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 is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/jhash.h>
45 #include <linux/hashtable.h>
46 #include <linux/rculist.h>
47 #include <linux/nodemask.h>
49 #include "workqueue_internal.h"
55 * A bound pool is either associated or disassociated with its CPU.
56 * While associated (!DISASSOCIATED), all workers are bound to the
57 * CPU and none has %WORKER_UNBOUND set and concurrency management
60 * While DISASSOCIATED, the cpu may be offline and all workers have
61 * %WORKER_UNBOUND set and concurrency management disabled, and may
62 * be executing on any CPU. The pool behaves as an unbound one.
64 * Note that DISASSOCIATED should be flipped only while holding
65 * manager_mutex to avoid changing binding state while
66 * create_worker() is in progress.
68 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
69 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
70 POOL_FREEZING = 1 << 3, /* freeze in progress */
73 WORKER_STARTED = 1 << 0, /* started */
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 -20.
102 RESCUER_NICE_LEVEL = -20,
103 HIGHPRI_NICE_LEVEL = -20,
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 * MG: pool->manager_mutex and pool->lock protected. Writes require both
125 * locks. Reads can happen under either lock.
127 * PL: wq_pool_mutex protected.
129 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
131 * WQ: wq->mutex protected.
133 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
135 * MD: wq_mayday_lock protected.
138 /* struct worker is defined in workqueue_internal.h */
141 spinlock_t lock; /* the pool lock */
142 int cpu; /* I: the associated cpu */
143 int node; /* I: the associated node ID */
144 int id; /* I: pool ID */
145 unsigned int flags; /* X: flags */
147 struct list_head worklist; /* L: list of pending works */
148 int nr_workers; /* L: total number of workers */
150 /* nr_idle includes the ones off idle_list for rebinding */
151 int nr_idle; /* L: currently idle ones */
153 struct list_head idle_list; /* X: list of idle workers */
154 struct timer_list idle_timer; /* L: worker idle timeout */
155 struct timer_list mayday_timer; /* L: SOS timer for workers */
157 /* a workers is either on busy_hash or idle_list, or the manager */
158 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
159 /* L: hash of busy workers */
161 /* see manage_workers() for details on the two manager mutexes */
162 struct mutex manager_arb; /* manager arbitration */
163 struct mutex manager_mutex; /* manager exclusion */
164 struct idr worker_idr; /* MG: worker IDs and iteration */
166 struct workqueue_attrs *attrs; /* I: worker attributes */
167 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
168 int refcnt; /* PL: refcnt for unbound pools */
171 * The current concurrency level. As it's likely to be accessed
172 * from other CPUs during try_to_wake_up(), put it in a separate
175 atomic_t nr_running ____cacheline_aligned_in_smp;
178 * Destruction of pool is sched-RCU protected to allow dereferences
179 * from get_work_pool().
182 } ____cacheline_aligned_in_smp;
185 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
186 * of work_struct->data are used for flags and the remaining high bits
187 * point to the pwq; thus, pwqs need to be aligned at two's power of the
188 * number of flag bits.
190 struct pool_workqueue {
191 struct worker_pool *pool; /* I: the associated pool */
192 struct workqueue_struct *wq; /* I: the owning workqueue */
193 int work_color; /* L: current color */
194 int flush_color; /* L: flushing color */
195 int refcnt; /* L: reference count */
196 int nr_in_flight[WORK_NR_COLORS];
197 /* L: nr of in_flight works */
198 int nr_active; /* L: nr of active works */
199 int max_active; /* L: max active works */
200 struct list_head delayed_works; /* L: delayed works */
201 struct list_head pwqs_node; /* WR: node on wq->pwqs */
202 struct list_head mayday_node; /* MD: node on wq->maydays */
205 * Release of unbound pwq is punted to system_wq. See put_pwq()
206 * and pwq_unbound_release_workfn() for details. pool_workqueue
207 * itself is also sched-RCU protected so that the first pwq can be
208 * determined without grabbing wq->mutex.
210 struct work_struct unbound_release_work;
212 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
215 * Structure used to wait for workqueue flush.
218 struct list_head list; /* WQ: list of flushers */
219 int flush_color; /* WQ: flush color waiting for */
220 struct completion done; /* flush completion */
226 * The externally visible workqueue. It relays the issued work items to
227 * the appropriate worker_pool through its pool_workqueues.
229 struct workqueue_struct {
230 struct list_head pwqs; /* WR: all pwqs of this wq */
231 struct list_head list; /* PL: list of all workqueues */
233 struct mutex mutex; /* protects this wq */
234 int work_color; /* WQ: current work color */
235 int flush_color; /* WQ: current flush color */
236 atomic_t nr_pwqs_to_flush; /* flush in progress */
237 struct wq_flusher *first_flusher; /* WQ: first flusher */
238 struct list_head flusher_queue; /* WQ: flush waiters */
239 struct list_head flusher_overflow; /* WQ: flush overflow list */
241 struct list_head maydays; /* MD: pwqs requesting rescue */
242 struct worker *rescuer; /* I: rescue worker */
244 int nr_drainers; /* WQ: drain in progress */
245 int saved_max_active; /* WQ: saved pwq max_active */
247 struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
250 struct wq_device *wq_dev; /* I: for sysfs interface */
252 #ifdef CONFIG_LOCKDEP
253 struct lockdep_map lockdep_map;
255 char name[WQ_NAME_LEN]; /* I: workqueue name */
257 /* hot fields used during command issue, aligned to cacheline */
258 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
259 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
262 static struct kmem_cache *pwq_cache;
264 static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
265 static cpumask_var_t *wq_numa_possible_cpumask;
266 /* possible CPUs of each node */
268 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
270 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
271 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
273 static LIST_HEAD(workqueues); /* PL: list of all workqueues */
274 static bool workqueue_freezing; /* PL: have wqs started freezing? */
276 /* the per-cpu worker pools */
277 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
280 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
282 /* PL: hash of all unbound pools keyed by pool->attrs */
283 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
285 /* I: attributes used when instantiating standard unbound pools on demand */
286 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
288 struct workqueue_struct *system_wq __read_mostly;
289 EXPORT_SYMBOL_GPL(system_wq);
290 struct workqueue_struct *system_highpri_wq __read_mostly;
291 EXPORT_SYMBOL_GPL(system_highpri_wq);
292 struct workqueue_struct *system_long_wq __read_mostly;
293 EXPORT_SYMBOL_GPL(system_long_wq);
294 struct workqueue_struct *system_unbound_wq __read_mostly;
295 EXPORT_SYMBOL_GPL(system_unbound_wq);
296 struct workqueue_struct *system_freezable_wq __read_mostly;
297 EXPORT_SYMBOL_GPL(system_freezable_wq);
299 static int worker_thread(void *__worker);
300 static void copy_workqueue_attrs(struct workqueue_attrs *to,
301 const struct workqueue_attrs *from);
303 #define CREATE_TRACE_POINTS
304 #include <trace/events/workqueue.h>
306 #define assert_rcu_or_pool_mutex() \
307 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
308 lockdep_is_held(&wq_pool_mutex), \
309 "sched RCU or wq_pool_mutex should be held")
311 #define assert_rcu_or_wq_mutex(wq) \
312 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
313 lockdep_is_held(&wq->mutex), \
314 "sched RCU or wq->mutex should be held")
316 #ifdef CONFIG_LOCKDEP
317 #define assert_manager_or_pool_lock(pool) \
318 WARN_ONCE(debug_locks && \
319 !lockdep_is_held(&(pool)->manager_mutex) && \
320 !lockdep_is_held(&(pool)->lock), \
321 "pool->manager_mutex or ->lock should be held")
323 #define assert_manager_or_pool_lock(pool) do { } while (0)
326 #define for_each_cpu_worker_pool(pool, cpu) \
327 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
328 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
332 * for_each_pool - iterate through all worker_pools in the system
333 * @pool: iteration cursor
334 * @pi: integer used for iteration
336 * This must be called either with wq_pool_mutex held or sched RCU read
337 * locked. If the pool needs to be used beyond the locking in effect, the
338 * caller is responsible for guaranteeing that the pool stays online.
340 * The if/else clause exists only for the lockdep assertion and can be
343 #define for_each_pool(pool, pi) \
344 idr_for_each_entry(&worker_pool_idr, pool, pi) \
345 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
349 * for_each_pool_worker - iterate through all workers of a worker_pool
350 * @worker: iteration cursor
351 * @wi: integer used for iteration
352 * @pool: worker_pool to iterate workers of
354 * This must be called with either @pool->manager_mutex or ->lock held.
356 * The if/else clause exists only for the lockdep assertion and can be
359 #define for_each_pool_worker(worker, wi, pool) \
360 idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
361 if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
365 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
366 * @pwq: iteration cursor
367 * @wq: the target workqueue
369 * This must be called either with wq->mutex held or sched RCU read locked.
370 * If the pwq needs to be used beyond the locking in effect, the caller is
371 * responsible for guaranteeing that the pwq stays online.
373 * The if/else clause exists only for the lockdep assertion and can be
376 #define for_each_pwq(pwq, wq) \
377 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
378 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
381 #ifdef CONFIG_DEBUG_OBJECTS_WORK
383 static struct debug_obj_descr work_debug_descr;
385 static void *work_debug_hint(void *addr)
387 return ((struct work_struct *) addr)->func;
391 * fixup_init is called when:
392 * - an active object is initialized
394 static int work_fixup_init(void *addr, enum debug_obj_state state)
396 struct work_struct *work = addr;
399 case ODEBUG_STATE_ACTIVE:
400 cancel_work_sync(work);
401 debug_object_init(work, &work_debug_descr);
409 * fixup_activate is called when:
410 * - an active object is activated
411 * - an unknown object is activated (might be a statically initialized object)
413 static int work_fixup_activate(void *addr, enum debug_obj_state state)
415 struct work_struct *work = addr;
419 case ODEBUG_STATE_NOTAVAILABLE:
421 * This is not really a fixup. The work struct was
422 * statically initialized. We just make sure that it
423 * is tracked in the object tracker.
425 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
426 debug_object_init(work, &work_debug_descr);
427 debug_object_activate(work, &work_debug_descr);
433 case ODEBUG_STATE_ACTIVE:
442 * fixup_free is called when:
443 * - an active object is freed
445 static int work_fixup_free(void *addr, enum debug_obj_state state)
447 struct work_struct *work = addr;
450 case ODEBUG_STATE_ACTIVE:
451 cancel_work_sync(work);
452 debug_object_free(work, &work_debug_descr);
459 static struct debug_obj_descr work_debug_descr = {
460 .name = "work_struct",
461 .debug_hint = work_debug_hint,
462 .fixup_init = work_fixup_init,
463 .fixup_activate = work_fixup_activate,
464 .fixup_free = work_fixup_free,
467 static inline void debug_work_activate(struct work_struct *work)
469 debug_object_activate(work, &work_debug_descr);
472 static inline void debug_work_deactivate(struct work_struct *work)
474 debug_object_deactivate(work, &work_debug_descr);
477 void __init_work(struct work_struct *work, int onstack)
480 debug_object_init_on_stack(work, &work_debug_descr);
482 debug_object_init(work, &work_debug_descr);
484 EXPORT_SYMBOL_GPL(__init_work);
486 void destroy_work_on_stack(struct work_struct *work)
488 debug_object_free(work, &work_debug_descr);
490 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
493 static inline void debug_work_activate(struct work_struct *work) { }
494 static inline void debug_work_deactivate(struct work_struct *work) { }
497 /* allocate ID and assign it to @pool */
498 static int worker_pool_assign_id(struct worker_pool *pool)
502 lockdep_assert_held(&wq_pool_mutex);
505 if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL))
507 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
508 } while (ret == -EAGAIN);
514 * first_pwq - return the first pool_workqueue of the specified workqueue
515 * @wq: the target workqueue
517 * This must be called either with wq->mutex held or sched RCU read locked.
518 * If the pwq needs to be used beyond the locking in effect, the caller is
519 * responsible for guaranteeing that the pwq stays online.
521 static struct pool_workqueue *first_pwq(struct workqueue_struct *wq)
523 assert_rcu_or_wq_mutex(wq);
524 return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue,
528 static unsigned int work_color_to_flags(int color)
530 return color << WORK_STRUCT_COLOR_SHIFT;
533 static int get_work_color(struct work_struct *work)
535 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
536 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
539 static int work_next_color(int color)
541 return (color + 1) % WORK_NR_COLORS;
545 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
546 * contain the pointer to the queued pwq. Once execution starts, the flag
547 * is cleared and the high bits contain OFFQ flags and pool ID.
549 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
550 * and clear_work_data() can be used to set the pwq, pool or clear
551 * work->data. These functions should only be called while the work is
552 * owned - ie. while the PENDING bit is set.
554 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
555 * corresponding to a work. Pool is available once the work has been
556 * queued anywhere after initialization until it is sync canceled. pwq is
557 * available only while the work item is queued.
559 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
560 * canceled. While being canceled, a work item may have its PENDING set
561 * but stay off timer and worklist for arbitrarily long and nobody should
562 * try to steal the PENDING bit.
564 static inline void set_work_data(struct work_struct *work, unsigned long data,
567 WARN_ON_ONCE(!work_pending(work));
568 atomic_long_set(&work->data, data | flags | work_static(work));
571 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
572 unsigned long extra_flags)
574 set_work_data(work, (unsigned long)pwq,
575 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
578 static void set_work_pool_and_keep_pending(struct work_struct *work,
581 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
582 WORK_STRUCT_PENDING);
585 static void set_work_pool_and_clear_pending(struct work_struct *work,
589 * The following wmb is paired with the implied mb in
590 * test_and_set_bit(PENDING) and ensures all updates to @work made
591 * here are visible to and precede any updates by the next PENDING
595 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
598 static void clear_work_data(struct work_struct *work)
600 smp_wmb(); /* see set_work_pool_and_clear_pending() */
601 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
604 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
606 unsigned long data = atomic_long_read(&work->data);
608 if (data & WORK_STRUCT_PWQ)
609 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
615 * get_work_pool - return the worker_pool a given work was associated with
616 * @work: the work item of interest
618 * Return the worker_pool @work was last associated with. %NULL if none.
620 * Pools are created and destroyed under wq_pool_mutex, and allows read
621 * access under sched-RCU read lock. As such, this function should be
622 * called under wq_pool_mutex or with preemption disabled.
624 * All fields of the returned pool are accessible as long as the above
625 * mentioned locking is in effect. If the returned pool needs to be used
626 * beyond the critical section, the caller is responsible for ensuring the
627 * returned pool is and stays online.
629 static struct worker_pool *get_work_pool(struct work_struct *work)
631 unsigned long data = atomic_long_read(&work->data);
634 assert_rcu_or_pool_mutex();
636 if (data & WORK_STRUCT_PWQ)
637 return ((struct pool_workqueue *)
638 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
640 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
641 if (pool_id == WORK_OFFQ_POOL_NONE)
644 return idr_find(&worker_pool_idr, pool_id);
648 * get_work_pool_id - return the worker pool ID a given work is associated with
649 * @work: the work item of interest
651 * Return the worker_pool ID @work was last associated with.
652 * %WORK_OFFQ_POOL_NONE if none.
654 static int get_work_pool_id(struct work_struct *work)
656 unsigned long data = atomic_long_read(&work->data);
658 if (data & WORK_STRUCT_PWQ)
659 return ((struct pool_workqueue *)
660 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
662 return data >> WORK_OFFQ_POOL_SHIFT;
665 static void mark_work_canceling(struct work_struct *work)
667 unsigned long pool_id = get_work_pool_id(work);
669 pool_id <<= WORK_OFFQ_POOL_SHIFT;
670 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
673 static bool work_is_canceling(struct work_struct *work)
675 unsigned long data = atomic_long_read(&work->data);
677 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
681 * Policy functions. These define the policies on how the global worker
682 * pools are managed. Unless noted otherwise, these functions assume that
683 * they're being called with pool->lock held.
686 static bool __need_more_worker(struct worker_pool *pool)
688 return !atomic_read(&pool->nr_running);
692 * Need to wake up a worker? Called from anything but currently
695 * Note that, because unbound workers never contribute to nr_running, this
696 * function will always return %true for unbound pools as long as the
697 * worklist isn't empty.
699 static bool need_more_worker(struct worker_pool *pool)
701 return !list_empty(&pool->worklist) && __need_more_worker(pool);
704 /* Can I start working? Called from busy but !running workers. */
705 static bool may_start_working(struct worker_pool *pool)
707 return pool->nr_idle;
710 /* Do I need to keep working? Called from currently running workers. */
711 static bool keep_working(struct worker_pool *pool)
713 return !list_empty(&pool->worklist) &&
714 atomic_read(&pool->nr_running) <= 1;
717 /* Do we need a new worker? Called from manager. */
718 static bool need_to_create_worker(struct worker_pool *pool)
720 return need_more_worker(pool) && !may_start_working(pool);
723 /* Do I need to be the manager? */
724 static bool need_to_manage_workers(struct worker_pool *pool)
726 return need_to_create_worker(pool) ||
727 (pool->flags & POOL_MANAGE_WORKERS);
730 /* Do we have too many workers and should some go away? */
731 static bool too_many_workers(struct worker_pool *pool)
733 bool managing = mutex_is_locked(&pool->manager_arb);
734 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
735 int nr_busy = pool->nr_workers - nr_idle;
738 * nr_idle and idle_list may disagree if idle rebinding is in
739 * progress. Never return %true if idle_list is empty.
741 if (list_empty(&pool->idle_list))
744 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
751 /* Return the first worker. Safe with preemption disabled */
752 static struct worker *first_worker(struct worker_pool *pool)
754 if (unlikely(list_empty(&pool->idle_list)))
757 return list_first_entry(&pool->idle_list, struct worker, entry);
761 * wake_up_worker - wake up an idle worker
762 * @pool: worker pool to wake worker from
764 * Wake up the first idle worker of @pool.
767 * spin_lock_irq(pool->lock).
769 static void wake_up_worker(struct worker_pool *pool)
771 struct worker *worker = first_worker(pool);
774 wake_up_process(worker->task);
778 * wq_worker_waking_up - a worker is waking up
779 * @task: task waking up
780 * @cpu: CPU @task is waking up to
782 * This function is called during try_to_wake_up() when a worker is
786 * spin_lock_irq(rq->lock)
788 void wq_worker_waking_up(struct task_struct *task, int cpu)
790 struct worker *worker = kthread_data(task);
792 if (!(worker->flags & WORKER_NOT_RUNNING)) {
793 WARN_ON_ONCE(worker->pool->cpu != cpu);
794 atomic_inc(&worker->pool->nr_running);
799 * wq_worker_sleeping - a worker is going to sleep
800 * @task: task going to sleep
801 * @cpu: CPU in question, must be the current CPU number
803 * This function is called during schedule() when a busy worker is
804 * going to sleep. Worker on the same cpu can be woken up by
805 * returning pointer to its task.
808 * spin_lock_irq(rq->lock)
811 * Worker task on @cpu to wake up, %NULL if none.
813 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
815 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
816 struct worker_pool *pool;
819 * Rescuers, which may not have all the fields set up like normal
820 * workers, also reach here, let's not access anything before
821 * checking NOT_RUNNING.
823 if (worker->flags & WORKER_NOT_RUNNING)
828 /* this can only happen on the local cpu */
829 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
833 * The counterpart of the following dec_and_test, implied mb,
834 * worklist not empty test sequence is in insert_work().
835 * Please read comment there.
837 * NOT_RUNNING is clear. This means that we're bound to and
838 * running on the local cpu w/ rq lock held and preemption
839 * disabled, which in turn means that none else could be
840 * manipulating idle_list, so dereferencing idle_list without pool
843 if (atomic_dec_and_test(&pool->nr_running) &&
844 !list_empty(&pool->worklist))
845 to_wakeup = first_worker(pool);
846 return to_wakeup ? to_wakeup->task : NULL;
850 * worker_set_flags - set worker flags and adjust nr_running accordingly
852 * @flags: flags to set
853 * @wakeup: wakeup an idle worker if necessary
855 * Set @flags in @worker->flags and adjust nr_running accordingly. If
856 * nr_running becomes zero and @wakeup is %true, an idle worker is
860 * spin_lock_irq(pool->lock)
862 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
865 struct worker_pool *pool = worker->pool;
867 WARN_ON_ONCE(worker->task != current);
870 * If transitioning into NOT_RUNNING, adjust nr_running and
871 * wake up an idle worker as necessary if requested by
874 if ((flags & WORKER_NOT_RUNNING) &&
875 !(worker->flags & WORKER_NOT_RUNNING)) {
877 if (atomic_dec_and_test(&pool->nr_running) &&
878 !list_empty(&pool->worklist))
879 wake_up_worker(pool);
881 atomic_dec(&pool->nr_running);
884 worker->flags |= flags;
888 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
890 * @flags: flags to clear
892 * Clear @flags in @worker->flags and adjust nr_running accordingly.
895 * spin_lock_irq(pool->lock)
897 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
899 struct worker_pool *pool = worker->pool;
900 unsigned int oflags = worker->flags;
902 WARN_ON_ONCE(worker->task != current);
904 worker->flags &= ~flags;
907 * If transitioning out of NOT_RUNNING, increment nr_running. Note
908 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
909 * of multiple flags, not a single flag.
911 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
912 if (!(worker->flags & WORKER_NOT_RUNNING))
913 atomic_inc(&pool->nr_running);
917 * find_worker_executing_work - find worker which is executing a work
918 * @pool: pool of interest
919 * @work: work to find worker for
921 * Find a worker which is executing @work on @pool by searching
922 * @pool->busy_hash which is keyed by the address of @work. For a worker
923 * to match, its current execution should match the address of @work and
924 * its work function. This is to avoid unwanted dependency between
925 * unrelated work executions through a work item being recycled while still
928 * This is a bit tricky. A work item may be freed once its execution
929 * starts and nothing prevents the freed area from being recycled for
930 * another work item. If the same work item address ends up being reused
931 * before the original execution finishes, workqueue will identify the
932 * recycled work item as currently executing and make it wait until the
933 * current execution finishes, introducing an unwanted dependency.
935 * This function checks the work item address and work function to avoid
936 * false positives. Note that this isn't complete as one may construct a
937 * work function which can introduce dependency onto itself through a
938 * recycled work item. Well, if somebody wants to shoot oneself in the
939 * foot that badly, there's only so much we can do, and if such deadlock
940 * actually occurs, it should be easy to locate the culprit work function.
943 * spin_lock_irq(pool->lock).
946 * Pointer to worker which is executing @work if found, NULL
949 static struct worker *find_worker_executing_work(struct worker_pool *pool,
950 struct work_struct *work)
952 struct worker *worker;
954 hash_for_each_possible(pool->busy_hash, worker, hentry,
956 if (worker->current_work == work &&
957 worker->current_func == work->func)
964 * move_linked_works - move linked works to a list
965 * @work: start of series of works to be scheduled
966 * @head: target list to append @work to
967 * @nextp: out paramter for nested worklist walking
969 * Schedule linked works starting from @work to @head. Work series to
970 * be scheduled starts at @work and includes any consecutive work with
971 * WORK_STRUCT_LINKED set in its predecessor.
973 * If @nextp is not NULL, it's updated to point to the next work of
974 * the last scheduled work. This allows move_linked_works() to be
975 * nested inside outer list_for_each_entry_safe().
978 * spin_lock_irq(pool->lock).
980 static void move_linked_works(struct work_struct *work, struct list_head *head,
981 struct work_struct **nextp)
983 struct work_struct *n;
986 * Linked worklist will always end before the end of the list,
987 * use NULL for list head.
989 list_for_each_entry_safe_from(work, n, NULL, entry) {
990 list_move_tail(&work->entry, head);
991 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
996 * If we're already inside safe list traversal and have moved
997 * multiple works to the scheduled queue, the next position
998 * needs to be updated.
1005 * get_pwq - get an extra reference on the specified pool_workqueue
1006 * @pwq: pool_workqueue to get
1008 * Obtain an extra reference on @pwq. The caller should guarantee that
1009 * @pwq has positive refcnt and be holding the matching pool->lock.
1011 static void get_pwq(struct pool_workqueue *pwq)
1013 lockdep_assert_held(&pwq->pool->lock);
1014 WARN_ON_ONCE(pwq->refcnt <= 0);
1019 * put_pwq - put a pool_workqueue reference
1020 * @pwq: pool_workqueue to put
1022 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1023 * destruction. The caller should be holding the matching pool->lock.
1025 static void put_pwq(struct pool_workqueue *pwq)
1027 lockdep_assert_held(&pwq->pool->lock);
1028 if (likely(--pwq->refcnt))
1030 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1033 * @pwq can't be released under pool->lock, bounce to
1034 * pwq_unbound_release_workfn(). This never recurses on the same
1035 * pool->lock as this path is taken only for unbound workqueues and
1036 * the release work item is scheduled on a per-cpu workqueue. To
1037 * avoid lockdep warning, unbound pool->locks are given lockdep
1038 * subclass of 1 in get_unbound_pool().
1040 schedule_work(&pwq->unbound_release_work);
1043 static void pwq_activate_delayed_work(struct work_struct *work)
1045 struct pool_workqueue *pwq = get_work_pwq(work);
1047 trace_workqueue_activate_work(work);
1048 move_linked_works(work, &pwq->pool->worklist, NULL);
1049 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1053 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1055 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1056 struct work_struct, entry);
1058 pwq_activate_delayed_work(work);
1062 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1063 * @pwq: pwq of interest
1064 * @color: color of work which left the queue
1066 * A work either has completed or is removed from pending queue,
1067 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1070 * spin_lock_irq(pool->lock).
1072 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1074 /* uncolored work items don't participate in flushing or nr_active */
1075 if (color == WORK_NO_COLOR)
1078 pwq->nr_in_flight[color]--;
1081 if (!list_empty(&pwq->delayed_works)) {
1082 /* one down, submit a delayed one */
1083 if (pwq->nr_active < pwq->max_active)
1084 pwq_activate_first_delayed(pwq);
1087 /* is flush in progress and are we at the flushing tip? */
1088 if (likely(pwq->flush_color != color))
1091 /* are there still in-flight works? */
1092 if (pwq->nr_in_flight[color])
1095 /* this pwq is done, clear flush_color */
1096 pwq->flush_color = -1;
1099 * If this was the last pwq, wake up the first flusher. It
1100 * will handle the rest.
1102 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1103 complete(&pwq->wq->first_flusher->done);
1109 * try_to_grab_pending - steal work item from worklist and disable irq
1110 * @work: work item to steal
1111 * @is_dwork: @work is a delayed_work
1112 * @flags: place to store irq state
1114 * Try to grab PENDING bit of @work. This function can handle @work in any
1115 * stable state - idle, on timer or on worklist. Return values are
1117 * 1 if @work was pending and we successfully stole PENDING
1118 * 0 if @work was idle and we claimed PENDING
1119 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1120 * -ENOENT if someone else is canceling @work, this state may persist
1121 * for arbitrarily long
1123 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1124 * interrupted while holding PENDING and @work off queue, irq must be
1125 * disabled on entry. This, combined with delayed_work->timer being
1126 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1128 * On successful return, >= 0, irq is disabled and the caller is
1129 * responsible for releasing it using local_irq_restore(*@flags).
1131 * This function is safe to call from any context including IRQ handler.
1133 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1134 unsigned long *flags)
1136 struct worker_pool *pool;
1137 struct pool_workqueue *pwq;
1139 local_irq_save(*flags);
1141 /* try to steal the timer if it exists */
1143 struct delayed_work *dwork = to_delayed_work(work);
1146 * dwork->timer is irqsafe. If del_timer() fails, it's
1147 * guaranteed that the timer is not queued anywhere and not
1148 * running on the local CPU.
1150 if (likely(del_timer(&dwork->timer)))
1154 /* try to claim PENDING the normal way */
1155 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1159 * The queueing is in progress, or it is already queued. Try to
1160 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1162 pool = get_work_pool(work);
1166 spin_lock(&pool->lock);
1168 * work->data is guaranteed to point to pwq only while the work
1169 * item is queued on pwq->wq, and both updating work->data to point
1170 * to pwq on queueing and to pool on dequeueing are done under
1171 * pwq->pool->lock. This in turn guarantees that, if work->data
1172 * points to pwq which is associated with a locked pool, the work
1173 * item is currently queued on that pool.
1175 pwq = get_work_pwq(work);
1176 if (pwq && pwq->pool == pool) {
1177 debug_work_deactivate(work);
1180 * A delayed work item cannot be grabbed directly because
1181 * it might have linked NO_COLOR work items which, if left
1182 * on the delayed_list, will confuse pwq->nr_active
1183 * management later on and cause stall. Make sure the work
1184 * item is activated before grabbing.
1186 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1187 pwq_activate_delayed_work(work);
1189 list_del_init(&work->entry);
1190 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1192 /* work->data points to pwq iff queued, point to pool */
1193 set_work_pool_and_keep_pending(work, pool->id);
1195 spin_unlock(&pool->lock);
1198 spin_unlock(&pool->lock);
1200 local_irq_restore(*flags);
1201 if (work_is_canceling(work))
1208 * insert_work - insert a work into a pool
1209 * @pwq: pwq @work belongs to
1210 * @work: work to insert
1211 * @head: insertion point
1212 * @extra_flags: extra WORK_STRUCT_* flags to set
1214 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1215 * work_struct flags.
1218 * spin_lock_irq(pool->lock).
1220 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1221 struct list_head *head, unsigned int extra_flags)
1223 struct worker_pool *pool = pwq->pool;
1225 /* we own @work, set data and link */
1226 set_work_pwq(work, pwq, extra_flags);
1227 list_add_tail(&work->entry, head);
1231 * Ensure either wq_worker_sleeping() sees the above
1232 * list_add_tail() or we see zero nr_running to avoid workers lying
1233 * around lazily while there are works to be processed.
1237 if (__need_more_worker(pool))
1238 wake_up_worker(pool);
1242 * Test whether @work is being queued from another work executing on the
1245 static bool is_chained_work(struct workqueue_struct *wq)
1247 struct worker *worker;
1249 worker = current_wq_worker();
1251 * Return %true iff I'm a worker execuing a work item on @wq. If
1252 * I'm @worker, it's safe to dereference it without locking.
1254 return worker && worker->current_pwq->wq == wq;
1257 static void __queue_work(int cpu, struct workqueue_struct *wq,
1258 struct work_struct *work)
1260 struct pool_workqueue *pwq;
1261 struct worker_pool *last_pool;
1262 struct list_head *worklist;
1263 unsigned int work_flags;
1264 unsigned int req_cpu = cpu;
1267 * While a work item is PENDING && off queue, a task trying to
1268 * steal the PENDING will busy-loop waiting for it to either get
1269 * queued or lose PENDING. Grabbing PENDING and queueing should
1270 * happen with IRQ disabled.
1272 WARN_ON_ONCE(!irqs_disabled());
1274 debug_work_activate(work);
1276 /* if dying, only works from the same workqueue are allowed */
1277 if (unlikely(wq->flags & __WQ_DRAINING) &&
1278 WARN_ON_ONCE(!is_chained_work(wq)))
1281 /* pwq which will be used unless @work is executing elsewhere */
1282 if (!(wq->flags & WQ_UNBOUND)) {
1283 if (cpu == WORK_CPU_UNBOUND)
1284 cpu = raw_smp_processor_id();
1285 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1287 pwq = first_pwq(wq);
1291 * If @work was previously on a different pool, it might still be
1292 * running there, in which case the work needs to be queued on that
1293 * pool to guarantee non-reentrancy.
1295 last_pool = get_work_pool(work);
1296 if (last_pool && last_pool != pwq->pool) {
1297 struct worker *worker;
1299 spin_lock(&last_pool->lock);
1301 worker = find_worker_executing_work(last_pool, work);
1303 if (worker && worker->current_pwq->wq == wq) {
1304 pwq = worker->current_pwq;
1306 /* meh... not running there, queue here */
1307 spin_unlock(&last_pool->lock);
1308 spin_lock(&pwq->pool->lock);
1311 spin_lock(&pwq->pool->lock);
1315 * pwq is determined and locked. For unbound pools, we could have
1316 * raced with pwq release and it could already be dead. If its
1317 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1318 * without another pwq replacing it as the first pwq or while a
1319 * work item is executing on it, so the retying is guaranteed to
1320 * make forward-progress.
1322 if (unlikely(!pwq->refcnt)) {
1323 if (wq->flags & WQ_UNBOUND) {
1324 spin_unlock(&pwq->pool->lock);
1329 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1333 /* pwq determined, queue */
1334 trace_workqueue_queue_work(req_cpu, pwq, work);
1336 if (WARN_ON(!list_empty(&work->entry))) {
1337 spin_unlock(&pwq->pool->lock);
1341 pwq->nr_in_flight[pwq->work_color]++;
1342 work_flags = work_color_to_flags(pwq->work_color);
1344 if (likely(pwq->nr_active < pwq->max_active)) {
1345 trace_workqueue_activate_work(work);
1347 worklist = &pwq->pool->worklist;
1349 work_flags |= WORK_STRUCT_DELAYED;
1350 worklist = &pwq->delayed_works;
1353 insert_work(pwq, work, worklist, work_flags);
1355 spin_unlock(&pwq->pool->lock);
1359 * queue_work_on - queue work on specific cpu
1360 * @cpu: CPU number to execute work on
1361 * @wq: workqueue to use
1362 * @work: work to queue
1364 * Returns %false if @work was already on a queue, %true otherwise.
1366 * We queue the work to a specific CPU, the caller must ensure it
1369 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1370 struct work_struct *work)
1373 unsigned long flags;
1375 local_irq_save(flags);
1377 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1378 __queue_work(cpu, wq, work);
1382 local_irq_restore(flags);
1385 EXPORT_SYMBOL_GPL(queue_work_on);
1387 void delayed_work_timer_fn(unsigned long __data)
1389 struct delayed_work *dwork = (struct delayed_work *)__data;
1391 /* should have been called from irqsafe timer with irq already off */
1392 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1394 EXPORT_SYMBOL(delayed_work_timer_fn);
1396 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1397 struct delayed_work *dwork, unsigned long delay)
1399 struct timer_list *timer = &dwork->timer;
1400 struct work_struct *work = &dwork->work;
1402 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1403 timer->data != (unsigned long)dwork);
1404 WARN_ON_ONCE(timer_pending(timer));
1405 WARN_ON_ONCE(!list_empty(&work->entry));
1408 * If @delay is 0, queue @dwork->work immediately. This is for
1409 * both optimization and correctness. The earliest @timer can
1410 * expire is on the closest next tick and delayed_work users depend
1411 * on that there's no such delay when @delay is 0.
1414 __queue_work(cpu, wq, &dwork->work);
1418 timer_stats_timer_set_start_info(&dwork->timer);
1422 timer->expires = jiffies + delay;
1424 if (unlikely(cpu != WORK_CPU_UNBOUND))
1425 add_timer_on(timer, cpu);
1431 * queue_delayed_work_on - queue work on specific CPU after delay
1432 * @cpu: CPU number to execute work on
1433 * @wq: workqueue to use
1434 * @dwork: work to queue
1435 * @delay: number of jiffies to wait before queueing
1437 * Returns %false if @work was already on a queue, %true otherwise. If
1438 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1441 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1442 struct delayed_work *dwork, unsigned long delay)
1444 struct work_struct *work = &dwork->work;
1446 unsigned long flags;
1448 /* read the comment in __queue_work() */
1449 local_irq_save(flags);
1451 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1452 __queue_delayed_work(cpu, wq, dwork, delay);
1456 local_irq_restore(flags);
1459 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1462 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1463 * @cpu: CPU number to execute work on
1464 * @wq: workqueue to use
1465 * @dwork: work to queue
1466 * @delay: number of jiffies to wait before queueing
1468 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1469 * modify @dwork's timer so that it expires after @delay. If @delay is
1470 * zero, @work is guaranteed to be scheduled immediately regardless of its
1473 * Returns %false if @dwork was idle and queued, %true if @dwork was
1474 * pending and its timer was modified.
1476 * This function is safe to call from any context including IRQ handler.
1477 * See try_to_grab_pending() for details.
1479 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1480 struct delayed_work *dwork, unsigned long delay)
1482 unsigned long flags;
1486 ret = try_to_grab_pending(&dwork->work, true, &flags);
1487 } while (unlikely(ret == -EAGAIN));
1489 if (likely(ret >= 0)) {
1490 __queue_delayed_work(cpu, wq, dwork, delay);
1491 local_irq_restore(flags);
1494 /* -ENOENT from try_to_grab_pending() becomes %true */
1497 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1500 * worker_enter_idle - enter idle state
1501 * @worker: worker which is entering idle state
1503 * @worker is entering idle state. Update stats and idle timer if
1507 * spin_lock_irq(pool->lock).
1509 static void worker_enter_idle(struct worker *worker)
1511 struct worker_pool *pool = worker->pool;
1513 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1514 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1515 (worker->hentry.next || worker->hentry.pprev)))
1518 /* can't use worker_set_flags(), also called from start_worker() */
1519 worker->flags |= WORKER_IDLE;
1521 worker->last_active = jiffies;
1523 /* idle_list is LIFO */
1524 list_add(&worker->entry, &pool->idle_list);
1526 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1527 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1530 * Sanity check nr_running. Because wq_unbind_fn() releases
1531 * pool->lock between setting %WORKER_UNBOUND and zapping
1532 * nr_running, the warning may trigger spuriously. Check iff
1533 * unbind is not in progress.
1535 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1536 pool->nr_workers == pool->nr_idle &&
1537 atomic_read(&pool->nr_running));
1541 * worker_leave_idle - leave idle state
1542 * @worker: worker which is leaving idle state
1544 * @worker is leaving idle state. Update stats.
1547 * spin_lock_irq(pool->lock).
1549 static void worker_leave_idle(struct worker *worker)
1551 struct worker_pool *pool = worker->pool;
1553 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1555 worker_clr_flags(worker, WORKER_IDLE);
1557 list_del_init(&worker->entry);
1561 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1562 * @pool: target worker_pool
1564 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1566 * Works which are scheduled while the cpu is online must at least be
1567 * scheduled to a worker which is bound to the cpu so that if they are
1568 * flushed from cpu callbacks while cpu is going down, they are
1569 * guaranteed to execute on the cpu.
1571 * This function is to be used by unbound workers and rescuers to bind
1572 * themselves to the target cpu and may race with cpu going down or
1573 * coming online. kthread_bind() can't be used because it may put the
1574 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1575 * verbatim as it's best effort and blocking and pool may be
1576 * [dis]associated in the meantime.
1578 * This function tries set_cpus_allowed() and locks pool and verifies the
1579 * binding against %POOL_DISASSOCIATED which is set during
1580 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1581 * enters idle state or fetches works without dropping lock, it can
1582 * guarantee the scheduling requirement described in the first paragraph.
1585 * Might sleep. Called without any lock but returns with pool->lock
1589 * %true if the associated pool is online (@worker is successfully
1590 * bound), %false if offline.
1592 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1593 __acquires(&pool->lock)
1597 * The following call may fail, succeed or succeed
1598 * without actually migrating the task to the cpu if
1599 * it races with cpu hotunplug operation. Verify
1600 * against POOL_DISASSOCIATED.
1602 if (!(pool->flags & POOL_DISASSOCIATED))
1603 set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1605 spin_lock_irq(&pool->lock);
1606 if (pool->flags & POOL_DISASSOCIATED)
1608 if (task_cpu(current) == pool->cpu &&
1609 cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask))
1611 spin_unlock_irq(&pool->lock);
1614 * We've raced with CPU hot[un]plug. Give it a breather
1615 * and retry migration. cond_resched() is required here;
1616 * otherwise, we might deadlock against cpu_stop trying to
1617 * bring down the CPU on non-preemptive kernel.
1624 static struct worker *alloc_worker(void)
1626 struct worker *worker;
1628 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1630 INIT_LIST_HEAD(&worker->entry);
1631 INIT_LIST_HEAD(&worker->scheduled);
1632 /* on creation a worker is in !idle && prep state */
1633 worker->flags = WORKER_PREP;
1639 * create_worker - create a new workqueue worker
1640 * @pool: pool the new worker will belong to
1642 * Create a new worker which is bound to @pool. The returned worker
1643 * can be started by calling start_worker() or destroyed using
1647 * Might sleep. Does GFP_KERNEL allocations.
1650 * Pointer to the newly created worker.
1652 static struct worker *create_worker(struct worker_pool *pool)
1654 struct worker *worker = NULL;
1658 lockdep_assert_held(&pool->manager_mutex);
1661 * ID is needed to determine kthread name. Allocate ID first
1662 * without installing the pointer.
1664 idr_preload(GFP_KERNEL);
1665 spin_lock_irq(&pool->lock);
1667 id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1669 spin_unlock_irq(&pool->lock);
1674 worker = alloc_worker();
1678 worker->pool = pool;
1682 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1683 pool->attrs->nice < 0 ? "H" : "");
1685 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1687 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1688 "kworker/%s", id_buf);
1689 if (IS_ERR(worker->task))
1693 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1694 * online CPUs. It'll be re-applied when any of the CPUs come up.
1696 set_user_nice(worker->task, pool->attrs->nice);
1697 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1699 /* prevent userland from meddling with cpumask of workqueue workers */
1700 worker->task->flags |= PF_NO_SETAFFINITY;
1703 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1704 * remains stable across this function. See the comments above the
1705 * flag definition for details.
1707 if (pool->flags & POOL_DISASSOCIATED)
1708 worker->flags |= WORKER_UNBOUND;
1710 /* successful, commit the pointer to idr */
1711 spin_lock_irq(&pool->lock);
1712 idr_replace(&pool->worker_idr, worker, worker->id);
1713 spin_unlock_irq(&pool->lock);
1719 spin_lock_irq(&pool->lock);
1720 idr_remove(&pool->worker_idr, id);
1721 spin_unlock_irq(&pool->lock);
1728 * start_worker - start a newly created worker
1729 * @worker: worker to start
1731 * Make the pool aware of @worker and start it.
1734 * spin_lock_irq(pool->lock).
1736 static void start_worker(struct worker *worker)
1738 worker->flags |= WORKER_STARTED;
1739 worker->pool->nr_workers++;
1740 worker_enter_idle(worker);
1741 wake_up_process(worker->task);
1745 * create_and_start_worker - create and start a worker for a pool
1746 * @pool: the target pool
1748 * Grab the managership of @pool and create and start a new worker for it.
1750 static int create_and_start_worker(struct worker_pool *pool)
1752 struct worker *worker;
1754 mutex_lock(&pool->manager_mutex);
1756 worker = create_worker(pool);
1758 spin_lock_irq(&pool->lock);
1759 start_worker(worker);
1760 spin_unlock_irq(&pool->lock);
1763 mutex_unlock(&pool->manager_mutex);
1765 return worker ? 0 : -ENOMEM;
1769 * destroy_worker - destroy a workqueue worker
1770 * @worker: worker to be destroyed
1772 * Destroy @worker and adjust @pool stats accordingly.
1775 * spin_lock_irq(pool->lock) which is released and regrabbed.
1777 static void destroy_worker(struct worker *worker)
1779 struct worker_pool *pool = worker->pool;
1781 lockdep_assert_held(&pool->manager_mutex);
1782 lockdep_assert_held(&pool->lock);
1784 /* sanity check frenzy */
1785 if (WARN_ON(worker->current_work) ||
1786 WARN_ON(!list_empty(&worker->scheduled)))
1789 if (worker->flags & WORKER_STARTED)
1791 if (worker->flags & WORKER_IDLE)
1794 list_del_init(&worker->entry);
1795 worker->flags |= WORKER_DIE;
1797 idr_remove(&pool->worker_idr, worker->id);
1799 spin_unlock_irq(&pool->lock);
1801 kthread_stop(worker->task);
1804 spin_lock_irq(&pool->lock);
1807 static void idle_worker_timeout(unsigned long __pool)
1809 struct worker_pool *pool = (void *)__pool;
1811 spin_lock_irq(&pool->lock);
1813 if (too_many_workers(pool)) {
1814 struct worker *worker;
1815 unsigned long expires;
1817 /* idle_list is kept in LIFO order, check the last one */
1818 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1819 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1821 if (time_before(jiffies, expires))
1822 mod_timer(&pool->idle_timer, expires);
1824 /* it's been idle for too long, wake up manager */
1825 pool->flags |= POOL_MANAGE_WORKERS;
1826 wake_up_worker(pool);
1830 spin_unlock_irq(&pool->lock);
1833 static void send_mayday(struct work_struct *work)
1835 struct pool_workqueue *pwq = get_work_pwq(work);
1836 struct workqueue_struct *wq = pwq->wq;
1838 lockdep_assert_held(&wq_mayday_lock);
1843 /* mayday mayday mayday */
1844 if (list_empty(&pwq->mayday_node)) {
1845 list_add_tail(&pwq->mayday_node, &wq->maydays);
1846 wake_up_process(wq->rescuer->task);
1850 static void pool_mayday_timeout(unsigned long __pool)
1852 struct worker_pool *pool = (void *)__pool;
1853 struct work_struct *work;
1855 spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1856 spin_lock(&pool->lock);
1858 if (need_to_create_worker(pool)) {
1860 * We've been trying to create a new worker but
1861 * haven't been successful. We might be hitting an
1862 * allocation deadlock. Send distress signals to
1865 list_for_each_entry(work, &pool->worklist, entry)
1869 spin_unlock(&pool->lock);
1870 spin_unlock_irq(&wq_mayday_lock);
1872 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1876 * maybe_create_worker - create a new worker if necessary
1877 * @pool: pool to create a new worker for
1879 * Create a new worker for @pool if necessary. @pool is guaranteed to
1880 * have at least one idle worker on return from this function. If
1881 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1882 * sent to all rescuers with works scheduled on @pool to resolve
1883 * possible allocation deadlock.
1885 * On return, need_to_create_worker() is guaranteed to be %false and
1886 * may_start_working() %true.
1889 * spin_lock_irq(pool->lock) which may be released and regrabbed
1890 * multiple times. Does GFP_KERNEL allocations. Called only from
1894 * %false if no action was taken and pool->lock stayed locked, %true
1897 static bool maybe_create_worker(struct worker_pool *pool)
1898 __releases(&pool->lock)
1899 __acquires(&pool->lock)
1901 if (!need_to_create_worker(pool))
1904 spin_unlock_irq(&pool->lock);
1906 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1907 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1910 struct worker *worker;
1912 worker = create_worker(pool);
1914 del_timer_sync(&pool->mayday_timer);
1915 spin_lock_irq(&pool->lock);
1916 start_worker(worker);
1917 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1922 if (!need_to_create_worker(pool))
1925 __set_current_state(TASK_INTERRUPTIBLE);
1926 schedule_timeout(CREATE_COOLDOWN);
1928 if (!need_to_create_worker(pool))
1932 del_timer_sync(&pool->mayday_timer);
1933 spin_lock_irq(&pool->lock);
1934 if (need_to_create_worker(pool))
1940 * maybe_destroy_worker - destroy workers which have been idle for a while
1941 * @pool: pool to destroy workers for
1943 * Destroy @pool workers which have been idle for longer than
1944 * IDLE_WORKER_TIMEOUT.
1947 * spin_lock_irq(pool->lock) which may be released and regrabbed
1948 * multiple times. Called only from manager.
1951 * %false if no action was taken and pool->lock stayed locked, %true
1954 static bool maybe_destroy_workers(struct worker_pool *pool)
1958 while (too_many_workers(pool)) {
1959 struct worker *worker;
1960 unsigned long expires;
1962 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1963 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1965 if (time_before(jiffies, expires)) {
1966 mod_timer(&pool->idle_timer, expires);
1970 destroy_worker(worker);
1978 * manage_workers - manage worker pool
1981 * Assume the manager role and manage the worker pool @worker belongs
1982 * to. At any given time, there can be only zero or one manager per
1983 * pool. The exclusion is handled automatically by this function.
1985 * The caller can safely start processing works on false return. On
1986 * true return, it's guaranteed that need_to_create_worker() is false
1987 * and may_start_working() is true.
1990 * spin_lock_irq(pool->lock) which may be released and regrabbed
1991 * multiple times. Does GFP_KERNEL allocations.
1994 * spin_lock_irq(pool->lock) which may be released and regrabbed
1995 * multiple times. Does GFP_KERNEL allocations.
1997 static bool manage_workers(struct worker *worker)
1999 struct worker_pool *pool = worker->pool;
2003 * Managership is governed by two mutexes - manager_arb and
2004 * manager_mutex. manager_arb handles arbitration of manager role.
2005 * Anyone who successfully grabs manager_arb wins the arbitration
2006 * and becomes the manager. mutex_trylock() on pool->manager_arb
2007 * failure while holding pool->lock reliably indicates that someone
2008 * else is managing the pool and the worker which failed trylock
2009 * can proceed to executing work items. This means that anyone
2010 * grabbing manager_arb is responsible for actually performing
2011 * manager duties. If manager_arb is grabbed and released without
2012 * actual management, the pool may stall indefinitely.
2014 * manager_mutex is used for exclusion of actual management
2015 * operations. The holder of manager_mutex can be sure that none
2016 * of management operations, including creation and destruction of
2017 * workers, won't take place until the mutex is released. Because
2018 * manager_mutex doesn't interfere with manager role arbitration,
2019 * it is guaranteed that the pool's management, while may be
2020 * delayed, won't be disturbed by someone else grabbing
2023 if (!mutex_trylock(&pool->manager_arb))
2027 * With manager arbitration won, manager_mutex would be free in
2028 * most cases. trylock first without dropping @pool->lock.
2030 if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2031 spin_unlock_irq(&pool->lock);
2032 mutex_lock(&pool->manager_mutex);
2036 pool->flags &= ~POOL_MANAGE_WORKERS;
2039 * Destroy and then create so that may_start_working() is true
2042 ret |= maybe_destroy_workers(pool);
2043 ret |= maybe_create_worker(pool);
2045 mutex_unlock(&pool->manager_mutex);
2046 mutex_unlock(&pool->manager_arb);
2051 * process_one_work - process single work
2053 * @work: work to process
2055 * Process @work. This function contains all the logics necessary to
2056 * process a single work including synchronization against and
2057 * interaction with other workers on the same cpu, queueing and
2058 * flushing. As long as context requirement is met, any worker can
2059 * call this function to process a work.
2062 * spin_lock_irq(pool->lock) which is released and regrabbed.
2064 static void process_one_work(struct worker *worker, struct work_struct *work)
2065 __releases(&pool->lock)
2066 __acquires(&pool->lock)
2068 struct pool_workqueue *pwq = get_work_pwq(work);
2069 struct worker_pool *pool = worker->pool;
2070 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2072 struct worker *collision;
2073 #ifdef CONFIG_LOCKDEP
2075 * It is permissible to free the struct work_struct from
2076 * inside the function that is called from it, this we need to
2077 * take into account for lockdep too. To avoid bogus "held
2078 * lock freed" warnings as well as problems when looking into
2079 * work->lockdep_map, make a copy and use that here.
2081 struct lockdep_map lockdep_map;
2083 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2086 * Ensure we're on the correct CPU. DISASSOCIATED test is
2087 * necessary to avoid spurious warnings from rescuers servicing the
2088 * unbound or a disassociated pool.
2090 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2091 !(pool->flags & POOL_DISASSOCIATED) &&
2092 raw_smp_processor_id() != pool->cpu);
2095 * A single work shouldn't be executed concurrently by
2096 * multiple workers on a single cpu. Check whether anyone is
2097 * already processing the work. If so, defer the work to the
2098 * currently executing one.
2100 collision = find_worker_executing_work(pool, work);
2101 if (unlikely(collision)) {
2102 move_linked_works(work, &collision->scheduled, NULL);
2106 /* claim and dequeue */
2107 debug_work_deactivate(work);
2108 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2109 worker->current_work = work;
2110 worker->current_func = work->func;
2111 worker->current_pwq = pwq;
2112 work_color = get_work_color(work);
2114 list_del_init(&work->entry);
2117 * CPU intensive works don't participate in concurrency
2118 * management. They're the scheduler's responsibility.
2120 if (unlikely(cpu_intensive))
2121 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2124 * Unbound pool isn't concurrency managed and work items should be
2125 * executed ASAP. Wake up another worker if necessary.
2127 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2128 wake_up_worker(pool);
2131 * Record the last pool and clear PENDING which should be the last
2132 * update to @work. Also, do this inside @pool->lock so that
2133 * PENDING and queued state changes happen together while IRQ is
2136 set_work_pool_and_clear_pending(work, pool->id);
2138 spin_unlock_irq(&pool->lock);
2140 lock_map_acquire_read(&pwq->wq->lockdep_map);
2141 lock_map_acquire(&lockdep_map);
2142 trace_workqueue_execute_start(work);
2143 worker->current_func(work);
2145 * While we must be careful to not use "work" after this, the trace
2146 * point will only record its address.
2148 trace_workqueue_execute_end(work);
2149 lock_map_release(&lockdep_map);
2150 lock_map_release(&pwq->wq->lockdep_map);
2152 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2153 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2154 " last function: %pf\n",
2155 current->comm, preempt_count(), task_pid_nr(current),
2156 worker->current_func);
2157 debug_show_held_locks(current);
2161 spin_lock_irq(&pool->lock);
2163 /* clear cpu intensive status */
2164 if (unlikely(cpu_intensive))
2165 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2167 /* we're done with it, release */
2168 hash_del(&worker->hentry);
2169 worker->current_work = NULL;
2170 worker->current_func = NULL;
2171 worker->current_pwq = NULL;
2172 pwq_dec_nr_in_flight(pwq, work_color);
2176 * process_scheduled_works - process scheduled works
2179 * Process all scheduled works. Please note that the scheduled list
2180 * may change while processing a work, so this function repeatedly
2181 * fetches a work from the top and executes it.
2184 * spin_lock_irq(pool->lock) which may be released and regrabbed
2187 static void process_scheduled_works(struct worker *worker)
2189 while (!list_empty(&worker->scheduled)) {
2190 struct work_struct *work = list_first_entry(&worker->scheduled,
2191 struct work_struct, entry);
2192 process_one_work(worker, work);
2197 * worker_thread - the worker thread function
2200 * The worker thread function. All workers belong to a worker_pool -
2201 * either a per-cpu one or dynamic unbound one. These workers process all
2202 * work items regardless of their specific target workqueue. The only
2203 * exception is work items which belong to workqueues with a rescuer which
2204 * will be explained in rescuer_thread().
2206 static int worker_thread(void *__worker)
2208 struct worker *worker = __worker;
2209 struct worker_pool *pool = worker->pool;
2211 /* tell the scheduler that this is a workqueue worker */
2212 worker->task->flags |= PF_WQ_WORKER;
2214 spin_lock_irq(&pool->lock);
2216 /* am I supposed to die? */
2217 if (unlikely(worker->flags & WORKER_DIE)) {
2218 spin_unlock_irq(&pool->lock);
2219 WARN_ON_ONCE(!list_empty(&worker->entry));
2220 worker->task->flags &= ~PF_WQ_WORKER;
2224 worker_leave_idle(worker);
2226 /* no more worker necessary? */
2227 if (!need_more_worker(pool))
2230 /* do we need to manage? */
2231 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2235 * ->scheduled list can only be filled while a worker is
2236 * preparing to process a work or actually processing it.
2237 * Make sure nobody diddled with it while I was sleeping.
2239 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2242 * Finish PREP stage. We're guaranteed to have at least one idle
2243 * worker or that someone else has already assumed the manager
2244 * role. This is where @worker starts participating in concurrency
2245 * management if applicable and concurrency management is restored
2246 * after being rebound. See rebind_workers() for details.
2248 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2251 struct work_struct *work =
2252 list_first_entry(&pool->worklist,
2253 struct work_struct, entry);
2255 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2256 /* optimization path, not strictly necessary */
2257 process_one_work(worker, work);
2258 if (unlikely(!list_empty(&worker->scheduled)))
2259 process_scheduled_works(worker);
2261 move_linked_works(work, &worker->scheduled, NULL);
2262 process_scheduled_works(worker);
2264 } while (keep_working(pool));
2266 worker_set_flags(worker, WORKER_PREP, false);
2268 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2272 * pool->lock is held and there's no work to process and no need to
2273 * manage, sleep. Workers are woken up only while holding
2274 * pool->lock or from local cpu, so setting the current state
2275 * before releasing pool->lock is enough to prevent losing any
2278 worker_enter_idle(worker);
2279 __set_current_state(TASK_INTERRUPTIBLE);
2280 spin_unlock_irq(&pool->lock);
2286 * rescuer_thread - the rescuer thread function
2289 * Workqueue rescuer thread function. There's one rescuer for each
2290 * workqueue which has WQ_MEM_RECLAIM set.
2292 * Regular work processing on a pool may block trying to create a new
2293 * worker which uses GFP_KERNEL allocation which has slight chance of
2294 * developing into deadlock if some works currently on the same queue
2295 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2296 * the problem rescuer solves.
2298 * When such condition is possible, the pool summons rescuers of all
2299 * workqueues which have works queued on the pool and let them process
2300 * those works so that forward progress can be guaranteed.
2302 * This should happen rarely.
2304 static int rescuer_thread(void *__rescuer)
2306 struct worker *rescuer = __rescuer;
2307 struct workqueue_struct *wq = rescuer->rescue_wq;
2308 struct list_head *scheduled = &rescuer->scheduled;
2310 set_user_nice(current, RESCUER_NICE_LEVEL);
2313 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2314 * doesn't participate in concurrency management.
2316 rescuer->task->flags |= PF_WQ_WORKER;
2318 set_current_state(TASK_INTERRUPTIBLE);
2320 if (kthread_should_stop()) {
2321 __set_current_state(TASK_RUNNING);
2322 rescuer->task->flags &= ~PF_WQ_WORKER;
2326 /* see whether any pwq is asking for help */
2327 spin_lock_irq(&wq_mayday_lock);
2329 while (!list_empty(&wq->maydays)) {
2330 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2331 struct pool_workqueue, mayday_node);
2332 struct worker_pool *pool = pwq->pool;
2333 struct work_struct *work, *n;
2335 __set_current_state(TASK_RUNNING);
2336 list_del_init(&pwq->mayday_node);
2338 spin_unlock_irq(&wq_mayday_lock);
2340 /* migrate to the target cpu if possible */
2341 worker_maybe_bind_and_lock(pool);
2342 rescuer->pool = pool;
2345 * Slurp in all works issued via this workqueue and
2348 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2349 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2350 if (get_work_pwq(work) == pwq)
2351 move_linked_works(work, scheduled, &n);
2353 process_scheduled_works(rescuer);
2356 * Leave this pool. If keep_working() is %true, notify a
2357 * regular worker; otherwise, we end up with 0 concurrency
2358 * and stalling the execution.
2360 if (keep_working(pool))
2361 wake_up_worker(pool);
2363 rescuer->pool = NULL;
2364 spin_unlock(&pool->lock);
2365 spin_lock(&wq_mayday_lock);
2368 spin_unlock_irq(&wq_mayday_lock);
2370 /* rescuers should never participate in concurrency management */
2371 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2377 struct work_struct work;
2378 struct completion done;
2381 static void wq_barrier_func(struct work_struct *work)
2383 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2384 complete(&barr->done);
2388 * insert_wq_barrier - insert a barrier work
2389 * @pwq: pwq to insert barrier into
2390 * @barr: wq_barrier to insert
2391 * @target: target work to attach @barr to
2392 * @worker: worker currently executing @target, NULL if @target is not executing
2394 * @barr is linked to @target such that @barr is completed only after
2395 * @target finishes execution. Please note that the ordering
2396 * guarantee is observed only with respect to @target and on the local
2399 * Currently, a queued barrier can't be canceled. This is because
2400 * try_to_grab_pending() can't determine whether the work to be
2401 * grabbed is at the head of the queue and thus can't clear LINKED
2402 * flag of the previous work while there must be a valid next work
2403 * after a work with LINKED flag set.
2405 * Note that when @worker is non-NULL, @target may be modified
2406 * underneath us, so we can't reliably determine pwq from @target.
2409 * spin_lock_irq(pool->lock).
2411 static void insert_wq_barrier(struct pool_workqueue *pwq,
2412 struct wq_barrier *barr,
2413 struct work_struct *target, struct worker *worker)
2415 struct list_head *head;
2416 unsigned int linked = 0;
2419 * debugobject calls are safe here even with pool->lock locked
2420 * as we know for sure that this will not trigger any of the
2421 * checks and call back into the fixup functions where we
2424 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2425 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2426 init_completion(&barr->done);
2429 * If @target is currently being executed, schedule the
2430 * barrier to the worker; otherwise, put it after @target.
2433 head = worker->scheduled.next;
2435 unsigned long *bits = work_data_bits(target);
2437 head = target->entry.next;
2438 /* there can already be other linked works, inherit and set */
2439 linked = *bits & WORK_STRUCT_LINKED;
2440 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2443 debug_work_activate(&barr->work);
2444 insert_work(pwq, &barr->work, head,
2445 work_color_to_flags(WORK_NO_COLOR) | linked);
2449 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2450 * @wq: workqueue being flushed
2451 * @flush_color: new flush color, < 0 for no-op
2452 * @work_color: new work color, < 0 for no-op
2454 * Prepare pwqs for workqueue flushing.
2456 * If @flush_color is non-negative, flush_color on all pwqs should be
2457 * -1. If no pwq has in-flight commands at the specified color, all
2458 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2459 * has in flight commands, its pwq->flush_color is set to
2460 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2461 * wakeup logic is armed and %true is returned.
2463 * The caller should have initialized @wq->first_flusher prior to
2464 * calling this function with non-negative @flush_color. If
2465 * @flush_color is negative, no flush color update is done and %false
2468 * If @work_color is non-negative, all pwqs should have the same
2469 * work_color which is previous to @work_color and all will be
2470 * advanced to @work_color.
2473 * mutex_lock(wq->mutex).
2476 * %true if @flush_color >= 0 and there's something to flush. %false
2479 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2480 int flush_color, int work_color)
2483 struct pool_workqueue *pwq;
2485 if (flush_color >= 0) {
2486 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2487 atomic_set(&wq->nr_pwqs_to_flush, 1);
2490 for_each_pwq(pwq, wq) {
2491 struct worker_pool *pool = pwq->pool;
2493 spin_lock_irq(&pool->lock);
2495 if (flush_color >= 0) {
2496 WARN_ON_ONCE(pwq->flush_color != -1);
2498 if (pwq->nr_in_flight[flush_color]) {
2499 pwq->flush_color = flush_color;
2500 atomic_inc(&wq->nr_pwqs_to_flush);
2505 if (work_color >= 0) {
2506 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2507 pwq->work_color = work_color;
2510 spin_unlock_irq(&pool->lock);
2513 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2514 complete(&wq->first_flusher->done);
2520 * flush_workqueue - ensure that any scheduled work has run to completion.
2521 * @wq: workqueue to flush
2523 * This function sleeps until all work items which were queued on entry
2524 * have finished execution, but it is not livelocked by new incoming ones.
2526 void flush_workqueue(struct workqueue_struct *wq)
2528 struct wq_flusher this_flusher = {
2529 .list = LIST_HEAD_INIT(this_flusher.list),
2531 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2535 lock_map_acquire(&wq->lockdep_map);
2536 lock_map_release(&wq->lockdep_map);
2538 mutex_lock(&wq->mutex);
2541 * Start-to-wait phase
2543 next_color = work_next_color(wq->work_color);
2545 if (next_color != wq->flush_color) {
2547 * Color space is not full. The current work_color
2548 * becomes our flush_color and work_color is advanced
2551 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2552 this_flusher.flush_color = wq->work_color;
2553 wq->work_color = next_color;
2555 if (!wq->first_flusher) {
2556 /* no flush in progress, become the first flusher */
2557 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2559 wq->first_flusher = &this_flusher;
2561 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2563 /* nothing to flush, done */
2564 wq->flush_color = next_color;
2565 wq->first_flusher = NULL;
2570 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2571 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2572 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2576 * Oops, color space is full, wait on overflow queue.
2577 * The next flush completion will assign us
2578 * flush_color and transfer to flusher_queue.
2580 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2583 mutex_unlock(&wq->mutex);
2585 wait_for_completion(&this_flusher.done);
2588 * Wake-up-and-cascade phase
2590 * First flushers are responsible for cascading flushes and
2591 * handling overflow. Non-first flushers can simply return.
2593 if (wq->first_flusher != &this_flusher)
2596 mutex_lock(&wq->mutex);
2598 /* we might have raced, check again with mutex held */
2599 if (wq->first_flusher != &this_flusher)
2602 wq->first_flusher = NULL;
2604 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2605 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2608 struct wq_flusher *next, *tmp;
2610 /* complete all the flushers sharing the current flush color */
2611 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2612 if (next->flush_color != wq->flush_color)
2614 list_del_init(&next->list);
2615 complete(&next->done);
2618 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2619 wq->flush_color != work_next_color(wq->work_color));
2621 /* this flush_color is finished, advance by one */
2622 wq->flush_color = work_next_color(wq->flush_color);
2624 /* one color has been freed, handle overflow queue */
2625 if (!list_empty(&wq->flusher_overflow)) {
2627 * Assign the same color to all overflowed
2628 * flushers, advance work_color and append to
2629 * flusher_queue. This is the start-to-wait
2630 * phase for these overflowed flushers.
2632 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2633 tmp->flush_color = wq->work_color;
2635 wq->work_color = work_next_color(wq->work_color);
2637 list_splice_tail_init(&wq->flusher_overflow,
2638 &wq->flusher_queue);
2639 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2642 if (list_empty(&wq->flusher_queue)) {
2643 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2648 * Need to flush more colors. Make the next flusher
2649 * the new first flusher and arm pwqs.
2651 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2652 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2654 list_del_init(&next->list);
2655 wq->first_flusher = next;
2657 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2661 * Meh... this color is already done, clear first
2662 * flusher and repeat cascading.
2664 wq->first_flusher = NULL;
2668 mutex_unlock(&wq->mutex);
2670 EXPORT_SYMBOL_GPL(flush_workqueue);
2673 * drain_workqueue - drain a workqueue
2674 * @wq: workqueue to drain
2676 * Wait until the workqueue becomes empty. While draining is in progress,
2677 * only chain queueing is allowed. IOW, only currently pending or running
2678 * work items on @wq can queue further work items on it. @wq is flushed
2679 * repeatedly until it becomes empty. The number of flushing is detemined
2680 * by the depth of chaining and should be relatively short. Whine if it
2683 void drain_workqueue(struct workqueue_struct *wq)
2685 unsigned int flush_cnt = 0;
2686 struct pool_workqueue *pwq;
2689 * __queue_work() needs to test whether there are drainers, is much
2690 * hotter than drain_workqueue() and already looks at @wq->flags.
2691 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2693 mutex_lock(&wq->mutex);
2694 if (!wq->nr_drainers++)
2695 wq->flags |= __WQ_DRAINING;
2696 mutex_unlock(&wq->mutex);
2698 flush_workqueue(wq);
2700 mutex_lock(&wq->mutex);
2702 for_each_pwq(pwq, wq) {
2705 spin_lock_irq(&pwq->pool->lock);
2706 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2707 spin_unlock_irq(&pwq->pool->lock);
2712 if (++flush_cnt == 10 ||
2713 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2714 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2715 wq->name, flush_cnt);
2717 mutex_unlock(&wq->mutex);
2721 if (!--wq->nr_drainers)
2722 wq->flags &= ~__WQ_DRAINING;
2723 mutex_unlock(&wq->mutex);
2725 EXPORT_SYMBOL_GPL(drain_workqueue);
2727 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2729 struct worker *worker = NULL;
2730 struct worker_pool *pool;
2731 struct pool_workqueue *pwq;
2735 local_irq_disable();
2736 pool = get_work_pool(work);
2742 spin_lock(&pool->lock);
2743 /* see the comment in try_to_grab_pending() with the same code */
2744 pwq = get_work_pwq(work);
2746 if (unlikely(pwq->pool != pool))
2749 worker = find_worker_executing_work(pool, work);
2752 pwq = worker->current_pwq;
2755 insert_wq_barrier(pwq, barr, work, worker);
2756 spin_unlock_irq(&pool->lock);
2759 * If @max_active is 1 or rescuer is in use, flushing another work
2760 * item on the same workqueue may lead to deadlock. Make sure the
2761 * flusher is not running on the same workqueue by verifying write
2764 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2765 lock_map_acquire(&pwq->wq->lockdep_map);
2767 lock_map_acquire_read(&pwq->wq->lockdep_map);
2768 lock_map_release(&pwq->wq->lockdep_map);
2772 spin_unlock_irq(&pool->lock);
2777 * flush_work - wait for a work to finish executing the last queueing instance
2778 * @work: the work to flush
2780 * Wait until @work has finished execution. @work is guaranteed to be idle
2781 * on return if it hasn't been requeued since flush started.
2784 * %true if flush_work() waited for the work to finish execution,
2785 * %false if it was already idle.
2787 bool flush_work(struct work_struct *work)
2789 struct wq_barrier barr;
2791 lock_map_acquire(&work->lockdep_map);
2792 lock_map_release(&work->lockdep_map);
2794 if (start_flush_work(work, &barr)) {
2795 wait_for_completion(&barr.done);
2796 destroy_work_on_stack(&barr.work);
2802 EXPORT_SYMBOL_GPL(flush_work);
2804 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2806 unsigned long flags;
2810 ret = try_to_grab_pending(work, is_dwork, &flags);
2812 * If someone else is canceling, wait for the same event it
2813 * would be waiting for before retrying.
2815 if (unlikely(ret == -ENOENT))
2817 } while (unlikely(ret < 0));
2819 /* tell other tasks trying to grab @work to back off */
2820 mark_work_canceling(work);
2821 local_irq_restore(flags);
2824 clear_work_data(work);
2829 * cancel_work_sync - cancel a work and wait for it to finish
2830 * @work: the work to cancel
2832 * Cancel @work and wait for its execution to finish. This function
2833 * can be used even if the work re-queues itself or migrates to
2834 * another workqueue. On return from this function, @work is
2835 * guaranteed to be not pending or executing on any CPU.
2837 * cancel_work_sync(&delayed_work->work) must not be used for
2838 * delayed_work's. Use cancel_delayed_work_sync() instead.
2840 * The caller must ensure that the workqueue on which @work was last
2841 * queued can't be destroyed before this function returns.
2844 * %true if @work was pending, %false otherwise.
2846 bool cancel_work_sync(struct work_struct *work)
2848 return __cancel_work_timer(work, false);
2850 EXPORT_SYMBOL_GPL(cancel_work_sync);
2853 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2854 * @dwork: the delayed work to flush
2856 * Delayed timer is cancelled and the pending work is queued for
2857 * immediate execution. Like flush_work(), this function only
2858 * considers the last queueing instance of @dwork.
2861 * %true if flush_work() waited for the work to finish execution,
2862 * %false if it was already idle.
2864 bool flush_delayed_work(struct delayed_work *dwork)
2866 local_irq_disable();
2867 if (del_timer_sync(&dwork->timer))
2868 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2870 return flush_work(&dwork->work);
2872 EXPORT_SYMBOL(flush_delayed_work);
2875 * cancel_delayed_work - cancel a delayed work
2876 * @dwork: delayed_work to cancel
2878 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2879 * and canceled; %false if wasn't pending. Note that the work callback
2880 * function may still be running on return, unless it returns %true and the
2881 * work doesn't re-arm itself. Explicitly flush or use
2882 * cancel_delayed_work_sync() to wait on it.
2884 * This function is safe to call from any context including IRQ handler.
2886 bool cancel_delayed_work(struct delayed_work *dwork)
2888 unsigned long flags;
2892 ret = try_to_grab_pending(&dwork->work, true, &flags);
2893 } while (unlikely(ret == -EAGAIN));
2895 if (unlikely(ret < 0))
2898 set_work_pool_and_clear_pending(&dwork->work,
2899 get_work_pool_id(&dwork->work));
2900 local_irq_restore(flags);
2903 EXPORT_SYMBOL(cancel_delayed_work);
2906 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2907 * @dwork: the delayed work cancel
2909 * This is cancel_work_sync() for delayed works.
2912 * %true if @dwork was pending, %false otherwise.
2914 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2916 return __cancel_work_timer(&dwork->work, true);
2918 EXPORT_SYMBOL(cancel_delayed_work_sync);
2921 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2922 * @func: the function to call
2924 * schedule_on_each_cpu() executes @func on each online CPU using the
2925 * system workqueue and blocks until all CPUs have completed.
2926 * schedule_on_each_cpu() is very slow.
2929 * 0 on success, -errno on failure.
2931 int schedule_on_each_cpu(work_func_t func)
2934 struct work_struct __percpu *works;
2936 works = alloc_percpu(struct work_struct);
2942 for_each_online_cpu(cpu) {
2943 struct work_struct *work = per_cpu_ptr(works, cpu);
2945 INIT_WORK(work, func);
2946 schedule_work_on(cpu, work);
2949 for_each_online_cpu(cpu)
2950 flush_work(per_cpu_ptr(works, cpu));
2958 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2960 * Forces execution of the kernel-global workqueue and blocks until its
2963 * Think twice before calling this function! It's very easy to get into
2964 * trouble if you don't take great care. Either of the following situations
2965 * will lead to deadlock:
2967 * One of the work items currently on the workqueue needs to acquire
2968 * a lock held by your code or its caller.
2970 * Your code is running in the context of a work routine.
2972 * They will be detected by lockdep when they occur, but the first might not
2973 * occur very often. It depends on what work items are on the workqueue and
2974 * what locks they need, which you have no control over.
2976 * In most situations flushing the entire workqueue is overkill; you merely
2977 * need to know that a particular work item isn't queued and isn't running.
2978 * In such cases you should use cancel_delayed_work_sync() or
2979 * cancel_work_sync() instead.
2981 void flush_scheduled_work(void)
2983 flush_workqueue(system_wq);
2985 EXPORT_SYMBOL(flush_scheduled_work);
2988 * execute_in_process_context - reliably execute the routine with user context
2989 * @fn: the function to execute
2990 * @ew: guaranteed storage for the execute work structure (must
2991 * be available when the work executes)
2993 * Executes the function immediately if process context is available,
2994 * otherwise schedules the function for delayed execution.
2996 * Returns: 0 - function was executed
2997 * 1 - function was scheduled for execution
2999 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3001 if (!in_interrupt()) {
3006 INIT_WORK(&ew->work, fn);
3007 schedule_work(&ew->work);
3011 EXPORT_SYMBOL_GPL(execute_in_process_context);
3015 * Workqueues with WQ_SYSFS flag set is visible to userland via
3016 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3017 * following attributes.
3019 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3020 * max_active RW int : maximum number of in-flight work items
3022 * Unbound workqueues have the following extra attributes.
3024 * id RO int : the associated pool ID
3025 * nice RW int : nice value of the workers
3026 * cpumask RW mask : bitmask of allowed CPUs for the workers
3029 struct workqueue_struct *wq;
3033 static struct workqueue_struct *dev_to_wq(struct device *dev)
3035 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3040 static ssize_t wq_per_cpu_show(struct device *dev,
3041 struct device_attribute *attr, char *buf)
3043 struct workqueue_struct *wq = dev_to_wq(dev);
3045 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3048 static ssize_t wq_max_active_show(struct device *dev,
3049 struct device_attribute *attr, char *buf)
3051 struct workqueue_struct *wq = dev_to_wq(dev);
3053 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3056 static ssize_t wq_max_active_store(struct device *dev,
3057 struct device_attribute *attr,
3058 const char *buf, size_t count)
3060 struct workqueue_struct *wq = dev_to_wq(dev);
3063 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3066 workqueue_set_max_active(wq, val);
3070 static struct device_attribute wq_sysfs_attrs[] = {
3071 __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3072 __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3076 static ssize_t wq_pool_id_show(struct device *dev,
3077 struct device_attribute *attr, char *buf)
3079 struct workqueue_struct *wq = dev_to_wq(dev);
3080 struct worker_pool *pool;
3083 rcu_read_lock_sched();
3084 pool = first_pwq(wq)->pool;
3085 written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id);
3086 rcu_read_unlock_sched();
3091 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3094 struct workqueue_struct *wq = dev_to_wq(dev);
3097 mutex_lock(&wq->mutex);
3098 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3099 mutex_unlock(&wq->mutex);
3104 /* prepare workqueue_attrs for sysfs store operations */
3105 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3107 struct workqueue_attrs *attrs;
3109 attrs = alloc_workqueue_attrs(GFP_KERNEL);
3113 mutex_lock(&wq->mutex);
3114 copy_workqueue_attrs(attrs, wq->unbound_attrs);
3115 mutex_unlock(&wq->mutex);
3119 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3120 const char *buf, size_t count)
3122 struct workqueue_struct *wq = dev_to_wq(dev);
3123 struct workqueue_attrs *attrs;
3126 attrs = wq_sysfs_prep_attrs(wq);
3130 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3131 attrs->nice >= -20 && attrs->nice <= 19)
3132 ret = apply_workqueue_attrs(wq, attrs);
3136 free_workqueue_attrs(attrs);
3137 return ret ?: count;
3140 static ssize_t wq_cpumask_show(struct device *dev,
3141 struct device_attribute *attr, char *buf)
3143 struct workqueue_struct *wq = dev_to_wq(dev);
3146 mutex_lock(&wq->mutex);
3147 written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3148 mutex_unlock(&wq->mutex);
3150 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3154 static ssize_t wq_cpumask_store(struct device *dev,
3155 struct device_attribute *attr,
3156 const char *buf, size_t count)
3158 struct workqueue_struct *wq = dev_to_wq(dev);
3159 struct workqueue_attrs *attrs;
3162 attrs = wq_sysfs_prep_attrs(wq);
3166 ret = cpumask_parse(buf, attrs->cpumask);
3168 ret = apply_workqueue_attrs(wq, attrs);
3170 free_workqueue_attrs(attrs);
3171 return ret ?: count;
3174 static struct device_attribute wq_sysfs_unbound_attrs[] = {
3175 __ATTR(pool_id, 0444, wq_pool_id_show, NULL),
3176 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3177 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3181 static struct bus_type wq_subsys = {
3182 .name = "workqueue",
3183 .dev_attrs = wq_sysfs_attrs,
3186 static int __init wq_sysfs_init(void)
3188 return subsys_virtual_register(&wq_subsys, NULL);
3190 core_initcall(wq_sysfs_init);
3192 static void wq_device_release(struct device *dev)
3194 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3200 * workqueue_sysfs_register - make a workqueue visible in sysfs
3201 * @wq: the workqueue to register
3203 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3204 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3205 * which is the preferred method.
3207 * Workqueue user should use this function directly iff it wants to apply
3208 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3209 * apply_workqueue_attrs() may race against userland updating the
3212 * Returns 0 on success, -errno on failure.
3214 int workqueue_sysfs_register(struct workqueue_struct *wq)
3216 struct wq_device *wq_dev;
3220 * Adjusting max_active or creating new pwqs by applyting
3221 * attributes breaks ordering guarantee. Disallow exposing ordered
3224 if (WARN_ON(wq->flags & __WQ_ORDERED))
3227 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3232 wq_dev->dev.bus = &wq_subsys;
3233 wq_dev->dev.init_name = wq->name;
3234 wq_dev->dev.release = wq_device_release;
3237 * unbound_attrs are created separately. Suppress uevent until
3238 * everything is ready.
3240 dev_set_uevent_suppress(&wq_dev->dev, true);
3242 ret = device_register(&wq_dev->dev);
3249 if (wq->flags & WQ_UNBOUND) {
3250 struct device_attribute *attr;
3252 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3253 ret = device_create_file(&wq_dev->dev, attr);
3255 device_unregister(&wq_dev->dev);
3262 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3267 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3268 * @wq: the workqueue to unregister
3270 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3272 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3274 struct wq_device *wq_dev = wq->wq_dev;
3280 device_unregister(&wq_dev->dev);
3282 #else /* CONFIG_SYSFS */
3283 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3284 #endif /* CONFIG_SYSFS */
3287 * free_workqueue_attrs - free a workqueue_attrs
3288 * @attrs: workqueue_attrs to free
3290 * Undo alloc_workqueue_attrs().
3292 void free_workqueue_attrs(struct workqueue_attrs *attrs)
3295 free_cpumask_var(attrs->cpumask);
3301 * alloc_workqueue_attrs - allocate a workqueue_attrs
3302 * @gfp_mask: allocation mask to use
3304 * Allocate a new workqueue_attrs, initialize with default settings and
3305 * return it. Returns NULL on failure.
3307 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3309 struct workqueue_attrs *attrs;
3311 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3314 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3317 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3320 free_workqueue_attrs(attrs);
3324 static void copy_workqueue_attrs(struct workqueue_attrs *to,
3325 const struct workqueue_attrs *from)
3327 to->nice = from->nice;
3328 cpumask_copy(to->cpumask, from->cpumask);
3331 /* hash value of the content of @attr */
3332 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3336 hash = jhash_1word(attrs->nice, hash);
3337 hash = jhash(cpumask_bits(attrs->cpumask),
3338 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3342 /* content equality test */
3343 static bool wqattrs_equal(const struct workqueue_attrs *a,
3344 const struct workqueue_attrs *b)
3346 if (a->nice != b->nice)
3348 if (!cpumask_equal(a->cpumask, b->cpumask))
3354 * init_worker_pool - initialize a newly zalloc'd worker_pool
3355 * @pool: worker_pool to initialize
3357 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3358 * Returns 0 on success, -errno on failure. Even on failure, all fields
3359 * inside @pool proper are initialized and put_unbound_pool() can be called
3360 * on @pool safely to release it.
3362 static int init_worker_pool(struct worker_pool *pool)
3364 spin_lock_init(&pool->lock);
3367 pool->node = NUMA_NO_NODE;
3368 pool->flags |= POOL_DISASSOCIATED;
3369 INIT_LIST_HEAD(&pool->worklist);
3370 INIT_LIST_HEAD(&pool->idle_list);
3371 hash_init(pool->busy_hash);
3373 init_timer_deferrable(&pool->idle_timer);
3374 pool->idle_timer.function = idle_worker_timeout;
3375 pool->idle_timer.data = (unsigned long)pool;
3377 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3378 (unsigned long)pool);
3380 mutex_init(&pool->manager_arb);
3381 mutex_init(&pool->manager_mutex);
3382 idr_init(&pool->worker_idr);
3384 INIT_HLIST_NODE(&pool->hash_node);
3387 /* shouldn't fail above this point */
3388 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3394 static void rcu_free_pool(struct rcu_head *rcu)
3396 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3398 idr_destroy(&pool->worker_idr);
3399 free_workqueue_attrs(pool->attrs);
3404 * put_unbound_pool - put a worker_pool
3405 * @pool: worker_pool to put
3407 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3408 * safe manner. get_unbound_pool() calls this function on its failure path
3409 * and this function should be able to release pools which went through,
3410 * successfully or not, init_worker_pool().
3412 * Should be called with wq_pool_mutex held.
3414 static void put_unbound_pool(struct worker_pool *pool)
3416 struct worker *worker;
3418 lockdep_assert_held(&wq_pool_mutex);
3424 if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3425 WARN_ON(!list_empty(&pool->worklist)))
3428 /* release id and unhash */
3430 idr_remove(&worker_pool_idr, pool->id);
3431 hash_del(&pool->hash_node);
3434 * Become the manager and destroy all workers. Grabbing
3435 * manager_arb prevents @pool's workers from blocking on
3438 mutex_lock(&pool->manager_arb);
3439 mutex_lock(&pool->manager_mutex);
3440 spin_lock_irq(&pool->lock);
3442 while ((worker = first_worker(pool)))
3443 destroy_worker(worker);
3444 WARN_ON(pool->nr_workers || pool->nr_idle);
3446 spin_unlock_irq(&pool->lock);
3447 mutex_unlock(&pool->manager_mutex);
3448 mutex_unlock(&pool->manager_arb);
3450 /* shut down the timers */
3451 del_timer_sync(&pool->idle_timer);
3452 del_timer_sync(&pool->mayday_timer);
3454 /* sched-RCU protected to allow dereferences from get_work_pool() */
3455 call_rcu_sched(&pool->rcu, rcu_free_pool);
3459 * get_unbound_pool - get a worker_pool with the specified attributes
3460 * @attrs: the attributes of the worker_pool to get
3462 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3463 * reference count and return it. If there already is a matching
3464 * worker_pool, it will be used; otherwise, this function attempts to
3465 * create a new one. On failure, returns NULL.
3467 * Should be called with wq_pool_mutex held.
3469 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3471 u32 hash = wqattrs_hash(attrs);
3472 struct worker_pool *pool;
3475 lockdep_assert_held(&wq_pool_mutex);
3477 /* do we already have a matching pool? */
3478 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3479 if (wqattrs_equal(pool->attrs, attrs)) {
3485 /* nope, create a new one */
3486 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3487 if (!pool || init_worker_pool(pool) < 0)
3490 if (workqueue_freezing)
3491 pool->flags |= POOL_FREEZING;
3493 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3494 copy_workqueue_attrs(pool->attrs, attrs);
3496 /* if cpumask is contained inside a NUMA node, we belong to that node */
3497 if (wq_numa_enabled) {
3498 for_each_node(node) {
3499 if (cpumask_subset(pool->attrs->cpumask,
3500 wq_numa_possible_cpumask[node])) {
3507 if (worker_pool_assign_id(pool) < 0)
3510 /* create and start the initial worker */
3511 if (create_and_start_worker(pool) < 0)
3515 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3520 put_unbound_pool(pool);
3524 static void rcu_free_pwq(struct rcu_head *rcu)
3526 kmem_cache_free(pwq_cache,
3527 container_of(rcu, struct pool_workqueue, rcu));
3531 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3532 * and needs to be destroyed.
3534 static void pwq_unbound_release_workfn(struct work_struct *work)
3536 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3537 unbound_release_work);
3538 struct workqueue_struct *wq = pwq->wq;
3539 struct worker_pool *pool = pwq->pool;
3542 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3546 * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3547 * necessary on release but do it anyway. It's easier to verify
3548 * and consistent with the linking path.
3550 mutex_lock(&wq->mutex);
3551 list_del_rcu(&pwq->pwqs_node);
3552 is_last = list_empty(&wq->pwqs);
3553 mutex_unlock(&wq->mutex);
3555 mutex_lock(&wq_pool_mutex);
3556 put_unbound_pool(pool);
3557 mutex_unlock(&wq_pool_mutex);
3559 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3562 * If we're the last pwq going away, @wq is already dead and no one
3563 * is gonna access it anymore. Free it.
3566 free_workqueue_attrs(wq->unbound_attrs);
3572 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3573 * @pwq: target pool_workqueue
3575 * If @pwq isn't freezing, set @pwq->max_active to the associated
3576 * workqueue's saved_max_active and activate delayed work items
3577 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3579 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3581 struct workqueue_struct *wq = pwq->wq;
3582 bool freezable = wq->flags & WQ_FREEZABLE;
3584 /* for @wq->saved_max_active */
3585 lockdep_assert_held(&wq->mutex);
3587 /* fast exit for non-freezable wqs */
3588 if (!freezable && pwq->max_active == wq->saved_max_active)
3591 spin_lock_irq(&pwq->pool->lock);
3593 if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3594 pwq->max_active = wq->saved_max_active;
3596 while (!list_empty(&pwq->delayed_works) &&
3597 pwq->nr_active < pwq->max_active)
3598 pwq_activate_first_delayed(pwq);
3601 * Need to kick a worker after thawed or an unbound wq's
3602 * max_active is bumped. It's a slow path. Do it always.
3604 wake_up_worker(pwq->pool);
3606 pwq->max_active = 0;
3609 spin_unlock_irq(&pwq->pool->lock);
3612 static void init_and_link_pwq(struct pool_workqueue *pwq,
3613 struct workqueue_struct *wq,
3614 struct worker_pool *pool,
3615 struct pool_workqueue **p_last_pwq)
3617 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3621 pwq->flush_color = -1;
3623 INIT_LIST_HEAD(&pwq->delayed_works);
3624 INIT_LIST_HEAD(&pwq->mayday_node);
3625 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3627 mutex_lock(&wq->mutex);
3630 * Set the matching work_color. This is synchronized with
3631 * wq->mutex to avoid confusing flush_workqueue().
3634 *p_last_pwq = first_pwq(wq);
3635 pwq->work_color = wq->work_color;
3637 /* sync max_active to the current setting */
3638 pwq_adjust_max_active(pwq);
3641 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3643 if (wq->flags & WQ_UNBOUND)
3644 copy_workqueue_attrs(wq->unbound_attrs, pool->attrs);
3646 mutex_unlock(&wq->mutex);
3650 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3651 * @wq: the target workqueue
3652 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3654 * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the
3655 * current attributes, a new pwq is created and made the first pwq which
3656 * will serve all new work items. Older pwqs are released as in-flight
3657 * work items finish. Note that a work item which repeatedly requeues
3658 * itself back-to-back will stay on its current pwq.
3660 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3663 int apply_workqueue_attrs(struct workqueue_struct *wq,
3664 const struct workqueue_attrs *attrs)
3666 struct workqueue_attrs *new_attrs;
3667 struct pool_workqueue *pwq = NULL, *last_pwq;
3668 struct worker_pool *pool;
3671 /* only unbound workqueues can change attributes */
3672 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3675 /* creating multiple pwqs breaks ordering guarantee */
3676 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3679 /* make a copy of @attrs and sanitize it */
3680 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3684 copy_workqueue_attrs(new_attrs, attrs);
3685 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3687 mutex_lock(&wq_pool_mutex);
3689 pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL);
3691 mutex_unlock(&wq_pool_mutex);
3695 pool = get_unbound_pool(new_attrs);
3697 mutex_unlock(&wq_pool_mutex);
3701 mutex_unlock(&wq_pool_mutex);
3703 init_and_link_pwq(pwq, wq, pool, &last_pwq);
3705 spin_lock_irq(&last_pwq->pool->lock);
3707 spin_unlock_irq(&last_pwq->pool->lock);
3713 free_workqueue_attrs(new_attrs);
3717 kmem_cache_free(pwq_cache, pwq);
3722 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3724 bool highpri = wq->flags & WQ_HIGHPRI;
3727 if (!(wq->flags & WQ_UNBOUND)) {
3728 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3732 for_each_possible_cpu(cpu) {
3733 struct pool_workqueue *pwq =
3734 per_cpu_ptr(wq->cpu_pwqs, cpu);
3735 struct worker_pool *cpu_pools =
3736 per_cpu(cpu_worker_pools, cpu);
3738 init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL);
3742 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3746 static int wq_clamp_max_active(int max_active, unsigned int flags,
3749 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3751 if (max_active < 1 || max_active > lim)
3752 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3753 max_active, name, 1, lim);
3755 return clamp_val(max_active, 1, lim);
3758 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3761 struct lock_class_key *key,
3762 const char *lock_name, ...)
3765 struct workqueue_struct *wq;
3766 struct pool_workqueue *pwq;
3768 /* allocate wq and format name */
3769 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
3773 if (flags & WQ_UNBOUND) {
3774 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3775 if (!wq->unbound_attrs)
3779 va_start(args, lock_name);
3780 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3783 max_active = max_active ?: WQ_DFL_ACTIVE;
3784 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3788 wq->saved_max_active = max_active;
3789 mutex_init(&wq->mutex);
3790 atomic_set(&wq->nr_pwqs_to_flush, 0);
3791 INIT_LIST_HEAD(&wq->pwqs);
3792 INIT_LIST_HEAD(&wq->flusher_queue);
3793 INIT_LIST_HEAD(&wq->flusher_overflow);
3794 INIT_LIST_HEAD(&wq->maydays);
3796 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3797 INIT_LIST_HEAD(&wq->list);
3799 if (alloc_and_link_pwqs(wq) < 0)
3803 * Workqueues which may be used during memory reclaim should
3804 * have a rescuer to guarantee forward progress.
3806 if (flags & WQ_MEM_RECLAIM) {
3807 struct worker *rescuer;
3809 rescuer = alloc_worker();
3813 rescuer->rescue_wq = wq;
3814 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3816 if (IS_ERR(rescuer->task)) {
3821 wq->rescuer = rescuer;
3822 rescuer->task->flags |= PF_NO_SETAFFINITY;
3823 wake_up_process(rescuer->task);
3826 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3830 * wq_pool_mutex protects global freeze state and workqueues list.
3831 * Grab it, adjust max_active and add the new @wq to workqueues
3834 mutex_lock(&wq_pool_mutex);
3836 mutex_lock(&wq->mutex);
3837 for_each_pwq(pwq, wq)
3838 pwq_adjust_max_active(pwq);
3839 mutex_unlock(&wq->mutex);
3841 list_add(&wq->list, &workqueues);
3843 mutex_unlock(&wq_pool_mutex);
3848 free_workqueue_attrs(wq->unbound_attrs);
3852 destroy_workqueue(wq);
3855 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3858 * destroy_workqueue - safely terminate a workqueue
3859 * @wq: target workqueue
3861 * Safely destroy a workqueue. All work currently pending will be done first.
3863 void destroy_workqueue(struct workqueue_struct *wq)
3865 struct pool_workqueue *pwq;
3867 /* drain it before proceeding with destruction */
3868 drain_workqueue(wq);
3871 mutex_lock(&wq->mutex);
3872 for_each_pwq(pwq, wq) {
3875 for (i = 0; i < WORK_NR_COLORS; i++) {
3876 if (WARN_ON(pwq->nr_in_flight[i])) {
3877 mutex_unlock(&wq->mutex);
3882 if (WARN_ON(pwq->refcnt > 1) ||
3883 WARN_ON(pwq->nr_active) ||
3884 WARN_ON(!list_empty(&pwq->delayed_works))) {
3885 mutex_unlock(&wq->mutex);
3889 mutex_unlock(&wq->mutex);
3892 * wq list is used to freeze wq, remove from list after
3893 * flushing is complete in case freeze races us.
3895 mutex_lock(&wq_pool_mutex);
3896 list_del_init(&wq->list);
3897 mutex_unlock(&wq_pool_mutex);
3899 workqueue_sysfs_unregister(wq);
3902 kthread_stop(wq->rescuer->task);
3907 if (!(wq->flags & WQ_UNBOUND)) {
3909 * The base ref is never dropped on per-cpu pwqs. Directly
3910 * free the pwqs and wq.
3912 free_percpu(wq->cpu_pwqs);
3916 * We're the sole accessor of @wq at this point. Directly
3917 * access the first pwq and put the base ref. As both pwqs
3918 * and pools are sched-RCU protected, the lock operations
3919 * are safe. @wq will be freed when the last pwq is
3922 pwq = list_first_entry(&wq->pwqs, struct pool_workqueue,
3924 spin_lock_irq(&pwq->pool->lock);
3926 spin_unlock_irq(&pwq->pool->lock);
3929 EXPORT_SYMBOL_GPL(destroy_workqueue);
3932 * workqueue_set_max_active - adjust max_active of a workqueue
3933 * @wq: target workqueue
3934 * @max_active: new max_active value.
3936 * Set max_active of @wq to @max_active.
3939 * Don't call from IRQ context.
3941 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3943 struct pool_workqueue *pwq;
3945 /* disallow meddling with max_active for ordered workqueues */
3946 if (WARN_ON(wq->flags & __WQ_ORDERED))
3949 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3951 mutex_lock(&wq->mutex);
3953 wq->saved_max_active = max_active;
3955 for_each_pwq(pwq, wq)
3956 pwq_adjust_max_active(pwq);
3958 mutex_unlock(&wq->mutex);
3960 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3963 * current_is_workqueue_rescuer - is %current workqueue rescuer?
3965 * Determine whether %current is a workqueue rescuer. Can be used from
3966 * work functions to determine whether it's being run off the rescuer task.
3968 bool current_is_workqueue_rescuer(void)
3970 struct worker *worker = current_wq_worker();
3972 return worker && worker->rescue_wq;
3976 * workqueue_congested - test whether a workqueue is congested
3977 * @cpu: CPU in question
3978 * @wq: target workqueue
3980 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3981 * no synchronization around this function and the test result is
3982 * unreliable and only useful as advisory hints or for debugging.
3985 * %true if congested, %false otherwise.
3987 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
3989 struct pool_workqueue *pwq;
3992 rcu_read_lock_sched();
3994 if (!(wq->flags & WQ_UNBOUND))
3995 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
3997 pwq = first_pwq(wq);
3999 ret = !list_empty(&pwq->delayed_works);
4000 rcu_read_unlock_sched();
4004 EXPORT_SYMBOL_GPL(workqueue_congested);
4007 * work_busy - test whether a work is currently pending or running
4008 * @work: the work to be tested
4010 * Test whether @work is currently pending or running. There is no
4011 * synchronization around this function and the test result is
4012 * unreliable and only useful as advisory hints or for debugging.
4015 * OR'd bitmask of WORK_BUSY_* bits.
4017 unsigned int work_busy(struct work_struct *work)
4019 struct worker_pool *pool;
4020 unsigned long flags;
4021 unsigned int ret = 0;
4023 if (work_pending(work))
4024 ret |= WORK_BUSY_PENDING;
4026 local_irq_save(flags);
4027 pool = get_work_pool(work);
4029 spin_lock(&pool->lock);
4030 if (find_worker_executing_work(pool, work))
4031 ret |= WORK_BUSY_RUNNING;
4032 spin_unlock(&pool->lock);
4034 local_irq_restore(flags);
4038 EXPORT_SYMBOL_GPL(work_busy);
4043 * There are two challenges in supporting CPU hotplug. Firstly, there
4044 * are a lot of assumptions on strong associations among work, pwq and
4045 * pool which make migrating pending and scheduled works very
4046 * difficult to implement without impacting hot paths. Secondly,
4047 * worker pools serve mix of short, long and very long running works making
4048 * blocked draining impractical.
4050 * This is solved by allowing the pools to be disassociated from the CPU
4051 * running as an unbound one and allowing it to be reattached later if the
4052 * cpu comes back online.
4055 static void wq_unbind_fn(struct work_struct *work)
4057 int cpu = smp_processor_id();
4058 struct worker_pool *pool;
4059 struct worker *worker;
4062 for_each_cpu_worker_pool(pool, cpu) {
4063 WARN_ON_ONCE(cpu != smp_processor_id());
4065 mutex_lock(&pool->manager_mutex);
4066 spin_lock_irq(&pool->lock);
4069 * We've blocked all manager operations. Make all workers
4070 * unbound and set DISASSOCIATED. Before this, all workers
4071 * except for the ones which are still executing works from
4072 * before the last CPU down must be on the cpu. After
4073 * this, they may become diasporas.
4075 for_each_pool_worker(worker, wi, pool)
4076 worker->flags |= WORKER_UNBOUND;
4078 pool->flags |= POOL_DISASSOCIATED;
4080 spin_unlock_irq(&pool->lock);
4081 mutex_unlock(&pool->manager_mutex);
4085 * Call schedule() so that we cross rq->lock and thus can guarantee
4086 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
4087 * as scheduler callbacks may be invoked from other cpus.
4092 * Sched callbacks are disabled now. Zap nr_running. After this,
4093 * nr_running stays zero and need_more_worker() and keep_working()
4094 * are always true as long as the worklist is not empty. Pools on
4095 * @cpu now behave as unbound (in terms of concurrency management)
4096 * pools which are served by workers tied to the CPU.
4098 * On return from this function, the current worker would trigger
4099 * unbound chain execution of pending work items if other workers
4102 for_each_cpu_worker_pool(pool, cpu)
4103 atomic_set(&pool->nr_running, 0);
4107 * rebind_workers - rebind all workers of a pool to the associated CPU
4108 * @pool: pool of interest
4110 * @pool->cpu is coming online. Rebind all workers to the CPU.
4112 static void rebind_workers(struct worker_pool *pool)
4114 struct worker *worker;
4117 lockdep_assert_held(&pool->manager_mutex);
4120 * Restore CPU affinity of all workers. As all idle workers should
4121 * be on the run-queue of the associated CPU before any local
4122 * wake-ups for concurrency management happen, restore CPU affinty
4123 * of all workers first and then clear UNBOUND. As we're called
4124 * from CPU_ONLINE, the following shouldn't fail.
4126 for_each_pool_worker(worker, wi, pool)
4127 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4128 pool->attrs->cpumask) < 0);
4130 spin_lock_irq(&pool->lock);
4132 for_each_pool_worker(worker, wi, pool) {
4133 unsigned int worker_flags = worker->flags;
4136 * A bound idle worker should actually be on the runqueue
4137 * of the associated CPU for local wake-ups targeting it to
4138 * work. Kick all idle workers so that they migrate to the
4139 * associated CPU. Doing this in the same loop as
4140 * replacing UNBOUND with REBOUND is safe as no worker will
4141 * be bound before @pool->lock is released.
4143 if (worker_flags & WORKER_IDLE)
4144 wake_up_process(worker->task);
4147 * We want to clear UNBOUND but can't directly call
4148 * worker_clr_flags() or adjust nr_running. Atomically
4149 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4150 * @worker will clear REBOUND using worker_clr_flags() when
4151 * it initiates the next execution cycle thus restoring
4152 * concurrency management. Note that when or whether
4153 * @worker clears REBOUND doesn't affect correctness.
4155 * ACCESS_ONCE() is necessary because @worker->flags may be
4156 * tested without holding any lock in
4157 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4158 * fail incorrectly leading to premature concurrency
4159 * management operations.
4161 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4162 worker_flags |= WORKER_REBOUND;
4163 worker_flags &= ~WORKER_UNBOUND;
4164 ACCESS_ONCE(worker->flags) = worker_flags;
4167 spin_unlock_irq(&pool->lock);
4171 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4172 * @pool: unbound pool of interest
4173 * @cpu: the CPU which is coming up
4175 * An unbound pool may end up with a cpumask which doesn't have any online
4176 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4177 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4178 * online CPU before, cpus_allowed of all its workers should be restored.
4180 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4182 static cpumask_t cpumask;
4183 struct worker *worker;
4186 lockdep_assert_held(&pool->manager_mutex);
4188 /* is @cpu allowed for @pool? */
4189 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4192 /* is @cpu the only online CPU? */
4193 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4194 if (cpumask_weight(&cpumask) != 1)
4197 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4198 for_each_pool_worker(worker, wi, pool)
4199 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4200 pool->attrs->cpumask) < 0);
4204 * Workqueues should be brought up before normal priority CPU notifiers.
4205 * This will be registered high priority CPU notifier.
4207 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
4208 unsigned long action,
4211 int cpu = (unsigned long)hcpu;
4212 struct worker_pool *pool;
4215 switch (action & ~CPU_TASKS_FROZEN) {
4216 case CPU_UP_PREPARE:
4217 for_each_cpu_worker_pool(pool, cpu) {
4218 if (pool->nr_workers)
4220 if (create_and_start_worker(pool) < 0)
4225 case CPU_DOWN_FAILED:
4227 mutex_lock(&wq_pool_mutex);
4229 for_each_pool(pool, pi) {
4230 mutex_lock(&pool->manager_mutex);
4232 if (pool->cpu == cpu) {
4233 spin_lock_irq(&pool->lock);
4234 pool->flags &= ~POOL_DISASSOCIATED;
4235 spin_unlock_irq(&pool->lock);
4237 rebind_workers(pool);
4238 } else if (pool->cpu < 0) {
4239 restore_unbound_workers_cpumask(pool, cpu);
4242 mutex_unlock(&pool->manager_mutex);
4245 mutex_unlock(&wq_pool_mutex);
4252 * Workqueues should be brought down after normal priority CPU notifiers.
4253 * This will be registered as low priority CPU notifier.
4255 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
4256 unsigned long action,
4259 int cpu = (unsigned long)hcpu;
4260 struct work_struct unbind_work;
4262 switch (action & ~CPU_TASKS_FROZEN) {
4263 case CPU_DOWN_PREPARE:
4264 /* unbinding should happen on the local CPU */
4265 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4266 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4267 flush_work(&unbind_work);
4275 struct work_for_cpu {
4276 struct work_struct work;
4282 static void work_for_cpu_fn(struct work_struct *work)
4284 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4286 wfc->ret = wfc->fn(wfc->arg);
4290 * work_on_cpu - run a function in user context on a particular cpu
4291 * @cpu: the cpu to run on
4292 * @fn: the function to run
4293 * @arg: the function arg
4295 * This will return the value @fn returns.
4296 * It is up to the caller to ensure that the cpu doesn't go offline.
4297 * The caller must not hold any locks which would prevent @fn from completing.
4299 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4301 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4303 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4304 schedule_work_on(cpu, &wfc.work);
4305 flush_work(&wfc.work);
4308 EXPORT_SYMBOL_GPL(work_on_cpu);
4309 #endif /* CONFIG_SMP */
4311 #ifdef CONFIG_FREEZER
4314 * freeze_workqueues_begin - begin freezing workqueues
4316 * Start freezing workqueues. After this function returns, all freezable
4317 * workqueues will queue new works to their delayed_works list instead of
4321 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4323 void freeze_workqueues_begin(void)
4325 struct worker_pool *pool;
4326 struct workqueue_struct *wq;
4327 struct pool_workqueue *pwq;
4330 mutex_lock(&wq_pool_mutex);
4332 WARN_ON_ONCE(workqueue_freezing);
4333 workqueue_freezing = true;
4336 for_each_pool(pool, pi) {
4337 spin_lock_irq(&pool->lock);
4338 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4339 pool->flags |= POOL_FREEZING;
4340 spin_unlock_irq(&pool->lock);
4343 list_for_each_entry(wq, &workqueues, list) {
4344 mutex_lock(&wq->mutex);
4345 for_each_pwq(pwq, wq)
4346 pwq_adjust_max_active(pwq);
4347 mutex_unlock(&wq->mutex);
4350 mutex_unlock(&wq_pool_mutex);
4354 * freeze_workqueues_busy - are freezable workqueues still busy?
4356 * Check whether freezing is complete. This function must be called
4357 * between freeze_workqueues_begin() and thaw_workqueues().
4360 * Grabs and releases wq_pool_mutex.
4363 * %true if some freezable workqueues are still busy. %false if freezing
4366 bool freeze_workqueues_busy(void)
4369 struct workqueue_struct *wq;
4370 struct pool_workqueue *pwq;
4372 mutex_lock(&wq_pool_mutex);
4374 WARN_ON_ONCE(!workqueue_freezing);
4376 list_for_each_entry(wq, &workqueues, list) {
4377 if (!(wq->flags & WQ_FREEZABLE))
4380 * nr_active is monotonically decreasing. It's safe
4381 * to peek without lock.
4383 rcu_read_lock_sched();
4384 for_each_pwq(pwq, wq) {
4385 WARN_ON_ONCE(pwq->nr_active < 0);
4386 if (pwq->nr_active) {
4388 rcu_read_unlock_sched();
4392 rcu_read_unlock_sched();
4395 mutex_unlock(&wq_pool_mutex);
4400 * thaw_workqueues - thaw workqueues
4402 * Thaw workqueues. Normal queueing is restored and all collected
4403 * frozen works are transferred to their respective pool worklists.
4406 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4408 void thaw_workqueues(void)
4410 struct workqueue_struct *wq;
4411 struct pool_workqueue *pwq;
4412 struct worker_pool *pool;
4415 mutex_lock(&wq_pool_mutex);
4417 if (!workqueue_freezing)
4420 /* clear FREEZING */
4421 for_each_pool(pool, pi) {
4422 spin_lock_irq(&pool->lock);
4423 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4424 pool->flags &= ~POOL_FREEZING;
4425 spin_unlock_irq(&pool->lock);
4428 /* restore max_active and repopulate worklist */
4429 list_for_each_entry(wq, &workqueues, list) {
4430 mutex_lock(&wq->mutex);
4431 for_each_pwq(pwq, wq)
4432 pwq_adjust_max_active(pwq);
4433 mutex_unlock(&wq->mutex);
4436 workqueue_freezing = false;
4438 mutex_unlock(&wq_pool_mutex);
4440 #endif /* CONFIG_FREEZER */
4442 static void __init wq_numa_init(void)
4447 /* determine NUMA pwq table len - highest node id + 1 */
4449 wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4451 if (num_possible_nodes() <= 1)
4455 * We want masks of possible CPUs of each node which isn't readily
4456 * available. Build one from cpu_to_node() which should have been
4457 * fully initialized by now.
4459 tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4463 BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, node));
4465 for_each_possible_cpu(cpu) {
4466 node = cpu_to_node(cpu);
4467 if (WARN_ON(node == NUMA_NO_NODE)) {
4468 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4469 /* happens iff arch is bonkers, let's just proceed */
4472 cpumask_set_cpu(cpu, tbl[node]);
4475 wq_numa_possible_cpumask = tbl;
4476 wq_numa_enabled = true;
4479 static int __init init_workqueues(void)
4481 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4484 /* make sure we have enough bits for OFFQ pool ID */
4485 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4486 WORK_CPU_END * NR_STD_WORKER_POOLS);
4488 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4490 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4492 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4493 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4497 /* initialize CPU pools */
4498 for_each_possible_cpu(cpu) {
4499 struct worker_pool *pool;
4502 for_each_cpu_worker_pool(pool, cpu) {
4503 BUG_ON(init_worker_pool(pool));
4505 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4506 pool->attrs->nice = std_nice[i++];
4507 pool->node = cpu_to_node(cpu);
4510 mutex_lock(&wq_pool_mutex);
4511 BUG_ON(worker_pool_assign_id(pool));
4512 mutex_unlock(&wq_pool_mutex);
4516 /* create the initial worker */
4517 for_each_online_cpu(cpu) {
4518 struct worker_pool *pool;
4520 for_each_cpu_worker_pool(pool, cpu) {
4521 pool->flags &= ~POOL_DISASSOCIATED;
4522 BUG_ON(create_and_start_worker(pool) < 0);
4526 /* create default unbound wq attrs */
4527 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4528 struct workqueue_attrs *attrs;
4530 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4531 attrs->nice = std_nice[i];
4532 unbound_std_wq_attrs[i] = attrs;
4535 system_wq = alloc_workqueue("events", 0, 0);
4536 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4537 system_long_wq = alloc_workqueue("events_long", 0, 0);
4538 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4539 WQ_UNBOUND_MAX_ACTIVE);
4540 system_freezable_wq = alloc_workqueue("events_freezable",
4542 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4543 !system_unbound_wq || !system_freezable_wq);
4546 early_initcall(init_workqueues);