2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_REBOUND = 1 << 8, /* worker was rebound */
81 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
84 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
86 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
89 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
102 RESCUER_NICE_LEVEL = MIN_NICE,
103 HIGHPRI_NICE_LEVEL = MIN_NICE,
109 * Structure fields follow one of the following exclusion rules.
111 * I: Modifiable by initialization/destruction paths and read-only for
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
117 * L: pool->lock protected. Access with pool->lock held.
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
124 * A: pool->attach_mutex protected.
126 * PL: wq_pool_mutex protected.
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
130 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
132 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133 * sched-RCU for reads.
135 * WQ: wq->mutex protected.
137 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
139 * MD: wq_mayday_lock protected.
142 /* struct worker is defined in workqueue_internal.h */
145 spinlock_t lock; /* the pool lock */
146 int cpu; /* I: the associated cpu */
147 int node; /* I: the associated node ID */
148 int id; /* I: pool ID */
149 unsigned int flags; /* X: flags */
151 unsigned long watchdog_ts; /* L: watchdog timestamp */
153 struct list_head worklist; /* L: list of pending works */
154 int nr_workers; /* L: total number of workers */
156 /* nr_idle includes the ones off idle_list for rebinding */
157 int nr_idle; /* L: currently idle ones */
159 struct list_head idle_list; /* X: list of idle workers */
160 struct timer_list idle_timer; /* L: worker idle timeout */
161 struct timer_list mayday_timer; /* L: SOS timer for workers */
163 /* a workers is either on busy_hash or idle_list, or the manager */
164 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
165 /* L: hash of busy workers */
167 /* see manage_workers() for details on the two manager mutexes */
168 struct mutex manager_arb; /* manager arbitration */
169 struct worker *manager; /* L: purely informational */
170 struct mutex attach_mutex; /* attach/detach exclusion */
171 struct list_head workers; /* A: attached workers */
172 struct completion *detach_completion; /* all workers detached */
174 struct ida worker_ida; /* worker IDs for task name */
176 struct workqueue_attrs *attrs; /* I: worker attributes */
177 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178 int refcnt; /* PL: refcnt for unbound pools */
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
185 atomic_t nr_running ____cacheline_aligned_in_smp;
188 * Destruction of pool is sched-RCU protected to allow dereferences
189 * from get_work_pool().
192 } ____cacheline_aligned_in_smp;
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
200 struct pool_workqueue {
201 struct worker_pool *pool; /* I: the associated pool */
202 struct workqueue_struct *wq; /* I: the owning workqueue */
203 int work_color; /* L: current color */
204 int flush_color; /* L: flushing color */
205 int refcnt; /* L: reference count */
206 int nr_in_flight[WORK_NR_COLORS];
207 /* L: nr of in_flight works */
208 int nr_active; /* L: nr of active works */
209 int max_active; /* L: max active works */
210 struct list_head delayed_works; /* L: delayed works */
211 struct list_head pwqs_node; /* WR: node on wq->pwqs */
212 struct list_head mayday_node; /* MD: node on wq->maydays */
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also sched-RCU protected so that the first pwq can be
218 * determined without grabbing wq->mutex.
220 struct work_struct unbound_release_work;
222 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
225 * Structure used to wait for workqueue flush.
228 struct list_head list; /* WQ: list of flushers */
229 int flush_color; /* WQ: flush color waiting for */
230 struct completion done; /* flush completion */
236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
239 struct workqueue_struct {
240 struct list_head pwqs; /* WR: all pwqs of this wq */
241 struct list_head list; /* PR: list of all workqueues */
243 struct mutex mutex; /* protects this wq */
244 int work_color; /* WQ: current work color */
245 int flush_color; /* WQ: current flush color */
246 atomic_t nr_pwqs_to_flush; /* flush in progress */
247 struct wq_flusher *first_flusher; /* WQ: first flusher */
248 struct list_head flusher_queue; /* WQ: flush waiters */
249 struct list_head flusher_overflow; /* WQ: flush overflow list */
251 struct list_head maydays; /* MD: pwqs requesting rescue */
252 struct worker *rescuer; /* I: rescue worker */
254 int nr_drainers; /* WQ: drain in progress */
255 int saved_max_active; /* WQ: saved pwq max_active */
257 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
258 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
261 struct wq_device *wq_dev; /* I: for sysfs interface */
263 #ifdef CONFIG_LOCKDEP
264 struct lockdep_map lockdep_map;
266 char name[WQ_NAME_LEN]; /* I: workqueue name */
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
278 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
281 static struct kmem_cache *pwq_cache;
283 static cpumask_var_t *wq_numa_possible_cpumask;
284 /* possible CPUs of each node */
286 static bool wq_disable_numa;
287 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
289 /* see the comment above the definition of WQ_POWER_EFFICIENT */
290 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
293 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
295 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
296 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
298 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
299 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
301 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
302 static bool workqueue_freezing; /* PL: have wqs started freezing? */
304 /* PL: allowable cpus for unbound wqs and work items */
305 static cpumask_var_t wq_unbound_cpumask;
307 /* CPU where unbound work was last round robin scheduled from this CPU */
308 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
311 * Local execution of unbound work items is no longer guaranteed. The
312 * following always forces round-robin CPU selection on unbound work items
313 * to uncover usages which depend on it.
315 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
316 static bool wq_debug_force_rr_cpu = true;
318 static bool wq_debug_force_rr_cpu = false;
320 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
322 /* the per-cpu worker pools */
323 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
325 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
327 /* PL: hash of all unbound pools keyed by pool->attrs */
328 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
330 /* I: attributes used when instantiating standard unbound pools on demand */
331 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
333 /* I: attributes used when instantiating ordered pools on demand */
334 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
336 struct workqueue_struct *system_wq __read_mostly;
337 EXPORT_SYMBOL(system_wq);
338 struct workqueue_struct *system_highpri_wq __read_mostly;
339 EXPORT_SYMBOL_GPL(system_highpri_wq);
340 struct workqueue_struct *system_long_wq __read_mostly;
341 EXPORT_SYMBOL_GPL(system_long_wq);
342 struct workqueue_struct *system_unbound_wq __read_mostly;
343 EXPORT_SYMBOL_GPL(system_unbound_wq);
344 struct workqueue_struct *system_freezable_wq __read_mostly;
345 EXPORT_SYMBOL_GPL(system_freezable_wq);
346 struct workqueue_struct *system_power_efficient_wq __read_mostly;
347 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
348 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
349 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
351 static int worker_thread(void *__worker);
352 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
354 #define CREATE_TRACE_POINTS
355 #include <trace/events/workqueue.h>
357 #define assert_rcu_or_pool_mutex() \
358 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
359 !lockdep_is_held(&wq_pool_mutex), \
360 "sched RCU or wq_pool_mutex should be held")
362 #define assert_rcu_or_wq_mutex(wq) \
363 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
364 !lockdep_is_held(&wq->mutex), \
365 "sched RCU or wq->mutex should be held")
367 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
368 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
369 !lockdep_is_held(&wq->mutex) && \
370 !lockdep_is_held(&wq_pool_mutex), \
371 "sched RCU, wq->mutex or wq_pool_mutex should be held")
373 #define for_each_cpu_worker_pool(pool, cpu) \
374 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
375 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
379 * for_each_pool - iterate through all worker_pools in the system
380 * @pool: iteration cursor
381 * @pi: integer used for iteration
383 * This must be called either with wq_pool_mutex held or sched RCU read
384 * locked. If the pool needs to be used beyond the locking in effect, the
385 * caller is responsible for guaranteeing that the pool stays online.
387 * The if/else clause exists only for the lockdep assertion and can be
390 #define for_each_pool(pool, pi) \
391 idr_for_each_entry(&worker_pool_idr, pool, pi) \
392 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
396 * for_each_pool_worker - iterate through all workers of a worker_pool
397 * @worker: iteration cursor
398 * @pool: worker_pool to iterate workers of
400 * This must be called with @pool->attach_mutex.
402 * The if/else clause exists only for the lockdep assertion and can be
405 #define for_each_pool_worker(worker, pool) \
406 list_for_each_entry((worker), &(pool)->workers, node) \
407 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
411 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
412 * @pwq: iteration cursor
413 * @wq: the target workqueue
415 * This must be called either with wq->mutex held or sched RCU read locked.
416 * If the pwq needs to be used beyond the locking in effect, the caller is
417 * responsible for guaranteeing that the pwq stays online.
419 * The if/else clause exists only for the lockdep assertion and can be
422 #define for_each_pwq(pwq, wq) \
423 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
424 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
427 #ifdef CONFIG_DEBUG_OBJECTS_WORK
429 static struct debug_obj_descr work_debug_descr;
431 static void *work_debug_hint(void *addr)
433 return ((struct work_struct *) addr)->func;
436 static bool work_is_static_object(void *addr)
438 struct work_struct *work = addr;
440 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
444 * fixup_init is called when:
445 * - an active object is initialized
447 static bool work_fixup_init(void *addr, enum debug_obj_state state)
449 struct work_struct *work = addr;
452 case ODEBUG_STATE_ACTIVE:
453 cancel_work_sync(work);
454 debug_object_init(work, &work_debug_descr);
462 * fixup_free is called when:
463 * - an active object is freed
465 static bool work_fixup_free(void *addr, enum debug_obj_state state)
467 struct work_struct *work = addr;
470 case ODEBUG_STATE_ACTIVE:
471 cancel_work_sync(work);
472 debug_object_free(work, &work_debug_descr);
479 static struct debug_obj_descr work_debug_descr = {
480 .name = "work_struct",
481 .debug_hint = work_debug_hint,
482 .is_static_object = work_is_static_object,
483 .fixup_init = work_fixup_init,
484 .fixup_free = work_fixup_free,
487 static inline void debug_work_activate(struct work_struct *work)
489 debug_object_activate(work, &work_debug_descr);
492 static inline void debug_work_deactivate(struct work_struct *work)
494 debug_object_deactivate(work, &work_debug_descr);
497 void __init_work(struct work_struct *work, int onstack)
500 debug_object_init_on_stack(work, &work_debug_descr);
502 debug_object_init(work, &work_debug_descr);
504 EXPORT_SYMBOL_GPL(__init_work);
506 void destroy_work_on_stack(struct work_struct *work)
508 debug_object_free(work, &work_debug_descr);
510 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
512 void destroy_delayed_work_on_stack(struct delayed_work *work)
514 destroy_timer_on_stack(&work->timer);
515 debug_object_free(&work->work, &work_debug_descr);
517 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
520 static inline void debug_work_activate(struct work_struct *work) { }
521 static inline void debug_work_deactivate(struct work_struct *work) { }
525 * worker_pool_assign_id - allocate ID and assing it to @pool
526 * @pool: the pool pointer of interest
528 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
529 * successfully, -errno on failure.
531 static int worker_pool_assign_id(struct worker_pool *pool)
535 lockdep_assert_held(&wq_pool_mutex);
537 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
547 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
548 * @wq: the target workqueue
551 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
553 * If the pwq needs to be used beyond the locking in effect, the caller is
554 * responsible for guaranteeing that the pwq stays online.
556 * Return: The unbound pool_workqueue for @node.
558 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
561 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
564 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
565 * delayed item is pending. The plan is to keep CPU -> NODE
566 * mapping valid and stable across CPU on/offlines. Once that
567 * happens, this workaround can be removed.
569 if (unlikely(node == NUMA_NO_NODE))
572 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
575 static unsigned int work_color_to_flags(int color)
577 return color << WORK_STRUCT_COLOR_SHIFT;
580 static int get_work_color(struct work_struct *work)
582 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
583 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
586 static int work_next_color(int color)
588 return (color + 1) % WORK_NR_COLORS;
592 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
593 * contain the pointer to the queued pwq. Once execution starts, the flag
594 * is cleared and the high bits contain OFFQ flags and pool ID.
596 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
597 * and clear_work_data() can be used to set the pwq, pool or clear
598 * work->data. These functions should only be called while the work is
599 * owned - ie. while the PENDING bit is set.
601 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
602 * corresponding to a work. Pool is available once the work has been
603 * queued anywhere after initialization until it is sync canceled. pwq is
604 * available only while the work item is queued.
606 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
607 * canceled. While being canceled, a work item may have its PENDING set
608 * but stay off timer and worklist for arbitrarily long and nobody should
609 * try to steal the PENDING bit.
611 static inline void set_work_data(struct work_struct *work, unsigned long data,
614 WARN_ON_ONCE(!work_pending(work));
615 atomic_long_set(&work->data, data | flags | work_static(work));
618 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
619 unsigned long extra_flags)
621 set_work_data(work, (unsigned long)pwq,
622 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
625 static void set_work_pool_and_keep_pending(struct work_struct *work,
628 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
629 WORK_STRUCT_PENDING);
632 static void set_work_pool_and_clear_pending(struct work_struct *work,
636 * The following wmb is paired with the implied mb in
637 * test_and_set_bit(PENDING) and ensures all updates to @work made
638 * here are visible to and precede any updates by the next PENDING
642 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
644 * The following mb guarantees that previous clear of a PENDING bit
645 * will not be reordered with any speculative LOADS or STORES from
646 * work->current_func, which is executed afterwards. This possible
647 * reordering can lead to a missed execution on attempt to qeueue
648 * the same @work. E.g. consider this case:
651 * ---------------------------- --------------------------------
653 * 1 STORE event_indicated
654 * 2 queue_work_on() {
655 * 3 test_and_set_bit(PENDING)
656 * 4 } set_..._and_clear_pending() {
657 * 5 set_work_data() # clear bit
659 * 7 work->current_func() {
660 * 8 LOAD event_indicated
663 * Without an explicit full barrier speculative LOAD on line 8 can
664 * be executed before CPU#0 does STORE on line 1. If that happens,
665 * CPU#0 observes the PENDING bit is still set and new execution of
666 * a @work is not queued in a hope, that CPU#1 will eventually
667 * finish the queued @work. Meanwhile CPU#1 does not see
668 * event_indicated is set, because speculative LOAD was executed
669 * before actual STORE.
674 static void clear_work_data(struct work_struct *work)
676 smp_wmb(); /* see set_work_pool_and_clear_pending() */
677 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
680 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
682 unsigned long data = atomic_long_read(&work->data);
684 if (data & WORK_STRUCT_PWQ)
685 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
691 * get_work_pool - return the worker_pool a given work was associated with
692 * @work: the work item of interest
694 * Pools are created and destroyed under wq_pool_mutex, and allows read
695 * access under sched-RCU read lock. As such, this function should be
696 * called under wq_pool_mutex or with preemption disabled.
698 * All fields of the returned pool are accessible as long as the above
699 * mentioned locking is in effect. If the returned pool needs to be used
700 * beyond the critical section, the caller is responsible for ensuring the
701 * returned pool is and stays online.
703 * Return: The worker_pool @work was last associated with. %NULL if none.
705 static struct worker_pool *get_work_pool(struct work_struct *work)
707 unsigned long data = atomic_long_read(&work->data);
710 assert_rcu_or_pool_mutex();
712 if (data & WORK_STRUCT_PWQ)
713 return ((struct pool_workqueue *)
714 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
716 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
717 if (pool_id == WORK_OFFQ_POOL_NONE)
720 return idr_find(&worker_pool_idr, pool_id);
724 * get_work_pool_id - return the worker pool ID a given work is associated with
725 * @work: the work item of interest
727 * Return: The worker_pool ID @work was last associated with.
728 * %WORK_OFFQ_POOL_NONE if none.
730 static int get_work_pool_id(struct work_struct *work)
732 unsigned long data = atomic_long_read(&work->data);
734 if (data & WORK_STRUCT_PWQ)
735 return ((struct pool_workqueue *)
736 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
738 return data >> WORK_OFFQ_POOL_SHIFT;
741 static void mark_work_canceling(struct work_struct *work)
743 unsigned long pool_id = get_work_pool_id(work);
745 pool_id <<= WORK_OFFQ_POOL_SHIFT;
746 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
749 static bool work_is_canceling(struct work_struct *work)
751 unsigned long data = atomic_long_read(&work->data);
753 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
757 * Policy functions. These define the policies on how the global worker
758 * pools are managed. Unless noted otherwise, these functions assume that
759 * they're being called with pool->lock held.
762 static bool __need_more_worker(struct worker_pool *pool)
764 return !atomic_read(&pool->nr_running);
768 * Need to wake up a worker? Called from anything but currently
771 * Note that, because unbound workers never contribute to nr_running, this
772 * function will always return %true for unbound pools as long as the
773 * worklist isn't empty.
775 static bool need_more_worker(struct worker_pool *pool)
777 return !list_empty(&pool->worklist) && __need_more_worker(pool);
780 /* Can I start working? Called from busy but !running workers. */
781 static bool may_start_working(struct worker_pool *pool)
783 return pool->nr_idle;
786 /* Do I need to keep working? Called from currently running workers. */
787 static bool keep_working(struct worker_pool *pool)
789 return !list_empty(&pool->worklist) &&
790 atomic_read(&pool->nr_running) <= 1;
793 /* Do we need a new worker? Called from manager. */
794 static bool need_to_create_worker(struct worker_pool *pool)
796 return need_more_worker(pool) && !may_start_working(pool);
799 /* Do we have too many workers and should some go away? */
800 static bool too_many_workers(struct worker_pool *pool)
802 bool managing = mutex_is_locked(&pool->manager_arb);
803 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
804 int nr_busy = pool->nr_workers - nr_idle;
806 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
813 /* Return the first idle worker. Safe with preemption disabled */
814 static struct worker *first_idle_worker(struct worker_pool *pool)
816 if (unlikely(list_empty(&pool->idle_list)))
819 return list_first_entry(&pool->idle_list, struct worker, entry);
823 * wake_up_worker - wake up an idle worker
824 * @pool: worker pool to wake worker from
826 * Wake up the first idle worker of @pool.
829 * spin_lock_irq(pool->lock).
831 static void wake_up_worker(struct worker_pool *pool)
833 struct worker *worker = first_idle_worker(pool);
836 wake_up_process(worker->task);
840 * wq_worker_waking_up - a worker is waking up
841 * @task: task waking up
842 * @cpu: CPU @task is waking up to
844 * This function is called during try_to_wake_up() when a worker is
848 * spin_lock_irq(rq->lock)
850 void wq_worker_waking_up(struct task_struct *task, int cpu)
852 struct worker *worker = kthread_data(task);
854 if (!(worker->flags & WORKER_NOT_RUNNING)) {
855 WARN_ON_ONCE(worker->pool->cpu != cpu);
856 atomic_inc(&worker->pool->nr_running);
861 * wq_worker_sleeping - a worker is going to sleep
862 * @task: task going to sleep
864 * This function is called during schedule() when a busy worker is
865 * going to sleep. Worker on the same cpu can be woken up by
866 * returning pointer to its task.
869 * spin_lock_irq(rq->lock)
872 * Worker task on @cpu to wake up, %NULL if none.
874 struct task_struct *wq_worker_sleeping(struct task_struct *task)
876 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
877 struct worker_pool *pool;
880 * Rescuers, which may not have all the fields set up like normal
881 * workers, also reach here, let's not access anything before
882 * checking NOT_RUNNING.
884 if (worker->flags & WORKER_NOT_RUNNING)
889 /* this can only happen on the local cpu */
890 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
894 * The counterpart of the following dec_and_test, implied mb,
895 * worklist not empty test sequence is in insert_work().
896 * Please read comment there.
898 * NOT_RUNNING is clear. This means that we're bound to and
899 * running on the local cpu w/ rq lock held and preemption
900 * disabled, which in turn means that none else could be
901 * manipulating idle_list, so dereferencing idle_list without pool
904 if (atomic_dec_and_test(&pool->nr_running) &&
905 !list_empty(&pool->worklist))
906 to_wakeup = first_idle_worker(pool);
907 return to_wakeup ? to_wakeup->task : NULL;
911 * worker_set_flags - set worker flags and adjust nr_running accordingly
913 * @flags: flags to set
915 * Set @flags in @worker->flags and adjust nr_running accordingly.
918 * spin_lock_irq(pool->lock)
920 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
922 struct worker_pool *pool = worker->pool;
924 WARN_ON_ONCE(worker->task != current);
926 /* If transitioning into NOT_RUNNING, adjust nr_running. */
927 if ((flags & WORKER_NOT_RUNNING) &&
928 !(worker->flags & WORKER_NOT_RUNNING)) {
929 atomic_dec(&pool->nr_running);
932 worker->flags |= flags;
936 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
938 * @flags: flags to clear
940 * Clear @flags in @worker->flags and adjust nr_running accordingly.
943 * spin_lock_irq(pool->lock)
945 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
947 struct worker_pool *pool = worker->pool;
948 unsigned int oflags = worker->flags;
950 WARN_ON_ONCE(worker->task != current);
952 worker->flags &= ~flags;
955 * If transitioning out of NOT_RUNNING, increment nr_running. Note
956 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
957 * of multiple flags, not a single flag.
959 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
960 if (!(worker->flags & WORKER_NOT_RUNNING))
961 atomic_inc(&pool->nr_running);
965 * find_worker_executing_work - find worker which is executing a work
966 * @pool: pool of interest
967 * @work: work to find worker for
969 * Find a worker which is executing @work on @pool by searching
970 * @pool->busy_hash which is keyed by the address of @work. For a worker
971 * to match, its current execution should match the address of @work and
972 * its work function. This is to avoid unwanted dependency between
973 * unrelated work executions through a work item being recycled while still
976 * This is a bit tricky. A work item may be freed once its execution
977 * starts and nothing prevents the freed area from being recycled for
978 * another work item. If the same work item address ends up being reused
979 * before the original execution finishes, workqueue will identify the
980 * recycled work item as currently executing and make it wait until the
981 * current execution finishes, introducing an unwanted dependency.
983 * This function checks the work item address and work function to avoid
984 * false positives. Note that this isn't complete as one may construct a
985 * work function which can introduce dependency onto itself through a
986 * recycled work item. Well, if somebody wants to shoot oneself in the
987 * foot that badly, there's only so much we can do, and if such deadlock
988 * actually occurs, it should be easy to locate the culprit work function.
991 * spin_lock_irq(pool->lock).
994 * Pointer to worker which is executing @work if found, %NULL
997 static struct worker *find_worker_executing_work(struct worker_pool *pool,
998 struct work_struct *work)
1000 struct worker *worker;
1002 hash_for_each_possible(pool->busy_hash, worker, hentry,
1003 (unsigned long)work)
1004 if (worker->current_work == work &&
1005 worker->current_func == work->func)
1012 * move_linked_works - move linked works to a list
1013 * @work: start of series of works to be scheduled
1014 * @head: target list to append @work to
1015 * @nextp: out parameter for nested worklist walking
1017 * Schedule linked works starting from @work to @head. Work series to
1018 * be scheduled starts at @work and includes any consecutive work with
1019 * WORK_STRUCT_LINKED set in its predecessor.
1021 * If @nextp is not NULL, it's updated to point to the next work of
1022 * the last scheduled work. This allows move_linked_works() to be
1023 * nested inside outer list_for_each_entry_safe().
1026 * spin_lock_irq(pool->lock).
1028 static void move_linked_works(struct work_struct *work, struct list_head *head,
1029 struct work_struct **nextp)
1031 struct work_struct *n;
1034 * Linked worklist will always end before the end of the list,
1035 * use NULL for list head.
1037 list_for_each_entry_safe_from(work, n, NULL, entry) {
1038 list_move_tail(&work->entry, head);
1039 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1044 * If we're already inside safe list traversal and have moved
1045 * multiple works to the scheduled queue, the next position
1046 * needs to be updated.
1053 * get_pwq - get an extra reference on the specified pool_workqueue
1054 * @pwq: pool_workqueue to get
1056 * Obtain an extra reference on @pwq. The caller should guarantee that
1057 * @pwq has positive refcnt and be holding the matching pool->lock.
1059 static void get_pwq(struct pool_workqueue *pwq)
1061 lockdep_assert_held(&pwq->pool->lock);
1062 WARN_ON_ONCE(pwq->refcnt <= 0);
1067 * put_pwq - put a pool_workqueue reference
1068 * @pwq: pool_workqueue to put
1070 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1071 * destruction. The caller should be holding the matching pool->lock.
1073 static void put_pwq(struct pool_workqueue *pwq)
1075 lockdep_assert_held(&pwq->pool->lock);
1076 if (likely(--pwq->refcnt))
1078 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1081 * @pwq can't be released under pool->lock, bounce to
1082 * pwq_unbound_release_workfn(). This never recurses on the same
1083 * pool->lock as this path is taken only for unbound workqueues and
1084 * the release work item is scheduled on a per-cpu workqueue. To
1085 * avoid lockdep warning, unbound pool->locks are given lockdep
1086 * subclass of 1 in get_unbound_pool().
1088 schedule_work(&pwq->unbound_release_work);
1092 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1093 * @pwq: pool_workqueue to put (can be %NULL)
1095 * put_pwq() with locking. This function also allows %NULL @pwq.
1097 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1101 * As both pwqs and pools are sched-RCU protected, the
1102 * following lock operations are safe.
1104 spin_lock_irq(&pwq->pool->lock);
1106 spin_unlock_irq(&pwq->pool->lock);
1110 static void pwq_activate_delayed_work(struct work_struct *work)
1112 struct pool_workqueue *pwq = get_work_pwq(work);
1114 trace_workqueue_activate_work(work);
1115 if (list_empty(&pwq->pool->worklist))
1116 pwq->pool->watchdog_ts = jiffies;
1117 move_linked_works(work, &pwq->pool->worklist, NULL);
1118 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1122 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1124 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1125 struct work_struct, entry);
1127 pwq_activate_delayed_work(work);
1131 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1132 * @pwq: pwq of interest
1133 * @color: color of work which left the queue
1135 * A work either has completed or is removed from pending queue,
1136 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1139 * spin_lock_irq(pool->lock).
1141 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1143 /* uncolored work items don't participate in flushing or nr_active */
1144 if (color == WORK_NO_COLOR)
1147 pwq->nr_in_flight[color]--;
1150 if (!list_empty(&pwq->delayed_works)) {
1151 /* one down, submit a delayed one */
1152 if (pwq->nr_active < pwq->max_active)
1153 pwq_activate_first_delayed(pwq);
1156 /* is flush in progress and are we at the flushing tip? */
1157 if (likely(pwq->flush_color != color))
1160 /* are there still in-flight works? */
1161 if (pwq->nr_in_flight[color])
1164 /* this pwq is done, clear flush_color */
1165 pwq->flush_color = -1;
1168 * If this was the last pwq, wake up the first flusher. It
1169 * will handle the rest.
1171 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1172 complete(&pwq->wq->first_flusher->done);
1178 * try_to_grab_pending - steal work item from worklist and disable irq
1179 * @work: work item to steal
1180 * @is_dwork: @work is a delayed_work
1181 * @flags: place to store irq state
1183 * Try to grab PENDING bit of @work. This function can handle @work in any
1184 * stable state - idle, on timer or on worklist.
1187 * 1 if @work was pending and we successfully stole PENDING
1188 * 0 if @work was idle and we claimed PENDING
1189 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1190 * -ENOENT if someone else is canceling @work, this state may persist
1191 * for arbitrarily long
1194 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1195 * interrupted while holding PENDING and @work off queue, irq must be
1196 * disabled on entry. This, combined with delayed_work->timer being
1197 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1199 * On successful return, >= 0, irq is disabled and the caller is
1200 * responsible for releasing it using local_irq_restore(*@flags).
1202 * This function is safe to call from any context including IRQ handler.
1204 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1205 unsigned long *flags)
1207 struct worker_pool *pool;
1208 struct pool_workqueue *pwq;
1210 local_irq_save(*flags);
1212 /* try to steal the timer if it exists */
1214 struct delayed_work *dwork = to_delayed_work(work);
1217 * dwork->timer is irqsafe. If del_timer() fails, it's
1218 * guaranteed that the timer is not queued anywhere and not
1219 * running on the local CPU.
1221 if (likely(del_timer(&dwork->timer)))
1225 /* try to claim PENDING the normal way */
1226 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1230 * The queueing is in progress, or it is already queued. Try to
1231 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1233 pool = get_work_pool(work);
1237 spin_lock(&pool->lock);
1239 * work->data is guaranteed to point to pwq only while the work
1240 * item is queued on pwq->wq, and both updating work->data to point
1241 * to pwq on queueing and to pool on dequeueing are done under
1242 * pwq->pool->lock. This in turn guarantees that, if work->data
1243 * points to pwq which is associated with a locked pool, the work
1244 * item is currently queued on that pool.
1246 pwq = get_work_pwq(work);
1247 if (pwq && pwq->pool == pool) {
1248 debug_work_deactivate(work);
1251 * A delayed work item cannot be grabbed directly because
1252 * it might have linked NO_COLOR work items which, if left
1253 * on the delayed_list, will confuse pwq->nr_active
1254 * management later on and cause stall. Make sure the work
1255 * item is activated before grabbing.
1257 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1258 pwq_activate_delayed_work(work);
1260 list_del_init(&work->entry);
1261 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1263 /* work->data points to pwq iff queued, point to pool */
1264 set_work_pool_and_keep_pending(work, pool->id);
1266 spin_unlock(&pool->lock);
1269 spin_unlock(&pool->lock);
1271 local_irq_restore(*flags);
1272 if (work_is_canceling(work))
1279 * insert_work - insert a work into a pool
1280 * @pwq: pwq @work belongs to
1281 * @work: work to insert
1282 * @head: insertion point
1283 * @extra_flags: extra WORK_STRUCT_* flags to set
1285 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1286 * work_struct flags.
1289 * spin_lock_irq(pool->lock).
1291 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1292 struct list_head *head, unsigned int extra_flags)
1294 struct worker_pool *pool = pwq->pool;
1296 /* we own @work, set data and link */
1297 set_work_pwq(work, pwq, extra_flags);
1298 list_add_tail(&work->entry, head);
1302 * Ensure either wq_worker_sleeping() sees the above
1303 * list_add_tail() or we see zero nr_running to avoid workers lying
1304 * around lazily while there are works to be processed.
1308 if (__need_more_worker(pool))
1309 wake_up_worker(pool);
1313 * Test whether @work is being queued from another work executing on the
1316 static bool is_chained_work(struct workqueue_struct *wq)
1318 struct worker *worker;
1320 worker = current_wq_worker();
1322 * Return %true iff I'm a worker execuing a work item on @wq. If
1323 * I'm @worker, it's safe to dereference it without locking.
1325 return worker && worker->current_pwq->wq == wq;
1329 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1330 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1331 * avoid perturbing sensitive tasks.
1333 static int wq_select_unbound_cpu(int cpu)
1335 static bool printed_dbg_warning;
1338 if (likely(!wq_debug_force_rr_cpu)) {
1339 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1341 } else if (!printed_dbg_warning) {
1342 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1343 printed_dbg_warning = true;
1346 if (cpumask_empty(wq_unbound_cpumask))
1349 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1350 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1351 if (unlikely(new_cpu >= nr_cpu_ids)) {
1352 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1353 if (unlikely(new_cpu >= nr_cpu_ids))
1356 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1361 static void __queue_work(int cpu, struct workqueue_struct *wq,
1362 struct work_struct *work)
1364 struct pool_workqueue *pwq;
1365 struct worker_pool *last_pool;
1366 struct list_head *worklist;
1367 unsigned int work_flags;
1368 unsigned int req_cpu = cpu;
1371 * While a work item is PENDING && off queue, a task trying to
1372 * steal the PENDING will busy-loop waiting for it to either get
1373 * queued or lose PENDING. Grabbing PENDING and queueing should
1374 * happen with IRQ disabled.
1376 WARN_ON_ONCE(!irqs_disabled());
1378 debug_work_activate(work);
1380 /* if draining, only works from the same workqueue are allowed */
1381 if (unlikely(wq->flags & __WQ_DRAINING) &&
1382 WARN_ON_ONCE(!is_chained_work(wq)))
1385 if (req_cpu == WORK_CPU_UNBOUND)
1386 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1388 /* pwq which will be used unless @work is executing elsewhere */
1389 if (!(wq->flags & WQ_UNBOUND))
1390 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1392 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1395 * If @work was previously on a different pool, it might still be
1396 * running there, in which case the work needs to be queued on that
1397 * pool to guarantee non-reentrancy.
1399 last_pool = get_work_pool(work);
1400 if (last_pool && last_pool != pwq->pool) {
1401 struct worker *worker;
1403 spin_lock(&last_pool->lock);
1405 worker = find_worker_executing_work(last_pool, work);
1407 if (worker && worker->current_pwq->wq == wq) {
1408 pwq = worker->current_pwq;
1410 /* meh... not running there, queue here */
1411 spin_unlock(&last_pool->lock);
1412 spin_lock(&pwq->pool->lock);
1415 spin_lock(&pwq->pool->lock);
1419 * pwq is determined and locked. For unbound pools, we could have
1420 * raced with pwq release and it could already be dead. If its
1421 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1422 * without another pwq replacing it in the numa_pwq_tbl or while
1423 * work items are executing on it, so the retrying is guaranteed to
1424 * make forward-progress.
1426 if (unlikely(!pwq->refcnt)) {
1427 if (wq->flags & WQ_UNBOUND) {
1428 spin_unlock(&pwq->pool->lock);
1433 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1437 /* pwq determined, queue */
1438 trace_workqueue_queue_work(req_cpu, pwq, work);
1440 if (WARN_ON(!list_empty(&work->entry))) {
1441 spin_unlock(&pwq->pool->lock);
1445 pwq->nr_in_flight[pwq->work_color]++;
1446 work_flags = work_color_to_flags(pwq->work_color);
1448 if (likely(pwq->nr_active < pwq->max_active)) {
1449 trace_workqueue_activate_work(work);
1451 worklist = &pwq->pool->worklist;
1452 if (list_empty(worklist))
1453 pwq->pool->watchdog_ts = jiffies;
1455 work_flags |= WORK_STRUCT_DELAYED;
1456 worklist = &pwq->delayed_works;
1459 insert_work(pwq, work, worklist, work_flags);
1461 spin_unlock(&pwq->pool->lock);
1465 * queue_work_on - queue work on specific cpu
1466 * @cpu: CPU number to execute work on
1467 * @wq: workqueue to use
1468 * @work: work to queue
1470 * We queue the work to a specific CPU, the caller must ensure it
1473 * Return: %false if @work was already on a queue, %true otherwise.
1475 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1476 struct work_struct *work)
1479 unsigned long flags;
1481 local_irq_save(flags);
1483 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1484 __queue_work(cpu, wq, work);
1488 local_irq_restore(flags);
1491 EXPORT_SYMBOL(queue_work_on);
1493 void delayed_work_timer_fn(unsigned long __data)
1495 struct delayed_work *dwork = (struct delayed_work *)__data;
1497 /* should have been called from irqsafe timer with irq already off */
1498 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1500 EXPORT_SYMBOL(delayed_work_timer_fn);
1502 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1503 struct delayed_work *dwork, unsigned long delay)
1505 struct timer_list *timer = &dwork->timer;
1506 struct work_struct *work = &dwork->work;
1508 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1509 timer->data != (unsigned long)dwork);
1510 WARN_ON_ONCE(timer_pending(timer));
1511 WARN_ON_ONCE(!list_empty(&work->entry));
1514 * If @delay is 0, queue @dwork->work immediately. This is for
1515 * both optimization and correctness. The earliest @timer can
1516 * expire is on the closest next tick and delayed_work users depend
1517 * on that there's no such delay when @delay is 0.
1520 __queue_work(cpu, wq, &dwork->work);
1524 timer_stats_timer_set_start_info(&dwork->timer);
1528 timer->expires = jiffies + delay;
1530 if (unlikely(cpu != WORK_CPU_UNBOUND))
1531 add_timer_on(timer, cpu);
1537 * queue_delayed_work_on - queue work on specific CPU after delay
1538 * @cpu: CPU number to execute work on
1539 * @wq: workqueue to use
1540 * @dwork: work to queue
1541 * @delay: number of jiffies to wait before queueing
1543 * Return: %false if @work was already on a queue, %true otherwise. If
1544 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1547 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1548 struct delayed_work *dwork, unsigned long delay)
1550 struct work_struct *work = &dwork->work;
1552 unsigned long flags;
1554 /* read the comment in __queue_work() */
1555 local_irq_save(flags);
1557 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1558 __queue_delayed_work(cpu, wq, dwork, delay);
1562 local_irq_restore(flags);
1565 EXPORT_SYMBOL(queue_delayed_work_on);
1568 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1569 * @cpu: CPU number to execute work on
1570 * @wq: workqueue to use
1571 * @dwork: work to queue
1572 * @delay: number of jiffies to wait before queueing
1574 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1575 * modify @dwork's timer so that it expires after @delay. If @delay is
1576 * zero, @work is guaranteed to be scheduled immediately regardless of its
1579 * Return: %false if @dwork was idle and queued, %true if @dwork was
1580 * pending and its timer was modified.
1582 * This function is safe to call from any context including IRQ handler.
1583 * See try_to_grab_pending() for details.
1585 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1586 struct delayed_work *dwork, unsigned long delay)
1588 unsigned long flags;
1592 ret = try_to_grab_pending(&dwork->work, true, &flags);
1593 } while (unlikely(ret == -EAGAIN));
1595 if (likely(ret >= 0)) {
1596 __queue_delayed_work(cpu, wq, dwork, delay);
1597 local_irq_restore(flags);
1600 /* -ENOENT from try_to_grab_pending() becomes %true */
1603 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1606 * worker_enter_idle - enter idle state
1607 * @worker: worker which is entering idle state
1609 * @worker is entering idle state. Update stats and idle timer if
1613 * spin_lock_irq(pool->lock).
1615 static void worker_enter_idle(struct worker *worker)
1617 struct worker_pool *pool = worker->pool;
1619 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1620 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1621 (worker->hentry.next || worker->hentry.pprev)))
1624 /* can't use worker_set_flags(), also called from create_worker() */
1625 worker->flags |= WORKER_IDLE;
1627 worker->last_active = jiffies;
1629 /* idle_list is LIFO */
1630 list_add(&worker->entry, &pool->idle_list);
1632 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1633 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1636 * Sanity check nr_running. Because wq_unbind_fn() releases
1637 * pool->lock between setting %WORKER_UNBOUND and zapping
1638 * nr_running, the warning may trigger spuriously. Check iff
1639 * unbind is not in progress.
1641 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1642 pool->nr_workers == pool->nr_idle &&
1643 atomic_read(&pool->nr_running));
1647 * worker_leave_idle - leave idle state
1648 * @worker: worker which is leaving idle state
1650 * @worker is leaving idle state. Update stats.
1653 * spin_lock_irq(pool->lock).
1655 static void worker_leave_idle(struct worker *worker)
1657 struct worker_pool *pool = worker->pool;
1659 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1661 worker_clr_flags(worker, WORKER_IDLE);
1663 list_del_init(&worker->entry);
1666 static struct worker *alloc_worker(int node)
1668 struct worker *worker;
1670 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1672 INIT_LIST_HEAD(&worker->entry);
1673 INIT_LIST_HEAD(&worker->scheduled);
1674 INIT_LIST_HEAD(&worker->node);
1675 /* on creation a worker is in !idle && prep state */
1676 worker->flags = WORKER_PREP;
1682 * worker_attach_to_pool() - attach a worker to a pool
1683 * @worker: worker to be attached
1684 * @pool: the target pool
1686 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1687 * cpu-binding of @worker are kept coordinated with the pool across
1690 static void worker_attach_to_pool(struct worker *worker,
1691 struct worker_pool *pool)
1693 mutex_lock(&pool->attach_mutex);
1696 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1697 * online CPUs. It'll be re-applied when any of the CPUs come up.
1699 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1702 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1703 * stable across this function. See the comments above the
1704 * flag definition for details.
1706 if (pool->flags & POOL_DISASSOCIATED)
1707 worker->flags |= WORKER_UNBOUND;
1709 list_add_tail(&worker->node, &pool->workers);
1711 mutex_unlock(&pool->attach_mutex);
1715 * worker_detach_from_pool() - detach a worker from its pool
1716 * @worker: worker which is attached to its pool
1717 * @pool: the pool @worker is attached to
1719 * Undo the attaching which had been done in worker_attach_to_pool(). The
1720 * caller worker shouldn't access to the pool after detached except it has
1721 * other reference to the pool.
1723 static void worker_detach_from_pool(struct worker *worker,
1724 struct worker_pool *pool)
1726 struct completion *detach_completion = NULL;
1728 mutex_lock(&pool->attach_mutex);
1729 list_del(&worker->node);
1730 if (list_empty(&pool->workers))
1731 detach_completion = pool->detach_completion;
1732 mutex_unlock(&pool->attach_mutex);
1734 /* clear leftover flags without pool->lock after it is detached */
1735 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1737 if (detach_completion)
1738 complete(detach_completion);
1742 * create_worker - create a new workqueue worker
1743 * @pool: pool the new worker will belong to
1745 * Create and start a new worker which is attached to @pool.
1748 * Might sleep. Does GFP_KERNEL allocations.
1751 * Pointer to the newly created worker.
1753 static struct worker *create_worker(struct worker_pool *pool)
1755 struct worker *worker = NULL;
1759 /* ID is needed to determine kthread name */
1760 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1764 worker = alloc_worker(pool->node);
1768 worker->pool = pool;
1772 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1773 pool->attrs->nice < 0 ? "H" : "");
1775 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1777 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1778 "kworker/%s", id_buf);
1779 if (IS_ERR(worker->task))
1782 set_user_nice(worker->task, pool->attrs->nice);
1783 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1785 /* successful, attach the worker to the pool */
1786 worker_attach_to_pool(worker, pool);
1788 /* start the newly created worker */
1789 spin_lock_irq(&pool->lock);
1790 worker->pool->nr_workers++;
1791 worker_enter_idle(worker);
1792 wake_up_process(worker->task);
1793 spin_unlock_irq(&pool->lock);
1799 ida_simple_remove(&pool->worker_ida, id);
1805 * destroy_worker - destroy a workqueue worker
1806 * @worker: worker to be destroyed
1808 * Destroy @worker and adjust @pool stats accordingly. The worker should
1812 * spin_lock_irq(pool->lock).
1814 static void destroy_worker(struct worker *worker)
1816 struct worker_pool *pool = worker->pool;
1818 lockdep_assert_held(&pool->lock);
1820 /* sanity check frenzy */
1821 if (WARN_ON(worker->current_work) ||
1822 WARN_ON(!list_empty(&worker->scheduled)) ||
1823 WARN_ON(!(worker->flags & WORKER_IDLE)))
1829 list_del_init(&worker->entry);
1830 worker->flags |= WORKER_DIE;
1831 wake_up_process(worker->task);
1834 static void idle_worker_timeout(unsigned long __pool)
1836 struct worker_pool *pool = (void *)__pool;
1838 spin_lock_irq(&pool->lock);
1840 while (too_many_workers(pool)) {
1841 struct worker *worker;
1842 unsigned long expires;
1844 /* idle_list is kept in LIFO order, check the last one */
1845 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1846 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1848 if (time_before(jiffies, expires)) {
1849 mod_timer(&pool->idle_timer, expires);
1853 destroy_worker(worker);
1856 spin_unlock_irq(&pool->lock);
1859 static void send_mayday(struct work_struct *work)
1861 struct pool_workqueue *pwq = get_work_pwq(work);
1862 struct workqueue_struct *wq = pwq->wq;
1864 lockdep_assert_held(&wq_mayday_lock);
1869 /* mayday mayday mayday */
1870 if (list_empty(&pwq->mayday_node)) {
1872 * If @pwq is for an unbound wq, its base ref may be put at
1873 * any time due to an attribute change. Pin @pwq until the
1874 * rescuer is done with it.
1877 list_add_tail(&pwq->mayday_node, &wq->maydays);
1878 wake_up_process(wq->rescuer->task);
1882 static void pool_mayday_timeout(unsigned long __pool)
1884 struct worker_pool *pool = (void *)__pool;
1885 struct work_struct *work;
1887 spin_lock_irq(&pool->lock);
1888 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1890 if (need_to_create_worker(pool)) {
1892 * We've been trying to create a new worker but
1893 * haven't been successful. We might be hitting an
1894 * allocation deadlock. Send distress signals to
1897 list_for_each_entry(work, &pool->worklist, entry)
1901 spin_unlock(&wq_mayday_lock);
1902 spin_unlock_irq(&pool->lock);
1904 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1908 * maybe_create_worker - create a new worker if necessary
1909 * @pool: pool to create a new worker for
1911 * Create a new worker for @pool if necessary. @pool is guaranteed to
1912 * have at least one idle worker on return from this function. If
1913 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1914 * sent to all rescuers with works scheduled on @pool to resolve
1915 * possible allocation deadlock.
1917 * On return, need_to_create_worker() is guaranteed to be %false and
1918 * may_start_working() %true.
1921 * spin_lock_irq(pool->lock) which may be released and regrabbed
1922 * multiple times. Does GFP_KERNEL allocations. Called only from
1925 static void maybe_create_worker(struct worker_pool *pool)
1926 __releases(&pool->lock)
1927 __acquires(&pool->lock)
1930 spin_unlock_irq(&pool->lock);
1932 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1933 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1936 if (create_worker(pool) || !need_to_create_worker(pool))
1939 schedule_timeout_interruptible(CREATE_COOLDOWN);
1941 if (!need_to_create_worker(pool))
1945 del_timer_sync(&pool->mayday_timer);
1946 spin_lock_irq(&pool->lock);
1948 * This is necessary even after a new worker was just successfully
1949 * created as @pool->lock was dropped and the new worker might have
1950 * already become busy.
1952 if (need_to_create_worker(pool))
1957 * manage_workers - manage worker pool
1960 * Assume the manager role and manage the worker pool @worker belongs
1961 * to. At any given time, there can be only zero or one manager per
1962 * pool. The exclusion is handled automatically by this function.
1964 * The caller can safely start processing works on false return. On
1965 * true return, it's guaranteed that need_to_create_worker() is false
1966 * and may_start_working() is true.
1969 * spin_lock_irq(pool->lock) which may be released and regrabbed
1970 * multiple times. Does GFP_KERNEL allocations.
1973 * %false if the pool doesn't need management and the caller can safely
1974 * start processing works, %true if management function was performed and
1975 * the conditions that the caller verified before calling the function may
1976 * no longer be true.
1978 static bool manage_workers(struct worker *worker)
1980 struct worker_pool *pool = worker->pool;
1983 * Anyone who successfully grabs manager_arb wins the arbitration
1984 * and becomes the manager. mutex_trylock() on pool->manager_arb
1985 * failure while holding pool->lock reliably indicates that someone
1986 * else is managing the pool and the worker which failed trylock
1987 * can proceed to executing work items. This means that anyone
1988 * grabbing manager_arb is responsible for actually performing
1989 * manager duties. If manager_arb is grabbed and released without
1990 * actual management, the pool may stall indefinitely.
1992 if (!mutex_trylock(&pool->manager_arb))
1994 pool->manager = worker;
1996 maybe_create_worker(pool);
1998 pool->manager = NULL;
1999 mutex_unlock(&pool->manager_arb);
2004 * process_one_work - process single work
2006 * @work: work to process
2008 * Process @work. This function contains all the logics necessary to
2009 * process a single work including synchronization against and
2010 * interaction with other workers on the same cpu, queueing and
2011 * flushing. As long as context requirement is met, any worker can
2012 * call this function to process a work.
2015 * spin_lock_irq(pool->lock) which is released and regrabbed.
2017 static void process_one_work(struct worker *worker, struct work_struct *work)
2018 __releases(&pool->lock)
2019 __acquires(&pool->lock)
2021 struct pool_workqueue *pwq = get_work_pwq(work);
2022 struct worker_pool *pool = worker->pool;
2023 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2025 struct worker *collision;
2026 #ifdef CONFIG_LOCKDEP
2028 * It is permissible to free the struct work_struct from
2029 * inside the function that is called from it, this we need to
2030 * take into account for lockdep too. To avoid bogus "held
2031 * lock freed" warnings as well as problems when looking into
2032 * work->lockdep_map, make a copy and use that here.
2034 struct lockdep_map lockdep_map;
2036 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2038 /* ensure we're on the correct CPU */
2039 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2040 raw_smp_processor_id() != pool->cpu);
2043 * A single work shouldn't be executed concurrently by
2044 * multiple workers on a single cpu. Check whether anyone is
2045 * already processing the work. If so, defer the work to the
2046 * currently executing one.
2048 collision = find_worker_executing_work(pool, work);
2049 if (unlikely(collision)) {
2050 move_linked_works(work, &collision->scheduled, NULL);
2054 /* claim and dequeue */
2055 debug_work_deactivate(work);
2056 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2057 worker->current_work = work;
2058 worker->current_func = work->func;
2059 worker->current_pwq = pwq;
2060 work_color = get_work_color(work);
2062 list_del_init(&work->entry);
2065 * CPU intensive works don't participate in concurrency management.
2066 * They're the scheduler's responsibility. This takes @worker out
2067 * of concurrency management and the next code block will chain
2068 * execution of the pending work items.
2070 if (unlikely(cpu_intensive))
2071 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2074 * Wake up another worker if necessary. The condition is always
2075 * false for normal per-cpu workers since nr_running would always
2076 * be >= 1 at this point. This is used to chain execution of the
2077 * pending work items for WORKER_NOT_RUNNING workers such as the
2078 * UNBOUND and CPU_INTENSIVE ones.
2080 if (need_more_worker(pool))
2081 wake_up_worker(pool);
2084 * Record the last pool and clear PENDING which should be the last
2085 * update to @work. Also, do this inside @pool->lock so that
2086 * PENDING and queued state changes happen together while IRQ is
2089 set_work_pool_and_clear_pending(work, pool->id);
2091 spin_unlock_irq(&pool->lock);
2093 lock_map_acquire_read(&pwq->wq->lockdep_map);
2094 lock_map_acquire(&lockdep_map);
2095 trace_workqueue_execute_start(work);
2096 worker->current_func(work);
2098 * While we must be careful to not use "work" after this, the trace
2099 * point will only record its address.
2101 trace_workqueue_execute_end(work);
2102 lock_map_release(&lockdep_map);
2103 lock_map_release(&pwq->wq->lockdep_map);
2105 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2106 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2107 " last function: %pf\n",
2108 current->comm, preempt_count(), task_pid_nr(current),
2109 worker->current_func);
2110 debug_show_held_locks(current);
2115 * The following prevents a kworker from hogging CPU on !PREEMPT
2116 * kernels, where a requeueing work item waiting for something to
2117 * happen could deadlock with stop_machine as such work item could
2118 * indefinitely requeue itself while all other CPUs are trapped in
2119 * stop_machine. At the same time, report a quiescent RCU state so
2120 * the same condition doesn't freeze RCU.
2122 cond_resched_rcu_qs();
2124 spin_lock_irq(&pool->lock);
2126 /* clear cpu intensive status */
2127 if (unlikely(cpu_intensive))
2128 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2130 /* we're done with it, release */
2131 hash_del(&worker->hentry);
2132 worker->current_work = NULL;
2133 worker->current_func = NULL;
2134 worker->current_pwq = NULL;
2135 worker->desc_valid = false;
2136 pwq_dec_nr_in_flight(pwq, work_color);
2140 * process_scheduled_works - process scheduled works
2143 * Process all scheduled works. Please note that the scheduled list
2144 * may change while processing a work, so this function repeatedly
2145 * fetches a work from the top and executes it.
2148 * spin_lock_irq(pool->lock) which may be released and regrabbed
2151 static void process_scheduled_works(struct worker *worker)
2153 while (!list_empty(&worker->scheduled)) {
2154 struct work_struct *work = list_first_entry(&worker->scheduled,
2155 struct work_struct, entry);
2156 process_one_work(worker, work);
2161 * worker_thread - the worker thread function
2164 * The worker thread function. All workers belong to a worker_pool -
2165 * either a per-cpu one or dynamic unbound one. These workers process all
2166 * work items regardless of their specific target workqueue. The only
2167 * exception is work items which belong to workqueues with a rescuer which
2168 * will be explained in rescuer_thread().
2172 static int worker_thread(void *__worker)
2174 struct worker *worker = __worker;
2175 struct worker_pool *pool = worker->pool;
2177 /* tell the scheduler that this is a workqueue worker */
2178 worker->task->flags |= PF_WQ_WORKER;
2180 spin_lock_irq(&pool->lock);
2182 /* am I supposed to die? */
2183 if (unlikely(worker->flags & WORKER_DIE)) {
2184 spin_unlock_irq(&pool->lock);
2185 WARN_ON_ONCE(!list_empty(&worker->entry));
2186 worker->task->flags &= ~PF_WQ_WORKER;
2188 set_task_comm(worker->task, "kworker/dying");
2189 ida_simple_remove(&pool->worker_ida, worker->id);
2190 worker_detach_from_pool(worker, pool);
2195 worker_leave_idle(worker);
2197 /* no more worker necessary? */
2198 if (!need_more_worker(pool))
2201 /* do we need to manage? */
2202 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2206 * ->scheduled list can only be filled while a worker is
2207 * preparing to process a work or actually processing it.
2208 * Make sure nobody diddled with it while I was sleeping.
2210 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2213 * Finish PREP stage. We're guaranteed to have at least one idle
2214 * worker or that someone else has already assumed the manager
2215 * role. This is where @worker starts participating in concurrency
2216 * management if applicable and concurrency management is restored
2217 * after being rebound. See rebind_workers() for details.
2219 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2222 struct work_struct *work =
2223 list_first_entry(&pool->worklist,
2224 struct work_struct, entry);
2226 pool->watchdog_ts = jiffies;
2228 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2229 /* optimization path, not strictly necessary */
2230 process_one_work(worker, work);
2231 if (unlikely(!list_empty(&worker->scheduled)))
2232 process_scheduled_works(worker);
2234 move_linked_works(work, &worker->scheduled, NULL);
2235 process_scheduled_works(worker);
2237 } while (keep_working(pool));
2239 worker_set_flags(worker, WORKER_PREP);
2242 * pool->lock is held and there's no work to process and no need to
2243 * manage, sleep. Workers are woken up only while holding
2244 * pool->lock or from local cpu, so setting the current state
2245 * before releasing pool->lock is enough to prevent losing any
2248 worker_enter_idle(worker);
2249 __set_current_state(TASK_INTERRUPTIBLE);
2250 spin_unlock_irq(&pool->lock);
2256 * rescuer_thread - the rescuer thread function
2259 * Workqueue rescuer thread function. There's one rescuer for each
2260 * workqueue which has WQ_MEM_RECLAIM set.
2262 * Regular work processing on a pool may block trying to create a new
2263 * worker which uses GFP_KERNEL allocation which has slight chance of
2264 * developing into deadlock if some works currently on the same queue
2265 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2266 * the problem rescuer solves.
2268 * When such condition is possible, the pool summons rescuers of all
2269 * workqueues which have works queued on the pool and let them process
2270 * those works so that forward progress can be guaranteed.
2272 * This should happen rarely.
2276 static int rescuer_thread(void *__rescuer)
2278 struct worker *rescuer = __rescuer;
2279 struct workqueue_struct *wq = rescuer->rescue_wq;
2280 struct list_head *scheduled = &rescuer->scheduled;
2283 set_user_nice(current, RESCUER_NICE_LEVEL);
2286 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2287 * doesn't participate in concurrency management.
2289 rescuer->task->flags |= PF_WQ_WORKER;
2291 set_current_state(TASK_INTERRUPTIBLE);
2294 * By the time the rescuer is requested to stop, the workqueue
2295 * shouldn't have any work pending, but @wq->maydays may still have
2296 * pwq(s) queued. This can happen by non-rescuer workers consuming
2297 * all the work items before the rescuer got to them. Go through
2298 * @wq->maydays processing before acting on should_stop so that the
2299 * list is always empty on exit.
2301 should_stop = kthread_should_stop();
2303 /* see whether any pwq is asking for help */
2304 spin_lock_irq(&wq_mayday_lock);
2306 while (!list_empty(&wq->maydays)) {
2307 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2308 struct pool_workqueue, mayday_node);
2309 struct worker_pool *pool = pwq->pool;
2310 struct work_struct *work, *n;
2313 __set_current_state(TASK_RUNNING);
2314 list_del_init(&pwq->mayday_node);
2316 spin_unlock_irq(&wq_mayday_lock);
2318 worker_attach_to_pool(rescuer, pool);
2320 spin_lock_irq(&pool->lock);
2321 rescuer->pool = pool;
2324 * Slurp in all works issued via this workqueue and
2327 WARN_ON_ONCE(!list_empty(scheduled));
2328 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2329 if (get_work_pwq(work) == pwq) {
2331 pool->watchdog_ts = jiffies;
2332 move_linked_works(work, scheduled, &n);
2337 if (!list_empty(scheduled)) {
2338 process_scheduled_works(rescuer);
2341 * The above execution of rescued work items could
2342 * have created more to rescue through
2343 * pwq_activate_first_delayed() or chained
2344 * queueing. Let's put @pwq back on mayday list so
2345 * that such back-to-back work items, which may be
2346 * being used to relieve memory pressure, don't
2347 * incur MAYDAY_INTERVAL delay inbetween.
2349 if (need_to_create_worker(pool)) {
2350 spin_lock(&wq_mayday_lock);
2352 list_move_tail(&pwq->mayday_node, &wq->maydays);
2353 spin_unlock(&wq_mayday_lock);
2358 * Put the reference grabbed by send_mayday(). @pool won't
2359 * go away while we're still attached to it.
2364 * Leave this pool. If need_more_worker() is %true, notify a
2365 * regular worker; otherwise, we end up with 0 concurrency
2366 * and stalling the execution.
2368 if (need_more_worker(pool))
2369 wake_up_worker(pool);
2371 rescuer->pool = NULL;
2372 spin_unlock_irq(&pool->lock);
2374 worker_detach_from_pool(rescuer, pool);
2376 spin_lock_irq(&wq_mayday_lock);
2379 spin_unlock_irq(&wq_mayday_lock);
2382 __set_current_state(TASK_RUNNING);
2383 rescuer->task->flags &= ~PF_WQ_WORKER;
2387 /* rescuers should never participate in concurrency management */
2388 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2394 * check_flush_dependency - check for flush dependency sanity
2395 * @target_wq: workqueue being flushed
2396 * @target_work: work item being flushed (NULL for workqueue flushes)
2398 * %current is trying to flush the whole @target_wq or @target_work on it.
2399 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2400 * reclaiming memory or running on a workqueue which doesn't have
2401 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2404 static void check_flush_dependency(struct workqueue_struct *target_wq,
2405 struct work_struct *target_work)
2407 work_func_t target_func = target_work ? target_work->func : NULL;
2408 struct worker *worker;
2410 if (target_wq->flags & WQ_MEM_RECLAIM)
2413 worker = current_wq_worker();
2415 WARN_ONCE(current->flags & PF_MEMALLOC,
2416 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2417 current->pid, current->comm, target_wq->name, target_func);
2418 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2419 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2420 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2421 worker->current_pwq->wq->name, worker->current_func,
2422 target_wq->name, target_func);
2426 struct work_struct work;
2427 struct completion done;
2428 struct task_struct *task; /* purely informational */
2431 static void wq_barrier_func(struct work_struct *work)
2433 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2434 complete(&barr->done);
2438 * insert_wq_barrier - insert a barrier work
2439 * @pwq: pwq to insert barrier into
2440 * @barr: wq_barrier to insert
2441 * @target: target work to attach @barr to
2442 * @worker: worker currently executing @target, NULL if @target is not executing
2444 * @barr is linked to @target such that @barr is completed only after
2445 * @target finishes execution. Please note that the ordering
2446 * guarantee is observed only with respect to @target and on the local
2449 * Currently, a queued barrier can't be canceled. This is because
2450 * try_to_grab_pending() can't determine whether the work to be
2451 * grabbed is at the head of the queue and thus can't clear LINKED
2452 * flag of the previous work while there must be a valid next work
2453 * after a work with LINKED flag set.
2455 * Note that when @worker is non-NULL, @target may be modified
2456 * underneath us, so we can't reliably determine pwq from @target.
2459 * spin_lock_irq(pool->lock).
2461 static void insert_wq_barrier(struct pool_workqueue *pwq,
2462 struct wq_barrier *barr,
2463 struct work_struct *target, struct worker *worker)
2465 struct list_head *head;
2466 unsigned int linked = 0;
2469 * debugobject calls are safe here even with pool->lock locked
2470 * as we know for sure that this will not trigger any of the
2471 * checks and call back into the fixup functions where we
2474 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2475 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2476 init_completion(&barr->done);
2477 barr->task = current;
2480 * If @target is currently being executed, schedule the
2481 * barrier to the worker; otherwise, put it after @target.
2484 head = worker->scheduled.next;
2486 unsigned long *bits = work_data_bits(target);
2488 head = target->entry.next;
2489 /* there can already be other linked works, inherit and set */
2490 linked = *bits & WORK_STRUCT_LINKED;
2491 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2494 debug_work_activate(&barr->work);
2495 insert_work(pwq, &barr->work, head,
2496 work_color_to_flags(WORK_NO_COLOR) | linked);
2500 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2501 * @wq: workqueue being flushed
2502 * @flush_color: new flush color, < 0 for no-op
2503 * @work_color: new work color, < 0 for no-op
2505 * Prepare pwqs for workqueue flushing.
2507 * If @flush_color is non-negative, flush_color on all pwqs should be
2508 * -1. If no pwq has in-flight commands at the specified color, all
2509 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2510 * has in flight commands, its pwq->flush_color is set to
2511 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2512 * wakeup logic is armed and %true is returned.
2514 * The caller should have initialized @wq->first_flusher prior to
2515 * calling this function with non-negative @flush_color. If
2516 * @flush_color is negative, no flush color update is done and %false
2519 * If @work_color is non-negative, all pwqs should have the same
2520 * work_color which is previous to @work_color and all will be
2521 * advanced to @work_color.
2524 * mutex_lock(wq->mutex).
2527 * %true if @flush_color >= 0 and there's something to flush. %false
2530 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2531 int flush_color, int work_color)
2534 struct pool_workqueue *pwq;
2536 if (flush_color >= 0) {
2537 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2538 atomic_set(&wq->nr_pwqs_to_flush, 1);
2541 for_each_pwq(pwq, wq) {
2542 struct worker_pool *pool = pwq->pool;
2544 spin_lock_irq(&pool->lock);
2546 if (flush_color >= 0) {
2547 WARN_ON_ONCE(pwq->flush_color != -1);
2549 if (pwq->nr_in_flight[flush_color]) {
2550 pwq->flush_color = flush_color;
2551 atomic_inc(&wq->nr_pwqs_to_flush);
2556 if (work_color >= 0) {
2557 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2558 pwq->work_color = work_color;
2561 spin_unlock_irq(&pool->lock);
2564 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2565 complete(&wq->first_flusher->done);
2571 * flush_workqueue - ensure that any scheduled work has run to completion.
2572 * @wq: workqueue to flush
2574 * This function sleeps until all work items which were queued on entry
2575 * have finished execution, but it is not livelocked by new incoming ones.
2577 void flush_workqueue(struct workqueue_struct *wq)
2579 struct wq_flusher this_flusher = {
2580 .list = LIST_HEAD_INIT(this_flusher.list),
2582 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2586 lock_map_acquire(&wq->lockdep_map);
2587 lock_map_release(&wq->lockdep_map);
2589 mutex_lock(&wq->mutex);
2592 * Start-to-wait phase
2594 next_color = work_next_color(wq->work_color);
2596 if (next_color != wq->flush_color) {
2598 * Color space is not full. The current work_color
2599 * becomes our flush_color and work_color is advanced
2602 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2603 this_flusher.flush_color = wq->work_color;
2604 wq->work_color = next_color;
2606 if (!wq->first_flusher) {
2607 /* no flush in progress, become the first flusher */
2608 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2610 wq->first_flusher = &this_flusher;
2612 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2614 /* nothing to flush, done */
2615 wq->flush_color = next_color;
2616 wq->first_flusher = NULL;
2621 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2622 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2623 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2627 * Oops, color space is full, wait on overflow queue.
2628 * The next flush completion will assign us
2629 * flush_color and transfer to flusher_queue.
2631 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2634 check_flush_dependency(wq, NULL);
2636 mutex_unlock(&wq->mutex);
2638 wait_for_completion(&this_flusher.done);
2641 * Wake-up-and-cascade phase
2643 * First flushers are responsible for cascading flushes and
2644 * handling overflow. Non-first flushers can simply return.
2646 if (wq->first_flusher != &this_flusher)
2649 mutex_lock(&wq->mutex);
2651 /* we might have raced, check again with mutex held */
2652 if (wq->first_flusher != &this_flusher)
2655 wq->first_flusher = NULL;
2657 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2658 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2661 struct wq_flusher *next, *tmp;
2663 /* complete all the flushers sharing the current flush color */
2664 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2665 if (next->flush_color != wq->flush_color)
2667 list_del_init(&next->list);
2668 complete(&next->done);
2671 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2672 wq->flush_color != work_next_color(wq->work_color));
2674 /* this flush_color is finished, advance by one */
2675 wq->flush_color = work_next_color(wq->flush_color);
2677 /* one color has been freed, handle overflow queue */
2678 if (!list_empty(&wq->flusher_overflow)) {
2680 * Assign the same color to all overflowed
2681 * flushers, advance work_color and append to
2682 * flusher_queue. This is the start-to-wait
2683 * phase for these overflowed flushers.
2685 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2686 tmp->flush_color = wq->work_color;
2688 wq->work_color = work_next_color(wq->work_color);
2690 list_splice_tail_init(&wq->flusher_overflow,
2691 &wq->flusher_queue);
2692 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2695 if (list_empty(&wq->flusher_queue)) {
2696 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2701 * Need to flush more colors. Make the next flusher
2702 * the new first flusher and arm pwqs.
2704 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2705 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2707 list_del_init(&next->list);
2708 wq->first_flusher = next;
2710 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2714 * Meh... this color is already done, clear first
2715 * flusher and repeat cascading.
2717 wq->first_flusher = NULL;
2721 mutex_unlock(&wq->mutex);
2723 EXPORT_SYMBOL(flush_workqueue);
2726 * drain_workqueue - drain a workqueue
2727 * @wq: workqueue to drain
2729 * Wait until the workqueue becomes empty. While draining is in progress,
2730 * only chain queueing is allowed. IOW, only currently pending or running
2731 * work items on @wq can queue further work items on it. @wq is flushed
2732 * repeatedly until it becomes empty. The number of flushing is determined
2733 * by the depth of chaining and should be relatively short. Whine if it
2736 void drain_workqueue(struct workqueue_struct *wq)
2738 unsigned int flush_cnt = 0;
2739 struct pool_workqueue *pwq;
2742 * __queue_work() needs to test whether there are drainers, is much
2743 * hotter than drain_workqueue() and already looks at @wq->flags.
2744 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2746 mutex_lock(&wq->mutex);
2747 if (!wq->nr_drainers++)
2748 wq->flags |= __WQ_DRAINING;
2749 mutex_unlock(&wq->mutex);
2751 flush_workqueue(wq);
2753 mutex_lock(&wq->mutex);
2755 for_each_pwq(pwq, wq) {
2758 spin_lock_irq(&pwq->pool->lock);
2759 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2760 spin_unlock_irq(&pwq->pool->lock);
2765 if (++flush_cnt == 10 ||
2766 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2767 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2768 wq->name, flush_cnt);
2770 mutex_unlock(&wq->mutex);
2774 if (!--wq->nr_drainers)
2775 wq->flags &= ~__WQ_DRAINING;
2776 mutex_unlock(&wq->mutex);
2778 EXPORT_SYMBOL_GPL(drain_workqueue);
2780 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2782 struct worker *worker = NULL;
2783 struct worker_pool *pool;
2784 struct pool_workqueue *pwq;
2788 local_irq_disable();
2789 pool = get_work_pool(work);
2795 spin_lock(&pool->lock);
2796 /* see the comment in try_to_grab_pending() with the same code */
2797 pwq = get_work_pwq(work);
2799 if (unlikely(pwq->pool != pool))
2802 worker = find_worker_executing_work(pool, work);
2805 pwq = worker->current_pwq;
2808 check_flush_dependency(pwq->wq, work);
2810 insert_wq_barrier(pwq, barr, work, worker);
2811 spin_unlock_irq(&pool->lock);
2814 * If @max_active is 1 or rescuer is in use, flushing another work
2815 * item on the same workqueue may lead to deadlock. Make sure the
2816 * flusher is not running on the same workqueue by verifying write
2819 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2820 lock_map_acquire(&pwq->wq->lockdep_map);
2822 lock_map_acquire_read(&pwq->wq->lockdep_map);
2823 lock_map_release(&pwq->wq->lockdep_map);
2827 spin_unlock_irq(&pool->lock);
2832 * flush_work - wait for a work to finish executing the last queueing instance
2833 * @work: the work to flush
2835 * Wait until @work has finished execution. @work is guaranteed to be idle
2836 * on return if it hasn't been requeued since flush started.
2839 * %true if flush_work() waited for the work to finish execution,
2840 * %false if it was already idle.
2842 bool flush_work(struct work_struct *work)
2844 struct wq_barrier barr;
2846 lock_map_acquire(&work->lockdep_map);
2847 lock_map_release(&work->lockdep_map);
2849 if (start_flush_work(work, &barr)) {
2850 wait_for_completion(&barr.done);
2851 destroy_work_on_stack(&barr.work);
2857 EXPORT_SYMBOL_GPL(flush_work);
2861 struct work_struct *work;
2864 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2866 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2868 if (cwait->work != key)
2870 return autoremove_wake_function(wait, mode, sync, key);
2873 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2875 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2876 unsigned long flags;
2880 ret = try_to_grab_pending(work, is_dwork, &flags);
2882 * If someone else is already canceling, wait for it to
2883 * finish. flush_work() doesn't work for PREEMPT_NONE
2884 * because we may get scheduled between @work's completion
2885 * and the other canceling task resuming and clearing
2886 * CANCELING - flush_work() will return false immediately
2887 * as @work is no longer busy, try_to_grab_pending() will
2888 * return -ENOENT as @work is still being canceled and the
2889 * other canceling task won't be able to clear CANCELING as
2890 * we're hogging the CPU.
2892 * Let's wait for completion using a waitqueue. As this
2893 * may lead to the thundering herd problem, use a custom
2894 * wake function which matches @work along with exclusive
2897 if (unlikely(ret == -ENOENT)) {
2898 struct cwt_wait cwait;
2900 init_wait(&cwait.wait);
2901 cwait.wait.func = cwt_wakefn;
2904 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2905 TASK_UNINTERRUPTIBLE);
2906 if (work_is_canceling(work))
2908 finish_wait(&cancel_waitq, &cwait.wait);
2910 } while (unlikely(ret < 0));
2912 /* tell other tasks trying to grab @work to back off */
2913 mark_work_canceling(work);
2914 local_irq_restore(flags);
2917 clear_work_data(work);
2920 * Paired with prepare_to_wait() above so that either
2921 * waitqueue_active() is visible here or !work_is_canceling() is
2925 if (waitqueue_active(&cancel_waitq))
2926 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2932 * cancel_work_sync - cancel a work and wait for it to finish
2933 * @work: the work to cancel
2935 * Cancel @work and wait for its execution to finish. This function
2936 * can be used even if the work re-queues itself or migrates to
2937 * another workqueue. On return from this function, @work is
2938 * guaranteed to be not pending or executing on any CPU.
2940 * cancel_work_sync(&delayed_work->work) must not be used for
2941 * delayed_work's. Use cancel_delayed_work_sync() instead.
2943 * The caller must ensure that the workqueue on which @work was last
2944 * queued can't be destroyed before this function returns.
2947 * %true if @work was pending, %false otherwise.
2949 bool cancel_work_sync(struct work_struct *work)
2951 return __cancel_work_timer(work, false);
2953 EXPORT_SYMBOL_GPL(cancel_work_sync);
2956 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2957 * @dwork: the delayed work to flush
2959 * Delayed timer is cancelled and the pending work is queued for
2960 * immediate execution. Like flush_work(), this function only
2961 * considers the last queueing instance of @dwork.
2964 * %true if flush_work() waited for the work to finish execution,
2965 * %false if it was already idle.
2967 bool flush_delayed_work(struct delayed_work *dwork)
2969 local_irq_disable();
2970 if (del_timer_sync(&dwork->timer))
2971 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2973 return flush_work(&dwork->work);
2975 EXPORT_SYMBOL(flush_delayed_work);
2977 static bool __cancel_work(struct work_struct *work, bool is_dwork)
2979 unsigned long flags;
2983 ret = try_to_grab_pending(work, is_dwork, &flags);
2984 } while (unlikely(ret == -EAGAIN));
2986 if (unlikely(ret < 0))
2989 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
2990 local_irq_restore(flags);
2995 * See cancel_delayed_work()
2997 bool cancel_work(struct work_struct *work)
2999 return __cancel_work(work, false);
3003 * cancel_delayed_work - cancel a delayed work
3004 * @dwork: delayed_work to cancel
3006 * Kill off a pending delayed_work.
3008 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3012 * The work callback function may still be running on return, unless
3013 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3014 * use cancel_delayed_work_sync() to wait on it.
3016 * This function is safe to call from any context including IRQ handler.
3018 bool cancel_delayed_work(struct delayed_work *dwork)
3020 return __cancel_work(&dwork->work, true);
3022 EXPORT_SYMBOL(cancel_delayed_work);
3025 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3026 * @dwork: the delayed work cancel
3028 * This is cancel_work_sync() for delayed works.
3031 * %true if @dwork was pending, %false otherwise.
3033 bool cancel_delayed_work_sync(struct delayed_work *dwork)
3035 return __cancel_work_timer(&dwork->work, true);
3037 EXPORT_SYMBOL(cancel_delayed_work_sync);
3040 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3041 * @func: the function to call
3043 * schedule_on_each_cpu() executes @func on each online CPU using the
3044 * system workqueue and blocks until all CPUs have completed.
3045 * schedule_on_each_cpu() is very slow.
3048 * 0 on success, -errno on failure.
3050 int schedule_on_each_cpu(work_func_t func)
3053 struct work_struct __percpu *works;
3055 works = alloc_percpu(struct work_struct);
3061 for_each_online_cpu(cpu) {
3062 struct work_struct *work = per_cpu_ptr(works, cpu);
3064 INIT_WORK(work, func);
3065 schedule_work_on(cpu, work);
3068 for_each_online_cpu(cpu)
3069 flush_work(per_cpu_ptr(works, cpu));
3077 * execute_in_process_context - reliably execute the routine with user context
3078 * @fn: the function to execute
3079 * @ew: guaranteed storage for the execute work structure (must
3080 * be available when the work executes)
3082 * Executes the function immediately if process context is available,
3083 * otherwise schedules the function for delayed execution.
3085 * Return: 0 - function was executed
3086 * 1 - function was scheduled for execution
3088 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3090 if (!in_interrupt()) {
3095 INIT_WORK(&ew->work, fn);
3096 schedule_work(&ew->work);
3100 EXPORT_SYMBOL_GPL(execute_in_process_context);
3103 * free_workqueue_attrs - free a workqueue_attrs
3104 * @attrs: workqueue_attrs to free
3106 * Undo alloc_workqueue_attrs().
3108 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3111 free_cpumask_var(attrs->cpumask);
3117 * alloc_workqueue_attrs - allocate a workqueue_attrs
3118 * @gfp_mask: allocation mask to use
3120 * Allocate a new workqueue_attrs, initialize with default settings and
3123 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3125 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3127 struct workqueue_attrs *attrs;
3129 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3132 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3135 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3138 free_workqueue_attrs(attrs);
3142 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3143 const struct workqueue_attrs *from)
3145 to->nice = from->nice;
3146 cpumask_copy(to->cpumask, from->cpumask);
3148 * Unlike hash and equality test, this function doesn't ignore
3149 * ->no_numa as it is used for both pool and wq attrs. Instead,
3150 * get_unbound_pool() explicitly clears ->no_numa after copying.
3152 to->no_numa = from->no_numa;
3155 /* hash value of the content of @attr */
3156 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3160 hash = jhash_1word(attrs->nice, hash);
3161 hash = jhash(cpumask_bits(attrs->cpumask),
3162 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3166 /* content equality test */
3167 static bool wqattrs_equal(const struct workqueue_attrs *a,
3168 const struct workqueue_attrs *b)
3170 if (a->nice != b->nice)
3172 if (!cpumask_equal(a->cpumask, b->cpumask))
3178 * init_worker_pool - initialize a newly zalloc'd worker_pool
3179 * @pool: worker_pool to initialize
3181 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3183 * Return: 0 on success, -errno on failure. Even on failure, all fields
3184 * inside @pool proper are initialized and put_unbound_pool() can be called
3185 * on @pool safely to release it.
3187 static int init_worker_pool(struct worker_pool *pool)
3189 spin_lock_init(&pool->lock);
3192 pool->node = NUMA_NO_NODE;
3193 pool->flags |= POOL_DISASSOCIATED;
3194 pool->watchdog_ts = jiffies;
3195 INIT_LIST_HEAD(&pool->worklist);
3196 INIT_LIST_HEAD(&pool->idle_list);
3197 hash_init(pool->busy_hash);
3199 init_timer_deferrable(&pool->idle_timer);
3200 pool->idle_timer.function = idle_worker_timeout;
3201 pool->idle_timer.data = (unsigned long)pool;
3203 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3204 (unsigned long)pool);
3206 mutex_init(&pool->manager_arb);
3207 mutex_init(&pool->attach_mutex);
3208 INIT_LIST_HEAD(&pool->workers);
3210 ida_init(&pool->worker_ida);
3211 INIT_HLIST_NODE(&pool->hash_node);
3214 /* shouldn't fail above this point */
3215 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3221 static void rcu_free_wq(struct rcu_head *rcu)
3223 struct workqueue_struct *wq =
3224 container_of(rcu, struct workqueue_struct, rcu);
3226 if (!(wq->flags & WQ_UNBOUND))
3227 free_percpu(wq->cpu_pwqs);
3229 free_workqueue_attrs(wq->unbound_attrs);
3235 static void rcu_free_pool(struct rcu_head *rcu)
3237 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3239 ida_destroy(&pool->worker_ida);
3240 free_workqueue_attrs(pool->attrs);
3245 * put_unbound_pool - put a worker_pool
3246 * @pool: worker_pool to put
3248 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3249 * safe manner. get_unbound_pool() calls this function on its failure path
3250 * and this function should be able to release pools which went through,
3251 * successfully or not, init_worker_pool().
3253 * Should be called with wq_pool_mutex held.
3255 static void put_unbound_pool(struct worker_pool *pool)
3257 DECLARE_COMPLETION_ONSTACK(detach_completion);
3258 struct worker *worker;
3260 lockdep_assert_held(&wq_pool_mutex);
3266 if (WARN_ON(!(pool->cpu < 0)) ||
3267 WARN_ON(!list_empty(&pool->worklist)))
3270 /* release id and unhash */
3272 idr_remove(&worker_pool_idr, pool->id);
3273 hash_del(&pool->hash_node);
3276 * Become the manager and destroy all workers. Grabbing
3277 * manager_arb prevents @pool's workers from blocking on
3280 mutex_lock(&pool->manager_arb);
3282 spin_lock_irq(&pool->lock);
3283 while ((worker = first_idle_worker(pool)))
3284 destroy_worker(worker);
3285 WARN_ON(pool->nr_workers || pool->nr_idle);
3286 spin_unlock_irq(&pool->lock);
3288 mutex_lock(&pool->attach_mutex);
3289 if (!list_empty(&pool->workers))
3290 pool->detach_completion = &detach_completion;
3291 mutex_unlock(&pool->attach_mutex);
3293 if (pool->detach_completion)
3294 wait_for_completion(pool->detach_completion);
3296 mutex_unlock(&pool->manager_arb);
3298 /* shut down the timers */
3299 del_timer_sync(&pool->idle_timer);
3300 del_timer_sync(&pool->mayday_timer);
3302 /* sched-RCU protected to allow dereferences from get_work_pool() */
3303 call_rcu_sched(&pool->rcu, rcu_free_pool);
3307 * get_unbound_pool - get a worker_pool with the specified attributes
3308 * @attrs: the attributes of the worker_pool to get
3310 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3311 * reference count and return it. If there already is a matching
3312 * worker_pool, it will be used; otherwise, this function attempts to
3315 * Should be called with wq_pool_mutex held.
3317 * Return: On success, a worker_pool with the same attributes as @attrs.
3318 * On failure, %NULL.
3320 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3322 u32 hash = wqattrs_hash(attrs);
3323 struct worker_pool *pool;
3325 int target_node = NUMA_NO_NODE;
3327 lockdep_assert_held(&wq_pool_mutex);
3329 /* do we already have a matching pool? */
3330 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3331 if (wqattrs_equal(pool->attrs, attrs)) {
3337 /* if cpumask is contained inside a NUMA node, we belong to that node */
3338 if (wq_numa_enabled) {
3339 for_each_node(node) {
3340 if (cpumask_subset(attrs->cpumask,
3341 wq_numa_possible_cpumask[node])) {
3348 /* nope, create a new one */
3349 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3350 if (!pool || init_worker_pool(pool) < 0)
3353 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3354 copy_workqueue_attrs(pool->attrs, attrs);
3355 pool->node = target_node;
3358 * no_numa isn't a worker_pool attribute, always clear it. See
3359 * 'struct workqueue_attrs' comments for detail.
3361 pool->attrs->no_numa = false;
3363 if (worker_pool_assign_id(pool) < 0)
3366 /* create and start the initial worker */
3367 if (!create_worker(pool))
3371 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3376 put_unbound_pool(pool);
3380 static void rcu_free_pwq(struct rcu_head *rcu)
3382 kmem_cache_free(pwq_cache,
3383 container_of(rcu, struct pool_workqueue, rcu));
3387 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3388 * and needs to be destroyed.
3390 static void pwq_unbound_release_workfn(struct work_struct *work)
3392 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3393 unbound_release_work);
3394 struct workqueue_struct *wq = pwq->wq;
3395 struct worker_pool *pool = pwq->pool;
3398 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3401 mutex_lock(&wq->mutex);
3402 list_del_rcu(&pwq->pwqs_node);
3403 is_last = list_empty(&wq->pwqs);
3404 mutex_unlock(&wq->mutex);
3406 mutex_lock(&wq_pool_mutex);
3407 put_unbound_pool(pool);
3408 mutex_unlock(&wq_pool_mutex);
3410 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3413 * If we're the last pwq going away, @wq is already dead and no one
3414 * is gonna access it anymore. Schedule RCU free.
3417 call_rcu_sched(&wq->rcu, rcu_free_wq);
3421 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3422 * @pwq: target pool_workqueue
3424 * If @pwq isn't freezing, set @pwq->max_active to the associated
3425 * workqueue's saved_max_active and activate delayed work items
3426 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3428 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3430 struct workqueue_struct *wq = pwq->wq;
3431 bool freezable = wq->flags & WQ_FREEZABLE;
3433 /* for @wq->saved_max_active */
3434 lockdep_assert_held(&wq->mutex);
3436 /* fast exit for non-freezable wqs */
3437 if (!freezable && pwq->max_active == wq->saved_max_active)
3440 spin_lock_irq(&pwq->pool->lock);
3443 * During [un]freezing, the caller is responsible for ensuring that
3444 * this function is called at least once after @workqueue_freezing
3445 * is updated and visible.
3447 if (!freezable || !workqueue_freezing) {
3448 pwq->max_active = wq->saved_max_active;
3450 while (!list_empty(&pwq->delayed_works) &&
3451 pwq->nr_active < pwq->max_active)
3452 pwq_activate_first_delayed(pwq);
3455 * Need to kick a worker after thawed or an unbound wq's
3456 * max_active is bumped. It's a slow path. Do it always.
3458 wake_up_worker(pwq->pool);
3460 pwq->max_active = 0;
3463 spin_unlock_irq(&pwq->pool->lock);
3466 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3467 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3468 struct worker_pool *pool)
3470 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3472 memset(pwq, 0, sizeof(*pwq));
3476 pwq->flush_color = -1;
3478 INIT_LIST_HEAD(&pwq->delayed_works);
3479 INIT_LIST_HEAD(&pwq->pwqs_node);
3480 INIT_LIST_HEAD(&pwq->mayday_node);
3481 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3484 /* sync @pwq with the current state of its associated wq and link it */
3485 static void link_pwq(struct pool_workqueue *pwq)
3487 struct workqueue_struct *wq = pwq->wq;
3489 lockdep_assert_held(&wq->mutex);
3491 /* may be called multiple times, ignore if already linked */
3492 if (!list_empty(&pwq->pwqs_node))
3495 /* set the matching work_color */
3496 pwq->work_color = wq->work_color;
3498 /* sync max_active to the current setting */
3499 pwq_adjust_max_active(pwq);
3502 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3505 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3506 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3507 const struct workqueue_attrs *attrs)
3509 struct worker_pool *pool;
3510 struct pool_workqueue *pwq;
3512 lockdep_assert_held(&wq_pool_mutex);
3514 pool = get_unbound_pool(attrs);
3518 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3520 put_unbound_pool(pool);
3524 init_pwq(pwq, wq, pool);
3529 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3530 * @attrs: the wq_attrs of the default pwq of the target workqueue
3531 * @node: the target NUMA node
3532 * @cpu_going_down: if >= 0, the CPU to consider as offline
3533 * @cpumask: outarg, the resulting cpumask
3535 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3536 * @cpu_going_down is >= 0, that cpu is considered offline during
3537 * calculation. The result is stored in @cpumask.
3539 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3540 * enabled and @node has online CPUs requested by @attrs, the returned
3541 * cpumask is the intersection of the possible CPUs of @node and
3544 * The caller is responsible for ensuring that the cpumask of @node stays
3547 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3550 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3551 int cpu_going_down, cpumask_t *cpumask)
3553 if (!wq_numa_enabled || attrs->no_numa)
3556 /* does @node have any online CPUs @attrs wants? */
3557 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3558 if (cpu_going_down >= 0)
3559 cpumask_clear_cpu(cpu_going_down, cpumask);
3561 if (cpumask_empty(cpumask))
3564 /* yeap, return possible CPUs in @node that @attrs wants */
3565 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3566 return !cpumask_equal(cpumask, attrs->cpumask);
3569 cpumask_copy(cpumask, attrs->cpumask);
3573 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3574 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3576 struct pool_workqueue *pwq)
3578 struct pool_workqueue *old_pwq;
3580 lockdep_assert_held(&wq_pool_mutex);
3581 lockdep_assert_held(&wq->mutex);
3583 /* link_pwq() can handle duplicate calls */
3586 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3587 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3591 /* context to store the prepared attrs & pwqs before applying */
3592 struct apply_wqattrs_ctx {
3593 struct workqueue_struct *wq; /* target workqueue */
3594 struct workqueue_attrs *attrs; /* attrs to apply */
3595 struct list_head list; /* queued for batching commit */
3596 struct pool_workqueue *dfl_pwq;
3597 struct pool_workqueue *pwq_tbl[];
3600 /* free the resources after success or abort */
3601 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3607 put_pwq_unlocked(ctx->pwq_tbl[node]);
3608 put_pwq_unlocked(ctx->dfl_pwq);
3610 free_workqueue_attrs(ctx->attrs);
3616 /* allocate the attrs and pwqs for later installation */
3617 static struct apply_wqattrs_ctx *
3618 apply_wqattrs_prepare(struct workqueue_struct *wq,
3619 const struct workqueue_attrs *attrs)
3621 struct apply_wqattrs_ctx *ctx;
3622 struct workqueue_attrs *new_attrs, *tmp_attrs;
3625 lockdep_assert_held(&wq_pool_mutex);
3627 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3630 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3631 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3632 if (!ctx || !new_attrs || !tmp_attrs)
3636 * Calculate the attrs of the default pwq.
3637 * If the user configured cpumask doesn't overlap with the
3638 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3640 copy_workqueue_attrs(new_attrs, attrs);
3641 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3642 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3643 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3646 * We may create multiple pwqs with differing cpumasks. Make a
3647 * copy of @new_attrs which will be modified and used to obtain
3650 copy_workqueue_attrs(tmp_attrs, new_attrs);
3653 * If something goes wrong during CPU up/down, we'll fall back to
3654 * the default pwq covering whole @attrs->cpumask. Always create
3655 * it even if we don't use it immediately.
3657 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3661 for_each_node(node) {
3662 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3663 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3664 if (!ctx->pwq_tbl[node])
3667 ctx->dfl_pwq->refcnt++;
3668 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3672 /* save the user configured attrs and sanitize it. */
3673 copy_workqueue_attrs(new_attrs, attrs);
3674 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3675 ctx->attrs = new_attrs;
3678 free_workqueue_attrs(tmp_attrs);
3682 free_workqueue_attrs(tmp_attrs);
3683 free_workqueue_attrs(new_attrs);
3684 apply_wqattrs_cleanup(ctx);
3688 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3689 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3693 /* all pwqs have been created successfully, let's install'em */
3694 mutex_lock(&ctx->wq->mutex);
3696 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3698 /* save the previous pwq and install the new one */
3700 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3701 ctx->pwq_tbl[node]);
3703 /* @dfl_pwq might not have been used, ensure it's linked */
3704 link_pwq(ctx->dfl_pwq);
3705 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3707 mutex_unlock(&ctx->wq->mutex);
3710 static void apply_wqattrs_lock(void)
3712 /* CPUs should stay stable across pwq creations and installations */
3714 mutex_lock(&wq_pool_mutex);
3717 static void apply_wqattrs_unlock(void)
3719 mutex_unlock(&wq_pool_mutex);
3723 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3724 const struct workqueue_attrs *attrs)
3726 struct apply_wqattrs_ctx *ctx;
3728 /* only unbound workqueues can change attributes */
3729 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3732 /* creating multiple pwqs breaks ordering guarantee */
3733 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3736 ctx = apply_wqattrs_prepare(wq, attrs);
3740 /* the ctx has been prepared successfully, let's commit it */
3741 apply_wqattrs_commit(ctx);
3742 apply_wqattrs_cleanup(ctx);
3748 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3749 * @wq: the target workqueue
3750 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3752 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3753 * machines, this function maps a separate pwq to each NUMA node with
3754 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3755 * NUMA node it was issued on. Older pwqs are released as in-flight work
3756 * items finish. Note that a work item which repeatedly requeues itself
3757 * back-to-back will stay on its current pwq.
3759 * Performs GFP_KERNEL allocations.
3761 * Return: 0 on success and -errno on failure.
3763 int apply_workqueue_attrs(struct workqueue_struct *wq,
3764 const struct workqueue_attrs *attrs)
3768 apply_wqattrs_lock();
3769 ret = apply_workqueue_attrs_locked(wq, attrs);
3770 apply_wqattrs_unlock();
3776 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3777 * @wq: the target workqueue
3778 * @cpu: the CPU coming up or going down
3779 * @online: whether @cpu is coming up or going down
3781 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3782 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3785 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3786 * falls back to @wq->dfl_pwq which may not be optimal but is always
3789 * Note that when the last allowed CPU of a NUMA node goes offline for a
3790 * workqueue with a cpumask spanning multiple nodes, the workers which were
3791 * already executing the work items for the workqueue will lose their CPU
3792 * affinity and may execute on any CPU. This is similar to how per-cpu
3793 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3794 * affinity, it's the user's responsibility to flush the work item from
3797 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3800 int node = cpu_to_node(cpu);
3801 int cpu_off = online ? -1 : cpu;
3802 struct pool_workqueue *old_pwq = NULL, *pwq;
3803 struct workqueue_attrs *target_attrs;
3806 lockdep_assert_held(&wq_pool_mutex);
3808 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3809 wq->unbound_attrs->no_numa)
3813 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3814 * Let's use a preallocated one. The following buf is protected by
3815 * CPU hotplug exclusion.
3817 target_attrs = wq_update_unbound_numa_attrs_buf;
3818 cpumask = target_attrs->cpumask;
3820 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3821 pwq = unbound_pwq_by_node(wq, node);
3824 * Let's determine what needs to be done. If the target cpumask is
3825 * different from the default pwq's, we need to compare it to @pwq's
3826 * and create a new one if they don't match. If the target cpumask
3827 * equals the default pwq's, the default pwq should be used.
3829 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3830 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3836 /* create a new pwq */
3837 pwq = alloc_unbound_pwq(wq, target_attrs);
3839 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3844 /* Install the new pwq. */
3845 mutex_lock(&wq->mutex);
3846 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3850 mutex_lock(&wq->mutex);
3851 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3852 get_pwq(wq->dfl_pwq);
3853 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3854 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3856 mutex_unlock(&wq->mutex);
3857 put_pwq_unlocked(old_pwq);
3860 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3862 bool highpri = wq->flags & WQ_HIGHPRI;
3865 if (!(wq->flags & WQ_UNBOUND)) {
3866 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3870 for_each_possible_cpu(cpu) {
3871 struct pool_workqueue *pwq =
3872 per_cpu_ptr(wq->cpu_pwqs, cpu);
3873 struct worker_pool *cpu_pools =
3874 per_cpu(cpu_worker_pools, cpu);
3876 init_pwq(pwq, wq, &cpu_pools[highpri]);
3878 mutex_lock(&wq->mutex);
3880 mutex_unlock(&wq->mutex);
3883 } else if (wq->flags & __WQ_ORDERED) {
3884 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3885 /* there should only be single pwq for ordering guarantee */
3886 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3887 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3888 "ordering guarantee broken for workqueue %s\n", wq->name);
3891 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3895 static int wq_clamp_max_active(int max_active, unsigned int flags,
3898 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3900 if (max_active < 1 || max_active > lim)
3901 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3902 max_active, name, 1, lim);
3904 return clamp_val(max_active, 1, lim);
3907 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3910 struct lock_class_key *key,
3911 const char *lock_name, ...)
3913 size_t tbl_size = 0;
3915 struct workqueue_struct *wq;
3916 struct pool_workqueue *pwq;
3918 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3919 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3920 flags |= WQ_UNBOUND;
3922 /* allocate wq and format name */
3923 if (flags & WQ_UNBOUND)
3924 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3926 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3930 if (flags & WQ_UNBOUND) {
3931 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3932 if (!wq->unbound_attrs)
3936 va_start(args, lock_name);
3937 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3940 max_active = max_active ?: WQ_DFL_ACTIVE;
3941 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3945 wq->saved_max_active = max_active;
3946 mutex_init(&wq->mutex);
3947 atomic_set(&wq->nr_pwqs_to_flush, 0);
3948 INIT_LIST_HEAD(&wq->pwqs);
3949 INIT_LIST_HEAD(&wq->flusher_queue);
3950 INIT_LIST_HEAD(&wq->flusher_overflow);
3951 INIT_LIST_HEAD(&wq->maydays);
3953 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3954 INIT_LIST_HEAD(&wq->list);
3956 if (alloc_and_link_pwqs(wq) < 0)
3960 * Workqueues which may be used during memory reclaim should
3961 * have a rescuer to guarantee forward progress.
3963 if (flags & WQ_MEM_RECLAIM) {
3964 struct worker *rescuer;
3966 rescuer = alloc_worker(NUMA_NO_NODE);
3970 rescuer->rescue_wq = wq;
3971 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3973 if (IS_ERR(rescuer->task)) {
3978 wq->rescuer = rescuer;
3979 kthread_bind_mask(rescuer->task, cpu_possible_mask);
3980 wake_up_process(rescuer->task);
3983 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3987 * wq_pool_mutex protects global freeze state and workqueues list.
3988 * Grab it, adjust max_active and add the new @wq to workqueues
3991 mutex_lock(&wq_pool_mutex);
3993 mutex_lock(&wq->mutex);
3994 for_each_pwq(pwq, wq)
3995 pwq_adjust_max_active(pwq);
3996 mutex_unlock(&wq->mutex);
3998 list_add_tail_rcu(&wq->list, &workqueues);
4000 mutex_unlock(&wq_pool_mutex);
4005 free_workqueue_attrs(wq->unbound_attrs);
4009 destroy_workqueue(wq);
4012 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4015 * destroy_workqueue - safely terminate a workqueue
4016 * @wq: target workqueue
4018 * Safely destroy a workqueue. All work currently pending will be done first.
4020 void destroy_workqueue(struct workqueue_struct *wq)
4022 struct pool_workqueue *pwq;
4025 /* drain it before proceeding with destruction */
4026 drain_workqueue(wq);
4029 mutex_lock(&wq->mutex);
4030 for_each_pwq(pwq, wq) {
4033 for (i = 0; i < WORK_NR_COLORS; i++) {
4034 if (WARN_ON(pwq->nr_in_flight[i])) {
4035 mutex_unlock(&wq->mutex);
4040 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4041 WARN_ON(pwq->nr_active) ||
4042 WARN_ON(!list_empty(&pwq->delayed_works))) {
4043 mutex_unlock(&wq->mutex);
4047 mutex_unlock(&wq->mutex);
4050 * wq list is used to freeze wq, remove from list after
4051 * flushing is complete in case freeze races us.
4053 mutex_lock(&wq_pool_mutex);
4054 list_del_rcu(&wq->list);
4055 mutex_unlock(&wq_pool_mutex);
4057 workqueue_sysfs_unregister(wq);
4060 kthread_stop(wq->rescuer->task);
4062 if (!(wq->flags & WQ_UNBOUND)) {
4064 * The base ref is never dropped on per-cpu pwqs. Directly
4065 * schedule RCU free.
4067 call_rcu_sched(&wq->rcu, rcu_free_wq);
4070 * We're the sole accessor of @wq at this point. Directly
4071 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4072 * @wq will be freed when the last pwq is released.
4074 for_each_node(node) {
4075 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4076 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4077 put_pwq_unlocked(pwq);
4081 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4082 * put. Don't access it afterwards.
4086 put_pwq_unlocked(pwq);
4089 EXPORT_SYMBOL_GPL(destroy_workqueue);
4092 * workqueue_set_max_active - adjust max_active of a workqueue
4093 * @wq: target workqueue
4094 * @max_active: new max_active value.
4096 * Set max_active of @wq to @max_active.
4099 * Don't call from IRQ context.
4101 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4103 struct pool_workqueue *pwq;
4105 /* disallow meddling with max_active for ordered workqueues */
4106 if (WARN_ON(wq->flags & __WQ_ORDERED))
4109 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4111 mutex_lock(&wq->mutex);
4113 wq->saved_max_active = max_active;
4115 for_each_pwq(pwq, wq)
4116 pwq_adjust_max_active(pwq);
4118 mutex_unlock(&wq->mutex);
4120 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4123 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4125 * Determine whether %current is a workqueue rescuer. Can be used from
4126 * work functions to determine whether it's being run off the rescuer task.
4128 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4130 bool current_is_workqueue_rescuer(void)
4132 struct worker *worker = current_wq_worker();
4134 return worker && worker->rescue_wq;
4138 * workqueue_congested - test whether a workqueue is congested
4139 * @cpu: CPU in question
4140 * @wq: target workqueue
4142 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4143 * no synchronization around this function and the test result is
4144 * unreliable and only useful as advisory hints or for debugging.
4146 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4147 * Note that both per-cpu and unbound workqueues may be associated with
4148 * multiple pool_workqueues which have separate congested states. A
4149 * workqueue being congested on one CPU doesn't mean the workqueue is also
4150 * contested on other CPUs / NUMA nodes.
4153 * %true if congested, %false otherwise.
4155 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4157 struct pool_workqueue *pwq;
4160 rcu_read_lock_sched();
4162 if (cpu == WORK_CPU_UNBOUND)
4163 cpu = smp_processor_id();
4165 if (!(wq->flags & WQ_UNBOUND))
4166 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4168 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4170 ret = !list_empty(&pwq->delayed_works);
4171 rcu_read_unlock_sched();
4175 EXPORT_SYMBOL_GPL(workqueue_congested);
4178 * work_busy - test whether a work is currently pending or running
4179 * @work: the work to be tested
4181 * Test whether @work is currently pending or running. There is no
4182 * synchronization around this function and the test result is
4183 * unreliable and only useful as advisory hints or for debugging.
4186 * OR'd bitmask of WORK_BUSY_* bits.
4188 unsigned int work_busy(struct work_struct *work)
4190 struct worker_pool *pool;
4191 unsigned long flags;
4192 unsigned int ret = 0;
4194 if (work_pending(work))
4195 ret |= WORK_BUSY_PENDING;
4197 local_irq_save(flags);
4198 pool = get_work_pool(work);
4200 spin_lock(&pool->lock);
4201 if (find_worker_executing_work(pool, work))
4202 ret |= WORK_BUSY_RUNNING;
4203 spin_unlock(&pool->lock);
4205 local_irq_restore(flags);
4209 EXPORT_SYMBOL_GPL(work_busy);
4212 * set_worker_desc - set description for the current work item
4213 * @fmt: printf-style format string
4214 * @...: arguments for the format string
4216 * This function can be called by a running work function to describe what
4217 * the work item is about. If the worker task gets dumped, this
4218 * information will be printed out together to help debugging. The
4219 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4221 void set_worker_desc(const char *fmt, ...)
4223 struct worker *worker = current_wq_worker();
4227 va_start(args, fmt);
4228 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4230 worker->desc_valid = true;
4235 * print_worker_info - print out worker information and description
4236 * @log_lvl: the log level to use when printing
4237 * @task: target task
4239 * If @task is a worker and currently executing a work item, print out the
4240 * name of the workqueue being serviced and worker description set with
4241 * set_worker_desc() by the currently executing work item.
4243 * This function can be safely called on any task as long as the
4244 * task_struct itself is accessible. While safe, this function isn't
4245 * synchronized and may print out mixups or garbages of limited length.
4247 void print_worker_info(const char *log_lvl, struct task_struct *task)
4249 work_func_t *fn = NULL;
4250 char name[WQ_NAME_LEN] = { };
4251 char desc[WORKER_DESC_LEN] = { };
4252 struct pool_workqueue *pwq = NULL;
4253 struct workqueue_struct *wq = NULL;
4254 bool desc_valid = false;
4255 struct worker *worker;
4257 if (!(task->flags & PF_WQ_WORKER))
4261 * This function is called without any synchronization and @task
4262 * could be in any state. Be careful with dereferences.
4264 worker = kthread_probe_data(task);
4267 * Carefully copy the associated workqueue's workfn and name. Keep
4268 * the original last '\0' in case the original contains garbage.
4270 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4271 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4272 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4273 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4275 /* copy worker description */
4276 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4278 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4280 if (fn || name[0] || desc[0]) {
4281 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4283 pr_cont(" (%s)", desc);
4288 static void pr_cont_pool_info(struct worker_pool *pool)
4290 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4291 if (pool->node != NUMA_NO_NODE)
4292 pr_cont(" node=%d", pool->node);
4293 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4296 static void pr_cont_work(bool comma, struct work_struct *work)
4298 if (work->func == wq_barrier_func) {
4299 struct wq_barrier *barr;
4301 barr = container_of(work, struct wq_barrier, work);
4303 pr_cont("%s BAR(%d)", comma ? "," : "",
4304 task_pid_nr(barr->task));
4306 pr_cont("%s %pf", comma ? "," : "", work->func);
4310 static void show_pwq(struct pool_workqueue *pwq)
4312 struct worker_pool *pool = pwq->pool;
4313 struct work_struct *work;
4314 struct worker *worker;
4315 bool has_in_flight = false, has_pending = false;
4318 pr_info(" pwq %d:", pool->id);
4319 pr_cont_pool_info(pool);
4321 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4322 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4324 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4325 if (worker->current_pwq == pwq) {
4326 has_in_flight = true;
4330 if (has_in_flight) {
4333 pr_info(" in-flight:");
4334 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4335 if (worker->current_pwq != pwq)
4338 pr_cont("%s %d%s:%pf", comma ? "," : "",
4339 task_pid_nr(worker->task),
4340 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4341 worker->current_func);
4342 list_for_each_entry(work, &worker->scheduled, entry)
4343 pr_cont_work(false, work);
4349 list_for_each_entry(work, &pool->worklist, entry) {
4350 if (get_work_pwq(work) == pwq) {
4358 pr_info(" pending:");
4359 list_for_each_entry(work, &pool->worklist, entry) {
4360 if (get_work_pwq(work) != pwq)
4363 pr_cont_work(comma, work);
4364 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4369 if (!list_empty(&pwq->delayed_works)) {
4372 pr_info(" delayed:");
4373 list_for_each_entry(work, &pwq->delayed_works, entry) {
4374 pr_cont_work(comma, work);
4375 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4382 * show_workqueue_state - dump workqueue state
4384 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4385 * all busy workqueues and pools.
4387 void show_workqueue_state(void)
4389 struct workqueue_struct *wq;
4390 struct worker_pool *pool;
4391 unsigned long flags;
4394 rcu_read_lock_sched();
4396 pr_info("Showing busy workqueues and worker pools:\n");
4398 list_for_each_entry_rcu(wq, &workqueues, list) {
4399 struct pool_workqueue *pwq;
4402 for_each_pwq(pwq, wq) {
4403 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4411 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4413 for_each_pwq(pwq, wq) {
4414 spin_lock_irqsave(&pwq->pool->lock, flags);
4415 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4417 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4421 for_each_pool(pool, pi) {
4422 struct worker *worker;
4425 spin_lock_irqsave(&pool->lock, flags);
4426 if (pool->nr_workers == pool->nr_idle)
4429 pr_info("pool %d:", pool->id);
4430 pr_cont_pool_info(pool);
4431 pr_cont(" hung=%us workers=%d",
4432 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4435 pr_cont(" manager: %d",
4436 task_pid_nr(pool->manager->task));
4437 list_for_each_entry(worker, &pool->idle_list, entry) {
4438 pr_cont(" %s%d", first ? "idle: " : "",
4439 task_pid_nr(worker->task));
4444 spin_unlock_irqrestore(&pool->lock, flags);
4447 rcu_read_unlock_sched();
4453 * There are two challenges in supporting CPU hotplug. Firstly, there
4454 * are a lot of assumptions on strong associations among work, pwq and
4455 * pool which make migrating pending and scheduled works very
4456 * difficult to implement without impacting hot paths. Secondly,
4457 * worker pools serve mix of short, long and very long running works making
4458 * blocked draining impractical.
4460 * This is solved by allowing the pools to be disassociated from the CPU
4461 * running as an unbound one and allowing it to be reattached later if the
4462 * cpu comes back online.
4465 static void wq_unbind_fn(struct work_struct *work)
4467 int cpu = smp_processor_id();
4468 struct worker_pool *pool;
4469 struct worker *worker;
4471 for_each_cpu_worker_pool(pool, cpu) {
4472 mutex_lock(&pool->attach_mutex);
4473 spin_lock_irq(&pool->lock);
4476 * We've blocked all attach/detach operations. Make all workers
4477 * unbound and set DISASSOCIATED. Before this, all workers
4478 * except for the ones which are still executing works from
4479 * before the last CPU down must be on the cpu. After
4480 * this, they may become diasporas.
4482 for_each_pool_worker(worker, pool)
4483 worker->flags |= WORKER_UNBOUND;
4485 pool->flags |= POOL_DISASSOCIATED;
4487 spin_unlock_irq(&pool->lock);
4488 mutex_unlock(&pool->attach_mutex);
4491 * Call schedule() so that we cross rq->lock and thus can
4492 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4493 * This is necessary as scheduler callbacks may be invoked
4499 * Sched callbacks are disabled now. Zap nr_running.
4500 * After this, nr_running stays zero and need_more_worker()
4501 * and keep_working() are always true as long as the
4502 * worklist is not empty. This pool now behaves as an
4503 * unbound (in terms of concurrency management) pool which
4504 * are served by workers tied to the pool.
4506 atomic_set(&pool->nr_running, 0);
4509 * With concurrency management just turned off, a busy
4510 * worker blocking could lead to lengthy stalls. Kick off
4511 * unbound chain execution of currently pending work items.
4513 spin_lock_irq(&pool->lock);
4514 wake_up_worker(pool);
4515 spin_unlock_irq(&pool->lock);
4520 * rebind_workers - rebind all workers of a pool to the associated CPU
4521 * @pool: pool of interest
4523 * @pool->cpu is coming online. Rebind all workers to the CPU.
4525 static void rebind_workers(struct worker_pool *pool)
4527 struct worker *worker;
4529 lockdep_assert_held(&pool->attach_mutex);
4532 * Restore CPU affinity of all workers. As all idle workers should
4533 * be on the run-queue of the associated CPU before any local
4534 * wake-ups for concurrency management happen, restore CPU affinity
4535 * of all workers first and then clear UNBOUND. As we're called
4536 * from CPU_ONLINE, the following shouldn't fail.
4538 for_each_pool_worker(worker, pool)
4539 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4540 pool->attrs->cpumask) < 0);
4542 spin_lock_irq(&pool->lock);
4545 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4546 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4547 * being reworked and this can go away in time.
4549 if (!(pool->flags & POOL_DISASSOCIATED)) {
4550 spin_unlock_irq(&pool->lock);
4554 pool->flags &= ~POOL_DISASSOCIATED;
4556 for_each_pool_worker(worker, pool) {
4557 unsigned int worker_flags = worker->flags;
4560 * A bound idle worker should actually be on the runqueue
4561 * of the associated CPU for local wake-ups targeting it to
4562 * work. Kick all idle workers so that they migrate to the
4563 * associated CPU. Doing this in the same loop as
4564 * replacing UNBOUND with REBOUND is safe as no worker will
4565 * be bound before @pool->lock is released.
4567 if (worker_flags & WORKER_IDLE)
4568 wake_up_process(worker->task);
4571 * We want to clear UNBOUND but can't directly call
4572 * worker_clr_flags() or adjust nr_running. Atomically
4573 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4574 * @worker will clear REBOUND using worker_clr_flags() when
4575 * it initiates the next execution cycle thus restoring
4576 * concurrency management. Note that when or whether
4577 * @worker clears REBOUND doesn't affect correctness.
4579 * ACCESS_ONCE() is necessary because @worker->flags may be
4580 * tested without holding any lock in
4581 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4582 * fail incorrectly leading to premature concurrency
4583 * management operations.
4585 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4586 worker_flags |= WORKER_REBOUND;
4587 worker_flags &= ~WORKER_UNBOUND;
4588 ACCESS_ONCE(worker->flags) = worker_flags;
4591 spin_unlock_irq(&pool->lock);
4595 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4596 * @pool: unbound pool of interest
4597 * @cpu: the CPU which is coming up
4599 * An unbound pool may end up with a cpumask which doesn't have any online
4600 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4601 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4602 * online CPU before, cpus_allowed of all its workers should be restored.
4604 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4606 static cpumask_t cpumask;
4607 struct worker *worker;
4609 lockdep_assert_held(&pool->attach_mutex);
4611 /* is @cpu allowed for @pool? */
4612 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4615 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4617 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4618 for_each_pool_worker(worker, pool)
4619 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4622 int workqueue_prepare_cpu(unsigned int cpu)
4624 struct worker_pool *pool;
4626 for_each_cpu_worker_pool(pool, cpu) {
4627 if (pool->nr_workers)
4629 if (!create_worker(pool))
4635 int workqueue_online_cpu(unsigned int cpu)
4637 struct worker_pool *pool;
4638 struct workqueue_struct *wq;
4641 mutex_lock(&wq_pool_mutex);
4643 for_each_pool(pool, pi) {
4644 mutex_lock(&pool->attach_mutex);
4646 if (pool->cpu == cpu)
4647 rebind_workers(pool);
4648 else if (pool->cpu < 0)
4649 restore_unbound_workers_cpumask(pool, cpu);
4651 mutex_unlock(&pool->attach_mutex);
4654 /* update NUMA affinity of unbound workqueues */
4655 list_for_each_entry(wq, &workqueues, list)
4656 wq_update_unbound_numa(wq, cpu, true);
4658 mutex_unlock(&wq_pool_mutex);
4662 int workqueue_offline_cpu(unsigned int cpu)
4664 struct work_struct unbind_work;
4665 struct workqueue_struct *wq;
4667 /* unbinding per-cpu workers should happen on the local CPU */
4668 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4669 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4671 /* update NUMA affinity of unbound workqueues */
4672 mutex_lock(&wq_pool_mutex);
4673 list_for_each_entry(wq, &workqueues, list)
4674 wq_update_unbound_numa(wq, cpu, false);
4675 mutex_unlock(&wq_pool_mutex);
4677 /* wait for per-cpu unbinding to finish */
4678 flush_work(&unbind_work);
4679 destroy_work_on_stack(&unbind_work);
4685 struct work_for_cpu {
4686 struct work_struct work;
4692 static void work_for_cpu_fn(struct work_struct *work)
4694 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4696 wfc->ret = wfc->fn(wfc->arg);
4700 * work_on_cpu - run a function in thread context on a particular cpu
4701 * @cpu: the cpu to run on
4702 * @fn: the function to run
4703 * @arg: the function arg
4705 * It is up to the caller to ensure that the cpu doesn't go offline.
4706 * The caller must not hold any locks which would prevent @fn from completing.
4708 * Return: The value @fn returns.
4710 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4712 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4714 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4715 schedule_work_on(cpu, &wfc.work);
4716 flush_work(&wfc.work);
4717 destroy_work_on_stack(&wfc.work);
4720 EXPORT_SYMBOL_GPL(work_on_cpu);
4721 #endif /* CONFIG_SMP */
4723 #ifdef CONFIG_FREEZER
4726 * freeze_workqueues_begin - begin freezing workqueues
4728 * Start freezing workqueues. After this function returns, all freezable
4729 * workqueues will queue new works to their delayed_works list instead of
4733 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4735 void freeze_workqueues_begin(void)
4737 struct workqueue_struct *wq;
4738 struct pool_workqueue *pwq;
4740 mutex_lock(&wq_pool_mutex);
4742 WARN_ON_ONCE(workqueue_freezing);
4743 workqueue_freezing = true;
4745 list_for_each_entry(wq, &workqueues, list) {
4746 mutex_lock(&wq->mutex);
4747 for_each_pwq(pwq, wq)
4748 pwq_adjust_max_active(pwq);
4749 mutex_unlock(&wq->mutex);
4752 mutex_unlock(&wq_pool_mutex);
4756 * freeze_workqueues_busy - are freezable workqueues still busy?
4758 * Check whether freezing is complete. This function must be called
4759 * between freeze_workqueues_begin() and thaw_workqueues().
4762 * Grabs and releases wq_pool_mutex.
4765 * %true if some freezable workqueues are still busy. %false if freezing
4768 bool freeze_workqueues_busy(void)
4771 struct workqueue_struct *wq;
4772 struct pool_workqueue *pwq;
4774 mutex_lock(&wq_pool_mutex);
4776 WARN_ON_ONCE(!workqueue_freezing);
4778 list_for_each_entry(wq, &workqueues, list) {
4779 if (!(wq->flags & WQ_FREEZABLE))
4782 * nr_active is monotonically decreasing. It's safe
4783 * to peek without lock.
4785 rcu_read_lock_sched();
4786 for_each_pwq(pwq, wq) {
4787 WARN_ON_ONCE(pwq->nr_active < 0);
4788 if (pwq->nr_active) {
4790 rcu_read_unlock_sched();
4794 rcu_read_unlock_sched();
4797 mutex_unlock(&wq_pool_mutex);
4802 * thaw_workqueues - thaw workqueues
4804 * Thaw workqueues. Normal queueing is restored and all collected
4805 * frozen works are transferred to their respective pool worklists.
4808 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4810 void thaw_workqueues(void)
4812 struct workqueue_struct *wq;
4813 struct pool_workqueue *pwq;
4815 mutex_lock(&wq_pool_mutex);
4817 if (!workqueue_freezing)
4820 workqueue_freezing = false;
4822 /* restore max_active and repopulate worklist */
4823 list_for_each_entry(wq, &workqueues, list) {
4824 mutex_lock(&wq->mutex);
4825 for_each_pwq(pwq, wq)
4826 pwq_adjust_max_active(pwq);
4827 mutex_unlock(&wq->mutex);
4831 mutex_unlock(&wq_pool_mutex);
4833 #endif /* CONFIG_FREEZER */
4835 static int workqueue_apply_unbound_cpumask(void)
4839 struct workqueue_struct *wq;
4840 struct apply_wqattrs_ctx *ctx, *n;
4842 lockdep_assert_held(&wq_pool_mutex);
4844 list_for_each_entry(wq, &workqueues, list) {
4845 if (!(wq->flags & WQ_UNBOUND))
4847 /* creating multiple pwqs breaks ordering guarantee */
4848 if (wq->flags & __WQ_ORDERED)
4851 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4857 list_add_tail(&ctx->list, &ctxs);
4860 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4862 apply_wqattrs_commit(ctx);
4863 apply_wqattrs_cleanup(ctx);
4870 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4871 * @cpumask: the cpumask to set
4873 * The low-level workqueues cpumask is a global cpumask that limits
4874 * the affinity of all unbound workqueues. This function check the @cpumask
4875 * and apply it to all unbound workqueues and updates all pwqs of them.
4877 * Retun: 0 - Success
4878 * -EINVAL - Invalid @cpumask
4879 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4881 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4884 cpumask_var_t saved_cpumask;
4886 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4889 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4890 if (!cpumask_empty(cpumask)) {
4891 apply_wqattrs_lock();
4893 /* save the old wq_unbound_cpumask. */
4894 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4896 /* update wq_unbound_cpumask at first and apply it to wqs. */
4897 cpumask_copy(wq_unbound_cpumask, cpumask);
4898 ret = workqueue_apply_unbound_cpumask();
4900 /* restore the wq_unbound_cpumask when failed. */
4902 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4904 apply_wqattrs_unlock();
4907 free_cpumask_var(saved_cpumask);
4913 * Workqueues with WQ_SYSFS flag set is visible to userland via
4914 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4915 * following attributes.
4917 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4918 * max_active RW int : maximum number of in-flight work items
4920 * Unbound workqueues have the following extra attributes.
4922 * id RO int : the associated pool ID
4923 * nice RW int : nice value of the workers
4924 * cpumask RW mask : bitmask of allowed CPUs for the workers
4927 struct workqueue_struct *wq;
4931 static struct workqueue_struct *dev_to_wq(struct device *dev)
4933 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4938 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4941 struct workqueue_struct *wq = dev_to_wq(dev);
4943 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4945 static DEVICE_ATTR_RO(per_cpu);
4947 static ssize_t max_active_show(struct device *dev,
4948 struct device_attribute *attr, char *buf)
4950 struct workqueue_struct *wq = dev_to_wq(dev);
4952 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4955 static ssize_t max_active_store(struct device *dev,
4956 struct device_attribute *attr, const char *buf,
4959 struct workqueue_struct *wq = dev_to_wq(dev);
4962 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4965 workqueue_set_max_active(wq, val);
4968 static DEVICE_ATTR_RW(max_active);
4970 static struct attribute *wq_sysfs_attrs[] = {
4971 &dev_attr_per_cpu.attr,
4972 &dev_attr_max_active.attr,
4975 ATTRIBUTE_GROUPS(wq_sysfs);
4977 static ssize_t wq_pool_ids_show(struct device *dev,
4978 struct device_attribute *attr, char *buf)
4980 struct workqueue_struct *wq = dev_to_wq(dev);
4981 const char *delim = "";
4982 int node, written = 0;
4984 rcu_read_lock_sched();
4985 for_each_node(node) {
4986 written += scnprintf(buf + written, PAGE_SIZE - written,
4987 "%s%d:%d", delim, node,
4988 unbound_pwq_by_node(wq, node)->pool->id);
4991 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4992 rcu_read_unlock_sched();
4997 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5000 struct workqueue_struct *wq = dev_to_wq(dev);
5003 mutex_lock(&wq->mutex);
5004 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5005 mutex_unlock(&wq->mutex);
5010 /* prepare workqueue_attrs for sysfs store operations */
5011 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5013 struct workqueue_attrs *attrs;
5015 lockdep_assert_held(&wq_pool_mutex);
5017 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5021 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5025 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5026 const char *buf, size_t count)
5028 struct workqueue_struct *wq = dev_to_wq(dev);
5029 struct workqueue_attrs *attrs;
5032 apply_wqattrs_lock();
5034 attrs = wq_sysfs_prep_attrs(wq);
5038 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5039 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5040 ret = apply_workqueue_attrs_locked(wq, attrs);
5045 apply_wqattrs_unlock();
5046 free_workqueue_attrs(attrs);
5047 return ret ?: count;
5050 static ssize_t wq_cpumask_show(struct device *dev,
5051 struct device_attribute *attr, char *buf)
5053 struct workqueue_struct *wq = dev_to_wq(dev);
5056 mutex_lock(&wq->mutex);
5057 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5058 cpumask_pr_args(wq->unbound_attrs->cpumask));
5059 mutex_unlock(&wq->mutex);
5063 static ssize_t wq_cpumask_store(struct device *dev,
5064 struct device_attribute *attr,
5065 const char *buf, size_t count)
5067 struct workqueue_struct *wq = dev_to_wq(dev);
5068 struct workqueue_attrs *attrs;
5071 apply_wqattrs_lock();
5073 attrs = wq_sysfs_prep_attrs(wq);
5077 ret = cpumask_parse(buf, attrs->cpumask);
5079 ret = apply_workqueue_attrs_locked(wq, attrs);
5082 apply_wqattrs_unlock();
5083 free_workqueue_attrs(attrs);
5084 return ret ?: count;
5087 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5090 struct workqueue_struct *wq = dev_to_wq(dev);
5093 mutex_lock(&wq->mutex);
5094 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5095 !wq->unbound_attrs->no_numa);
5096 mutex_unlock(&wq->mutex);
5101 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5102 const char *buf, size_t count)
5104 struct workqueue_struct *wq = dev_to_wq(dev);
5105 struct workqueue_attrs *attrs;
5106 int v, ret = -ENOMEM;
5108 apply_wqattrs_lock();
5110 attrs = wq_sysfs_prep_attrs(wq);
5115 if (sscanf(buf, "%d", &v) == 1) {
5116 attrs->no_numa = !v;
5117 ret = apply_workqueue_attrs_locked(wq, attrs);
5121 apply_wqattrs_unlock();
5122 free_workqueue_attrs(attrs);
5123 return ret ?: count;
5126 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5127 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5128 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5129 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5130 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5134 static struct bus_type wq_subsys = {
5135 .name = "workqueue",
5136 .dev_groups = wq_sysfs_groups,
5139 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5140 struct device_attribute *attr, char *buf)
5144 mutex_lock(&wq_pool_mutex);
5145 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5146 cpumask_pr_args(wq_unbound_cpumask));
5147 mutex_unlock(&wq_pool_mutex);
5152 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5153 struct device_attribute *attr, const char *buf, size_t count)
5155 cpumask_var_t cpumask;
5158 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5161 ret = cpumask_parse(buf, cpumask);
5163 ret = workqueue_set_unbound_cpumask(cpumask);
5165 free_cpumask_var(cpumask);
5166 return ret ? ret : count;
5169 static struct device_attribute wq_sysfs_cpumask_attr =
5170 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5171 wq_unbound_cpumask_store);
5173 static int __init wq_sysfs_init(void)
5177 err = subsys_virtual_register(&wq_subsys, NULL);
5181 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5183 core_initcall(wq_sysfs_init);
5185 static void wq_device_release(struct device *dev)
5187 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5193 * workqueue_sysfs_register - make a workqueue visible in sysfs
5194 * @wq: the workqueue to register
5196 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5197 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5198 * which is the preferred method.
5200 * Workqueue user should use this function directly iff it wants to apply
5201 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5202 * apply_workqueue_attrs() may race against userland updating the
5205 * Return: 0 on success, -errno on failure.
5207 int workqueue_sysfs_register(struct workqueue_struct *wq)
5209 struct wq_device *wq_dev;
5213 * Adjusting max_active or creating new pwqs by applying
5214 * attributes breaks ordering guarantee. Disallow exposing ordered
5217 if (WARN_ON(wq->flags & __WQ_ORDERED))
5220 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5225 wq_dev->dev.bus = &wq_subsys;
5226 wq_dev->dev.release = wq_device_release;
5227 dev_set_name(&wq_dev->dev, "%s", wq->name);
5230 * unbound_attrs are created separately. Suppress uevent until
5231 * everything is ready.
5233 dev_set_uevent_suppress(&wq_dev->dev, true);
5235 ret = device_register(&wq_dev->dev);
5242 if (wq->flags & WQ_UNBOUND) {
5243 struct device_attribute *attr;
5245 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5246 ret = device_create_file(&wq_dev->dev, attr);
5248 device_unregister(&wq_dev->dev);
5255 dev_set_uevent_suppress(&wq_dev->dev, false);
5256 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5261 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5262 * @wq: the workqueue to unregister
5264 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5266 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5268 struct wq_device *wq_dev = wq->wq_dev;
5274 device_unregister(&wq_dev->dev);
5276 #else /* CONFIG_SYSFS */
5277 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5278 #endif /* CONFIG_SYSFS */
5281 * Workqueue watchdog.
5283 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5284 * flush dependency, a concurrency managed work item which stays RUNNING
5285 * indefinitely. Workqueue stalls can be very difficult to debug as the
5286 * usual warning mechanisms don't trigger and internal workqueue state is
5289 * Workqueue watchdog monitors all worker pools periodically and dumps
5290 * state if some pools failed to make forward progress for a while where
5291 * forward progress is defined as the first item on ->worklist changing.
5293 * This mechanism is controlled through the kernel parameter
5294 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5295 * corresponding sysfs parameter file.
5297 #ifdef CONFIG_WQ_WATCHDOG
5299 static void wq_watchdog_timer_fn(unsigned long data);
5301 static unsigned long wq_watchdog_thresh = 30;
5302 static struct timer_list wq_watchdog_timer =
5303 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5305 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5306 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5308 static void wq_watchdog_reset_touched(void)
5312 wq_watchdog_touched = jiffies;
5313 for_each_possible_cpu(cpu)
5314 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5317 static void wq_watchdog_timer_fn(unsigned long data)
5319 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5320 bool lockup_detected = false;
5321 struct worker_pool *pool;
5329 for_each_pool(pool, pi) {
5330 unsigned long pool_ts, touched, ts;
5332 if (list_empty(&pool->worklist))
5335 /* get the latest of pool and touched timestamps */
5336 pool_ts = READ_ONCE(pool->watchdog_ts);
5337 touched = READ_ONCE(wq_watchdog_touched);
5339 if (time_after(pool_ts, touched))
5344 if (pool->cpu >= 0) {
5345 unsigned long cpu_touched =
5346 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5348 if (time_after(cpu_touched, ts))
5353 if (time_after(jiffies, ts + thresh)) {
5354 lockup_detected = true;
5355 pr_emerg("BUG: workqueue lockup - pool");
5356 pr_cont_pool_info(pool);
5357 pr_cont(" stuck for %us!\n",
5358 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5364 if (lockup_detected)
5365 show_workqueue_state();
5367 wq_watchdog_reset_touched();
5368 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5371 void wq_watchdog_touch(int cpu)
5374 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5376 wq_watchdog_touched = jiffies;
5379 static void wq_watchdog_set_thresh(unsigned long thresh)
5381 wq_watchdog_thresh = 0;
5382 del_timer_sync(&wq_watchdog_timer);
5385 wq_watchdog_thresh = thresh;
5386 wq_watchdog_reset_touched();
5387 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5391 static int wq_watchdog_param_set_thresh(const char *val,
5392 const struct kernel_param *kp)
5394 unsigned long thresh;
5397 ret = kstrtoul(val, 0, &thresh);
5402 wq_watchdog_set_thresh(thresh);
5404 wq_watchdog_thresh = thresh;
5409 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5410 .set = wq_watchdog_param_set_thresh,
5411 .get = param_get_ulong,
5414 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5417 static void wq_watchdog_init(void)
5419 wq_watchdog_set_thresh(wq_watchdog_thresh);
5422 #else /* CONFIG_WQ_WATCHDOG */
5424 static inline void wq_watchdog_init(void) { }
5426 #endif /* CONFIG_WQ_WATCHDOG */
5428 static void __init wq_numa_init(void)
5433 if (num_possible_nodes() <= 1)
5436 if (wq_disable_numa) {
5437 pr_info("workqueue: NUMA affinity support disabled\n");
5441 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5442 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5445 * We want masks of possible CPUs of each node which isn't readily
5446 * available. Build one from cpu_to_node() which should have been
5447 * fully initialized by now.
5449 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5453 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5454 node_online(node) ? node : NUMA_NO_NODE));
5456 for_each_possible_cpu(cpu) {
5457 node = cpu_to_node(cpu);
5458 if (WARN_ON(node == NUMA_NO_NODE)) {
5459 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5460 /* happens iff arch is bonkers, let's just proceed */
5463 cpumask_set_cpu(cpu, tbl[node]);
5466 wq_numa_possible_cpumask = tbl;
5467 wq_numa_enabled = true;
5470 static int __init init_workqueues(void)
5472 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5475 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5477 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5478 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5480 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5484 /* initialize CPU pools */
5485 for_each_possible_cpu(cpu) {
5486 struct worker_pool *pool;
5489 for_each_cpu_worker_pool(pool, cpu) {
5490 BUG_ON(init_worker_pool(pool));
5492 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5493 pool->attrs->nice = std_nice[i++];
5494 pool->node = cpu_to_node(cpu);
5497 mutex_lock(&wq_pool_mutex);
5498 BUG_ON(worker_pool_assign_id(pool));
5499 mutex_unlock(&wq_pool_mutex);
5503 /* create the initial worker */
5504 for_each_online_cpu(cpu) {
5505 struct worker_pool *pool;
5507 for_each_cpu_worker_pool(pool, cpu) {
5508 pool->flags &= ~POOL_DISASSOCIATED;
5509 BUG_ON(!create_worker(pool));
5513 /* create default unbound and ordered wq attrs */
5514 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5515 struct workqueue_attrs *attrs;
5517 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5518 attrs->nice = std_nice[i];
5519 unbound_std_wq_attrs[i] = attrs;
5522 * An ordered wq should have only one pwq as ordering is
5523 * guaranteed by max_active which is enforced by pwqs.
5524 * Turn off NUMA so that dfl_pwq is used for all nodes.
5526 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5527 attrs->nice = std_nice[i];
5528 attrs->no_numa = true;
5529 ordered_wq_attrs[i] = attrs;
5532 system_wq = alloc_workqueue("events", 0, 0);
5533 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5534 system_long_wq = alloc_workqueue("events_long", 0, 0);
5535 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5536 WQ_UNBOUND_MAX_ACTIVE);
5537 system_freezable_wq = alloc_workqueue("events_freezable",
5539 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5540 WQ_POWER_EFFICIENT, 0);
5541 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5542 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5544 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5545 !system_unbound_wq || !system_freezable_wq ||
5546 !system_power_efficient_wq ||
5547 !system_freezable_power_efficient_wq);
5553 early_initcall(init_workqueues);