 * being used to relieve memory pressure, don't
2371 * incur MAYDAY_INTERVAL delay inbetween.
2373 if (need_to_create_worker(pool)) {
2374 spin_lock(&wq_mayday_lock);
2376 * Queue iff we aren't racing destruction
2377 * and somebody else hasn't queued it already.
2379 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2381 list_add_tail(&pwq->mayday_node, &wq->maydays);
2383 spin_unlock(&wq_mayday_lock);
2388 * Put the reference grabbed by send_mayday(). @pool won't
2389 * go away while we're still attached to it.
2394 * Leave this pool. If need_more_worker() is %true, notify a
2395 * regular worker; otherwise, we end up with 0 concurrency
2396 * and stalling the execution.
2398 if (need_more_worker(pool))
2399 wake_up_worker(pool);
2401 rescuer->pool = NULL;
2402 spin_unlock_irq(&pool->lock);
2404 worker_detach_from_pool(rescuer, pool);
2406 spin_lock_irq(&wq_mayday_lock);
2409 spin_unlock_irq(&wq_mayday_lock);
2412 __set_current_state(TASK_RUNNING);
2413 rescuer->task->flags &= ~PF_WQ_WORKER;
2417 /* rescuers should never participate in concurrency management */
2418 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2424 * check_flush_dependency - check for flush dependency sanity
2425 * @target_wq: workqueue being flushed
2426 * @target_work: work item being flushed (NULL for workqueue flushes)
2428 * %current is trying to flush the whole @target_wq or @target_work on it.
2429 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2430 * reclaiming memory or running on a workqueue which doesn't have
2431 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2434 static void check_flush_dependency(struct workqueue_struct *target_wq,
2435 struct work_struct *target_work)
2437 work_func_t target_func = target_work ? target_work->func : NULL;
2438 struct worker *worker;
2440 if (target_wq->flags & WQ_MEM_RECLAIM)
2443 worker = current_wq_worker();
2445 WARN_ONCE(current->flags & PF_MEMALLOC,
2446 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2447 current->pid, current->comm, target_wq->name, target_func);
2448 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2449 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2450 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2451 worker->current_pwq->wq->name, worker->current_func,
2452 target_wq->name, target_func);
2456 struct work_struct work;
2457 struct completion done;
2458 struct task_struct *task; /* purely informational */
2461 static void wq_barrier_func(struct work_struct *work)
2463 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2464 complete(&barr->done);
2468 * insert_wq_barrier - insert a barrier work
2469 * @pwq: pwq to insert barrier into
2470 * @barr: wq_barrier to insert
2471 * @target: target work to attach @barr to
2472 * @worker: worker currently executing @target, NULL if @target is not executing
2474 * @barr is linked to @target such that @barr is completed only after
2475 * @target finishes execution. Please note that the ordering
2476 * guarantee is observed only with respect to @target and on the local
2479 * Currently, a queued barrier can't be canceled. This is because
2480 * try_to_grab_pending() can't determine whether the work to be
2481 * grabbed is at the head of the queue and thus can't clear LINKED
2482 * flag of the previous work while there must be a valid next work
2483 * after a work with LINKED flag set.
2485 * Note that when @worker is non-NULL, @target may be modified
2486 * underneath us, so we can't reliably determine pwq from @target.
2489 * spin_lock_irq(pool->lock).
2491 static void insert_wq_barrier(struct pool_workqueue *pwq,
2492 struct wq_barrier *barr,
2493 struct work_struct *target, struct worker *worker)
2495 struct list_head *head;
2496 unsigned int linked = 0;
2499 * debugobject calls are safe here even with pool->lock locked
2500 * as we know for sure that this will not trigger any of the
2501 * checks and call back into the fixup functions where we
2504 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2505 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2508 * Explicitly init the crosslock for wq_barrier::done, make its lock
2509 * key a subkey of the corresponding work. As a result we won't
2510 * build a dependency between wq_barrier::done and unrelated work.
2512 lockdep_init_map_crosslock((struct lockdep_map *)&barr->done.map,
2513 "(complete)wq_barr::done",
2514 target->lockdep_map.key, 1);
2515 __init_completion(&barr->done);
2516 barr->task = current;
2519 * If @target is currently being executed, schedule the
2520 * barrier to the worker; otherwise, put it after @target.
2523 head = worker->scheduled.next;
2525 unsigned long *bits = work_data_bits(target);
2527 head = target->entry.next;
2528 /* there can already be other linked works, inherit and set */
2529 linked = *bits & WORK_STRUCT_LINKED;
2530 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2533 debug_work_activate(&barr->work);
2534 insert_work(pwq, &barr->work, head,
2535 work_color_to_flags(WORK_NO_COLOR) | linked);
2539 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2540 * @wq: workqueue being flushed
2541 * @flush_color: new flush color, < 0 for no-op
2542 * @work_color: new work color, < 0 for no-op
2544 * Prepare pwqs for workqueue flushing.
2546 * If @flush_color is non-negative, flush_color on all pwqs should be
2547 * -1. If no pwq has in-flight commands at the specified color, all
2548 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2549 * has in flight commands, its pwq->flush_color is set to
2550 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2551 * wakeup logic is armed and %true is returned.
2553 * The caller should have initialized @wq->first_flusher prior to
2554 * calling this function with non-negative @flush_color. If
2555 * @flush_color is negative, no flush color update is done and %false
2558 * If @work_color is non-negative, all pwqs should have the same
2559 * work_color which is previous to @work_color and all will be
2560 * advanced to @work_color.
2563 * mutex_lock(wq->mutex).
2566 * %true if @flush_color >= 0 and there's something to flush. %false
2569 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2570 int flush_color, int work_color)
2573 struct pool_workqueue *pwq;
2575 if (flush_color >= 0) {
2576 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2577 atomic_set(&wq->nr_pwqs_to_flush, 1);
2580 for_each_pwq(pwq, wq) {
2581 struct worker_pool *pool = pwq->pool;
2583 spin_lock_irq(&pool->lock);
2585 if (flush_color >= 0) {
2586 WARN_ON_ONCE(pwq->flush_color != -1);
2588 if (pwq->nr_in_flight[flush_color]) {
2589 pwq->flush_color = flush_color;
2590 atomic_inc(&wq->nr_pwqs_to_flush);
2595 if (work_color >= 0) {
2596 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2597 pwq->work_color = work_color;
2600 spin_unlock_irq(&pool->lock);
2603 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2604 complete(&wq->first_flusher->done);
2610 * flush_workqueue - ensure that any scheduled work has run to completion.
2611 * @wq: workqueue to flush
2613 * This function sleeps until all work items which were queued on entry
2614 * have finished execution, but it is not livelocked by new incoming ones.
2616 void flush_workqueue(struct workqueue_struct *wq)
2618 struct wq_flusher this_flusher = {
2619 .list = LIST_HEAD_INIT(this_flusher.list),
2621 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2625 if (WARN_ON(!wq_online))
2628 lock_map_acquire(&wq->lockdep_map);
2629 lock_map_release(&wq->lockdep_map);
2631 mutex_lock(&wq->mutex);
2634 * Start-to-wait phase
2636 next_color = work_next_color(wq->work_color);
2638 if (next_color != wq->flush_color) {
2640 * Color space is not full. The current work_color
2641 * becomes our flush_color and work_color is advanced
2644 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2645 this_flusher.flush_color = wq->work_color;
2646 wq->work_color = next_color;
2648 if (!wq->first_flusher) {
2649 /* no flush in progress, become the first flusher */
2650 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2652 wq->first_flusher = &this_flusher;
2654 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2656 /* nothing to flush, done */
2657 wq->flush_color = next_color;
2658 wq->first_flusher = NULL;
2663 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2664 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2665 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2669 * Oops, color space is full, wait on overflow queue.
2670 * The next flush completion will assign us
2671 * flush_color and transfer to flusher_queue.
2673 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2676 check_flush_dependency(wq, NULL);
2678 mutex_unlock(&wq->mutex);
2680 wait_for_completion(&this_flusher.done);
2683 * Wake-up-and-cascade phase
2685 * First flushers are responsible for cascading flushes and
2686 * handling overflow. Non-first flushers can simply return.
2688 if (wq->first_flusher != &this_flusher)
2691 mutex_lock(&wq->mutex);
2693 /* we might have raced, check again with mutex held */
2694 if (wq->first_flusher != &this_flusher)
2697 wq->first_flusher = NULL;
2699 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2700 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2703 struct wq_flusher *next, *tmp;
2705 /* complete all the flushers sharing the current flush color */
2706 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2707 if (next->flush_color != wq->flush_color)
2709 list_del_init(&next->list);
2710 complete(&next->done);
2713 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2714 wq->flush_color != work_next_color(wq->work_color));
2716 /* this flush_color is finished, advance by one */
2717 wq->flush_color = work_next_color(wq->flush_color);
2719 /* one color has been freed, handle overflow queue */
2720 if (!list_empty(&wq->flusher_overflow)) {
2722 * Assign the same color to all overflowed
2723 * flushers, advance work_color and append to
2724 * flusher_queue. This is the start-to-wait
2725 * phase for these overflowed flushers.
2727 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2728 tmp->flush_color = wq->work_color;
2730 wq->work_color = work_next_color(wq->work_color);
2732 list_splice_tail_init(&wq->flusher_overflow,
2733 &wq->flusher_queue);
2734 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2737 if (list_empty(&wq->flusher_queue)) {
2738 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2743 * Need to flush more colors. Make the next flusher
2744 * the new first flusher and arm pwqs.
2746 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2747 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2749 list_del_init(&next->list);
2750 wq->first_flusher = next;
2752 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2756 * Meh... this color is already done, clear first
2757 * flusher and repeat cascading.
2759 wq->first_flusher = NULL;
2763 mutex_unlock(&wq->mutex);
2765 EXPORT_SYMBOL(flush_workqueue);
2768 * drain_workqueue - drain a workqueue
2769 * @wq: workqueue to drain
2771 * Wait until the workqueue becomes empty. While draining is in progress,
2772 * only chain queueing is allowed. IOW, only currently pending or running
2773 * work items on @wq can queue further work items on it. @wq is flushed
2774 * repeatedly until it becomes empty. The number of flushing is determined
2775 * by the depth of chaining and should be relatively short. Whine if it
2778 void drain_workqueue(struct workqueue_struct *wq)
2780 unsigned int flush_cnt = 0;
2781 struct pool_workqueue *pwq;
2784 * __queue_work() needs to test whether there are drainers, is much
2785 * hotter than drain_workqueue() and already looks at @wq->flags.
2786 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2788 mutex_lock(&wq->mutex);
2789 if (!wq->nr_drainers++)
2790 wq->flags |= __WQ_DRAINING;
2791 mutex_unlock(&wq->mutex);
2793 flush_workqueue(wq);
2795 mutex_lock(&wq->mutex);
2797 for_each_pwq(pwq, wq) {
2800 spin_lock_irq(&pwq->pool->lock);
2801 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2802 spin_unlock_irq(&pwq->pool->lock);
2807 if (++flush_cnt == 10 ||
2808 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2809 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2810 wq->name, flush_cnt);
2812 mutex_unlock(&wq->mutex);
2816 if (!--wq->nr_drainers)
2817 wq->flags &= ~__WQ_DRAINING;
2818 mutex_unlock(&wq->mutex);
2820 EXPORT_SYMBOL_GPL(drain_workqueue);
2822 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2824 struct worker *worker = NULL;
2825 struct worker_pool *pool;
2826 struct pool_workqueue *pwq;
2830 local_irq_disable();
2831 pool = get_work_pool(work);
2837 spin_lock(&pool->lock);
2838 /* see the comment in try_to_grab_pending() with the same code */
2839 pwq = get_work_pwq(work);
2841 if (unlikely(pwq->pool != pool))
2844 worker = find_worker_executing_work(pool, work);
2847 pwq = worker->current_pwq;
2850 check_flush_dependency(pwq->wq, work);
2852 insert_wq_barrier(pwq, barr, work, worker);
2853 spin_unlock_irq(&pool->lock);
2856 * Force a lock recursion deadlock when using flush_work() inside a
2857 * single-threaded or rescuer equipped workqueue.
2859 * For single threaded workqueues the deadlock happens when the work
2860 * is after the work issuing the flush_work(). For rescuer equipped
2861 * workqueues the deadlock happens when the rescuer stalls, blocking
2864 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) {
2865 lock_map_acquire(&pwq->wq->lockdep_map);
2866 lock_map_release(&pwq->wq->lockdep_map);
2871 spin_unlock_irq(&pool->lock);
2876 * flush_work - wait for a work to finish executing the last queueing instance
2877 * @work: the work to flush
2879 * Wait until @work has finished execution. @work is guaranteed to be idle
2880 * on return if it hasn't been requeued since flush started.
2883 * %true if flush_work() waited for the work to finish execution,
2884 * %false if it was already idle.
2886 bool flush_work(struct work_struct *work)
2888 struct wq_barrier barr;
2890 if (WARN_ON(!wq_online))
2893 lock_map_acquire(&work->lockdep_map);
2894 lock_map_release(&work->lockdep_map);
2896 if (start_flush_work(work, &barr)) {
2897 wait_for_completion(&barr.done);
2898 destroy_work_on_stack(&barr.work);
2904 EXPORT_SYMBOL_GPL(flush_work);
2907 wait_queue_entry_t wait;
2908 struct work_struct *work;
2911 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2913 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2915 if (cwait->work != key)
2917 return autoremove_wake_function(wait, mode, sync, key);
2920 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2922 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2923 unsigned long flags;
2927 ret = try_to_grab_pending(work, is_dwork, &flags);
2929 * If someone else is already canceling, wait for it to
2930 * finish. flush_work() doesn't work for PREEMPT_NONE
2931 * because we may get scheduled between @work's completion
2932 * and the other canceling task resuming and clearing
2933 * CANCELING - flush_work() will return false immediately
2934 * as @work is no longer busy, try_to_grab_pending() will
2935 * return -ENOENT as @work is still being canceled and the
2936 * other canceling task won't be able to clear CANCELING as
2937 * we're hogging the CPU.
2939 * Let's wait for completion using a waitqueue. As this
2940 * may lead to the thundering herd problem, use a custom
2941 * wake function which matches @work along with exclusive
2944 if (unlikely(ret == -ENOENT)) {
2945 struct cwt_wait cwait;
2947 init_wait(&cwait.wait);
2948 cwait.wait.func = cwt_wakefn;
2951 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2952 TASK_UNINTERRUPTIBLE);
2953 if (work_is_canceling(work))
2955 finish_wait(&cancel_waitq, &cwait.wait);
2957 } while (unlikely(ret < 0));
2959 /* tell other tasks trying to grab @work to back off */
2960 mark_work_canceling(work);
2961 local_irq_restore(flags);
2964 * This allows canceling during early boot. We know that @work
2970 clear_work_data(work);
2973 * Paired with prepare_to_wait() above so that either
2974 * waitqueue_active() is visible here or !work_is_canceling() is
2978 if (waitqueue_active(&cancel_waitq))
2979 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2985 * cancel_work_sync - cancel a work and wait for it to finish
2986 * @work: the work to cancel
2988 * Cancel @work and wait for its execution to finish. This function
2989 * can be used even if the work re-queues itself or migrates to
2990 * another workqueue. On return from this function, @work is
2991 * guaranteed to be not pending or executing on any CPU.
2993 * cancel_work_sync(&delayed_work->work) must not be used for
2994 * delayed_work's. Use cancel_delayed_work_sync() instead.
2996 * The caller must ensure that the workqueue on which @work was last
2997 * queued can't be destroyed before this function returns.
3000 * %true if @work was pending, %false otherwise.
3002 bool cancel_work_sync(struct work_struct *work)
3004 return __cancel_work_timer(work, false);
3006 EXPORT_SYMBOL_GPL(cancel_work_sync);
3009 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3010 * @dwork: the delayed work to flush
3012 * Delayed timer is cancelled and the pending work is queued for
3013 * immediate execution. Like flush_work(), this function only
3014 * considers the last queueing instance of @dwork.
3017 * %true if flush_work() waited for the work to finish execution,
3018 * %false if it was already idle.
3020 bool flush_delayed_work(struct delayed_work *dwork)
3022 local_irq_disable();
3023 if (del_timer_sync(&dwork->timer))
3024 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3026 return flush_work(&dwork->work);
3028 EXPORT_SYMBOL(flush_delayed_work);
3030 static bool __cancel_work(struct work_struct *work, bool is_dwork)
3032 unsigned long flags;
3036 ret = try_to_grab_pending(work, is_dwork, &flags);
3037 } while (unlikely(ret == -EAGAIN));
3039 if (unlikely(ret < 0))
3042 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3043 local_irq_restore(flags);
3048 * See cancel_delayed_work()
3050 bool cancel_work(struct work_struct *work)
3052 return __cancel_work(work, false);
3056 * cancel_delayed_work - cancel a delayed work
3057 * @dwork: delayed_work to cancel
3059 * Kill off a pending delayed_work.
3061 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3065 * The work callback function may still be running on return, unless
3066 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3067 * use cancel_delayed_work_sync() to wait on it.
3069 * This function is safe to call from any context including IRQ handler.
3071 bool cancel_delayed_work(struct delayed_work *dwork)
3073 return __cancel_work(&dwork->work, true);
3075 EXPORT_SYMBOL(cancel_delayed_work);
3078 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3079 * @dwork: the delayed work cancel
3081 * This is cancel_work_sync() for delayed works.
3084 * %true if @dwork was pending, %false otherwise.
3086 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3088 return __cancel_work_timer(&dwork->work, true);
3090 EXPORT_SYMBOL(cancel_delayed_work_sync);
3093 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3094 * @func: the function to call
3096 * schedule_on_each_cpu() executes @func on each online CPU using the
3097 * system workqueue and blocks until all CPUs have completed.
3098 * schedule_on_each_cpu() is very slow.
3101 * 0 on success, -errno on failure.
3103 int schedule_on_each_cpu(work_func_t func)
3106 struct work_struct __percpu *works;
3108 works = alloc_percpu(struct work_struct);
3114 for_each_online_cpu(cpu) {
3115 struct work_struct *work = per_cpu_ptr(works, cpu);
3117 INIT_WORK(work, func);
3118 schedule_work_on(cpu, work);
3121 for_each_online_cpu(cpu)
3122 flush_work(per_cpu_ptr(works, cpu));
3130 * execute_in_process_context - reliably execute the routine with user context
3131 * @fn: the function to execute
3132 * @ew: guaranteed storage for the execute work structure (must
3133 * be available when the work executes)
3135 * Executes the function immediately if process context is available,
3136 * otherwise schedules the function for delayed execution.
3138 * Return: 0 - function was executed
3139 * 1 - function was scheduled for execution
3141 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3143 if (!in_interrupt()) {
3148 INIT_WORK(&ew->work, fn);
3149 schedule_work(&ew->work);
3153 EXPORT_SYMBOL_GPL(execute_in_process_context);
3156 * free_workqueue_attrs - free a workqueue_attrs
3157 * @attrs: workqueue_attrs to free
3159 * Undo alloc_workqueue_attrs().
3161 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3164 free_cpumask_var(attrs->cpumask);
3170 * alloc_workqueue_attrs - allocate a workqueue_attrs
3171 * @gfp_mask: allocation mask to use
3173 * Allocate a new workqueue_attrs, initialize with default settings and
3176 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3178 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3180 struct workqueue_attrs *attrs;
3182 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3185 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3188 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3191 free_workqueue_attrs(attrs);
3195 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3196 const struct workqueue_attrs *from)
3198 to->nice = from->nice;
3199 cpumask_copy(to->cpumask, from->cpumask);
3201 * Unlike hash and equality test, this function doesn't ignore
3202 * ->no_numa as it is used for both pool and wq attrs. Instead,
3203 * get_unbound_pool() explicitly clears ->no_numa after copying.
3205 to->no_numa = from->no_numa;
3208 /* hash value of the content of @attr */
3209 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3213 hash = jhash_1word(attrs->nice, hash);
3214 hash = jhash(cpumask_bits(attrs->cpumask),
3215 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3219 /* content equality test */
3220 static bool wqattrs_equal(const struct workqueue_attrs *a,
3221 const struct workqueue_attrs *b)
3223 if (a->nice != b->nice)
3225 if (!cpumask_equal(a->cpumask, b->cpumask))
3231 * init_worker_pool - initialize a newly zalloc'd worker_pool
3232 * @pool: worker_pool to initialize
3234 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3236 * Return: 0 on success, -errno on failure. Even on failure, all fields
3237 * inside @pool proper are initialized and put_unbound_pool() can be called
3238 * on @pool safely to release it.
3240 static int init_worker_pool(struct worker_pool *pool)
3242 spin_lock_init(&pool->lock);
3245 pool->node = NUMA_NO_NODE;
3246 pool->flags |= POOL_DISASSOCIATED;
3247 pool->watchdog_ts = jiffies;
3248 INIT_LIST_HEAD(&pool->worklist);
3249 INIT_LIST_HEAD(&pool->idle_list);
3250 hash_init(pool->busy_hash);
3252 setup_deferrable_timer(&pool->idle_timer, idle_worker_timeout,
3253 (unsigned long)pool);
3255 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3256 (unsigned long)pool);
3258 mutex_init(&pool->attach_mutex);
3259 INIT_LIST_HEAD(&pool->workers);
3261 ida_init(&pool->worker_ida);
3262 INIT_HLIST_NODE(&pool->hash_node);
3265 /* shouldn't fail above this point */
3266 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3272 static void rcu_free_wq(struct rcu_head *rcu)
3274 struct workqueue_struct *wq =
3275 container_of(rcu, struct workqueue_struct, rcu);
3277 if (!(wq->flags & WQ_UNBOUND))
3278 free_percpu(wq->cpu_pwqs);
3280 free_workqueue_attrs(wq->unbound_attrs);
3286 static void rcu_free_pool(struct rcu_head *rcu)
3288 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3290 ida_destroy(&pool->worker_ida);
3291 free_workqueue_attrs(pool->attrs);
3296 * put_unbound_pool - put a worker_pool
3297 * @pool: worker_pool to put
3299 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3300 * safe manner. get_unbound_pool() calls this function on its failure path
3301 * and this function should be able to release pools which went through,
3302 * successfully or not, init_worker_pool().
3304 * Should be called with wq_pool_mutex held.
3306 static void put_unbound_pool(struct worker_pool *pool)
3308 DECLARE_COMPLETION_ONSTACK(detach_completion);
3309 struct worker *worker;
3311 lockdep_assert_held(&wq_pool_mutex);
3317 if (WARN_ON(!(pool->cpu < 0)) ||
3318 WARN_ON(!list_empty(&pool->worklist)))
3321 /* release id and unhash */
3323 idr_remove(&worker_pool_idr, pool->id);
3324 hash_del(&pool->hash_node);
3327 * Become the manager and destroy all workers. This prevents
3328 * @pool's workers from blocking on attach_mutex. We're the last
3329 * manager and @pool gets freed with the flag set.
3331 spin_lock_irq(&pool->lock);
3332 wait_event_lock_irq(wq_manager_wait,
3333 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3334 pool->flags |= POOL_MANAGER_ACTIVE;
3336 while ((worker = first_idle_worker(pool)))
3337 destroy_worker(worker);
3338 WARN_ON(pool->nr_workers || pool->nr_idle);
3339 spin_unlock_irq(&pool->lock);
3341 mutex_lock(&pool->attach_mutex);
3342 if (!list_empty(&pool->workers))
3343 pool->detach_completion = &detach_completion;
3344 mutex_unlock(&pool->attach_mutex);
3346 if (pool->detach_completion)
3347 wait_for_completion(pool->detach_completion);
3349 /* shut down the timers */
3350 del_timer_sync(&pool->idle_timer);
3351 del_timer_sync(&pool->mayday_timer);
3353 /* sched-RCU protected to allow dereferences from get_work_pool() */
3354 call_rcu_sched(&pool->rcu, rcu_free_pool);
3358 * get_unbound_pool - get a worker_pool with the specified attributes
3359 * @attrs: the attributes of the worker_pool to get
3361 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3362 * reference count and return it. If there already is a matching
3363 * worker_pool, it will be used; otherwise, this function attempts to
3366 * Should be called with wq_pool_mutex held.
3368 * Return: On success, a worker_pool with the same attributes as @attrs.
3369 * On failure, %NULL.
3371 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3373 u32 hash = wqattrs_hash(attrs);
3374 struct worker_pool *pool;
3376 int target_node = NUMA_NO_NODE;
3378 lockdep_assert_held(&wq_pool_mutex);
3380 /* do we already have a matching pool? */
3381 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3382 if (wqattrs_equal(pool->attrs, attrs)) {
3388 /* if cpumask is contained inside a NUMA node, we belong to that node */
3389 if (wq_numa_enabled) {
3390 for_each_node(node) {
3391 if (cpumask_subset(attrs->cpumask,
3392 wq_numa_possible_cpumask[node])) {
3399 /* nope, create a new one */
3400 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3401 if (!pool || init_worker_pool(pool) < 0)
3404 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3405 copy_workqueue_attrs(pool->attrs, attrs);
3406 pool->node = target_node;
3409 * no_numa isn't a worker_pool attribute, always clear it. See
3410 * 'struct workqueue_attrs' comments for detail.
3412 pool->attrs->no_numa = false;
3414 if (worker_pool_assign_id(pool) < 0)
3417 /* create and start the initial worker */
3418 if (wq_online && !create_worker(pool))
3422 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3427 put_unbound_pool(pool);
3431 static void rcu_free_pwq(struct rcu_head *rcu)
3433 kmem_cache_free(pwq_cache,
3434 container_of(rcu, struct pool_workqueue, rcu));
3438 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3439 * and needs to be destroyed.
3441 static void pwq_unbound_release_workfn(struct work_struct *work)
3443 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3444 unbound_release_work);
3445 struct workqueue_struct *wq = pwq->wq;
3446 struct worker_pool *pool = pwq->pool;
3447 bool is_last = false;
3450 * when @pwq is not linked, it doesn't hold any reference to the
3451 * @wq, and @wq is invalid to access.
3453 if (!list_empty(&pwq->pwqs_node)) {
3454 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3457 mutex_lock(&wq->mutex);
3458 list_del_rcu(&pwq->pwqs_node);
3459 is_last = list_empty(&wq->pwqs);
3460 mutex_unlock(&wq->mutex);
3463 mutex_lock(&wq_pool_mutex);
3464 put_unbound_pool(pool);
3465 mutex_unlock(&wq_pool_mutex);
3467 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3470 * If we're the last pwq going away, @wq is already dead and no one
3471 * is gonna access it anymore. Schedule RCU free.
3474 call_rcu_sched(&wq->rcu, rcu_free_wq);
3478 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3479 * @pwq: target pool_workqueue
3481 * If @pwq isn't freezing, set @pwq->max_active to the associated
3482 * workqueue's saved_max_active and activate delayed work items
3483 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3485 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3487 struct workqueue_struct *wq = pwq->wq;
3488 bool freezable = wq->flags & WQ_FREEZABLE;
3489 unsigned long flags;
3491 /* for @wq->saved_max_active */
3492 lockdep_assert_held(&wq->mutex);
3494 /* fast exit for non-freezable wqs */
3495 if (!freezable && pwq->max_active == wq->saved_max_active)
3498 /* this function can be called during early boot w/ irq disabled */
3499 spin_lock_irqsave(&pwq->pool->lock, flags);
3502 * During [un]freezing, the caller is responsible for ensuring that
3503 * this function is called at least once after @workqueue_freezing
3504 * is updated and visible.
3506 if (!freezable || !workqueue_freezing) {
3509 pwq->max_active = wq->saved_max_active;
3511 while (!list_empty(&pwq->delayed_works) &&
3512 pwq->nr_active < pwq->max_active) {
3513 pwq_activate_first_delayed(pwq);
3518 * Need to kick a worker after thawed or an unbound wq's
3519 * max_active is bumped. In realtime scenarios, always kicking a
3520 * worker will cause interference on the isolated cpu cores, so
3521 * let's kick iff work items were activated.
3524 wake_up_worker(pwq->pool);
3526 pwq->max_active = 0;
3529 spin_unlock_irqrestore(&pwq->pool->lock, flags);
3532 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3533 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3534 struct worker_pool *pool)
3536 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3538 memset(pwq, 0, sizeof(*pwq));
3542 pwq->flush_color = -1;
3544 INIT_LIST_HEAD(&pwq->delayed_works);
3545 INIT_LIST_HEAD(&pwq->pwqs_node);
3546 INIT_LIST_HEAD(&pwq->mayday_node);
3547 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3550 /* sync @pwq with the current state of its associated wq and link it */
3551 static void link_pwq(struct pool_workqueue *pwq)
3553 struct workqueue_struct *wq = pwq->wq;
3555 lockdep_assert_held(&wq->mutex);
3557 /* may be called multiple times, ignore if already linked */
3558 if (!list_empty(&pwq->pwqs_node))
3561 /* set the matching work_color */
3562 pwq->work_color = wq->work_color;
3564 /* sync max_active to the current setting */
3565 pwq_adjust_max_active(pwq);
3568 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3571 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3572 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3573 const struct workqueue_attrs *attrs)
3575 struct worker_pool *pool;
3576 struct pool_workqueue *pwq;
3578 lockdep_assert_held(&wq_pool_mutex);
3580 pool = get_unbound_pool(attrs);
3584 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3586 put_unbound_pool(pool);
3590 init_pwq(pwq, wq, pool);
3595 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3596 * @attrs: the wq_attrs of the default pwq of the target workqueue
3597 * @node: the target NUMA node
3598 * @cpu_going_down: if >= 0, the CPU to consider as offline
3599 * @cpumask: outarg, the resulting cpumask
3601 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3602 * @cpu_going_down is >= 0, that cpu is considered offline during
3603 * calculation. The result is stored in @cpumask.
3605 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3606 * enabled and @node has online CPUs requested by @attrs, the returned
3607 * cpumask is the intersection of the possible CPUs of @node and
3610 * The caller is responsible for ensuring that the cpumask of @node stays
3613 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3616 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3617 int cpu_going_down, cpumask_t *cpumask)
3619 if (!wq_numa_enabled || attrs->no_numa)
3622 /* does @node have any online CPUs @attrs wants? */
3623 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3624 if (cpu_going_down >= 0)
3625 cpumask_clear_cpu(cpu_going_down, cpumask);
3627 if (cpumask_empty(cpumask))
3630 /* yeap, return possible CPUs in @node that @attrs wants */
3631 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3633 if (cpumask_empty(cpumask)) {
3634 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3635 "possible intersect\n");
3639 return !cpumask_equal(cpumask, attrs->cpumask);
3642 cpumask_copy(cpumask, attrs->cpumask);
3646 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3647 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3649 struct pool_workqueue *pwq)
3651 struct pool_workqueue *old_pwq;
3653 lockdep_assert_held(&wq_pool_mutex);
3654 lockdep_assert_held(&wq->mutex);
3656 /* link_pwq() can handle duplicate calls */
3659 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3660 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3664 /* context to store the prepared attrs & pwqs before applying */
3665 struct apply_wqattrs_ctx {
3666 struct workqueue_struct *wq; /* target workqueue */
3667 struct workqueue_attrs *attrs; /* attrs to apply */
3668 struct list_head list; /* queued for batching commit */
3669 struct pool_workqueue *dfl_pwq;
3670 struct pool_workqueue *pwq_tbl[];
3673 /* free the resources after success or abort */
3674 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3680 put_pwq_unlocked(ctx->pwq_tbl[node]);
3681 put_pwq_unlocked(ctx->dfl_pwq);
3683 free_workqueue_attrs(ctx->attrs);
3689 /* allocate the attrs and pwqs for later installation */
3690 static struct apply_wqattrs_ctx *
3691 apply_wqattrs_prepare(struct workqueue_struct *wq,
3692 const struct workqueue_attrs *attrs)
3694 struct apply_wqattrs_ctx *ctx;
3695 struct workqueue_attrs *new_attrs, *tmp_attrs;
3698 lockdep_assert_held(&wq_pool_mutex);
3700 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3703 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3704 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3705 if (!ctx || !new_attrs || !tmp_attrs)
3709 * Calculate the attrs of the default pwq.
3710 * If the user configured cpumask doesn't overlap with the
3711 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3713 copy_workqueue_attrs(new_attrs, attrs);
3714 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3715 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3716 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3719 * We may create multiple pwqs with differing cpumasks. Make a
3720 * copy of @new_attrs which will be modified and used to obtain
3723 copy_workqueue_attrs(tmp_attrs, new_attrs);
3726 * If something goes wrong during CPU up/down, we'll fall back to
3727 * the default pwq covering whole @attrs->cpumask. Always create
3728 * it even if we don't use it immediately.
3730 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3734 for_each_node(node) {
3735 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3736 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3737 if (!ctx->pwq_tbl[node])
3740 ctx->dfl_pwq->refcnt++;
3741 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3745 /* save the user configured attrs and sanitize it. */
3746 copy_workqueue_attrs(new_attrs, attrs);
3747 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3748 ctx->attrs = new_attrs;
3751 free_workqueue_attrs(tmp_attrs);
3755 free_workqueue_attrs(tmp_attrs);
3756 free_workqueue_attrs(new_attrs);
3757 apply_wqattrs_cleanup(ctx);
3761 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3762 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3766 /* all pwqs have been created successfully, let's install'em */
3767 mutex_lock(&ctx->wq->mutex);
3769 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3771 /* save the previous pwq and install the new one */
3773 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3774 ctx->pwq_tbl[node]);
3776 /* @dfl_pwq might not have been used, ensure it's linked */
3777 link_pwq(ctx->dfl_pwq);
3778 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3780 mutex_unlock(&ctx->wq->mutex);
3783 static void apply_wqattrs_lock(void)
3785 /* CPUs should stay stable across pwq creations and installations */
3787 mutex_lock(&wq_pool_mutex);
3790 static void apply_wqattrs_unlock(void)
3792 mutex_unlock(&wq_pool_mutex);
3796 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3797 const struct workqueue_attrs *attrs)
3799 struct apply_wqattrs_ctx *ctx;
3801 /* only unbound workqueues can change attributes */
3802 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3805 /* creating multiple pwqs breaks ordering guarantee */
3806 if (!list_empty(&wq->pwqs)) {
3807 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3810 wq->flags &= ~__WQ_ORDERED;
3813 ctx = apply_wqattrs_prepare(wq, attrs);
3817 /* the ctx has been prepared successfully, let's commit it */
3818 apply_wqattrs_commit(ctx);
3819 apply_wqattrs_cleanup(ctx);
3825 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3826 * @wq: the target workqueue
3827 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3829 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3830 * machines, this function maps a separate pwq to each NUMA node with
3831 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3832 * NUMA node it was issued on. Older pwqs are released as in-flight work
3833 * items finish. Note that a work item which repeatedly requeues itself
3834 * back-to-back will stay on its current pwq.
3836 * Performs GFP_KERNEL allocations.
3838 * Return: 0 on success and -errno on failure.
3840 int apply_workqueue_attrs(struct workqueue_struct *wq,
3841 const struct workqueue_attrs *attrs)
3845 apply_wqattrs_lock();
3846 ret = apply_workqueue_attrs_locked(wq, attrs);
3847 apply_wqattrs_unlock();
3853 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3854 * @wq: the target workqueue
3855 * @cpu: the CPU coming up or going down
3856 * @online: whether @cpu is coming up or going down
3858 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3859 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3862 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3863 * falls back to @wq->dfl_pwq which may not be optimal but is always
3866 * Note that when the last allowed CPU of a NUMA node goes offline for a
3867 * workqueue with a cpumask spanning multiple nodes, the workers which were
3868 * already executing the work items for the workqueue will lose their CPU
3869 * affinity and may execute on any CPU. This is similar to how per-cpu
3870 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3871 * affinity, it's the user's responsibility to flush the work item from
3874 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3877 int node = cpu_to_node(cpu);
3878 int cpu_off = online ? -1 : cpu;
3879 struct pool_workqueue *old_pwq = NULL, *pwq;
3880 struct workqueue_attrs *target_attrs;
3883 lockdep_assert_held(&wq_pool_mutex);
3885 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3886 wq->unbound_attrs->no_numa)
3890 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3891 * Let's use a preallocated one. The following buf is protected by
3892 * CPU hotplug exclusion.
3894 target_attrs = wq_update_unbound_numa_attrs_buf;
3895 cpumask = target_attrs->cpumask;
3897 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3898 pwq = unbound_pwq_by_node(wq, node);
3901 * Let's determine what needs to be done. If the target cpumask is
3902 * different from the default pwq's, we need to compare it to @pwq's
3903 * and create a new one if they don't match. If the target cpumask
3904 * equals the default pwq's, the default pwq should be used.
3906 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3907 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3913 /* create a new pwq */
3914 pwq = alloc_unbound_pwq(wq, target_attrs);
3916 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3921 /* Install the new pwq. */
3922 mutex_lock(&wq->mutex);
3923 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3927 mutex_lock(&wq->mutex);
3928 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3929 get_pwq(wq->dfl_pwq);
3930 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3931 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3933 mutex_unlock(&wq->mutex);
3934 put_pwq_unlocked(old_pwq);
3937 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3939 bool highpri = wq->flags & WQ_HIGHPRI;
3942 if (!(wq->flags & WQ_UNBOUND)) {
3943 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3947 for_each_possible_cpu(cpu) {
3948 struct pool_workqueue *pwq =
3949 per_cpu_ptr(wq->cpu_pwqs, cpu);
3950 struct worker_pool *cpu_pools =
3951 per_cpu(cpu_worker_pools, cpu);
3953 init_pwq(pwq, wq, &cpu_pools[highpri]);
3955 mutex_lock(&wq->mutex);
3957 mutex_unlock(&wq->mutex);
3960 } else if (wq->flags & __WQ_ORDERED) {
3961 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3962 /* there should only be single pwq for ordering guarantee */
3963 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3964 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3965 "ordering guarantee broken for workqueue %s\n", wq->name);
3968 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3972 static int wq_clamp_max_active(int max_active, unsigned int flags,
3975 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3977 if (max_active < 1 || max_active > lim)
3978 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3979 max_active, name, 1, lim);
3981 return clamp_val(max_active, 1, lim);
3984 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3987 struct lock_class_key *key,
3988 const char *lock_name, ...)
3990 size_t tbl_size = 0;
3992 struct workqueue_struct *wq;
3993 struct pool_workqueue *pwq;
3996 * Unbound && max_active == 1 used to imply ordered, which is no
3997 * longer the case on NUMA machines due to per-node pools. While
3998 * alloc_ordered_workqueue() is the right way to create an ordered
3999 * workqueue, keep the previous behavior to avoid subtle breakages
4002 if ((flags & WQ_UNBOUND) && max_active == 1)
4003 flags |= __WQ_ORDERED;
4005 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4006 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4007 flags |= WQ_UNBOUND;
4009 /* allocate wq and format name */
4010 if (flags & WQ_UNBOUND)
4011 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4013 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4017 if (flags & WQ_UNBOUND) {
4018 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4019 if (!wq->unbound_attrs)
4023 va_start(args, lock_name);
4024 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4027 max_active = max_active ?: WQ_DFL_ACTIVE;
4028 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4032 wq->saved_max_active = max_active;
4033 mutex_init(&wq->mutex);
4034 atomic_set(&wq->nr_pwqs_to_flush, 0);
4035 INIT_LIST_HEAD(&wq->pwqs);
4036 INIT_LIST_HEAD(&wq->flusher_queue);
4037 INIT_LIST_HEAD(&wq->flusher_overflow);
4038 INIT_LIST_HEAD(&wq->maydays);
4040 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4041 INIT_LIST_HEAD(&wq->list);
4043 if (alloc_and_link_pwqs(wq) < 0)
4047 * Workqueues which may be used during memory reclaim should
4048 * have a rescuer to guarantee forward progress.
4050 if (flags & WQ_MEM_RECLAIM) {
4051 struct worker *rescuer;
4053 rescuer = alloc_worker(NUMA_NO_NODE);
4057 rescuer->rescue_wq = wq;
4058 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4060 if (IS_ERR(rescuer->task)) {
4065 wq->rescuer = rescuer;
4066 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4067 wake_up_process(rescuer->task);
4070 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4074 * wq_pool_mutex protects global freeze state and workqueues list.
4075 * Grab it, adjust max_active and add the new @wq to workqueues
4078 mutex_lock(&wq_pool_mutex);
4080 mutex_lock(&wq->mutex);
4081 for_each_pwq(pwq, wq)
4082 pwq_adjust_max_active(pwq);
4083 mutex_unlock(&wq->mutex);
4085 list_add_tail_rcu(&wq->list, &workqueues);
4087 mutex_unlock(&wq_pool_mutex);
4092 free_workqueue_attrs(wq->unbound_attrs);
4096 destroy_workqueue(wq);
4099 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4102 * destroy_workqueue - safely terminate a workqueue
4103 * @wq: target workqueue
4105 * Safely destroy a workqueue. All work currently pending will be done first.
4107 void destroy_workqueue(struct workqueue_struct *wq)
4109 struct pool_workqueue *pwq;
4113 * Remove it from sysfs first so that sanity check failure doesn't
4114 * lead to sysfs name conflicts.
4116 workqueue_sysfs_unregister(wq);
4118 /* drain it before proceeding with destruction */
4119 drain_workqueue(wq);
4121 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4123 struct worker *rescuer = wq->rescuer;
4125 /* this prevents new queueing */
4126 spin_lock_irq(&wq_mayday_lock);
4128 spin_unlock_irq(&wq_mayday_lock);
4130 /* rescuer will empty maydays list before exiting */
4131 kthread_stop(rescuer->task);
4136 mutex_lock(&wq->mutex);
4137 for_each_pwq(pwq, wq) {
4140 for (i = 0; i < WORK_NR_COLORS; i++) {
4141 if (WARN_ON(pwq->nr_in_flight[i])) {
4142 mutex_unlock(&wq->mutex);
4143 show_workqueue_state();
4148 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4149 WARN_ON(pwq->nr_active) ||
4150 WARN_ON(!list_empty(&pwq->delayed_works))) {
4151 mutex_unlock(&wq->mutex);
4152 show_workqueue_state();
4156 mutex_unlock(&wq->mutex);
4159 * wq list is used to freeze wq, remove from list after
4160 * flushing is complete in case freeze races us.
4162 mutex_lock(&wq_pool_mutex);
4163 list_del_rcu(&wq->list);
4164 mutex_unlock(&wq_pool_mutex);
4166 if (!(wq->flags & WQ_UNBOUND)) {
4168 * The base ref is never dropped on per-cpu pwqs. Directly
4169 * schedule RCU free.
4171 call_rcu_sched(&wq->rcu, rcu_free_wq);
4174 * We're the sole accessor of @wq at this point. Directly
4175 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4176 * @wq will be freed when the last pwq is released.
4178 for_each_node(node) {
4179 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4180 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4181 put_pwq_unlocked(pwq);
4185 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4186 * put. Don't access it afterwards.
4190 put_pwq_unlocked(pwq);
4193 EXPORT_SYMBOL_GPL(destroy_workqueue);
4196 * workqueue_set_max_active - adjust max_active of a workqueue
4197 * @wq: target workqueue
4198 * @max_active: new max_active value.
4200 * Set max_active of @wq to @max_active.
4203 * Don't call from IRQ context.
4205 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4207 struct pool_workqueue *pwq;
4209 /* disallow meddling with max_active for ordered workqueues */
4210 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4213 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4215 mutex_lock(&wq->mutex);
4217 wq->flags &= ~__WQ_ORDERED;
4218 wq->saved_max_active = max_active;
4220 for_each_pwq(pwq, wq)
4221 pwq_adjust_max_active(pwq);
4223 mutex_unlock(&wq->mutex);
4225 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4228 * current_work - retrieve %current task's work struct
4230 * Determine if %current task is a workqueue worker and what it's working on.
4231 * Useful to find out the context that the %current task is running in.
4233 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4235 struct work_struct *current_work(void)
4237 struct worker *worker = current_wq_worker();
4239 return worker ? worker->current_work : NULL;
4241 EXPORT_SYMBOL(current_work);
4244 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4246 * Determine whether %current is a workqueue rescuer. Can be used from
4247 * work functions to determine whether it's being run off the rescuer task.
4249 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4251 bool current_is_workqueue_rescuer(void)
4253 struct worker *worker = current_wq_worker();
4255 return worker && worker->rescue_wq;
4259 * workqueue_congested - test whether a workqueue is congested
4260 * @cpu: CPU in question
4261 * @wq: target workqueue
4263 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4264 * no synchronization around this function and the test result is
4265 * unreliable and only useful as advisory hints or for debugging.
4267 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4268 * Note that both per-cpu and unbound workqueues may be associated with
4269 * multiple pool_workqueues which have separate congested states. A
4270 * workqueue being congested on one CPU doesn't mean the workqueue is also
4271 * contested on other CPUs / NUMA nodes.
4274 * %true if congested, %false otherwise.
4276 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4278 struct pool_workqueue *pwq;
4281 rcu_read_lock_sched();
4283 if (cpu == WORK_CPU_UNBOUND)
4284 cpu = smp_processor_id();
4286 if (!(wq->flags & WQ_UNBOUND))
4287 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4289 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4291 ret = !list_empty(&pwq->delayed_works);
4292 rcu_read_unlock_sched();
4296 EXPORT_SYMBOL_GPL(workqueue_congested);
4299 * work_busy - test whether a work is currently pending or running
4300 * @work: the work to be tested
4302 * Test whether @work is currently pending or running. There is no
4303 * synchronization around this function and the test result is
4304 * unreliable and only useful as advisory hints or for debugging.
4307 * OR'd bitmask of WORK_BUSY_* bits.
4309 unsigned int work_busy(struct work_struct *work)
4311 struct worker_pool *pool;
4312 unsigned long flags;
4313 unsigned int ret = 0;
4315 if (work_pending(work))
4316 ret |= WORK_BUSY_PENDING;
4318 local_irq_save(flags);
4319 pool = get_work_pool(work);
4321 spin_lock(&pool->lock);
4322 if (find_worker_executing_work(pool, work))
4323 ret |= WORK_BUSY_RUNNING;
4324 spin_unlock(&pool->lock);
4326 local_irq_restore(flags);
4330 EXPORT_SYMBOL_GPL(work_busy);
4333 * set_worker_desc - set description for the current work item
4334 * @fmt: printf-style format string
4335 * @...: arguments for the format string
4337 * This function can be called by a running work function to describe what
4338 * the work item is about. If the worker task gets dumped, this
4339 * information will be printed out together to help debugging. The
4340 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4342 void set_worker_desc(const char *fmt, ...)
4344 struct worker *worker = current_wq_worker();
4348 va_start(args, fmt);
4349 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4351 worker->desc_valid = true;
4356 * print_worker_info - print out worker information and description
4357 * @log_lvl: the log level to use when printing
4358 * @task: target task
4360 * If @task is a worker and currently executing a work item, print out the
4361 * name of the workqueue being serviced and worker description set with
4362 * set_worker_desc() by the currently executing work item.
4364 * This function can be safely called on any task as long as the
4365 * task_struct itself is accessible. While safe, this function isn't
4366 * synchronized and may print out mixups or garbages of limited length.
4368 void print_worker_info(const char *log_lvl, struct task_struct *task)
4370 work_func_t *fn = NULL;
4371 char name[WQ_NAME_LEN] = { };
4372 char desc[WORKER_DESC_LEN] = { };
4373 struct pool_workqueue *pwq = NULL;
4374 struct workqueue_struct *wq = NULL;
4375 bool desc_valid = false;
4376 struct worker *worker;
4378 if (!(task->flags & PF_WQ_WORKER))
4382 * This function is called without any synchronization and @task
4383 * could be in any state. Be careful with dereferences.
4385 worker = kthread_probe_data(task);
4388 * Carefully copy the associated workqueue's workfn and name. Keep
4389 * the original last '\0' in case the original contains garbage.
4391 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4392 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4393 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4394 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4396 /* copy worker description */
4397 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4399 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4401 if (fn || name[0] || desc[0]) {
4402 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4404 pr_cont(" (%s)", desc);
4409 static void pr_cont_pool_info(struct worker_pool *pool)
4411 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4412 if (pool->node != NUMA_NO_NODE)
4413 pr_cont(" node=%d", pool->node);
4414 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4417 static void pr_cont_work(bool comma, struct work_struct *work)
4419 if (work->func == wq_barrier_func) {
4420 struct wq_barrier *barr;
4422 barr = container_of(work, struct wq_barrier, work);
4424 pr_cont("%s BAR(%d)", comma ? "," : "",
4425 task_pid_nr(barr->task));
4427 pr_cont("%s %pf", comma ? "," : "", work->func);
4431 static void show_pwq(struct pool_workqueue *pwq)
4433 struct worker_pool *pool = pwq->pool;
4434 struct work_struct *work;
4435 struct worker *worker;
4436 bool has_in_flight = false, has_pending = false;
4439 pr_info(" pwq %d:", pool->id);
4440 pr_cont_pool_info(pool);
4442 pr_cont(" active=%d/%d refcnt=%d%s\n",
4443 pwq->nr_active, pwq->max_active, pwq->refcnt,
4444 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4446 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4447 if (worker->current_pwq == pwq) {
4448 has_in_flight = true;
4452 if (has_in_flight) {
4455 pr_info(" in-flight:");
4456 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4457 if (worker->current_pwq != pwq)
4460 pr_cont("%s %d%s:%pf", comma ? "," : "",
4461 task_pid_nr(worker->task),
4462 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4463 worker->current_func);
4464 list_for_each_entry(work, &worker->scheduled, entry)
4465 pr_cont_work(false, work);
4471 list_for_each_entry(work, &pool->worklist, entry) {
4472 if (get_work_pwq(work) == pwq) {
4480 pr_info(" pending:");
4481 list_for_each_entry(work, &pool->worklist, entry) {
4482 if (get_work_pwq(work) != pwq)
4485 pr_cont_work(comma, work);
4486 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4491 if (!list_empty(&pwq->delayed_works)) {
4494 pr_info(" delayed:");
4495 list_for_each_entry(work, &pwq->delayed_works, entry) {
4496 pr_cont_work(comma, work);
4497 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4504 * show_workqueue_state - dump workqueue state
4506 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4507 * all busy workqueues and pools.
4509 void show_workqueue_state(void)
4511 struct workqueue_struct *wq;
4512 struct worker_pool *pool;
4513 unsigned long flags;
4516 rcu_read_lock_sched();
4518 pr_info("Showing busy workqueues and worker pools:\n");
4520 list_for_each_entry_rcu(wq, &workqueues, list) {
4521 struct pool_workqueue *pwq;
4524 for_each_pwq(pwq, wq) {
4525 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4533 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4535 for_each_pwq(pwq, wq) {
4536 spin_lock_irqsave(&pwq->pool->lock, flags);
4537 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4539 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4541 * We could be printing a lot from atomic context, e.g.
4542 * sysrq-t -> show_workqueue_state(). Avoid triggering
4545 touch_nmi_watchdog();
4549 for_each_pool(pool, pi) {
4550 struct worker *worker;
4553 spin_lock_irqsave(&pool->lock, flags);
4554 if (pool->nr_workers == pool->nr_idle)
4557 pr_info("pool %d:", pool->id);
4558 pr_cont_pool_info(pool);
4559 pr_cont(" hung=%us workers=%d",
4560 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4563 pr_cont(" manager: %d",
4564 task_pid_nr(pool->manager->task));
4565 list_for_each_entry(worker, &pool->idle_list, entry) {
4566 pr_cont(" %s%d", first ? "idle: " : "",
4567 task_pid_nr(worker->task));
4572 spin_unlock_irqrestore(&pool->lock, flags);
4574 * We could be printing a lot from atomic context, e.g.
4575 * sysrq-t -> show_workqueue_state(). Avoid triggering
4578 touch_nmi_watchdog();
4581 rcu_read_unlock_sched();
4587 * There are two challenges in supporting CPU hotplug. Firstly, there
4588 * are a lot of assumptions on strong associations among work, pwq and
4589 * pool which make migrating pending and scheduled works very
4590 * difficult to implement without impacting hot paths. Secondly,
4591 * worker pools serve mix of short, long and very long running works making
4592 * blocked draining impractical.
4594 * This is solved by allowing the pools to be disassociated from the CPU
4595 * running as an unbound one and allowing it to be reattached later if the
4596 * cpu comes back online.
4599 static void wq_unbind_fn(struct work_struct *work)
4601 int cpu = smp_processor_id();
4602 struct worker_pool *pool;
4603 struct worker *worker;
4605 for_each_cpu_worker_pool(pool, cpu) {
4606 mutex_lock(&pool->attach_mutex);
4607 spin_lock_irq(&pool->lock);
4610 * We've blocked all attach/detach operations. Make all workers
4611 * unbound and set DISASSOCIATED. Before this, all workers
4612 * except for the ones which are still executing works from
4613 * before the last CPU down must be on the cpu. After
4614 * this, they may become diasporas.
4616 for_each_pool_worker(worker, pool)
4617 worker->flags |= WORKER_UNBOUND;
4619 pool->flags |= POOL_DISASSOCIATED;
4621 spin_unlock_irq(&pool->lock);
4622 mutex_unlock(&pool->attach_mutex);
4625 * Call schedule() so that we cross rq->lock and thus can
4626 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4627 * This is necessary as scheduler callbacks may be invoked
4633 * Sched callbacks are disabled now. Zap nr_running.
4634 * After this, nr_running stays zero and need_more_worker()
4635 * and keep_working() are always true as long as the
4636 * worklist is not empty. This pool now behaves as an
4637 * unbound (in terms of concurrency management) pool which
4638 * are served by workers tied to the pool.
4640 atomic_set(&pool->nr_running, 0);
4643 * With concurrency management just turned off, a busy
4644 * worker blocking could lead to lengthy stalls. Kick off
4645 * unbound chain execution of currently pending work items.
4647 spin_lock_irq(&pool->lock);
4648 wake_up_worker(pool);
4649 spin_unlock_irq(&pool->lock);
4654 * rebind_workers - rebind all workers of a pool to the associated CPU
4655 * @pool: pool of interest
4657 * @pool->cpu is coming online. Rebind all workers to the CPU.
4659 static void rebind_workers(struct worker_pool *pool)
4661 struct worker *worker;
4663 lockdep_assert_held(&pool->attach_mutex);
4666 * Restore CPU affinity of all workers. As all idle workers should
4667 * be on the run-queue of the associated CPU before any local
4668 * wake-ups for concurrency management happen, restore CPU affinity
4669 * of all workers first and then clear UNBOUND. As we're called
4670 * from CPU_ONLINE, the following shouldn't fail.
4672 for_each_pool_worker(worker, pool)
4673 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4674 pool->attrs->cpumask) < 0);
4676 spin_lock_irq(&pool->lock);
4679 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4680 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4681 * being reworked and this can go away in time.
4683 if (!(pool->flags & POOL_DISASSOCIATED)) {
4684 spin_unlock_irq(&pool->lock);
4688 pool->flags &= ~POOL_DISASSOCIATED;
4690 for_each_pool_worker(worker, pool) {
4691 unsigned int worker_flags = worker->flags;
4694 * A bound idle worker should actually be on the runqueue
4695 * of the associated CPU for local wake-ups targeting it to
4696 * work. Kick all idle workers so that they migrate to the
4697 * associated CPU. Doing this in the same loop as
4698 * replacing UNBOUND with REBOUND is safe as no worker will
4699 * be bound before @pool->lock is released.
4701 if (worker_flags & WORKER_IDLE)
4702 wake_up_process(worker->task);
4705 * We want to clear UNBOUND but can't directly call
4706 * worker_clr_flags() or adjust nr_running. Atomically
4707 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4708 * @worker will clear REBOUND using worker_clr_flags() when
4709 * it initiates the next execution cycle thus restoring
4710 * concurrency management. Note that when or whether
4711 * @worker clears REBOUND doesn't affect correctness.
4713 * ACCESS_ONCE() is necessary because @worker->flags may be
4714 * tested without holding any lock in
4715 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4716 * fail incorrectly leading to premature concurrency
4717 * management operations.
4719 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4720 worker_flags |= WORKER_REBOUND;
4721 worker_flags &= ~WORKER_UNBOUND;
4722 ACCESS_ONCE(worker->flags) = worker_flags;
4725 spin_unlock_irq(&pool->lock);
4729 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4730 * @pool: unbound pool of interest
4731 * @cpu: the CPU which is coming up
4733 * An unbound pool may end up with a cpumask which doesn't have any online
4734 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4735 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4736 * online CPU before, cpus_allowed of all its workers should be restored.
4738 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4740 static cpumask_t cpumask;
4741 struct worker *worker;
4743 lockdep_assert_held(&pool->attach_mutex);
4745 /* is @cpu allowed for @pool? */
4746 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4749 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4751 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4752 for_each_pool_worker(worker, pool)
4753 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4756 int workqueue_prepare_cpu(unsigned int cpu)
4758 struct worker_pool *pool;
4760 for_each_cpu_worker_pool(pool, cpu) {
4761 if (pool->nr_workers)
4763 if (!create_worker(pool))
4769 int workqueue_online_cpu(unsigned int cpu)
4771 struct worker_pool *pool;
4772 struct workqueue_struct *wq;
4775 mutex_lock(&wq_pool_mutex);
4777 for_each_pool(pool, pi) {
4778 mutex_lock(&pool->attach_mutex);
4780 if (pool->cpu == cpu)
4781 rebind_workers(pool);
4782 else if (pool->cpu < 0)
4783 restore_unbound_workers_cpumask(pool, cpu);
4785 mutex_unlock(&pool->attach_mutex);
4788 /* update NUMA affinity of unbound workqueues */
4789 list_for_each_entry(wq, &workqueues, list)
4790 wq_update_unbound_numa(wq, cpu, true);
4792 mutex_unlock(&wq_pool_mutex);
4796 int workqueue_offline_cpu(unsigned int cpu)
4798 struct work_struct unbind_work;
4799 struct workqueue_struct *wq;
4801 /* unbinding per-cpu workers should happen on the local CPU */
4802 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4803 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4805 /* update NUMA affinity of unbound workqueues */
4806 mutex_lock(&wq_pool_mutex);
4807 list_for_each_entry(wq, &workqueues, list)
4808 wq_update_unbound_numa(wq, cpu, false);
4809 mutex_unlock(&wq_pool_mutex);
4811 /* wait for per-cpu unbinding to finish */
4812 flush_work(&unbind_work);
4813 destroy_work_on_stack(&unbind_work);
4819 struct work_for_cpu {
4820 struct work_struct work;
4826 static void work_for_cpu_fn(struct work_struct *work)
4828 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4830 wfc->ret = wfc->fn(wfc->arg);
4834 * work_on_cpu - run a function in thread context on a particular cpu
4835 * @cpu: the cpu to run on
4836 * @fn: the function to run
4837 * @arg: the function arg
4839 * It is up to the caller to ensure that the cpu doesn't go offline.
4840 * The caller must not hold any locks which would prevent @fn from completing.
4842 * Return: The value @fn returns.
4844 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4846 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4848 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4849 schedule_work_on(cpu, &wfc.work);
4850 flush_work(&wfc.work);
4851 destroy_work_on_stack(&wfc.work);
4854 EXPORT_SYMBOL_GPL(work_on_cpu);
4857 * work_on_cpu_safe - run a function in thread context on a particular cpu
4858 * @cpu: the cpu to run on
4859 * @fn: the function to run
4860 * @arg: the function argument
4862 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
4863 * any locks which would prevent @fn from completing.
4865 * Return: The value @fn returns.
4867 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
4872 if (cpu_online(cpu))
4873 ret = work_on_cpu(cpu, fn, arg);
4877 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
4878 #endif /* CONFIG_SMP */
4880 #ifdef CONFIG_FREEZER
4883 * freeze_workqueues_begin - begin freezing workqueues
4885 * Start freezing workqueues. After this function returns, all freezable
4886 * workqueues will queue new works to their delayed_works list instead of
4890 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4892 void freeze_workqueues_begin(void)
4894 struct workqueue_struct *wq;
4895 struct pool_workqueue *pwq;
4897 mutex_lock(&wq_pool_mutex);
4899 WARN_ON_ONCE(workqueue_freezing);
4900 workqueue_freezing = true;
4902 list_for_each_entry(wq, &workqueues, list) {
4903 mutex_lock(&wq->mutex);
4904 for_each_pwq(pwq, wq)
4905 pwq_adjust_max_active(pwq);
4906 mutex_unlock(&wq->mutex);
4909 mutex_unlock(&wq_pool_mutex);
4913 * freeze_workqueues_busy - are freezable workqueues still busy?
4915 * Check whether freezing is complete. This function must be called
4916 * between freeze_workqueues_begin() and thaw_workqueues().
4919 * Grabs and releases wq_pool_mutex.
4922 * %true if some freezable workqueues are still busy. %false if freezing
4925 bool freeze_workqueues_busy(void)
4928 struct workqueue_struct *wq;
4929 struct pool_workqueue *pwq;
4931 mutex_lock(&wq_pool_mutex);
4933 WARN_ON_ONCE(!workqueue_freezing);
4935 list_for_each_entry(wq, &workqueues, list) {
4936 if (!(wq->flags & WQ_FREEZABLE))
4939 * nr_active is monotonically decreasing. It's safe
4940 * to peek without lock.
4942 rcu_read_lock_sched();
4943 for_each_pwq(pwq, wq) {
4944 WARN_ON_ONCE(pwq->nr_active < 0);
4945 if (pwq->nr_active) {
4947 rcu_read_unlock_sched();
4951 rcu_read_unlock_sched();
4954 mutex_unlock(&wq_pool_mutex);
4959 * thaw_workqueues - thaw workqueues
4961 * Thaw workqueues. Normal queueing is restored and all collected
4962 * frozen works are transferred to their respective pool worklists.
4965 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4967 void thaw_workqueues(void)
4969 struct workqueue_struct *wq;
4970 struct pool_workqueue *pwq;
4972 mutex_lock(&wq_pool_mutex);
4974 if (!workqueue_freezing)
4977 workqueue_freezing = false;
4979 /* restore max_active and repopulate worklist */
4980 list_for_each_entry(wq, &workqueues, list) {
4981 mutex_lock(&wq->mutex);
4982 for_each_pwq(pwq, wq)
4983 pwq_adjust_max_active(pwq);
4984 mutex_unlock(&wq->mutex);
4988 mutex_unlock(&wq_pool_mutex);
4990 #endif /* CONFIG_FREEZER */
4992 static int workqueue_apply_unbound_cpumask(void)
4996 struct workqueue_struct *wq;
4997 struct apply_wqattrs_ctx *ctx, *n;
4999 lockdep_assert_held(&wq_pool_mutex);
5001 list_for_each_entry(wq, &workqueues, list) {
5002 if (!(wq->flags & WQ_UNBOUND))
5004 /* creating multiple pwqs breaks ordering guarantee */
5005 if (wq->flags & __WQ_ORDERED)
5008 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5014 list_add_tail(&ctx->list, &ctxs);
5017 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5019 apply_wqattrs_commit(ctx);
5020 apply_wqattrs_cleanup(ctx);
5027 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5028 * @cpumask: the cpumask to set
5030 * The low-level workqueues cpumask is a global cpumask that limits
5031 * the affinity of all unbound workqueues. This function check the @cpumask
5032 * and apply it to all unbound workqueues and updates all pwqs of them.
5034 * Retun: 0 - Success
5035 * -EINVAL - Invalid @cpumask
5036 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5038 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5041 cpumask_var_t saved_cpumask;
5043 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5046 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5047 if (!cpumask_empty(cpumask)) {
5048 apply_wqattrs_lock();
5050 /* save the old wq_unbound_cpumask. */
5051 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5053 /* update wq_unbound_cpumask at first and apply it to wqs. */
5054 cpumask_copy(wq_unbound_cpumask, cpumask);
5055 ret = workqueue_apply_unbound_cpumask();
5057 /* restore the wq_unbound_cpumask when failed. */
5059 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5061 apply_wqattrs_unlock();
5064 free_cpumask_var(saved_cpumask);
5070 * Workqueues with WQ_SYSFS flag set is visible to userland via
5071 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5072 * following attributes.
5074 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5075 * max_active RW int : maximum number of in-flight work items
5077 * Unbound workqueues have the following extra attributes.
5079 * id RO int : the associated pool ID
5080 * nice RW int : nice value of the workers
5081 * cpumask RW mask : bitmask of allowed CPUs for the workers
5084 struct workqueue_struct *wq;
5088 static struct workqueue_struct *dev_to_wq(struct device *dev)
5090 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5095 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5098 struct workqueue_struct *wq = dev_to_wq(dev);
5100 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5102 static DEVICE_ATTR_RO(per_cpu);
5104 static ssize_t max_active_show(struct device *dev,
5105 struct device_attribute *attr, char *buf)
5107 struct workqueue_struct *wq = dev_to_wq(dev);
5109 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5112 static ssize_t max_active_store(struct device *dev,
5113 struct device_attribute *attr, const char *buf,
5116 struct workqueue_struct *wq = dev_to_wq(dev);
5119 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5122 workqueue_set_max_active(wq, val);
5125 static DEVICE_ATTR_RW(max_active);
5127 static struct attribute *wq_sysfs_attrs[] = {
5128 &dev_attr_per_cpu.attr,
5129 &dev_attr_max_active.attr,
5132 ATTRIBUTE_GROUPS(wq_sysfs);
5134 static ssize_t wq_pool_ids_show(struct device *dev,
5135 struct device_attribute *attr, char *buf)
5137 struct workqueue_struct *wq = dev_to_wq(dev);
5138 const char *delim = "";
5139 int node, written = 0;
5141 rcu_read_lock_sched();
5142 for_each_node(node) {
5143 written += scnprintf(buf + written, PAGE_SIZE - written,
5144 "%s%d:%d", delim, node,
5145 unbound_pwq_by_node(wq, node)->pool->id);
5148 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5149 rcu_read_unlock_sched();
5154 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5157 struct workqueue_struct *wq = dev_to_wq(dev);
5160 mutex_lock(&wq->mutex);
5161 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5162 mutex_unlock(&wq->mutex);
5167 /* prepare workqueue_attrs for sysfs store operations */
5168 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5170 struct workqueue_attrs *attrs;
5172 lockdep_assert_held(&wq_pool_mutex);
5174 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5178 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5182 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5183 const char *buf, size_t count)
5185 struct workqueue_struct *wq = dev_to_wq(dev);
5186 struct workqueue_attrs *attrs;
5189 apply_wqattrs_lock();
5191 attrs = wq_sysfs_prep_attrs(wq);
5195 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5196 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5197 ret = apply_workqueue_attrs_locked(wq, attrs);
5202 apply_wqattrs_unlock();
5203 free_workqueue_attrs(attrs);
5204 return ret ?: count;
5207 static ssize_t wq_cpumask_show(struct device *dev,
5208 struct device_attribute *attr, char *buf)
5210 struct workqueue_struct *wq = dev_to_wq(dev);
5213 mutex_lock(&wq->mutex);
5214 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5215 cpumask_pr_args(wq->unbound_attrs->cpumask));
5216 mutex_unlock(&wq->mutex);
5220 static ssize_t wq_cpumask_store(struct device *dev,
5221 struct device_attribute *attr,
5222 const char *buf, size_t count)
5224 struct workqueue_struct *wq = dev_to_wq(dev);
5225 struct workqueue_attrs *attrs;
5228 apply_wqattrs_lock();
5230 attrs = wq_sysfs_prep_attrs(wq);
5234 ret = cpumask_parse(buf, attrs->cpumask);
5236 ret = apply_workqueue_attrs_locked(wq, attrs);
5239 apply_wqattrs_unlock();
5240 free_workqueue_attrs(attrs);
5241 return ret ?: count;
5244 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5247 struct workqueue_struct *wq = dev_to_wq(dev);
5250 mutex_lock(&wq->mutex);
5251 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5252 !wq->unbound_attrs->no_numa);
5253 mutex_unlock(&wq->mutex);
5258 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5259 const char *buf, size_t count)
5261 struct workqueue_struct *wq = dev_to_wq(dev);
5262 struct workqueue_attrs *attrs;
5263 int v, ret = -ENOMEM;
5265 apply_wqattrs_lock();
5267 attrs = wq_sysfs_prep_attrs(wq);
5272 if (sscanf(buf, "%d", &v) == 1) {
5273 attrs->no_numa = !v;
5274 ret = apply_workqueue_attrs_locked(wq, attrs);
5278 apply_wqattrs_unlock();
5279 free_workqueue_attrs(attrs);
5280 return ret ?: count;
5283 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5284 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5285 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5286 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5287 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5291 static struct bus_type wq_subsys = {
5292 .name = "workqueue",
5293 .dev_groups = wq_sysfs_groups,
5296 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5297 struct device_attribute *attr, char *buf)
5301 mutex_lock(&wq_pool_mutex);
5302 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5303 cpumask_pr_args(wq_unbound_cpumask));
5304 mutex_unlock(&wq_pool_mutex);
5309 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5310 struct device_attribute *attr, const char *buf, size_t count)
5312 cpumask_var_t cpumask;
5315 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5318 ret = cpumask_parse(buf, cpumask);
5320 ret = workqueue_set_unbound_cpumask(cpumask);
5322 free_cpumask_var(cpumask);
5323 return ret ? ret : count;
5326 static struct device_attribute wq_sysfs_cpumask_attr =
5327 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5328 wq_unbound_cpumask_store);
5330 static int __init wq_sysfs_init(void)
5334 err = subsys_virtual_register(&wq_subsys, NULL);
5338 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5340 core_initcall(wq_sysfs_init);
5342 static void wq_device_release(struct device *dev)
5344 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5350 * workqueue_sysfs_register - make a workqueue visible in sysfs
5351 * @wq: the workqueue to register
5353 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5354 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5355 * which is the preferred method.
5357 * Workqueue user should use this function directly iff it wants to apply
5358 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5359 * apply_workqueue_attrs() may race against userland updating the
5362 * Return: 0 on success, -errno on failure.
5364 int workqueue_sysfs_register(struct workqueue_struct *wq)
5366 struct wq_device *wq_dev;
5370 * Adjusting max_active or creating new pwqs by applying
5371 * attributes breaks ordering guarantee. Disallow exposing ordered
5374 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5377 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5382 wq_dev->dev.bus = &wq_subsys;
5383 wq_dev->dev.release = wq_device_release;
5384 dev_set_name(&wq_dev->dev, "%s", wq->name);
5387 * unbound_attrs are created separately. Suppress uevent until
5388 * everything is ready.
5390 dev_set_uevent_suppress(&wq_dev->dev, true);
5392 ret = device_register(&wq_dev->dev);
5394 put_device(&wq_dev->dev);
5399 if (wq->flags & WQ_UNBOUND) {
5400 struct device_attribute *attr;
5402 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5403 ret = device_create_file(&wq_dev->dev, attr);
5405 device_unregister(&wq_dev->dev);
5412 dev_set_uevent_suppress(&wq_dev->dev, false);
5413 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5418 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5419 * @wq: the workqueue to unregister
5421 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5423 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5425 struct wq_device *wq_dev = wq->wq_dev;
5431 device_unregister(&wq_dev->dev);
5433 #else /* CONFIG_SYSFS */
5434 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5435 #endif /* CONFIG_SYSFS */
5438 * Workqueue watchdog.
5440 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5441 * flush dependency, a concurrency managed work item which stays RUNNING
5442 * indefinitely. Workqueue stalls can be very difficult to debug as the
5443 * usual warning mechanisms don't trigger and internal workqueue state is
5446 * Workqueue watchdog monitors all worker pools periodically and dumps
5447 * state if some pools failed to make forward progress for a while where
5448 * forward progress is defined as the first item on ->worklist changing.
5450 * This mechanism is controlled through the kernel parameter
5451 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5452 * corresponding sysfs parameter file.
5454 #ifdef CONFIG_WQ_WATCHDOG
5456 static void wq_watchdog_timer_fn(unsigned long data);
5458 static unsigned long wq_watchdog_thresh = 30;
5459 static struct timer_list wq_watchdog_timer =
5460 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5462 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5463 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5465 static void wq_watchdog_reset_touched(void)
5469 wq_watchdog_touched = jiffies;
5470 for_each_possible_cpu(cpu)
5471 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5474 static void wq_watchdog_timer_fn(unsigned long data)
5476 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5477 bool lockup_detected = false;
5478 unsigned long now = jiffies;
5479 struct worker_pool *pool;
5487 for_each_pool(pool, pi) {
5488 unsigned long pool_ts, touched, ts;
5490 if (list_empty(&pool->worklist))
5494 * If a virtual machine is stopped by the host it can look to
5495 * the watchdog like a stall.
5497 kvm_check_and_clear_guest_paused();
5499 /* get the latest of pool and touched timestamps */
5500 pool_ts = READ_ONCE(pool->watchdog_ts);
5501 touched = READ_ONCE(wq_watchdog_touched);
5503 if (time_after(pool_ts, touched))
5508 if (pool->cpu >= 0) {
5509 unsigned long cpu_touched =
5510 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5512 if (time_after(cpu_touched, ts))
5517 if (time_after(now, ts + thresh)) {
5518 lockup_detected = true;
5519 pr_emerg("BUG: workqueue lockup - pool");
5520 pr_cont_pool_info(pool);
5521 pr_cont(" stuck for %us!\n",
5522 jiffies_to_msecs(now - pool_ts) / 1000);
5528 if (lockup_detected)
5529 show_workqueue_state();
5531 wq_watchdog_reset_touched();
5532 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5535 notrace void wq_watchdog_touch(int cpu)
5538 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5540 wq_watchdog_touched = jiffies;
5543 static void wq_watchdog_set_thresh(unsigned long thresh)
5545 wq_watchdog_thresh = 0;
5546 del_timer_sync(&wq_watchdog_timer);
5549 wq_watchdog_thresh = thresh;
5550 wq_watchdog_reset_touched();
5551 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5555 static int wq_watchdog_param_set_thresh(const char *val,
5556 const struct kernel_param *kp)
5558 unsigned long thresh;
5561 ret = kstrtoul(val, 0, &thresh);
5566 wq_watchdog_set_thresh(thresh);
5568 wq_watchdog_thresh = thresh;
5573 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5574 .set = wq_watchdog_param_set_thresh,
5575 .get = param_get_ulong,
5578 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5581 static void wq_watchdog_init(void)
5583 wq_watchdog_set_thresh(wq_watchdog_thresh);
5586 #else /* CONFIG_WQ_WATCHDOG */
5588 static inline void wq_watchdog_init(void) { }
5590 #endif /* CONFIG_WQ_WATCHDOG */
5592 static void __init wq_numa_init(void)
5597 if (num_possible_nodes() <= 1)
5600 if (wq_disable_numa) {
5601 pr_info("workqueue: NUMA affinity support disabled\n");
5605 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5606 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5609 * We want masks of possible CPUs of each node which isn't readily
5610 * available. Build one from cpu_to_node() which should have been
5611 * fully initialized by now.
5613 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5617 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5618 node_online(node) ? node : NUMA_NO_NODE));
5620 for_each_possible_cpu(cpu) {
5621 node = cpu_to_node(cpu);
5622 if (WARN_ON(node == NUMA_NO_NODE)) {
5623 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5624 /* happens iff arch is bonkers, let's just proceed */
5627 cpumask_set_cpu(cpu, tbl[node]);
5630 wq_numa_possible_cpumask = tbl;
5631 wq_numa_enabled = true;
5635 * workqueue_init_early - early init for workqueue subsystem
5637 * This is the first half of two-staged workqueue subsystem initialization
5638 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5639 * idr are up. It sets up all the data structures and system workqueues
5640 * and allows early boot code to create workqueues and queue/cancel work
5641 * items. Actual work item execution starts only after kthreads can be
5642 * created and scheduled right before early initcalls.
5644 int __init workqueue_init_early(void)
5646 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5649 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5651 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5652 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5654 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5656 /* initialize CPU pools */
5657 for_each_possible_cpu(cpu) {
5658 struct worker_pool *pool;
5661 for_each_cpu_worker_pool(pool, cpu) {
5662 BUG_ON(init_worker_pool(pool));
5664 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5665 pool->attrs->nice = std_nice[i++];
5666 pool->node = cpu_to_node(cpu);
5669 mutex_lock(&wq_pool_mutex);
5670 BUG_ON(worker_pool_assign_id(pool));
5671 mutex_unlock(&wq_pool_mutex);
5675 /* create default unbound and ordered wq attrs */
5676 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5677 struct workqueue_attrs *attrs;
5679 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5680 attrs->nice = std_nice[i];
5681 unbound_std_wq_attrs[i] = attrs;
5684 * An ordered wq should have only one pwq as ordering is
5685 * guaranteed by max_active which is enforced by pwqs.
5686 * Turn off NUMA so that dfl_pwq is used for all nodes.
5688 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5689 attrs->nice = std_nice[i];
5690 attrs->no_numa = true;
5691 ordered_wq_attrs[i] = attrs;
5694 system_wq = alloc_workqueue("events", 0, 0);
5695 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5696 system_long_wq = alloc_workqueue("events_long", 0, 0);
5697 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5698 WQ_UNBOUND_MAX_ACTIVE);
5699 system_freezable_wq = alloc_workqueue("events_freezable",
5701 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5702 WQ_POWER_EFFICIENT, 0);
5703 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5704 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5706 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5707 !system_unbound_wq || !system_freezable_wq ||
5708 !system_power_efficient_wq ||
5709 !system_freezable_power_efficient_wq);
5715 * workqueue_init - bring workqueue subsystem fully online
5717 * This is the latter half of two-staged workqueue subsystem initialization
5718 * and invoked as soon as kthreads can be created and scheduled.
5719 * Workqueues have been created and work items queued on them, but there
5720 * are no kworkers executing the work items yet. Populate the worker pools
5721 * with the initial workers and enable future kworker creations.
5723 int __init workqueue_init(void)
5725 struct workqueue_struct *wq;
5726 struct worker_pool *pool;
5730 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5731 * CPU to node mapping may not be available that early on some
5732 * archs such as power and arm64. As per-cpu pools created
5733 * previously could be missing node hint and unbound pools NUMA
5734 * affinity, fix them up.
5738 mutex_lock(&wq_pool_mutex);
5740 for_each_possible_cpu(cpu) {
5741 for_each_cpu_worker_pool(pool, cpu) {
5742 pool->node = cpu_to_node(cpu);
5746 list_for_each_entry(wq, &workqueues, list)
5747 wq_update_unbound_numa(wq, smp_processor_id(), true);
5749 mutex_unlock(&wq_pool_mutex);
5751 /* create the initial workers */
5752 for_each_online_cpu(cpu) {
5753 for_each_cpu_worker_pool(pool, cpu) {
5754 pool->flags &= ~POOL_DISASSOCIATED;
5755 BUG_ON(!create_worker(pool));
5759 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5760 BUG_ON(!create_worker(pool));