2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
51 #include <linux/bug.h>
52 #include <linux/delay.h>
54 #include "workqueue_internal.h"
60 * A bound pool is either associated or disassociated with its CPU.
61 * While associated (!DISASSOCIATED), all workers are bound to the
62 * CPU and none has %WORKER_UNBOUND set and concurrency management
65 * While DISASSOCIATED, the cpu may be offline and all workers have
66 * %WORKER_UNBOUND set and concurrency management disabled, and may
67 * be executing on any CPU. The pool behaves as an unbound one.
69 * Note that DISASSOCIATED should be flipped only while holding
70 * attach_mutex to avoid changing binding state while
71 * worker_attach_to_pool() is in progress.
73 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
74 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
77 WORKER_DIE = 1 << 1, /* die die die */
78 WORKER_IDLE = 1 << 2, /* is idle */
79 WORKER_PREP = 1 << 3, /* preparing to run works */
80 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
81 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
82 WORKER_REBOUND = 1 << 8, /* worker was rebound */
84 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
85 WORKER_UNBOUND | WORKER_REBOUND,
87 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
89 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
90 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
92 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
93 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
95 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
96 /* call for help after 10ms
98 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
99 CREATE_COOLDOWN = HZ, /* time to breath after fail */
102 * Rescue workers are used only on emergencies and shared by
103 * all cpus. Give MIN_NICE.
105 RESCUER_NICE_LEVEL = MIN_NICE,
106 HIGHPRI_NICE_LEVEL = MIN_NICE,
112 * Structure fields follow one of the following exclusion rules.
114 * I: Modifiable by initialization/destruction paths and read-only for
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
120 * L: pool->lock protected. Access with pool->lock held.
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
127 * A: pool->attach_mutex protected.
129 * PL: wq_pool_mutex protected.
131 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
133 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
135 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
136 * sched-RCU for reads.
138 * WQ: wq->mutex protected.
140 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
142 * MD: wq_mayday_lock protected.
145 /* struct worker is defined in workqueue_internal.h */
148 spinlock_t lock; /* the pool lock */
149 int cpu; /* I: the associated cpu */
150 int node; /* I: the associated node ID */
151 int id; /* I: pool ID */
152 unsigned int flags; /* X: flags */
154 unsigned long watchdog_ts; /* L: watchdog timestamp */
156 struct list_head worklist; /* L: list of pending works */
157 int nr_workers; /* L: total number of workers */
159 /* nr_idle includes the ones off idle_list for rebinding */
160 int nr_idle; /* L: currently idle ones */
162 struct list_head idle_list; /* X: list of idle workers */
163 struct timer_list idle_timer; /* L: worker idle timeout */
164 struct timer_list mayday_timer; /* L: SOS timer for workers */
166 /* a workers is either on busy_hash or idle_list, or the manager */
167 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
168 /* L: hash of busy workers */
170 /* see manage_workers() for details on the two manager mutexes */
171 struct worker *manager; /* L: purely informational */
172 struct mutex attach_mutex; /* attach/detach exclusion */
173 struct list_head workers; /* A: attached workers */
174 struct completion *detach_completion; /* all workers detached */
176 struct ida worker_ida; /* worker IDs for task name */
178 struct workqueue_attrs *attrs; /* I: worker attributes */
179 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
180 int refcnt; /* PL: refcnt for unbound pools */
183 * The current concurrency level. As it's likely to be accessed
184 * from other CPUs during try_to_wake_up(), put it in a separate
187 atomic_t nr_running ____cacheline_aligned_in_smp;
190 * Destruction of pool is sched-RCU protected to allow dereferences
191 * from get_work_pool().
194 } ____cacheline_aligned_in_smp;
197 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
198 * of work_struct->data are used for flags and the remaining high bits
199 * point to the pwq; thus, pwqs need to be aligned at two's power of the
200 * number of flag bits.
202 struct pool_workqueue {
203 struct worker_pool *pool; /* I: the associated pool */
204 struct workqueue_struct *wq; /* I: the owning workqueue */
205 int work_color; /* L: current color */
206 int flush_color; /* L: flushing color */
207 int refcnt; /* L: reference count */
208 int nr_in_flight[WORK_NR_COLORS];
209 /* L: nr of in_flight works */
210 int nr_active; /* L: nr of active works */
211 int max_active; /* L: max active works */
212 struct list_head delayed_works; /* L: delayed works */
213 struct list_head pwqs_node; /* WR: node on wq->pwqs */
214 struct list_head mayday_node; /* MD: node on wq->maydays */
217 * Release of unbound pwq is punted to system_wq. See put_pwq()
218 * and pwq_unbound_release_workfn() for details. pool_workqueue
219 * itself is also sched-RCU protected so that the first pwq can be
220 * determined without grabbing wq->mutex.
222 struct work_struct unbound_release_work;
224 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
227 * Structure used to wait for workqueue flush.
230 struct list_head list; /* WQ: list of flushers */
231 int flush_color; /* WQ: flush color waiting for */
232 struct completion done; /* flush completion */
238 * The externally visible workqueue. It relays the issued work items to
239 * the appropriate worker_pool through its pool_workqueues.
241 struct workqueue_struct {
242 struct list_head pwqs; /* WR: all pwqs of this wq */
243 struct list_head list; /* PR: list of all workqueues */
245 struct mutex mutex; /* protects this wq */
246 int work_color; /* WQ: current work color */
247 int flush_color; /* WQ: current flush color */
248 atomic_t nr_pwqs_to_flush; /* flush in progress */
249 struct wq_flusher *first_flusher; /* WQ: first flusher */
250 struct list_head flusher_queue; /* WQ: flush waiters */
251 struct list_head flusher_overflow; /* WQ: flush overflow list */
253 struct list_head maydays; /* MD: pwqs requesting rescue */
254 struct worker *rescuer; /* I: rescue worker */
256 int nr_drainers; /* WQ: drain in progress */
257 int saved_max_active; /* WQ: saved pwq max_active */
259 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
260 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
263 struct wq_device *wq_dev; /* I: for sysfs interface */
265 #ifdef CONFIG_LOCKDEP
266 struct lockdep_map lockdep_map;
268 char name[WQ_NAME_LEN]; /* I: workqueue name */
271 * Destruction of workqueue_struct is sched-RCU protected to allow
272 * walking the workqueues list without grabbing wq_pool_mutex.
273 * This is used to dump all workqueues from sysrq.
277 /* hot fields used during command issue, aligned to cacheline */
278 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
279 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
280 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
283 static struct kmem_cache *pwq_cache;
285 static cpumask_var_t *wq_numa_possible_cpumask;
286 /* possible CPUs of each node */
288 static bool wq_disable_numa;
289 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
291 /* see the comment above the definition of WQ_POWER_EFFICIENT */
292 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
293 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
295 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
297 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
298 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
300 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
301 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
302 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
304 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
305 static bool workqueue_freezing; /* PL: have wqs started freezing? */
307 static cpumask_var_t wq_unbound_cpumask; /* PL: low level cpumask for all unbound wqs */
309 /* the per-cpu worker pools */
310 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
313 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
315 /* PL: hash of all unbound pools keyed by pool->attrs */
316 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
318 /* I: attributes used when instantiating standard unbound pools on demand */
319 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
321 /* I: attributes used when instantiating ordered pools on demand */
322 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
324 struct workqueue_struct *system_wq __read_mostly;
325 EXPORT_SYMBOL(system_wq);
326 struct workqueue_struct *system_highpri_wq __read_mostly;
327 EXPORT_SYMBOL_GPL(system_highpri_wq);
328 struct workqueue_struct *system_long_wq __read_mostly;
329 EXPORT_SYMBOL_GPL(system_long_wq);
330 struct workqueue_struct *system_unbound_wq __read_mostly;
331 EXPORT_SYMBOL_GPL(system_unbound_wq);
332 struct workqueue_struct *system_freezable_wq __read_mostly;
333 EXPORT_SYMBOL_GPL(system_freezable_wq);
334 struct workqueue_struct *system_power_efficient_wq __read_mostly;
335 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
336 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
337 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
339 static int worker_thread(void *__worker);
340 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
342 #define CREATE_TRACE_POINTS
343 #include <trace/events/workqueue.h>
345 #define assert_rcu_or_pool_mutex() \
346 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
347 !lockdep_is_held(&wq_pool_mutex), \
348 "sched RCU or wq_pool_mutex should be held")
350 #define assert_rcu_or_wq_mutex(wq) \
351 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
352 !lockdep_is_held(&wq->mutex), \
353 "sched RCU or wq->mutex should be held")
355 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
356 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
357 !lockdep_is_held(&wq->mutex) && \
358 !lockdep_is_held(&wq_pool_mutex), \
359 "sched RCU, wq->mutex or wq_pool_mutex should be held")
361 #define for_each_cpu_worker_pool(pool, cpu) \
362 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
363 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
367 * for_each_pool - iterate through all worker_pools in the system
368 * @pool: iteration cursor
369 * @pi: integer used for iteration
371 * This must be called either with wq_pool_mutex held or sched RCU read
372 * locked. If the pool needs to be used beyond the locking in effect, the
373 * caller is responsible for guaranteeing that the pool stays online.
375 * The if/else clause exists only for the lockdep assertion and can be
378 #define for_each_pool(pool, pi) \
379 idr_for_each_entry(&worker_pool_idr, pool, pi) \
380 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
384 * for_each_pool_worker - iterate through all workers of a worker_pool
385 * @worker: iteration cursor
386 * @pool: worker_pool to iterate workers of
388 * This must be called with @pool->attach_mutex.
390 * The if/else clause exists only for the lockdep assertion and can be
393 #define for_each_pool_worker(worker, pool) \
394 list_for_each_entry((worker), &(pool)->workers, node) \
395 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
399 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
400 * @pwq: iteration cursor
401 * @wq: the target workqueue
403 * This must be called either with wq->mutex held or sched RCU read locked.
404 * If the pwq needs to be used beyond the locking in effect, the caller is
405 * responsible for guaranteeing that the pwq stays online.
407 * The if/else clause exists only for the lockdep assertion and can be
410 #define for_each_pwq(pwq, wq) \
411 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
412 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
415 #ifdef CONFIG_DEBUG_OBJECTS_WORK
417 static struct debug_obj_descr work_debug_descr;
419 static void *work_debug_hint(void *addr)
421 return ((struct work_struct *) addr)->func;
425 * fixup_init is called when:
426 * - an active object is initialized
428 static int work_fixup_init(void *addr, enum debug_obj_state state)
430 struct work_struct *work = addr;
433 case ODEBUG_STATE_ACTIVE:
434 cancel_work_sync(work);
435 debug_object_init(work, &work_debug_descr);
443 * fixup_activate is called when:
444 * - an active object is activated
445 * - an unknown object is activated (might be a statically initialized object)
447 static int work_fixup_activate(void *addr, enum debug_obj_state state)
449 struct work_struct *work = addr;
453 case ODEBUG_STATE_NOTAVAILABLE:
455 * This is not really a fixup. The work struct was
456 * statically initialized. We just make sure that it
457 * is tracked in the object tracker.
459 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
460 debug_object_init(work, &work_debug_descr);
461 debug_object_activate(work, &work_debug_descr);
467 case ODEBUG_STATE_ACTIVE:
476 * fixup_free is called when:
477 * - an active object is freed
479 static int work_fixup_free(void *addr, enum debug_obj_state state)
481 struct work_struct *work = addr;
484 case ODEBUG_STATE_ACTIVE:
485 cancel_work_sync(work);
486 debug_object_free(work, &work_debug_descr);
493 static struct debug_obj_descr work_debug_descr = {
494 .name = "work_struct",
495 .debug_hint = work_debug_hint,
496 .fixup_init = work_fixup_init,
497 .fixup_activate = work_fixup_activate,
498 .fixup_free = work_fixup_free,
501 static inline void debug_work_activate(struct work_struct *work)
503 debug_object_activate(work, &work_debug_descr);
506 static inline void debug_work_deactivate(struct work_struct *work)
508 debug_object_deactivate(work, &work_debug_descr);
511 void __init_work(struct work_struct *work, int onstack)
514 debug_object_init_on_stack(work, &work_debug_descr);
516 debug_object_init(work, &work_debug_descr);
518 EXPORT_SYMBOL_GPL(__init_work);
520 void destroy_work_on_stack(struct work_struct *work)
522 debug_object_free(work, &work_debug_descr);
524 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
526 void destroy_delayed_work_on_stack(struct delayed_work *work)
528 destroy_timer_on_stack(&work->timer);
529 debug_object_free(&work->work, &work_debug_descr);
531 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
534 static inline void debug_work_activate(struct work_struct *work) { }
535 static inline void debug_work_deactivate(struct work_struct *work) { }
539 * worker_pool_assign_id - allocate ID and assing it to @pool
540 * @pool: the pool pointer of interest
542 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
543 * successfully, -errno on failure.
545 static int worker_pool_assign_id(struct worker_pool *pool)
549 lockdep_assert_held(&wq_pool_mutex);
551 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
561 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
562 * @wq: the target workqueue
565 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
567 * If the pwq needs to be used beyond the locking in effect, the caller is
568 * responsible for guaranteeing that the pwq stays online.
570 * Return: The unbound pool_workqueue for @node.
572 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
575 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
578 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
579 * delayed item is pending. The plan is to keep CPU -> NODE
580 * mapping valid and stable across CPU on/offlines. Once that
581 * happens, this workaround can be removed.
583 if (unlikely(node == NUMA_NO_NODE))
586 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
589 static unsigned int work_color_to_flags(int color)
591 return color << WORK_STRUCT_COLOR_SHIFT;
594 static int get_work_color(struct work_struct *work)
596 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
597 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
600 static int work_next_color(int color)
602 return (color + 1) % WORK_NR_COLORS;
606 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
607 * contain the pointer to the queued pwq. Once execution starts, the flag
608 * is cleared and the high bits contain OFFQ flags and pool ID.
610 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
611 * and clear_work_data() can be used to set the pwq, pool or clear
612 * work->data. These functions should only be called while the work is
613 * owned - ie. while the PENDING bit is set.
615 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
616 * corresponding to a work. Pool is available once the work has been
617 * queued anywhere after initialization until it is sync canceled. pwq is
618 * available only while the work item is queued.
620 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
621 * canceled. While being canceled, a work item may have its PENDING set
622 * but stay off timer and worklist for arbitrarily long and nobody should
623 * try to steal the PENDING bit.
625 static inline void set_work_data(struct work_struct *work, unsigned long data,
628 WARN_ON_ONCE(!work_pending(work));
629 atomic_long_set(&work->data, data | flags | work_static(work));
632 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
633 unsigned long extra_flags)
635 set_work_data(work, (unsigned long)pwq,
636 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
639 static void set_work_pool_and_keep_pending(struct work_struct *work,
642 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
643 WORK_STRUCT_PENDING);
646 static void set_work_pool_and_clear_pending(struct work_struct *work,
650 * The following wmb is paired with the implied mb in
651 * test_and_set_bit(PENDING) and ensures all updates to @work made
652 * here are visible to and precede any updates by the next PENDING
656 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
658 * The following mb guarantees that previous clear of a PENDING bit
659 * will not be reordered with any speculative LOADS or STORES from
660 * work->current_func, which is executed afterwards. This possible
661 * reordering can lead to a missed execution on attempt to qeueue
662 * the same @work. E.g. consider this case:
665 * ---------------------------- --------------------------------
667 * 1 STORE event_indicated
668 * 2 queue_work_on() {
669 * 3 test_and_set_bit(PENDING)
670 * 4 } set_..._and_clear_pending() {
671 * 5 set_work_data() # clear bit
673 * 7 work->current_func() {
674 * 8 LOAD event_indicated
677 * Without an explicit full barrier speculative LOAD on line 8 can
678 * be executed before CPU#0 does STORE on line 1. If that happens,
679 * CPU#0 observes the PENDING bit is still set and new execution of
680 * a @work is not queued in a hope, that CPU#1 will eventually
681 * finish the queued @work. Meanwhile CPU#1 does not see
682 * event_indicated is set, because speculative LOAD was executed
683 * before actual STORE.
688 static void clear_work_data(struct work_struct *work)
690 smp_wmb(); /* see set_work_pool_and_clear_pending() */
691 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
694 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
696 unsigned long data = atomic_long_read(&work->data);
698 if (data & WORK_STRUCT_PWQ)
699 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
705 * get_work_pool - return the worker_pool a given work was associated with
706 * @work: the work item of interest
708 * Pools are created and destroyed under wq_pool_mutex, and allows read
709 * access under sched-RCU read lock. As such, this function should be
710 * called under wq_pool_mutex or with preemption disabled.
712 * All fields of the returned pool are accessible as long as the above
713 * mentioned locking is in effect. If the returned pool needs to be used
714 * beyond the critical section, the caller is responsible for ensuring the
715 * returned pool is and stays online.
717 * Return: The worker_pool @work was last associated with. %NULL if none.
719 static struct worker_pool *get_work_pool(struct work_struct *work)
721 unsigned long data = atomic_long_read(&work->data);
724 assert_rcu_or_pool_mutex();
726 if (data & WORK_STRUCT_PWQ)
727 return ((struct pool_workqueue *)
728 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
730 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
731 if (pool_id == WORK_OFFQ_POOL_NONE)
734 return idr_find(&worker_pool_idr, pool_id);
738 * get_work_pool_id - return the worker pool ID a given work is associated with
739 * @work: the work item of interest
741 * Return: The worker_pool ID @work was last associated with.
742 * %WORK_OFFQ_POOL_NONE if none.
744 static int get_work_pool_id(struct work_struct *work)
746 unsigned long data = atomic_long_read(&work->data);
748 if (data & WORK_STRUCT_PWQ)
749 return ((struct pool_workqueue *)
750 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
752 return data >> WORK_OFFQ_POOL_SHIFT;
755 static void mark_work_canceling(struct work_struct *work)
757 unsigned long pool_id = get_work_pool_id(work);
759 pool_id <<= WORK_OFFQ_POOL_SHIFT;
760 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
763 static bool work_is_canceling(struct work_struct *work)
765 unsigned long data = atomic_long_read(&work->data);
767 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
771 * Policy functions. These define the policies on how the global worker
772 * pools are managed. Unless noted otherwise, these functions assume that
773 * they're being called with pool->lock held.
776 static bool __need_more_worker(struct worker_pool *pool)
778 return !atomic_read(&pool->nr_running);
782 * Need to wake up a worker? Called from anything but currently
785 * Note that, because unbound workers never contribute to nr_running, this
786 * function will always return %true for unbound pools as long as the
787 * worklist isn't empty.
789 static bool need_more_worker(struct worker_pool *pool)
791 return !list_empty(&pool->worklist) && __need_more_worker(pool);
794 /* Can I start working? Called from busy but !running workers. */
795 static bool may_start_working(struct worker_pool *pool)
797 return pool->nr_idle;
800 /* Do I need to keep working? Called from currently running workers. */
801 static bool keep_working(struct worker_pool *pool)
803 return !list_empty(&pool->worklist) &&
804 atomic_read(&pool->nr_running) <= 1;
807 /* Do we need a new worker? Called from manager. */
808 static bool need_to_create_worker(struct worker_pool *pool)
810 return need_more_worker(pool) && !may_start_working(pool);
813 /* Do we have too many workers and should some go away? */
814 static bool too_many_workers(struct worker_pool *pool)
816 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
817 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
818 int nr_busy = pool->nr_workers - nr_idle;
820 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
827 /* Return the first idle worker. Safe with preemption disabled */
828 static struct worker *first_idle_worker(struct worker_pool *pool)
830 if (unlikely(list_empty(&pool->idle_list)))
833 return list_first_entry(&pool->idle_list, struct worker, entry);
837 * wake_up_worker - wake up an idle worker
838 * @pool: worker pool to wake worker from
840 * Wake up the first idle worker of @pool.
843 * spin_lock_irq(pool->lock).
845 static void wake_up_worker(struct worker_pool *pool)
847 struct worker *worker = first_idle_worker(pool);
850 wake_up_process(worker->task);
854 * wq_worker_waking_up - a worker is waking up
855 * @task: task waking up
856 * @cpu: CPU @task is waking up to
858 * This function is called during try_to_wake_up() when a worker is
862 * spin_lock_irq(rq->lock)
864 void wq_worker_waking_up(struct task_struct *task, int cpu)
866 struct worker *worker = kthread_data(task);
868 if (!(worker->flags & WORKER_NOT_RUNNING)) {
869 WARN_ON_ONCE(worker->pool->cpu != cpu);
870 atomic_inc(&worker->pool->nr_running);
875 * wq_worker_sleeping - a worker is going to sleep
876 * @task: task going to sleep
877 * @cpu: CPU in question, must be the current CPU number
879 * This function is called during schedule() when a busy worker is
880 * going to sleep. Worker on the same cpu can be woken up by
881 * returning pointer to its task.
884 * spin_lock_irq(rq->lock)
887 * Worker task on @cpu to wake up, %NULL if none.
889 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
891 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
892 struct worker_pool *pool;
895 * Rescuers, which may not have all the fields set up like normal
896 * workers, also reach here, let's not access anything before
897 * checking NOT_RUNNING.
899 if (worker->flags & WORKER_NOT_RUNNING)
904 /* this can only happen on the local cpu */
905 if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu))
909 * The counterpart of the following dec_and_test, implied mb,
910 * worklist not empty test sequence is in insert_work().
911 * Please read comment there.
913 * NOT_RUNNING is clear. This means that we're bound to and
914 * running on the local cpu w/ rq lock held and preemption
915 * disabled, which in turn means that none else could be
916 * manipulating idle_list, so dereferencing idle_list without pool
919 if (atomic_dec_and_test(&pool->nr_running) &&
920 !list_empty(&pool->worklist))
921 to_wakeup = first_idle_worker(pool);
922 return to_wakeup ? to_wakeup->task : NULL;
926 * worker_set_flags - set worker flags and adjust nr_running accordingly
928 * @flags: flags to set
930 * Set @flags in @worker->flags and adjust nr_running accordingly.
933 * spin_lock_irq(pool->lock)
935 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
937 struct worker_pool *pool = worker->pool;
939 WARN_ON_ONCE(worker->task != current);
941 /* If transitioning into NOT_RUNNING, adjust nr_running. */
942 if ((flags & WORKER_NOT_RUNNING) &&
943 !(worker->flags & WORKER_NOT_RUNNING)) {
944 atomic_dec(&pool->nr_running);
947 worker->flags |= flags;
951 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
953 * @flags: flags to clear
955 * Clear @flags in @worker->flags and adjust nr_running accordingly.
958 * spin_lock_irq(pool->lock)
960 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
962 struct worker_pool *pool = worker->pool;
963 unsigned int oflags = worker->flags;
965 WARN_ON_ONCE(worker->task != current);
967 worker->flags &= ~flags;
970 * If transitioning out of NOT_RUNNING, increment nr_running. Note
971 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
972 * of multiple flags, not a single flag.
974 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
975 if (!(worker->flags & WORKER_NOT_RUNNING))
976 atomic_inc(&pool->nr_running);
980 * find_worker_executing_work - find worker which is executing a work
981 * @pool: pool of interest
982 * @work: work to find worker for
984 * Find a worker which is executing @work on @pool by searching
985 * @pool->busy_hash which is keyed by the address of @work. For a worker
986 * to match, its current execution should match the address of @work and
987 * its work function. This is to avoid unwanted dependency between
988 * unrelated work executions through a work item being recycled while still
991 * This is a bit tricky. A work item may be freed once its execution
992 * starts and nothing prevents the freed area from being recycled for
993 * another work item. If the same work item address ends up being reused
994 * before the original execution finishes, workqueue will identify the
995 * recycled work item as currently executing and make it wait until the
996 * current execution finishes, introducing an unwanted dependency.
998 * This function checks the work item address and work function to avoid
999 * false positives. Note that this isn't complete as one may construct a
1000 * work function which can introduce dependency onto itself through a
1001 * recycled work item. Well, if somebody wants to shoot oneself in the
1002 * foot that badly, there's only so much we can do, and if such deadlock
1003 * actually occurs, it should be easy to locate the culprit work function.
1006 * spin_lock_irq(pool->lock).
1009 * Pointer to worker which is executing @work if found, %NULL
1012 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1013 struct work_struct *work)
1015 struct worker *worker;
1017 hash_for_each_possible(pool->busy_hash, worker, hentry,
1018 (unsigned long)work)
1019 if (worker->current_work == work &&
1020 worker->current_func == work->func)
1027 * move_linked_works - move linked works to a list
1028 * @work: start of series of works to be scheduled
1029 * @head: target list to append @work to
1030 * @nextp: out parameter for nested worklist walking
1032 * Schedule linked works starting from @work to @head. Work series to
1033 * be scheduled starts at @work and includes any consecutive work with
1034 * WORK_STRUCT_LINKED set in its predecessor.
1036 * If @nextp is not NULL, it's updated to point to the next work of
1037 * the last scheduled work. This allows move_linked_works() to be
1038 * nested inside outer list_for_each_entry_safe().
1041 * spin_lock_irq(pool->lock).
1043 static void move_linked_works(struct work_struct *work, struct list_head *head,
1044 struct work_struct **nextp)
1046 struct work_struct *n;
1049 * Linked worklist will always end before the end of the list,
1050 * use NULL for list head.
1052 list_for_each_entry_safe_from(work, n, NULL, entry) {
1053 list_move_tail(&work->entry, head);
1054 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1059 * If we're already inside safe list traversal and have moved
1060 * multiple works to the scheduled queue, the next position
1061 * needs to be updated.
1068 * get_pwq - get an extra reference on the specified pool_workqueue
1069 * @pwq: pool_workqueue to get
1071 * Obtain an extra reference on @pwq. The caller should guarantee that
1072 * @pwq has positive refcnt and be holding the matching pool->lock.
1074 static void get_pwq(struct pool_workqueue *pwq)
1076 lockdep_assert_held(&pwq->pool->lock);
1077 WARN_ON_ONCE(pwq->refcnt <= 0);
1082 * put_pwq - put a pool_workqueue reference
1083 * @pwq: pool_workqueue to put
1085 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1086 * destruction. The caller should be holding the matching pool->lock.
1088 static void put_pwq(struct pool_workqueue *pwq)
1090 lockdep_assert_held(&pwq->pool->lock);
1091 if (likely(--pwq->refcnt))
1093 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1096 * @pwq can't be released under pool->lock, bounce to
1097 * pwq_unbound_release_workfn(). This never recurses on the same
1098 * pool->lock as this path is taken only for unbound workqueues and
1099 * the release work item is scheduled on a per-cpu workqueue. To
1100 * avoid lockdep warning, unbound pool->locks are given lockdep
1101 * subclass of 1 in get_unbound_pool().
1103 schedule_work(&pwq->unbound_release_work);
1107 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1108 * @pwq: pool_workqueue to put (can be %NULL)
1110 * put_pwq() with locking. This function also allows %NULL @pwq.
1112 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1116 * As both pwqs and pools are sched-RCU protected, the
1117 * following lock operations are safe.
1119 spin_lock_irq(&pwq->pool->lock);
1121 spin_unlock_irq(&pwq->pool->lock);
1125 static void pwq_activate_delayed_work(struct work_struct *work)
1127 struct pool_workqueue *pwq = get_work_pwq(work);
1129 trace_workqueue_activate_work(work);
1130 if (list_empty(&pwq->pool->worklist))
1131 pwq->pool->watchdog_ts = jiffies;
1132 move_linked_works(work, &pwq->pool->worklist, NULL);
1133 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1137 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1139 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1140 struct work_struct, entry);
1142 pwq_activate_delayed_work(work);
1146 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1147 * @pwq: pwq of interest
1148 * @color: color of work which left the queue
1150 * A work either has completed or is removed from pending queue,
1151 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1154 * spin_lock_irq(pool->lock).
1156 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1158 /* uncolored work items don't participate in flushing or nr_active */
1159 if (color == WORK_NO_COLOR)
1162 pwq->nr_in_flight[color]--;
1165 if (!list_empty(&pwq->delayed_works)) {
1166 /* one down, submit a delayed one */
1167 if (pwq->nr_active < pwq->max_active)
1168 pwq_activate_first_delayed(pwq);
1171 /* is flush in progress and are we at the flushing tip? */
1172 if (likely(pwq->flush_color != color))
1175 /* are there still in-flight works? */
1176 if (pwq->nr_in_flight[color])
1179 /* this pwq is done, clear flush_color */
1180 pwq->flush_color = -1;
1183 * If this was the last pwq, wake up the first flusher. It
1184 * will handle the rest.
1186 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1187 complete(&pwq->wq->first_flusher->done);
1193 * try_to_grab_pending - steal work item from worklist and disable irq
1194 * @work: work item to steal
1195 * @is_dwork: @work is a delayed_work
1196 * @flags: place to store irq state
1198 * Try to grab PENDING bit of @work. This function can handle @work in any
1199 * stable state - idle, on timer or on worklist.
1202 * 1 if @work was pending and we successfully stole PENDING
1203 * 0 if @work was idle and we claimed PENDING
1204 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1205 * -ENOENT if someone else is canceling @work, this state may persist
1206 * for arbitrarily long
1209 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1210 * interrupted while holding PENDING and @work off queue, irq must be
1211 * disabled on entry. This, combined with delayed_work->timer being
1212 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1214 * On successful return, >= 0, irq is disabled and the caller is
1215 * responsible for releasing it using local_irq_restore(*@flags).
1217 * This function is safe to call from any context including IRQ handler.
1219 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1220 unsigned long *flags)
1222 struct worker_pool *pool;
1223 struct pool_workqueue *pwq;
1225 local_irq_save(*flags);
1227 /* try to steal the timer if it exists */
1229 struct delayed_work *dwork = to_delayed_work(work);
1232 * dwork->timer is irqsafe. If del_timer() fails, it's
1233 * guaranteed that the timer is not queued anywhere and not
1234 * running on the local CPU.
1236 if (likely(del_timer(&dwork->timer)))
1240 /* try to claim PENDING the normal way */
1241 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1245 * The queueing is in progress, or it is already queued. Try to
1246 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1248 pool = get_work_pool(work);
1252 spin_lock(&pool->lock);
1254 * work->data is guaranteed to point to pwq only while the work
1255 * item is queued on pwq->wq, and both updating work->data to point
1256 * to pwq on queueing and to pool on dequeueing are done under
1257 * pwq->pool->lock. This in turn guarantees that, if work->data
1258 * points to pwq which is associated with a locked pool, the work
1259 * item is currently queued on that pool.
1261 pwq = get_work_pwq(work);
1262 if (pwq && pwq->pool == pool) {
1263 debug_work_deactivate(work);
1266 * A delayed work item cannot be grabbed directly because
1267 * it might have linked NO_COLOR work items which, if left
1268 * on the delayed_list, will confuse pwq->nr_active
1269 * management later on and cause stall. Make sure the work
1270 * item is activated before grabbing.
1272 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1273 pwq_activate_delayed_work(work);
1275 list_del_init(&work->entry);
1276 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1278 /* work->data points to pwq iff queued, point to pool */
1279 set_work_pool_and_keep_pending(work, pool->id);
1281 spin_unlock(&pool->lock);
1284 spin_unlock(&pool->lock);
1286 local_irq_restore(*flags);
1287 if (work_is_canceling(work))
1291 * The queueing is in progress in another context. If we keep
1292 * taking the pool->lock in a busy loop, the other context may
1293 * never get the lock. Give 1 usec delay to avoid this contention.
1300 * insert_work - insert a work into a pool
1301 * @pwq: pwq @work belongs to
1302 * @work: work to insert
1303 * @head: insertion point
1304 * @extra_flags: extra WORK_STRUCT_* flags to set
1306 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1307 * work_struct flags.
1310 * spin_lock_irq(pool->lock).
1312 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1313 struct list_head *head, unsigned int extra_flags)
1315 struct worker_pool *pool = pwq->pool;
1317 /* we own @work, set data and link */
1318 set_work_pwq(work, pwq, extra_flags);
1319 list_add_tail(&work->entry, head);
1323 * Ensure either wq_worker_sleeping() sees the above
1324 * list_add_tail() or we see zero nr_running to avoid workers lying
1325 * around lazily while there are works to be processed.
1329 if (__need_more_worker(pool))
1330 wake_up_worker(pool);
1334 * Test whether @work is being queued from another work executing on the
1337 static bool is_chained_work(struct workqueue_struct *wq)
1339 struct worker *worker;
1341 worker = current_wq_worker();
1343 * Return %true iff I'm a worker execuing a work item on @wq. If
1344 * I'm @worker, it's safe to dereference it without locking.
1346 return worker && worker->current_pwq->wq == wq;
1349 static void __queue_work(int cpu, struct workqueue_struct *wq,
1350 struct work_struct *work)
1352 struct pool_workqueue *pwq;
1353 struct worker_pool *last_pool;
1354 struct list_head *worklist;
1355 unsigned int work_flags;
1356 unsigned int req_cpu = cpu;
1359 * While a work item is PENDING && off queue, a task trying to
1360 * steal the PENDING will busy-loop waiting for it to either get
1361 * queued or lose PENDING. Grabbing PENDING and queueing should
1362 * happen with IRQ disabled.
1364 WARN_ON_ONCE(!irqs_disabled());
1366 debug_work_activate(work);
1368 /* if draining, only works from the same workqueue are allowed */
1369 if (unlikely(wq->flags & __WQ_DRAINING) &&
1370 WARN_ON_ONCE(!is_chained_work(wq)))
1373 if (req_cpu == WORK_CPU_UNBOUND)
1374 cpu = raw_smp_processor_id();
1376 /* pwq which will be used unless @work is executing elsewhere */
1377 if (!(wq->flags & WQ_UNBOUND))
1378 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1380 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1383 * If @work was previously on a different pool, it might still be
1384 * running there, in which case the work needs to be queued on that
1385 * pool to guarantee non-reentrancy.
1387 last_pool = get_work_pool(work);
1388 if (last_pool && last_pool != pwq->pool) {
1389 struct worker *worker;
1391 spin_lock(&last_pool->lock);
1393 worker = find_worker_executing_work(last_pool, work);
1395 if (worker && worker->current_pwq->wq == wq) {
1396 pwq = worker->current_pwq;
1398 /* meh... not running there, queue here */
1399 spin_unlock(&last_pool->lock);
1400 spin_lock(&pwq->pool->lock);
1403 spin_lock(&pwq->pool->lock);
1407 * pwq is determined and locked. For unbound pools, we could have
1408 * raced with pwq release and it could already be dead. If its
1409 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1410 * without another pwq replacing it in the numa_pwq_tbl or while
1411 * work items are executing on it, so the retrying is guaranteed to
1412 * make forward-progress.
1414 if (unlikely(!pwq->refcnt)) {
1415 if (wq->flags & WQ_UNBOUND) {
1416 spin_unlock(&pwq->pool->lock);
1421 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1425 /* pwq determined, queue */
1426 trace_workqueue_queue_work(req_cpu, pwq, work);
1428 if (WARN_ON(!list_empty(&work->entry))) {
1429 spin_unlock(&pwq->pool->lock);
1433 pwq->nr_in_flight[pwq->work_color]++;
1434 work_flags = work_color_to_flags(pwq->work_color);
1436 if (likely(pwq->nr_active < pwq->max_active)) {
1437 trace_workqueue_activate_work(work);
1439 worklist = &pwq->pool->worklist;
1440 if (list_empty(worklist))
1441 pwq->pool->watchdog_ts = jiffies;
1443 work_flags |= WORK_STRUCT_DELAYED;
1444 worklist = &pwq->delayed_works;
1447 insert_work(pwq, work, worklist, work_flags);
1449 spin_unlock(&pwq->pool->lock);
1453 * queue_work_on - queue work on specific cpu
1454 * @cpu: CPU number to execute work on
1455 * @wq: workqueue to use
1456 * @work: work to queue
1458 * We queue the work to a specific CPU, the caller must ensure it
1461 * Return: %false if @work was already on a queue, %true otherwise.
1463 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1464 struct work_struct *work)
1467 unsigned long flags;
1469 local_irq_save(flags);
1471 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1472 __queue_work(cpu, wq, work);
1476 local_irq_restore(flags);
1479 EXPORT_SYMBOL(queue_work_on);
1481 void delayed_work_timer_fn(unsigned long __data)
1483 struct delayed_work *dwork = (struct delayed_work *)__data;
1485 /* should have been called from irqsafe timer with irq already off */
1486 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1488 EXPORT_SYMBOL(delayed_work_timer_fn);
1490 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1491 struct delayed_work *dwork, unsigned long delay)
1493 struct timer_list *timer = &dwork->timer;
1494 struct work_struct *work = &dwork->work;
1497 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1498 timer->data != (unsigned long)dwork);
1499 WARN_ON_ONCE(timer_pending(timer));
1500 WARN_ON_ONCE(!list_empty(&work->entry));
1503 * If @delay is 0, queue @dwork->work immediately. This is for
1504 * both optimization and correctness. The earliest @timer can
1505 * expire is on the closest next tick and delayed_work users depend
1506 * on that there's no such delay when @delay is 0.
1509 __queue_work(cpu, wq, &dwork->work);
1515 timer->expires = jiffies + delay;
1517 if (unlikely(cpu != WORK_CPU_UNBOUND))
1518 add_timer_on(timer, cpu);
1524 * queue_delayed_work_on - queue work on specific CPU after delay
1525 * @cpu: CPU number to execute work on
1526 * @wq: workqueue to use
1527 * @dwork: work to queue
1528 * @delay: number of jiffies to wait before queueing
1530 * Return: %false if @work was already on a queue, %true otherwise. If
1531 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1534 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1535 struct delayed_work *dwork, unsigned long delay)
1537 struct work_struct *work = &dwork->work;
1539 unsigned long flags;
1541 /* read the comment in __queue_work() */
1542 local_irq_save(flags);
1544 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1545 __queue_delayed_work(cpu, wq, dwork, delay);
1549 local_irq_restore(flags);
1552 EXPORT_SYMBOL(queue_delayed_work_on);
1555 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1556 * @cpu: CPU number to execute work on
1557 * @wq: workqueue to use
1558 * @dwork: work to queue
1559 * @delay: number of jiffies to wait before queueing
1561 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1562 * modify @dwork's timer so that it expires after @delay. If @delay is
1563 * zero, @work is guaranteed to be scheduled immediately regardless of its
1566 * Return: %false if @dwork was idle and queued, %true if @dwork was
1567 * pending and its timer was modified.
1569 * This function is safe to call from any context including IRQ handler.
1570 * See try_to_grab_pending() for details.
1572 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1573 struct delayed_work *dwork, unsigned long delay)
1575 unsigned long flags;
1579 ret = try_to_grab_pending(&dwork->work, true, &flags);
1580 } while (unlikely(ret == -EAGAIN));
1582 if (likely(ret >= 0)) {
1583 __queue_delayed_work(cpu, wq, dwork, delay);
1584 local_irq_restore(flags);
1587 /* -ENOENT from try_to_grab_pending() becomes %true */
1590 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1593 * worker_enter_idle - enter idle state
1594 * @worker: worker which is entering idle state
1596 * @worker is entering idle state. Update stats and idle timer if
1600 * spin_lock_irq(pool->lock).
1602 static void worker_enter_idle(struct worker *worker)
1604 struct worker_pool *pool = worker->pool;
1606 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1607 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1608 (worker->hentry.next || worker->hentry.pprev)))
1611 /* can't use worker_set_flags(), also called from create_worker() */
1612 worker->flags |= WORKER_IDLE;
1614 worker->last_active = jiffies;
1616 /* idle_list is LIFO */
1617 list_add(&worker->entry, &pool->idle_list);
1619 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1620 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1623 * Sanity check nr_running. Because wq_unbind_fn() releases
1624 * pool->lock between setting %WORKER_UNBOUND and zapping
1625 * nr_running, the warning may trigger spuriously. Check iff
1626 * unbind is not in progress.
1628 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1629 pool->nr_workers == pool->nr_idle &&
1630 atomic_read(&pool->nr_running));
1634 * worker_leave_idle - leave idle state
1635 * @worker: worker which is leaving idle state
1637 * @worker is leaving idle state. Update stats.
1640 * spin_lock_irq(pool->lock).
1642 static void worker_leave_idle(struct worker *worker)
1644 struct worker_pool *pool = worker->pool;
1646 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1648 worker_clr_flags(worker, WORKER_IDLE);
1650 list_del_init(&worker->entry);
1653 static struct worker *alloc_worker(int node)
1655 struct worker *worker;
1657 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1659 INIT_LIST_HEAD(&worker->entry);
1660 INIT_LIST_HEAD(&worker->scheduled);
1661 INIT_LIST_HEAD(&worker->node);
1662 /* on creation a worker is in !idle && prep state */
1663 worker->flags = WORKER_PREP;
1669 * worker_attach_to_pool() - attach a worker to a pool
1670 * @worker: worker to be attached
1671 * @pool: the target pool
1673 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1674 * cpu-binding of @worker are kept coordinated with the pool across
1677 static void worker_attach_to_pool(struct worker *worker,
1678 struct worker_pool *pool)
1680 mutex_lock(&pool->attach_mutex);
1683 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1684 * online CPUs. It'll be re-applied when any of the CPUs come up.
1686 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1689 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1690 * stable across this function. See the comments above the
1691 * flag definition for details.
1693 if (pool->flags & POOL_DISASSOCIATED)
1694 worker->flags |= WORKER_UNBOUND;
1696 list_add_tail(&worker->node, &pool->workers);
1698 mutex_unlock(&pool->attach_mutex);
1702 * worker_detach_from_pool() - detach a worker from its pool
1703 * @worker: worker which is attached to its pool
1704 * @pool: the pool @worker is attached to
1706 * Undo the attaching which had been done in worker_attach_to_pool(). The
1707 * caller worker shouldn't access to the pool after detached except it has
1708 * other reference to the pool.
1710 static void worker_detach_from_pool(struct worker *worker,
1711 struct worker_pool *pool)
1713 struct completion *detach_completion = NULL;
1715 mutex_lock(&pool->attach_mutex);
1716 list_del(&worker->node);
1717 if (list_empty(&pool->workers))
1718 detach_completion = pool->detach_completion;
1719 mutex_unlock(&pool->attach_mutex);
1721 /* clear leftover flags without pool->lock after it is detached */
1722 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1724 if (detach_completion)
1725 complete(detach_completion);
1729 * create_worker - create a new workqueue worker
1730 * @pool: pool the new worker will belong to
1732 * Create and start a new worker which is attached to @pool.
1735 * Might sleep. Does GFP_KERNEL allocations.
1738 * Pointer to the newly created worker.
1740 static struct worker *create_worker(struct worker_pool *pool)
1742 struct worker *worker = NULL;
1746 /* ID is needed to determine kthread name */
1747 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1751 worker = alloc_worker(pool->node);
1755 worker->pool = pool;
1759 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1760 pool->attrs->nice < 0 ? "H" : "");
1762 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1764 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1765 "kworker/%s", id_buf);
1766 if (IS_ERR(worker->task))
1769 set_user_nice(worker->task, pool->attrs->nice);
1770 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1772 /* successful, attach the worker to the pool */
1773 worker_attach_to_pool(worker, pool);
1775 /* start the newly created worker */
1776 spin_lock_irq(&pool->lock);
1777 worker->pool->nr_workers++;
1778 worker_enter_idle(worker);
1779 wake_up_process(worker->task);
1780 spin_unlock_irq(&pool->lock);
1786 ida_simple_remove(&pool->worker_ida, id);
1792 * destroy_worker - destroy a workqueue worker
1793 * @worker: worker to be destroyed
1795 * Destroy @worker and adjust @pool stats accordingly. The worker should
1799 * spin_lock_irq(pool->lock).
1801 static void destroy_worker(struct worker *worker)
1803 struct worker_pool *pool = worker->pool;
1805 lockdep_assert_held(&pool->lock);
1807 /* sanity check frenzy */
1808 if (WARN_ON(worker->current_work) ||
1809 WARN_ON(!list_empty(&worker->scheduled)) ||
1810 WARN_ON(!(worker->flags & WORKER_IDLE)))
1816 list_del_init(&worker->entry);
1817 worker->flags |= WORKER_DIE;
1818 wake_up_process(worker->task);
1821 static void idle_worker_timeout(unsigned long __pool)
1823 struct worker_pool *pool = (void *)__pool;
1825 spin_lock_irq(&pool->lock);
1827 while (too_many_workers(pool)) {
1828 struct worker *worker;
1829 unsigned long expires;
1831 /* idle_list is kept in LIFO order, check the last one */
1832 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1833 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1835 if (time_before(jiffies, expires)) {
1836 mod_timer(&pool->idle_timer, expires);
1840 destroy_worker(worker);
1843 spin_unlock_irq(&pool->lock);
1846 static void send_mayday(struct work_struct *work)
1848 struct pool_workqueue *pwq = get_work_pwq(work);
1849 struct workqueue_struct *wq = pwq->wq;
1851 lockdep_assert_held(&wq_mayday_lock);
1856 /* mayday mayday mayday */
1857 if (list_empty(&pwq->mayday_node)) {
1859 * If @pwq is for an unbound wq, its base ref may be put at
1860 * any time due to an attribute change. Pin @pwq until the
1861 * rescuer is done with it.
1864 list_add_tail(&pwq->mayday_node, &wq->maydays);
1865 wake_up_process(wq->rescuer->task);
1869 static void pool_mayday_timeout(unsigned long __pool)
1871 struct worker_pool *pool = (void *)__pool;
1872 struct work_struct *work;
1874 spin_lock_irq(&pool->lock);
1875 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1877 if (need_to_create_worker(pool)) {
1879 * We've been trying to create a new worker but
1880 * haven't been successful. We might be hitting an
1881 * allocation deadlock. Send distress signals to
1884 list_for_each_entry(work, &pool->worklist, entry)
1888 spin_unlock(&wq_mayday_lock);
1889 spin_unlock_irq(&pool->lock);
1891 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1895 * maybe_create_worker - create a new worker if necessary
1896 * @pool: pool to create a new worker for
1898 * Create a new worker for @pool if necessary. @pool is guaranteed to
1899 * have at least one idle worker on return from this function. If
1900 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1901 * sent to all rescuers with works scheduled on @pool to resolve
1902 * possible allocation deadlock.
1904 * On return, need_to_create_worker() is guaranteed to be %false and
1905 * may_start_working() %true.
1908 * spin_lock_irq(pool->lock) which may be released and regrabbed
1909 * multiple times. Does GFP_KERNEL allocations. Called only from
1912 static void maybe_create_worker(struct worker_pool *pool)
1913 __releases(&pool->lock)
1914 __acquires(&pool->lock)
1917 spin_unlock_irq(&pool->lock);
1919 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1920 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1923 if (create_worker(pool) || !need_to_create_worker(pool))
1926 schedule_timeout_interruptible(CREATE_COOLDOWN);
1928 if (!need_to_create_worker(pool))
1932 del_timer_sync(&pool->mayday_timer);
1933 spin_lock_irq(&pool->lock);
1935 * This is necessary even after a new worker was just successfully
1936 * created as @pool->lock was dropped and the new worker might have
1937 * already become busy.
1939 if (need_to_create_worker(pool))
1944 * manage_workers - manage worker pool
1947 * Assume the manager role and manage the worker pool @worker belongs
1948 * to. At any given time, there can be only zero or one manager per
1949 * pool. The exclusion is handled automatically by this function.
1951 * The caller can safely start processing works on false return. On
1952 * true return, it's guaranteed that need_to_create_worker() is false
1953 * and may_start_working() is true.
1956 * spin_lock_irq(pool->lock) which may be released and regrabbed
1957 * multiple times. Does GFP_KERNEL allocations.
1960 * %false if the pool doesn't need management and the caller can safely
1961 * start processing works, %true if management function was performed and
1962 * the conditions that the caller verified before calling the function may
1963 * no longer be true.
1965 static bool manage_workers(struct worker *worker)
1967 struct worker_pool *pool = worker->pool;
1969 if (pool->flags & POOL_MANAGER_ACTIVE)
1972 pool->flags |= POOL_MANAGER_ACTIVE;
1973 pool->manager = worker;
1975 maybe_create_worker(pool);
1977 pool->manager = NULL;
1978 pool->flags &= ~POOL_MANAGER_ACTIVE;
1979 wake_up(&wq_manager_wait);
1984 * process_one_work - process single work
1986 * @work: work to process
1988 * Process @work. This function contains all the logics necessary to
1989 * process a single work including synchronization against and
1990 * interaction with other workers on the same cpu, queueing and
1991 * flushing. As long as context requirement is met, any worker can
1992 * call this function to process a work.
1995 * spin_lock_irq(pool->lock) which is released and regrabbed.
1997 static void process_one_work(struct worker *worker, struct work_struct *work)
1998 __releases(&pool->lock)
1999 __acquires(&pool->lock)
2001 struct pool_workqueue *pwq = get_work_pwq(work);
2002 struct worker_pool *pool = worker->pool;
2003 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2005 struct worker *collision;
2006 #ifdef CONFIG_LOCKDEP
2008 * It is permissible to free the struct work_struct from
2009 * inside the function that is called from it, this we need to
2010 * take into account for lockdep too. To avoid bogus "held
2011 * lock freed" warnings as well as problems when looking into
2012 * work->lockdep_map, make a copy and use that here.
2014 struct lockdep_map lockdep_map;
2016 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2018 /* ensure we're on the correct CPU */
2019 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2020 raw_smp_processor_id() != pool->cpu);
2023 * A single work shouldn't be executed concurrently by
2024 * multiple workers on a single cpu. Check whether anyone is
2025 * already processing the work. If so, defer the work to the
2026 * currently executing one.
2028 collision = find_worker_executing_work(pool, work);
2029 if (unlikely(collision)) {
2030 move_linked_works(work, &collision->scheduled, NULL);
2034 /* claim and dequeue */
2035 debug_work_deactivate(work);
2036 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2037 worker->current_work = work;
2038 worker->current_func = work->func;
2039 worker->current_pwq = pwq;
2040 work_color = get_work_color(work);
2042 list_del_init(&work->entry);
2045 * CPU intensive works don't participate in concurrency management.
2046 * They're the scheduler's responsibility. This takes @worker out
2047 * of concurrency management and the next code block will chain
2048 * execution of the pending work items.
2050 if (unlikely(cpu_intensive))
2051 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2054 * Wake up another worker if necessary. The condition is always
2055 * false for normal per-cpu workers since nr_running would always
2056 * be >= 1 at this point. This is used to chain execution of the
2057 * pending work items for WORKER_NOT_RUNNING workers such as the
2058 * UNBOUND and CPU_INTENSIVE ones.
2060 if (need_more_worker(pool))
2061 wake_up_worker(pool);
2064 * Record the last pool and clear PENDING which should be the last
2065 * update to @work. Also, do this inside @pool->lock so that
2066 * PENDING and queued state changes happen together while IRQ is
2069 set_work_pool_and_clear_pending(work, pool->id);
2071 spin_unlock_irq(&pool->lock);
2073 lock_map_acquire_read(&pwq->wq->lockdep_map);
2074 lock_map_acquire(&lockdep_map);
2075 trace_workqueue_execute_start(work);
2076 worker->current_func(work);
2078 * While we must be careful to not use "work" after this, the trace
2079 * point will only record its address.
2081 trace_workqueue_execute_end(work);
2082 lock_map_release(&lockdep_map);
2083 lock_map_release(&pwq->wq->lockdep_map);
2085 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2086 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2087 " last function: %pf\n",
2088 current->comm, preempt_count(), task_pid_nr(current),
2089 worker->current_func);
2090 debug_show_held_locks(current);
2091 BUG_ON(PANIC_CORRUPTION);
2096 * The following prevents a kworker from hogging CPU on !PREEMPT
2097 * kernels, where a requeueing work item waiting for something to
2098 * happen could deadlock with stop_machine as such work item could
2099 * indefinitely requeue itself while all other CPUs are trapped in
2100 * stop_machine. At the same time, report a quiescent RCU state so
2101 * the same condition doesn't freeze RCU.
2103 cond_resched_rcu_qs();
2105 spin_lock_irq(&pool->lock);
2107 /* clear cpu intensive status */
2108 if (unlikely(cpu_intensive))
2109 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2111 /* we're done with it, release */
2112 hash_del(&worker->hentry);
2113 worker->current_work = NULL;
2114 worker->current_func = NULL;
2115 worker->current_pwq = NULL;
2116 worker->desc_valid = false;
2117 pwq_dec_nr_in_flight(pwq, work_color);
2121 * process_scheduled_works - process scheduled works
2124 * Process all scheduled works. Please note that the scheduled list
2125 * may change while processing a work, so this function repeatedly
2126 * fetches a work from the top and executes it.
2129 * spin_lock_irq(pool->lock) which may be released and regrabbed
2132 static void process_scheduled_works(struct worker *worker)
2134 while (!list_empty(&worker->scheduled)) {
2135 struct work_struct *work = list_first_entry(&worker->scheduled,
2136 struct work_struct, entry);
2137 process_one_work(worker, work);
2142 * worker_thread - the worker thread function
2145 * The worker thread function. All workers belong to a worker_pool -
2146 * either a per-cpu one or dynamic unbound one. These workers process all
2147 * work items regardless of their specific target workqueue. The only
2148 * exception is work items which belong to workqueues with a rescuer which
2149 * will be explained in rescuer_thread().
2153 static int worker_thread(void *__worker)
2155 struct worker *worker = __worker;
2156 struct worker_pool *pool = worker->pool;
2158 /* tell the scheduler that this is a workqueue worker */
2159 worker->task->flags |= PF_WQ_WORKER;
2161 spin_lock_irq(&pool->lock);
2163 /* am I supposed to die? */
2164 if (unlikely(worker->flags & WORKER_DIE)) {
2165 spin_unlock_irq(&pool->lock);
2166 WARN_ON_ONCE(!list_empty(&worker->entry));
2167 worker->task->flags &= ~PF_WQ_WORKER;
2169 set_task_comm(worker->task, "kworker/dying");
2170 ida_simple_remove(&pool->worker_ida, worker->id);
2171 worker_detach_from_pool(worker, pool);
2176 worker_leave_idle(worker);
2178 /* no more worker necessary? */
2179 if (!need_more_worker(pool))
2182 /* do we need to manage? */
2183 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2187 * ->scheduled list can only be filled while a worker is
2188 * preparing to process a work or actually processing it.
2189 * Make sure nobody diddled with it while I was sleeping.
2191 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2194 * Finish PREP stage. We're guaranteed to have at least one idle
2195 * worker or that someone else has already assumed the manager
2196 * role. This is where @worker starts participating in concurrency
2197 * management if applicable and concurrency management is restored
2198 * after being rebound. See rebind_workers() for details.
2200 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2203 struct work_struct *work =
2204 list_first_entry(&pool->worklist,
2205 struct work_struct, entry);
2207 pool->watchdog_ts = jiffies;
2209 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2210 /* optimization path, not strictly necessary */
2211 process_one_work(worker, work);
2212 if (unlikely(!list_empty(&worker->scheduled)))
2213 process_scheduled_works(worker);
2215 move_linked_works(work, &worker->scheduled, NULL);
2216 process_scheduled_works(worker);
2218 } while (keep_working(pool));
2220 worker_set_flags(worker, WORKER_PREP);
2223 * pool->lock is held and there's no work to process and no need to
2224 * manage, sleep. Workers are woken up only while holding
2225 * pool->lock or from local cpu, so setting the current state
2226 * before releasing pool->lock is enough to prevent losing any
2229 worker_enter_idle(worker);
2230 __set_current_state(TASK_INTERRUPTIBLE);
2231 spin_unlock_irq(&pool->lock);
2237 * rescuer_thread - the rescuer thread function
2240 * Workqueue rescuer thread function. There's one rescuer for each
2241 * workqueue which has WQ_MEM_RECLAIM set.
2243 * Regular work processing on a pool may block trying to create a new
2244 * worker which uses GFP_KERNEL allocation which has slight chance of
2245 * developing into deadlock if some works currently on the same queue
2246 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2247 * the problem rescuer solves.
2249 * When such condition is possible, the pool summons rescuers of all
2250 * workqueues which have works queued on the pool and let them process
2251 * those works so that forward progress can be guaranteed.
2253 * This should happen rarely.
2257 static int rescuer_thread(void *__rescuer)
2259 struct worker *rescuer = __rescuer;
2260 struct workqueue_struct *wq = rescuer->rescue_wq;
2261 struct list_head *scheduled = &rescuer->scheduled;
2264 set_user_nice(current, RESCUER_NICE_LEVEL);
2267 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2268 * doesn't participate in concurrency management.
2270 rescuer->task->flags |= PF_WQ_WORKER;
2272 set_current_state(TASK_INTERRUPTIBLE);
2275 * By the time the rescuer is requested to stop, the workqueue
2276 * shouldn't have any work pending, but @wq->maydays may still have
2277 * pwq(s) queued. This can happen by non-rescuer workers consuming
2278 * all the work items before the rescuer got to them. Go through
2279 * @wq->maydays processing before acting on should_stop so that the
2280 * list is always empty on exit.
2282 should_stop = kthread_should_stop();
2284 /* see whether any pwq is asking for help */
2285 spin_lock_irq(&wq_mayday_lock);
2287 while (!list_empty(&wq->maydays)) {
2288 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2289 struct pool_workqueue, mayday_node);
2290 struct worker_pool *pool = pwq->pool;
2291 struct work_struct *work, *n;
2294 __set_current_state(TASK_RUNNING);
2295 list_del_init(&pwq->mayday_node);
2297 spin_unlock_irq(&wq_mayday_lock);
2299 worker_attach_to_pool(rescuer, pool);
2301 spin_lock_irq(&pool->lock);
2302 rescuer->pool = pool;
2305 * Slurp in all works issued via this workqueue and
2308 WARN_ON_ONCE(!list_empty(scheduled));
2309 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2310 if (get_work_pwq(work) == pwq) {
2312 pool->watchdog_ts = jiffies;
2313 move_linked_works(work, scheduled, &n);
2318 if (!list_empty(scheduled)) {
2319 process_scheduled_works(rescuer);
2322 * The above execution of rescued work items could
2323 * have created more to rescue through
2324 * pwq_activate_first_delayed() or chained
2325 * queueing. Let's put @pwq back on mayday list so
2326 * that such back-to-back work items, which may be
2327 * being used to relieve memory pressure, don't
2328 * incur MAYDAY_INTERVAL delay inbetween.
2330 if (need_to_create_worker(pool)) {
2331 spin_lock(&wq_mayday_lock);
2333 list_move_tail(&pwq->mayday_node, &wq->maydays);
2334 spin_unlock(&wq_mayday_lock);
2339 * Put the reference grabbed by send_mayday(). @pool won't
2340 * go away while we're still attached to it.
2345 * Leave this pool. If need_more_worker() is %true, notify a
2346 * regular worker; otherwise, we end up with 0 concurrency
2347 * and stalling the execution.
2349 if (need_more_worker(pool))
2350 wake_up_worker(pool);
2352 rescuer->pool = NULL;
2353 spin_unlock_irq(&pool->lock);
2355 worker_detach_from_pool(rescuer, pool);
2357 spin_lock_irq(&wq_mayday_lock);
2360 spin_unlock_irq(&wq_mayday_lock);
2363 __set_current_state(TASK_RUNNING);
2364 rescuer->task->flags &= ~PF_WQ_WORKER;
2368 /* rescuers should never participate in concurrency management */
2369 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2375 struct work_struct work;
2376 struct completion done;
2377 struct task_struct *task; /* purely informational */
2380 static void wq_barrier_func(struct work_struct *work)
2382 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2383 complete(&barr->done);
2387 * insert_wq_barrier - insert a barrier work
2388 * @pwq: pwq to insert barrier into
2389 * @barr: wq_barrier to insert
2390 * @target: target work to attach @barr to
2391 * @worker: worker currently executing @target, NULL if @target is not executing
2393 * @barr is linked to @target such that @barr is completed only after
2394 * @target finishes execution. Please note that the ordering
2395 * guarantee is observed only with respect to @target and on the local
2398 * Currently, a queued barrier can't be canceled. This is because
2399 * try_to_grab_pending() can't determine whether the work to be
2400 * grabbed is at the head of the queue and thus can't clear LINKED
2401 * flag of the previous work while there must be a valid next work
2402 * after a work with LINKED flag set.
2404 * Note that when @worker is non-NULL, @target may be modified
2405 * underneath us, so we can't reliably determine pwq from @target.
2408 * spin_lock_irq(pool->lock).
2410 static void insert_wq_barrier(struct pool_workqueue *pwq,
2411 struct wq_barrier *barr,
2412 struct work_struct *target, struct worker *worker)
2414 struct list_head *head;
2415 unsigned int linked = 0;
2418 * debugobject calls are safe here even with pool->lock locked
2419 * as we know for sure that this will not trigger any of the
2420 * checks and call back into the fixup functions where we
2423 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2424 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2425 init_completion(&barr->done);
2426 barr->task = current;
2429 * If @target is currently being executed, schedule the
2430 * barrier to the worker; otherwise, put it after @target.
2433 head = worker->scheduled.next;
2435 unsigned long *bits = work_data_bits(target);
2437 head = target->entry.next;
2438 /* there can already be other linked works, inherit and set */
2439 linked = *bits & WORK_STRUCT_LINKED;
2440 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2443 debug_work_activate(&barr->work);
2444 insert_work(pwq, &barr->work, head,
2445 work_color_to_flags(WORK_NO_COLOR) | linked);
2449 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2450 * @wq: workqueue being flushed
2451 * @flush_color: new flush color, < 0 for no-op
2452 * @work_color: new work color, < 0 for no-op
2454 * Prepare pwqs for workqueue flushing.
2456 * If @flush_color is non-negative, flush_color on all pwqs should be
2457 * -1. If no pwq has in-flight commands at the specified color, all
2458 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2459 * has in flight commands, its pwq->flush_color is set to
2460 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2461 * wakeup logic is armed and %true is returned.
2463 * The caller should have initialized @wq->first_flusher prior to
2464 * calling this function with non-negative @flush_color. If
2465 * @flush_color is negative, no flush color update is done and %false
2468 * If @work_color is non-negative, all pwqs should have the same
2469 * work_color which is previous to @work_color and all will be
2470 * advanced to @work_color.
2473 * mutex_lock(wq->mutex).
2476 * %true if @flush_color >= 0 and there's something to flush. %false
2479 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2480 int flush_color, int work_color)
2483 struct pool_workqueue *pwq;
2485 if (flush_color >= 0) {
2486 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2487 atomic_set(&wq->nr_pwqs_to_flush, 1);
2490 for_each_pwq(pwq, wq) {
2491 struct worker_pool *pool = pwq->pool;
2493 spin_lock_irq(&pool->lock);
2495 if (flush_color >= 0) {
2496 WARN_ON_ONCE(pwq->flush_color != -1);
2498 if (pwq->nr_in_flight[flush_color]) {
2499 pwq->flush_color = flush_color;
2500 atomic_inc(&wq->nr_pwqs_to_flush);
2505 if (work_color >= 0) {
2506 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2507 pwq->work_color = work_color;
2510 spin_unlock_irq(&pool->lock);
2513 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2514 complete(&wq->first_flusher->done);
2520 * flush_workqueue - ensure that any scheduled work has run to completion.
2521 * @wq: workqueue to flush
2523 * This function sleeps until all work items which were queued on entry
2524 * have finished execution, but it is not livelocked by new incoming ones.
2526 void flush_workqueue(struct workqueue_struct *wq)
2528 struct wq_flusher this_flusher = {
2529 .list = LIST_HEAD_INIT(this_flusher.list),
2531 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2535 lock_map_acquire(&wq->lockdep_map);
2536 lock_map_release(&wq->lockdep_map);
2538 mutex_lock(&wq->mutex);
2541 * Start-to-wait phase
2543 next_color = work_next_color(wq->work_color);
2545 if (next_color != wq->flush_color) {
2547 * Color space is not full. The current work_color
2548 * becomes our flush_color and work_color is advanced
2551 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2552 this_flusher.flush_color = wq->work_color;
2553 wq->work_color = next_color;
2555 if (!wq->first_flusher) {
2556 /* no flush in progress, become the first flusher */
2557 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2559 wq->first_flusher = &this_flusher;
2561 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2563 /* nothing to flush, done */
2564 wq->flush_color = next_color;
2565 wq->first_flusher = NULL;
2570 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2571 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2572 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2576 * Oops, color space is full, wait on overflow queue.
2577 * The next flush completion will assign us
2578 * flush_color and transfer to flusher_queue.
2580 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2583 mutex_unlock(&wq->mutex);
2585 wait_for_completion(&this_flusher.done);
2588 * Wake-up-and-cascade phase
2590 * First flushers are responsible for cascading flushes and
2591 * handling overflow. Non-first flushers can simply return.
2593 if (wq->first_flusher != &this_flusher)
2596 mutex_lock(&wq->mutex);
2598 /* we might have raced, check again with mutex held */
2599 if (wq->first_flusher != &this_flusher)
2602 wq->first_flusher = NULL;
2604 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2605 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2608 struct wq_flusher *next, *tmp;
2610 /* complete all the flushers sharing the current flush color */
2611 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2612 if (next->flush_color != wq->flush_color)
2614 list_del_init(&next->list);
2615 complete(&next->done);
2618 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2619 wq->flush_color != work_next_color(wq->work_color));
2621 /* this flush_color is finished, advance by one */
2622 wq->flush_color = work_next_color(wq->flush_color);
2624 /* one color has been freed, handle overflow queue */
2625 if (!list_empty(&wq->flusher_overflow)) {
2627 * Assign the same color to all overflowed
2628 * flushers, advance work_color and append to
2629 * flusher_queue. This is the start-to-wait
2630 * phase for these overflowed flushers.
2632 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2633 tmp->flush_color = wq->work_color;
2635 wq->work_color = work_next_color(wq->work_color);
2637 list_splice_tail_init(&wq->flusher_overflow,
2638 &wq->flusher_queue);
2639 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2642 if (list_empty(&wq->flusher_queue)) {
2643 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2648 * Need to flush more colors. Make the next flusher
2649 * the new first flusher and arm pwqs.
2651 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2652 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2654 list_del_init(&next->list);
2655 wq->first_flusher = next;
2657 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2661 * Meh... this color is already done, clear first
2662 * flusher and repeat cascading.
2664 wq->first_flusher = NULL;
2668 mutex_unlock(&wq->mutex);
2670 EXPORT_SYMBOL(flush_workqueue);
2673 * drain_workqueue - drain a workqueue
2674 * @wq: workqueue to drain
2676 * Wait until the workqueue becomes empty. While draining is in progress,
2677 * only chain queueing is allowed. IOW, only currently pending or running
2678 * work items on @wq can queue further work items on it. @wq is flushed
2679 * repeatedly until it becomes empty. The number of flushing is determined
2680 * by the depth of chaining and should be relatively short. Whine if it
2683 void drain_workqueue(struct workqueue_struct *wq)
2685 unsigned int flush_cnt = 0;
2686 struct pool_workqueue *pwq;
2689 * __queue_work() needs to test whether there are drainers, is much
2690 * hotter than drain_workqueue() and already looks at @wq->flags.
2691 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2693 mutex_lock(&wq->mutex);
2694 if (!wq->nr_drainers++)
2695 wq->flags |= __WQ_DRAINING;
2696 mutex_unlock(&wq->mutex);
2698 flush_workqueue(wq);
2700 mutex_lock(&wq->mutex);
2702 for_each_pwq(pwq, wq) {
2705 spin_lock_irq(&pwq->pool->lock);
2706 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2707 spin_unlock_irq(&pwq->pool->lock);
2712 if (++flush_cnt == 10 ||
2713 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2714 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2715 wq->name, flush_cnt);
2717 mutex_unlock(&wq->mutex);
2721 if (!--wq->nr_drainers)
2722 wq->flags &= ~__WQ_DRAINING;
2723 mutex_unlock(&wq->mutex);
2725 EXPORT_SYMBOL_GPL(drain_workqueue);
2727 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2729 struct worker *worker = NULL;
2730 struct worker_pool *pool;
2731 struct pool_workqueue *pwq;
2735 local_irq_disable();
2736 pool = get_work_pool(work);
2742 spin_lock(&pool->lock);
2743 /* see the comment in try_to_grab_pending() with the same code */
2744 pwq = get_work_pwq(work);
2746 if (unlikely(pwq->pool != pool))
2749 worker = find_worker_executing_work(pool, work);
2752 pwq = worker->current_pwq;
2755 insert_wq_barrier(pwq, barr, work, worker);
2756 spin_unlock_irq(&pool->lock);
2759 * If @max_active is 1 or rescuer is in use, flushing another work
2760 * item on the same workqueue may lead to deadlock. Make sure the
2761 * flusher is not running on the same workqueue by verifying write
2764 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2765 lock_map_acquire(&pwq->wq->lockdep_map);
2767 lock_map_acquire_read(&pwq->wq->lockdep_map);
2768 lock_map_release(&pwq->wq->lockdep_map);
2772 spin_unlock_irq(&pool->lock);
2777 * flush_work - wait for a work to finish executing the last queueing instance
2778 * @work: the work to flush
2780 * Wait until @work has finished execution. @work is guaranteed to be idle
2781 * on return if it hasn't been requeued since flush started.
2784 * %true if flush_work() waited for the work to finish execution,
2785 * %false if it was already idle.
2787 bool flush_work(struct work_struct *work)
2789 struct wq_barrier barr;
2791 lock_map_acquire(&work->lockdep_map);
2792 lock_map_release(&work->lockdep_map);
2794 if (start_flush_work(work, &barr)) {
2795 wait_for_completion(&barr.done);
2796 destroy_work_on_stack(&barr.work);
2802 EXPORT_SYMBOL_GPL(flush_work);
2806 struct work_struct *work;
2809 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2811 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2813 if (cwait->work != key)
2815 return autoremove_wake_function(wait, mode, sync, key);
2818 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2820 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2821 unsigned long flags;
2825 ret = try_to_grab_pending(work, is_dwork, &flags);
2827 * If someone else is already canceling, wait for it to
2828 * finish. flush_work() doesn't work for PREEMPT_NONE
2829 * because we may get scheduled between @work's completion
2830 * and the other canceling task resuming and clearing
2831 * CANCELING - flush_work() will return false immediately
2832 * as @work is no longer busy, try_to_grab_pending() will
2833 * return -ENOENT as @work is still being canceled and the
2834 * other canceling task won't be able to clear CANCELING as
2835 * we're hogging the CPU.
2837 * Let's wait for completion using a waitqueue. As this
2838 * may lead to the thundering herd problem, use a custom
2839 * wake function which matches @work along with exclusive
2842 if (unlikely(ret == -ENOENT)) {
2843 struct cwt_wait cwait;
2845 init_wait(&cwait.wait);
2846 cwait.wait.func = cwt_wakefn;
2849 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2850 TASK_UNINTERRUPTIBLE);
2851 if (work_is_canceling(work))
2853 finish_wait(&cancel_waitq, &cwait.wait);
2855 } while (unlikely(ret < 0));
2857 /* tell other tasks trying to grab @work to back off */
2858 mark_work_canceling(work);
2859 local_irq_restore(flags);
2862 clear_work_data(work);
2865 * Paired with prepare_to_wait() above so that either
2866 * waitqueue_active() is visible here or !work_is_canceling() is
2870 if (waitqueue_active(&cancel_waitq))
2871 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2877 * cancel_work_sync - cancel a work and wait for it to finish
2878 * @work: the work to cancel
2880 * Cancel @work and wait for its execution to finish. This function
2881 * can be used even if the work re-queues itself or migrates to
2882 * another workqueue. On return from this function, @work is
2883 * guaranteed to be not pending or executing on any CPU.
2885 * cancel_work_sync(&delayed_work->work) must not be used for
2886 * delayed_work's. Use cancel_delayed_work_sync() instead.
2888 * The caller must ensure that the workqueue on which @work was last
2889 * queued can't be destroyed before this function returns.
2892 * %true if @work was pending, %false otherwise.
2894 bool cancel_work_sync(struct work_struct *work)
2896 return __cancel_work_timer(work, false);
2898 EXPORT_SYMBOL_GPL(cancel_work_sync);
2901 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2902 * @dwork: the delayed work to flush
2904 * Delayed timer is cancelled and the pending work is queued for
2905 * immediate execution. Like flush_work(), this function only
2906 * considers the last queueing instance of @dwork.
2909 * %true if flush_work() waited for the work to finish execution,
2910 * %false if it was already idle.
2912 bool flush_delayed_work(struct delayed_work *dwork)
2914 local_irq_disable();
2915 if (del_timer_sync(&dwork->timer))
2916 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2918 return flush_work(&dwork->work);
2920 EXPORT_SYMBOL(flush_delayed_work);
2922 static bool __cancel_work(struct work_struct *work, bool is_dwork)
2924 unsigned long flags;
2928 ret = try_to_grab_pending(work, is_dwork, &flags);
2929 } while (unlikely(ret == -EAGAIN));
2931 if (unlikely(ret < 0))
2934 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
2935 local_irq_restore(flags);
2940 * See cancel_delayed_work()
2942 bool cancel_work(struct work_struct *work)
2944 return __cancel_work(work, false);
2948 * cancel_delayed_work - cancel a delayed work
2949 * @dwork: delayed_work to cancel
2951 * Kill off a pending delayed_work.
2953 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2957 * The work callback function may still be running on return, unless
2958 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2959 * use cancel_delayed_work_sync() to wait on it.
2961 * This function is safe to call from any context including IRQ handler.
2963 bool cancel_delayed_work(struct delayed_work *dwork)
2965 return __cancel_work(&dwork->work, true);
2967 EXPORT_SYMBOL(cancel_delayed_work);
2970 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2971 * @dwork: the delayed work cancel
2973 * This is cancel_work_sync() for delayed works.
2976 * %true if @dwork was pending, %false otherwise.
2978 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2980 return __cancel_work_timer(&dwork->work, true);
2982 EXPORT_SYMBOL(cancel_delayed_work_sync);
2985 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2986 * @func: the function to call
2988 * schedule_on_each_cpu() executes @func on each online CPU using the
2989 * system workqueue and blocks until all CPUs have completed.
2990 * schedule_on_each_cpu() is very slow.
2993 * 0 on success, -errno on failure.
2995 int schedule_on_each_cpu(work_func_t func)
2998 struct work_struct __percpu *works;
3000 works = alloc_percpu(struct work_struct);
3006 for_each_online_cpu(cpu) {
3007 struct work_struct *work = per_cpu_ptr(works, cpu);
3009 INIT_WORK(work, func);
3010 schedule_work_on(cpu, work);
3013 for_each_online_cpu(cpu)
3014 flush_work(per_cpu_ptr(works, cpu));
3022 * execute_in_process_context - reliably execute the routine with user context
3023 * @fn: the function to execute
3024 * @ew: guaranteed storage for the execute work structure (must
3025 * be available when the work executes)
3027 * Executes the function immediately if process context is available,
3028 * otherwise schedules the function for delayed execution.
3030 * Return: 0 - function was executed
3031 * 1 - function was scheduled for execution
3033 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3035 if (!in_interrupt()) {
3040 INIT_WORK(&ew->work, fn);
3041 schedule_work(&ew->work);
3045 EXPORT_SYMBOL_GPL(execute_in_process_context);
3048 * free_workqueue_attrs - free a workqueue_attrs
3049 * @attrs: workqueue_attrs to free
3051 * Undo alloc_workqueue_attrs().
3053 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3056 free_cpumask_var(attrs->cpumask);
3062 * alloc_workqueue_attrs - allocate a workqueue_attrs
3063 * @gfp_mask: allocation mask to use
3065 * Allocate a new workqueue_attrs, initialize with default settings and
3068 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3070 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3072 struct workqueue_attrs *attrs;
3074 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3077 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3080 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3083 free_workqueue_attrs(attrs);
3087 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3088 const struct workqueue_attrs *from)
3090 to->nice = from->nice;
3091 cpumask_copy(to->cpumask, from->cpumask);
3093 * Unlike hash and equality test, this function doesn't ignore
3094 * ->no_numa as it is used for both pool and wq attrs. Instead,
3095 * get_unbound_pool() explicitly clears ->no_numa after copying.
3097 to->no_numa = from->no_numa;
3100 /* hash value of the content of @attr */
3101 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3105 hash = jhash_1word(attrs->nice, hash);
3106 hash = jhash(cpumask_bits(attrs->cpumask),
3107 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3111 /* content equality test */
3112 static bool wqattrs_equal(const struct workqueue_attrs *a,
3113 const struct workqueue_attrs *b)
3115 if (a->nice != b->nice)
3117 if (!cpumask_equal(a->cpumask, b->cpumask))
3123 * init_worker_pool - initialize a newly zalloc'd worker_pool
3124 * @pool: worker_pool to initialize
3126 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3128 * Return: 0 on success, -errno on failure. Even on failure, all fields
3129 * inside @pool proper are initialized and put_unbound_pool() can be called
3130 * on @pool safely to release it.
3132 static int init_worker_pool(struct worker_pool *pool)
3134 spin_lock_init(&pool->lock);
3137 pool->node = NUMA_NO_NODE;
3138 pool->flags |= POOL_DISASSOCIATED;
3139 pool->watchdog_ts = jiffies;
3140 INIT_LIST_HEAD(&pool->worklist);
3141 INIT_LIST_HEAD(&pool->idle_list);
3142 hash_init(pool->busy_hash);
3144 init_timer_deferrable(&pool->idle_timer);
3145 pool->idle_timer.function = idle_worker_timeout;
3146 pool->idle_timer.data = (unsigned long)pool;
3148 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3149 (unsigned long)pool);
3151 mutex_init(&pool->attach_mutex);
3152 INIT_LIST_HEAD(&pool->workers);
3154 ida_init(&pool->worker_ida);
3155 INIT_HLIST_NODE(&pool->hash_node);
3158 /* shouldn't fail above this point */
3159 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3165 static void rcu_free_wq(struct rcu_head *rcu)
3167 struct workqueue_struct *wq =
3168 container_of(rcu, struct workqueue_struct, rcu);
3170 if (!(wq->flags & WQ_UNBOUND))
3171 free_percpu(wq->cpu_pwqs);
3173 free_workqueue_attrs(wq->unbound_attrs);
3179 static void rcu_free_pool(struct rcu_head *rcu)
3181 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3183 ida_destroy(&pool->worker_ida);
3184 free_workqueue_attrs(pool->attrs);
3189 * put_unbound_pool - put a worker_pool
3190 * @pool: worker_pool to put
3192 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3193 * safe manner. get_unbound_pool() calls this function on its failure path
3194 * and this function should be able to release pools which went through,
3195 * successfully or not, init_worker_pool().
3197 * Should be called with wq_pool_mutex held.
3199 static void put_unbound_pool(struct worker_pool *pool)
3201 DECLARE_COMPLETION_ONSTACK(detach_completion);
3202 struct worker *worker;
3204 lockdep_assert_held(&wq_pool_mutex);
3210 if (WARN_ON(!(pool->cpu < 0)) ||
3211 WARN_ON(!list_empty(&pool->worklist)))
3214 /* release id and unhash */
3216 idr_remove(&worker_pool_idr, pool->id);
3217 hash_del(&pool->hash_node);
3220 * Become the manager and destroy all workers. This prevents
3221 * @pool's workers from blocking on attach_mutex. We're the last
3222 * manager and @pool gets freed with the flag set.
3224 spin_lock_irq(&pool->lock);
3225 wait_event_lock_irq(wq_manager_wait,
3226 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3227 pool->flags |= POOL_MANAGER_ACTIVE;
3229 while ((worker = first_idle_worker(pool)))
3230 destroy_worker(worker);
3231 WARN_ON(pool->nr_workers || pool->nr_idle);
3232 spin_unlock_irq(&pool->lock);
3234 mutex_lock(&pool->attach_mutex);
3235 if (!list_empty(&pool->workers))
3236 pool->detach_completion = &detach_completion;
3237 mutex_unlock(&pool->attach_mutex);
3239 if (pool->detach_completion)
3240 wait_for_completion(pool->detach_completion);
3242 /* shut down the timers */
3243 del_timer_sync(&pool->idle_timer);
3244 del_timer_sync(&pool->mayday_timer);
3246 /* sched-RCU protected to allow dereferences from get_work_pool() */
3247 call_rcu_sched(&pool->rcu, rcu_free_pool);
3251 * get_unbound_pool - get a worker_pool with the specified attributes
3252 * @attrs: the attributes of the worker_pool to get
3254 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3255 * reference count and return it. If there already is a matching
3256 * worker_pool, it will be used; otherwise, this function attempts to
3259 * Should be called with wq_pool_mutex held.
3261 * Return: On success, a worker_pool with the same attributes as @attrs.
3262 * On failure, %NULL.
3264 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3266 u32 hash = wqattrs_hash(attrs);
3267 struct worker_pool *pool;
3269 int target_node = NUMA_NO_NODE;
3271 lockdep_assert_held(&wq_pool_mutex);
3273 /* do we already have a matching pool? */
3274 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3275 if (wqattrs_equal(pool->attrs, attrs)) {
3281 /* if cpumask is contained inside a NUMA node, we belong to that node */
3282 if (wq_numa_enabled) {
3283 for_each_node(node) {
3284 if (cpumask_subset(attrs->cpumask,
3285 wq_numa_possible_cpumask[node])) {
3292 /* nope, create a new one */
3293 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3294 if (!pool || init_worker_pool(pool) < 0)
3297 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3298 copy_workqueue_attrs(pool->attrs, attrs);
3299 pool->node = target_node;
3302 * no_numa isn't a worker_pool attribute, always clear it. See
3303 * 'struct workqueue_attrs' comments for detail.
3305 pool->attrs->no_numa = false;
3307 if (worker_pool_assign_id(pool) < 0)
3310 /* create and start the initial worker */
3311 if (!create_worker(pool))
3315 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3320 put_unbound_pool(pool);
3324 static void rcu_free_pwq(struct rcu_head *rcu)
3326 kmem_cache_free(pwq_cache,
3327 container_of(rcu, struct pool_workqueue, rcu));
3331 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3332 * and needs to be destroyed.
3334 static void pwq_unbound_release_workfn(struct work_struct *work)
3336 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3337 unbound_release_work);
3338 struct workqueue_struct *wq = pwq->wq;
3339 struct worker_pool *pool = pwq->pool;
3342 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3345 mutex_lock(&wq->mutex);
3346 list_del_rcu(&pwq->pwqs_node);
3347 is_last = list_empty(&wq->pwqs);
3348 mutex_unlock(&wq->mutex);
3350 mutex_lock(&wq_pool_mutex);
3351 put_unbound_pool(pool);
3352 mutex_unlock(&wq_pool_mutex);
3354 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3357 * If we're the last pwq going away, @wq is already dead and no one
3358 * is gonna access it anymore. Schedule RCU free.
3361 call_rcu_sched(&wq->rcu, rcu_free_wq);
3365 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3366 * @pwq: target pool_workqueue
3368 * If @pwq isn't freezing, set @pwq->max_active to the associated
3369 * workqueue's saved_max_active and activate delayed work items
3370 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3372 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3374 struct workqueue_struct *wq = pwq->wq;
3375 bool freezable = wq->flags & WQ_FREEZABLE;
3377 /* for @wq->saved_max_active */
3378 lockdep_assert_held(&wq->mutex);
3380 /* fast exit for non-freezable wqs */
3381 if (!freezable && pwq->max_active == wq->saved_max_active)
3384 spin_lock_irq(&pwq->pool->lock);
3387 * During [un]freezing, the caller is responsible for ensuring that
3388 * this function is called at least once after @workqueue_freezing
3389 * is updated and visible.
3391 if (!freezable || !workqueue_freezing) {
3392 pwq->max_active = wq->saved_max_active;
3394 while (!list_empty(&pwq->delayed_works) &&
3395 pwq->nr_active < pwq->max_active)
3396 pwq_activate_first_delayed(pwq);
3399 * Need to kick a worker after thawed or an unbound wq's
3400 * max_active is bumped. It's a slow path. Do it always.
3402 wake_up_worker(pwq->pool);
3404 pwq->max_active = 0;
3407 spin_unlock_irq(&pwq->pool->lock);
3410 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3411 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3412 struct worker_pool *pool)
3414 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3416 memset(pwq, 0, sizeof(*pwq));
3420 pwq->flush_color = -1;
3422 INIT_LIST_HEAD(&pwq->delayed_works);
3423 INIT_LIST_HEAD(&pwq->pwqs_node);
3424 INIT_LIST_HEAD(&pwq->mayday_node);
3425 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3428 /* sync @pwq with the current state of its associated wq and link it */
3429 static void link_pwq(struct pool_workqueue *pwq)
3431 struct workqueue_struct *wq = pwq->wq;
3433 lockdep_assert_held(&wq->mutex);
3435 /* may be called multiple times, ignore if already linked */
3436 if (!list_empty(&pwq->pwqs_node))
3439 /* set the matching work_color */
3440 pwq->work_color = wq->work_color;
3442 /* sync max_active to the current setting */
3443 pwq_adjust_max_active(pwq);
3446 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3449 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3450 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3451 const struct workqueue_attrs *attrs)
3453 struct worker_pool *pool;
3454 struct pool_workqueue *pwq;
3456 lockdep_assert_held(&wq_pool_mutex);
3458 pool = get_unbound_pool(attrs);
3462 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3464 put_unbound_pool(pool);
3468 init_pwq(pwq, wq, pool);
3473 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3474 * @attrs: the wq_attrs of the default pwq of the target workqueue
3475 * @node: the target NUMA node
3476 * @cpu_going_down: if >= 0, the CPU to consider as offline
3477 * @cpumask: outarg, the resulting cpumask
3479 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3480 * @cpu_going_down is >= 0, that cpu is considered offline during
3481 * calculation. The result is stored in @cpumask.
3483 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3484 * enabled and @node has online CPUs requested by @attrs, the returned
3485 * cpumask is the intersection of the possible CPUs of @node and
3488 * The caller is responsible for ensuring that the cpumask of @node stays
3491 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3494 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3495 int cpu_going_down, cpumask_t *cpumask)
3497 if (!wq_numa_enabled || attrs->no_numa)
3500 /* does @node have any online CPUs @attrs wants? */
3501 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3502 if (cpu_going_down >= 0)
3503 cpumask_clear_cpu(cpu_going_down, cpumask);
3505 if (cpumask_empty(cpumask))
3508 /* yeap, return possible CPUs in @node that @attrs wants */
3509 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3510 return !cpumask_equal(cpumask, attrs->cpumask);
3513 cpumask_copy(cpumask, attrs->cpumask);
3517 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3518 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3520 struct pool_workqueue *pwq)
3522 struct pool_workqueue *old_pwq;
3524 lockdep_assert_held(&wq_pool_mutex);
3525 lockdep_assert_held(&wq->mutex);
3527 /* link_pwq() can handle duplicate calls */
3530 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3531 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3535 /* context to store the prepared attrs & pwqs before applying */
3536 struct apply_wqattrs_ctx {
3537 struct workqueue_struct *wq; /* target workqueue */
3538 struct workqueue_attrs *attrs; /* attrs to apply */
3539 struct list_head list; /* queued for batching commit */
3540 struct pool_workqueue *dfl_pwq;
3541 struct pool_workqueue *pwq_tbl[];
3544 /* free the resources after success or abort */
3545 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3551 put_pwq_unlocked(ctx->pwq_tbl[node]);
3552 put_pwq_unlocked(ctx->dfl_pwq);
3554 free_workqueue_attrs(ctx->attrs);
3560 /* allocate the attrs and pwqs for later installation */
3561 static struct apply_wqattrs_ctx *
3562 apply_wqattrs_prepare(struct workqueue_struct *wq,
3563 const struct workqueue_attrs *attrs)
3565 struct apply_wqattrs_ctx *ctx;
3566 struct workqueue_attrs *new_attrs, *tmp_attrs;
3569 lockdep_assert_held(&wq_pool_mutex);
3571 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3574 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3575 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3576 if (!ctx || !new_attrs || !tmp_attrs)
3580 * Calculate the attrs of the default pwq.
3581 * If the user configured cpumask doesn't overlap with the
3582 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3584 copy_workqueue_attrs(new_attrs, attrs);
3585 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3586 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3587 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3590 * We may create multiple pwqs with differing cpumasks. Make a
3591 * copy of @new_attrs which will be modified and used to obtain
3594 copy_workqueue_attrs(tmp_attrs, new_attrs);
3597 * If something goes wrong during CPU up/down, we'll fall back to
3598 * the default pwq covering whole @attrs->cpumask. Always create
3599 * it even if we don't use it immediately.
3601 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3605 for_each_node(node) {
3606 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3607 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3608 if (!ctx->pwq_tbl[node])
3611 ctx->dfl_pwq->refcnt++;
3612 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3616 /* save the user configured attrs and sanitize it. */
3617 copy_workqueue_attrs(new_attrs, attrs);
3618 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3619 ctx->attrs = new_attrs;
3622 free_workqueue_attrs(tmp_attrs);
3626 free_workqueue_attrs(tmp_attrs);
3627 free_workqueue_attrs(new_attrs);
3628 apply_wqattrs_cleanup(ctx);
3632 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3633 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3637 /* all pwqs have been created successfully, let's install'em */
3638 mutex_lock(&ctx->wq->mutex);
3640 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3642 /* save the previous pwq and install the new one */
3644 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3645 ctx->pwq_tbl[node]);
3647 /* @dfl_pwq might not have been used, ensure it's linked */
3648 link_pwq(ctx->dfl_pwq);
3649 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3651 mutex_unlock(&ctx->wq->mutex);
3654 static void apply_wqattrs_lock(void)
3656 /* CPUs should stay stable across pwq creations and installations */
3658 mutex_lock(&wq_pool_mutex);
3661 static void apply_wqattrs_unlock(void)
3663 mutex_unlock(&wq_pool_mutex);
3667 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3668 const struct workqueue_attrs *attrs)
3670 struct apply_wqattrs_ctx *ctx;
3673 /* only unbound workqueues can change attributes */
3674 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3677 /* creating multiple pwqs breaks ordering guarantee */
3678 if (!list_empty(&wq->pwqs)) {
3679 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
3682 wq->flags &= ~__WQ_ORDERED;
3685 ctx = apply_wqattrs_prepare(wq, attrs);
3687 /* the ctx has been prepared successfully, let's commit it */
3689 apply_wqattrs_commit(ctx);
3693 apply_wqattrs_cleanup(ctx);
3699 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3700 * @wq: the target workqueue
3701 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3703 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3704 * machines, this function maps a separate pwq to each NUMA node with
3705 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3706 * NUMA node it was issued on. Older pwqs are released as in-flight work
3707 * items finish. Note that a work item which repeatedly requeues itself
3708 * back-to-back will stay on its current pwq.
3710 * Performs GFP_KERNEL allocations.
3712 * Return: 0 on success and -errno on failure.
3714 int apply_workqueue_attrs(struct workqueue_struct *wq,
3715 const struct workqueue_attrs *attrs)
3719 apply_wqattrs_lock();
3720 ret = apply_workqueue_attrs_locked(wq, attrs);
3721 apply_wqattrs_unlock();
3727 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3728 * @wq: the target workqueue
3729 * @cpu: the CPU coming up or going down
3730 * @online: whether @cpu is coming up or going down
3732 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3733 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3736 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3737 * falls back to @wq->dfl_pwq which may not be optimal but is always
3740 * Note that when the last allowed CPU of a NUMA node goes offline for a
3741 * workqueue with a cpumask spanning multiple nodes, the workers which were
3742 * already executing the work items for the workqueue will lose their CPU
3743 * affinity and may execute on any CPU. This is similar to how per-cpu
3744 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3745 * affinity, it's the user's responsibility to flush the work item from
3748 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3751 int node = cpu_to_node(cpu);
3752 int cpu_off = online ? -1 : cpu;
3753 struct pool_workqueue *old_pwq = NULL, *pwq;
3754 struct workqueue_attrs *target_attrs;
3757 lockdep_assert_held(&wq_pool_mutex);
3759 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3760 wq->unbound_attrs->no_numa)
3764 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3765 * Let's use a preallocated one. The following buf is protected by
3766 * CPU hotplug exclusion.
3768 target_attrs = wq_update_unbound_numa_attrs_buf;
3769 cpumask = target_attrs->cpumask;
3771 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3772 pwq = unbound_pwq_by_node(wq, node);
3775 * Let's determine what needs to be done. If the target cpumask is
3776 * different from the default pwq's, we need to compare it to @pwq's
3777 * and create a new one if they don't match. If the target cpumask
3778 * equals the default pwq's, the default pwq should be used.
3780 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3781 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3787 /* create a new pwq */
3788 pwq = alloc_unbound_pwq(wq, target_attrs);
3790 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3795 /* Install the new pwq. */
3796 mutex_lock(&wq->mutex);
3797 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3801 mutex_lock(&wq->mutex);
3802 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3803 get_pwq(wq->dfl_pwq);
3804 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3805 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3807 mutex_unlock(&wq->mutex);
3808 put_pwq_unlocked(old_pwq);
3811 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3813 bool highpri = wq->flags & WQ_HIGHPRI;
3816 if (!(wq->flags & WQ_UNBOUND)) {
3817 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3821 for_each_possible_cpu(cpu) {
3822 struct pool_workqueue *pwq =
3823 per_cpu_ptr(wq->cpu_pwqs, cpu);
3824 struct worker_pool *cpu_pools =
3825 per_cpu(cpu_worker_pools, cpu);
3827 init_pwq(pwq, wq, &cpu_pools[highpri]);
3829 mutex_lock(&wq->mutex);
3831 mutex_unlock(&wq->mutex);
3834 } else if (wq->flags & __WQ_ORDERED) {
3835 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3836 /* there should only be single pwq for ordering guarantee */
3837 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3838 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3839 "ordering guarantee broken for workqueue %s\n", wq->name);
3842 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3846 static int wq_clamp_max_active(int max_active, unsigned int flags,
3849 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3851 if (max_active < 1 || max_active > lim)
3852 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3853 max_active, name, 1, lim);
3855 return clamp_val(max_active, 1, lim);
3858 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3861 struct lock_class_key *key,
3862 const char *lock_name, ...)
3864 size_t tbl_size = 0;
3866 struct workqueue_struct *wq;
3867 struct pool_workqueue *pwq;
3870 * Unbound && max_active == 1 used to imply ordered, which is no
3871 * longer the case on NUMA machines due to per-node pools. While
3872 * alloc_ordered_workqueue() is the right way to create an ordered
3873 * workqueue, keep the previous behavior to avoid subtle breakages
3876 if ((flags & WQ_UNBOUND) && max_active == 1)
3877 flags |= __WQ_ORDERED;
3879 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3880 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3881 flags |= WQ_UNBOUND;
3883 /* allocate wq and format name */
3884 if (flags & WQ_UNBOUND)
3885 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3887 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3891 if (flags & WQ_UNBOUND) {
3892 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3893 if (!wq->unbound_attrs)
3897 va_start(args, lock_name);
3898 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3901 max_active = max_active ?: WQ_DFL_ACTIVE;
3902 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3906 wq->saved_max_active = max_active;
3907 mutex_init(&wq->mutex);
3908 atomic_set(&wq->nr_pwqs_to_flush, 0);
3909 INIT_LIST_HEAD(&wq->pwqs);
3910 INIT_LIST_HEAD(&wq->flusher_queue);
3911 INIT_LIST_HEAD(&wq->flusher_overflow);
3912 INIT_LIST_HEAD(&wq->maydays);
3914 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3915 INIT_LIST_HEAD(&wq->list);
3917 if (alloc_and_link_pwqs(wq) < 0)
3921 * Workqueues which may be used during memory reclaim should
3922 * have a rescuer to guarantee forward progress.
3924 if (flags & WQ_MEM_RECLAIM) {
3925 struct worker *rescuer;
3927 rescuer = alloc_worker(NUMA_NO_NODE);
3931 rescuer->rescue_wq = wq;
3932 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3934 if (IS_ERR(rescuer->task)) {
3939 wq->rescuer = rescuer;
3940 kthread_bind_mask(rescuer->task, cpu_possible_mask);
3941 wake_up_process(rescuer->task);
3944 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3948 * wq_pool_mutex protects global freeze state and workqueues list.
3949 * Grab it, adjust max_active and add the new @wq to workqueues
3952 mutex_lock(&wq_pool_mutex);
3954 mutex_lock(&wq->mutex);
3955 for_each_pwq(pwq, wq)
3956 pwq_adjust_max_active(pwq);
3957 mutex_unlock(&wq->mutex);
3959 list_add_tail_rcu(&wq->list, &workqueues);
3961 mutex_unlock(&wq_pool_mutex);
3966 free_workqueue_attrs(wq->unbound_attrs);
3970 destroy_workqueue(wq);
3973 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3976 * destroy_workqueue - safely terminate a workqueue
3977 * @wq: target workqueue
3979 * Safely destroy a workqueue. All work currently pending will be done first.
3981 void destroy_workqueue(struct workqueue_struct *wq)
3983 struct pool_workqueue *pwq;
3986 /* drain it before proceeding with destruction */
3987 drain_workqueue(wq);
3990 mutex_lock(&wq->mutex);
3991 for_each_pwq(pwq, wq) {
3994 for (i = 0; i < WORK_NR_COLORS; i++) {
3995 if (WARN_ON(pwq->nr_in_flight[i])) {
3996 mutex_unlock(&wq->mutex);
4001 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4002 WARN_ON(pwq->nr_active) ||
4003 WARN_ON(!list_empty(&pwq->delayed_works))) {
4004 mutex_unlock(&wq->mutex);
4008 mutex_unlock(&wq->mutex);
4011 * wq list is used to freeze wq, remove from list after
4012 * flushing is complete in case freeze races us.
4014 mutex_lock(&wq_pool_mutex);
4015 list_del_rcu(&wq->list);
4016 mutex_unlock(&wq_pool_mutex);
4018 workqueue_sysfs_unregister(wq);
4021 kthread_stop(wq->rescuer->task);
4023 if (!(wq->flags & WQ_UNBOUND)) {
4025 * The base ref is never dropped on per-cpu pwqs. Directly
4026 * schedule RCU free.
4028 call_rcu_sched(&wq->rcu, rcu_free_wq);
4031 * We're the sole accessor of @wq at this point. Directly
4032 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4033 * @wq will be freed when the last pwq is released.
4035 for_each_node(node) {
4036 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4037 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4038 put_pwq_unlocked(pwq);
4042 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4043 * put. Don't access it afterwards.
4047 put_pwq_unlocked(pwq);
4050 EXPORT_SYMBOL_GPL(destroy_workqueue);
4053 * workqueue_set_max_active - adjust max_active of a workqueue
4054 * @wq: target workqueue
4055 * @max_active: new max_active value.
4057 * Set max_active of @wq to @max_active.
4060 * Don't call from IRQ context.
4062 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4064 struct pool_workqueue *pwq;
4066 /* disallow meddling with max_active for ordered workqueues */
4067 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4070 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4072 mutex_lock(&wq->mutex);
4074 wq->flags &= ~__WQ_ORDERED;
4075 wq->saved_max_active = max_active;
4077 for_each_pwq(pwq, wq)
4078 pwq_adjust_max_active(pwq);
4080 mutex_unlock(&wq->mutex);
4082 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4085 * current_work - retrieve %current task's work struct
4087 * Determine if %current task is a workqueue worker and what it's working on.
4088 * Useful to find out the context that the %current task is running in.
4090 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4092 struct work_struct *current_work(void)
4094 struct worker *worker = current_wq_worker();
4096 return worker ? worker->current_work : NULL;
4098 EXPORT_SYMBOL(current_work);
4101 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4103 * Determine whether %current is a workqueue rescuer. Can be used from
4104 * work functions to determine whether it's being run off the rescuer task.
4106 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4108 bool current_is_workqueue_rescuer(void)
4110 struct worker *worker = current_wq_worker();
4112 return worker && worker->rescue_wq;
4116 * workqueue_congested - test whether a workqueue is congested
4117 * @cpu: CPU in question
4118 * @wq: target workqueue
4120 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4121 * no synchronization around this function and the test result is
4122 * unreliable and only useful as advisory hints or for debugging.
4124 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4125 * Note that both per-cpu and unbound workqueues may be associated with
4126 * multiple pool_workqueues which have separate congested states. A
4127 * workqueue being congested on one CPU doesn't mean the workqueue is also
4128 * contested on other CPUs / NUMA nodes.
4131 * %true if congested, %false otherwise.
4133 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4135 struct pool_workqueue *pwq;
4138 rcu_read_lock_sched();
4140 if (cpu == WORK_CPU_UNBOUND)
4141 cpu = smp_processor_id();
4143 if (!(wq->flags & WQ_UNBOUND))
4144 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4146 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4148 ret = !list_empty(&pwq->delayed_works);
4149 rcu_read_unlock_sched();
4153 EXPORT_SYMBOL_GPL(workqueue_congested);
4156 * work_busy - test whether a work is currently pending or running
4157 * @work: the work to be tested
4159 * Test whether @work is currently pending or running. There is no
4160 * synchronization around this function and the test result is
4161 * unreliable and only useful as advisory hints or for debugging.
4164 * OR'd bitmask of WORK_BUSY_* bits.
4166 unsigned int work_busy(struct work_struct *work)
4168 struct worker_pool *pool;
4169 unsigned long flags;
4170 unsigned int ret = 0;
4172 if (work_pending(work))
4173 ret |= WORK_BUSY_PENDING;
4175 local_irq_save(flags);
4176 pool = get_work_pool(work);
4178 spin_lock(&pool->lock);
4179 if (find_worker_executing_work(pool, work))
4180 ret |= WORK_BUSY_RUNNING;
4181 spin_unlock(&pool->lock);
4183 local_irq_restore(flags);
4187 EXPORT_SYMBOL_GPL(work_busy);
4190 * set_worker_desc - set description for the current work item
4191 * @fmt: printf-style format string
4192 * @...: arguments for the format string
4194 * This function can be called by a running work function to describe what
4195 * the work item is about. If the worker task gets dumped, this
4196 * information will be printed out together to help debugging. The
4197 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4199 void set_worker_desc(const char *fmt, ...)
4201 struct worker *worker = current_wq_worker();
4205 va_start(args, fmt);
4206 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4208 worker->desc_valid = true;
4213 * print_worker_info - print out worker information and description
4214 * @log_lvl: the log level to use when printing
4215 * @task: target task
4217 * If @task is a worker and currently executing a work item, print out the
4218 * name of the workqueue being serviced and worker description set with
4219 * set_worker_desc() by the currently executing work item.
4221 * This function can be safely called on any task as long as the
4222 * task_struct itself is accessible. While safe, this function isn't
4223 * synchronized and may print out mixups or garbages of limited length.
4225 void print_worker_info(const char *log_lvl, struct task_struct *task)
4227 work_func_t *fn = NULL;
4228 char name[WQ_NAME_LEN] = { };
4229 char desc[WORKER_DESC_LEN] = { };
4230 struct pool_workqueue *pwq = NULL;
4231 struct workqueue_struct *wq = NULL;
4232 bool desc_valid = false;
4233 struct worker *worker;
4235 if (!(task->flags & PF_WQ_WORKER))
4239 * This function is called without any synchronization and @task
4240 * could be in any state. Be careful with dereferences.
4242 worker = probe_kthread_data(task);
4245 * Carefully copy the associated workqueue's workfn and name. Keep
4246 * the original last '\0' in case the original contains garbage.
4248 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4249 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4250 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4251 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4253 /* copy worker description */
4254 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4256 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4258 if (fn || name[0] || desc[0]) {
4259 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4261 pr_cont(" (%s)", desc);
4266 static void pr_cont_pool_info(struct worker_pool *pool)
4268 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4269 if (pool->node != NUMA_NO_NODE)
4270 pr_cont(" node=%d", pool->node);
4271 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4274 static void pr_cont_work(bool comma, struct work_struct *work)
4276 if (work->func == wq_barrier_func) {
4277 struct wq_barrier *barr;
4279 barr = container_of(work, struct wq_barrier, work);
4281 pr_cont("%s BAR(%d)", comma ? "," : "",
4282 task_pid_nr(barr->task));
4284 pr_cont("%s %pf", comma ? "," : "", work->func);
4288 static void show_pwq(struct pool_workqueue *pwq)
4290 struct worker_pool *pool = pwq->pool;
4291 struct work_struct *work;
4292 struct worker *worker;
4293 bool has_in_flight = false, has_pending = false;
4296 pr_info(" pwq %d:", pool->id);
4297 pr_cont_pool_info(pool);
4299 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4300 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4302 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4303 if (worker->current_pwq == pwq) {
4304 has_in_flight = true;
4308 if (has_in_flight) {
4311 pr_info(" in-flight:");
4312 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4313 if (worker->current_pwq != pwq)
4316 pr_cont("%s %d%s:%pf", comma ? "," : "",
4317 task_pid_nr(worker->task),
4318 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4319 worker->current_func);
4320 list_for_each_entry(work, &worker->scheduled, entry)
4321 pr_cont_work(false, work);
4327 list_for_each_entry(work, &pool->worklist, entry) {
4328 if (get_work_pwq(work) == pwq) {
4336 pr_info(" pending:");
4337 list_for_each_entry(work, &pool->worklist, entry) {
4338 if (get_work_pwq(work) != pwq)
4341 pr_cont_work(comma, work);
4342 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4347 if (!list_empty(&pwq->delayed_works)) {
4350 pr_info(" delayed:");
4351 list_for_each_entry(work, &pwq->delayed_works, entry) {
4352 pr_cont_work(comma, work);
4353 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4360 * show_workqueue_state - dump workqueue state
4362 * Called from a sysrq handler and prints out all busy workqueues and
4365 void show_workqueue_state(void)
4367 struct workqueue_struct *wq;
4368 struct worker_pool *pool;
4369 unsigned long flags;
4372 rcu_read_lock_sched();
4374 pr_info("Showing busy workqueues and worker pools:\n");
4376 list_for_each_entry_rcu(wq, &workqueues, list) {
4377 struct pool_workqueue *pwq;
4380 for_each_pwq(pwq, wq) {
4381 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4389 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4391 for_each_pwq(pwq, wq) {
4392 spin_lock_irqsave(&pwq->pool->lock, flags);
4393 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4395 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4399 for_each_pool(pool, pi) {
4400 struct worker *worker;
4403 spin_lock_irqsave(&pool->lock, flags);
4404 if (pool->nr_workers == pool->nr_idle)
4407 pr_info("pool %d:", pool->id);
4408 pr_cont_pool_info(pool);
4409 pr_cont(" hung=%us workers=%d",
4410 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4413 pr_cont(" manager: %d",
4414 task_pid_nr(pool->manager->task));
4415 list_for_each_entry(worker, &pool->idle_list, entry) {
4416 pr_cont(" %s%d", first ? "idle: " : "",
4417 task_pid_nr(worker->task));
4422 spin_unlock_irqrestore(&pool->lock, flags);
4425 rcu_read_unlock_sched();
4431 * There are two challenges in supporting CPU hotplug. Firstly, there
4432 * are a lot of assumptions on strong associations among work, pwq and
4433 * pool which make migrating pending and scheduled works very
4434 * difficult to implement without impacting hot paths. Secondly,
4435 * worker pools serve mix of short, long and very long running works making
4436 * blocked draining impractical.
4438 * This is solved by allowing the pools to be disassociated from the CPU
4439 * running as an unbound one and allowing it to be reattached later if the
4440 * cpu comes back online.
4443 static void wq_unbind_fn(struct work_struct *work)
4445 int cpu = smp_processor_id();
4446 struct worker_pool *pool;
4447 struct worker *worker;
4449 for_each_cpu_worker_pool(pool, cpu) {
4450 mutex_lock(&pool->attach_mutex);
4451 spin_lock_irq(&pool->lock);
4454 * We've blocked all attach/detach operations. Make all workers
4455 * unbound and set DISASSOCIATED. Before this, all workers
4456 * except for the ones which are still executing works from
4457 * before the last CPU down must be on the cpu. After
4458 * this, they may become diasporas.
4460 for_each_pool_worker(worker, pool)
4461 worker->flags |= WORKER_UNBOUND;
4463 pool->flags |= POOL_DISASSOCIATED;
4465 spin_unlock_irq(&pool->lock);
4466 mutex_unlock(&pool->attach_mutex);
4469 * Call schedule() so that we cross rq->lock and thus can
4470 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4471 * This is necessary as scheduler callbacks may be invoked
4477 * Sched callbacks are disabled now. Zap nr_running.
4478 * After this, nr_running stays zero and need_more_worker()
4479 * and keep_working() are always true as long as the
4480 * worklist is not empty. This pool now behaves as an
4481 * unbound (in terms of concurrency management) pool which
4482 * are served by workers tied to the pool.
4484 atomic_set(&pool->nr_running, 0);
4487 * With concurrency management just turned off, a busy
4488 * worker blocking could lead to lengthy stalls. Kick off
4489 * unbound chain execution of currently pending work items.
4491 spin_lock_irq(&pool->lock);
4492 wake_up_worker(pool);
4493 spin_unlock_irq(&pool->lock);
4498 * rebind_workers - rebind all workers of a pool to the associated CPU
4499 * @pool: pool of interest
4501 * @pool->cpu is coming online. Rebind all workers to the CPU.
4503 static void rebind_workers(struct worker_pool *pool)
4505 struct worker *worker;
4507 lockdep_assert_held(&pool->attach_mutex);
4510 * Restore CPU affinity of all workers. As all idle workers should
4511 * be on the run-queue of the associated CPU before any local
4512 * wake-ups for concurrency management happen, restore CPU affinity
4513 * of all workers first and then clear UNBOUND. As we're called
4514 * from CPU_ONLINE, the following shouldn't fail.
4516 for_each_pool_worker(worker, pool)
4517 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4518 pool->attrs->cpumask) < 0);
4520 spin_lock_irq(&pool->lock);
4523 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4524 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4525 * being reworked and this can go away in time.
4527 if (!(pool->flags & POOL_DISASSOCIATED)) {
4528 spin_unlock_irq(&pool->lock);
4532 pool->flags &= ~POOL_DISASSOCIATED;
4534 for_each_pool_worker(worker, pool) {
4535 unsigned int worker_flags = worker->flags;
4538 * A bound idle worker should actually be on the runqueue
4539 * of the associated CPU for local wake-ups targeting it to
4540 * work. Kick all idle workers so that they migrate to the
4541 * associated CPU. Doing this in the same loop as
4542 * replacing UNBOUND with REBOUND is safe as no worker will
4543 * be bound before @pool->lock is released.
4545 if (worker_flags & WORKER_IDLE)
4546 wake_up_process(worker->task);
4549 * We want to clear UNBOUND but can't directly call
4550 * worker_clr_flags() or adjust nr_running. Atomically
4551 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4552 * @worker will clear REBOUND using worker_clr_flags() when
4553 * it initiates the next execution cycle thus restoring
4554 * concurrency management. Note that when or whether
4555 * @worker clears REBOUND doesn't affect correctness.
4557 * ACCESS_ONCE() is necessary because @worker->flags may be
4558 * tested without holding any lock in
4559 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4560 * fail incorrectly leading to premature concurrency
4561 * management operations.
4563 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4564 worker_flags |= WORKER_REBOUND;
4565 worker_flags &= ~WORKER_UNBOUND;
4566 ACCESS_ONCE(worker->flags) = worker_flags;
4569 spin_unlock_irq(&pool->lock);
4573 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4574 * @pool: unbound pool of interest
4575 * @cpu: the CPU which is coming up
4577 * An unbound pool may end up with a cpumask which doesn't have any online
4578 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4579 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4580 * online CPU before, cpus_allowed of all its workers should be restored.
4582 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4584 static cpumask_t cpumask;
4585 struct worker *worker;
4587 lockdep_assert_held(&pool->attach_mutex);
4589 /* is @cpu allowed for @pool? */
4590 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4593 /* is @cpu the only online CPU? */
4594 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4595 if (cpumask_weight(&cpumask) != 1)
4598 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4599 for_each_pool_worker(worker, pool)
4600 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4601 pool->attrs->cpumask) < 0);
4605 * Workqueues should be brought up before normal priority CPU notifiers.
4606 * This will be registered high priority CPU notifier.
4608 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4609 unsigned long action,
4612 int cpu = (unsigned long)hcpu;
4613 struct worker_pool *pool;
4614 struct workqueue_struct *wq;
4617 switch (action & ~CPU_TASKS_FROZEN) {
4618 case CPU_UP_PREPARE:
4619 for_each_cpu_worker_pool(pool, cpu) {
4620 if (pool->nr_workers)
4622 if (!create_worker(pool))
4627 case CPU_DOWN_FAILED:
4629 mutex_lock(&wq_pool_mutex);
4631 for_each_pool(pool, pi) {
4632 mutex_lock(&pool->attach_mutex);
4634 if (pool->cpu == cpu)
4635 rebind_workers(pool);
4636 else if (pool->cpu < 0)
4637 restore_unbound_workers_cpumask(pool, cpu);
4639 mutex_unlock(&pool->attach_mutex);
4642 /* update NUMA affinity of unbound workqueues */
4643 list_for_each_entry(wq, &workqueues, list)
4644 wq_update_unbound_numa(wq, cpu, true);
4646 mutex_unlock(&wq_pool_mutex);
4653 * Workqueues should be brought down after normal priority CPU notifiers.
4654 * This will be registered as low priority CPU notifier.
4656 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4657 unsigned long action,
4660 int cpu = (unsigned long)hcpu;
4661 struct work_struct unbind_work;
4662 struct workqueue_struct *wq;
4664 switch (action & ~CPU_TASKS_FROZEN) {
4665 case CPU_DOWN_PREPARE:
4666 /* unbinding per-cpu workers should happen on the local CPU */
4667 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4668 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4670 /* update NUMA affinity of unbound workqueues */
4671 mutex_lock(&wq_pool_mutex);
4672 list_for_each_entry(wq, &workqueues, list)
4673 wq_update_unbound_numa(wq, cpu, false);
4674 mutex_unlock(&wq_pool_mutex);
4676 /* wait for per-cpu unbinding to finish */
4677 flush_work(&unbind_work);
4678 destroy_work_on_stack(&unbind_work);
4686 struct work_for_cpu {
4687 struct work_struct work;
4693 static void work_for_cpu_fn(struct work_struct *work)
4695 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4697 wfc->ret = wfc->fn(wfc->arg);
4701 * work_on_cpu - run a function in user context on a particular cpu
4702 * @cpu: the cpu to run on
4703 * @fn: the function to run
4704 * @arg: the function arg
4706 * It is up to the caller to ensure that the cpu doesn't go offline.
4707 * The caller must not hold any locks which would prevent @fn from completing.
4709 * Return: The value @fn returns.
4711 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4713 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4715 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4716 schedule_work_on(cpu, &wfc.work);
4717 flush_work(&wfc.work);
4718 destroy_work_on_stack(&wfc.work);
4721 EXPORT_SYMBOL_GPL(work_on_cpu);
4722 #endif /* CONFIG_SMP */
4724 #ifdef CONFIG_FREEZER
4727 * freeze_workqueues_begin - begin freezing workqueues
4729 * Start freezing workqueues. After this function returns, all freezable
4730 * workqueues will queue new works to their delayed_works list instead of
4734 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4736 void freeze_workqueues_begin(void)
4738 struct workqueue_struct *wq;
4739 struct pool_workqueue *pwq;
4741 mutex_lock(&wq_pool_mutex);
4743 WARN_ON_ONCE(workqueue_freezing);
4744 workqueue_freezing = true;
4746 list_for_each_entry(wq, &workqueues, list) {
4747 mutex_lock(&wq->mutex);
4748 for_each_pwq(pwq, wq)
4749 pwq_adjust_max_active(pwq);
4750 mutex_unlock(&wq->mutex);
4753 mutex_unlock(&wq_pool_mutex);
4757 * freeze_workqueues_busy - are freezable workqueues still busy?
4759 * Check whether freezing is complete. This function must be called
4760 * between freeze_workqueues_begin() and thaw_workqueues().
4763 * Grabs and releases wq_pool_mutex.
4766 * %true if some freezable workqueues are still busy. %false if freezing
4769 bool freeze_workqueues_busy(void)
4772 struct workqueue_struct *wq;
4773 struct pool_workqueue *pwq;
4775 mutex_lock(&wq_pool_mutex);
4777 WARN_ON_ONCE(!workqueue_freezing);
4779 list_for_each_entry(wq, &workqueues, list) {
4780 if (!(wq->flags & WQ_FREEZABLE))
4783 * nr_active is monotonically decreasing. It's safe
4784 * to peek without lock.
4786 rcu_read_lock_sched();
4787 for_each_pwq(pwq, wq) {
4788 WARN_ON_ONCE(pwq->nr_active < 0);
4789 if (pwq->nr_active) {
4791 rcu_read_unlock_sched();
4795 rcu_read_unlock_sched();
4798 mutex_unlock(&wq_pool_mutex);
4803 * thaw_workqueues - thaw workqueues
4805 * Thaw workqueues. Normal queueing is restored and all collected
4806 * frozen works are transferred to their respective pool worklists.
4809 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4811 void thaw_workqueues(void)
4813 struct workqueue_struct *wq;
4814 struct pool_workqueue *pwq;
4816 mutex_lock(&wq_pool_mutex);
4818 if (!workqueue_freezing)
4821 workqueue_freezing = false;
4823 /* restore max_active and repopulate worklist */
4824 list_for_each_entry(wq, &workqueues, list) {
4825 mutex_lock(&wq->mutex);
4826 for_each_pwq(pwq, wq)
4827 pwq_adjust_max_active(pwq);
4828 mutex_unlock(&wq->mutex);
4832 mutex_unlock(&wq_pool_mutex);
4834 #endif /* CONFIG_FREEZER */
4836 static int workqueue_apply_unbound_cpumask(void)
4840 struct workqueue_struct *wq;
4841 struct apply_wqattrs_ctx *ctx, *n;
4843 lockdep_assert_held(&wq_pool_mutex);
4845 list_for_each_entry(wq, &workqueues, list) {
4846 if (!(wq->flags & WQ_UNBOUND))
4848 /* creating multiple pwqs breaks ordering guarantee */
4849 if (wq->flags & __WQ_ORDERED)
4852 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4858 list_add_tail(&ctx->list, &ctxs);
4861 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4863 apply_wqattrs_commit(ctx);
4864 apply_wqattrs_cleanup(ctx);
4871 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4872 * @cpumask: the cpumask to set
4874 * The low-level workqueues cpumask is a global cpumask that limits
4875 * the affinity of all unbound workqueues. This function check the @cpumask
4876 * and apply it to all unbound workqueues and updates all pwqs of them.
4878 * Retun: 0 - Success
4879 * -EINVAL - Invalid @cpumask
4880 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4882 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4885 cpumask_var_t saved_cpumask;
4887 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4890 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4891 if (!cpumask_empty(cpumask)) {
4892 apply_wqattrs_lock();
4894 /* save the old wq_unbound_cpumask. */
4895 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4897 /* update wq_unbound_cpumask at first and apply it to wqs. */
4898 cpumask_copy(wq_unbound_cpumask, cpumask);
4899 ret = workqueue_apply_unbound_cpumask();
4901 /* restore the wq_unbound_cpumask when failed. */
4903 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4905 apply_wqattrs_unlock();
4908 free_cpumask_var(saved_cpumask);
4914 * Workqueues with WQ_SYSFS flag set is visible to userland via
4915 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4916 * following attributes.
4918 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4919 * max_active RW int : maximum number of in-flight work items
4921 * Unbound workqueues have the following extra attributes.
4923 * id RO int : the associated pool ID
4924 * nice RW int : nice value of the workers
4925 * cpumask RW mask : bitmask of allowed CPUs for the workers
4928 struct workqueue_struct *wq;
4932 static struct workqueue_struct *dev_to_wq(struct device *dev)
4934 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4939 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4942 struct workqueue_struct *wq = dev_to_wq(dev);
4944 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4946 static DEVICE_ATTR_RO(per_cpu);
4948 static ssize_t max_active_show(struct device *dev,
4949 struct device_attribute *attr, char *buf)
4951 struct workqueue_struct *wq = dev_to_wq(dev);
4953 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4956 static ssize_t max_active_store(struct device *dev,
4957 struct device_attribute *attr, const char *buf,
4960 struct workqueue_struct *wq = dev_to_wq(dev);
4963 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4966 workqueue_set_max_active(wq, val);
4969 static DEVICE_ATTR_RW(max_active);
4971 static struct attribute *wq_sysfs_attrs[] = {
4972 &dev_attr_per_cpu.attr,
4973 &dev_attr_max_active.attr,
4976 ATTRIBUTE_GROUPS(wq_sysfs);
4978 static ssize_t wq_pool_ids_show(struct device *dev,
4979 struct device_attribute *attr, char *buf)
4981 struct workqueue_struct *wq = dev_to_wq(dev);
4982 const char *delim = "";
4983 int node, written = 0;
4985 rcu_read_lock_sched();
4986 for_each_node(node) {
4987 written += scnprintf(buf + written, PAGE_SIZE - written,
4988 "%s%d:%d", delim, node,
4989 unbound_pwq_by_node(wq, node)->pool->id);
4992 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4993 rcu_read_unlock_sched();
4998 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5001 struct workqueue_struct *wq = dev_to_wq(dev);
5004 mutex_lock(&wq->mutex);
5005 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5006 mutex_unlock(&wq->mutex);
5011 /* prepare workqueue_attrs for sysfs store operations */
5012 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5014 struct workqueue_attrs *attrs;
5016 lockdep_assert_held(&wq_pool_mutex);
5018 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5022 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5026 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5027 const char *buf, size_t count)
5029 struct workqueue_struct *wq = dev_to_wq(dev);
5030 struct workqueue_attrs *attrs;
5033 apply_wqattrs_lock();
5035 attrs = wq_sysfs_prep_attrs(wq);
5039 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5040 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5041 ret = apply_workqueue_attrs_locked(wq, attrs);
5046 apply_wqattrs_unlock();
5047 free_workqueue_attrs(attrs);
5048 return ret ?: count;
5051 static ssize_t wq_cpumask_show(struct device *dev,
5052 struct device_attribute *attr, char *buf)
5054 struct workqueue_struct *wq = dev_to_wq(dev);
5057 mutex_lock(&wq->mutex);
5058 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5059 cpumask_pr_args(wq->unbound_attrs->cpumask));
5060 mutex_unlock(&wq->mutex);
5064 static ssize_t wq_cpumask_store(struct device *dev,
5065 struct device_attribute *attr,
5066 const char *buf, size_t count)
5068 struct workqueue_struct *wq = dev_to_wq(dev);
5069 struct workqueue_attrs *attrs;
5072 apply_wqattrs_lock();
5074 attrs = wq_sysfs_prep_attrs(wq);
5078 ret = cpumask_parse(buf, attrs->cpumask);
5080 ret = apply_workqueue_attrs_locked(wq, attrs);
5083 apply_wqattrs_unlock();
5084 free_workqueue_attrs(attrs);
5085 return ret ?: count;
5088 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5091 struct workqueue_struct *wq = dev_to_wq(dev);
5094 mutex_lock(&wq->mutex);
5095 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5096 !wq->unbound_attrs->no_numa);
5097 mutex_unlock(&wq->mutex);
5102 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5103 const char *buf, size_t count)
5105 struct workqueue_struct *wq = dev_to_wq(dev);
5106 struct workqueue_attrs *attrs;
5107 int v, ret = -ENOMEM;
5109 apply_wqattrs_lock();
5111 attrs = wq_sysfs_prep_attrs(wq);
5116 if (sscanf(buf, "%d", &v) == 1) {
5117 attrs->no_numa = !v;
5118 ret = apply_workqueue_attrs_locked(wq, attrs);
5122 apply_wqattrs_unlock();
5123 free_workqueue_attrs(attrs);
5124 return ret ?: count;
5127 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5128 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5129 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5130 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5131 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5135 static struct bus_type wq_subsys = {
5136 .name = "workqueue",
5137 .dev_groups = wq_sysfs_groups,
5140 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5141 struct device_attribute *attr, char *buf)
5145 mutex_lock(&wq_pool_mutex);
5146 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5147 cpumask_pr_args(wq_unbound_cpumask));
5148 mutex_unlock(&wq_pool_mutex);
5153 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5154 struct device_attribute *attr, const char *buf, size_t count)
5156 cpumask_var_t cpumask;
5159 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5162 ret = cpumask_parse(buf, cpumask);
5164 ret = workqueue_set_unbound_cpumask(cpumask);
5166 free_cpumask_var(cpumask);
5167 return ret ? ret : count;
5170 static struct device_attribute wq_sysfs_cpumask_attr =
5171 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5172 wq_unbound_cpumask_store);
5174 static int __init wq_sysfs_init(void)
5178 err = subsys_virtual_register(&wq_subsys, NULL);
5182 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5184 core_initcall(wq_sysfs_init);
5186 static void wq_device_release(struct device *dev)
5188 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5194 * workqueue_sysfs_register - make a workqueue visible in sysfs
5195 * @wq: the workqueue to register
5197 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5198 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5199 * which is the preferred method.
5201 * Workqueue user should use this function directly iff it wants to apply
5202 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5203 * apply_workqueue_attrs() may race against userland updating the
5206 * Return: 0 on success, -errno on failure.
5208 int workqueue_sysfs_register(struct workqueue_struct *wq)
5210 struct wq_device *wq_dev;
5214 * Adjusting max_active or creating new pwqs by applying
5215 * attributes breaks ordering guarantee. Disallow exposing ordered
5218 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5221 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5226 wq_dev->dev.bus = &wq_subsys;
5227 wq_dev->dev.init_name = wq->name;
5228 wq_dev->dev.release = wq_device_release;
5231 * unbound_attrs are created separately. Suppress uevent until
5232 * everything is ready.
5234 dev_set_uevent_suppress(&wq_dev->dev, true);
5236 ret = device_register(&wq_dev->dev);
5238 put_device(&wq_dev->dev);
5243 if (wq->flags & WQ_UNBOUND) {
5244 struct device_attribute *attr;
5246 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5247 ret = device_create_file(&wq_dev->dev, attr);
5249 device_unregister(&wq_dev->dev);
5256 dev_set_uevent_suppress(&wq_dev->dev, false);
5257 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5262 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5263 * @wq: the workqueue to unregister
5265 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5267 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5269 struct wq_device *wq_dev = wq->wq_dev;
5275 device_unregister(&wq_dev->dev);
5277 #else /* CONFIG_SYSFS */
5278 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5279 #endif /* CONFIG_SYSFS */
5282 * Workqueue watchdog.
5284 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5285 * flush dependency, a concurrency managed work item which stays RUNNING
5286 * indefinitely. Workqueue stalls can be very difficult to debug as the
5287 * usual warning mechanisms don't trigger and internal workqueue state is
5290 * Workqueue watchdog monitors all worker pools periodically and dumps
5291 * state if some pools failed to make forward progress for a while where
5292 * forward progress is defined as the first item on ->worklist changing.
5294 * This mechanism is controlled through the kernel parameter
5295 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5296 * corresponding sysfs parameter file.
5298 #ifdef CONFIG_WQ_WATCHDOG
5300 static void wq_watchdog_timer_fn(unsigned long data);
5302 static unsigned long wq_watchdog_thresh = 30;
5303 static struct timer_list wq_watchdog_timer =
5304 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5306 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5307 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5309 static void wq_watchdog_reset_touched(void)
5313 wq_watchdog_touched = jiffies;
5314 for_each_possible_cpu(cpu)
5315 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5318 static void wq_watchdog_timer_fn(unsigned long data)
5320 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5321 bool lockup_detected = false;
5322 struct worker_pool *pool;
5330 for_each_pool(pool, pi) {
5331 unsigned long pool_ts, touched, ts;
5333 if (list_empty(&pool->worklist))
5336 /* get the latest of pool and touched timestamps */
5337 pool_ts = READ_ONCE(pool->watchdog_ts);
5338 touched = READ_ONCE(wq_watchdog_touched);
5340 if (time_after(pool_ts, touched))
5345 if (pool->cpu >= 0) {
5346 unsigned long cpu_touched =
5347 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5349 if (time_after(cpu_touched, ts))
5354 if (time_after(jiffies, ts + thresh)) {
5355 lockup_detected = true;
5356 pr_emerg("BUG: workqueue lockup - pool");
5357 pr_cont_pool_info(pool);
5358 pr_cont(" stuck for %us!\n",
5359 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5365 if (lockup_detected)
5366 show_workqueue_state();
5368 wq_watchdog_reset_touched();
5369 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5372 void wq_watchdog_touch(int cpu)
5375 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5377 wq_watchdog_touched = jiffies;
5380 static void wq_watchdog_set_thresh(unsigned long thresh)
5382 wq_watchdog_thresh = 0;
5383 del_timer_sync(&wq_watchdog_timer);
5386 wq_watchdog_thresh = thresh;
5387 wq_watchdog_reset_touched();
5388 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5392 static int wq_watchdog_param_set_thresh(const char *val,
5393 const struct kernel_param *kp)
5395 unsigned long thresh;
5398 ret = kstrtoul(val, 0, &thresh);
5403 wq_watchdog_set_thresh(thresh);
5405 wq_watchdog_thresh = thresh;
5410 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5411 .set = wq_watchdog_param_set_thresh,
5412 .get = param_get_ulong,
5415 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5418 static void wq_watchdog_init(void)
5420 wq_watchdog_set_thresh(wq_watchdog_thresh);
5423 #else /* CONFIG_WQ_WATCHDOG */
5425 static inline void wq_watchdog_init(void) { }
5427 #endif /* CONFIG_WQ_WATCHDOG */
5429 static void __init wq_numa_init(void)
5434 if (num_possible_nodes() <= 1)
5437 if (wq_disable_numa) {
5438 pr_info("workqueue: NUMA affinity support disabled\n");
5442 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5443 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5446 * We want masks of possible CPUs of each node which isn't readily
5447 * available. Build one from cpu_to_node() which should have been
5448 * fully initialized by now.
5450 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5454 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5455 node_online(node) ? node : NUMA_NO_NODE));
5457 for_each_possible_cpu(cpu) {
5458 node = cpu_to_node(cpu);
5459 if (WARN_ON(node == NUMA_NO_NODE)) {
5460 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5461 /* happens iff arch is bonkers, let's just proceed */
5464 cpumask_set_cpu(cpu, tbl[node]);
5467 wq_numa_possible_cpumask = tbl;
5468 wq_numa_enabled = true;
5471 static int __init init_workqueues(void)
5473 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5476 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5478 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5479 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5481 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5483 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5484 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5488 /* initialize CPU pools */
5489 for_each_possible_cpu(cpu) {
5490 struct worker_pool *pool;
5493 for_each_cpu_worker_pool(pool, cpu) {
5494 BUG_ON(init_worker_pool(pool));
5496 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5497 pool->attrs->nice = std_nice[i++];
5498 pool->node = cpu_to_node(cpu);
5501 mutex_lock(&wq_pool_mutex);
5502 BUG_ON(worker_pool_assign_id(pool));
5503 mutex_unlock(&wq_pool_mutex);
5507 /* create the initial worker */
5508 for_each_online_cpu(cpu) {
5509 struct worker_pool *pool;
5511 for_each_cpu_worker_pool(pool, cpu) {
5512 pool->flags &= ~POOL_DISASSOCIATED;
5513 BUG_ON(!create_worker(pool));
5517 /* create default unbound and ordered wq attrs */
5518 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5519 struct workqueue_attrs *attrs;
5521 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5522 attrs->nice = std_nice[i];
5523 unbound_std_wq_attrs[i] = attrs;
5526 * An ordered wq should have only one pwq as ordering is
5527 * guaranteed by max_active which is enforced by pwqs.
5528 * Turn off NUMA so that dfl_pwq is used for all nodes.
5530 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5531 attrs->nice = std_nice[i];
5532 attrs->no_numa = true;
5533 ordered_wq_attrs[i] = attrs;
5536 system_wq = alloc_workqueue("events", 0, 0);
5537 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5538 system_long_wq = alloc_workqueue("events_long", 0, 0);
5539 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5540 WQ_UNBOUND_MAX_ACTIVE);
5541 system_freezable_wq = alloc_workqueue("events_freezable",
5543 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5544 WQ_POWER_EFFICIENT, 0);
5545 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5546 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5548 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5549 !system_unbound_wq || !system_freezable_wq ||
5550 !system_power_efficient_wq ||
5551 !system_freezable_power_efficient_wq);
5557 early_initcall(init_workqueues);