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>
48 #include "workqueue_internal.h"
54 * A bound pool is either associated or disassociated with its CPU.
55 * While associated (!DISASSOCIATED), all workers are bound to the
56 * CPU and none has %WORKER_UNBOUND set and concurrency management
59 * While DISASSOCIATED, the cpu may be offline and all workers have
60 * %WORKER_UNBOUND set and concurrency management disabled, and may
61 * be executing on any CPU. The pool behaves as an unbound one.
63 * Note that DISASSOCIATED can be flipped only while holding
64 * assoc_mutex to avoid changing binding state while
65 * create_worker() is in progress.
67 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
68 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
69 POOL_FREEZING
= 1 << 3, /* freeze in progress */
72 WORKER_STARTED
= 1 << 0, /* started */
73 WORKER_DIE
= 1 << 1, /* die die die */
74 WORKER_IDLE
= 1 << 2, /* is idle */
75 WORKER_PREP
= 1 << 3, /* preparing to run works */
76 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
77 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
79 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_UNBOUND
|
82 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
84 UNBOUND_POOL_HASH_ORDER
= 6, /* hashed by pool->attrs */
85 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
87 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
88 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
90 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
91 /* call for help after 10ms
93 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
94 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
97 * Rescue workers are used only on emergencies and shared by
100 RESCUER_NICE_LEVEL
= -20,
101 HIGHPRI_NICE_LEVEL
= -20,
105 * Structure fields follow one of the following exclusion rules.
107 * I: Modifiable by initialization/destruction paths and read-only for
110 * P: Preemption protected. Disabling preemption is enough and should
111 * only be modified and accessed from the local cpu.
113 * L: pool->lock protected. Access with pool->lock held.
115 * X: During normal operation, modification requires pool->lock and should
116 * be done only from local cpu. Either disabling preemption on local
117 * cpu or grabbing pool->lock is enough for read access. If
118 * POOL_DISASSOCIATED is set, it's identical to L.
120 * F: wq->flush_mutex protected.
122 * W: workqueue_lock protected.
124 * R: workqueue_lock protected for writes. Sched-RCU protected for reads.
126 * FR: wq->flush_mutex and workqueue_lock protected for writes. Sched-RCU
127 * protected for reads.
130 /* struct worker is defined in workqueue_internal.h */
133 spinlock_t lock
; /* the pool lock */
134 int cpu
; /* I: the associated cpu */
135 int id
; /* I: pool ID */
136 unsigned int flags
; /* X: flags */
138 struct list_head worklist
; /* L: list of pending works */
139 int nr_workers
; /* L: total number of workers */
141 /* nr_idle includes the ones off idle_list for rebinding */
142 int nr_idle
; /* L: currently idle ones */
144 struct list_head idle_list
; /* X: list of idle workers */
145 struct timer_list idle_timer
; /* L: worker idle timeout */
146 struct timer_list mayday_timer
; /* L: SOS timer for workers */
148 /* workers are chained either in busy_hash or idle_list */
149 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
150 /* L: hash of busy workers */
152 struct mutex manager_arb
; /* manager arbitration */
153 struct mutex assoc_mutex
; /* protect POOL_DISASSOCIATED */
154 struct ida worker_ida
; /* L: for worker IDs */
156 struct workqueue_attrs
*attrs
; /* I: worker attributes */
157 struct hlist_node hash_node
; /* R: unbound_pool_hash node */
158 int refcnt
; /* refcnt for unbound pools */
161 * The current concurrency level. As it's likely to be accessed
162 * from other CPUs during try_to_wake_up(), put it in a separate
165 atomic_t nr_running ____cacheline_aligned_in_smp
;
168 * Destruction of pool is sched-RCU protected to allow dereferences
169 * from get_work_pool().
172 } ____cacheline_aligned_in_smp
;
175 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
176 * of work_struct->data are used for flags and the remaining high bits
177 * point to the pwq; thus, pwqs need to be aligned at two's power of the
178 * number of flag bits.
180 struct pool_workqueue
{
181 struct worker_pool
*pool
; /* I: the associated pool */
182 struct workqueue_struct
*wq
; /* I: the owning workqueue */
183 int work_color
; /* L: current color */
184 int flush_color
; /* L: flushing color */
185 int refcnt
; /* L: reference count */
186 int nr_in_flight
[WORK_NR_COLORS
];
187 /* L: nr of in_flight works */
188 int nr_active
; /* L: nr of active works */
189 int max_active
; /* L: max active works */
190 struct list_head delayed_works
; /* L: delayed works */
191 struct list_head pwqs_node
; /* FR: node on wq->pwqs */
192 struct list_head mayday_node
; /* W: node on wq->maydays */
195 * Release of unbound pwq is punted to system_wq. See put_pwq()
196 * and pwq_unbound_release_workfn() for details. pool_workqueue
197 * itself is also sched-RCU protected so that the first pwq can be
198 * determined without grabbing workqueue_lock.
200 struct work_struct unbound_release_work
;
202 } __aligned(1 << WORK_STRUCT_FLAG_BITS
);
205 * Structure used to wait for workqueue flush.
208 struct list_head list
; /* F: list of flushers */
209 int flush_color
; /* F: flush color waiting for */
210 struct completion done
; /* flush completion */
216 * The externally visible workqueue abstraction is an array of
217 * per-CPU workqueues:
219 struct workqueue_struct
{
220 unsigned int flags
; /* W: WQ_* flags */
221 struct pool_workqueue __percpu
*cpu_pwqs
; /* I: per-cpu pwq's */
222 struct list_head pwqs
; /* FR: all pwqs of this wq */
223 struct list_head list
; /* W: list of all workqueues */
225 struct mutex flush_mutex
; /* protects wq flushing */
226 int work_color
; /* F: current work color */
227 int flush_color
; /* F: current flush color */
228 atomic_t nr_pwqs_to_flush
; /* flush in progress */
229 struct wq_flusher
*first_flusher
; /* F: first flusher */
230 struct list_head flusher_queue
; /* F: flush waiters */
231 struct list_head flusher_overflow
; /* F: flush overflow list */
233 struct list_head maydays
; /* W: pwqs requesting rescue */
234 struct worker
*rescuer
; /* I: rescue worker */
236 int nr_drainers
; /* W: drain in progress */
237 int saved_max_active
; /* W: saved pwq max_active */
240 struct wq_device
*wq_dev
; /* I: for sysfs interface */
242 #ifdef CONFIG_LOCKDEP
243 struct lockdep_map lockdep_map
;
245 char name
[]; /* I: workqueue name */
248 static struct kmem_cache
*pwq_cache
;
250 /* hash of all unbound pools keyed by pool->attrs */
251 static DEFINE_HASHTABLE(unbound_pool_hash
, UNBOUND_POOL_HASH_ORDER
);
253 static struct workqueue_attrs
*unbound_std_wq_attrs
[NR_STD_WORKER_POOLS
];
255 struct workqueue_struct
*system_wq __read_mostly
;
256 EXPORT_SYMBOL_GPL(system_wq
);
257 struct workqueue_struct
*system_highpri_wq __read_mostly
;
258 EXPORT_SYMBOL_GPL(system_highpri_wq
);
259 struct workqueue_struct
*system_long_wq __read_mostly
;
260 EXPORT_SYMBOL_GPL(system_long_wq
);
261 struct workqueue_struct
*system_unbound_wq __read_mostly
;
262 EXPORT_SYMBOL_GPL(system_unbound_wq
);
263 struct workqueue_struct
*system_freezable_wq __read_mostly
;
264 EXPORT_SYMBOL_GPL(system_freezable_wq
);
266 #define CREATE_TRACE_POINTS
267 #include <trace/events/workqueue.h>
269 #define assert_rcu_or_wq_lock() \
270 rcu_lockdep_assert(rcu_read_lock_sched_held() || \
271 lockdep_is_held(&workqueue_lock), \
272 "sched RCU or workqueue lock should be held")
274 #define for_each_cpu_worker_pool(pool, cpu) \
275 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
276 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
279 #define for_each_busy_worker(worker, i, pool) \
280 hash_for_each(pool->busy_hash, i, worker, hentry)
283 * for_each_pool - iterate through all worker_pools in the system
284 * @pool: iteration cursor
285 * @id: integer used for iteration
287 * This must be called either with workqueue_lock held or sched RCU read
288 * locked. If the pool needs to be used beyond the locking in effect, the
289 * caller is responsible for guaranteeing that the pool stays online.
291 * The if/else clause exists only for the lockdep assertion and can be
294 #define for_each_pool(pool, id) \
295 idr_for_each_entry(&worker_pool_idr, pool, id) \
296 if (({ assert_rcu_or_wq_lock(); false; })) { } \
300 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
301 * @pwq: iteration cursor
302 * @wq: the target workqueue
304 * This must be called either with workqueue_lock held or sched RCU read
305 * locked. If the pwq needs to be used beyond the locking in effect, the
306 * caller is responsible for guaranteeing that the pwq stays online.
308 * The if/else clause exists only for the lockdep assertion and can be
311 #define for_each_pwq(pwq, wq) \
312 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
313 if (({ assert_rcu_or_wq_lock(); false; })) { } \
316 #ifdef CONFIG_DEBUG_OBJECTS_WORK
318 static struct debug_obj_descr work_debug_descr
;
320 static void *work_debug_hint(void *addr
)
322 return ((struct work_struct
*) addr
)->func
;
326 * fixup_init is called when:
327 * - an active object is initialized
329 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
331 struct work_struct
*work
= addr
;
334 case ODEBUG_STATE_ACTIVE
:
335 cancel_work_sync(work
);
336 debug_object_init(work
, &work_debug_descr
);
344 * fixup_activate is called when:
345 * - an active object is activated
346 * - an unknown object is activated (might be a statically initialized object)
348 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
350 struct work_struct
*work
= addr
;
354 case ODEBUG_STATE_NOTAVAILABLE
:
356 * This is not really a fixup. The work struct was
357 * statically initialized. We just make sure that it
358 * is tracked in the object tracker.
360 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
361 debug_object_init(work
, &work_debug_descr
);
362 debug_object_activate(work
, &work_debug_descr
);
368 case ODEBUG_STATE_ACTIVE
:
377 * fixup_free is called when:
378 * - an active object is freed
380 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
382 struct work_struct
*work
= addr
;
385 case ODEBUG_STATE_ACTIVE
:
386 cancel_work_sync(work
);
387 debug_object_free(work
, &work_debug_descr
);
394 static struct debug_obj_descr work_debug_descr
= {
395 .name
= "work_struct",
396 .debug_hint
= work_debug_hint
,
397 .fixup_init
= work_fixup_init
,
398 .fixup_activate
= work_fixup_activate
,
399 .fixup_free
= work_fixup_free
,
402 static inline void debug_work_activate(struct work_struct
*work
)
404 debug_object_activate(work
, &work_debug_descr
);
407 static inline void debug_work_deactivate(struct work_struct
*work
)
409 debug_object_deactivate(work
, &work_debug_descr
);
412 void __init_work(struct work_struct
*work
, int onstack
)
415 debug_object_init_on_stack(work
, &work_debug_descr
);
417 debug_object_init(work
, &work_debug_descr
);
419 EXPORT_SYMBOL_GPL(__init_work
);
421 void destroy_work_on_stack(struct work_struct
*work
)
423 debug_object_free(work
, &work_debug_descr
);
425 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
428 static inline void debug_work_activate(struct work_struct
*work
) { }
429 static inline void debug_work_deactivate(struct work_struct
*work
) { }
432 /* Serializes the accesses to the list of workqueues. */
433 static DEFINE_SPINLOCK(workqueue_lock
);
434 static LIST_HEAD(workqueues
);
435 static bool workqueue_freezing
; /* W: have wqs started freezing? */
438 * The CPU and unbound standard worker pools. The unbound ones have
439 * POOL_DISASSOCIATED set, and their workers have WORKER_UNBOUND set.
441 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
],
445 * idr of all pools. Modifications are protected by workqueue_lock. Read
446 * accesses are protected by sched-RCU protected.
448 static DEFINE_IDR(worker_pool_idr
);
450 static int worker_thread(void *__worker
);
451 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
452 const struct workqueue_attrs
*from
);
454 /* allocate ID and assign it to @pool */
455 static int worker_pool_assign_id(struct worker_pool
*pool
)
460 if (!idr_pre_get(&worker_pool_idr
, GFP_KERNEL
))
463 spin_lock_irq(&workqueue_lock
);
464 ret
= idr_get_new(&worker_pool_idr
, pool
, &pool
->id
);
465 spin_unlock_irq(&workqueue_lock
);
466 } while (ret
== -EAGAIN
);
472 * first_pwq - return the first pool_workqueue of the specified workqueue
473 * @wq: the target workqueue
475 * This must be called either with workqueue_lock held or sched RCU read
476 * locked. If the pwq needs to be used beyond the locking in effect, the
477 * caller is responsible for guaranteeing that the pwq stays online.
479 static struct pool_workqueue
*first_pwq(struct workqueue_struct
*wq
)
481 assert_rcu_or_wq_lock();
482 return list_first_or_null_rcu(&wq
->pwqs
, struct pool_workqueue
,
486 static unsigned int work_color_to_flags(int color
)
488 return color
<< WORK_STRUCT_COLOR_SHIFT
;
491 static int get_work_color(struct work_struct
*work
)
493 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
494 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
497 static int work_next_color(int color
)
499 return (color
+ 1) % WORK_NR_COLORS
;
503 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
504 * contain the pointer to the queued pwq. Once execution starts, the flag
505 * is cleared and the high bits contain OFFQ flags and pool ID.
507 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
508 * and clear_work_data() can be used to set the pwq, pool or clear
509 * work->data. These functions should only be called while the work is
510 * owned - ie. while the PENDING bit is set.
512 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
513 * corresponding to a work. Pool is available once the work has been
514 * queued anywhere after initialization until it is sync canceled. pwq is
515 * available only while the work item is queued.
517 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
518 * canceled. While being canceled, a work item may have its PENDING set
519 * but stay off timer and worklist for arbitrarily long and nobody should
520 * try to steal the PENDING bit.
522 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
525 WARN_ON_ONCE(!work_pending(work
));
526 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
529 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
530 unsigned long extra_flags
)
532 set_work_data(work
, (unsigned long)pwq
,
533 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
536 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
539 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
540 WORK_STRUCT_PENDING
);
543 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
547 * The following wmb is paired with the implied mb in
548 * test_and_set_bit(PENDING) and ensures all updates to @work made
549 * here are visible to and precede any updates by the next PENDING
553 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
556 static void clear_work_data(struct work_struct
*work
)
558 smp_wmb(); /* see set_work_pool_and_clear_pending() */
559 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
562 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
564 unsigned long data
= atomic_long_read(&work
->data
);
566 if (data
& WORK_STRUCT_PWQ
)
567 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
573 * get_work_pool - return the worker_pool a given work was associated with
574 * @work: the work item of interest
576 * Return the worker_pool @work was last associated with. %NULL if none.
578 * Pools are created and destroyed under workqueue_lock, and allows read
579 * access under sched-RCU read lock. As such, this function should be
580 * called under workqueue_lock or with preemption disabled.
582 * All fields of the returned pool are accessible as long as the above
583 * mentioned locking is in effect. If the returned pool needs to be used
584 * beyond the critical section, the caller is responsible for ensuring the
585 * returned pool is and stays online.
587 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
589 unsigned long data
= atomic_long_read(&work
->data
);
592 assert_rcu_or_wq_lock();
594 if (data
& WORK_STRUCT_PWQ
)
595 return ((struct pool_workqueue
*)
596 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
598 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
599 if (pool_id
== WORK_OFFQ_POOL_NONE
)
602 return idr_find(&worker_pool_idr
, pool_id
);
606 * get_work_pool_id - return the worker pool ID a given work is associated with
607 * @work: the work item of interest
609 * Return the worker_pool ID @work was last associated with.
610 * %WORK_OFFQ_POOL_NONE if none.
612 static int get_work_pool_id(struct work_struct
*work
)
614 unsigned long data
= atomic_long_read(&work
->data
);
616 if (data
& WORK_STRUCT_PWQ
)
617 return ((struct pool_workqueue
*)
618 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
620 return data
>> WORK_OFFQ_POOL_SHIFT
;
623 static void mark_work_canceling(struct work_struct
*work
)
625 unsigned long pool_id
= get_work_pool_id(work
);
627 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
628 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
631 static bool work_is_canceling(struct work_struct
*work
)
633 unsigned long data
= atomic_long_read(&work
->data
);
635 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
639 * Policy functions. These define the policies on how the global worker
640 * pools are managed. Unless noted otherwise, these functions assume that
641 * they're being called with pool->lock held.
644 static bool __need_more_worker(struct worker_pool
*pool
)
646 return !atomic_read(&pool
->nr_running
);
650 * Need to wake up a worker? Called from anything but currently
653 * Note that, because unbound workers never contribute to nr_running, this
654 * function will always return %true for unbound pools as long as the
655 * worklist isn't empty.
657 static bool need_more_worker(struct worker_pool
*pool
)
659 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
662 /* Can I start working? Called from busy but !running workers. */
663 static bool may_start_working(struct worker_pool
*pool
)
665 return pool
->nr_idle
;
668 /* Do I need to keep working? Called from currently running workers. */
669 static bool keep_working(struct worker_pool
*pool
)
671 return !list_empty(&pool
->worklist
) &&
672 atomic_read(&pool
->nr_running
) <= 1;
675 /* Do we need a new worker? Called from manager. */
676 static bool need_to_create_worker(struct worker_pool
*pool
)
678 return need_more_worker(pool
) && !may_start_working(pool
);
681 /* Do I need to be the manager? */
682 static bool need_to_manage_workers(struct worker_pool
*pool
)
684 return need_to_create_worker(pool
) ||
685 (pool
->flags
& POOL_MANAGE_WORKERS
);
688 /* Do we have too many workers and should some go away? */
689 static bool too_many_workers(struct worker_pool
*pool
)
691 bool managing
= mutex_is_locked(&pool
->manager_arb
);
692 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
693 int nr_busy
= pool
->nr_workers
- nr_idle
;
696 * nr_idle and idle_list may disagree if idle rebinding is in
697 * progress. Never return %true if idle_list is empty.
699 if (list_empty(&pool
->idle_list
))
702 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
709 /* Return the first worker. Safe with preemption disabled */
710 static struct worker
*first_worker(struct worker_pool
*pool
)
712 if (unlikely(list_empty(&pool
->idle_list
)))
715 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
719 * wake_up_worker - wake up an idle worker
720 * @pool: worker pool to wake worker from
722 * Wake up the first idle worker of @pool.
725 * spin_lock_irq(pool->lock).
727 static void wake_up_worker(struct worker_pool
*pool
)
729 struct worker
*worker
= first_worker(pool
);
732 wake_up_process(worker
->task
);
736 * wq_worker_waking_up - a worker is waking up
737 * @task: task waking up
738 * @cpu: CPU @task is waking up to
740 * This function is called during try_to_wake_up() when a worker is
744 * spin_lock_irq(rq->lock)
746 void wq_worker_waking_up(struct task_struct
*task
, int cpu
)
748 struct worker
*worker
= kthread_data(task
);
750 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
751 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
752 atomic_inc(&worker
->pool
->nr_running
);
757 * wq_worker_sleeping - a worker is going to sleep
758 * @task: task going to sleep
759 * @cpu: CPU in question, must be the current CPU number
761 * This function is called during schedule() when a busy worker is
762 * going to sleep. Worker on the same cpu can be woken up by
763 * returning pointer to its task.
766 * spin_lock_irq(rq->lock)
769 * Worker task on @cpu to wake up, %NULL if none.
771 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
, int cpu
)
773 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
774 struct worker_pool
*pool
;
777 * Rescuers, which may not have all the fields set up like normal
778 * workers, also reach here, let's not access anything before
779 * checking NOT_RUNNING.
781 if (worker
->flags
& WORKER_NOT_RUNNING
)
786 /* this can only happen on the local cpu */
787 if (WARN_ON_ONCE(cpu
!= raw_smp_processor_id()))
791 * The counterpart of the following dec_and_test, implied mb,
792 * worklist not empty test sequence is in insert_work().
793 * Please read comment there.
795 * NOT_RUNNING is clear. This means that we're bound to and
796 * running on the local cpu w/ rq lock held and preemption
797 * disabled, which in turn means that none else could be
798 * manipulating idle_list, so dereferencing idle_list without pool
801 if (atomic_dec_and_test(&pool
->nr_running
) &&
802 !list_empty(&pool
->worklist
))
803 to_wakeup
= first_worker(pool
);
804 return to_wakeup
? to_wakeup
->task
: NULL
;
808 * worker_set_flags - set worker flags and adjust nr_running accordingly
810 * @flags: flags to set
811 * @wakeup: wakeup an idle worker if necessary
813 * Set @flags in @worker->flags and adjust nr_running accordingly. If
814 * nr_running becomes zero and @wakeup is %true, an idle worker is
818 * spin_lock_irq(pool->lock)
820 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
823 struct worker_pool
*pool
= worker
->pool
;
825 WARN_ON_ONCE(worker
->task
!= current
);
828 * If transitioning into NOT_RUNNING, adjust nr_running and
829 * wake up an idle worker as necessary if requested by
832 if ((flags
& WORKER_NOT_RUNNING
) &&
833 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
835 if (atomic_dec_and_test(&pool
->nr_running
) &&
836 !list_empty(&pool
->worklist
))
837 wake_up_worker(pool
);
839 atomic_dec(&pool
->nr_running
);
842 worker
->flags
|= flags
;
846 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
848 * @flags: flags to clear
850 * Clear @flags in @worker->flags and adjust nr_running accordingly.
853 * spin_lock_irq(pool->lock)
855 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
857 struct worker_pool
*pool
= worker
->pool
;
858 unsigned int oflags
= worker
->flags
;
860 WARN_ON_ONCE(worker
->task
!= current
);
862 worker
->flags
&= ~flags
;
865 * If transitioning out of NOT_RUNNING, increment nr_running. Note
866 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
867 * of multiple flags, not a single flag.
869 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
870 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
871 atomic_inc(&pool
->nr_running
);
875 * find_worker_executing_work - find worker which is executing a work
876 * @pool: pool of interest
877 * @work: work to find worker for
879 * Find a worker which is executing @work on @pool by searching
880 * @pool->busy_hash which is keyed by the address of @work. For a worker
881 * to match, its current execution should match the address of @work and
882 * its work function. This is to avoid unwanted dependency between
883 * unrelated work executions through a work item being recycled while still
886 * This is a bit tricky. A work item may be freed once its execution
887 * starts and nothing prevents the freed area from being recycled for
888 * another work item. If the same work item address ends up being reused
889 * before the original execution finishes, workqueue will identify the
890 * recycled work item as currently executing and make it wait until the
891 * current execution finishes, introducing an unwanted dependency.
893 * This function checks the work item address, work function and workqueue
894 * to avoid false positives. Note that this isn't complete as one may
895 * construct a work function which can introduce dependency onto itself
896 * through a recycled work item. Well, if somebody wants to shoot oneself
897 * in the foot that badly, there's only so much we can do, and if such
898 * deadlock actually occurs, it should be easy to locate the culprit work
902 * spin_lock_irq(pool->lock).
905 * Pointer to worker which is executing @work if found, NULL
908 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
909 struct work_struct
*work
)
911 struct worker
*worker
;
913 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
915 if (worker
->current_work
== work
&&
916 worker
->current_func
== work
->func
)
923 * move_linked_works - move linked works to a list
924 * @work: start of series of works to be scheduled
925 * @head: target list to append @work to
926 * @nextp: out paramter for nested worklist walking
928 * Schedule linked works starting from @work to @head. Work series to
929 * be scheduled starts at @work and includes any consecutive work with
930 * WORK_STRUCT_LINKED set in its predecessor.
932 * If @nextp is not NULL, it's updated to point to the next work of
933 * the last scheduled work. This allows move_linked_works() to be
934 * nested inside outer list_for_each_entry_safe().
937 * spin_lock_irq(pool->lock).
939 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
940 struct work_struct
**nextp
)
942 struct work_struct
*n
;
945 * Linked worklist will always end before the end of the list,
946 * use NULL for list head.
948 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
949 list_move_tail(&work
->entry
, head
);
950 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
955 * If we're already inside safe list traversal and have moved
956 * multiple works to the scheduled queue, the next position
957 * needs to be updated.
964 * get_pwq - get an extra reference on the specified pool_workqueue
965 * @pwq: pool_workqueue to get
967 * Obtain an extra reference on @pwq. The caller should guarantee that
968 * @pwq has positive refcnt and be holding the matching pool->lock.
970 static void get_pwq(struct pool_workqueue
*pwq
)
972 lockdep_assert_held(&pwq
->pool
->lock
);
973 WARN_ON_ONCE(pwq
->refcnt
<= 0);
978 * put_pwq - put a pool_workqueue reference
979 * @pwq: pool_workqueue to put
981 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
982 * destruction. The caller should be holding the matching pool->lock.
984 static void put_pwq(struct pool_workqueue
*pwq
)
986 lockdep_assert_held(&pwq
->pool
->lock
);
987 if (likely(--pwq
->refcnt
))
989 if (WARN_ON_ONCE(!(pwq
->wq
->flags
& WQ_UNBOUND
)))
992 * @pwq can't be released under pool->lock, bounce to
993 * pwq_unbound_release_workfn(). This never recurses on the same
994 * pool->lock as this path is taken only for unbound workqueues and
995 * the release work item is scheduled on a per-cpu workqueue. To
996 * avoid lockdep warning, unbound pool->locks are given lockdep
997 * subclass of 1 in get_unbound_pool().
999 schedule_work(&pwq
->unbound_release_work
);
1002 static void pwq_activate_delayed_work(struct work_struct
*work
)
1004 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1006 trace_workqueue_activate_work(work
);
1007 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
1008 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1012 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
1014 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
1015 struct work_struct
, entry
);
1017 pwq_activate_delayed_work(work
);
1021 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1022 * @pwq: pwq of interest
1023 * @color: color of work which left the queue
1025 * A work either has completed or is removed from pending queue,
1026 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1029 * spin_lock_irq(pool->lock).
1031 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
1033 /* uncolored work items don't participate in flushing or nr_active */
1034 if (color
== WORK_NO_COLOR
)
1037 pwq
->nr_in_flight
[color
]--;
1040 if (!list_empty(&pwq
->delayed_works
)) {
1041 /* one down, submit a delayed one */
1042 if (pwq
->nr_active
< pwq
->max_active
)
1043 pwq_activate_first_delayed(pwq
);
1046 /* is flush in progress and are we at the flushing tip? */
1047 if (likely(pwq
->flush_color
!= color
))
1050 /* are there still in-flight works? */
1051 if (pwq
->nr_in_flight
[color
])
1054 /* this pwq is done, clear flush_color */
1055 pwq
->flush_color
= -1;
1058 * If this was the last pwq, wake up the first flusher. It
1059 * will handle the rest.
1061 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1062 complete(&pwq
->wq
->first_flusher
->done
);
1068 * try_to_grab_pending - steal work item from worklist and disable irq
1069 * @work: work item to steal
1070 * @is_dwork: @work is a delayed_work
1071 * @flags: place to store irq state
1073 * Try to grab PENDING bit of @work. This function can handle @work in any
1074 * stable state - idle, on timer or on worklist. Return values are
1076 * 1 if @work was pending and we successfully stole PENDING
1077 * 0 if @work was idle and we claimed PENDING
1078 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1079 * -ENOENT if someone else is canceling @work, this state may persist
1080 * for arbitrarily long
1082 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1083 * interrupted while holding PENDING and @work off queue, irq must be
1084 * disabled on entry. This, combined with delayed_work->timer being
1085 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1087 * On successful return, >= 0, irq is disabled and the caller is
1088 * responsible for releasing it using local_irq_restore(*@flags).
1090 * This function is safe to call from any context including IRQ handler.
1092 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1093 unsigned long *flags
)
1095 struct worker_pool
*pool
;
1096 struct pool_workqueue
*pwq
;
1098 local_irq_save(*flags
);
1100 /* try to steal the timer if it exists */
1102 struct delayed_work
*dwork
= to_delayed_work(work
);
1105 * dwork->timer is irqsafe. If del_timer() fails, it's
1106 * guaranteed that the timer is not queued anywhere and not
1107 * running on the local CPU.
1109 if (likely(del_timer(&dwork
->timer
)))
1113 /* try to claim PENDING the normal way */
1114 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1118 * The queueing is in progress, or it is already queued. Try to
1119 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1121 pool
= get_work_pool(work
);
1125 spin_lock(&pool
->lock
);
1127 * work->data is guaranteed to point to pwq only while the work
1128 * item is queued on pwq->wq, and both updating work->data to point
1129 * to pwq on queueing and to pool on dequeueing are done under
1130 * pwq->pool->lock. This in turn guarantees that, if work->data
1131 * points to pwq which is associated with a locked pool, the work
1132 * item is currently queued on that pool.
1134 pwq
= get_work_pwq(work
);
1135 if (pwq
&& pwq
->pool
== pool
) {
1136 debug_work_deactivate(work
);
1139 * A delayed work item cannot be grabbed directly because
1140 * it might have linked NO_COLOR work items which, if left
1141 * on the delayed_list, will confuse pwq->nr_active
1142 * management later on and cause stall. Make sure the work
1143 * item is activated before grabbing.
1145 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1146 pwq_activate_delayed_work(work
);
1148 list_del_init(&work
->entry
);
1149 pwq_dec_nr_in_flight(get_work_pwq(work
), get_work_color(work
));
1151 /* work->data points to pwq iff queued, point to pool */
1152 set_work_pool_and_keep_pending(work
, pool
->id
);
1154 spin_unlock(&pool
->lock
);
1157 spin_unlock(&pool
->lock
);
1159 local_irq_restore(*flags
);
1160 if (work_is_canceling(work
))
1167 * insert_work - insert a work into a pool
1168 * @pwq: pwq @work belongs to
1169 * @work: work to insert
1170 * @head: insertion point
1171 * @extra_flags: extra WORK_STRUCT_* flags to set
1173 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1174 * work_struct flags.
1177 * spin_lock_irq(pool->lock).
1179 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1180 struct list_head
*head
, unsigned int extra_flags
)
1182 struct worker_pool
*pool
= pwq
->pool
;
1184 /* we own @work, set data and link */
1185 set_work_pwq(work
, pwq
, extra_flags
);
1186 list_add_tail(&work
->entry
, head
);
1190 * Ensure either worker_sched_deactivated() sees the above
1191 * list_add_tail() or we see zero nr_running to avoid workers
1192 * lying around lazily while there are works to be processed.
1196 if (__need_more_worker(pool
))
1197 wake_up_worker(pool
);
1201 * Test whether @work is being queued from another work executing on the
1204 static bool is_chained_work(struct workqueue_struct
*wq
)
1206 struct worker
*worker
;
1208 worker
= current_wq_worker();
1210 * Return %true iff I'm a worker execuing a work item on @wq. If
1211 * I'm @worker, it's safe to dereference it without locking.
1213 return worker
&& worker
->current_pwq
->wq
== wq
;
1216 static void __queue_work(int cpu
, struct workqueue_struct
*wq
,
1217 struct work_struct
*work
)
1219 struct pool_workqueue
*pwq
;
1220 struct worker_pool
*last_pool
;
1221 struct list_head
*worklist
;
1222 unsigned int work_flags
;
1223 unsigned int req_cpu
= cpu
;
1226 * While a work item is PENDING && off queue, a task trying to
1227 * steal the PENDING will busy-loop waiting for it to either get
1228 * queued or lose PENDING. Grabbing PENDING and queueing should
1229 * happen with IRQ disabled.
1231 WARN_ON_ONCE(!irqs_disabled());
1233 debug_work_activate(work
);
1235 /* if dying, only works from the same workqueue are allowed */
1236 if (unlikely(wq
->flags
& __WQ_DRAINING
) &&
1237 WARN_ON_ONCE(!is_chained_work(wq
)))
1240 /* pwq which will be used unless @work is executing elsewhere */
1241 if (!(wq
->flags
& WQ_UNBOUND
)) {
1242 if (cpu
== WORK_CPU_UNBOUND
)
1243 cpu
= raw_smp_processor_id();
1244 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
1246 pwq
= first_pwq(wq
);
1250 * If @work was previously on a different pool, it might still be
1251 * running there, in which case the work needs to be queued on that
1252 * pool to guarantee non-reentrancy.
1254 last_pool
= get_work_pool(work
);
1255 if (last_pool
&& last_pool
!= pwq
->pool
) {
1256 struct worker
*worker
;
1258 spin_lock(&last_pool
->lock
);
1260 worker
= find_worker_executing_work(last_pool
, work
);
1262 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1263 pwq
= worker
->current_pwq
;
1265 /* meh... not running there, queue here */
1266 spin_unlock(&last_pool
->lock
);
1267 spin_lock(&pwq
->pool
->lock
);
1270 spin_lock(&pwq
->pool
->lock
);
1274 * pwq is determined and locked. For unbound pools, we could have
1275 * raced with pwq release and it could already be dead. If its
1276 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1277 * without another pwq replacing it as the first pwq or while a
1278 * work item is executing on it, so the retying is guaranteed to
1279 * make forward-progress.
1281 if (unlikely(!pwq
->refcnt
)) {
1282 if (wq
->flags
& WQ_UNBOUND
) {
1283 spin_unlock(&pwq
->pool
->lock
);
1288 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1292 /* pwq determined, queue */
1293 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1295 if (WARN_ON(!list_empty(&work
->entry
))) {
1296 spin_unlock(&pwq
->pool
->lock
);
1300 pwq
->nr_in_flight
[pwq
->work_color
]++;
1301 work_flags
= work_color_to_flags(pwq
->work_color
);
1303 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1304 trace_workqueue_activate_work(work
);
1306 worklist
= &pwq
->pool
->worklist
;
1308 work_flags
|= WORK_STRUCT_DELAYED
;
1309 worklist
= &pwq
->delayed_works
;
1312 insert_work(pwq
, work
, worklist
, work_flags
);
1314 spin_unlock(&pwq
->pool
->lock
);
1318 * queue_work_on - queue work on specific cpu
1319 * @cpu: CPU number to execute work on
1320 * @wq: workqueue to use
1321 * @work: work to queue
1323 * Returns %false if @work was already on a queue, %true otherwise.
1325 * We queue the work to a specific CPU, the caller must ensure it
1328 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1329 struct work_struct
*work
)
1332 unsigned long flags
;
1334 local_irq_save(flags
);
1336 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1337 __queue_work(cpu
, wq
, work
);
1341 local_irq_restore(flags
);
1344 EXPORT_SYMBOL_GPL(queue_work_on
);
1347 * queue_work - queue work on a workqueue
1348 * @wq: workqueue to use
1349 * @work: work to queue
1351 * Returns %false if @work was already on a queue, %true otherwise.
1353 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1354 * it can be processed by another CPU.
1356 bool queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1358 return queue_work_on(WORK_CPU_UNBOUND
, wq
, work
);
1360 EXPORT_SYMBOL_GPL(queue_work
);
1362 void delayed_work_timer_fn(unsigned long __data
)
1364 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1366 /* should have been called from irqsafe timer with irq already off */
1367 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1369 EXPORT_SYMBOL(delayed_work_timer_fn
);
1371 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1372 struct delayed_work
*dwork
, unsigned long delay
)
1374 struct timer_list
*timer
= &dwork
->timer
;
1375 struct work_struct
*work
= &dwork
->work
;
1377 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1378 timer
->data
!= (unsigned long)dwork
);
1379 WARN_ON_ONCE(timer_pending(timer
));
1380 WARN_ON_ONCE(!list_empty(&work
->entry
));
1383 * If @delay is 0, queue @dwork->work immediately. This is for
1384 * both optimization and correctness. The earliest @timer can
1385 * expire is on the closest next tick and delayed_work users depend
1386 * on that there's no such delay when @delay is 0.
1389 __queue_work(cpu
, wq
, &dwork
->work
);
1393 timer_stats_timer_set_start_info(&dwork
->timer
);
1397 timer
->expires
= jiffies
+ delay
;
1399 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1400 add_timer_on(timer
, cpu
);
1406 * queue_delayed_work_on - queue work on specific CPU after delay
1407 * @cpu: CPU number to execute work on
1408 * @wq: workqueue to use
1409 * @dwork: work to queue
1410 * @delay: number of jiffies to wait before queueing
1412 * Returns %false if @work was already on a queue, %true otherwise. If
1413 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1416 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1417 struct delayed_work
*dwork
, unsigned long delay
)
1419 struct work_struct
*work
= &dwork
->work
;
1421 unsigned long flags
;
1423 /* read the comment in __queue_work() */
1424 local_irq_save(flags
);
1426 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1427 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1431 local_irq_restore(flags
);
1434 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1437 * queue_delayed_work - queue work on a workqueue after delay
1438 * @wq: workqueue to use
1439 * @dwork: delayable work to queue
1440 * @delay: number of jiffies to wait before queueing
1442 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1444 bool queue_delayed_work(struct workqueue_struct
*wq
,
1445 struct delayed_work
*dwork
, unsigned long delay
)
1447 return queue_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1449 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1452 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1453 * @cpu: CPU number to execute work on
1454 * @wq: workqueue to use
1455 * @dwork: work to queue
1456 * @delay: number of jiffies to wait before queueing
1458 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1459 * modify @dwork's timer so that it expires after @delay. If @delay is
1460 * zero, @work is guaranteed to be scheduled immediately regardless of its
1463 * Returns %false if @dwork was idle and queued, %true if @dwork was
1464 * pending and its timer was modified.
1466 * This function is safe to call from any context including IRQ handler.
1467 * See try_to_grab_pending() for details.
1469 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1470 struct delayed_work
*dwork
, unsigned long delay
)
1472 unsigned long flags
;
1476 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1477 } while (unlikely(ret
== -EAGAIN
));
1479 if (likely(ret
>= 0)) {
1480 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1481 local_irq_restore(flags
);
1484 /* -ENOENT from try_to_grab_pending() becomes %true */
1487 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1490 * mod_delayed_work - modify delay of or queue a delayed work
1491 * @wq: workqueue to use
1492 * @dwork: work to queue
1493 * @delay: number of jiffies to wait before queueing
1495 * mod_delayed_work_on() on local CPU.
1497 bool mod_delayed_work(struct workqueue_struct
*wq
, struct delayed_work
*dwork
,
1498 unsigned long delay
)
1500 return mod_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1502 EXPORT_SYMBOL_GPL(mod_delayed_work
);
1505 * worker_enter_idle - enter idle state
1506 * @worker: worker which is entering idle state
1508 * @worker is entering idle state. Update stats and idle timer if
1512 * spin_lock_irq(pool->lock).
1514 static void worker_enter_idle(struct worker
*worker
)
1516 struct worker_pool
*pool
= worker
->pool
;
1518 if (WARN_ON_ONCE(worker
->flags
& WORKER_IDLE
) ||
1519 WARN_ON_ONCE(!list_empty(&worker
->entry
) &&
1520 (worker
->hentry
.next
|| worker
->hentry
.pprev
)))
1523 /* can't use worker_set_flags(), also called from start_worker() */
1524 worker
->flags
|= WORKER_IDLE
;
1526 worker
->last_active
= jiffies
;
1528 /* idle_list is LIFO */
1529 list_add(&worker
->entry
, &pool
->idle_list
);
1531 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1532 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1535 * Sanity check nr_running. Because wq_unbind_fn() releases
1536 * pool->lock between setting %WORKER_UNBOUND and zapping
1537 * nr_running, the warning may trigger spuriously. Check iff
1538 * unbind is not in progress.
1540 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1541 pool
->nr_workers
== pool
->nr_idle
&&
1542 atomic_read(&pool
->nr_running
));
1546 * worker_leave_idle - leave idle state
1547 * @worker: worker which is leaving idle state
1549 * @worker is leaving idle state. Update stats.
1552 * spin_lock_irq(pool->lock).
1554 static void worker_leave_idle(struct worker
*worker
)
1556 struct worker_pool
*pool
= worker
->pool
;
1558 if (WARN_ON_ONCE(!(worker
->flags
& WORKER_IDLE
)))
1560 worker_clr_flags(worker
, WORKER_IDLE
);
1562 list_del_init(&worker
->entry
);
1566 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1567 * @pool: target worker_pool
1569 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1571 * Works which are scheduled while the cpu is online must at least be
1572 * scheduled to a worker which is bound to the cpu so that if they are
1573 * flushed from cpu callbacks while cpu is going down, they are
1574 * guaranteed to execute on the cpu.
1576 * This function is to be used by unbound workers and rescuers to bind
1577 * themselves to the target cpu and may race with cpu going down or
1578 * coming online. kthread_bind() can't be used because it may put the
1579 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1580 * verbatim as it's best effort and blocking and pool may be
1581 * [dis]associated in the meantime.
1583 * This function tries set_cpus_allowed() and locks pool and verifies the
1584 * binding against %POOL_DISASSOCIATED which is set during
1585 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1586 * enters idle state or fetches works without dropping lock, it can
1587 * guarantee the scheduling requirement described in the first paragraph.
1590 * Might sleep. Called without any lock but returns with pool->lock
1594 * %true if the associated pool is online (@worker is successfully
1595 * bound), %false if offline.
1597 static bool worker_maybe_bind_and_lock(struct worker_pool
*pool
)
1598 __acquires(&pool
->lock
)
1602 * The following call may fail, succeed or succeed
1603 * without actually migrating the task to the cpu if
1604 * it races with cpu hotunplug operation. Verify
1605 * against POOL_DISASSOCIATED.
1607 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1608 set_cpus_allowed_ptr(current
, pool
->attrs
->cpumask
);
1610 spin_lock_irq(&pool
->lock
);
1611 if (pool
->flags
& POOL_DISASSOCIATED
)
1613 if (task_cpu(current
) == pool
->cpu
&&
1614 cpumask_equal(¤t
->cpus_allowed
, pool
->attrs
->cpumask
))
1616 spin_unlock_irq(&pool
->lock
);
1619 * We've raced with CPU hot[un]plug. Give it a breather
1620 * and retry migration. cond_resched() is required here;
1621 * otherwise, we might deadlock against cpu_stop trying to
1622 * bring down the CPU on non-preemptive kernel.
1630 * Rebind an idle @worker to its CPU. worker_thread() will test
1631 * list_empty(@worker->entry) before leaving idle and call this function.
1633 static void idle_worker_rebind(struct worker
*worker
)
1635 /* CPU may go down again inbetween, clear UNBOUND only on success */
1636 if (worker_maybe_bind_and_lock(worker
->pool
))
1637 worker_clr_flags(worker
, WORKER_UNBOUND
);
1639 /* rebind complete, become available again */
1640 list_add(&worker
->entry
, &worker
->pool
->idle_list
);
1641 spin_unlock_irq(&worker
->pool
->lock
);
1645 * Function for @worker->rebind.work used to rebind unbound busy workers to
1646 * the associated cpu which is coming back online. This is scheduled by
1647 * cpu up but can race with other cpu hotplug operations and may be
1648 * executed twice without intervening cpu down.
1650 static void busy_worker_rebind_fn(struct work_struct
*work
)
1652 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1654 if (worker_maybe_bind_and_lock(worker
->pool
))
1655 worker_clr_flags(worker
, WORKER_UNBOUND
);
1657 spin_unlock_irq(&worker
->pool
->lock
);
1661 * rebind_workers - rebind all workers of a pool to the associated CPU
1662 * @pool: pool of interest
1664 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1665 * is different for idle and busy ones.
1667 * Idle ones will be removed from the idle_list and woken up. They will
1668 * add themselves back after completing rebind. This ensures that the
1669 * idle_list doesn't contain any unbound workers when re-bound busy workers
1670 * try to perform local wake-ups for concurrency management.
1672 * Busy workers can rebind after they finish their current work items.
1673 * Queueing the rebind work item at the head of the scheduled list is
1674 * enough. Note that nr_running will be properly bumped as busy workers
1677 * On return, all non-manager workers are scheduled for rebind - see
1678 * manage_workers() for the manager special case. Any idle worker
1679 * including the manager will not appear on @idle_list until rebind is
1680 * complete, making local wake-ups safe.
1682 static void rebind_workers(struct worker_pool
*pool
)
1684 struct worker
*worker
, *n
;
1687 lockdep_assert_held(&pool
->assoc_mutex
);
1688 lockdep_assert_held(&pool
->lock
);
1690 /* dequeue and kick idle ones */
1691 list_for_each_entry_safe(worker
, n
, &pool
->idle_list
, entry
) {
1693 * idle workers should be off @pool->idle_list until rebind
1694 * is complete to avoid receiving premature local wake-ups.
1696 list_del_init(&worker
->entry
);
1699 * worker_thread() will see the above dequeuing and call
1700 * idle_worker_rebind().
1702 wake_up_process(worker
->task
);
1705 /* rebind busy workers */
1706 for_each_busy_worker(worker
, i
, pool
) {
1707 struct work_struct
*rebind_work
= &worker
->rebind_work
;
1708 struct workqueue_struct
*wq
;
1710 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
1711 work_data_bits(rebind_work
)))
1714 debug_work_activate(rebind_work
);
1717 * wq doesn't really matter but let's keep @worker->pool
1718 * and @pwq->pool consistent for sanity.
1720 if (worker
->pool
->attrs
->nice
< 0)
1721 wq
= system_highpri_wq
;
1725 insert_work(per_cpu_ptr(wq
->cpu_pwqs
, pool
->cpu
), rebind_work
,
1726 worker
->scheduled
.next
,
1727 work_color_to_flags(WORK_NO_COLOR
));
1731 static struct worker
*alloc_worker(void)
1733 struct worker
*worker
;
1735 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1737 INIT_LIST_HEAD(&worker
->entry
);
1738 INIT_LIST_HEAD(&worker
->scheduled
);
1739 INIT_WORK(&worker
->rebind_work
, busy_worker_rebind_fn
);
1740 /* on creation a worker is in !idle && prep state */
1741 worker
->flags
= WORKER_PREP
;
1747 * create_worker - create a new workqueue worker
1748 * @pool: pool the new worker will belong to
1750 * Create a new worker which is bound to @pool. The returned worker
1751 * can be started by calling start_worker() or destroyed using
1755 * Might sleep. Does GFP_KERNEL allocations.
1758 * Pointer to the newly created worker.
1760 static struct worker
*create_worker(struct worker_pool
*pool
)
1762 const char *pri
= pool
->attrs
->nice
< 0 ? "H" : "";
1763 struct worker
*worker
= NULL
;
1766 spin_lock_irq(&pool
->lock
);
1767 while (ida_get_new(&pool
->worker_ida
, &id
)) {
1768 spin_unlock_irq(&pool
->lock
);
1769 if (!ida_pre_get(&pool
->worker_ida
, GFP_KERNEL
))
1771 spin_lock_irq(&pool
->lock
);
1773 spin_unlock_irq(&pool
->lock
);
1775 worker
= alloc_worker();
1779 worker
->pool
= pool
;
1783 worker
->task
= kthread_create_on_node(worker_thread
,
1784 worker
, cpu_to_node(pool
->cpu
),
1785 "kworker/%d:%d%s", pool
->cpu
, id
, pri
);
1787 worker
->task
= kthread_create(worker_thread
, worker
,
1790 if (IS_ERR(worker
->task
))
1793 set_user_nice(worker
->task
, pool
->attrs
->nice
);
1794 set_cpus_allowed_ptr(worker
->task
, pool
->attrs
->cpumask
);
1797 * %PF_THREAD_BOUND is used to prevent userland from meddling with
1798 * cpumask of workqueue workers. This is an abuse. We need
1799 * %PF_NO_SETAFFINITY.
1801 worker
->task
->flags
|= PF_THREAD_BOUND
;
1804 * The caller is responsible for ensuring %POOL_DISASSOCIATED
1805 * remains stable across this function. See the comments above the
1806 * flag definition for details.
1808 if (pool
->flags
& POOL_DISASSOCIATED
)
1809 worker
->flags
|= WORKER_UNBOUND
;
1814 spin_lock_irq(&pool
->lock
);
1815 ida_remove(&pool
->worker_ida
, id
);
1816 spin_unlock_irq(&pool
->lock
);
1823 * start_worker - start a newly created worker
1824 * @worker: worker to start
1826 * Make the pool aware of @worker and start it.
1829 * spin_lock_irq(pool->lock).
1831 static void start_worker(struct worker
*worker
)
1833 worker
->flags
|= WORKER_STARTED
;
1834 worker
->pool
->nr_workers
++;
1835 worker_enter_idle(worker
);
1836 wake_up_process(worker
->task
);
1840 * destroy_worker - destroy a workqueue worker
1841 * @worker: worker to be destroyed
1843 * Destroy @worker and adjust @pool stats accordingly.
1846 * spin_lock_irq(pool->lock) which is released and regrabbed.
1848 static void destroy_worker(struct worker
*worker
)
1850 struct worker_pool
*pool
= worker
->pool
;
1851 int id
= worker
->id
;
1853 /* sanity check frenzy */
1854 if (WARN_ON(worker
->current_work
) ||
1855 WARN_ON(!list_empty(&worker
->scheduled
)))
1858 if (worker
->flags
& WORKER_STARTED
)
1860 if (worker
->flags
& WORKER_IDLE
)
1863 list_del_init(&worker
->entry
);
1864 worker
->flags
|= WORKER_DIE
;
1866 spin_unlock_irq(&pool
->lock
);
1868 kthread_stop(worker
->task
);
1871 spin_lock_irq(&pool
->lock
);
1872 ida_remove(&pool
->worker_ida
, id
);
1875 static void idle_worker_timeout(unsigned long __pool
)
1877 struct worker_pool
*pool
= (void *)__pool
;
1879 spin_lock_irq(&pool
->lock
);
1881 if (too_many_workers(pool
)) {
1882 struct worker
*worker
;
1883 unsigned long expires
;
1885 /* idle_list is kept in LIFO order, check the last one */
1886 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1887 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1889 if (time_before(jiffies
, expires
))
1890 mod_timer(&pool
->idle_timer
, expires
);
1892 /* it's been idle for too long, wake up manager */
1893 pool
->flags
|= POOL_MANAGE_WORKERS
;
1894 wake_up_worker(pool
);
1898 spin_unlock_irq(&pool
->lock
);
1901 static void send_mayday(struct work_struct
*work
)
1903 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1904 struct workqueue_struct
*wq
= pwq
->wq
;
1906 lockdep_assert_held(&workqueue_lock
);
1911 /* mayday mayday mayday */
1912 if (list_empty(&pwq
->mayday_node
)) {
1913 list_add_tail(&pwq
->mayday_node
, &wq
->maydays
);
1914 wake_up_process(wq
->rescuer
->task
);
1918 static void pool_mayday_timeout(unsigned long __pool
)
1920 struct worker_pool
*pool
= (void *)__pool
;
1921 struct work_struct
*work
;
1923 spin_lock_irq(&workqueue_lock
); /* for wq->maydays */
1924 spin_lock(&pool
->lock
);
1926 if (need_to_create_worker(pool
)) {
1928 * We've been trying to create a new worker but
1929 * haven't been successful. We might be hitting an
1930 * allocation deadlock. Send distress signals to
1933 list_for_each_entry(work
, &pool
->worklist
, entry
)
1937 spin_unlock(&pool
->lock
);
1938 spin_unlock_irq(&workqueue_lock
);
1940 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1944 * maybe_create_worker - create a new worker if necessary
1945 * @pool: pool to create a new worker for
1947 * Create a new worker for @pool if necessary. @pool is guaranteed to
1948 * have at least one idle worker on return from this function. If
1949 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1950 * sent to all rescuers with works scheduled on @pool to resolve
1951 * possible allocation deadlock.
1953 * On return, need_to_create_worker() is guaranteed to be false and
1954 * may_start_working() true.
1957 * spin_lock_irq(pool->lock) which may be released and regrabbed
1958 * multiple times. Does GFP_KERNEL allocations. Called only from
1962 * false if no action was taken and pool->lock stayed locked, true
1965 static bool maybe_create_worker(struct worker_pool
*pool
)
1966 __releases(&pool
->lock
)
1967 __acquires(&pool
->lock
)
1969 if (!need_to_create_worker(pool
))
1972 spin_unlock_irq(&pool
->lock
);
1974 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1975 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1978 struct worker
*worker
;
1980 worker
= create_worker(pool
);
1982 del_timer_sync(&pool
->mayday_timer
);
1983 spin_lock_irq(&pool
->lock
);
1984 start_worker(worker
);
1985 if (WARN_ON_ONCE(need_to_create_worker(pool
)))
1990 if (!need_to_create_worker(pool
))
1993 __set_current_state(TASK_INTERRUPTIBLE
);
1994 schedule_timeout(CREATE_COOLDOWN
);
1996 if (!need_to_create_worker(pool
))
2000 del_timer_sync(&pool
->mayday_timer
);
2001 spin_lock_irq(&pool
->lock
);
2002 if (need_to_create_worker(pool
))
2008 * maybe_destroy_worker - destroy workers which have been idle for a while
2009 * @pool: pool to destroy workers for
2011 * Destroy @pool workers which have been idle for longer than
2012 * IDLE_WORKER_TIMEOUT.
2015 * spin_lock_irq(pool->lock) which may be released and regrabbed
2016 * multiple times. Called only from manager.
2019 * false if no action was taken and pool->lock stayed locked, true
2022 static bool maybe_destroy_workers(struct worker_pool
*pool
)
2026 while (too_many_workers(pool
)) {
2027 struct worker
*worker
;
2028 unsigned long expires
;
2030 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2031 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2033 if (time_before(jiffies
, expires
)) {
2034 mod_timer(&pool
->idle_timer
, expires
);
2038 destroy_worker(worker
);
2046 * manage_workers - manage worker pool
2049 * Assume the manager role and manage the worker pool @worker belongs
2050 * to. At any given time, there can be only zero or one manager per
2051 * pool. The exclusion is handled automatically by this function.
2053 * The caller can safely start processing works on false return. On
2054 * true return, it's guaranteed that need_to_create_worker() is false
2055 * and may_start_working() is true.
2058 * spin_lock_irq(pool->lock) which may be released and regrabbed
2059 * multiple times. Does GFP_KERNEL allocations.
2062 * spin_lock_irq(pool->lock) which may be released and regrabbed
2063 * multiple times. Does GFP_KERNEL allocations.
2065 static bool manage_workers(struct worker
*worker
)
2067 struct worker_pool
*pool
= worker
->pool
;
2070 if (!mutex_trylock(&pool
->manager_arb
))
2074 * To simplify both worker management and CPU hotplug, hold off
2075 * management while hotplug is in progress. CPU hotplug path can't
2076 * grab @pool->manager_arb to achieve this because that can lead to
2077 * idle worker depletion (all become busy thinking someone else is
2078 * managing) which in turn can result in deadlock under extreme
2079 * circumstances. Use @pool->assoc_mutex to synchronize manager
2080 * against CPU hotplug.
2082 * assoc_mutex would always be free unless CPU hotplug is in
2083 * progress. trylock first without dropping @pool->lock.
2085 if (unlikely(!mutex_trylock(&pool
->assoc_mutex
))) {
2086 spin_unlock_irq(&pool
->lock
);
2087 mutex_lock(&pool
->assoc_mutex
);
2089 * CPU hotplug could have happened while we were waiting
2090 * for assoc_mutex. Hotplug itself can't handle us
2091 * because manager isn't either on idle or busy list, and
2092 * @pool's state and ours could have deviated.
2094 * As hotplug is now excluded via assoc_mutex, we can
2095 * simply try to bind. It will succeed or fail depending
2096 * on @pool's current state. Try it and adjust
2097 * %WORKER_UNBOUND accordingly.
2099 if (worker_maybe_bind_and_lock(pool
))
2100 worker
->flags
&= ~WORKER_UNBOUND
;
2102 worker
->flags
|= WORKER_UNBOUND
;
2107 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2110 * Destroy and then create so that may_start_working() is true
2113 ret
|= maybe_destroy_workers(pool
);
2114 ret
|= maybe_create_worker(pool
);
2116 mutex_unlock(&pool
->assoc_mutex
);
2117 mutex_unlock(&pool
->manager_arb
);
2122 * process_one_work - process single work
2124 * @work: work to process
2126 * Process @work. This function contains all the logics necessary to
2127 * process a single work including synchronization against and
2128 * interaction with other workers on the same cpu, queueing and
2129 * flushing. As long as context requirement is met, any worker can
2130 * call this function to process a work.
2133 * spin_lock_irq(pool->lock) which is released and regrabbed.
2135 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2136 __releases(&pool
->lock
)
2137 __acquires(&pool
->lock
)
2139 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2140 struct worker_pool
*pool
= worker
->pool
;
2141 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2143 struct worker
*collision
;
2144 #ifdef CONFIG_LOCKDEP
2146 * It is permissible to free the struct work_struct from
2147 * inside the function that is called from it, this we need to
2148 * take into account for lockdep too. To avoid bogus "held
2149 * lock freed" warnings as well as problems when looking into
2150 * work->lockdep_map, make a copy and use that here.
2152 struct lockdep_map lockdep_map
;
2154 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2157 * Ensure we're on the correct CPU. DISASSOCIATED test is
2158 * necessary to avoid spurious warnings from rescuers servicing the
2159 * unbound or a disassociated pool.
2161 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2162 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2163 raw_smp_processor_id() != pool
->cpu
);
2166 * A single work shouldn't be executed concurrently by
2167 * multiple workers on a single cpu. Check whether anyone is
2168 * already processing the work. If so, defer the work to the
2169 * currently executing one.
2171 collision
= find_worker_executing_work(pool
, work
);
2172 if (unlikely(collision
)) {
2173 move_linked_works(work
, &collision
->scheduled
, NULL
);
2177 /* claim and dequeue */
2178 debug_work_deactivate(work
);
2179 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2180 worker
->current_work
= work
;
2181 worker
->current_func
= work
->func
;
2182 worker
->current_pwq
= pwq
;
2183 work_color
= get_work_color(work
);
2185 list_del_init(&work
->entry
);
2188 * CPU intensive works don't participate in concurrency
2189 * management. They're the scheduler's responsibility.
2191 if (unlikely(cpu_intensive
))
2192 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2195 * Unbound pool isn't concurrency managed and work items should be
2196 * executed ASAP. Wake up another worker if necessary.
2198 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2199 wake_up_worker(pool
);
2202 * Record the last pool and clear PENDING which should be the last
2203 * update to @work. Also, do this inside @pool->lock so that
2204 * PENDING and queued state changes happen together while IRQ is
2207 set_work_pool_and_clear_pending(work
, pool
->id
);
2209 spin_unlock_irq(&pool
->lock
);
2211 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2212 lock_map_acquire(&lockdep_map
);
2213 trace_workqueue_execute_start(work
);
2214 worker
->current_func(work
);
2216 * While we must be careful to not use "work" after this, the trace
2217 * point will only record its address.
2219 trace_workqueue_execute_end(work
);
2220 lock_map_release(&lockdep_map
);
2221 lock_map_release(&pwq
->wq
->lockdep_map
);
2223 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2224 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2225 " last function: %pf\n",
2226 current
->comm
, preempt_count(), task_pid_nr(current
),
2227 worker
->current_func
);
2228 debug_show_held_locks(current
);
2232 spin_lock_irq(&pool
->lock
);
2234 /* clear cpu intensive status */
2235 if (unlikely(cpu_intensive
))
2236 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2238 /* we're done with it, release */
2239 hash_del(&worker
->hentry
);
2240 worker
->current_work
= NULL
;
2241 worker
->current_func
= NULL
;
2242 worker
->current_pwq
= NULL
;
2243 pwq_dec_nr_in_flight(pwq
, work_color
);
2247 * process_scheduled_works - process scheduled works
2250 * Process all scheduled works. Please note that the scheduled list
2251 * may change while processing a work, so this function repeatedly
2252 * fetches a work from the top and executes it.
2255 * spin_lock_irq(pool->lock) which may be released and regrabbed
2258 static void process_scheduled_works(struct worker
*worker
)
2260 while (!list_empty(&worker
->scheduled
)) {
2261 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2262 struct work_struct
, entry
);
2263 process_one_work(worker
, work
);
2268 * worker_thread - the worker thread function
2271 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2272 * of these per each cpu. These workers process all works regardless of
2273 * their specific target workqueue. The only exception is works which
2274 * belong to workqueues with a rescuer which will be explained in
2277 static int worker_thread(void *__worker
)
2279 struct worker
*worker
= __worker
;
2280 struct worker_pool
*pool
= worker
->pool
;
2282 /* tell the scheduler that this is a workqueue worker */
2283 worker
->task
->flags
|= PF_WQ_WORKER
;
2285 spin_lock_irq(&pool
->lock
);
2287 /* we are off idle list if destruction or rebind is requested */
2288 if (unlikely(list_empty(&worker
->entry
))) {
2289 spin_unlock_irq(&pool
->lock
);
2291 /* if DIE is set, destruction is requested */
2292 if (worker
->flags
& WORKER_DIE
) {
2293 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2297 /* otherwise, rebind */
2298 idle_worker_rebind(worker
);
2302 worker_leave_idle(worker
);
2304 /* no more worker necessary? */
2305 if (!need_more_worker(pool
))
2308 /* do we need to manage? */
2309 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2313 * ->scheduled list can only be filled while a worker is
2314 * preparing to process a work or actually processing it.
2315 * Make sure nobody diddled with it while I was sleeping.
2317 WARN_ON_ONCE(!list_empty(&worker
->scheduled
));
2320 * When control reaches this point, we're guaranteed to have
2321 * at least one idle worker or that someone else has already
2322 * assumed the manager role.
2324 worker_clr_flags(worker
, WORKER_PREP
);
2327 struct work_struct
*work
=
2328 list_first_entry(&pool
->worklist
,
2329 struct work_struct
, entry
);
2331 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2332 /* optimization path, not strictly necessary */
2333 process_one_work(worker
, work
);
2334 if (unlikely(!list_empty(&worker
->scheduled
)))
2335 process_scheduled_works(worker
);
2337 move_linked_works(work
, &worker
->scheduled
, NULL
);
2338 process_scheduled_works(worker
);
2340 } while (keep_working(pool
));
2342 worker_set_flags(worker
, WORKER_PREP
, false);
2344 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2348 * pool->lock is held and there's no work to process and no need to
2349 * manage, sleep. Workers are woken up only while holding
2350 * pool->lock or from local cpu, so setting the current state
2351 * before releasing pool->lock is enough to prevent losing any
2354 worker_enter_idle(worker
);
2355 __set_current_state(TASK_INTERRUPTIBLE
);
2356 spin_unlock_irq(&pool
->lock
);
2362 * rescuer_thread - the rescuer thread function
2365 * Workqueue rescuer thread function. There's one rescuer for each
2366 * workqueue which has WQ_MEM_RECLAIM set.
2368 * Regular work processing on a pool may block trying to create a new
2369 * worker which uses GFP_KERNEL allocation which has slight chance of
2370 * developing into deadlock if some works currently on the same queue
2371 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2372 * the problem rescuer solves.
2374 * When such condition is possible, the pool summons rescuers of all
2375 * workqueues which have works queued on the pool and let them process
2376 * those works so that forward progress can be guaranteed.
2378 * This should happen rarely.
2380 static int rescuer_thread(void *__rescuer
)
2382 struct worker
*rescuer
= __rescuer
;
2383 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2384 struct list_head
*scheduled
= &rescuer
->scheduled
;
2386 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2389 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2390 * doesn't participate in concurrency management.
2392 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2394 set_current_state(TASK_INTERRUPTIBLE
);
2396 if (kthread_should_stop()) {
2397 __set_current_state(TASK_RUNNING
);
2398 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2402 /* see whether any pwq is asking for help */
2403 spin_lock_irq(&workqueue_lock
);
2405 while (!list_empty(&wq
->maydays
)) {
2406 struct pool_workqueue
*pwq
= list_first_entry(&wq
->maydays
,
2407 struct pool_workqueue
, mayday_node
);
2408 struct worker_pool
*pool
= pwq
->pool
;
2409 struct work_struct
*work
, *n
;
2411 __set_current_state(TASK_RUNNING
);
2412 list_del_init(&pwq
->mayday_node
);
2414 spin_unlock_irq(&workqueue_lock
);
2416 /* migrate to the target cpu if possible */
2417 worker_maybe_bind_and_lock(pool
);
2418 rescuer
->pool
= pool
;
2421 * Slurp in all works issued via this workqueue and
2424 WARN_ON_ONCE(!list_empty(&rescuer
->scheduled
));
2425 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2426 if (get_work_pwq(work
) == pwq
)
2427 move_linked_works(work
, scheduled
, &n
);
2429 process_scheduled_works(rescuer
);
2432 * Leave this pool. If keep_working() is %true, notify a
2433 * regular worker; otherwise, we end up with 0 concurrency
2434 * and stalling the execution.
2436 if (keep_working(pool
))
2437 wake_up_worker(pool
);
2439 rescuer
->pool
= NULL
;
2440 spin_unlock(&pool
->lock
);
2441 spin_lock(&workqueue_lock
);
2444 spin_unlock_irq(&workqueue_lock
);
2446 /* rescuers should never participate in concurrency management */
2447 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2453 struct work_struct work
;
2454 struct completion done
;
2457 static void wq_barrier_func(struct work_struct
*work
)
2459 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2460 complete(&barr
->done
);
2464 * insert_wq_barrier - insert a barrier work
2465 * @pwq: pwq to insert barrier into
2466 * @barr: wq_barrier to insert
2467 * @target: target work to attach @barr to
2468 * @worker: worker currently executing @target, NULL if @target is not executing
2470 * @barr is linked to @target such that @barr is completed only after
2471 * @target finishes execution. Please note that the ordering
2472 * guarantee is observed only with respect to @target and on the local
2475 * Currently, a queued barrier can't be canceled. This is because
2476 * try_to_grab_pending() can't determine whether the work to be
2477 * grabbed is at the head of the queue and thus can't clear LINKED
2478 * flag of the previous work while there must be a valid next work
2479 * after a work with LINKED flag set.
2481 * Note that when @worker is non-NULL, @target may be modified
2482 * underneath us, so we can't reliably determine pwq from @target.
2485 * spin_lock_irq(pool->lock).
2487 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2488 struct wq_barrier
*barr
,
2489 struct work_struct
*target
, struct worker
*worker
)
2491 struct list_head
*head
;
2492 unsigned int linked
= 0;
2495 * debugobject calls are safe here even with pool->lock locked
2496 * as we know for sure that this will not trigger any of the
2497 * checks and call back into the fixup functions where we
2500 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2501 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2502 init_completion(&barr
->done
);
2505 * If @target is currently being executed, schedule the
2506 * barrier to the worker; otherwise, put it after @target.
2509 head
= worker
->scheduled
.next
;
2511 unsigned long *bits
= work_data_bits(target
);
2513 head
= target
->entry
.next
;
2514 /* there can already be other linked works, inherit and set */
2515 linked
= *bits
& WORK_STRUCT_LINKED
;
2516 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2519 debug_work_activate(&barr
->work
);
2520 insert_work(pwq
, &barr
->work
, head
,
2521 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2525 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2526 * @wq: workqueue being flushed
2527 * @flush_color: new flush color, < 0 for no-op
2528 * @work_color: new work color, < 0 for no-op
2530 * Prepare pwqs for workqueue flushing.
2532 * If @flush_color is non-negative, flush_color on all pwqs should be
2533 * -1. If no pwq has in-flight commands at the specified color, all
2534 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2535 * has in flight commands, its pwq->flush_color is set to
2536 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2537 * wakeup logic is armed and %true is returned.
2539 * The caller should have initialized @wq->first_flusher prior to
2540 * calling this function with non-negative @flush_color. If
2541 * @flush_color is negative, no flush color update is done and %false
2544 * If @work_color is non-negative, all pwqs should have the same
2545 * work_color which is previous to @work_color and all will be
2546 * advanced to @work_color.
2549 * mutex_lock(wq->flush_mutex).
2552 * %true if @flush_color >= 0 and there's something to flush. %false
2555 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2556 int flush_color
, int work_color
)
2559 struct pool_workqueue
*pwq
;
2561 if (flush_color
>= 0) {
2562 WARN_ON_ONCE(atomic_read(&wq
->nr_pwqs_to_flush
));
2563 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2566 local_irq_disable();
2568 for_each_pwq(pwq
, wq
) {
2569 struct worker_pool
*pool
= pwq
->pool
;
2571 spin_lock(&pool
->lock
);
2573 if (flush_color
>= 0) {
2574 WARN_ON_ONCE(pwq
->flush_color
!= -1);
2576 if (pwq
->nr_in_flight
[flush_color
]) {
2577 pwq
->flush_color
= flush_color
;
2578 atomic_inc(&wq
->nr_pwqs_to_flush
);
2583 if (work_color
>= 0) {
2584 WARN_ON_ONCE(work_color
!= work_next_color(pwq
->work_color
));
2585 pwq
->work_color
= work_color
;
2588 spin_unlock(&pool
->lock
);
2593 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2594 complete(&wq
->first_flusher
->done
);
2600 * flush_workqueue - ensure that any scheduled work has run to completion.
2601 * @wq: workqueue to flush
2603 * Forces execution of the workqueue and blocks until its completion.
2604 * This is typically used in driver shutdown handlers.
2606 * We sleep until all works which were queued on entry have been handled,
2607 * but we are not livelocked by new incoming ones.
2609 void flush_workqueue(struct workqueue_struct
*wq
)
2611 struct wq_flusher this_flusher
= {
2612 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2614 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2618 lock_map_acquire(&wq
->lockdep_map
);
2619 lock_map_release(&wq
->lockdep_map
);
2621 mutex_lock(&wq
->flush_mutex
);
2624 * Start-to-wait phase
2626 next_color
= work_next_color(wq
->work_color
);
2628 if (next_color
!= wq
->flush_color
) {
2630 * Color space is not full. The current work_color
2631 * becomes our flush_color and work_color is advanced
2634 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
));
2635 this_flusher
.flush_color
= wq
->work_color
;
2636 wq
->work_color
= next_color
;
2638 if (!wq
->first_flusher
) {
2639 /* no flush in progress, become the first flusher */
2640 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2642 wq
->first_flusher
= &this_flusher
;
2644 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2646 /* nothing to flush, done */
2647 wq
->flush_color
= next_color
;
2648 wq
->first_flusher
= NULL
;
2653 WARN_ON_ONCE(wq
->flush_color
== this_flusher
.flush_color
);
2654 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2655 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2659 * Oops, color space is full, wait on overflow queue.
2660 * The next flush completion will assign us
2661 * flush_color and transfer to flusher_queue.
2663 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2666 mutex_unlock(&wq
->flush_mutex
);
2668 wait_for_completion(&this_flusher
.done
);
2671 * Wake-up-and-cascade phase
2673 * First flushers are responsible for cascading flushes and
2674 * handling overflow. Non-first flushers can simply return.
2676 if (wq
->first_flusher
!= &this_flusher
)
2679 mutex_lock(&wq
->flush_mutex
);
2681 /* we might have raced, check again with mutex held */
2682 if (wq
->first_flusher
!= &this_flusher
)
2685 wq
->first_flusher
= NULL
;
2687 WARN_ON_ONCE(!list_empty(&this_flusher
.list
));
2688 WARN_ON_ONCE(wq
->flush_color
!= this_flusher
.flush_color
);
2691 struct wq_flusher
*next
, *tmp
;
2693 /* complete all the flushers sharing the current flush color */
2694 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2695 if (next
->flush_color
!= wq
->flush_color
)
2697 list_del_init(&next
->list
);
2698 complete(&next
->done
);
2701 WARN_ON_ONCE(!list_empty(&wq
->flusher_overflow
) &&
2702 wq
->flush_color
!= work_next_color(wq
->work_color
));
2704 /* this flush_color is finished, advance by one */
2705 wq
->flush_color
= work_next_color(wq
->flush_color
);
2707 /* one color has been freed, handle overflow queue */
2708 if (!list_empty(&wq
->flusher_overflow
)) {
2710 * Assign the same color to all overflowed
2711 * flushers, advance work_color and append to
2712 * flusher_queue. This is the start-to-wait
2713 * phase for these overflowed flushers.
2715 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2716 tmp
->flush_color
= wq
->work_color
;
2718 wq
->work_color
= work_next_color(wq
->work_color
);
2720 list_splice_tail_init(&wq
->flusher_overflow
,
2721 &wq
->flusher_queue
);
2722 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2725 if (list_empty(&wq
->flusher_queue
)) {
2726 WARN_ON_ONCE(wq
->flush_color
!= wq
->work_color
);
2731 * Need to flush more colors. Make the next flusher
2732 * the new first flusher and arm pwqs.
2734 WARN_ON_ONCE(wq
->flush_color
== wq
->work_color
);
2735 WARN_ON_ONCE(wq
->flush_color
!= next
->flush_color
);
2737 list_del_init(&next
->list
);
2738 wq
->first_flusher
= next
;
2740 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2744 * Meh... this color is already done, clear first
2745 * flusher and repeat cascading.
2747 wq
->first_flusher
= NULL
;
2751 mutex_unlock(&wq
->flush_mutex
);
2753 EXPORT_SYMBOL_GPL(flush_workqueue
);
2756 * drain_workqueue - drain a workqueue
2757 * @wq: workqueue to drain
2759 * Wait until the workqueue becomes empty. While draining is in progress,
2760 * only chain queueing is allowed. IOW, only currently pending or running
2761 * work items on @wq can queue further work items on it. @wq is flushed
2762 * repeatedly until it becomes empty. The number of flushing is detemined
2763 * by the depth of chaining and should be relatively short. Whine if it
2766 void drain_workqueue(struct workqueue_struct
*wq
)
2768 unsigned int flush_cnt
= 0;
2769 struct pool_workqueue
*pwq
;
2772 * __queue_work() needs to test whether there are drainers, is much
2773 * hotter than drain_workqueue() and already looks at @wq->flags.
2774 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2776 spin_lock_irq(&workqueue_lock
);
2777 if (!wq
->nr_drainers
++)
2778 wq
->flags
|= __WQ_DRAINING
;
2779 spin_unlock_irq(&workqueue_lock
);
2781 flush_workqueue(wq
);
2783 local_irq_disable();
2785 for_each_pwq(pwq
, wq
) {
2788 spin_lock(&pwq
->pool
->lock
);
2789 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2790 spin_unlock(&pwq
->pool
->lock
);
2795 if (++flush_cnt
== 10 ||
2796 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2797 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2798 wq
->name
, flush_cnt
);
2804 spin_lock(&workqueue_lock
);
2805 if (!--wq
->nr_drainers
)
2806 wq
->flags
&= ~__WQ_DRAINING
;
2807 spin_unlock(&workqueue_lock
);
2811 EXPORT_SYMBOL_GPL(drain_workqueue
);
2813 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2815 struct worker
*worker
= NULL
;
2816 struct worker_pool
*pool
;
2817 struct pool_workqueue
*pwq
;
2821 local_irq_disable();
2822 pool
= get_work_pool(work
);
2828 spin_lock(&pool
->lock
);
2829 /* see the comment in try_to_grab_pending() with the same code */
2830 pwq
= get_work_pwq(work
);
2832 if (unlikely(pwq
->pool
!= pool
))
2835 worker
= find_worker_executing_work(pool
, work
);
2838 pwq
= worker
->current_pwq
;
2841 insert_wq_barrier(pwq
, barr
, work
, worker
);
2842 spin_unlock_irq(&pool
->lock
);
2845 * If @max_active is 1 or rescuer is in use, flushing another work
2846 * item on the same workqueue may lead to deadlock. Make sure the
2847 * flusher is not running on the same workqueue by verifying write
2850 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->rescuer
)
2851 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2853 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2854 lock_map_release(&pwq
->wq
->lockdep_map
);
2858 spin_unlock_irq(&pool
->lock
);
2863 * flush_work - wait for a work to finish executing the last queueing instance
2864 * @work: the work to flush
2866 * Wait until @work has finished execution. @work is guaranteed to be idle
2867 * on return if it hasn't been requeued since flush started.
2870 * %true if flush_work() waited for the work to finish execution,
2871 * %false if it was already idle.
2873 bool flush_work(struct work_struct
*work
)
2875 struct wq_barrier barr
;
2877 lock_map_acquire(&work
->lockdep_map
);
2878 lock_map_release(&work
->lockdep_map
);
2880 if (start_flush_work(work
, &barr
)) {
2881 wait_for_completion(&barr
.done
);
2882 destroy_work_on_stack(&barr
.work
);
2888 EXPORT_SYMBOL_GPL(flush_work
);
2890 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2892 unsigned long flags
;
2896 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2898 * If someone else is canceling, wait for the same event it
2899 * would be waiting for before retrying.
2901 if (unlikely(ret
== -ENOENT
))
2903 } while (unlikely(ret
< 0));
2905 /* tell other tasks trying to grab @work to back off */
2906 mark_work_canceling(work
);
2907 local_irq_restore(flags
);
2910 clear_work_data(work
);
2915 * cancel_work_sync - cancel a work and wait for it to finish
2916 * @work: the work to cancel
2918 * Cancel @work and wait for its execution to finish. This function
2919 * can be used even if the work re-queues itself or migrates to
2920 * another workqueue. On return from this function, @work is
2921 * guaranteed to be not pending or executing on any CPU.
2923 * cancel_work_sync(&delayed_work->work) must not be used for
2924 * delayed_work's. Use cancel_delayed_work_sync() instead.
2926 * The caller must ensure that the workqueue on which @work was last
2927 * queued can't be destroyed before this function returns.
2930 * %true if @work was pending, %false otherwise.
2932 bool cancel_work_sync(struct work_struct
*work
)
2934 return __cancel_work_timer(work
, false);
2936 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2939 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2940 * @dwork: the delayed work to flush
2942 * Delayed timer is cancelled and the pending work is queued for
2943 * immediate execution. Like flush_work(), this function only
2944 * considers the last queueing instance of @dwork.
2947 * %true if flush_work() waited for the work to finish execution,
2948 * %false if it was already idle.
2950 bool flush_delayed_work(struct delayed_work
*dwork
)
2952 local_irq_disable();
2953 if (del_timer_sync(&dwork
->timer
))
2954 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2956 return flush_work(&dwork
->work
);
2958 EXPORT_SYMBOL(flush_delayed_work
);
2961 * cancel_delayed_work - cancel a delayed work
2962 * @dwork: delayed_work to cancel
2964 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2965 * and canceled; %false if wasn't pending. Note that the work callback
2966 * function may still be running on return, unless it returns %true and the
2967 * work doesn't re-arm itself. Explicitly flush or use
2968 * cancel_delayed_work_sync() to wait on it.
2970 * This function is safe to call from any context including IRQ handler.
2972 bool cancel_delayed_work(struct delayed_work
*dwork
)
2974 unsigned long flags
;
2978 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2979 } while (unlikely(ret
== -EAGAIN
));
2981 if (unlikely(ret
< 0))
2984 set_work_pool_and_clear_pending(&dwork
->work
,
2985 get_work_pool_id(&dwork
->work
));
2986 local_irq_restore(flags
);
2989 EXPORT_SYMBOL(cancel_delayed_work
);
2992 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2993 * @dwork: the delayed work cancel
2995 * This is cancel_work_sync() for delayed works.
2998 * %true if @dwork was pending, %false otherwise.
3000 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3002 return __cancel_work_timer(&dwork
->work
, true);
3004 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3007 * schedule_work_on - put work task on a specific cpu
3008 * @cpu: cpu to put the work task on
3009 * @work: job to be done
3011 * This puts a job on a specific cpu
3013 bool schedule_work_on(int cpu
, struct work_struct
*work
)
3015 return queue_work_on(cpu
, system_wq
, work
);
3017 EXPORT_SYMBOL(schedule_work_on
);
3020 * schedule_work - put work task in global workqueue
3021 * @work: job to be done
3023 * Returns %false if @work was already on the kernel-global workqueue and
3026 * This puts a job in the kernel-global workqueue if it was not already
3027 * queued and leaves it in the same position on the kernel-global
3028 * workqueue otherwise.
3030 bool schedule_work(struct work_struct
*work
)
3032 return queue_work(system_wq
, work
);
3034 EXPORT_SYMBOL(schedule_work
);
3037 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3039 * @dwork: job to be done
3040 * @delay: number of jiffies to wait
3042 * After waiting for a given time this puts a job in the kernel-global
3043 * workqueue on the specified CPU.
3045 bool schedule_delayed_work_on(int cpu
, struct delayed_work
*dwork
,
3046 unsigned long delay
)
3048 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
3050 EXPORT_SYMBOL(schedule_delayed_work_on
);
3053 * schedule_delayed_work - put work task in global workqueue after delay
3054 * @dwork: job to be done
3055 * @delay: number of jiffies to wait or 0 for immediate execution
3057 * After waiting for a given time this puts a job in the kernel-global
3060 bool schedule_delayed_work(struct delayed_work
*dwork
, unsigned long delay
)
3062 return queue_delayed_work(system_wq
, dwork
, delay
);
3064 EXPORT_SYMBOL(schedule_delayed_work
);
3067 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3068 * @func: the function to call
3070 * schedule_on_each_cpu() executes @func on each online CPU using the
3071 * system workqueue and blocks until all CPUs have completed.
3072 * schedule_on_each_cpu() is very slow.
3075 * 0 on success, -errno on failure.
3077 int schedule_on_each_cpu(work_func_t func
)
3080 struct work_struct __percpu
*works
;
3082 works
= alloc_percpu(struct work_struct
);
3088 for_each_online_cpu(cpu
) {
3089 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3091 INIT_WORK(work
, func
);
3092 schedule_work_on(cpu
, work
);
3095 for_each_online_cpu(cpu
)
3096 flush_work(per_cpu_ptr(works
, cpu
));
3104 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3106 * Forces execution of the kernel-global workqueue and blocks until its
3109 * Think twice before calling this function! It's very easy to get into
3110 * trouble if you don't take great care. Either of the following situations
3111 * will lead to deadlock:
3113 * One of the work items currently on the workqueue needs to acquire
3114 * a lock held by your code or its caller.
3116 * Your code is running in the context of a work routine.
3118 * They will be detected by lockdep when they occur, but the first might not
3119 * occur very often. It depends on what work items are on the workqueue and
3120 * what locks they need, which you have no control over.
3122 * In most situations flushing the entire workqueue is overkill; you merely
3123 * need to know that a particular work item isn't queued and isn't running.
3124 * In such cases you should use cancel_delayed_work_sync() or
3125 * cancel_work_sync() instead.
3127 void flush_scheduled_work(void)
3129 flush_workqueue(system_wq
);
3131 EXPORT_SYMBOL(flush_scheduled_work
);
3134 * execute_in_process_context - reliably execute the routine with user context
3135 * @fn: the function to execute
3136 * @ew: guaranteed storage for the execute work structure (must
3137 * be available when the work executes)
3139 * Executes the function immediately if process context is available,
3140 * otherwise schedules the function for delayed execution.
3142 * Returns: 0 - function was executed
3143 * 1 - function was scheduled for execution
3145 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3147 if (!in_interrupt()) {
3152 INIT_WORK(&ew
->work
, fn
);
3153 schedule_work(&ew
->work
);
3157 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3159 int keventd_up(void)
3161 return system_wq
!= NULL
;
3166 * Workqueues with WQ_SYSFS flag set is visible to userland via
3167 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3168 * following attributes.
3170 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3171 * max_active RW int : maximum number of in-flight work items
3173 * Unbound workqueues have the following extra attributes.
3175 * id RO int : the associated pool ID
3176 * nice RW int : nice value of the workers
3177 * cpumask RW mask : bitmask of allowed CPUs for the workers
3180 struct workqueue_struct
*wq
;
3184 static struct workqueue_struct
*dev_to_wq(struct device
*dev
)
3186 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3191 static ssize_t
wq_per_cpu_show(struct device
*dev
,
3192 struct device_attribute
*attr
, char *buf
)
3194 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3196 return scnprintf(buf
, PAGE_SIZE
, "%d\n", (bool)!(wq
->flags
& WQ_UNBOUND
));
3199 static ssize_t
wq_max_active_show(struct device
*dev
,
3200 struct device_attribute
*attr
, char *buf
)
3202 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3204 return scnprintf(buf
, PAGE_SIZE
, "%d\n", wq
->saved_max_active
);
3207 static ssize_t
wq_max_active_store(struct device
*dev
,
3208 struct device_attribute
*attr
,
3209 const char *buf
, size_t count
)
3211 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3214 if (sscanf(buf
, "%d", &val
) != 1 || val
<= 0)
3217 workqueue_set_max_active(wq
, val
);
3221 static struct device_attribute wq_sysfs_attrs
[] = {
3222 __ATTR(per_cpu
, 0444, wq_per_cpu_show
, NULL
),
3223 __ATTR(max_active
, 0644, wq_max_active_show
, wq_max_active_store
),
3227 static ssize_t
wq_pool_id_show(struct device
*dev
,
3228 struct device_attribute
*attr
, char *buf
)
3230 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3231 struct worker_pool
*pool
;
3234 rcu_read_lock_sched();
3235 pool
= first_pwq(wq
)->pool
;
3236 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n", pool
->id
);
3237 rcu_read_unlock_sched();
3242 static ssize_t
wq_nice_show(struct device
*dev
, struct device_attribute
*attr
,
3245 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3248 rcu_read_lock_sched();
3249 written
= scnprintf(buf
, PAGE_SIZE
, "%d\n",
3250 first_pwq(wq
)->pool
->attrs
->nice
);
3251 rcu_read_unlock_sched();
3256 /* prepare workqueue_attrs for sysfs store operations */
3257 static struct workqueue_attrs
*wq_sysfs_prep_attrs(struct workqueue_struct
*wq
)
3259 struct workqueue_attrs
*attrs
;
3261 attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3265 rcu_read_lock_sched();
3266 copy_workqueue_attrs(attrs
, first_pwq(wq
)->pool
->attrs
);
3267 rcu_read_unlock_sched();
3271 static ssize_t
wq_nice_store(struct device
*dev
, struct device_attribute
*attr
,
3272 const char *buf
, size_t count
)
3274 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3275 struct workqueue_attrs
*attrs
;
3278 attrs
= wq_sysfs_prep_attrs(wq
);
3282 if (sscanf(buf
, "%d", &attrs
->nice
) == 1 &&
3283 attrs
->nice
>= -20 && attrs
->nice
<= 19)
3284 ret
= apply_workqueue_attrs(wq
, attrs
);
3288 free_workqueue_attrs(attrs
);
3289 return ret
?: count
;
3292 static ssize_t
wq_cpumask_show(struct device
*dev
,
3293 struct device_attribute
*attr
, char *buf
)
3295 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3298 rcu_read_lock_sched();
3299 written
= cpumask_scnprintf(buf
, PAGE_SIZE
,
3300 first_pwq(wq
)->pool
->attrs
->cpumask
);
3301 rcu_read_unlock_sched();
3303 written
+= scnprintf(buf
+ written
, PAGE_SIZE
- written
, "\n");
3307 static ssize_t
wq_cpumask_store(struct device
*dev
,
3308 struct device_attribute
*attr
,
3309 const char *buf
, size_t count
)
3311 struct workqueue_struct
*wq
= dev_to_wq(dev
);
3312 struct workqueue_attrs
*attrs
;
3315 attrs
= wq_sysfs_prep_attrs(wq
);
3319 ret
= cpumask_parse(buf
, attrs
->cpumask
);
3321 ret
= apply_workqueue_attrs(wq
, attrs
);
3323 free_workqueue_attrs(attrs
);
3324 return ret
?: count
;
3327 static struct device_attribute wq_sysfs_unbound_attrs
[] = {
3328 __ATTR(pool_id
, 0444, wq_pool_id_show
, NULL
),
3329 __ATTR(nice
, 0644, wq_nice_show
, wq_nice_store
),
3330 __ATTR(cpumask
, 0644, wq_cpumask_show
, wq_cpumask_store
),
3334 static struct bus_type wq_subsys
= {
3335 .name
= "workqueue",
3336 .dev_attrs
= wq_sysfs_attrs
,
3339 static int __init
wq_sysfs_init(void)
3341 return subsys_virtual_register(&wq_subsys
, NULL
);
3343 core_initcall(wq_sysfs_init
);
3345 static void wq_device_release(struct device
*dev
)
3347 struct wq_device
*wq_dev
= container_of(dev
, struct wq_device
, dev
);
3353 * workqueue_sysfs_register - make a workqueue visible in sysfs
3354 * @wq: the workqueue to register
3356 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3357 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3358 * which is the preferred method.
3360 * Workqueue user should use this function directly iff it wants to apply
3361 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3362 * apply_workqueue_attrs() may race against userland updating the
3365 * Returns 0 on success, -errno on failure.
3367 int workqueue_sysfs_register(struct workqueue_struct
*wq
)
3369 struct wq_device
*wq_dev
;
3373 * Adjusting max_active or creating new pwqs by applyting
3374 * attributes breaks ordering guarantee. Disallow exposing ordered
3377 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
3380 wq
->wq_dev
= wq_dev
= kzalloc(sizeof(*wq_dev
), GFP_KERNEL
);
3385 wq_dev
->dev
.bus
= &wq_subsys
;
3386 wq_dev
->dev
.init_name
= wq
->name
;
3387 wq_dev
->dev
.release
= wq_device_release
;
3390 * unbound_attrs are created separately. Suppress uevent until
3391 * everything is ready.
3393 dev_set_uevent_suppress(&wq_dev
->dev
, true);
3395 ret
= device_register(&wq_dev
->dev
);
3402 if (wq
->flags
& WQ_UNBOUND
) {
3403 struct device_attribute
*attr
;
3405 for (attr
= wq_sysfs_unbound_attrs
; attr
->attr
.name
; attr
++) {
3406 ret
= device_create_file(&wq_dev
->dev
, attr
);
3408 device_unregister(&wq_dev
->dev
);
3415 kobject_uevent(&wq_dev
->dev
.kobj
, KOBJ_ADD
);
3420 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3421 * @wq: the workqueue to unregister
3423 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3425 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
)
3427 struct wq_device
*wq_dev
= wq
->wq_dev
;
3433 device_unregister(&wq_dev
->dev
);
3435 #else /* CONFIG_SYSFS */
3436 static void workqueue_sysfs_unregister(struct workqueue_struct
*wq
) { }
3437 #endif /* CONFIG_SYSFS */
3440 * free_workqueue_attrs - free a workqueue_attrs
3441 * @attrs: workqueue_attrs to free
3443 * Undo alloc_workqueue_attrs().
3445 void free_workqueue_attrs(struct workqueue_attrs
*attrs
)
3448 free_cpumask_var(attrs
->cpumask
);
3454 * alloc_workqueue_attrs - allocate a workqueue_attrs
3455 * @gfp_mask: allocation mask to use
3457 * Allocate a new workqueue_attrs, initialize with default settings and
3458 * return it. Returns NULL on failure.
3460 struct workqueue_attrs
*alloc_workqueue_attrs(gfp_t gfp_mask
)
3462 struct workqueue_attrs
*attrs
;
3464 attrs
= kzalloc(sizeof(*attrs
), gfp_mask
);
3467 if (!alloc_cpumask_var(&attrs
->cpumask
, gfp_mask
))
3470 cpumask_setall(attrs
->cpumask
);
3473 free_workqueue_attrs(attrs
);
3477 static void copy_workqueue_attrs(struct workqueue_attrs
*to
,
3478 const struct workqueue_attrs
*from
)
3480 to
->nice
= from
->nice
;
3481 cpumask_copy(to
->cpumask
, from
->cpumask
);
3485 * Hacky implementation of jhash of bitmaps which only considers the
3486 * specified number of bits. We probably want a proper implementation in
3487 * include/linux/jhash.h.
3489 static u32
jhash_bitmap(const unsigned long *bitmap
, int bits
, u32 hash
)
3491 int nr_longs
= bits
/ BITS_PER_LONG
;
3492 int nr_leftover
= bits
% BITS_PER_LONG
;
3493 unsigned long leftover
= 0;
3496 hash
= jhash(bitmap
, nr_longs
* sizeof(long), hash
);
3498 bitmap_copy(&leftover
, bitmap
+ nr_longs
, nr_leftover
);
3499 hash
= jhash(&leftover
, sizeof(long), hash
);
3504 /* hash value of the content of @attr */
3505 static u32
wqattrs_hash(const struct workqueue_attrs
*attrs
)
3509 hash
= jhash_1word(attrs
->nice
, hash
);
3510 hash
= jhash_bitmap(cpumask_bits(attrs
->cpumask
), nr_cpu_ids
, hash
);
3514 /* content equality test */
3515 static bool wqattrs_equal(const struct workqueue_attrs
*a
,
3516 const struct workqueue_attrs
*b
)
3518 if (a
->nice
!= b
->nice
)
3520 if (!cpumask_equal(a
->cpumask
, b
->cpumask
))
3526 * init_worker_pool - initialize a newly zalloc'd worker_pool
3527 * @pool: worker_pool to initialize
3529 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3530 * Returns 0 on success, -errno on failure. Even on failure, all fields
3531 * inside @pool proper are initialized and put_unbound_pool() can be called
3532 * on @pool safely to release it.
3534 static int init_worker_pool(struct worker_pool
*pool
)
3536 spin_lock_init(&pool
->lock
);
3539 pool
->flags
|= POOL_DISASSOCIATED
;
3540 INIT_LIST_HEAD(&pool
->worklist
);
3541 INIT_LIST_HEAD(&pool
->idle_list
);
3542 hash_init(pool
->busy_hash
);
3544 init_timer_deferrable(&pool
->idle_timer
);
3545 pool
->idle_timer
.function
= idle_worker_timeout
;
3546 pool
->idle_timer
.data
= (unsigned long)pool
;
3548 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3549 (unsigned long)pool
);
3551 mutex_init(&pool
->manager_arb
);
3552 mutex_init(&pool
->assoc_mutex
);
3553 ida_init(&pool
->worker_ida
);
3555 INIT_HLIST_NODE(&pool
->hash_node
);
3558 /* shouldn't fail above this point */
3559 pool
->attrs
= alloc_workqueue_attrs(GFP_KERNEL
);
3565 static void rcu_free_pool(struct rcu_head
*rcu
)
3567 struct worker_pool
*pool
= container_of(rcu
, struct worker_pool
, rcu
);
3569 ida_destroy(&pool
->worker_ida
);
3570 free_workqueue_attrs(pool
->attrs
);
3575 * put_unbound_pool - put a worker_pool
3576 * @pool: worker_pool to put
3578 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3581 static void put_unbound_pool(struct worker_pool
*pool
)
3583 struct worker
*worker
;
3585 spin_lock_irq(&workqueue_lock
);
3586 if (--pool
->refcnt
) {
3587 spin_unlock_irq(&workqueue_lock
);
3592 if (WARN_ON(!(pool
->flags
& POOL_DISASSOCIATED
)) ||
3593 WARN_ON(!list_empty(&pool
->worklist
))) {
3594 spin_unlock_irq(&workqueue_lock
);
3598 /* release id and unhash */
3600 idr_remove(&worker_pool_idr
, pool
->id
);
3601 hash_del(&pool
->hash_node
);
3603 spin_unlock_irq(&workqueue_lock
);
3605 /* lock out manager and destroy all workers */
3606 mutex_lock(&pool
->manager_arb
);
3607 spin_lock_irq(&pool
->lock
);
3609 while ((worker
= first_worker(pool
)))
3610 destroy_worker(worker
);
3611 WARN_ON(pool
->nr_workers
|| pool
->nr_idle
);
3613 spin_unlock_irq(&pool
->lock
);
3614 mutex_unlock(&pool
->manager_arb
);
3616 /* shut down the timers */
3617 del_timer_sync(&pool
->idle_timer
);
3618 del_timer_sync(&pool
->mayday_timer
);
3620 /* sched-RCU protected to allow dereferences from get_work_pool() */
3621 call_rcu_sched(&pool
->rcu
, rcu_free_pool
);
3625 * get_unbound_pool - get a worker_pool with the specified attributes
3626 * @attrs: the attributes of the worker_pool to get
3628 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3629 * reference count and return it. If there already is a matching
3630 * worker_pool, it will be used; otherwise, this function attempts to
3631 * create a new one. On failure, returns NULL.
3633 static struct worker_pool
*get_unbound_pool(const struct workqueue_attrs
*attrs
)
3635 static DEFINE_MUTEX(create_mutex
);
3636 u32 hash
= wqattrs_hash(attrs
);
3637 struct worker_pool
*pool
;
3638 struct worker
*worker
;
3640 mutex_lock(&create_mutex
);
3642 /* do we already have a matching pool? */
3643 spin_lock_irq(&workqueue_lock
);
3644 hash_for_each_possible(unbound_pool_hash
, pool
, hash_node
, hash
) {
3645 if (wqattrs_equal(pool
->attrs
, attrs
)) {
3650 spin_unlock_irq(&workqueue_lock
);
3652 /* nope, create a new one */
3653 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
3654 if (!pool
|| init_worker_pool(pool
) < 0)
3657 lockdep_set_subclass(&pool
->lock
, 1); /* see put_pwq() */
3658 copy_workqueue_attrs(pool
->attrs
, attrs
);
3660 if (worker_pool_assign_id(pool
) < 0)
3663 /* create and start the initial worker */
3664 worker
= create_worker(pool
);
3668 spin_lock_irq(&pool
->lock
);
3669 start_worker(worker
);
3670 spin_unlock_irq(&pool
->lock
);
3673 spin_lock_irq(&workqueue_lock
);
3674 hash_add(unbound_pool_hash
, &pool
->hash_node
, hash
);
3676 spin_unlock_irq(&workqueue_lock
);
3677 mutex_unlock(&create_mutex
);
3680 mutex_unlock(&create_mutex
);
3682 put_unbound_pool(pool
);
3686 static void rcu_free_pwq(struct rcu_head
*rcu
)
3688 kmem_cache_free(pwq_cache
,
3689 container_of(rcu
, struct pool_workqueue
, rcu
));
3693 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3694 * and needs to be destroyed.
3696 static void pwq_unbound_release_workfn(struct work_struct
*work
)
3698 struct pool_workqueue
*pwq
= container_of(work
, struct pool_workqueue
,
3699 unbound_release_work
);
3700 struct workqueue_struct
*wq
= pwq
->wq
;
3701 struct worker_pool
*pool
= pwq
->pool
;
3703 if (WARN_ON_ONCE(!(wq
->flags
& WQ_UNBOUND
)))
3707 * Unlink @pwq. Synchronization against flush_mutex isn't strictly
3708 * necessary on release but do it anyway. It's easier to verify
3709 * and consistent with the linking path.
3711 mutex_lock(&wq
->flush_mutex
);
3712 spin_lock_irq(&workqueue_lock
);
3713 list_del_rcu(&pwq
->pwqs_node
);
3714 spin_unlock_irq(&workqueue_lock
);
3715 mutex_unlock(&wq
->flush_mutex
);
3717 put_unbound_pool(pool
);
3718 call_rcu_sched(&pwq
->rcu
, rcu_free_pwq
);
3721 * If we're the last pwq going away, @wq is already dead and no one
3722 * is gonna access it anymore. Free it.
3724 if (list_empty(&wq
->pwqs
))
3729 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3730 * @pwq: target pool_workqueue
3732 * If @pwq isn't freezing, set @pwq->max_active to the associated
3733 * workqueue's saved_max_active and activate delayed work items
3734 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3736 static void pwq_adjust_max_active(struct pool_workqueue
*pwq
)
3738 struct workqueue_struct
*wq
= pwq
->wq
;
3739 bool freezable
= wq
->flags
& WQ_FREEZABLE
;
3741 /* for @wq->saved_max_active */
3742 lockdep_assert_held(&workqueue_lock
);
3744 /* fast exit for non-freezable wqs */
3745 if (!freezable
&& pwq
->max_active
== wq
->saved_max_active
)
3748 spin_lock(&pwq
->pool
->lock
);
3750 if (!freezable
|| !(pwq
->pool
->flags
& POOL_FREEZING
)) {
3751 pwq
->max_active
= wq
->saved_max_active
;
3753 while (!list_empty(&pwq
->delayed_works
) &&
3754 pwq
->nr_active
< pwq
->max_active
)
3755 pwq_activate_first_delayed(pwq
);
3757 pwq
->max_active
= 0;
3760 spin_unlock(&pwq
->pool
->lock
);
3763 static void init_and_link_pwq(struct pool_workqueue
*pwq
,
3764 struct workqueue_struct
*wq
,
3765 struct worker_pool
*pool
,
3766 struct pool_workqueue
**p_last_pwq
)
3768 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3772 pwq
->flush_color
= -1;
3774 pwq
->max_active
= wq
->saved_max_active
;
3775 INIT_LIST_HEAD(&pwq
->delayed_works
);
3776 INIT_LIST_HEAD(&pwq
->mayday_node
);
3777 INIT_WORK(&pwq
->unbound_release_work
, pwq_unbound_release_workfn
);
3780 * Link @pwq and set the matching work_color. This is synchronized
3781 * with flush_mutex to avoid confusing flush_workqueue().
3783 mutex_lock(&wq
->flush_mutex
);
3784 spin_lock_irq(&workqueue_lock
);
3787 *p_last_pwq
= first_pwq(wq
);
3788 pwq
->work_color
= wq
->work_color
;
3789 list_add_rcu(&pwq
->pwqs_node
, &wq
->pwqs
);
3791 spin_unlock_irq(&workqueue_lock
);
3792 mutex_unlock(&wq
->flush_mutex
);
3796 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3797 * @wq: the target workqueue
3798 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3800 * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the
3801 * current attributes, a new pwq is created and made the first pwq which
3802 * will serve all new work items. Older pwqs are released as in-flight
3803 * work items finish. Note that a work item which repeatedly requeues
3804 * itself back-to-back will stay on its current pwq.
3806 * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on
3809 int apply_workqueue_attrs(struct workqueue_struct
*wq
,
3810 const struct workqueue_attrs
*attrs
)
3812 struct pool_workqueue
*pwq
, *last_pwq
;
3813 struct worker_pool
*pool
;
3815 /* only unbound workqueues can change attributes */
3816 if (WARN_ON(!(wq
->flags
& WQ_UNBOUND
)))
3819 /* creating multiple pwqs breaks ordering guarantee */
3820 if (WARN_ON((wq
->flags
& __WQ_ORDERED
) && !list_empty(&wq
->pwqs
)))
3823 pwq
= kmem_cache_zalloc(pwq_cache
, GFP_KERNEL
);
3827 pool
= get_unbound_pool(attrs
);
3829 kmem_cache_free(pwq_cache
, pwq
);
3833 init_and_link_pwq(pwq
, wq
, pool
, &last_pwq
);
3835 spin_lock_irq(&last_pwq
->pool
->lock
);
3837 spin_unlock_irq(&last_pwq
->pool
->lock
);
3843 static int alloc_and_link_pwqs(struct workqueue_struct
*wq
)
3845 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
3848 if (!(wq
->flags
& WQ_UNBOUND
)) {
3849 wq
->cpu_pwqs
= alloc_percpu(struct pool_workqueue
);
3853 for_each_possible_cpu(cpu
) {
3854 struct pool_workqueue
*pwq
=
3855 per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
3856 struct worker_pool
*cpu_pools
=
3857 per_cpu(cpu_worker_pools
, cpu
);
3859 init_and_link_pwq(pwq
, wq
, &cpu_pools
[highpri
], NULL
);
3863 return apply_workqueue_attrs(wq
, unbound_std_wq_attrs
[highpri
]);
3867 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3870 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3872 if (max_active
< 1 || max_active
> lim
)
3873 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3874 max_active
, name
, 1, lim
);
3876 return clamp_val(max_active
, 1, lim
);
3879 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3882 struct lock_class_key
*key
,
3883 const char *lock_name
, ...)
3885 va_list args
, args1
;
3886 struct workqueue_struct
*wq
;
3887 struct pool_workqueue
*pwq
;
3890 /* determine namelen, allocate wq and format name */
3891 va_start(args
, lock_name
);
3892 va_copy(args1
, args
);
3893 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
3895 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
3899 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
3903 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3904 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3908 wq
->saved_max_active
= max_active
;
3909 mutex_init(&wq
->flush_mutex
);
3910 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
3911 INIT_LIST_HEAD(&wq
->pwqs
);
3912 INIT_LIST_HEAD(&wq
->flusher_queue
);
3913 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3914 INIT_LIST_HEAD(&wq
->maydays
);
3916 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3917 INIT_LIST_HEAD(&wq
->list
);
3919 if (alloc_and_link_pwqs(wq
) < 0)
3923 * Workqueues which may be used during memory reclaim should
3924 * have a rescuer to guarantee forward progress.
3926 if (flags
& WQ_MEM_RECLAIM
) {
3927 struct worker
*rescuer
;
3929 rescuer
= alloc_worker();
3933 rescuer
->rescue_wq
= wq
;
3934 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3936 if (IS_ERR(rescuer
->task
)) {
3941 wq
->rescuer
= rescuer
;
3942 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
3943 wake_up_process(rescuer
->task
);
3946 if ((wq
->flags
& WQ_SYSFS
) && workqueue_sysfs_register(wq
))
3950 * workqueue_lock protects global freeze state and workqueues list.
3951 * Grab it, adjust max_active and add the new workqueue to
3954 spin_lock_irq(&workqueue_lock
);
3956 for_each_pwq(pwq
, wq
)
3957 pwq_adjust_max_active(pwq
);
3959 list_add(&wq
->list
, &workqueues
);
3961 spin_unlock_irq(&workqueue_lock
);
3969 destroy_workqueue(wq
);
3972 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3975 * destroy_workqueue - safely terminate a workqueue
3976 * @wq: target workqueue
3978 * Safely destroy a workqueue. All work currently pending will be done first.
3980 void destroy_workqueue(struct workqueue_struct
*wq
)
3982 struct pool_workqueue
*pwq
;
3984 /* drain it before proceeding with destruction */
3985 drain_workqueue(wq
);
3987 spin_lock_irq(&workqueue_lock
);
3990 for_each_pwq(pwq
, wq
) {
3993 for (i
= 0; i
< WORK_NR_COLORS
; i
++) {
3994 if (WARN_ON(pwq
->nr_in_flight
[i
])) {
3995 spin_unlock_irq(&workqueue_lock
);
4000 if (WARN_ON(pwq
->refcnt
> 1) ||
4001 WARN_ON(pwq
->nr_active
) ||
4002 WARN_ON(!list_empty(&pwq
->delayed_works
))) {
4003 spin_unlock_irq(&workqueue_lock
);
4009 * wq list is used to freeze wq, remove from list after
4010 * flushing is complete in case freeze races us.
4012 list_del_init(&wq
->list
);
4014 spin_unlock_irq(&workqueue_lock
);
4016 workqueue_sysfs_unregister(wq
);
4019 kthread_stop(wq
->rescuer
->task
);
4024 if (!(wq
->flags
& WQ_UNBOUND
)) {
4026 * The base ref is never dropped on per-cpu pwqs. Directly
4027 * free the pwqs and wq.
4029 free_percpu(wq
->cpu_pwqs
);
4033 * We're the sole accessor of @wq at this point. Directly
4034 * access the first pwq and put the base ref. As both pwqs
4035 * and pools are sched-RCU protected, the lock operations
4036 * are safe. @wq will be freed when the last pwq is
4039 pwq
= list_first_entry(&wq
->pwqs
, struct pool_workqueue
,
4041 spin_lock_irq(&pwq
->pool
->lock
);
4043 spin_unlock_irq(&pwq
->pool
->lock
);
4046 EXPORT_SYMBOL_GPL(destroy_workqueue
);
4049 * workqueue_set_max_active - adjust max_active of a workqueue
4050 * @wq: target workqueue
4051 * @max_active: new max_active value.
4053 * Set max_active of @wq to @max_active.
4056 * Don't call from IRQ context.
4058 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
4060 struct pool_workqueue
*pwq
;
4062 /* disallow meddling with max_active for ordered workqueues */
4063 if (WARN_ON(wq
->flags
& __WQ_ORDERED
))
4066 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
4068 spin_lock_irq(&workqueue_lock
);
4070 wq
->saved_max_active
= max_active
;
4072 for_each_pwq(pwq
, wq
)
4073 pwq_adjust_max_active(pwq
);
4075 spin_unlock_irq(&workqueue_lock
);
4077 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
4080 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4082 * Determine whether %current is a workqueue rescuer. Can be used from
4083 * work functions to determine whether it's being run off the rescuer task.
4085 bool current_is_workqueue_rescuer(void)
4087 struct worker
*worker
= current_wq_worker();
4089 return worker
&& worker
== worker
->current_pwq
->wq
->rescuer
;
4093 * workqueue_congested - test whether a workqueue is congested
4094 * @cpu: CPU in question
4095 * @wq: target workqueue
4097 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4098 * no synchronization around this function and the test result is
4099 * unreliable and only useful as advisory hints or for debugging.
4102 * %true if congested, %false otherwise.
4104 bool workqueue_congested(int cpu
, struct workqueue_struct
*wq
)
4106 struct pool_workqueue
*pwq
;
4111 if (!(wq
->flags
& WQ_UNBOUND
))
4112 pwq
= per_cpu_ptr(wq
->cpu_pwqs
, cpu
);
4114 pwq
= first_pwq(wq
);
4116 ret
= !list_empty(&pwq
->delayed_works
);
4121 EXPORT_SYMBOL_GPL(workqueue_congested
);
4124 * work_busy - test whether a work is currently pending or running
4125 * @work: the work to be tested
4127 * Test whether @work is currently pending or running. There is no
4128 * synchronization around this function and the test result is
4129 * unreliable and only useful as advisory hints or for debugging.
4132 * OR'd bitmask of WORK_BUSY_* bits.
4134 unsigned int work_busy(struct work_struct
*work
)
4136 struct worker_pool
*pool
;
4137 unsigned long flags
;
4138 unsigned int ret
= 0;
4140 if (work_pending(work
))
4141 ret
|= WORK_BUSY_PENDING
;
4143 local_irq_save(flags
);
4144 pool
= get_work_pool(work
);
4146 spin_lock(&pool
->lock
);
4147 if (find_worker_executing_work(pool
, work
))
4148 ret
|= WORK_BUSY_RUNNING
;
4149 spin_unlock(&pool
->lock
);
4151 local_irq_restore(flags
);
4155 EXPORT_SYMBOL_GPL(work_busy
);
4160 * There are two challenges in supporting CPU hotplug. Firstly, there
4161 * are a lot of assumptions on strong associations among work, pwq and
4162 * pool which make migrating pending and scheduled works very
4163 * difficult to implement without impacting hot paths. Secondly,
4164 * worker pools serve mix of short, long and very long running works making
4165 * blocked draining impractical.
4167 * This is solved by allowing the pools to be disassociated from the CPU
4168 * running as an unbound one and allowing it to be reattached later if the
4169 * cpu comes back online.
4172 static void wq_unbind_fn(struct work_struct
*work
)
4174 int cpu
= smp_processor_id();
4175 struct worker_pool
*pool
;
4176 struct worker
*worker
;
4179 for_each_cpu_worker_pool(pool
, cpu
) {
4180 WARN_ON_ONCE(cpu
!= smp_processor_id());
4182 mutex_lock(&pool
->assoc_mutex
);
4183 spin_lock_irq(&pool
->lock
);
4186 * We've claimed all manager positions. Make all workers
4187 * unbound and set DISASSOCIATED. Before this, all workers
4188 * except for the ones which are still executing works from
4189 * before the last CPU down must be on the cpu. After
4190 * this, they may become diasporas.
4192 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
4193 worker
->flags
|= WORKER_UNBOUND
;
4195 for_each_busy_worker(worker
, i
, pool
)
4196 worker
->flags
|= WORKER_UNBOUND
;
4198 pool
->flags
|= POOL_DISASSOCIATED
;
4200 spin_unlock_irq(&pool
->lock
);
4201 mutex_unlock(&pool
->assoc_mutex
);
4205 * Call schedule() so that we cross rq->lock and thus can guarantee
4206 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
4207 * as scheduler callbacks may be invoked from other cpus.
4212 * Sched callbacks are disabled now. Zap nr_running. After this,
4213 * nr_running stays zero and need_more_worker() and keep_working()
4214 * are always true as long as the worklist is not empty. Pools on
4215 * @cpu now behave as unbound (in terms of concurrency management)
4216 * pools which are served by workers tied to the CPU.
4218 * On return from this function, the current worker would trigger
4219 * unbound chain execution of pending work items if other workers
4222 for_each_cpu_worker_pool(pool
, cpu
)
4223 atomic_set(&pool
->nr_running
, 0);
4227 * Workqueues should be brought up before normal priority CPU notifiers.
4228 * This will be registered high priority CPU notifier.
4230 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
4231 unsigned long action
,
4234 int cpu
= (unsigned long)hcpu
;
4235 struct worker_pool
*pool
;
4237 switch (action
& ~CPU_TASKS_FROZEN
) {
4238 case CPU_UP_PREPARE
:
4239 for_each_cpu_worker_pool(pool
, cpu
) {
4240 struct worker
*worker
;
4242 if (pool
->nr_workers
)
4245 worker
= create_worker(pool
);
4249 spin_lock_irq(&pool
->lock
);
4250 start_worker(worker
);
4251 spin_unlock_irq(&pool
->lock
);
4255 case CPU_DOWN_FAILED
:
4257 for_each_cpu_worker_pool(pool
, cpu
) {
4258 mutex_lock(&pool
->assoc_mutex
);
4259 spin_lock_irq(&pool
->lock
);
4261 pool
->flags
&= ~POOL_DISASSOCIATED
;
4262 rebind_workers(pool
);
4264 spin_unlock_irq(&pool
->lock
);
4265 mutex_unlock(&pool
->assoc_mutex
);
4273 * Workqueues should be brought down after normal priority CPU notifiers.
4274 * This will be registered as low priority CPU notifier.
4276 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
4277 unsigned long action
,
4280 int cpu
= (unsigned long)hcpu
;
4281 struct work_struct unbind_work
;
4283 switch (action
& ~CPU_TASKS_FROZEN
) {
4284 case CPU_DOWN_PREPARE
:
4285 /* unbinding should happen on the local CPU */
4286 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
4287 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
4288 flush_work(&unbind_work
);
4296 struct work_for_cpu
{
4297 struct work_struct work
;
4303 static void work_for_cpu_fn(struct work_struct
*work
)
4305 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
4307 wfc
->ret
= wfc
->fn(wfc
->arg
);
4311 * work_on_cpu - run a function in user context on a particular cpu
4312 * @cpu: the cpu to run on
4313 * @fn: the function to run
4314 * @arg: the function arg
4316 * This will return the value @fn returns.
4317 * It is up to the caller to ensure that the cpu doesn't go offline.
4318 * The caller must not hold any locks which would prevent @fn from completing.
4320 long work_on_cpu(int cpu
, long (*fn
)(void *), void *arg
)
4322 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
4324 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
4325 schedule_work_on(cpu
, &wfc
.work
);
4326 flush_work(&wfc
.work
);
4329 EXPORT_SYMBOL_GPL(work_on_cpu
);
4330 #endif /* CONFIG_SMP */
4332 #ifdef CONFIG_FREEZER
4335 * freeze_workqueues_begin - begin freezing workqueues
4337 * Start freezing workqueues. After this function returns, all freezable
4338 * workqueues will queue new works to their frozen_works list instead of
4342 * Grabs and releases workqueue_lock and pool->lock's.
4344 void freeze_workqueues_begin(void)
4346 struct worker_pool
*pool
;
4347 struct workqueue_struct
*wq
;
4348 struct pool_workqueue
*pwq
;
4351 spin_lock_irq(&workqueue_lock
);
4353 WARN_ON_ONCE(workqueue_freezing
);
4354 workqueue_freezing
= true;
4357 for_each_pool(pool
, id
) {
4358 spin_lock(&pool
->lock
);
4359 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
4360 pool
->flags
|= POOL_FREEZING
;
4361 spin_unlock(&pool
->lock
);
4364 /* suppress further executions by setting max_active to zero */
4365 list_for_each_entry(wq
, &workqueues
, list
) {
4366 for_each_pwq(pwq
, wq
)
4367 pwq_adjust_max_active(pwq
);
4370 spin_unlock_irq(&workqueue_lock
);
4374 * freeze_workqueues_busy - are freezable workqueues still busy?
4376 * Check whether freezing is complete. This function must be called
4377 * between freeze_workqueues_begin() and thaw_workqueues().
4380 * Grabs and releases workqueue_lock.
4383 * %true if some freezable workqueues are still busy. %false if freezing
4386 bool freeze_workqueues_busy(void)
4389 struct workqueue_struct
*wq
;
4390 struct pool_workqueue
*pwq
;
4392 spin_lock_irq(&workqueue_lock
);
4394 WARN_ON_ONCE(!workqueue_freezing
);
4396 list_for_each_entry(wq
, &workqueues
, list
) {
4397 if (!(wq
->flags
& WQ_FREEZABLE
))
4400 * nr_active is monotonically decreasing. It's safe
4401 * to peek without lock.
4403 for_each_pwq(pwq
, wq
) {
4404 WARN_ON_ONCE(pwq
->nr_active
< 0);
4405 if (pwq
->nr_active
) {
4412 spin_unlock_irq(&workqueue_lock
);
4417 * thaw_workqueues - thaw workqueues
4419 * Thaw workqueues. Normal queueing is restored and all collected
4420 * frozen works are transferred to their respective pool worklists.
4423 * Grabs and releases workqueue_lock and pool->lock's.
4425 void thaw_workqueues(void)
4427 struct workqueue_struct
*wq
;
4428 struct pool_workqueue
*pwq
;
4429 struct worker_pool
*pool
;
4432 spin_lock_irq(&workqueue_lock
);
4434 if (!workqueue_freezing
)
4437 /* clear FREEZING */
4438 for_each_pool(pool
, id
) {
4439 spin_lock(&pool
->lock
);
4440 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
4441 pool
->flags
&= ~POOL_FREEZING
;
4442 spin_unlock(&pool
->lock
);
4445 /* restore max_active and repopulate worklist */
4446 list_for_each_entry(wq
, &workqueues
, list
) {
4447 for_each_pwq(pwq
, wq
)
4448 pwq_adjust_max_active(pwq
);
4452 for_each_pool(pool
, id
) {
4453 spin_lock(&pool
->lock
);
4454 wake_up_worker(pool
);
4455 spin_unlock(&pool
->lock
);
4458 workqueue_freezing
= false;
4460 spin_unlock_irq(&workqueue_lock
);
4462 #endif /* CONFIG_FREEZER */
4464 static int __init
init_workqueues(void)
4466 int std_nice
[NR_STD_WORKER_POOLS
] = { 0, HIGHPRI_NICE_LEVEL
};
4469 /* make sure we have enough bits for OFFQ pool ID */
4470 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
4471 WORK_CPU_END
* NR_STD_WORKER_POOLS
);
4473 WARN_ON(__alignof__(struct pool_workqueue
) < __alignof__(long long));
4475 pwq_cache
= KMEM_CACHE(pool_workqueue
, SLAB_PANIC
);
4477 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
4478 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
4480 /* initialize CPU pools */
4481 for_each_possible_cpu(cpu
) {
4482 struct worker_pool
*pool
;
4485 for_each_cpu_worker_pool(pool
, cpu
) {
4486 BUG_ON(init_worker_pool(pool
));
4488 cpumask_copy(pool
->attrs
->cpumask
, cpumask_of(cpu
));
4489 pool
->attrs
->nice
= std_nice
[i
++];
4492 BUG_ON(worker_pool_assign_id(pool
));
4496 /* create the initial worker */
4497 for_each_online_cpu(cpu
) {
4498 struct worker_pool
*pool
;
4500 for_each_cpu_worker_pool(pool
, cpu
) {
4501 struct worker
*worker
;
4503 pool
->flags
&= ~POOL_DISASSOCIATED
;
4505 worker
= create_worker(pool
);
4507 spin_lock_irq(&pool
->lock
);
4508 start_worker(worker
);
4509 spin_unlock_irq(&pool
->lock
);
4513 /* create default unbound wq attrs */
4514 for (i
= 0; i
< NR_STD_WORKER_POOLS
; i
++) {
4515 struct workqueue_attrs
*attrs
;
4517 BUG_ON(!(attrs
= alloc_workqueue_attrs(GFP_KERNEL
)));
4519 attrs
->nice
= std_nice
[i
];
4520 cpumask_setall(attrs
->cpumask
);
4522 unbound_std_wq_attrs
[i
] = attrs
;
4525 system_wq
= alloc_workqueue("events", 0, 0);
4526 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
4527 system_long_wq
= alloc_workqueue("events_long", 0, 0);
4528 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
4529 WQ_UNBOUND_MAX_ACTIVE
);
4530 system_freezable_wq
= alloc_workqueue("events_freezable",
4532 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
4533 !system_unbound_wq
|| !system_freezable_wq
);
4536 early_initcall(init_workqueues
);