2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
24 * Please read Documentation/workqueue.txt for details.
27 #include <linux/export.h>
28 #include <linux/kernel.h>
29 #include <linux/sched.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/completion.h>
33 #include <linux/workqueue.h>
34 #include <linux/slab.h>
35 #include <linux/cpu.h>
36 #include <linux/notifier.h>
37 #include <linux/kthread.h>
38 #include <linux/hardirq.h>
39 #include <linux/mempolicy.h>
40 #include <linux/freezer.h>
41 #include <linux/kallsyms.h>
42 #include <linux/debug_locks.h>
43 #include <linux/lockdep.h>
44 #include <linux/idr.h>
45 #include <linux/jhash.h>
46 #include <linux/hashtable.h>
47 #include <linux/rculist.h>
48 #include <linux/nodemask.h>
49 #include <linux/moduleparam.h>
50 #include <linux/uaccess.h>
52 #include "workqueue_internal.h"
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_REBOUND = 1 << 8, /* worker was rebound */
81 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
84 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
86 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
89 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
102 RESCUER_NICE_LEVEL = MIN_NICE,
103 HIGHPRI_NICE_LEVEL = MIN_NICE,
109 * Structure fields follow one of the following exclusion rules.
111 * I: Modifiable by initialization/destruction paths and read-only for
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
117 * L: pool->lock protected. Access with pool->lock held.
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
124 * A: pool->attach_mutex protected.
126 * PL: wq_pool_mutex protected.
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
130 * WQ: wq->mutex protected.
132 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
134 * MD: wq_mayday_lock protected.
137 /* struct worker is defined in workqueue_internal.h */
140 spinlock_t lock; /* the pool lock */
141 int cpu; /* I: the associated cpu */
142 int node; /* I: the associated node ID */
143 int id; /* I: pool ID */
144 unsigned int flags; /* X: flags */
146 struct list_head worklist; /* L: list of pending works */
147 int nr_workers; /* L: total number of workers */
149 /* nr_idle includes the ones off idle_list for rebinding */
150 int nr_idle; /* L: currently idle ones */
152 struct list_head idle_list; /* X: list of idle workers */
153 struct timer_list idle_timer; /* L: worker idle timeout */
154 struct timer_list mayday_timer; /* L: SOS timer for workers */
156 /* a workers is either on busy_hash or idle_list, or the manager */
157 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
158 /* L: hash of busy workers */
160 /* see manage_workers() for details on the two manager mutexes */
161 struct mutex manager_arb; /* manager arbitration */
162 struct worker *manager; /* L: purely informational */
163 struct mutex attach_mutex; /* attach/detach exclusion */
164 struct list_head workers; /* A: attached workers */
165 struct completion *detach_completion; /* all workers detached */
167 struct ida worker_ida; /* worker IDs for task name */
169 struct workqueue_attrs *attrs; /* I: worker attributes */
170 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
171 int refcnt; /* PL: refcnt for unbound pools */
174 * The current concurrency level. As it's likely to be accessed
175 * from other CPUs during try_to_wake_up(), put it in a separate
178 atomic_t nr_running ____cacheline_aligned_in_smp;
181 * Destruction of pool is sched-RCU protected to allow dereferences
182 * from get_work_pool().
185 } ____cacheline_aligned_in_smp;
188 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
189 * of work_struct->data are used for flags and the remaining high bits
190 * point to the pwq; thus, pwqs need to be aligned at two's power of the
191 * number of flag bits.
193 struct pool_workqueue {
194 struct worker_pool *pool; /* I: the associated pool */
195 struct workqueue_struct *wq; /* I: the owning workqueue */
196 int work_color; /* L: current color */
197 int flush_color; /* L: flushing color */
198 int refcnt; /* L: reference count */
199 int nr_in_flight[WORK_NR_COLORS];
200 /* L: nr of in_flight works */
201 int nr_active; /* L: nr of active works */
202 int max_active; /* L: max active works */
203 struct list_head delayed_works; /* L: delayed works */
204 struct list_head pwqs_node; /* WR: node on wq->pwqs */
205 struct list_head mayday_node; /* MD: node on wq->maydays */
208 * Release of unbound pwq is punted to system_wq. See put_pwq()
209 * and pwq_unbound_release_workfn() for details. pool_workqueue
210 * itself is also sched-RCU protected so that the first pwq can be
211 * determined without grabbing wq->mutex.
213 struct work_struct unbound_release_work;
215 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
218 * Structure used to wait for workqueue flush.
221 struct list_head list; /* WQ: list of flushers */
222 int flush_color; /* WQ: flush color waiting for */
223 struct completion done; /* flush completion */
229 * The externally visible workqueue. It relays the issued work items to
230 * the appropriate worker_pool through its pool_workqueues.
232 struct workqueue_struct {
233 struct list_head pwqs; /* WR: all pwqs of this wq */
234 struct list_head list; /* PR: list of all workqueues */
236 struct mutex mutex; /* protects this wq */
237 int work_color; /* WQ: current work color */
238 int flush_color; /* WQ: current flush color */
239 atomic_t nr_pwqs_to_flush; /* flush in progress */
240 struct wq_flusher *first_flusher; /* WQ: first flusher */
241 struct list_head flusher_queue; /* WQ: flush waiters */
242 struct list_head flusher_overflow; /* WQ: flush overflow list */
244 struct list_head maydays; /* MD: pwqs requesting rescue */
245 struct worker *rescuer; /* I: rescue worker */
247 int nr_drainers; /* WQ: drain in progress */
248 int saved_max_active; /* WQ: saved pwq max_active */
250 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
251 struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
254 struct wq_device *wq_dev; /* I: for sysfs interface */
256 #ifdef CONFIG_LOCKDEP
257 struct lockdep_map lockdep_map;
259 char name[WQ_NAME_LEN]; /* I: workqueue name */
262 * Destruction of workqueue_struct is sched-RCU protected to allow
263 * walking the workqueues list without grabbing wq_pool_mutex.
264 * This is used to dump all workqueues from sysrq.
268 /* hot fields used during command issue, aligned to cacheline */
269 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
270 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
271 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
274 static struct kmem_cache *pwq_cache;
276 static cpumask_var_t *wq_numa_possible_cpumask;
277 /* possible CPUs of each node */
279 static bool wq_disable_numa;
280 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
282 /* see the comment above the definition of WQ_POWER_EFFICIENT */
283 #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
284 static bool wq_power_efficient = true;
286 static bool wq_power_efficient;
289 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
291 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
293 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
294 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
296 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
297 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
299 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
300 static bool workqueue_freezing; /* PL: have wqs started freezing? */
302 static cpumask_var_t wq_unbound_cpumask; /* PL: low level cpumask for all unbound wqs */
304 /* the per-cpu worker pools */
305 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
308 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
310 /* PL: hash of all unbound pools keyed by pool->attrs */
311 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
313 /* I: attributes used when instantiating standard unbound pools on demand */
314 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
316 /* I: attributes used when instantiating ordered pools on demand */
317 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
319 struct workqueue_struct *system_wq __read_mostly;
320 EXPORT_SYMBOL(system_wq);
321 struct workqueue_struct *system_highpri_wq __read_mostly;
322 EXPORT_SYMBOL_GPL(system_highpri_wq);
323 struct workqueue_struct *system_long_wq __read_mostly;
324 EXPORT_SYMBOL_GPL(system_long_wq);
325 struct workqueue_struct *system_unbound_wq __read_mostly;
326 EXPORT_SYMBOL_GPL(system_unbound_wq);
327 struct workqueue_struct *system_freezable_wq __read_mostly;
328 EXPORT_SYMBOL_GPL(system_freezable_wq);
329 struct workqueue_struct *system_power_efficient_wq __read_mostly;
330 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
331 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
332 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
334 static int worker_thread(void *__worker);
335 static void copy_workqueue_attrs(struct workqueue_attrs *to,
336 const struct workqueue_attrs *from);
337 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
339 #define CREATE_TRACE_POINTS
340 #include <trace/events/workqueue.h>
342 #define assert_rcu_or_pool_mutex() \
343 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
344 lockdep_is_held(&wq_pool_mutex), \
345 "sched RCU or wq_pool_mutex should be held")
347 #define assert_rcu_or_wq_mutex(wq) \
348 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
349 lockdep_is_held(&wq->mutex), \
350 "sched RCU or wq->mutex should be held")
352 #define for_each_cpu_worker_pool(pool, cpu) \
353 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
354 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
358 * for_each_pool - iterate through all worker_pools in the system
359 * @pool: iteration cursor
360 * @pi: integer used for iteration
362 * This must be called either with wq_pool_mutex held or sched RCU read
363 * locked. If the pool needs to be used beyond the locking in effect, the
364 * caller is responsible for guaranteeing that the pool stays online.
366 * The if/else clause exists only for the lockdep assertion and can be
369 #define for_each_pool(pool, pi) \
370 idr_for_each_entry(&worker_pool_idr, pool, pi) \
371 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
375 * for_each_pool_worker - iterate through all workers of a worker_pool
376 * @worker: iteration cursor
377 * @pool: worker_pool to iterate workers of
379 * This must be called with @pool->attach_mutex.
381 * The if/else clause exists only for the lockdep assertion and can be
384 #define for_each_pool_worker(worker, pool) \
385 list_for_each_entry((worker), &(pool)->workers, node) \
386 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
390 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
391 * @pwq: iteration cursor
392 * @wq: the target workqueue
394 * This must be called either with wq->mutex held or sched RCU read locked.
395 * If the pwq needs to be used beyond the locking in effect, the caller is
396 * responsible for guaranteeing that the pwq stays online.
398 * The if/else clause exists only for the lockdep assertion and can be
401 #define for_each_pwq(pwq, wq) \
402 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
403 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
406 #ifdef CONFIG_DEBUG_OBJECTS_WORK
408 static struct debug_obj_descr work_debug_descr;
410 static void *work_debug_hint(void *addr)
412 return ((struct work_struct *) addr)->func;
416 * fixup_init is called when:
417 * - an active object is initialized
419 static int work_fixup_init(void *addr, enum debug_obj_state state)
421 struct work_struct *work = addr;
424 case ODEBUG_STATE_ACTIVE:
425 cancel_work_sync(work);
426 debug_object_init(work, &work_debug_descr);
434 * fixup_activate is called when:
435 * - an active object is activated
436 * - an unknown object is activated (might be a statically initialized object)
438 static int work_fixup_activate(void *addr, enum debug_obj_state state)
440 struct work_struct *work = addr;
444 case ODEBUG_STATE_NOTAVAILABLE:
446 * This is not really a fixup. The work struct was
447 * statically initialized. We just make sure that it
448 * is tracked in the object tracker.
450 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
451 debug_object_init(work, &work_debug_descr);
452 debug_object_activate(work, &work_debug_descr);
458 case ODEBUG_STATE_ACTIVE:
467 * fixup_free is called when:
468 * - an active object is freed
470 static int work_fixup_free(void *addr, enum debug_obj_state state)
472 struct work_struct *work = addr;
475 case ODEBUG_STATE_ACTIVE:
476 cancel_work_sync(work);
477 debug_object_free(work, &work_debug_descr);
484 static struct debug_obj_descr work_debug_descr = {
485 .name = "work_struct",
486 .debug_hint = work_debug_hint,
487 .fixup_init = work_fixup_init,
488 .fixup_activate = work_fixup_activate,
489 .fixup_free = work_fixup_free,
492 static inline void debug_work_activate(struct work_struct *work)
494 debug_object_activate(work, &work_debug_descr);
497 static inline void debug_work_deactivate(struct work_struct *work)
499 debug_object_deactivate(work, &work_debug_descr);
502 void __init_work(struct work_struct *work, int onstack)
505 debug_object_init_on_stack(work, &work_debug_descr);
507 debug_object_init(work, &work_debug_descr);
509 EXPORT_SYMBOL_GPL(__init_work);
511 void destroy_work_on_stack(struct work_struct *work)
513 debug_object_free(work, &work_debug_descr);
515 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
517 void destroy_delayed_work_on_stack(struct delayed_work *work)
519 destroy_timer_on_stack(&work->timer);
520 debug_object_free(&work->work, &work_debug_descr);
522 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
525 static inline void debug_work_activate(struct work_struct *work) { }
526 static inline void debug_work_deactivate(struct work_struct *work) { }
530 * worker_pool_assign_id - allocate ID and assing it to @pool
531 * @pool: the pool pointer of interest
533 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
534 * successfully, -errno on failure.
536 static int worker_pool_assign_id(struct worker_pool *pool)
540 lockdep_assert_held(&wq_pool_mutex);
542 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
552 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
553 * @wq: the target workqueue
556 * This must be called either with pwq_lock held or sched RCU read locked.
557 * If the pwq needs to be used beyond the locking in effect, the caller is
558 * responsible for guaranteeing that the pwq stays online.
560 * Return: The unbound pool_workqueue for @node.
562 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
565 assert_rcu_or_wq_mutex(wq);
566 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
569 static unsigned int work_color_to_flags(int color)
571 return color << WORK_STRUCT_COLOR_SHIFT;
574 static int get_work_color(struct work_struct *work)
576 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
577 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
580 static int work_next_color(int color)
582 return (color + 1) % WORK_NR_COLORS;
586 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
587 * contain the pointer to the queued pwq. Once execution starts, the flag
588 * is cleared and the high bits contain OFFQ flags and pool ID.
590 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
591 * and clear_work_data() can be used to set the pwq, pool or clear
592 * work->data. These functions should only be called while the work is
593 * owned - ie. while the PENDING bit is set.
595 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
596 * corresponding to a work. Pool is available once the work has been
597 * queued anywhere after initialization until it is sync canceled. pwq is
598 * available only while the work item is queued.
600 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
601 * canceled. While being canceled, a work item may have its PENDING set
602 * but stay off timer and worklist for arbitrarily long and nobody should
603 * try to steal the PENDING bit.
605 static inline void set_work_data(struct work_struct *work, unsigned long data,
608 WARN_ON_ONCE(!work_pending(work));
609 atomic_long_set(&work->data, data | flags | work_static(work));
612 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
613 unsigned long extra_flags)
615 set_work_data(work, (unsigned long)pwq,
616 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
619 static void set_work_pool_and_keep_pending(struct work_struct *work,
622 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
623 WORK_STRUCT_PENDING);
626 static void set_work_pool_and_clear_pending(struct work_struct *work,
630 * The following wmb is paired with the implied mb in
631 * test_and_set_bit(PENDING) and ensures all updates to @work made
632 * here are visible to and precede any updates by the next PENDING
636 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
639 static void clear_work_data(struct work_struct *work)
641 smp_wmb(); /* see set_work_pool_and_clear_pending() */
642 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
645 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
647 unsigned long data = atomic_long_read(&work->data);
649 if (data & WORK_STRUCT_PWQ)
650 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
656 * get_work_pool - return the worker_pool a given work was associated with
657 * @work: the work item of interest
659 * Pools are created and destroyed under wq_pool_mutex, and allows read
660 * access under sched-RCU read lock. As such, this function should be
661 * called under wq_pool_mutex or with preemption disabled.
663 * All fields of the returned pool are accessible as long as the above
664 * mentioned locking is in effect. If the returned pool needs to be used
665 * beyond the critical section, the caller is responsible for ensuring the
666 * returned pool is and stays online.
668 * Return: The worker_pool @work was last associated with. %NULL if none.
670 static struct worker_pool *get_work_pool(struct work_struct *work)
672 unsigned long data = atomic_long_read(&work->data);
675 assert_rcu_or_pool_mutex();
677 if (data & WORK_STRUCT_PWQ)
678 return ((struct pool_workqueue *)
679 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
681 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
682 if (pool_id == WORK_OFFQ_POOL_NONE)
685 return idr_find(&worker_pool_idr, pool_id);
689 * get_work_pool_id - return the worker pool ID a given work is associated with
690 * @work: the work item of interest
692 * Return: The worker_pool ID @work was last associated with.
693 * %WORK_OFFQ_POOL_NONE if none.
695 static int get_work_pool_id(struct work_struct *work)
697 unsigned long data = atomic_long_read(&work->data);
699 if (data & WORK_STRUCT_PWQ)
700 return ((struct pool_workqueue *)
701 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
703 return data >> WORK_OFFQ_POOL_SHIFT;
706 static void mark_work_canceling(struct work_struct *work)
708 unsigned long pool_id = get_work_pool_id(work);
710 pool_id <<= WORK_OFFQ_POOL_SHIFT;
711 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
714 static bool work_is_canceling(struct work_struct *work)
716 unsigned long data = atomic_long_read(&work->data);
718 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
722 * Policy functions. These define the policies on how the global worker
723 * pools are managed. Unless noted otherwise, these functions assume that
724 * they're being called with pool->lock held.
727 static bool __need_more_worker(struct worker_pool *pool)
729 return !atomic_read(&pool->nr_running);
733 * Need to wake up a worker? Called from anything but currently
736 * Note that, because unbound workers never contribute to nr_running, this
737 * function will always return %true for unbound pools as long as the
738 * worklist isn't empty.
740 static bool need_more_worker(struct worker_pool *pool)
742 return !list_empty(&pool->worklist) && __need_more_worker(pool);
745 /* Can I start working? Called from busy but !running workers. */
746 static bool may_start_working(struct worker_pool *pool)
748 return pool->nr_idle;
751 /* Do I need to keep working? Called from currently running workers. */
752 static bool keep_working(struct worker_pool *pool)
754 return !list_empty(&pool->worklist) &&
755 atomic_read(&pool->nr_running) <= 1;
758 /* Do we need a new worker? Called from manager. */
759 static bool need_to_create_worker(struct worker_pool *pool)
761 return need_more_worker(pool) && !may_start_working(pool);
764 /* Do we have too many workers and should some go away? */
765 static bool too_many_workers(struct worker_pool *pool)
767 bool managing = mutex_is_locked(&pool->manager_arb);
768 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
769 int nr_busy = pool->nr_workers - nr_idle;
771 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
778 /* Return the first idle worker. Safe with preemption disabled */
779 static struct worker *first_idle_worker(struct worker_pool *pool)
781 if (unlikely(list_empty(&pool->idle_list)))
784 return list_first_entry(&pool->idle_list, struct worker, entry);
788 * wake_up_worker - wake up an idle worker
789 * @pool: worker pool to wake worker from
791 * Wake up the first idle worker of @pool.
794 * spin_lock_irq(pool->lock).
796 static void wake_up_worker(struct worker_pool *pool)
798 struct worker *worker = first_idle_worker(pool);
801 wake_up_process(worker->task);
805 * wq_worker_waking_up - a worker is waking up
806 * @task: task waking up
807 * @cpu: CPU @task is waking up to
809 * This function is called during try_to_wake_up() when a worker is
813 * spin_lock_irq(rq->lock)
815 void wq_worker_waking_up(struct task_struct *task, int cpu)
817 struct worker *worker = kthread_data(task);
819 if (!(worker->flags & WORKER_NOT_RUNNING)) {
820 WARN_ON_ONCE(worker->pool->cpu != cpu);
821 atomic_inc(&worker->pool->nr_running);
826 * wq_worker_sleeping - a worker is going to sleep
827 * @task: task going to sleep
828 * @cpu: CPU in question, must be the current CPU number
830 * This function is called during schedule() when a busy worker is
831 * going to sleep. Worker on the same cpu can be woken up by
832 * returning pointer to its task.
835 * spin_lock_irq(rq->lock)
838 * Worker task on @cpu to wake up, %NULL if none.
840 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
842 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
843 struct worker_pool *pool;
846 * Rescuers, which may not have all the fields set up like normal
847 * workers, also reach here, let's not access anything before
848 * checking NOT_RUNNING.
850 if (worker->flags & WORKER_NOT_RUNNING)
855 /* this can only happen on the local cpu */
856 if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu))
860 * The counterpart of the following dec_and_test, implied mb,
861 * worklist not empty test sequence is in insert_work().
862 * Please read comment there.
864 * NOT_RUNNING is clear. This means that we're bound to and
865 * running on the local cpu w/ rq lock held and preemption
866 * disabled, which in turn means that none else could be
867 * manipulating idle_list, so dereferencing idle_list without pool
870 if (atomic_dec_and_test(&pool->nr_running) &&
871 !list_empty(&pool->worklist))
872 to_wakeup = first_idle_worker(pool);
873 return to_wakeup ? to_wakeup->task : NULL;
877 * worker_set_flags - set worker flags and adjust nr_running accordingly
879 * @flags: flags to set
881 * Set @flags in @worker->flags and adjust nr_running accordingly.
884 * spin_lock_irq(pool->lock)
886 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
888 struct worker_pool *pool = worker->pool;
890 WARN_ON_ONCE(worker->task != current);
892 /* If transitioning into NOT_RUNNING, adjust nr_running. */
893 if ((flags & WORKER_NOT_RUNNING) &&
894 !(worker->flags & WORKER_NOT_RUNNING)) {
895 atomic_dec(&pool->nr_running);
898 worker->flags |= flags;
902 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
904 * @flags: flags to clear
906 * Clear @flags in @worker->flags and adjust nr_running accordingly.
909 * spin_lock_irq(pool->lock)
911 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
913 struct worker_pool *pool = worker->pool;
914 unsigned int oflags = worker->flags;
916 WARN_ON_ONCE(worker->task != current);
918 worker->flags &= ~flags;
921 * If transitioning out of NOT_RUNNING, increment nr_running. Note
922 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
923 * of multiple flags, not a single flag.
925 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
926 if (!(worker->flags & WORKER_NOT_RUNNING))
927 atomic_inc(&pool->nr_running);
931 * find_worker_executing_work - find worker which is executing a work
932 * @pool: pool of interest
933 * @work: work to find worker for
935 * Find a worker which is executing @work on @pool by searching
936 * @pool->busy_hash which is keyed by the address of @work. For a worker
937 * to match, its current execution should match the address of @work and
938 * its work function. This is to avoid unwanted dependency between
939 * unrelated work executions through a work item being recycled while still
942 * This is a bit tricky. A work item may be freed once its execution
943 * starts and nothing prevents the freed area from being recycled for
944 * another work item. If the same work item address ends up being reused
945 * before the original execution finishes, workqueue will identify the
946 * recycled work item as currently executing and make it wait until the
947 * current execution finishes, introducing an unwanted dependency.
949 * This function checks the work item address and work function to avoid
950 * false positives. Note that this isn't complete as one may construct a
951 * work function which can introduce dependency onto itself through a
952 * recycled work item. Well, if somebody wants to shoot oneself in the
953 * foot that badly, there's only so much we can do, and if such deadlock
954 * actually occurs, it should be easy to locate the culprit work function.
957 * spin_lock_irq(pool->lock).
960 * Pointer to worker which is executing @work if found, %NULL
963 static struct worker *find_worker_executing_work(struct worker_pool *pool,
964 struct work_struct *work)
966 struct worker *worker;
968 hash_for_each_possible(pool->busy_hash, worker, hentry,
970 if (worker->current_work == work &&
971 worker->current_func == work->func)
978 * move_linked_works - move linked works to a list
979 * @work: start of series of works to be scheduled
980 * @head: target list to append @work to
981 * @nextp: out paramter for nested worklist walking
983 * Schedule linked works starting from @work to @head. Work series to
984 * be scheduled starts at @work and includes any consecutive work with
985 * WORK_STRUCT_LINKED set in its predecessor.
987 * If @nextp is not NULL, it's updated to point to the next work of
988 * the last scheduled work. This allows move_linked_works() to be
989 * nested inside outer list_for_each_entry_safe().
992 * spin_lock_irq(pool->lock).
994 static void move_linked_works(struct work_struct *work, struct list_head *head,
995 struct work_struct **nextp)
997 struct work_struct *n;
1000 * Linked worklist will always end before the end of the list,
1001 * use NULL for list head.
1003 list_for_each_entry_safe_from(work, n, NULL, entry) {
1004 list_move_tail(&work->entry, head);
1005 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1010 * If we're already inside safe list traversal and have moved
1011 * multiple works to the scheduled queue, the next position
1012 * needs to be updated.
1019 * get_pwq - get an extra reference on the specified pool_workqueue
1020 * @pwq: pool_workqueue to get
1022 * Obtain an extra reference on @pwq. The caller should guarantee that
1023 * @pwq has positive refcnt and be holding the matching pool->lock.
1025 static void get_pwq(struct pool_workqueue *pwq)
1027 lockdep_assert_held(&pwq->pool->lock);
1028 WARN_ON_ONCE(pwq->refcnt <= 0);
1033 * put_pwq - put a pool_workqueue reference
1034 * @pwq: pool_workqueue to put
1036 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1037 * destruction. The caller should be holding the matching pool->lock.
1039 static void put_pwq(struct pool_workqueue *pwq)
1041 lockdep_assert_held(&pwq->pool->lock);
1042 if (likely(--pwq->refcnt))
1044 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1047 * @pwq can't be released under pool->lock, bounce to
1048 * pwq_unbound_release_workfn(). This never recurses on the same
1049 * pool->lock as this path is taken only for unbound workqueues and
1050 * the release work item is scheduled on a per-cpu workqueue. To
1051 * avoid lockdep warning, unbound pool->locks are given lockdep
1052 * subclass of 1 in get_unbound_pool().
1054 schedule_work(&pwq->unbound_release_work);
1058 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1059 * @pwq: pool_workqueue to put (can be %NULL)
1061 * put_pwq() with locking. This function also allows %NULL @pwq.
1063 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1067 * As both pwqs and pools are sched-RCU protected, the
1068 * following lock operations are safe.
1070 spin_lock_irq(&pwq->pool->lock);
1072 spin_unlock_irq(&pwq->pool->lock);
1076 static void pwq_activate_delayed_work(struct work_struct *work)
1078 struct pool_workqueue *pwq = get_work_pwq(work);
1080 trace_workqueue_activate_work(work);
1081 move_linked_works(work, &pwq->pool->worklist, NULL);
1082 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1086 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1088 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1089 struct work_struct, entry);
1091 pwq_activate_delayed_work(work);
1095 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1096 * @pwq: pwq of interest
1097 * @color: color of work which left the queue
1099 * A work either has completed or is removed from pending queue,
1100 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1103 * spin_lock_irq(pool->lock).
1105 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1107 /* uncolored work items don't participate in flushing or nr_active */
1108 if (color == WORK_NO_COLOR)
1111 pwq->nr_in_flight[color]--;
1114 if (!list_empty(&pwq->delayed_works)) {
1115 /* one down, submit a delayed one */
1116 if (pwq->nr_active < pwq->max_active)
1117 pwq_activate_first_delayed(pwq);
1120 /* is flush in progress and are we at the flushing tip? */
1121 if (likely(pwq->flush_color != color))
1124 /* are there still in-flight works? */
1125 if (pwq->nr_in_flight[color])
1128 /* this pwq is done, clear flush_color */
1129 pwq->flush_color = -1;
1132 * If this was the last pwq, wake up the first flusher. It
1133 * will handle the rest.
1135 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1136 complete(&pwq->wq->first_flusher->done);
1142 * try_to_grab_pending - steal work item from worklist and disable irq
1143 * @work: work item to steal
1144 * @is_dwork: @work is a delayed_work
1145 * @flags: place to store irq state
1147 * Try to grab PENDING bit of @work. This function can handle @work in any
1148 * stable state - idle, on timer or on worklist.
1151 * 1 if @work was pending and we successfully stole PENDING
1152 * 0 if @work was idle and we claimed PENDING
1153 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1154 * -ENOENT if someone else is canceling @work, this state may persist
1155 * for arbitrarily long
1158 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1159 * interrupted while holding PENDING and @work off queue, irq must be
1160 * disabled on entry. This, combined with delayed_work->timer being
1161 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1163 * On successful return, >= 0, irq is disabled and the caller is
1164 * responsible for releasing it using local_irq_restore(*@flags).
1166 * This function is safe to call from any context including IRQ handler.
1168 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1169 unsigned long *flags)
1171 struct worker_pool *pool;
1172 struct pool_workqueue *pwq;
1174 local_irq_save(*flags);
1176 /* try to steal the timer if it exists */
1178 struct delayed_work *dwork = to_delayed_work(work);
1181 * dwork->timer is irqsafe. If del_timer() fails, it's
1182 * guaranteed that the timer is not queued anywhere and not
1183 * running on the local CPU.
1185 if (likely(del_timer(&dwork->timer)))
1189 /* try to claim PENDING the normal way */
1190 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1194 * The queueing is in progress, or it is already queued. Try to
1195 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1197 pool = get_work_pool(work);
1201 spin_lock(&pool->lock);
1203 * work->data is guaranteed to point to pwq only while the work
1204 * item is queued on pwq->wq, and both updating work->data to point
1205 * to pwq on queueing and to pool on dequeueing are done under
1206 * pwq->pool->lock. This in turn guarantees that, if work->data
1207 * points to pwq which is associated with a locked pool, the work
1208 * item is currently queued on that pool.
1210 pwq = get_work_pwq(work);
1211 if (pwq && pwq->pool == pool) {
1212 debug_work_deactivate(work);
1215 * A delayed work item cannot be grabbed directly because
1216 * it might have linked NO_COLOR work items which, if left
1217 * on the delayed_list, will confuse pwq->nr_active
1218 * management later on and cause stall. Make sure the work
1219 * item is activated before grabbing.
1221 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1222 pwq_activate_delayed_work(work);
1224 list_del_init(&work->entry);
1225 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1227 /* work->data points to pwq iff queued, point to pool */
1228 set_work_pool_and_keep_pending(work, pool->id);
1230 spin_unlock(&pool->lock);
1233 spin_unlock(&pool->lock);
1235 local_irq_restore(*flags);
1236 if (work_is_canceling(work))
1243 * insert_work - insert a work into a pool
1244 * @pwq: pwq @work belongs to
1245 * @work: work to insert
1246 * @head: insertion point
1247 * @extra_flags: extra WORK_STRUCT_* flags to set
1249 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1250 * work_struct flags.
1253 * spin_lock_irq(pool->lock).
1255 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1256 struct list_head *head, unsigned int extra_flags)
1258 struct worker_pool *pool = pwq->pool;
1260 /* we own @work, set data and link */
1261 set_work_pwq(work, pwq, extra_flags);
1262 list_add_tail(&work->entry, head);
1266 * Ensure either wq_worker_sleeping() sees the above
1267 * list_add_tail() or we see zero nr_running to avoid workers lying
1268 * around lazily while there are works to be processed.
1272 if (__need_more_worker(pool))
1273 wake_up_worker(pool);
1277 * Test whether @work is being queued from another work executing on the
1280 static bool is_chained_work(struct workqueue_struct *wq)
1282 struct worker *worker;
1284 worker = current_wq_worker();
1286 * Return %true iff I'm a worker execuing a work item on @wq. If
1287 * I'm @worker, it's safe to dereference it without locking.
1289 return worker && worker->current_pwq->wq == wq;
1292 static void __queue_work(int cpu, struct workqueue_struct *wq,
1293 struct work_struct *work)
1295 struct pool_workqueue *pwq;
1296 struct worker_pool *last_pool;
1297 struct list_head *worklist;
1298 unsigned int work_flags;
1299 unsigned int req_cpu = cpu;
1302 * While a work item is PENDING && off queue, a task trying to
1303 * steal the PENDING will busy-loop waiting for it to either get
1304 * queued or lose PENDING. Grabbing PENDING and queueing should
1305 * happen with IRQ disabled.
1307 WARN_ON_ONCE(!irqs_disabled());
1309 debug_work_activate(work);
1311 /* if draining, only works from the same workqueue are allowed */
1312 if (unlikely(wq->flags & __WQ_DRAINING) &&
1313 WARN_ON_ONCE(!is_chained_work(wq)))
1316 if (req_cpu == WORK_CPU_UNBOUND)
1317 cpu = raw_smp_processor_id();
1319 /* pwq which will be used unless @work is executing elsewhere */
1320 if (!(wq->flags & WQ_UNBOUND))
1321 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1323 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1326 * If @work was previously on a different pool, it might still be
1327 * running there, in which case the work needs to be queued on that
1328 * pool to guarantee non-reentrancy.
1330 last_pool = get_work_pool(work);
1331 if (last_pool && last_pool != pwq->pool) {
1332 struct worker *worker;
1334 spin_lock(&last_pool->lock);
1336 worker = find_worker_executing_work(last_pool, work);
1338 if (worker && worker->current_pwq->wq == wq) {
1339 pwq = worker->current_pwq;
1341 /* meh... not running there, queue here */
1342 spin_unlock(&last_pool->lock);
1343 spin_lock(&pwq->pool->lock);
1346 spin_lock(&pwq->pool->lock);
1350 * pwq is determined and locked. For unbound pools, we could have
1351 * raced with pwq release and it could already be dead. If its
1352 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1353 * without another pwq replacing it in the numa_pwq_tbl or while
1354 * work items are executing on it, so the retrying is guaranteed to
1355 * make forward-progress.
1357 if (unlikely(!pwq->refcnt)) {
1358 if (wq->flags & WQ_UNBOUND) {
1359 spin_unlock(&pwq->pool->lock);
1364 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1368 /* pwq determined, queue */
1369 trace_workqueue_queue_work(req_cpu, pwq, work);
1371 if (WARN_ON(!list_empty(&work->entry))) {
1372 spin_unlock(&pwq->pool->lock);
1376 pwq->nr_in_flight[pwq->work_color]++;
1377 work_flags = work_color_to_flags(pwq->work_color);
1379 if (likely(pwq->nr_active < pwq->max_active)) {
1380 trace_workqueue_activate_work(work);
1382 worklist = &pwq->pool->worklist;
1384 work_flags |= WORK_STRUCT_DELAYED;
1385 worklist = &pwq->delayed_works;
1388 insert_work(pwq, work, worklist, work_flags);
1390 spin_unlock(&pwq->pool->lock);
1394 * queue_work_on - queue work on specific cpu
1395 * @cpu: CPU number to execute work on
1396 * @wq: workqueue to use
1397 * @work: work to queue
1399 * We queue the work to a specific CPU, the caller must ensure it
1402 * Return: %false if @work was already on a queue, %true otherwise.
1404 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1405 struct work_struct *work)
1408 unsigned long flags;
1410 local_irq_save(flags);
1412 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1413 __queue_work(cpu, wq, work);
1417 local_irq_restore(flags);
1420 EXPORT_SYMBOL(queue_work_on);
1422 void delayed_work_timer_fn(unsigned long __data)
1424 struct delayed_work *dwork = (struct delayed_work *)__data;
1426 /* should have been called from irqsafe timer with irq already off */
1427 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1429 EXPORT_SYMBOL(delayed_work_timer_fn);
1431 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1432 struct delayed_work *dwork, unsigned long delay)
1434 struct timer_list *timer = &dwork->timer;
1435 struct work_struct *work = &dwork->work;
1437 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1438 timer->data != (unsigned long)dwork);
1439 WARN_ON_ONCE(timer_pending(timer));
1440 WARN_ON_ONCE(!list_empty(&work->entry));
1443 * If @delay is 0, queue @dwork->work immediately. This is for
1444 * both optimization and correctness. The earliest @timer can
1445 * expire is on the closest next tick and delayed_work users depend
1446 * on that there's no such delay when @delay is 0.
1449 __queue_work(cpu, wq, &dwork->work);
1453 timer_stats_timer_set_start_info(&dwork->timer);
1457 timer->expires = jiffies + delay;
1459 if (unlikely(cpu != WORK_CPU_UNBOUND))
1460 add_timer_on(timer, cpu);
1466 * queue_delayed_work_on - queue work on specific CPU after delay
1467 * @cpu: CPU number to execute work on
1468 * @wq: workqueue to use
1469 * @dwork: work to queue
1470 * @delay: number of jiffies to wait before queueing
1472 * Return: %false if @work was already on a queue, %true otherwise. If
1473 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1476 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1477 struct delayed_work *dwork, unsigned long delay)
1479 struct work_struct *work = &dwork->work;
1481 unsigned long flags;
1483 /* read the comment in __queue_work() */
1484 local_irq_save(flags);
1486 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1487 __queue_delayed_work(cpu, wq, dwork, delay);
1491 local_irq_restore(flags);
1494 EXPORT_SYMBOL(queue_delayed_work_on);
1497 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1498 * @cpu: CPU number to execute work on
1499 * @wq: workqueue to use
1500 * @dwork: work to queue
1501 * @delay: number of jiffies to wait before queueing
1503 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1504 * modify @dwork's timer so that it expires after @delay. If @delay is
1505 * zero, @work is guaranteed to be scheduled immediately regardless of its
1508 * Return: %false if @dwork was idle and queued, %true if @dwork was
1509 * pending and its timer was modified.
1511 * This function is safe to call from any context including IRQ handler.
1512 * See try_to_grab_pending() for details.
1514 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1515 struct delayed_work *dwork, unsigned long delay)
1517 unsigned long flags;
1521 ret = try_to_grab_pending(&dwork->work, true, &flags);
1522 } while (unlikely(ret == -EAGAIN));
1524 if (likely(ret >= 0)) {
1525 __queue_delayed_work(cpu, wq, dwork, delay);
1526 local_irq_restore(flags);
1529 /* -ENOENT from try_to_grab_pending() becomes %true */
1532 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1535 * worker_enter_idle - enter idle state
1536 * @worker: worker which is entering idle state
1538 * @worker is entering idle state. Update stats and idle timer if
1542 * spin_lock_irq(pool->lock).
1544 static void worker_enter_idle(struct worker *worker)
1546 struct worker_pool *pool = worker->pool;
1548 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1549 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1550 (worker->hentry.next || worker->hentry.pprev)))
1553 /* can't use worker_set_flags(), also called from create_worker() */
1554 worker->flags |= WORKER_IDLE;
1556 worker->last_active = jiffies;
1558 /* idle_list is LIFO */
1559 list_add(&worker->entry, &pool->idle_list);
1561 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1562 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1565 * Sanity check nr_running. Because wq_unbind_fn() releases
1566 * pool->lock between setting %WORKER_UNBOUND and zapping
1567 * nr_running, the warning may trigger spuriously. Check iff
1568 * unbind is not in progress.
1570 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1571 pool->nr_workers == pool->nr_idle &&
1572 atomic_read(&pool->nr_running));
1576 * worker_leave_idle - leave idle state
1577 * @worker: worker which is leaving idle state
1579 * @worker is leaving idle state. Update stats.
1582 * spin_lock_irq(pool->lock).
1584 static void worker_leave_idle(struct worker *worker)
1586 struct worker_pool *pool = worker->pool;
1588 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1590 worker_clr_flags(worker, WORKER_IDLE);
1592 list_del_init(&worker->entry);
1595 static struct worker *alloc_worker(int node)
1597 struct worker *worker;
1599 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1601 INIT_LIST_HEAD(&worker->entry);
1602 INIT_LIST_HEAD(&worker->scheduled);
1603 INIT_LIST_HEAD(&worker->node);
1604 /* on creation a worker is in !idle && prep state */
1605 worker->flags = WORKER_PREP;
1611 * worker_attach_to_pool() - attach a worker to a pool
1612 * @worker: worker to be attached
1613 * @pool: the target pool
1615 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1616 * cpu-binding of @worker are kept coordinated with the pool across
1619 static void worker_attach_to_pool(struct worker *worker,
1620 struct worker_pool *pool)
1622 mutex_lock(&pool->attach_mutex);
1625 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1626 * online CPUs. It'll be re-applied when any of the CPUs come up.
1628 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1631 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1632 * stable across this function. See the comments above the
1633 * flag definition for details.
1635 if (pool->flags & POOL_DISASSOCIATED)
1636 worker->flags |= WORKER_UNBOUND;
1638 list_add_tail(&worker->node, &pool->workers);
1640 mutex_unlock(&pool->attach_mutex);
1644 * worker_detach_from_pool() - detach a worker from its pool
1645 * @worker: worker which is attached to its pool
1646 * @pool: the pool @worker is attached to
1648 * Undo the attaching which had been done in worker_attach_to_pool(). The
1649 * caller worker shouldn't access to the pool after detached except it has
1650 * other reference to the pool.
1652 static void worker_detach_from_pool(struct worker *worker,
1653 struct worker_pool *pool)
1655 struct completion *detach_completion = NULL;
1657 mutex_lock(&pool->attach_mutex);
1658 list_del(&worker->node);
1659 if (list_empty(&pool->workers))
1660 detach_completion = pool->detach_completion;
1661 mutex_unlock(&pool->attach_mutex);
1663 /* clear leftover flags without pool->lock after it is detached */
1664 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1666 if (detach_completion)
1667 complete(detach_completion);
1671 * create_worker - create a new workqueue worker
1672 * @pool: pool the new worker will belong to
1674 * Create and start a new worker which is attached to @pool.
1677 * Might sleep. Does GFP_KERNEL allocations.
1680 * Pointer to the newly created worker.
1682 static struct worker *create_worker(struct worker_pool *pool)
1684 struct worker *worker = NULL;
1688 /* ID is needed to determine kthread name */
1689 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1693 worker = alloc_worker(pool->node);
1697 worker->pool = pool;
1701 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1702 pool->attrs->nice < 0 ? "H" : "");
1704 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1706 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1707 "kworker/%s", id_buf);
1708 if (IS_ERR(worker->task))
1711 set_user_nice(worker->task, pool->attrs->nice);
1713 /* prevent userland from meddling with cpumask of workqueue workers */
1714 worker->task->flags |= PF_NO_SETAFFINITY;
1716 /* successful, attach the worker to the pool */
1717 worker_attach_to_pool(worker, pool);
1719 /* start the newly created worker */
1720 spin_lock_irq(&pool->lock);
1721 worker->pool->nr_workers++;
1722 worker_enter_idle(worker);
1723 wake_up_process(worker->task);
1724 spin_unlock_irq(&pool->lock);
1730 ida_simple_remove(&pool->worker_ida, id);
1736 * destroy_worker - destroy a workqueue worker
1737 * @worker: worker to be destroyed
1739 * Destroy @worker and adjust @pool stats accordingly. The worker should
1743 * spin_lock_irq(pool->lock).
1745 static void destroy_worker(struct worker *worker)
1747 struct worker_pool *pool = worker->pool;
1749 lockdep_assert_held(&pool->lock);
1751 /* sanity check frenzy */
1752 if (WARN_ON(worker->current_work) ||
1753 WARN_ON(!list_empty(&worker->scheduled)) ||
1754 WARN_ON(!(worker->flags & WORKER_IDLE)))
1760 list_del_init(&worker->entry);
1761 worker->flags |= WORKER_DIE;
1762 wake_up_process(worker->task);
1765 static void idle_worker_timeout(unsigned long __pool)
1767 struct worker_pool *pool = (void *)__pool;
1769 spin_lock_irq(&pool->lock);
1771 while (too_many_workers(pool)) {
1772 struct worker *worker;
1773 unsigned long expires;
1775 /* idle_list is kept in LIFO order, check the last one */
1776 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1777 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1779 if (time_before(jiffies, expires)) {
1780 mod_timer(&pool->idle_timer, expires);
1784 destroy_worker(worker);
1787 spin_unlock_irq(&pool->lock);
1790 static void send_mayday(struct work_struct *work)
1792 struct pool_workqueue *pwq = get_work_pwq(work);
1793 struct workqueue_struct *wq = pwq->wq;
1795 lockdep_assert_held(&wq_mayday_lock);
1800 /* mayday mayday mayday */
1801 if (list_empty(&pwq->mayday_node)) {
1803 * If @pwq is for an unbound wq, its base ref may be put at
1804 * any time due to an attribute change. Pin @pwq until the
1805 * rescuer is done with it.
1808 list_add_tail(&pwq->mayday_node, &wq->maydays);
1809 wake_up_process(wq->rescuer->task);
1813 static void pool_mayday_timeout(unsigned long __pool)
1815 struct worker_pool *pool = (void *)__pool;
1816 struct work_struct *work;
1818 spin_lock_irq(&pool->lock);
1819 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1821 if (need_to_create_worker(pool)) {
1823 * We've been trying to create a new worker but
1824 * haven't been successful. We might be hitting an
1825 * allocation deadlock. Send distress signals to
1828 list_for_each_entry(work, &pool->worklist, entry)
1832 spin_unlock(&wq_mayday_lock);
1833 spin_unlock_irq(&pool->lock);
1835 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1839 * maybe_create_worker - create a new worker if necessary
1840 * @pool: pool to create a new worker for
1842 * Create a new worker for @pool if necessary. @pool is guaranteed to
1843 * have at least one idle worker on return from this function. If
1844 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1845 * sent to all rescuers with works scheduled on @pool to resolve
1846 * possible allocation deadlock.
1848 * On return, need_to_create_worker() is guaranteed to be %false and
1849 * may_start_working() %true.
1852 * spin_lock_irq(pool->lock) which may be released and regrabbed
1853 * multiple times. Does GFP_KERNEL allocations. Called only from
1856 static void maybe_create_worker(struct worker_pool *pool)
1857 __releases(&pool->lock)
1858 __acquires(&pool->lock)
1861 spin_unlock_irq(&pool->lock);
1863 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1864 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1867 if (create_worker(pool) || !need_to_create_worker(pool))
1870 schedule_timeout_interruptible(CREATE_COOLDOWN);
1872 if (!need_to_create_worker(pool))
1876 del_timer_sync(&pool->mayday_timer);
1877 spin_lock_irq(&pool->lock);
1879 * This is necessary even after a new worker was just successfully
1880 * created as @pool->lock was dropped and the new worker might have
1881 * already become busy.
1883 if (need_to_create_worker(pool))
1888 * manage_workers - manage worker pool
1891 * Assume the manager role and manage the worker pool @worker belongs
1892 * to. At any given time, there can be only zero or one manager per
1893 * pool. The exclusion is handled automatically by this function.
1895 * The caller can safely start processing works on false return. On
1896 * true return, it's guaranteed that need_to_create_worker() is false
1897 * and may_start_working() is true.
1900 * spin_lock_irq(pool->lock) which may be released and regrabbed
1901 * multiple times. Does GFP_KERNEL allocations.
1904 * %false if the pool doesn't need management and the caller can safely
1905 * start processing works, %true if management function was performed and
1906 * the conditions that the caller verified before calling the function may
1907 * no longer be true.
1909 static bool manage_workers(struct worker *worker)
1911 struct worker_pool *pool = worker->pool;
1914 * Anyone who successfully grabs manager_arb wins the arbitration
1915 * and becomes the manager. mutex_trylock() on pool->manager_arb
1916 * failure while holding pool->lock reliably indicates that someone
1917 * else is managing the pool and the worker which failed trylock
1918 * can proceed to executing work items. This means that anyone
1919 * grabbing manager_arb is responsible for actually performing
1920 * manager duties. If manager_arb is grabbed and released without
1921 * actual management, the pool may stall indefinitely.
1923 if (!mutex_trylock(&pool->manager_arb))
1925 pool->manager = worker;
1927 maybe_create_worker(pool);
1929 pool->manager = NULL;
1930 mutex_unlock(&pool->manager_arb);
1935 * process_one_work - process single work
1937 * @work: work to process
1939 * Process @work. This function contains all the logics necessary to
1940 * process a single work including synchronization against and
1941 * interaction with other workers on the same cpu, queueing and
1942 * flushing. As long as context requirement is met, any worker can
1943 * call this function to process a work.
1946 * spin_lock_irq(pool->lock) which is released and regrabbed.
1948 static void process_one_work(struct worker *worker, struct work_struct *work)
1949 __releases(&pool->lock)
1950 __acquires(&pool->lock)
1952 struct pool_workqueue *pwq = get_work_pwq(work);
1953 struct worker_pool *pool = worker->pool;
1954 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
1956 struct worker *collision;
1957 #ifdef CONFIG_LOCKDEP
1959 * It is permissible to free the struct work_struct from
1960 * inside the function that is called from it, this we need to
1961 * take into account for lockdep too. To avoid bogus "held
1962 * lock freed" warnings as well as problems when looking into
1963 * work->lockdep_map, make a copy and use that here.
1965 struct lockdep_map lockdep_map;
1967 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
1969 /* ensure we're on the correct CPU */
1970 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1971 raw_smp_processor_id() != pool->cpu);
1974 * A single work shouldn't be executed concurrently by
1975 * multiple workers on a single cpu. Check whether anyone is
1976 * already processing the work. If so, defer the work to the
1977 * currently executing one.
1979 collision = find_worker_executing_work(pool, work);
1980 if (unlikely(collision)) {
1981 move_linked_works(work, &collision->scheduled, NULL);
1985 /* claim and dequeue */
1986 debug_work_deactivate(work);
1987 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
1988 worker->current_work = work;
1989 worker->current_func = work->func;
1990 worker->current_pwq = pwq;
1991 work_color = get_work_color(work);
1993 list_del_init(&work->entry);
1996 * CPU intensive works don't participate in concurrency management.
1997 * They're the scheduler's responsibility. This takes @worker out
1998 * of concurrency management and the next code block will chain
1999 * execution of the pending work items.
2001 if (unlikely(cpu_intensive))
2002 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2005 * Wake up another worker if necessary. The condition is always
2006 * false for normal per-cpu workers since nr_running would always
2007 * be >= 1 at this point. This is used to chain execution of the
2008 * pending work items for WORKER_NOT_RUNNING workers such as the
2009 * UNBOUND and CPU_INTENSIVE ones.
2011 if (need_more_worker(pool))
2012 wake_up_worker(pool);
2015 * Record the last pool and clear PENDING which should be the last
2016 * update to @work. Also, do this inside @pool->lock so that
2017 * PENDING and queued state changes happen together while IRQ is
2020 set_work_pool_and_clear_pending(work, pool->id);
2022 spin_unlock_irq(&pool->lock);
2024 lock_map_acquire_read(&pwq->wq->lockdep_map);
2025 lock_map_acquire(&lockdep_map);
2026 trace_workqueue_execute_start(work);
2027 worker->current_func(work);
2029 * While we must be careful to not use "work" after this, the trace
2030 * point will only record its address.
2032 trace_workqueue_execute_end(work);
2033 lock_map_release(&lockdep_map);
2034 lock_map_release(&pwq->wq->lockdep_map);
2036 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2037 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2038 " last function: %pf\n",
2039 current->comm, preempt_count(), task_pid_nr(current),
2040 worker->current_func);
2041 debug_show_held_locks(current);
2046 * The following prevents a kworker from hogging CPU on !PREEMPT
2047 * kernels, where a requeueing work item waiting for something to
2048 * happen could deadlock with stop_machine as such work item could
2049 * indefinitely requeue itself while all other CPUs are trapped in
2050 * stop_machine. At the same time, report a quiescent RCU state so
2051 * the same condition doesn't freeze RCU.
2053 cond_resched_rcu_qs();
2055 spin_lock_irq(&pool->lock);
2057 /* clear cpu intensive status */
2058 if (unlikely(cpu_intensive))
2059 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2061 /* we're done with it, release */
2062 hash_del(&worker->hentry);
2063 worker->current_work = NULL;
2064 worker->current_func = NULL;
2065 worker->current_pwq = NULL;
2066 worker->desc_valid = false;
2067 pwq_dec_nr_in_flight(pwq, work_color);
2071 * process_scheduled_works - process scheduled works
2074 * Process all scheduled works. Please note that the scheduled list
2075 * may change while processing a work, so this function repeatedly
2076 * fetches a work from the top and executes it.
2079 * spin_lock_irq(pool->lock) which may be released and regrabbed
2082 static void process_scheduled_works(struct worker *worker)
2084 while (!list_empty(&worker->scheduled)) {
2085 struct work_struct *work = list_first_entry(&worker->scheduled,
2086 struct work_struct, entry);
2087 process_one_work(worker, work);
2092 * worker_thread - the worker thread function
2095 * The worker thread function. All workers belong to a worker_pool -
2096 * either a per-cpu one or dynamic unbound one. These workers process all
2097 * work items regardless of their specific target workqueue. The only
2098 * exception is work items which belong to workqueues with a rescuer which
2099 * will be explained in rescuer_thread().
2103 static int worker_thread(void *__worker)
2105 struct worker *worker = __worker;
2106 struct worker_pool *pool = worker->pool;
2108 /* tell the scheduler that this is a workqueue worker */
2109 worker->task->flags |= PF_WQ_WORKER;
2111 spin_lock_irq(&pool->lock);
2113 /* am I supposed to die? */
2114 if (unlikely(worker->flags & WORKER_DIE)) {
2115 spin_unlock_irq(&pool->lock);
2116 WARN_ON_ONCE(!list_empty(&worker->entry));
2117 worker->task->flags &= ~PF_WQ_WORKER;
2119 set_task_comm(worker->task, "kworker/dying");
2120 ida_simple_remove(&pool->worker_ida, worker->id);
2121 worker_detach_from_pool(worker, pool);
2126 worker_leave_idle(worker);
2128 /* no more worker necessary? */
2129 if (!need_more_worker(pool))
2132 /* do we need to manage? */
2133 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2137 * ->scheduled list can only be filled while a worker is
2138 * preparing to process a work or actually processing it.
2139 * Make sure nobody diddled with it while I was sleeping.
2141 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2144 * Finish PREP stage. We're guaranteed to have at least one idle
2145 * worker or that someone else has already assumed the manager
2146 * role. This is where @worker starts participating in concurrency
2147 * management if applicable and concurrency management is restored
2148 * after being rebound. See rebind_workers() for details.
2150 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2153 struct work_struct *work =
2154 list_first_entry(&pool->worklist,
2155 struct work_struct, entry);
2157 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2158 /* optimization path, not strictly necessary */
2159 process_one_work(worker, work);
2160 if (unlikely(!list_empty(&worker->scheduled)))
2161 process_scheduled_works(worker);
2163 move_linked_works(work, &worker->scheduled, NULL);
2164 process_scheduled_works(worker);
2166 } while (keep_working(pool));
2168 worker_set_flags(worker, WORKER_PREP);
2171 * pool->lock is held and there's no work to process and no need to
2172 * manage, sleep. Workers are woken up only while holding
2173 * pool->lock or from local cpu, so setting the current state
2174 * before releasing pool->lock is enough to prevent losing any
2177 worker_enter_idle(worker);
2178 __set_current_state(TASK_INTERRUPTIBLE);
2179 spin_unlock_irq(&pool->lock);
2185 * rescuer_thread - the rescuer thread function
2188 * Workqueue rescuer thread function. There's one rescuer for each
2189 * workqueue which has WQ_MEM_RECLAIM set.
2191 * Regular work processing on a pool may block trying to create a new
2192 * worker which uses GFP_KERNEL allocation which has slight chance of
2193 * developing into deadlock if some works currently on the same queue
2194 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2195 * the problem rescuer solves.
2197 * When such condition is possible, the pool summons rescuers of all
2198 * workqueues which have works queued on the pool and let them process
2199 * those works so that forward progress can be guaranteed.
2201 * This should happen rarely.
2205 static int rescuer_thread(void *__rescuer)
2207 struct worker *rescuer = __rescuer;
2208 struct workqueue_struct *wq = rescuer->rescue_wq;
2209 struct list_head *scheduled = &rescuer->scheduled;
2212 set_user_nice(current, RESCUER_NICE_LEVEL);
2215 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2216 * doesn't participate in concurrency management.
2218 rescuer->task->flags |= PF_WQ_WORKER;
2220 set_current_state(TASK_INTERRUPTIBLE);
2223 * By the time the rescuer is requested to stop, the workqueue
2224 * shouldn't have any work pending, but @wq->maydays may still have
2225 * pwq(s) queued. This can happen by non-rescuer workers consuming
2226 * all the work items before the rescuer got to them. Go through
2227 * @wq->maydays processing before acting on should_stop so that the
2228 * list is always empty on exit.
2230 should_stop = kthread_should_stop();
2232 /* see whether any pwq is asking for help */
2233 spin_lock_irq(&wq_mayday_lock);
2235 while (!list_empty(&wq->maydays)) {
2236 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2237 struct pool_workqueue, mayday_node);
2238 struct worker_pool *pool = pwq->pool;
2239 struct work_struct *work, *n;
2241 __set_current_state(TASK_RUNNING);
2242 list_del_init(&pwq->mayday_node);
2244 spin_unlock_irq(&wq_mayday_lock);
2246 worker_attach_to_pool(rescuer, pool);
2248 spin_lock_irq(&pool->lock);
2249 rescuer->pool = pool;
2252 * Slurp in all works issued via this workqueue and
2255 WARN_ON_ONCE(!list_empty(scheduled));
2256 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2257 if (get_work_pwq(work) == pwq)
2258 move_linked_works(work, scheduled, &n);
2260 if (!list_empty(scheduled)) {
2261 process_scheduled_works(rescuer);
2264 * The above execution of rescued work items could
2265 * have created more to rescue through
2266 * pwq_activate_first_delayed() or chained
2267 * queueing. Let's put @pwq back on mayday list so
2268 * that such back-to-back work items, which may be
2269 * being used to relieve memory pressure, don't
2270 * incur MAYDAY_INTERVAL delay inbetween.
2272 if (need_to_create_worker(pool)) {
2273 spin_lock(&wq_mayday_lock);
2275 list_move_tail(&pwq->mayday_node, &wq->maydays);
2276 spin_unlock(&wq_mayday_lock);
2281 * Put the reference grabbed by send_mayday(). @pool won't
2282 * go away while we're still attached to it.
2287 * Leave this pool. If need_more_worker() is %true, notify a
2288 * regular worker; otherwise, we end up with 0 concurrency
2289 * and stalling the execution.
2291 if (need_more_worker(pool))
2292 wake_up_worker(pool);
2294 rescuer->pool = NULL;
2295 spin_unlock_irq(&pool->lock);
2297 worker_detach_from_pool(rescuer, pool);
2299 spin_lock_irq(&wq_mayday_lock);
2302 spin_unlock_irq(&wq_mayday_lock);
2305 __set_current_state(TASK_RUNNING);
2306 rescuer->task->flags &= ~PF_WQ_WORKER;
2310 /* rescuers should never participate in concurrency management */
2311 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2317 struct work_struct work;
2318 struct completion done;
2319 struct task_struct *task; /* purely informational */
2322 static void wq_barrier_func(struct work_struct *work)
2324 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2325 complete(&barr->done);
2329 * insert_wq_barrier - insert a barrier work
2330 * @pwq: pwq to insert barrier into
2331 * @barr: wq_barrier to insert
2332 * @target: target work to attach @barr to
2333 * @worker: worker currently executing @target, NULL if @target is not executing
2335 * @barr is linked to @target such that @barr is completed only after
2336 * @target finishes execution. Please note that the ordering
2337 * guarantee is observed only with respect to @target and on the local
2340 * Currently, a queued barrier can't be canceled. This is because
2341 * try_to_grab_pending() can't determine whether the work to be
2342 * grabbed is at the head of the queue and thus can't clear LINKED
2343 * flag of the previous work while there must be a valid next work
2344 * after a work with LINKED flag set.
2346 * Note that when @worker is non-NULL, @target may be modified
2347 * underneath us, so we can't reliably determine pwq from @target.
2350 * spin_lock_irq(pool->lock).
2352 static void insert_wq_barrier(struct pool_workqueue *pwq,
2353 struct wq_barrier *barr,
2354 struct work_struct *target, struct worker *worker)
2356 struct list_head *head;
2357 unsigned int linked = 0;
2360 * debugobject calls are safe here even with pool->lock locked
2361 * as we know for sure that this will not trigger any of the
2362 * checks and call back into the fixup functions where we
2365 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2366 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2367 init_completion(&barr->done);
2368 barr->task = current;
2371 * If @target is currently being executed, schedule the
2372 * barrier to the worker; otherwise, put it after @target.
2375 head = worker->scheduled.next;
2377 unsigned long *bits = work_data_bits(target);
2379 head = target->entry.next;
2380 /* there can already be other linked works, inherit and set */
2381 linked = *bits & WORK_STRUCT_LINKED;
2382 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2385 debug_work_activate(&barr->work);
2386 insert_work(pwq, &barr->work, head,
2387 work_color_to_flags(WORK_NO_COLOR) | linked);
2391 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2392 * @wq: workqueue being flushed
2393 * @flush_color: new flush color, < 0 for no-op
2394 * @work_color: new work color, < 0 for no-op
2396 * Prepare pwqs for workqueue flushing.
2398 * If @flush_color is non-negative, flush_color on all pwqs should be
2399 * -1. If no pwq has in-flight commands at the specified color, all
2400 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2401 * has in flight commands, its pwq->flush_color is set to
2402 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2403 * wakeup logic is armed and %true is returned.
2405 * The caller should have initialized @wq->first_flusher prior to
2406 * calling this function with non-negative @flush_color. If
2407 * @flush_color is negative, no flush color update is done and %false
2410 * If @work_color is non-negative, all pwqs should have the same
2411 * work_color which is previous to @work_color and all will be
2412 * advanced to @work_color.
2415 * mutex_lock(wq->mutex).
2418 * %true if @flush_color >= 0 and there's something to flush. %false
2421 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2422 int flush_color, int work_color)
2425 struct pool_workqueue *pwq;
2427 if (flush_color >= 0) {
2428 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2429 atomic_set(&wq->nr_pwqs_to_flush, 1);
2432 for_each_pwq(pwq, wq) {
2433 struct worker_pool *pool = pwq->pool;
2435 spin_lock_irq(&pool->lock);
2437 if (flush_color >= 0) {
2438 WARN_ON_ONCE(pwq->flush_color != -1);
2440 if (pwq->nr_in_flight[flush_color]) {
2441 pwq->flush_color = flush_color;
2442 atomic_inc(&wq->nr_pwqs_to_flush);
2447 if (work_color >= 0) {
2448 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2449 pwq->work_color = work_color;
2452 spin_unlock_irq(&pool->lock);
2455 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2456 complete(&wq->first_flusher->done);
2462 * flush_workqueue - ensure that any scheduled work has run to completion.
2463 * @wq: workqueue to flush
2465 * This function sleeps until all work items which were queued on entry
2466 * have finished execution, but it is not livelocked by new incoming ones.
2468 void flush_workqueue(struct workqueue_struct *wq)
2470 struct wq_flusher this_flusher = {
2471 .list = LIST_HEAD_INIT(this_flusher.list),
2473 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2477 lock_map_acquire(&wq->lockdep_map);
2478 lock_map_release(&wq->lockdep_map);
2480 mutex_lock(&wq->mutex);
2483 * Start-to-wait phase
2485 next_color = work_next_color(wq->work_color);
2487 if (next_color != wq->flush_color) {
2489 * Color space is not full. The current work_color
2490 * becomes our flush_color and work_color is advanced
2493 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2494 this_flusher.flush_color = wq->work_color;
2495 wq->work_color = next_color;
2497 if (!wq->first_flusher) {
2498 /* no flush in progress, become the first flusher */
2499 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2501 wq->first_flusher = &this_flusher;
2503 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2505 /* nothing to flush, done */
2506 wq->flush_color = next_color;
2507 wq->first_flusher = NULL;
2512 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2513 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2514 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2518 * Oops, color space is full, wait on overflow queue.
2519 * The next flush completion will assign us
2520 * flush_color and transfer to flusher_queue.
2522 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2525 mutex_unlock(&wq->mutex);
2527 wait_for_completion(&this_flusher.done);
2530 * Wake-up-and-cascade phase
2532 * First flushers are responsible for cascading flushes and
2533 * handling overflow. Non-first flushers can simply return.
2535 if (wq->first_flusher != &this_flusher)
2538 mutex_lock(&wq->mutex);
2540 /* we might have raced, check again with mutex held */
2541 if (wq->first_flusher != &this_flusher)
2544 wq->first_flusher = NULL;
2546 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2547 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2550 struct wq_flusher *next, *tmp;
2552 /* complete all the flushers sharing the current flush color */
2553 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2554 if (next->flush_color != wq->flush_color)
2556 list_del_init(&next->list);
2557 complete(&next->done);
2560 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2561 wq->flush_color != work_next_color(wq->work_color));
2563 /* this flush_color is finished, advance by one */
2564 wq->flush_color = work_next_color(wq->flush_color);
2566 /* one color has been freed, handle overflow queue */
2567 if (!list_empty(&wq->flusher_overflow)) {
2569 * Assign the same color to all overflowed
2570 * flushers, advance work_color and append to
2571 * flusher_queue. This is the start-to-wait
2572 * phase for these overflowed flushers.
2574 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2575 tmp->flush_color = wq->work_color;
2577 wq->work_color = work_next_color(wq->work_color);
2579 list_splice_tail_init(&wq->flusher_overflow,
2580 &wq->flusher_queue);
2581 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2584 if (list_empty(&wq->flusher_queue)) {
2585 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2590 * Need to flush more colors. Make the next flusher
2591 * the new first flusher and arm pwqs.
2593 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2594 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2596 list_del_init(&next->list);
2597 wq->first_flusher = next;
2599 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2603 * Meh... this color is already done, clear first
2604 * flusher and repeat cascading.
2606 wq->first_flusher = NULL;
2610 mutex_unlock(&wq->mutex);
2612 EXPORT_SYMBOL_GPL(flush_workqueue);
2615 * drain_workqueue - drain a workqueue
2616 * @wq: workqueue to drain
2618 * Wait until the workqueue becomes empty. While draining is in progress,
2619 * only chain queueing is allowed. IOW, only currently pending or running
2620 * work items on @wq can queue further work items on it. @wq is flushed
2621 * repeatedly until it becomes empty. The number of flushing is determined
2622 * by the depth of chaining and should be relatively short. Whine if it
2625 void drain_workqueue(struct workqueue_struct *wq)
2627 unsigned int flush_cnt = 0;
2628 struct pool_workqueue *pwq;
2631 * __queue_work() needs to test whether there are drainers, is much
2632 * hotter than drain_workqueue() and already looks at @wq->flags.
2633 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2635 mutex_lock(&wq->mutex);
2636 if (!wq->nr_drainers++)
2637 wq->flags |= __WQ_DRAINING;
2638 mutex_unlock(&wq->mutex);
2640 flush_workqueue(wq);
2642 mutex_lock(&wq->mutex);
2644 for_each_pwq(pwq, wq) {
2647 spin_lock_irq(&pwq->pool->lock);
2648 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2649 spin_unlock_irq(&pwq->pool->lock);
2654 if (++flush_cnt == 10 ||
2655 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2656 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2657 wq->name, flush_cnt);
2659 mutex_unlock(&wq->mutex);
2663 if (!--wq->nr_drainers)
2664 wq->flags &= ~__WQ_DRAINING;
2665 mutex_unlock(&wq->mutex);
2667 EXPORT_SYMBOL_GPL(drain_workqueue);
2669 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2671 struct worker *worker = NULL;
2672 struct worker_pool *pool;
2673 struct pool_workqueue *pwq;
2677 local_irq_disable();
2678 pool = get_work_pool(work);
2684 spin_lock(&pool->lock);
2685 /* see the comment in try_to_grab_pending() with the same code */
2686 pwq = get_work_pwq(work);
2688 if (unlikely(pwq->pool != pool))
2691 worker = find_worker_executing_work(pool, work);
2694 pwq = worker->current_pwq;
2697 insert_wq_barrier(pwq, barr, work, worker);
2698 spin_unlock_irq(&pool->lock);
2701 * If @max_active is 1 or rescuer is in use, flushing another work
2702 * item on the same workqueue may lead to deadlock. Make sure the
2703 * flusher is not running on the same workqueue by verifying write
2706 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2707 lock_map_acquire(&pwq->wq->lockdep_map);
2709 lock_map_acquire_read(&pwq->wq->lockdep_map);
2710 lock_map_release(&pwq->wq->lockdep_map);
2714 spin_unlock_irq(&pool->lock);
2719 * flush_work - wait for a work to finish executing the last queueing instance
2720 * @work: the work to flush
2722 * Wait until @work has finished execution. @work is guaranteed to be idle
2723 * on return if it hasn't been requeued since flush started.
2726 * %true if flush_work() waited for the work to finish execution,
2727 * %false if it was already idle.
2729 bool flush_work(struct work_struct *work)
2731 struct wq_barrier barr;
2733 lock_map_acquire(&work->lockdep_map);
2734 lock_map_release(&work->lockdep_map);
2736 if (start_flush_work(work, &barr)) {
2737 wait_for_completion(&barr.done);
2738 destroy_work_on_stack(&barr.work);
2744 EXPORT_SYMBOL_GPL(flush_work);
2748 struct work_struct *work;
2751 static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2753 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2755 if (cwait->work != key)
2757 return autoremove_wake_function(wait, mode, sync, key);
2760 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2762 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2763 unsigned long flags;
2767 ret = try_to_grab_pending(work, is_dwork, &flags);
2769 * If someone else is already canceling, wait for it to
2770 * finish. flush_work() doesn't work for PREEMPT_NONE
2771 * because we may get scheduled between @work's completion
2772 * and the other canceling task resuming and clearing
2773 * CANCELING - flush_work() will return false immediately
2774 * as @work is no longer busy, try_to_grab_pending() will
2775 * return -ENOENT as @work is still being canceled and the
2776 * other canceling task won't be able to clear CANCELING as
2777 * we're hogging the CPU.
2779 * Let's wait for completion using a waitqueue. As this
2780 * may lead to the thundering herd problem, use a custom
2781 * wake function which matches @work along with exclusive
2784 if (unlikely(ret == -ENOENT)) {
2785 struct cwt_wait cwait;
2787 init_wait(&cwait.wait);
2788 cwait.wait.func = cwt_wakefn;
2791 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2792 TASK_UNINTERRUPTIBLE);
2793 if (work_is_canceling(work))
2795 finish_wait(&cancel_waitq, &cwait.wait);
2797 } while (unlikely(ret < 0));
2799 /* tell other tasks trying to grab @work to back off */
2800 mark_work_canceling(work);
2801 local_irq_restore(flags);
2804 clear_work_data(work);
2807 * Paired with prepare_to_wait() above so that either
2808 * waitqueue_active() is visible here or !work_is_canceling() is
2812 if (waitqueue_active(&cancel_waitq))
2813 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2819 * cancel_work_sync - cancel a work and wait for it to finish
2820 * @work: the work to cancel
2822 * Cancel @work and wait for its execution to finish. This function
2823 * can be used even if the work re-queues itself or migrates to
2824 * another workqueue. On return from this function, @work is
2825 * guaranteed to be not pending or executing on any CPU.
2827 * cancel_work_sync(&delayed_work->work) must not be used for
2828 * delayed_work's. Use cancel_delayed_work_sync() instead.
2830 * The caller must ensure that the workqueue on which @work was last
2831 * queued can't be destroyed before this function returns.
2834 * %true if @work was pending, %false otherwise.
2836 bool cancel_work_sync(struct work_struct *work)
2838 return __cancel_work_timer(work, false);
2840 EXPORT_SYMBOL_GPL(cancel_work_sync);
2843 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2844 * @dwork: the delayed work to flush
2846 * Delayed timer is cancelled and the pending work is queued for
2847 * immediate execution. Like flush_work(), this function only
2848 * considers the last queueing instance of @dwork.
2851 * %true if flush_work() waited for the work to finish execution,
2852 * %false if it was already idle.
2854 bool flush_delayed_work(struct delayed_work *dwork)
2856 local_irq_disable();
2857 if (del_timer_sync(&dwork->timer))
2858 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2860 return flush_work(&dwork->work);
2862 EXPORT_SYMBOL(flush_delayed_work);
2865 * cancel_delayed_work - cancel a delayed work
2866 * @dwork: delayed_work to cancel
2868 * Kill off a pending delayed_work.
2870 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2874 * The work callback function may still be running on return, unless
2875 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2876 * use cancel_delayed_work_sync() to wait on it.
2878 * This function is safe to call from any context including IRQ handler.
2880 bool cancel_delayed_work(struct delayed_work *dwork)
2882 unsigned long flags;
2886 ret = try_to_grab_pending(&dwork->work, true, &flags);
2887 } while (unlikely(ret == -EAGAIN));
2889 if (unlikely(ret < 0))
2892 set_work_pool_and_clear_pending(&dwork->work,
2893 get_work_pool_id(&dwork->work));
2894 local_irq_restore(flags);
2897 EXPORT_SYMBOL(cancel_delayed_work);
2900 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2901 * @dwork: the delayed work cancel
2903 * This is cancel_work_sync() for delayed works.
2906 * %true if @dwork was pending, %false otherwise.
2908 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2910 return __cancel_work_timer(&dwork->work, true);
2912 EXPORT_SYMBOL(cancel_delayed_work_sync);
2915 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2916 * @func: the function to call
2918 * schedule_on_each_cpu() executes @func on each online CPU using the
2919 * system workqueue and blocks until all CPUs have completed.
2920 * schedule_on_each_cpu() is very slow.
2923 * 0 on success, -errno on failure.
2925 int schedule_on_each_cpu(work_func_t func)
2928 struct work_struct __percpu *works;
2930 works = alloc_percpu(struct work_struct);
2936 for_each_online_cpu(cpu) {
2937 struct work_struct *work = per_cpu_ptr(works, cpu);
2939 INIT_WORK(work, func);
2940 schedule_work_on(cpu, work);
2943 for_each_online_cpu(cpu)
2944 flush_work(per_cpu_ptr(works, cpu));
2952 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2954 * Forces execution of the kernel-global workqueue and blocks until its
2957 * Think twice before calling this function! It's very easy to get into
2958 * trouble if you don't take great care. Either of the following situations
2959 * will lead to deadlock:
2961 * One of the work items currently on the workqueue needs to acquire
2962 * a lock held by your code or its caller.
2964 * Your code is running in the context of a work routine.
2966 * They will be detected by lockdep when they occur, but the first might not
2967 * occur very often. It depends on what work items are on the workqueue and
2968 * what locks they need, which you have no control over.
2970 * In most situations flushing the entire workqueue is overkill; you merely
2971 * need to know that a particular work item isn't queued and isn't running.
2972 * In such cases you should use cancel_delayed_work_sync() or
2973 * cancel_work_sync() instead.
2975 void flush_scheduled_work(void)
2977 flush_workqueue(system_wq);
2979 EXPORT_SYMBOL(flush_scheduled_work);
2982 * execute_in_process_context - reliably execute the routine with user context
2983 * @fn: the function to execute
2984 * @ew: guaranteed storage for the execute work structure (must
2985 * be available when the work executes)
2987 * Executes the function immediately if process context is available,
2988 * otherwise schedules the function for delayed execution.
2990 * Return: 0 - function was executed
2991 * 1 - function was scheduled for execution
2993 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2995 if (!in_interrupt()) {
3000 INIT_WORK(&ew->work, fn);
3001 schedule_work(&ew->work);
3005 EXPORT_SYMBOL_GPL(execute_in_process_context);
3008 * free_workqueue_attrs - free a workqueue_attrs
3009 * @attrs: workqueue_attrs to free
3011 * Undo alloc_workqueue_attrs().
3013 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3016 free_cpumask_var(attrs->cpumask);
3022 * alloc_workqueue_attrs - allocate a workqueue_attrs
3023 * @gfp_mask: allocation mask to use
3025 * Allocate a new workqueue_attrs, initialize with default settings and
3028 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3030 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3032 struct workqueue_attrs *attrs;
3034 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3037 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3040 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3043 free_workqueue_attrs(attrs);
3047 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3048 const struct workqueue_attrs *from)
3050 to->nice = from->nice;
3051 cpumask_copy(to->cpumask, from->cpumask);
3053 * Unlike hash and equality test, this function doesn't ignore
3054 * ->no_numa as it is used for both pool and wq attrs. Instead,
3055 * get_unbound_pool() explicitly clears ->no_numa after copying.
3057 to->no_numa = from->no_numa;
3060 /* hash value of the content of @attr */
3061 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3065 hash = jhash_1word(attrs->nice, hash);
3066 hash = jhash(cpumask_bits(attrs->cpumask),
3067 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3071 /* content equality test */
3072 static bool wqattrs_equal(const struct workqueue_attrs *a,
3073 const struct workqueue_attrs *b)
3075 if (a->nice != b->nice)
3077 if (!cpumask_equal(a->cpumask, b->cpumask))
3083 * init_worker_pool - initialize a newly zalloc'd worker_pool
3084 * @pool: worker_pool to initialize
3086 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3088 * Return: 0 on success, -errno on failure. Even on failure, all fields
3089 * inside @pool proper are initialized and put_unbound_pool() can be called
3090 * on @pool safely to release it.
3092 static int init_worker_pool(struct worker_pool *pool)
3094 spin_lock_init(&pool->lock);
3097 pool->node = NUMA_NO_NODE;
3098 pool->flags |= POOL_DISASSOCIATED;
3099 INIT_LIST_HEAD(&pool->worklist);
3100 INIT_LIST_HEAD(&pool->idle_list);
3101 hash_init(pool->busy_hash);
3103 init_timer_deferrable(&pool->idle_timer);
3104 pool->idle_timer.function = idle_worker_timeout;
3105 pool->idle_timer.data = (unsigned long)pool;
3107 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3108 (unsigned long)pool);
3110 mutex_init(&pool->manager_arb);
3111 mutex_init(&pool->attach_mutex);
3112 INIT_LIST_HEAD(&pool->workers);
3114 ida_init(&pool->worker_ida);
3115 INIT_HLIST_NODE(&pool->hash_node);
3118 /* shouldn't fail above this point */
3119 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3125 static void rcu_free_wq(struct rcu_head *rcu)
3127 struct workqueue_struct *wq =
3128 container_of(rcu, struct workqueue_struct, rcu);
3130 if (!(wq->flags & WQ_UNBOUND))
3131 free_percpu(wq->cpu_pwqs);
3133 free_workqueue_attrs(wq->unbound_attrs);
3139 static void rcu_free_pool(struct rcu_head *rcu)
3141 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3143 ida_destroy(&pool->worker_ida);
3144 free_workqueue_attrs(pool->attrs);
3149 * put_unbound_pool - put a worker_pool
3150 * @pool: worker_pool to put
3152 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3153 * safe manner. get_unbound_pool() calls this function on its failure path
3154 * and this function should be able to release pools which went through,
3155 * successfully or not, init_worker_pool().
3157 * Should be called with wq_pool_mutex held.
3159 static void put_unbound_pool(struct worker_pool *pool)
3161 DECLARE_COMPLETION_ONSTACK(detach_completion);
3162 struct worker *worker;
3164 lockdep_assert_held(&wq_pool_mutex);
3170 if (WARN_ON(!(pool->cpu < 0)) ||
3171 WARN_ON(!list_empty(&pool->worklist)))
3174 /* release id and unhash */
3176 idr_remove(&worker_pool_idr, pool->id);
3177 hash_del(&pool->hash_node);
3180 * Become the manager and destroy all workers. Grabbing
3181 * manager_arb prevents @pool's workers from blocking on
3184 mutex_lock(&pool->manager_arb);
3186 spin_lock_irq(&pool->lock);
3187 while ((worker = first_idle_worker(pool)))
3188 destroy_worker(worker);
3189 WARN_ON(pool->nr_workers || pool->nr_idle);
3190 spin_unlock_irq(&pool->lock);
3192 mutex_lock(&pool->attach_mutex);
3193 if (!list_empty(&pool->workers))
3194 pool->detach_completion = &detach_completion;
3195 mutex_unlock(&pool->attach_mutex);
3197 if (pool->detach_completion)
3198 wait_for_completion(pool->detach_completion);
3200 mutex_unlock(&pool->manager_arb);
3202 /* shut down the timers */
3203 del_timer_sync(&pool->idle_timer);
3204 del_timer_sync(&pool->mayday_timer);
3206 /* sched-RCU protected to allow dereferences from get_work_pool() */
3207 call_rcu_sched(&pool->rcu, rcu_free_pool);
3211 * get_unbound_pool - get a worker_pool with the specified attributes
3212 * @attrs: the attributes of the worker_pool to get
3214 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3215 * reference count and return it. If there already is a matching
3216 * worker_pool, it will be used; otherwise, this function attempts to
3219 * Should be called with wq_pool_mutex held.
3221 * Return: On success, a worker_pool with the same attributes as @attrs.
3222 * On failure, %NULL.
3224 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3226 u32 hash = wqattrs_hash(attrs);
3227 struct worker_pool *pool;
3230 lockdep_assert_held(&wq_pool_mutex);
3232 /* do we already have a matching pool? */
3233 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3234 if (wqattrs_equal(pool->attrs, attrs)) {
3240 /* nope, create a new one */
3241 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3242 if (!pool || init_worker_pool(pool) < 0)
3245 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3246 copy_workqueue_attrs(pool->attrs, attrs);
3249 * no_numa isn't a worker_pool attribute, always clear it. See
3250 * 'struct workqueue_attrs' comments for detail.
3252 pool->attrs->no_numa = false;
3254 /* if cpumask is contained inside a NUMA node, we belong to that node */
3255 if (wq_numa_enabled) {
3256 for_each_node(node) {
3257 if (cpumask_subset(pool->attrs->cpumask,
3258 wq_numa_possible_cpumask[node])) {
3265 if (worker_pool_assign_id(pool) < 0)
3268 /* create and start the initial worker */
3269 if (!create_worker(pool))
3273 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3278 put_unbound_pool(pool);
3282 static void rcu_free_pwq(struct rcu_head *rcu)
3284 kmem_cache_free(pwq_cache,
3285 container_of(rcu, struct pool_workqueue, rcu));
3289 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3290 * and needs to be destroyed.
3292 static void pwq_unbound_release_workfn(struct work_struct *work)
3294 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3295 unbound_release_work);
3296 struct workqueue_struct *wq = pwq->wq;
3297 struct worker_pool *pool = pwq->pool;
3300 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3303 mutex_lock(&wq->mutex);
3304 list_del_rcu(&pwq->pwqs_node);
3305 is_last = list_empty(&wq->pwqs);
3306 mutex_unlock(&wq->mutex);
3308 mutex_lock(&wq_pool_mutex);
3309 put_unbound_pool(pool);
3310 mutex_unlock(&wq_pool_mutex);
3312 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3315 * If we're the last pwq going away, @wq is already dead and no one
3316 * is gonna access it anymore. Schedule RCU free.
3319 call_rcu_sched(&wq->rcu, rcu_free_wq);
3323 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3324 * @pwq: target pool_workqueue
3326 * If @pwq isn't freezing, set @pwq->max_active to the associated
3327 * workqueue's saved_max_active and activate delayed work items
3328 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3330 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3332 struct workqueue_struct *wq = pwq->wq;
3333 bool freezable = wq->flags & WQ_FREEZABLE;
3335 /* for @wq->saved_max_active */
3336 lockdep_assert_held(&wq->mutex);
3338 /* fast exit for non-freezable wqs */
3339 if (!freezable && pwq->max_active == wq->saved_max_active)
3342 spin_lock_irq(&pwq->pool->lock);
3345 * During [un]freezing, the caller is responsible for ensuring that
3346 * this function is called at least once after @workqueue_freezing
3347 * is updated and visible.
3349 if (!freezable || !workqueue_freezing) {
3350 pwq->max_active = wq->saved_max_active;
3352 while (!list_empty(&pwq->delayed_works) &&
3353 pwq->nr_active < pwq->max_active)
3354 pwq_activate_first_delayed(pwq);
3357 * Need to kick a worker after thawed or an unbound wq's
3358 * max_active is bumped. It's a slow path. Do it always.
3360 wake_up_worker(pwq->pool);
3362 pwq->max_active = 0;
3365 spin_unlock_irq(&pwq->pool->lock);
3368 /* initialize newly alloced @pwq which is associated with @wq and @pool */
3369 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3370 struct worker_pool *pool)
3372 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3374 memset(pwq, 0, sizeof(*pwq));
3378 pwq->flush_color = -1;
3380 INIT_LIST_HEAD(&pwq->delayed_works);
3381 INIT_LIST_HEAD(&pwq->pwqs_node);
3382 INIT_LIST_HEAD(&pwq->mayday_node);
3383 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3386 /* sync @pwq with the current state of its associated wq and link it */
3387 static void link_pwq(struct pool_workqueue *pwq)
3389 struct workqueue_struct *wq = pwq->wq;
3391 lockdep_assert_held(&wq->mutex);
3393 /* may be called multiple times, ignore if already linked */
3394 if (!list_empty(&pwq->pwqs_node))
3397 /* set the matching work_color */
3398 pwq->work_color = wq->work_color;
3400 /* sync max_active to the current setting */
3401 pwq_adjust_max_active(pwq);
3404 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3407 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3408 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3409 const struct workqueue_attrs *attrs)
3411 struct worker_pool *pool;
3412 struct pool_workqueue *pwq;
3414 lockdep_assert_held(&wq_pool_mutex);
3416 pool = get_unbound_pool(attrs);
3420 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3422 put_unbound_pool(pool);
3426 init_pwq(pwq, wq, pool);
3431 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3432 * @attrs: the wq_attrs of the default pwq of the target workqueue
3433 * @node: the target NUMA node
3434 * @cpu_going_down: if >= 0, the CPU to consider as offline
3435 * @cpumask: outarg, the resulting cpumask
3437 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3438 * @cpu_going_down is >= 0, that cpu is considered offline during
3439 * calculation. The result is stored in @cpumask.
3441 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3442 * enabled and @node has online CPUs requested by @attrs, the returned
3443 * cpumask is the intersection of the possible CPUs of @node and
3446 * The caller is responsible for ensuring that the cpumask of @node stays
3449 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3452 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3453 int cpu_going_down, cpumask_t *cpumask)
3455 if (!wq_numa_enabled || attrs->no_numa)
3458 /* does @node have any online CPUs @attrs wants? */
3459 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3460 if (cpu_going_down >= 0)
3461 cpumask_clear_cpu(cpu_going_down, cpumask);
3463 if (cpumask_empty(cpumask))
3466 /* yeap, return possible CPUs in @node that @attrs wants */
3467 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3468 return !cpumask_equal(cpumask, attrs->cpumask);
3471 cpumask_copy(cpumask, attrs->cpumask);
3475 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3476 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3478 struct pool_workqueue *pwq)
3480 struct pool_workqueue *old_pwq;
3482 lockdep_assert_held(&wq->mutex);
3484 /* link_pwq() can handle duplicate calls */
3487 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3488 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3492 /* context to store the prepared attrs & pwqs before applying */
3493 struct apply_wqattrs_ctx {
3494 struct workqueue_struct *wq; /* target workqueue */
3495 struct workqueue_attrs *attrs; /* attrs to apply */
3496 struct list_head list; /* queued for batching commit */
3497 struct pool_workqueue *dfl_pwq;
3498 struct pool_workqueue *pwq_tbl[];
3501 /* free the resources after success or abort */
3502 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3508 put_pwq_unlocked(ctx->pwq_tbl[node]);
3509 put_pwq_unlocked(ctx->dfl_pwq);
3511 free_workqueue_attrs(ctx->attrs);
3517 /* allocate the attrs and pwqs for later installation */
3518 static struct apply_wqattrs_ctx *
3519 apply_wqattrs_prepare(struct workqueue_struct *wq,
3520 const struct workqueue_attrs *attrs)
3522 struct apply_wqattrs_ctx *ctx;
3523 struct workqueue_attrs *new_attrs, *tmp_attrs;
3526 lockdep_assert_held(&wq_pool_mutex);
3528 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3531 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3532 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3533 if (!ctx || !new_attrs || !tmp_attrs)
3537 * Calculate the attrs of the default pwq.
3538 * If the user configured cpumask doesn't overlap with the
3539 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3541 copy_workqueue_attrs(new_attrs, attrs);
3542 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3543 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3544 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3547 * We may create multiple pwqs with differing cpumasks. Make a
3548 * copy of @new_attrs which will be modified and used to obtain
3551 copy_workqueue_attrs(tmp_attrs, new_attrs);
3554 * If something goes wrong during CPU up/down, we'll fall back to
3555 * the default pwq covering whole @attrs->cpumask. Always create
3556 * it even if we don't use it immediately.
3558 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3562 for_each_node(node) {
3563 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3564 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3565 if (!ctx->pwq_tbl[node])
3568 ctx->dfl_pwq->refcnt++;
3569 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3573 /* save the user configured attrs and sanitize it. */
3574 copy_workqueue_attrs(new_attrs, attrs);
3575 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3576 ctx->attrs = new_attrs;
3579 free_workqueue_attrs(tmp_attrs);
3583 free_workqueue_attrs(tmp_attrs);
3584 free_workqueue_attrs(new_attrs);
3585 apply_wqattrs_cleanup(ctx);
3589 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3590 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3594 /* all pwqs have been created successfully, let's install'em */
3595 mutex_lock(&ctx->wq->mutex);
3597 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3599 /* save the previous pwq and install the new one */
3601 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3602 ctx->pwq_tbl[node]);
3604 /* @dfl_pwq might not have been used, ensure it's linked */
3605 link_pwq(ctx->dfl_pwq);
3606 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3608 mutex_unlock(&ctx->wq->mutex);
3612 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3613 * @wq: the target workqueue
3614 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3616 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3617 * machines, this function maps a separate pwq to each NUMA node with
3618 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3619 * NUMA node it was issued on. Older pwqs are released as in-flight work
3620 * items finish. Note that a work item which repeatedly requeues itself
3621 * back-to-back will stay on its current pwq.
3623 * Performs GFP_KERNEL allocations.
3625 * Return: 0 on success and -errno on failure.
3627 int apply_workqueue_attrs(struct workqueue_struct *wq,
3628 const struct workqueue_attrs *attrs)
3630 struct apply_wqattrs_ctx *ctx;
3633 /* only unbound workqueues can change attributes */
3634 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3637 /* creating multiple pwqs breaks ordering guarantee */
3638 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3642 * CPUs should stay stable across pwq creations and installations.
3643 * Pin CPUs, determine the target cpumask for each node and create
3648 mutex_lock(&wq_pool_mutex);
3649 ctx = apply_wqattrs_prepare(wq, attrs);
3650 mutex_unlock(&wq_pool_mutex);
3652 /* the ctx has been prepared successfully, let's commit it */
3654 apply_wqattrs_commit(ctx);
3660 apply_wqattrs_cleanup(ctx);
3666 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3667 * @wq: the target workqueue
3668 * @cpu: the CPU coming up or going down
3669 * @online: whether @cpu is coming up or going down
3671 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3672 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3675 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3676 * falls back to @wq->dfl_pwq which may not be optimal but is always
3679 * Note that when the last allowed CPU of a NUMA node goes offline for a
3680 * workqueue with a cpumask spanning multiple nodes, the workers which were
3681 * already executing the work items for the workqueue will lose their CPU
3682 * affinity and may execute on any CPU. This is similar to how per-cpu
3683 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3684 * affinity, it's the user's responsibility to flush the work item from
3687 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3690 int node = cpu_to_node(cpu);
3691 int cpu_off = online ? -1 : cpu;
3692 struct pool_workqueue *old_pwq = NULL, *pwq;
3693 struct workqueue_attrs *target_attrs;
3696 lockdep_assert_held(&wq_pool_mutex);
3698 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
3702 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3703 * Let's use a preallocated one. The following buf is protected by
3704 * CPU hotplug exclusion.
3706 target_attrs = wq_update_unbound_numa_attrs_buf;
3707 cpumask = target_attrs->cpumask;
3709 mutex_lock(&wq->mutex);
3710 if (wq->unbound_attrs->no_numa)
3713 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3714 pwq = unbound_pwq_by_node(wq, node);
3717 * Let's determine what needs to be done. If the target cpumask is
3718 * different from the default pwq's, we need to compare it to @pwq's
3719 * and create a new one if they don't match. If the target cpumask
3720 * equals the default pwq's, the default pwq should be used.
3722 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3723 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3729 mutex_unlock(&wq->mutex);
3731 /* create a new pwq */
3732 pwq = alloc_unbound_pwq(wq, target_attrs);
3734 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3736 mutex_lock(&wq->mutex);
3741 * Install the new pwq. As this function is called only from CPU
3742 * hotplug callbacks and applying a new attrs is wrapped with
3743 * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
3746 mutex_lock(&wq->mutex);
3747 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3751 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3752 get_pwq(wq->dfl_pwq);
3753 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3754 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3756 mutex_unlock(&wq->mutex);
3757 put_pwq_unlocked(old_pwq);
3760 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3762 bool highpri = wq->flags & WQ_HIGHPRI;
3765 if (!(wq->flags & WQ_UNBOUND)) {
3766 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3770 for_each_possible_cpu(cpu) {
3771 struct pool_workqueue *pwq =
3772 per_cpu_ptr(wq->cpu_pwqs, cpu);
3773 struct worker_pool *cpu_pools =
3774 per_cpu(cpu_worker_pools, cpu);
3776 init_pwq(pwq, wq, &cpu_pools[highpri]);
3778 mutex_lock(&wq->mutex);
3780 mutex_unlock(&wq->mutex);
3783 } else if (wq->flags & __WQ_ORDERED) {
3784 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3785 /* there should only be single pwq for ordering guarantee */
3786 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3787 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3788 "ordering guarantee broken for workqueue %s\n", wq->name);
3791 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3795 static int wq_clamp_max_active(int max_active, unsigned int flags,
3798 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3800 if (max_active < 1 || max_active > lim)
3801 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3802 max_active, name, 1, lim);
3804 return clamp_val(max_active, 1, lim);
3807 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3810 struct lock_class_key *key,
3811 const char *lock_name, ...)
3813 size_t tbl_size = 0;
3815 struct workqueue_struct *wq;
3816 struct pool_workqueue *pwq;
3818 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3819 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3820 flags |= WQ_UNBOUND;
3822 /* allocate wq and format name */
3823 if (flags & WQ_UNBOUND)
3824 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3826 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3830 if (flags & WQ_UNBOUND) {
3831 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3832 if (!wq->unbound_attrs)
3836 va_start(args, lock_name);
3837 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3840 max_active = max_active ?: WQ_DFL_ACTIVE;
3841 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3845 wq->saved_max_active = max_active;
3846 mutex_init(&wq->mutex);
3847 atomic_set(&wq->nr_pwqs_to_flush, 0);
3848 INIT_LIST_HEAD(&wq->pwqs);
3849 INIT_LIST_HEAD(&wq->flusher_queue);
3850 INIT_LIST_HEAD(&wq->flusher_overflow);
3851 INIT_LIST_HEAD(&wq->maydays);
3853 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3854 INIT_LIST_HEAD(&wq->list);
3856 if (alloc_and_link_pwqs(wq) < 0)
3860 * Workqueues which may be used during memory reclaim should
3861 * have a rescuer to guarantee forward progress.
3863 if (flags & WQ_MEM_RECLAIM) {
3864 struct worker *rescuer;
3866 rescuer = alloc_worker(NUMA_NO_NODE);
3870 rescuer->rescue_wq = wq;
3871 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3873 if (IS_ERR(rescuer->task)) {
3878 wq->rescuer = rescuer;
3879 rescuer->task->flags |= PF_NO_SETAFFINITY;
3880 wake_up_process(rescuer->task);
3883 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3887 * wq_pool_mutex protects global freeze state and workqueues list.
3888 * Grab it, adjust max_active and add the new @wq to workqueues
3891 mutex_lock(&wq_pool_mutex);
3893 mutex_lock(&wq->mutex);
3894 for_each_pwq(pwq, wq)
3895 pwq_adjust_max_active(pwq);
3896 mutex_unlock(&wq->mutex);
3898 list_add_tail_rcu(&wq->list, &workqueues);
3900 mutex_unlock(&wq_pool_mutex);
3905 free_workqueue_attrs(wq->unbound_attrs);
3909 destroy_workqueue(wq);
3912 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3915 * destroy_workqueue - safely terminate a workqueue
3916 * @wq: target workqueue
3918 * Safely destroy a workqueue. All work currently pending will be done first.
3920 void destroy_workqueue(struct workqueue_struct *wq)
3922 struct pool_workqueue *pwq;
3925 /* drain it before proceeding with destruction */
3926 drain_workqueue(wq);
3929 mutex_lock(&wq->mutex);
3930 for_each_pwq(pwq, wq) {
3933 for (i = 0; i < WORK_NR_COLORS; i++) {
3934 if (WARN_ON(pwq->nr_in_flight[i])) {
3935 mutex_unlock(&wq->mutex);
3940 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
3941 WARN_ON(pwq->nr_active) ||
3942 WARN_ON(!list_empty(&pwq->delayed_works))) {
3943 mutex_unlock(&wq->mutex);
3947 mutex_unlock(&wq->mutex);
3950 * wq list is used to freeze wq, remove from list after
3951 * flushing is complete in case freeze races us.
3953 mutex_lock(&wq_pool_mutex);
3954 list_del_rcu(&wq->list);
3955 mutex_unlock(&wq_pool_mutex);
3957 workqueue_sysfs_unregister(wq);
3960 kthread_stop(wq->rescuer->task);
3962 if (!(wq->flags & WQ_UNBOUND)) {
3964 * The base ref is never dropped on per-cpu pwqs. Directly
3965 * schedule RCU free.
3967 call_rcu_sched(&wq->rcu, rcu_free_wq);
3970 * We're the sole accessor of @wq at this point. Directly
3971 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
3972 * @wq will be freed when the last pwq is released.
3974 for_each_node(node) {
3975 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3976 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
3977 put_pwq_unlocked(pwq);
3981 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
3982 * put. Don't access it afterwards.
3986 put_pwq_unlocked(pwq);
3989 EXPORT_SYMBOL_GPL(destroy_workqueue);
3992 * workqueue_set_max_active - adjust max_active of a workqueue
3993 * @wq: target workqueue
3994 * @max_active: new max_active value.
3996 * Set max_active of @wq to @max_active.
3999 * Don't call from IRQ context.
4001 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4003 struct pool_workqueue *pwq;
4005 /* disallow meddling with max_active for ordered workqueues */
4006 if (WARN_ON(wq->flags & __WQ_ORDERED))
4009 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4011 mutex_lock(&wq->mutex);
4013 wq->saved_max_active = max_active;
4015 for_each_pwq(pwq, wq)
4016 pwq_adjust_max_active(pwq);
4018 mutex_unlock(&wq->mutex);
4020 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4023 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4025 * Determine whether %current is a workqueue rescuer. Can be used from
4026 * work functions to determine whether it's being run off the rescuer task.
4028 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4030 bool current_is_workqueue_rescuer(void)
4032 struct worker *worker = current_wq_worker();
4034 return worker && worker->rescue_wq;
4038 * workqueue_congested - test whether a workqueue is congested
4039 * @cpu: CPU in question
4040 * @wq: target workqueue
4042 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4043 * no synchronization around this function and the test result is
4044 * unreliable and only useful as advisory hints or for debugging.
4046 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4047 * Note that both per-cpu and unbound workqueues may be associated with
4048 * multiple pool_workqueues which have separate congested states. A
4049 * workqueue being congested on one CPU doesn't mean the workqueue is also
4050 * contested on other CPUs / NUMA nodes.
4053 * %true if congested, %false otherwise.
4055 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4057 struct pool_workqueue *pwq;
4060 rcu_read_lock_sched();
4062 if (cpu == WORK_CPU_UNBOUND)
4063 cpu = smp_processor_id();
4065 if (!(wq->flags & WQ_UNBOUND))
4066 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4068 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4070 ret = !list_empty(&pwq->delayed_works);
4071 rcu_read_unlock_sched();
4075 EXPORT_SYMBOL_GPL(workqueue_congested);
4078 * work_busy - test whether a work is currently pending or running
4079 * @work: the work to be tested
4081 * Test whether @work is currently pending or running. There is no
4082 * synchronization around this function and the test result is
4083 * unreliable and only useful as advisory hints or for debugging.
4086 * OR'd bitmask of WORK_BUSY_* bits.
4088 unsigned int work_busy(struct work_struct *work)
4090 struct worker_pool *pool;
4091 unsigned long flags;
4092 unsigned int ret = 0;
4094 if (work_pending(work))
4095 ret |= WORK_BUSY_PENDING;
4097 local_irq_save(flags);
4098 pool = get_work_pool(work);
4100 spin_lock(&pool->lock);
4101 if (find_worker_executing_work(pool, work))
4102 ret |= WORK_BUSY_RUNNING;
4103 spin_unlock(&pool->lock);
4105 local_irq_restore(flags);
4109 EXPORT_SYMBOL_GPL(work_busy);
4112 * set_worker_desc - set description for the current work item
4113 * @fmt: printf-style format string
4114 * @...: arguments for the format string
4116 * This function can be called by a running work function to describe what
4117 * the work item is about. If the worker task gets dumped, this
4118 * information will be printed out together to help debugging. The
4119 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4121 void set_worker_desc(const char *fmt, ...)
4123 struct worker *worker = current_wq_worker();
4127 va_start(args, fmt);
4128 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4130 worker->desc_valid = true;
4135 * print_worker_info - print out worker information and description
4136 * @log_lvl: the log level to use when printing
4137 * @task: target task
4139 * If @task is a worker and currently executing a work item, print out the
4140 * name of the workqueue being serviced and worker description set with
4141 * set_worker_desc() by the currently executing work item.
4143 * This function can be safely called on any task as long as the
4144 * task_struct itself is accessible. While safe, this function isn't
4145 * synchronized and may print out mixups or garbages of limited length.
4147 void print_worker_info(const char *log_lvl, struct task_struct *task)
4149 work_func_t *fn = NULL;
4150 char name[WQ_NAME_LEN] = { };
4151 char desc[WORKER_DESC_LEN] = { };
4152 struct pool_workqueue *pwq = NULL;
4153 struct workqueue_struct *wq = NULL;
4154 bool desc_valid = false;
4155 struct worker *worker;
4157 if (!(task->flags & PF_WQ_WORKER))
4161 * This function is called without any synchronization and @task
4162 * could be in any state. Be careful with dereferences.
4164 worker = probe_kthread_data(task);
4167 * Carefully copy the associated workqueue's workfn and name. Keep
4168 * the original last '\0' in case the original contains garbage.
4170 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4171 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4172 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4173 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4175 /* copy worker description */
4176 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4178 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4180 if (fn || name[0] || desc[0]) {
4181 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4183 pr_cont(" (%s)", desc);
4188 static void pr_cont_pool_info(struct worker_pool *pool)
4190 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4191 if (pool->node != NUMA_NO_NODE)
4192 pr_cont(" node=%d", pool->node);
4193 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4196 static void pr_cont_work(bool comma, struct work_struct *work)
4198 if (work->func == wq_barrier_func) {
4199 struct wq_barrier *barr;
4201 barr = container_of(work, struct wq_barrier, work);
4203 pr_cont("%s BAR(%d)", comma ? "," : "",
4204 task_pid_nr(barr->task));
4206 pr_cont("%s %pf", comma ? "," : "", work->func);
4210 static void show_pwq(struct pool_workqueue *pwq)
4212 struct worker_pool *pool = pwq->pool;
4213 struct work_struct *work;
4214 struct worker *worker;
4215 bool has_in_flight = false, has_pending = false;
4218 pr_info(" pwq %d:", pool->id);
4219 pr_cont_pool_info(pool);
4221 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4222 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4224 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4225 if (worker->current_pwq == pwq) {
4226 has_in_flight = true;
4230 if (has_in_flight) {
4233 pr_info(" in-flight:");
4234 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4235 if (worker->current_pwq != pwq)
4238 pr_cont("%s %d%s:%pf", comma ? "," : "",
4239 task_pid_nr(worker->task),
4240 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4241 worker->current_func);
4242 list_for_each_entry(work, &worker->scheduled, entry)
4243 pr_cont_work(false, work);
4249 list_for_each_entry(work, &pool->worklist, entry) {
4250 if (get_work_pwq(work) == pwq) {
4258 pr_info(" pending:");
4259 list_for_each_entry(work, &pool->worklist, entry) {
4260 if (get_work_pwq(work) != pwq)
4263 pr_cont_work(comma, work);
4264 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4269 if (!list_empty(&pwq->delayed_works)) {
4272 pr_info(" delayed:");
4273 list_for_each_entry(work, &pwq->delayed_works, entry) {
4274 pr_cont_work(comma, work);
4275 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4282 * show_workqueue_state - dump workqueue state
4284 * Called from a sysrq handler and prints out all busy workqueues and
4287 void show_workqueue_state(void)
4289 struct workqueue_struct *wq;
4290 struct worker_pool *pool;
4291 unsigned long flags;
4294 rcu_read_lock_sched();
4296 pr_info("Showing busy workqueues and worker pools:\n");
4298 list_for_each_entry_rcu(wq, &workqueues, list) {
4299 struct pool_workqueue *pwq;
4302 for_each_pwq(pwq, wq) {
4303 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4311 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4313 for_each_pwq(pwq, wq) {
4314 spin_lock_irqsave(&pwq->pool->lock, flags);
4315 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4317 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4321 for_each_pool(pool, pi) {
4322 struct worker *worker;
4325 spin_lock_irqsave(&pool->lock, flags);
4326 if (pool->nr_workers == pool->nr_idle)
4329 pr_info("pool %d:", pool->id);
4330 pr_cont_pool_info(pool);
4331 pr_cont(" workers=%d", pool->nr_workers);
4333 pr_cont(" manager: %d",
4334 task_pid_nr(pool->manager->task));
4335 list_for_each_entry(worker, &pool->idle_list, entry) {
4336 pr_cont(" %s%d", first ? "idle: " : "",
4337 task_pid_nr(worker->task));
4342 spin_unlock_irqrestore(&pool->lock, flags);
4345 rcu_read_unlock_sched();
4351 * There are two challenges in supporting CPU hotplug. Firstly, there
4352 * are a lot of assumptions on strong associations among work, pwq and
4353 * pool which make migrating pending and scheduled works very
4354 * difficult to implement without impacting hot paths. Secondly,
4355 * worker pools serve mix of short, long and very long running works making
4356 * blocked draining impractical.
4358 * This is solved by allowing the pools to be disassociated from the CPU
4359 * running as an unbound one and allowing it to be reattached later if the
4360 * cpu comes back online.
4363 static void wq_unbind_fn(struct work_struct *work)
4365 int cpu = smp_processor_id();
4366 struct worker_pool *pool;
4367 struct worker *worker;
4369 for_each_cpu_worker_pool(pool, cpu) {
4370 mutex_lock(&pool->attach_mutex);
4371 spin_lock_irq(&pool->lock);
4374 * We've blocked all attach/detach operations. Make all workers
4375 * unbound and set DISASSOCIATED. Before this, all workers
4376 * except for the ones which are still executing works from
4377 * before the last CPU down must be on the cpu. After
4378 * this, they may become diasporas.
4380 for_each_pool_worker(worker, pool)
4381 worker->flags |= WORKER_UNBOUND;
4383 pool->flags |= POOL_DISASSOCIATED;
4385 spin_unlock_irq(&pool->lock);
4386 mutex_unlock(&pool->attach_mutex);
4389 * Call schedule() so that we cross rq->lock and thus can
4390 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4391 * This is necessary as scheduler callbacks may be invoked
4397 * Sched callbacks are disabled now. Zap nr_running.
4398 * After this, nr_running stays zero and need_more_worker()
4399 * and keep_working() are always true as long as the
4400 * worklist is not empty. This pool now behaves as an
4401 * unbound (in terms of concurrency management) pool which
4402 * are served by workers tied to the pool.
4404 atomic_set(&pool->nr_running, 0);
4407 * With concurrency management just turned off, a busy
4408 * worker blocking could lead to lengthy stalls. Kick off
4409 * unbound chain execution of currently pending work items.
4411 spin_lock_irq(&pool->lock);
4412 wake_up_worker(pool);
4413 spin_unlock_irq(&pool->lock);
4418 * rebind_workers - rebind all workers of a pool to the associated CPU
4419 * @pool: pool of interest
4421 * @pool->cpu is coming online. Rebind all workers to the CPU.
4423 static void rebind_workers(struct worker_pool *pool)
4425 struct worker *worker;
4427 lockdep_assert_held(&pool->attach_mutex);
4430 * Restore CPU affinity of all workers. As all idle workers should
4431 * be on the run-queue of the associated CPU before any local
4432 * wake-ups for concurrency management happen, restore CPU affinty
4433 * of all workers first and then clear UNBOUND. As we're called
4434 * from CPU_ONLINE, the following shouldn't fail.
4436 for_each_pool_worker(worker, pool)
4437 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4438 pool->attrs->cpumask) < 0);
4440 spin_lock_irq(&pool->lock);
4441 pool->flags &= ~POOL_DISASSOCIATED;
4443 for_each_pool_worker(worker, pool) {
4444 unsigned int worker_flags = worker->flags;
4447 * A bound idle worker should actually be on the runqueue
4448 * of the associated CPU for local wake-ups targeting it to
4449 * work. Kick all idle workers so that they migrate to the
4450 * associated CPU. Doing this in the same loop as
4451 * replacing UNBOUND with REBOUND is safe as no worker will
4452 * be bound before @pool->lock is released.
4454 if (worker_flags & WORKER_IDLE)
4455 wake_up_process(worker->task);
4458 * We want to clear UNBOUND but can't directly call
4459 * worker_clr_flags() or adjust nr_running. Atomically
4460 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4461 * @worker will clear REBOUND using worker_clr_flags() when
4462 * it initiates the next execution cycle thus restoring
4463 * concurrency management. Note that when or whether
4464 * @worker clears REBOUND doesn't affect correctness.
4466 * ACCESS_ONCE() is necessary because @worker->flags may be
4467 * tested without holding any lock in
4468 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4469 * fail incorrectly leading to premature concurrency
4470 * management operations.
4472 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4473 worker_flags |= WORKER_REBOUND;
4474 worker_flags &= ~WORKER_UNBOUND;
4475 ACCESS_ONCE(worker->flags) = worker_flags;
4478 spin_unlock_irq(&pool->lock);
4482 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4483 * @pool: unbound pool of interest
4484 * @cpu: the CPU which is coming up
4486 * An unbound pool may end up with a cpumask which doesn't have any online
4487 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4488 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4489 * online CPU before, cpus_allowed of all its workers should be restored.
4491 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4493 static cpumask_t cpumask;
4494 struct worker *worker;
4496 lockdep_assert_held(&pool->attach_mutex);
4498 /* is @cpu allowed for @pool? */
4499 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4502 /* is @cpu the only online CPU? */
4503 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4504 if (cpumask_weight(&cpumask) != 1)
4507 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4508 for_each_pool_worker(worker, pool)
4509 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4510 pool->attrs->cpumask) < 0);
4514 * Workqueues should be brought up before normal priority CPU notifiers.
4515 * This will be registered high priority CPU notifier.
4517 static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4518 unsigned long action,
4521 int cpu = (unsigned long)hcpu;
4522 struct worker_pool *pool;
4523 struct workqueue_struct *wq;
4526 switch (action & ~CPU_TASKS_FROZEN) {
4527 case CPU_UP_PREPARE:
4528 for_each_cpu_worker_pool(pool, cpu) {
4529 if (pool->nr_workers)
4531 if (!create_worker(pool))
4536 case CPU_DOWN_FAILED:
4538 mutex_lock(&wq_pool_mutex);
4540 for_each_pool(pool, pi) {
4541 mutex_lock(&pool->attach_mutex);
4543 if (pool->cpu == cpu)
4544 rebind_workers(pool);
4545 else if (pool->cpu < 0)
4546 restore_unbound_workers_cpumask(pool, cpu);
4548 mutex_unlock(&pool->attach_mutex);
4551 /* update NUMA affinity of unbound workqueues */
4552 list_for_each_entry(wq, &workqueues, list)
4553 wq_update_unbound_numa(wq, cpu, true);
4555 mutex_unlock(&wq_pool_mutex);
4562 * Workqueues should be brought down after normal priority CPU notifiers.
4563 * This will be registered as low priority CPU notifier.
4565 static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4566 unsigned long action,
4569 int cpu = (unsigned long)hcpu;
4570 struct work_struct unbind_work;
4571 struct workqueue_struct *wq;
4573 switch (action & ~CPU_TASKS_FROZEN) {
4574 case CPU_DOWN_PREPARE:
4575 /* unbinding per-cpu workers should happen on the local CPU */
4576 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4577 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4579 /* update NUMA affinity of unbound workqueues */
4580 mutex_lock(&wq_pool_mutex);
4581 list_for_each_entry(wq, &workqueues, list)
4582 wq_update_unbound_numa(wq, cpu, false);
4583 mutex_unlock(&wq_pool_mutex);
4585 /* wait for per-cpu unbinding to finish */
4586 flush_work(&unbind_work);
4587 destroy_work_on_stack(&unbind_work);
4595 struct work_for_cpu {
4596 struct work_struct work;
4602 static void work_for_cpu_fn(struct work_struct *work)
4604 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4606 wfc->ret = wfc->fn(wfc->arg);
4610 * work_on_cpu - run a function in user context on a particular cpu
4611 * @cpu: the cpu to run on
4612 * @fn: the function to run
4613 * @arg: the function arg
4615 * It is up to the caller to ensure that the cpu doesn't go offline.
4616 * The caller must not hold any locks which would prevent @fn from completing.
4618 * Return: The value @fn returns.
4620 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4622 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4624 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4625 schedule_work_on(cpu, &wfc.work);
4626 flush_work(&wfc.work);
4627 destroy_work_on_stack(&wfc.work);
4630 EXPORT_SYMBOL_GPL(work_on_cpu);
4631 #endif /* CONFIG_SMP */
4633 #ifdef CONFIG_FREEZER
4636 * freeze_workqueues_begin - begin freezing workqueues
4638 * Start freezing workqueues. After this function returns, all freezable
4639 * workqueues will queue new works to their delayed_works list instead of
4643 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4645 void freeze_workqueues_begin(void)
4647 struct workqueue_struct *wq;
4648 struct pool_workqueue *pwq;
4650 mutex_lock(&wq_pool_mutex);
4652 WARN_ON_ONCE(workqueue_freezing);
4653 workqueue_freezing = true;
4655 list_for_each_entry(wq, &workqueues, list) {
4656 mutex_lock(&wq->mutex);
4657 for_each_pwq(pwq, wq)
4658 pwq_adjust_max_active(pwq);
4659 mutex_unlock(&wq->mutex);
4662 mutex_unlock(&wq_pool_mutex);
4666 * freeze_workqueues_busy - are freezable workqueues still busy?
4668 * Check whether freezing is complete. This function must be called
4669 * between freeze_workqueues_begin() and thaw_workqueues().
4672 * Grabs and releases wq_pool_mutex.
4675 * %true if some freezable workqueues are still busy. %false if freezing
4678 bool freeze_workqueues_busy(void)
4681 struct workqueue_struct *wq;
4682 struct pool_workqueue *pwq;
4684 mutex_lock(&wq_pool_mutex);
4686 WARN_ON_ONCE(!workqueue_freezing);
4688 list_for_each_entry(wq, &workqueues, list) {
4689 if (!(wq->flags & WQ_FREEZABLE))
4692 * nr_active is monotonically decreasing. It's safe
4693 * to peek without lock.
4695 rcu_read_lock_sched();
4696 for_each_pwq(pwq, wq) {
4697 WARN_ON_ONCE(pwq->nr_active < 0);
4698 if (pwq->nr_active) {
4700 rcu_read_unlock_sched();
4704 rcu_read_unlock_sched();
4707 mutex_unlock(&wq_pool_mutex);
4712 * thaw_workqueues - thaw workqueues
4714 * Thaw workqueues. Normal queueing is restored and all collected
4715 * frozen works are transferred to their respective pool worklists.
4718 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4720 void thaw_workqueues(void)
4722 struct workqueue_struct *wq;
4723 struct pool_workqueue *pwq;
4725 mutex_lock(&wq_pool_mutex);
4727 if (!workqueue_freezing)
4730 workqueue_freezing = false;
4732 /* restore max_active and repopulate worklist */
4733 list_for_each_entry(wq, &workqueues, list) {
4734 mutex_lock(&wq->mutex);
4735 for_each_pwq(pwq, wq)
4736 pwq_adjust_max_active(pwq);
4737 mutex_unlock(&wq->mutex);
4741 mutex_unlock(&wq_pool_mutex);
4743 #endif /* CONFIG_FREEZER */
4745 static int workqueue_apply_unbound_cpumask(void)
4749 struct workqueue_struct *wq;
4750 struct apply_wqattrs_ctx *ctx, *n;
4752 lockdep_assert_held(&wq_pool_mutex);
4754 list_for_each_entry(wq, &workqueues, list) {
4755 if (!(wq->flags & WQ_UNBOUND))
4757 /* creating multiple pwqs breaks ordering guarantee */
4758 if (wq->flags & __WQ_ORDERED)
4761 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4767 list_add_tail(&ctx->list, &ctxs);
4770 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4772 apply_wqattrs_commit(ctx);
4773 apply_wqattrs_cleanup(ctx);
4780 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4781 * @cpumask: the cpumask to set
4783 * The low-level workqueues cpumask is a global cpumask that limits
4784 * the affinity of all unbound workqueues. This function check the @cpumask
4785 * and apply it to all unbound workqueues and updates all pwqs of them.
4787 * Retun: 0 - Success
4788 * -EINVAL - Invalid @cpumask
4789 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4791 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4794 cpumask_var_t saved_cpumask;
4796 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4800 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4801 if (!cpumask_empty(cpumask)) {
4802 mutex_lock(&wq_pool_mutex);
4804 /* save the old wq_unbound_cpumask. */
4805 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4807 /* update wq_unbound_cpumask at first and apply it to wqs. */
4808 cpumask_copy(wq_unbound_cpumask, cpumask);
4809 ret = workqueue_apply_unbound_cpumask();
4811 /* restore the wq_unbound_cpumask when failed. */
4813 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4815 mutex_unlock(&wq_pool_mutex);
4819 free_cpumask_var(saved_cpumask);
4825 * Workqueues with WQ_SYSFS flag set is visible to userland via
4826 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4827 * following attributes.
4829 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4830 * max_active RW int : maximum number of in-flight work items
4832 * Unbound workqueues have the following extra attributes.
4834 * id RO int : the associated pool ID
4835 * nice RW int : nice value of the workers
4836 * cpumask RW mask : bitmask of allowed CPUs for the workers
4839 struct workqueue_struct *wq;
4843 static struct workqueue_struct *dev_to_wq(struct device *dev)
4845 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4850 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4853 struct workqueue_struct *wq = dev_to_wq(dev);
4855 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4857 static DEVICE_ATTR_RO(per_cpu);
4859 static ssize_t max_active_show(struct device *dev,
4860 struct device_attribute *attr, char *buf)
4862 struct workqueue_struct *wq = dev_to_wq(dev);
4864 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4867 static ssize_t max_active_store(struct device *dev,
4868 struct device_attribute *attr, const char *buf,
4871 struct workqueue_struct *wq = dev_to_wq(dev);
4874 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
4877 workqueue_set_max_active(wq, val);
4880 static DEVICE_ATTR_RW(max_active);
4882 static struct attribute *wq_sysfs_attrs[] = {
4883 &dev_attr_per_cpu.attr,
4884 &dev_attr_max_active.attr,
4887 ATTRIBUTE_GROUPS(wq_sysfs);
4889 static ssize_t wq_pool_ids_show(struct device *dev,
4890 struct device_attribute *attr, char *buf)
4892 struct workqueue_struct *wq = dev_to_wq(dev);
4893 const char *delim = "";
4894 int node, written = 0;
4896 rcu_read_lock_sched();
4897 for_each_node(node) {
4898 written += scnprintf(buf + written, PAGE_SIZE - written,
4899 "%s%d:%d", delim, node,
4900 unbound_pwq_by_node(wq, node)->pool->id);
4903 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
4904 rcu_read_unlock_sched();
4909 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
4912 struct workqueue_struct *wq = dev_to_wq(dev);
4915 mutex_lock(&wq->mutex);
4916 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
4917 mutex_unlock(&wq->mutex);
4922 /* prepare workqueue_attrs for sysfs store operations */
4923 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
4925 struct workqueue_attrs *attrs;
4927 attrs = alloc_workqueue_attrs(GFP_KERNEL);
4931 mutex_lock(&wq->mutex);
4932 copy_workqueue_attrs(attrs, wq->unbound_attrs);
4933 mutex_unlock(&wq->mutex);
4937 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
4938 const char *buf, size_t count)
4940 struct workqueue_struct *wq = dev_to_wq(dev);
4941 struct workqueue_attrs *attrs;
4944 attrs = wq_sysfs_prep_attrs(wq);
4948 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
4949 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
4950 ret = apply_workqueue_attrs(wq, attrs);
4954 free_workqueue_attrs(attrs);
4955 return ret ?: count;
4958 static ssize_t wq_cpumask_show(struct device *dev,
4959 struct device_attribute *attr, char *buf)
4961 struct workqueue_struct *wq = dev_to_wq(dev);
4964 mutex_lock(&wq->mutex);
4965 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
4966 cpumask_pr_args(wq->unbound_attrs->cpumask));
4967 mutex_unlock(&wq->mutex);
4971 static ssize_t wq_cpumask_store(struct device *dev,
4972 struct device_attribute *attr,
4973 const char *buf, size_t count)
4975 struct workqueue_struct *wq = dev_to_wq(dev);
4976 struct workqueue_attrs *attrs;
4979 attrs = wq_sysfs_prep_attrs(wq);
4983 ret = cpumask_parse(buf, attrs->cpumask);
4985 ret = apply_workqueue_attrs(wq, attrs);
4987 free_workqueue_attrs(attrs);
4988 return ret ?: count;
4991 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
4994 struct workqueue_struct *wq = dev_to_wq(dev);
4997 mutex_lock(&wq->mutex);
4998 written = scnprintf(buf, PAGE_SIZE, "%d\n",
4999 !wq->unbound_attrs->no_numa);
5000 mutex_unlock(&wq->mutex);
5005 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5006 const char *buf, size_t count)
5008 struct workqueue_struct *wq = dev_to_wq(dev);
5009 struct workqueue_attrs *attrs;
5012 attrs = wq_sysfs_prep_attrs(wq);
5017 if (sscanf(buf, "%d", &v) == 1) {
5018 attrs->no_numa = !v;
5019 ret = apply_workqueue_attrs(wq, attrs);
5022 free_workqueue_attrs(attrs);
5023 return ret ?: count;
5026 static struct device_attribute wq_sysfs_unbound_attrs[] = {
5027 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5028 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5029 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5030 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5034 static struct bus_type wq_subsys = {
5035 .name = "workqueue",
5036 .dev_groups = wq_sysfs_groups,
5039 static ssize_t wq_unbound_cpumask_show(struct device *dev,
5040 struct device_attribute *attr, char *buf)
5044 mutex_lock(&wq_pool_mutex);
5045 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5046 cpumask_pr_args(wq_unbound_cpumask));
5047 mutex_unlock(&wq_pool_mutex);
5052 static ssize_t wq_unbound_cpumask_store(struct device *dev,
5053 struct device_attribute *attr, const char *buf, size_t count)
5055 cpumask_var_t cpumask;
5058 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5061 ret = cpumask_parse(buf, cpumask);
5063 ret = workqueue_set_unbound_cpumask(cpumask);
5065 free_cpumask_var(cpumask);
5066 return ret ? ret : count;
5069 static struct device_attribute wq_sysfs_cpumask_attr =
5070 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5071 wq_unbound_cpumask_store);
5073 static int __init wq_sysfs_init(void)
5077 err = subsys_virtual_register(&wq_subsys, NULL);
5081 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5083 core_initcall(wq_sysfs_init);
5085 static void wq_device_release(struct device *dev)
5087 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5093 * workqueue_sysfs_register - make a workqueue visible in sysfs
5094 * @wq: the workqueue to register
5096 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5097 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5098 * which is the preferred method.
5100 * Workqueue user should use this function directly iff it wants to apply
5101 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5102 * apply_workqueue_attrs() may race against userland updating the
5105 * Return: 0 on success, -errno on failure.
5107 int workqueue_sysfs_register(struct workqueue_struct *wq)
5109 struct wq_device *wq_dev;
5113 * Adjusting max_active or creating new pwqs by applyting
5114 * attributes breaks ordering guarantee. Disallow exposing ordered
5117 if (WARN_ON(wq->flags & __WQ_ORDERED))
5120 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5125 wq_dev->dev.bus = &wq_subsys;
5126 wq_dev->dev.init_name = wq->name;
5127 wq_dev->dev.release = wq_device_release;
5130 * unbound_attrs are created separately. Suppress uevent until
5131 * everything is ready.
5133 dev_set_uevent_suppress(&wq_dev->dev, true);
5135 ret = device_register(&wq_dev->dev);
5142 if (wq->flags & WQ_UNBOUND) {
5143 struct device_attribute *attr;
5145 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5146 ret = device_create_file(&wq_dev->dev, attr);
5148 device_unregister(&wq_dev->dev);
5155 dev_set_uevent_suppress(&wq_dev->dev, false);
5156 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5161 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5162 * @wq: the workqueue to unregister
5164 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5166 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5168 struct wq_device *wq_dev = wq->wq_dev;
5174 device_unregister(&wq_dev->dev);
5176 #else /* CONFIG_SYSFS */
5177 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5178 #endif /* CONFIG_SYSFS */
5180 static void __init wq_numa_init(void)
5185 if (num_possible_nodes() <= 1)
5188 if (wq_disable_numa) {
5189 pr_info("workqueue: NUMA affinity support disabled\n");
5193 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5194 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5197 * We want masks of possible CPUs of each node which isn't readily
5198 * available. Build one from cpu_to_node() which should have been
5199 * fully initialized by now.
5201 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5205 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5206 node_online(node) ? node : NUMA_NO_NODE));
5208 for_each_possible_cpu(cpu) {
5209 node = cpu_to_node(cpu);
5210 if (WARN_ON(node == NUMA_NO_NODE)) {
5211 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5212 /* happens iff arch is bonkers, let's just proceed */
5215 cpumask_set_cpu(cpu, tbl[node]);
5218 wq_numa_possible_cpumask = tbl;
5219 wq_numa_enabled = true;
5222 static int __init init_workqueues(void)
5224 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5227 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5229 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5230 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5232 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5234 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5235 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5239 /* initialize CPU pools */
5240 for_each_possible_cpu(cpu) {
5241 struct worker_pool *pool;
5244 for_each_cpu_worker_pool(pool, cpu) {
5245 BUG_ON(init_worker_pool(pool));
5247 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5248 pool->attrs->nice = std_nice[i++];
5249 pool->node = cpu_to_node(cpu);
5252 mutex_lock(&wq_pool_mutex);
5253 BUG_ON(worker_pool_assign_id(pool));
5254 mutex_unlock(&wq_pool_mutex);
5258 /* create the initial worker */
5259 for_each_online_cpu(cpu) {
5260 struct worker_pool *pool;
5262 for_each_cpu_worker_pool(pool, cpu) {
5263 pool->flags &= ~POOL_DISASSOCIATED;
5264 BUG_ON(!create_worker(pool));
5268 /* create default unbound and ordered wq attrs */
5269 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5270 struct workqueue_attrs *attrs;
5272 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5273 attrs->nice = std_nice[i];
5274 unbound_std_wq_attrs[i] = attrs;
5277 * An ordered wq should have only one pwq as ordering is
5278 * guaranteed by max_active which is enforced by pwqs.
5279 * Turn off NUMA so that dfl_pwq is used for all nodes.
5281 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5282 attrs->nice = std_nice[i];
5283 attrs->no_numa = true;
5284 ordered_wq_attrs[i] = attrs;
5287 system_wq = alloc_workqueue("events", 0, 0);
5288 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5289 system_long_wq = alloc_workqueue("events_long", 0, 0);
5290 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5291 WQ_UNBOUND_MAX_ACTIVE);
5292 system_freezable_wq = alloc_workqueue("events_freezable",
5294 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5295 WQ_POWER_EFFICIENT, 0);
5296 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5297 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5299 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5300 !system_unbound_wq || !system_freezable_wq ||
5301 !system_power_efficient_wq ||
5302 !system_freezable_power_efficient_wq);
5305 early_initcall(init_workqueues);